Last active
October 25, 2016 23:00
-
-
Save moskewcz/25d5d545e2ce9f1819a02238f3e5dfd0 to your computer and use it in GitHub Desktop.
This file has been truncated, but you can view the full file.
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),filts=(dims=(out_chan=16,in_chan=192,y=1,x=1)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=1,chan=16,y=28,x=28)),out_chans=(tn=uint32_t,v=16),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00073968 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=192,y=1,x=1,out_chan=16)),in=(dims=(chan=192,pel=784)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=1,y=28,x=28)),out=(dims=(chan=16,pel=784)),out_pels=(tn=none,dims=(img=1,y=28,x=28)),out_ref=(dims=(img=1,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=98,out_chan_tile=2,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000117024 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=192,y=1,x=1,out_chan=16)),in=(dims=(chan=192,pel=784)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=1,y=28,x=28)),out=(dims=(chan=16,pel=784)),out_pels=(tn=none,dims=(img=1,y=28,x=28)),out_ref=(dims=(img=1,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=98,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00087296 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=7,blk_in_chan=192,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=192,y=28,x=28)),out=(dims=(img=1,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=7,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000176832 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=7,blk_in_chan=192,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=192,y=28,x=28)),out=(dims=(img=1,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=7,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00087184 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=7,blk_in_chan=192,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=192,y=28,x=28)),out=(dims=(img=1,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=7,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000177376 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=7,blk_in_chan=192,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=192,y=28,x=28)),out=(dims=(img=1,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=7,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00085888 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000151296 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00085008 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000150912 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00247568 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00038224 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00115456 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=4,blk_in_chan=192,blk_y=64,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=192,y=28,x=28)),out=(dims=(img=1,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=4,out_chan_blk=1,blk_y=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000204416 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=4,blk_in_chan=192,blk_y=64,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=192,y=28,x=28)),out=(dims=(img=1,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=4,out_chan_blk=1,blk_y=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000145728 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000177184 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=7,blk_in_chan=192,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=192,y=28,x=28)),out=(dims=(img=1,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=7,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000176352 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=7,blk_in_chan=192,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=192,y=28,x=28)),out=(dims=(img=1,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=7,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000146656 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00014672 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00037408 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000208576 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=4,blk_in_chan=192,blk_y=64,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=192,y=28,x=28)),out=(dims=(img=1,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=4,out_chan_blk=1,blk_y=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),filts=(dims=(out_chan=16,in_chan=192,y=1,x=1)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=5,chan=16,y=28,x=28)),out_chans=(tn=uint32_t,v=16),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0007344 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=192,y=1,x=1,out_chan=16)),in=(dims=(chan=192,pel=3920)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=5,y=28,x=28)),out=(dims=(chan=16,pel=4096)),out_pels=(tn=none,dims=(img=5,y=28,x=28)),out_ref=(dims=(img=5,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=128,out_chan_tile=2,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000146336 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=192,y=1,x=1,out_chan=16)),in=(dims=(chan=192,pel=3920)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=5,y=28,x=28)),out=(dims=(chan=16,pel=4096)),out_pels=(tn=none,dims=(img=5,y=28,x=28)),out_ref=(dims=(img=5,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=128,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00090864 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=9,blk_bx=7,blk_in_chan=192,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=192,y=28,x=28)),out=(dims=(img=5,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=9,blk_bx=7,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000208 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=9,blk_bx=7,blk_in_chan=192,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=192,y=28,x=28)),out=(dims=(img=5,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=9,blk_bx=7,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0009096 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=9,blk_bx=7,blk_in_chan=192,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=192,y=28,x=28)),out=(dims=(img=5,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=9,blk_bx=7,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000206816 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=9,blk_bx=7,blk_in_chan=192,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=192,y=28,x=28)),out=(dims=(img=5,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=9,blk_bx=7,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00096416 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000158432 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00095744 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00015872 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0028272 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000386528 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0011992 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),flags=(tn=uint32_t),in=(dims=(blk_bline=3,blk_bx=4,blk_in_chan=192,blk_y=64,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=192,y=28,x=28)),out=(dims=(img=5,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=3,blk_bx=4,out_chan_blk=1,blk_y=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000248352 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),flags=(tn=uint32_t),in=(dims=(blk_bline=3,blk_bx=4,blk_in_chan=192,blk_y=64,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=192,y=28,x=28)),out=(dims=(img=5,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=3,blk_bx=4,out_chan_blk=1,blk_y=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 9.7248e-05 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00020464 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=9,blk_bx=7,blk_in_chan=192,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=192,y=28,x=28)),out=(dims=(img=5,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=9,blk_bx=7,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000204416 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=9,blk_bx=7,blk_in_chan=192,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=192,y=28,x=28)),out=(dims=(img=5,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=9,blk_bx=7,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000152192 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000151488 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000378304 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000254304 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),flags=(tn=uint32_t),in=(dims=(blk_bline=3,blk_bx=4,blk_in_chan=192,blk_y=64,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=192,y=28,x=28)),out=(dims=(img=5,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=3,blk_bx=4,out_chan_blk=1,blk_y=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),filts=(dims=(out_chan=16,in_chan=192,y=1,x=1)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=20,chan=16,y=28,x=28)),out_chans=(tn=uint32_t,v=16),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00072832 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=192,y=1,x=1,out_chan=16)),in=(dims=(chan=192,pel=15680)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=20,y=28,x=28)),out=(dims=(chan=16,pel=16384)),out_pels=(tn=none,dims=(img=20,y=28,x=28)),out_ref=(dims=(img=20,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=128,out_chan_tile=2,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000270272 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=192,y=1,x=1,out_chan=16)),in=(dims=(chan=192,pel=15680)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=20,y=28,x=28)),out=(dims=(chan=16,pel=16384)),out_pels=(tn=none,dims=(img=20,y=28,x=28)),out_ref=(dims=(img=20,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=128,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00102704 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=7,blk_in_chan=192,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=192,y=28,x=28)),out=(dims=(img=20,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=35,blk_bx=7,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000256736 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=7,blk_in_chan=192,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=192,y=28,x=28)),out=(dims=(img=20,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=35,blk_bx=7,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00062048 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=7,blk_in_chan=192,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=192,y=28,x=28)),out=(dims=(img=20,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=35,blk_bx=7,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000257152 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=7,blk_in_chan=192,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=192,y=28,x=28)),out=(dims=(img=20,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=35,blk_bx=7,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00107824 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00023056 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00107792 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000229152 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.002792 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000516672 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00127536 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),flags=(tn=uint32_t),in=(dims=(blk_bline=9,blk_bx=4,blk_in_chan=192,blk_y=64,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=192,y=28,x=28)),out=(dims=(img=20,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=9,blk_bx=4,out_chan_blk=1,blk_y=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00029072 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),flags=(tn=uint32_t),in=(dims=(blk_bline=9,blk_bx=4,blk_in_chan=192,blk_y=64,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=192,y=28,x=28)),out=(dims=(img=20,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=9,blk_bx=4,out_chan_blk=1,blk_y=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000125632 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000255648 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=7,blk_in_chan=192,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=192,y=28,x=28)),out=(dims=(img=20,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=35,blk_bx=7,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00025664 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=7,blk_in_chan=192,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=192,y=28,x=28)),out=(dims=(img=20,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=35,blk_bx=7,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000227072 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000226368 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00050576 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000296032 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),flags=(tn=uint32_t),in=(dims=(blk_bline=9,blk_bx=4,blk_in_chan=192,blk_y=64,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=192,y=28,x=28)),out=(dims=(img=20,chan=16,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=9,blk_bx=4,out_chan_blk=1,blk_y=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),filts=(dims=(out_chan=16,in_chan=480,y=1,x=1)),in=(dims=(img=1,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=1,chan=16,y=14,x=14)),out_chans=(tn=uint32_t,v=16),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00168768 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=480,y=1,x=1,out_chan=16)),in=(dims=(chan=480,pel=200)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=1,y=14,x=14)),out=(dims=(chan=16,pel=200)),out_pels=(tn=none,dims=(img=1,y=14,x=14)),out_ref=(dims=(img=1,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=25,out_chan_tile=2,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000224352 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=480,y=1,x=1,out_chan=16)),in=(dims=(chan=480,pel=200)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=1,y=14,x=14)),out=(dims=(chan=16,pel=200)),out_pels=(tn=none,dims=(img=1,y=14,x=14)),out_ref=(dims=(img=1,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=25,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00203856 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=4,blk_in_chan=480,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=480,y=14,x=14)),out=(dims=(img=1,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=4,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000425216 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=4,blk_in_chan=480,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=480,y=14,x=14)),out=(dims=(img=1,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=4,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00203328 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=4,blk_in_chan=480,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=480,y=14,x=14)),out=(dims=(img=1,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=4,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000424544 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=4,blk_in_chan=480,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=480,y=14,x=14)),out=(dims=(img=1,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=4,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00191568 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=1,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000368384 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=1,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00189664 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=1,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000366784 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=1,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00594208 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),in=(dims=(img=1,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=25,out_chan_tile=2,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000909408 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),in=(dims=(img=1,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=25,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0026208 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=2,blk_in_chan=480,blk_y=25,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=480,y=14,x=14)),out=(dims=(img=1,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=2,out_chan_blk=1,blk_y=25,out_chan_tile=2,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000477184 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=2,blk_in_chan=480,blk_y=25,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=480,y=14,x=14)),out=(dims=(img=1,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=2,out_chan_blk=1,blk_y=25,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000120416 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00043024 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=4,blk_in_chan=480,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=480,y=14,x=14)),out=(dims=(img=1,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=4,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000428064 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=4,blk_in_chan=480,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=480,y=14,x=14)),out=(dims=(img=1,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=4,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000355712 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=1,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000355904 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=1,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000878048 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),in=(dims=(img=1,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=25,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000466016 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=2,blk_in_chan=480,blk_y=25,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=480,y=14,x=14)),out=(dims=(img=1,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=2,out_chan_blk=1,blk_y=25,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),filts=(dims=(out_chan=16,in_chan=480,y=1,x=1)),in=(dims=(img=5,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=5,chan=16,y=14,x=14)),out_chans=(tn=uint32_t,v=16),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00174176 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=480,y=1,x=1,out_chan=16)),in=(dims=(chan=480,pel=984)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=5,y=14,x=14)),out=(dims=(chan=16,pel=984)),out_pels=(tn=none,dims=(img=5,y=14,x=14)),out_ref=(dims=(img=5,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=123,out_chan_tile=2,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000293408 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=480,y=1,x=1,out_chan=16)),in=(dims=(chan=480,pel=984)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=5,y=14,x=14)),out=(dims=(chan=16,pel=984)),out_pels=(tn=none,dims=(img=5,y=14,x=14)),out_ref=(dims=(img=5,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=123,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0020672 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=4,blk_in_chan=480,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=480,y=14,x=14)),out=(dims=(img=5,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=4,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00049008 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=4,blk_in_chan=480,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=480,y=14,x=14)),out=(dims=(img=5,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=4,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0020664 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=4,blk_in_chan=480,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=480,y=14,x=14)),out=(dims=(img=5,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=4,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000489536 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=4,blk_in_chan=480,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=480,y=14,x=14)),out=(dims=(img=5,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=4,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00202016 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=5,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00036544 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=5,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00200224 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=5,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000366752 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=5,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0063024 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),in=(dims=(img=5,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00093152 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),in=(dims=(img=5,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00288192 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=2,blk_in_chan=480,blk_y=64,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=480,y=14,x=14)),out=(dims=(img=5,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=2,out_chan_blk=1,blk_y=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000586528 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=2,blk_in_chan=480,blk_y=64,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=480,y=14,x=14)),out=(dims=(img=5,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=2,out_chan_blk=1,blk_y=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000126848 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000492512 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=4,blk_in_chan=480,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=480,y=14,x=14)),out=(dims=(img=5,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=4,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000489952 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=4,blk_in_chan=480,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=480,y=14,x=14)),out=(dims=(img=5,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=4,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000355424 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=5,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000354016 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=5,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000907776 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),in=(dims=(img=5,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000592512 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=2,blk_in_chan=480,blk_y=64,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=480,y=14,x=14)),out=(dims=(img=5,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=2,out_chan_blk=1,blk_y=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),filts=(dims=(out_chan=16,in_chan=480,y=1,x=1)),in=(dims=(img=20,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=20,chan=16,y=14,x=14)),out_chans=(tn=uint32_t,v=16),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00169504 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=480,y=1,x=1,out_chan=16)),in=(dims=(chan=480,pel=3920)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=20,y=14,x=14)),out=(dims=(chan=16,pel=4096)),out_pels=(tn=none,dims=(img=20,y=14,x=14)),out_ref=(dims=(img=20,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=128,out_chan_tile=2,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000364224 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=480,y=1,x=1,out_chan=16)),in=(dims=(chan=480,pel=3920)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=20,y=14,x=14)),out=(dims=(chan=16,pel=4096)),out_pels=(tn=none,dims=(img=20,y=14,x=14)),out_ref=(dims=(img=20,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=128,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00230608 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=480,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=480,y=14,x=14)),out=(dims=(img=20,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000521664 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=480,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=480,y=14,x=14)),out=(dims=(img=20,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.001396 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=480,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=480,y=14,x=14)),out=(dims=(img=20,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000521536 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=480,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=480,y=14,x=14)),out=(dims=(img=20,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00235168 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=20,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000460096 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=20,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00235328 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=20,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000459872 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=20,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00682496 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),in=(dims=(img=20,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.0011767 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),in=(dims=(img=20,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00290272 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=2,blk_in_chan=480,blk_y=64,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=480,y=14,x=14)),out=(dims=(img=20,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=2,out_chan_blk=1,blk_y=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000613184 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=2,blk_in_chan=480,blk_y=64,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=480,y=14,x=14)),out=(dims=(img=20,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=2,out_chan_blk=1,blk_y=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00014592 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00053296 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=480,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=480,y=14,x=14)),out=(dims=(img=20,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000523136 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=480,blk_y=16,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=480,y=14,x=14)),out=(dims=(img=20,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00044864 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=20,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00044832 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=4)),in=(dims=(img=20,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=16,out_chan_tile=4,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00115539 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),in=(dims=(img=20,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.0006304 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=16)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=2)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=2,blk_in_chan=480,blk_y=64,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=480,y=14,x=14)),out=(dims=(img=20,chan=16,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=2,out_chan_blk=1,blk_y=64,out_chan_tile=2,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),filts=(dims=(out_chan=24,in_chan=512,y=1,x=1)),in=(dims=(img=1,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=1,chan=24,y=14,x=14)),out_chans=(tn=uint32_t,v=24),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00183504 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=512,y=1,x=1,out_chan=24)),in=(dims=(chan=512,pel=200)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=1,y=14,x=14)),out=(dims=(chan=24,pel=200)),out_pels=(tn=none,dims=(img=1,y=14,x=14)),out_ref=(dims=(img=1,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=25,out_chan_tile=3,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00023088 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=512,y=1,x=1,out_chan=24)),in=(dims=(chan=512,pel=200)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=1,y=14,x=14)),out=(dims=(chan=24,pel=200)),out_pels=(tn=none,dims=(img=1,y=14,x=14)),out_ref=(dims=(img=1,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=25,out_chan_tile=3,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00223424 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=4,blk_in_chan=512,blk_y=10,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=512,y=14,x=14)),out=(dims=(img=1,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=4,out_chan_blk=1,blk_y=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000487296 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=4,blk_in_chan=512,blk_y=10,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=512,y=14,x=14)),out=(dims=(img=1,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=4,out_chan_blk=1,blk_y=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0022232 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=4,blk_in_chan=512,blk_y=10,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=512,y=14,x=14)),out=(dims=(img=1,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=4,out_chan_blk=1,blk_y=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000489664 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=4,blk_in_chan=512,blk_y=10,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=512,y=14,x=14)),out=(dims=(img=1,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=4,out_chan_blk=1,blk_y=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00182992 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),in=(dims=(img=1,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=5,out_chan_blk=1,pels_tile=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000388032 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),in=(dims=(img=1,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=5,out_chan_blk=1,pels_tile=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0018128 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),in=(dims=(img=1,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=5,out_chan_blk=1,pels_tile=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000387232 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),in=(dims=(img=1,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=5,out_chan_blk=1,pels_tile=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00490128 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=3)),in=(dims=(img=1,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=25,out_chan_tile=3,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000815808 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=3)),in=(dims=(img=1,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=25,out_chan_tile=3,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00317728 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=3)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=2,blk_in_chan=512,blk_y=25,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=512,y=14,x=14)),out=(dims=(img=1,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=2,out_chan_blk=1,blk_y=25,out_chan_tile=3,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000648128 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=3)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=2,blk_in_chan=512,blk_y=25,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=512,y=14,x=14)),out=(dims=(img=1,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=2,out_chan_blk=1,blk_y=25,out_chan_tile=3,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000124896 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000477824 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=4,blk_in_chan=512,blk_y=10,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=512,y=14,x=14)),out=(dims=(img=1,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=4,out_chan_blk=1,blk_y=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000476064 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=4,blk_in_chan=512,blk_y=10,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=512,y=14,x=14)),out=(dims=(img=1,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=4,out_chan_blk=1,blk_y=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000387584 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),in=(dims=(img=1,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=5,out_chan_blk=1,pels_tile=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000387616 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),in=(dims=(img=1,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=5,out_chan_blk=1,pels_tile=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000809376 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=3)),in=(dims=(img=1,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=25,out_chan_tile=3,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00065008 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=3)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=2,blk_in_chan=512,blk_y=25,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=512,y=14,x=14)),out=(dims=(img=1,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=2,out_chan_blk=1,blk_y=25,out_chan_tile=3,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),filts=(dims=(out_chan=24,in_chan=512,y=1,x=1)),in=(dims=(img=5,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=5,chan=24,y=14,x=14)),out_chans=(tn=uint32_t,v=24),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00184528 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=512,y=1,x=1,out_chan=24)),in=(dims=(chan=512,pel=984)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=5,y=14,x=14)),out=(dims=(chan=24,pel=1360)),out_pels=(tn=none,dims=(img=5,y=14,x=14)),out_ref=(dims=(img=5,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=85,out_chan_tile=3,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000308192 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=512,y=1,x=1,out_chan=24)),in=(dims=(chan=512,pel=984)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=5,y=14,x=14)),out=(dims=(chan=24,pel=1360)),out_pels=(tn=none,dims=(img=5,y=14,x=14)),out_ref=(dims=(img=5,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=85,out_chan_tile=3,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00227216 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=4,blk_in_chan=512,blk_y=10,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=512,y=14,x=14)),out=(dims=(img=5,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=7,blk_bx=4,out_chan_blk=1,blk_y=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000557408 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=4,blk_in_chan=512,blk_y=10,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=512,y=14,x=14)),out=(dims=(img=5,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=7,blk_bx=4,out_chan_blk=1,blk_y=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00227856 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=4,blk_in_chan=512,blk_y=10,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=512,y=14,x=14)),out=(dims=(img=5,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=7,blk_bx=4,out_chan_blk=1,blk_y=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000557312 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=4,blk_in_chan=512,blk_y=10,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=512,y=14,x=14)),out=(dims=(img=5,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=7,blk_bx=4,out_chan_blk=1,blk_y=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00195664 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),in=(dims=(img=5,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000388 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),in=(dims=(img=5,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00194448 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),in=(dims=(img=5,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000385408 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),in=(dims=(img=5,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0051912 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=3)),in=(dims=(img=5,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=3,out_chan_blk=1,pels_tile=42,out_chan_tile=3,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00083328 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=3)),in=(dims=(img=5,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=3,out_chan_blk=1,pels_tile=42,out_chan_tile=3,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00368928 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=3)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=2,blk_in_chan=512,blk_y=42,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=512,y=14,x=14)),out=(dims=(img=5,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=2,out_chan_blk=1,blk_y=42,out_chan_tile=3,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000819904 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=3)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=2,blk_in_chan=512,blk_y=42,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=512,y=14,x=14)),out=(dims=(img=5,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=2,out_chan_blk=1,blk_y=42,out_chan_tile=3,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000129248 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000544928 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=4,blk_in_chan=512,blk_y=10,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=512,y=14,x=14)),out=(dims=(img=5,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=7,blk_bx=4,out_chan_blk=1,blk_y=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000543296 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=4,blk_in_chan=512,blk_y=10,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=512,y=14,x=14)),out=(dims=(img=5,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=7,blk_bx=4,out_chan_blk=1,blk_y=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000387072 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),in=(dims=(img=5,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000386784 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),in=(dims=(img=5,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.0008216 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=3)),in=(dims=(img=5,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=3,out_chan_blk=1,pels_tile=42,out_chan_tile=3,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000820128 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=3)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=2,blk_in_chan=512,blk_y=42,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=512,y=14,x=14)),out=(dims=(img=5,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=2,out_chan_blk=1,blk_y=42,out_chan_tile=3,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),filts=(dims=(out_chan=24,in_chan=512,y=1,x=1)),in=(dims=(img=20,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=20,chan=24,y=14,x=14)),out_chans=(tn=uint32_t,v=24),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00181568 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=512,y=1,x=1,out_chan=24)),in=(dims=(chan=512,pel=3920)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=20,y=14,x=14)),out=(dims=(chan=24,pel=4080)),out_pels=(tn=none,dims=(img=20,y=14,x=14)),out_ref=(dims=(img=20,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=6,out_chan_blk=1,pels_tile=85,out_chan_tile=3,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000341376 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=512,y=1,x=1,out_chan=24)),in=(dims=(chan=512,pel=3920)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=20,y=14,x=14)),out=(dims=(chan=24,pel=4080)),out_pels=(tn=none,dims=(img=20,y=14,x=14)),out_ref=(dims=(img=20,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=6,out_chan_blk=1,pels_tile=85,out_chan_tile=3,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0025288 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=28,blk_bx=4,blk_in_chan=512,blk_y=10,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=512,y=14,x=14)),out=(dims=(img=20,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=28,blk_bx=4,out_chan_blk=1,blk_y=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000604672 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=28,blk_bx=4,blk_in_chan=512,blk_y=10,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=512,y=14,x=14)),out=(dims=(img=20,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=28,blk_bx=4,out_chan_blk=1,blk_y=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0015272 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=28,blk_bx=4,blk_in_chan=512,blk_y=10,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=512,y=14,x=14)),out=(dims=(img=20,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=28,blk_bx=4,out_chan_blk=1,blk_y=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000604416 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=28,blk_bx=4,blk_in_chan=512,blk_y=10,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=512,y=14,x=14)),out=(dims=(img=20,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=28,blk_bx=4,out_chan_blk=1,blk_y=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00229952 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),in=(dims=(img=20,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=98,out_chan_blk=1,pels_tile=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000502304 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),in=(dims=(img=20,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=98,out_chan_blk=1,pels_tile=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00231984 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),in=(dims=(img=20,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=98,out_chan_blk=1,pels_tile=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000500352 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),in=(dims=(img=20,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=98,out_chan_blk=1,pels_tile=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0055288 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=3)),in=(dims=(img=20,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=12,out_chan_blk=1,pels_tile=42,out_chan_tile=3,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.0010305 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=3)),in=(dims=(img=20,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=12,out_chan_blk=1,pels_tile=42,out_chan_tile=3,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00387248 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=3)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=2,blk_in_chan=512,blk_y=42,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=512,y=14,x=14)),out=(dims=(img=20,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=7,blk_bx=2,out_chan_blk=1,blk_y=42,out_chan_tile=3,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.0008432 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=3)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=2,blk_in_chan=512,blk_y=42,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=512,y=14,x=14)),out=(dims=(img=20,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=7,blk_bx=2,out_chan_blk=1,blk_y=42,out_chan_tile=3,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000149792 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000597664 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=28,blk_bx=4,blk_in_chan=512,blk_y=10,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=512,y=14,x=14)),out=(dims=(img=20,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=28,blk_bx=4,out_chan_blk=1,blk_y=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000590432 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=28,blk_bx=4,blk_in_chan=512,blk_y=10,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=512,y=14,x=14)),out=(dims=(img=20,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=28,blk_bx=4,out_chan_blk=1,blk_y=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000501152 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),in=(dims=(img=20,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=98,out_chan_blk=1,pels_tile=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000495456 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=6)),in=(dims=(img=20,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=98,out_chan_blk=1,pels_tile=10,out_chan_tile=6,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00102006 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=3)),in=(dims=(img=20,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=12,out_chan_blk=1,pels_tile=42,out_chan_tile=3,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000842976 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=24)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=3)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=2,blk_in_chan=512,blk_y=42,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=512,y=14,x=14)),out=(dims=(img=20,chan=24,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=7,blk_bx=2,out_chan_blk=1,blk_y=42,out_chan_tile=3,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),filts=(dims=(out_chan=32,in_chan=192,y=1,x=1)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=1,chan=32,y=28,x=28)),out_chans=(tn=uint32_t,v=32),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
| 0105000000666C6F61740101DADA000000000000000004000000010000000062000003000000696D672000000010030000040000006368616E1C0000001C00000001000000791C00000001000000010000007805000000666C6F61740062000000000000018D905F1F69F48806000000002DCDD54338000000295B8E494A830D49787CBB48FF87C448121C6748D4666148B840234853C22348BE5BC84794C8CB47F209D5475A5DAD47DF67B04762EAAB4765A6834760A68147359C8C476C508547C6BF7B47EFBB774761DC6D473F144B478965324744BA5947B5E54747A58225475A6E234705882647A4D6184716CB1347C09F1647BA601C4754201947C46A5444AE99C14441105C44B2AEF24412CCB744B2F10A457F688844B3B3A644533531453714394280DC6F420000000000000000000000002DA83A4215DC6A4200000000000000009D2FDA4200000000246A4D430000000000000000 | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00075024 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=192,y=1,x=1,out_chan=32)),in=(dims=(chan=192,pel=784)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=1,y=28,x=28)),out=(dims=(chan=32,pel=1024)),out_pels=(tn=none,dims=(img=1,y=28,x=28)),out_ref=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=64,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000101984 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=192,y=1,x=1,out_chan=32)),in=(dims=(chan=192,pel=784)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=1,y=28,x=28)),out=(dims=(chan=32,pel=1024)),out_pels=(tn=none,dims=(img=1,y=28,x=28)),out_ref=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=64,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.000944 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=7,blk_in_chan=192,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=192,y=28,x=28)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000181824 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=7,blk_in_chan=192,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=192,y=28,x=28)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00093104 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=7,blk_in_chan=192,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=192,y=28,x=28)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000180832 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=7,blk_in_chan=192,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=192,y=28,x=28)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00083408 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.0001488 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0008216 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000148672 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00160592 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000255456 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00106016 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=4,blk_in_chan=192,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=192,y=28,x=28)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=4,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000205536 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=4,blk_in_chan=192,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=192,y=28,x=28)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=4,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 8.9024e-05 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000183168 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=7,blk_in_chan=192,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=192,y=28,x=28)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000181536 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=7,blk_in_chan=192,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=192,y=28,x=28)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000148192 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000147232 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00025712 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00021296 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=4,blk_in_chan=192,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=192,y=28,x=28)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=4,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),filts=(dims=(out_chan=32,in_chan=192,y=1,x=1)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=5,chan=32,y=28,x=28)),out_chans=(tn=uint32_t,v=32),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00074016 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=192,y=1,x=1,out_chan=32)),in=(dims=(chan=192,pel=3920)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=5,y=28,x=28)),out=(dims=(chan=32,pel=4096)),out_pels=(tn=none,dims=(img=5,y=28,x=28)),out_ref=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=64,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000138624 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=192,y=1,x=1,out_chan=32)),in=(dims=(chan=192,pel=3920)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=5,y=28,x=28)),out=(dims=(chan=32,pel=4096)),out_pels=(tn=none,dims=(img=5,y=28,x=28)),out_ref=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=64,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00093856 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=7,blk_in_chan=192,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=192,y=28,x=28)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000220384 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=7,blk_in_chan=192,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=192,y=28,x=28)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00093808 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=7,blk_in_chan=192,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=192,y=28,x=28)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000220096 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=7,blk_in_chan=192,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=192,y=28,x=28)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00095424 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000157024 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00095504 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000157888 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00177184 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000257504 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00113744 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=4,blk_in_chan=192,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=192,y=28,x=28)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=4,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000241344 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=4,blk_in_chan=192,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=192,y=28,x=28)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=4,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 9.5168e-05 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000220032 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=7,blk_in_chan=192,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=192,y=28,x=28)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000218432 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=7,blk_in_chan=192,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=192,y=28,x=28)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000172064 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000171232 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.0002632 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000247968 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=4,blk_in_chan=192,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=192,y=28,x=28)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=4,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),filts=(dims=(out_chan=32,in_chan=192,y=1,x=1)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=20,chan=32,y=28,x=28)),out_chans=(tn=uint32_t,v=32),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00076112 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=192,y=1,x=1,out_chan=32)),in=(dims=(chan=192,pel=15680)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=20,y=28,x=28)),out=(dims=(chan=32,pel=15872)),out_pels=(tn=none,dims=(img=20,y=28,x=28)),out_ref=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=64,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000302784 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=192,y=1,x=1,out_chan=32)),in=(dims=(chan=192,pel=15680)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=20,y=28,x=28)),out=(dims=(chan=32,pel=15872)),out_pels=(tn=none,dims=(img=20,y=28,x=28)),out_ref=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=64,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00125088 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=7,blk_in_chan=192,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=192,y=28,x=28)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000311456 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=7,blk_in_chan=192,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=192,y=28,x=28)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.000696 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=7,blk_in_chan=192,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=192,y=28,x=28)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000312288 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=7,blk_in_chan=192,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=192,y=28,x=28)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00108304 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=490,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000422528 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=490,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00108368 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=490,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000418752 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=490,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00175392 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000362336 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.001224 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=192,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=192,y=28,x=28)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000320416 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=192,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=192,y=28,x=28)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00013472 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000306272 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=7,blk_in_chan=192,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=192,y=28,x=28)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000306112 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=7,blk_in_chan=192,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=192,y=28,x=28)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000249856 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=490,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.0002488 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=490,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000364992 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00032288 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=192,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=192,y=28,x=28)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),filts=(dims=(out_chan=32,in_chan=256,y=1,x=1)),in=(dims=(img=1,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=1,chan=32,y=28,x=28)),out_chans=(tn=uint32_t,v=32),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00096096 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=256,y=1,x=1,out_chan=32)),in=(dims=(chan=256,pel=784)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=1,y=28,x=28)),out=(dims=(chan=32,pel=1024)),out_pels=(tn=none,dims=(img=1,y=28,x=28)),out_ref=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=64,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000128704 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=256,y=1,x=1,out_chan=32)),in=(dims=(chan=256,pel=784)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=1,y=28,x=28)),out=(dims=(chan=32,pel=1024)),out_pels=(tn=none,dims=(img=1,y=28,x=28)),out_ref=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=64,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00115776 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=7,blk_in_chan=256,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=256,y=28,x=28)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00023792 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=7,blk_in_chan=256,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=256,y=28,x=28)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00115392 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=7,blk_in_chan=256,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=256,y=28,x=28)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00023664 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=7,blk_in_chan=256,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=256,y=28,x=28)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00104416 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000195936 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00103824 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000194656 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00206112 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=1,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000332928 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=1,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00140016 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=4,blk_in_chan=256,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=256,y=28,x=28)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=4,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000266752 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=4,blk_in_chan=256,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=256,y=28,x=28)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=4,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 9.616e-05 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00024256 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=7,blk_in_chan=256,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=256,y=28,x=28)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000240128 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=7,blk_in_chan=256,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=256,y=28,x=28)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000195488 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000193344 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000335936 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=1,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00027872 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=4,blk_in_chan=256,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=256,y=28,x=28)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=4,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),filts=(dims=(out_chan=32,in_chan=256,y=1,x=1)),in=(dims=(img=5,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=5,chan=32,y=28,x=28)),out_chans=(tn=uint32_t,v=32),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00095552 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=256,y=1,x=1,out_chan=32)),in=(dims=(chan=256,pel=3920)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=5,y=28,x=28)),out=(dims=(chan=32,pel=4096)),out_pels=(tn=none,dims=(img=5,y=28,x=28)),out_ref=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=64,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000180736 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=256,y=1,x=1,out_chan=32)),in=(dims=(chan=256,pel=3920)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=5,y=28,x=28)),out=(dims=(chan=32,pel=4096)),out_pels=(tn=none,dims=(img=5,y=28,x=28)),out_ref=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=64,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0011744 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=7,blk_in_chan=256,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=256,y=28,x=28)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00028928 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=7,blk_in_chan=256,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=256,y=28,x=28)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00116816 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=7,blk_in_chan=256,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=256,y=28,x=28)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000288768 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=7,blk_in_chan=256,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=256,y=28,x=28)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00125072 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000246336 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00125056 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000249056 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00227296 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=5,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000400896 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=5,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00150672 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=4,blk_in_chan=256,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=256,y=28,x=28)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=4,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00031376 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=4,blk_in_chan=256,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=256,y=28,x=28)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=4,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000111648 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000292288 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=7,blk_in_chan=256,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=256,y=28,x=28)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000292736 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=7,blk_in_chan=256,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=256,y=28,x=28)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00024208 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000240384 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.0004024 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=5,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00032752 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=4,blk_in_chan=256,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=256,y=28,x=28)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=4,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),filts=(dims=(out_chan=32,in_chan=256,y=1,x=1)),in=(dims=(img=20,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=20,chan=32,y=28,x=28)),out_chans=(tn=uint32_t,v=32),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00097264 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=256,y=1,x=1,out_chan=32)),in=(dims=(chan=256,pel=15680)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=20,y=28,x=28)),out=(dims=(chan=32,pel=15872)),out_pels=(tn=none,dims=(img=20,y=28,x=28)),out_ref=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=64,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000389568 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=256,y=1,x=1,out_chan=32)),in=(dims=(chan=256,pel=15680)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=20,y=28,x=28)),out=(dims=(chan=32,pel=15872)),out_pels=(tn=none,dims=(img=20,y=28,x=28)),out_ref=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=64,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00147728 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=7,blk_in_chan=256,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=256,y=28,x=28)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000395008 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=7,blk_in_chan=256,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=256,y=28,x=28)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0008576 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=7,blk_in_chan=256,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=256,y=28,x=28)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000395296 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=7,blk_in_chan=256,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=256,y=28,x=28)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00137392 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=490,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00054976 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=490,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00137344 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=490,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000547872 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=490,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00225216 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=20,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000461984 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=20,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00159104 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=256,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=256,y=28,x=28)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000395136 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=256,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=256,y=28,x=28)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000165248 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00040304 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=7,blk_in_chan=256,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=256,y=28,x=28)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000401184 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=7,blk_in_chan=256,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=256,y=28,x=28)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000324768 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=490,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000325024 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=490,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000466048 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=20,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000403904 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=256,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=256,y=28,x=28)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),filts=(dims=(out_chan=32,in_chan=16,y=5,x=5)),in=(dims=(img=1,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),kern_sz=(tn=none,dims=(y=5,x=5)),out=(dims=(img=1,chan=32,y=28,x=28)),out_chans=(tn=uint32_t,v=32),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00164336 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=16,y=5,x=5,out_chan=32)),in=(dims=(chan=16,pel=1024)),in_pad=(tn=none,dims=(y=2,x=2)),in_pels=(tn=none,dims=(img=1,y=30,x=30)),out=(dims=(chan=32,pel=1024)),out_pels=(tn=none,dims=(img=1,y=30,x=30)),out_ref=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=64,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000231104 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=16,y=5,x=5,out_chan=32)),in=(dims=(chan=16,pel=1024)),in_pad=(tn=none,dims=(y=2,x=2)),in_pels=(tn=none,dims=(img=1,y=30,x=30)),out=(dims=(chan=32,pel=1024)),out_pels=(tn=none,dims=(img=1,y=30,x=30)),out_ref=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=64,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00081696 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=7,blk_in_chan=16,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=16,y=28,x=28)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 6.8992e-05 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=7,blk_in_chan=16,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=16,y=28,x=28)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00080896 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=7,blk_in_chan=16,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=16,y=28,x=28)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 6.8544e-05 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=7,blk_in_chan=16,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=16,y=28,x=28)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00195664 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000362304 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00194112 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000362112 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00323616 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=1,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000536832 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=1,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00134352 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=4,blk_in_chan=16,blk_y=36,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=16,y=28,x=28)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=4,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000110528 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=4,blk_in_chan=16,blk_y=36,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=16,y=28,x=28)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=4,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000267936 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 6.88e-05 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=7,blk_in_chan=16,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=16,y=28,x=28)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 6.7712e-05 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=7,blk_in_chan=16,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=16,y=28,x=28)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00032272 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000322496 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000423168 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=1,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00010304 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=4,blk_in_chan=16,blk_y=36,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=16,y=28,x=28)),out=(dims=(img=1,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=4,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),filts=(dims=(out_chan=32,in_chan=16,y=5,x=5)),in=(dims=(img=5,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),kern_sz=(tn=none,dims=(y=5,x=5)),out=(dims=(img=5,chan=32,y=28,x=28)),out_chans=(tn=uint32_t,v=32),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0016528 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=16,y=5,x=5,out_chan=32)),in=(dims=(chan=16,pel=4624)),in_pad=(tn=none,dims=(y=2,x=2)),in_pels=(tn=none,dims=(img=5,y=30,x=30)),out=(dims=(chan=32,pel=4608)),out_pels=(tn=none,dims=(img=5,y=30,x=30)),out_ref=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=9,out_chan_blk=1,pels_tile=64,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000261504 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=16,y=5,x=5,out_chan=32)),in=(dims=(chan=16,pel=4624)),in_pad=(tn=none,dims=(y=2,x=2)),in_pels=(tn=none,dims=(img=5,y=30,x=30)),out=(dims=(chan=32,pel=4608)),out_pels=(tn=none,dims=(img=5,y=30,x=30)),out_ref=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=9,out_chan_blk=1,pels_tile=64,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00079824 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=7,blk_in_chan=16,blk_y=16,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=16,y=28,x=28)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 9.1264e-05 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=7,blk_in_chan=16,blk_y=16,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=16,y=28,x=28)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0007936 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=7,blk_in_chan=16,blk_y=16,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=16,y=28,x=28)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 9.04e-05 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=7,blk_in_chan=16,blk_y=16,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=16,y=28,x=28)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0019496 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000375776 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00193168 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000375456 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00325936 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=5,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000540288 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=5,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00134416 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=4,blk_in_chan=16,blk_y=40,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=16,y=28,x=28)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=4,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000116064 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=4,blk_in_chan=16,blk_y=40,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=16,y=28,x=28)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=4,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000127776 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 8.9056e-05 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=7,blk_in_chan=16,blk_y=16,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=16,y=28,x=28)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 8.7712e-05 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=7,blk_in_chan=16,blk_y=16,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=16,y=28,x=28)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000344832 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000344896 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000431232 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=5,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000108288 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=4,blk_in_chan=16,blk_y=40,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=16,y=28,x=28)),out=(dims=(img=5,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=4,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),filts=(dims=(out_chan=32,in_chan=16,y=5,x=5)),in=(dims=(img=20,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),kern_sz=(tn=none,dims=(y=5,x=5)),out=(dims=(img=20,chan=32,y=28,x=28)),out_chans=(tn=uint32_t,v=32),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00165056 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=16,y=5,x=5,out_chan=32)),in=(dims=(chan=16,pel=18128)),in_pad=(tn=none,dims=(y=2,x=2)),in_pels=(tn=none,dims=(img=20,y=30,x=30)),out=(dims=(chan=32,pel=18432)),out_pels=(tn=none,dims=(img=20,y=30,x=30)),out_ref=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=36,out_chan_blk=1,pels_tile=64,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00062064 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=16,y=5,x=5,out_chan=32)),in=(dims=(chan=16,pel=18128)),in_pad=(tn=none,dims=(y=2,x=2)),in_pels=(tn=none,dims=(img=20,y=30,x=30)),out=(dims=(chan=32,pel=18432)),out_pels=(tn=none,dims=(img=20,y=30,x=30)),out_ref=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=36,out_chan_blk=1,pels_tile=64,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00093456 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=7,blk_in_chan=16,blk_y=16,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=16,y=28,x=28)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000243104 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=7,blk_in_chan=16,blk_y=16,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=16,y=28,x=28)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00093728 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=7,blk_in_chan=16,blk_y=16,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=16,y=28,x=28)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000238368 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=7,blk_in_chan=16,blk_y=16,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=16,y=28,x=28)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00222656 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=490,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000799648 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=490,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00135248 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=490,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000797216 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=490,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00325936 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=20,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000602976 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=20,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00137328 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=16,blk_y=40,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=16,y=28,x=28)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000199584 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=16,blk_y=40,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=16,y=28,x=28)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000217824 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000225472 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=7,blk_in_chan=16,blk_y=16,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=16,y=28,x=28)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000225696 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=7,blk_in_chan=16,blk_y=16,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=16,y=28,x=28)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=7,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000530464 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=490,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000535392 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=490,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000516256 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=20,chan=16,y=28,x=28)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000190592 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=16,blk_y=40,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=16,y=28,x=28)),out=(dims=(img=20,chan=32,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),filts=(dims=(out_chan=32,in_chan=512,y=1,x=1)),in=(dims=(img=1,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=1,chan=32,y=14,x=14)),out_chans=(tn=uint32_t,v=32),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00177424 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=512,y=1,x=1,out_chan=32)),in=(dims=(chan=512,pel=200)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=1,y=14,x=14)),out=(dims=(chan=32,pel=200)),out_pels=(tn=none,dims=(img=1,y=14,x=14)),out_ref=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=25,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000217824 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=512,y=1,x=1,out_chan=32)),in=(dims=(chan=512,pel=200)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=1,y=14,x=14)),out=(dims=(chan=32,pel=200)),out_pels=(tn=none,dims=(img=1,y=14,x=14)),out_ref=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=25,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00215408 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=4,blk_in_chan=512,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=512,y=14,x=14)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000472288 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=4,blk_in_chan=512,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=512,y=14,x=14)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00216832 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=4,blk_in_chan=512,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=512,y=14,x=14)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000471104 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=4,blk_in_chan=512,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=512,y=14,x=14)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0019176 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=7,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000382528 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=7,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00191264 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=7,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000383552 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=7,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0039968 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=1,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=25,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000635232 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=1,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=25,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00266112 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=2,blk_in_chan=512,blk_y=25,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=512,y=14,x=14)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=2,out_chan_blk=1,blk_y=25,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00051664 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=2,blk_in_chan=512,blk_y=25,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=512,y=14,x=14)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=2,out_chan_blk=1,blk_y=25,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000125472 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000467232 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=4,blk_in_chan=512,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=512,y=14,x=14)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000466208 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=4,blk_in_chan=512,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=512,y=14,x=14)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000381216 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=7,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00038048 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=7,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000642528 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=1,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=25,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000499264 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=2,blk_in_chan=512,blk_y=25,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=512,y=14,x=14)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=2,out_chan_blk=1,blk_y=25,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),filts=(dims=(out_chan=32,in_chan=512,y=1,x=1)),in=(dims=(img=5,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=5,chan=32,y=14,x=14)),out_chans=(tn=uint32_t,v=32),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00179936 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=512,y=1,x=1,out_chan=32)),in=(dims=(chan=512,pel=984)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=5,y=14,x=14)),out=(dims=(chan=32,pel=1024)),out_pels=(tn=none,dims=(img=5,y=14,x=14)),out_ref=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=64,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000296672 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=512,y=1,x=1,out_chan=32)),in=(dims=(chan=512,pel=984)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=5,y=14,x=14)),out=(dims=(chan=32,pel=1024)),out_pels=(tn=none,dims=(img=5,y=14,x=14)),out_ref=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=64,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00224944 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=9,blk_bx=4,blk_in_chan=512,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=512,y=14,x=14)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=9,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000540832 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=9,blk_bx=4,blk_in_chan=512,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=512,y=14,x=14)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=9,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00222656 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=9,blk_bx=4,blk_in_chan=512,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=512,y=14,x=14)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=9,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000540096 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=9,blk_bx=4,blk_in_chan=512,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=512,y=14,x=14)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=9,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00203248 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000384832 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0020344 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000384736 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00419184 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=5,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00064432 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=5,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0028264 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=3,blk_bx=2,blk_in_chan=512,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=512,y=14,x=14)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=3,blk_bx=2,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000600608 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=3,blk_bx=2,blk_in_chan=512,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=512,y=14,x=14)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=3,blk_bx=2,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000128576 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000535488 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=9,blk_bx=4,blk_in_chan=512,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=512,y=14,x=14)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=9,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00053248 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=9,blk_bx=4,blk_in_chan=512,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=512,y=14,x=14)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=9,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.0003824 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000384352 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00065024 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=5,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00061824 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=3,blk_bx=2,blk_in_chan=512,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=512,y=14,x=14)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=3,blk_bx=2,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),filts=(dims=(out_chan=32,in_chan=512,y=1,x=1)),in=(dims=(img=20,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=20,chan=32,y=14,x=14)),out_chans=(tn=uint32_t,v=32),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00176512 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=512,y=1,x=1,out_chan=32)),in=(dims=(chan=512,pel=3920)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=20,y=14,x=14)),out=(dims=(chan=32,pel=4096)),out_pels=(tn=none,dims=(img=20,y=14,x=14)),out_ref=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=64,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000340704 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=512,y=1,x=1,out_chan=32)),in=(dims=(chan=512,pel=3920)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=20,y=14,x=14)),out=(dims=(chan=32,pel=4096)),out_pels=(tn=none,dims=(img=20,y=14,x=14)),out_ref=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=64,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00247696 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=4,blk_in_chan=512,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=512,y=14,x=14)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=35,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00058976 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=4,blk_in_chan=512,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=512,y=14,x=14)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=35,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00522112 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=4,blk_in_chan=512,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=512,y=14,x=14)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=35,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000589376 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=4,blk_in_chan=512,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=512,y=14,x=14)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=35,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00243216 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000499168 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00230288 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000499392 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0043256 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=20,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000779808 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=20,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00301664 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=9,blk_bx=2,blk_in_chan=512,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=512,y=14,x=14)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=9,blk_bx=2,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000633376 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=9,blk_bx=2,blk_in_chan=512,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=512,y=14,x=14)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=9,blk_bx=2,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000152 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000594112 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=4,blk_in_chan=512,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=512,y=14,x=14)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=35,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00059472 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=4,blk_in_chan=512,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=512,y=14,x=14)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=35,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000497248 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000497568 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000785824 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=20,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00065008 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=9,blk_bx=2,blk_in_chan=512,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=512,y=14,x=14)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=9,blk_bx=2,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),filts=(dims=(out_chan=32,in_chan=528,y=1,x=1)),in=(dims=(img=1,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=1,chan=32,y=14,x=14)),out_chans=(tn=uint32_t,v=32),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0018648 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=528,y=1,x=1,out_chan=32)),in=(dims=(chan=528,pel=200)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=1,y=14,x=14)),out=(dims=(chan=32,pel=200)),out_pels=(tn=none,dims=(img=1,y=14,x=14)),out_ref=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=25,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000221472 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=528,y=1,x=1,out_chan=32)),in=(dims=(chan=528,pel=200)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=1,y=14,x=14)),out=(dims=(chan=32,pel=200)),out_pels=(tn=none,dims=(img=1,y=14,x=14)),out_ref=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=25,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00232288 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=4,blk_in_chan=528,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=528,y=14,x=14)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000485472 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=4,blk_in_chan=528,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=528,y=14,x=14)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00228352 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=4,blk_in_chan=528,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=528,y=14,x=14)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000484288 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=4,blk_in_chan=528,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=528,y=14,x=14)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0006232 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=7,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.0003936 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=7,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00062 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=7,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000395904 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=7,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00407968 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=1,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=25,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000653504 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=1,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=25,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00273936 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=2,blk_in_chan=528,blk_y=25,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=528,y=14,x=14)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=2,out_chan_blk=1,blk_y=25,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000528448 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=2,blk_in_chan=528,blk_y=25,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=528,y=14,x=14)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=2,out_chan_blk=1,blk_y=25,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000129184 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000485248 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=4,blk_in_chan=528,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=528,y=14,x=14)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000484224 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=4,blk_in_chan=528,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=528,y=14,x=14)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000392256 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=7,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.0003936 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=7,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000661216 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=1,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=25,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000507424 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=2,blk_in_chan=528,blk_y=25,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=528,y=14,x=14)),out=(dims=(img=1,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=2,out_chan_blk=1,blk_y=25,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),filts=(dims=(out_chan=32,in_chan=528,y=1,x=1)),in=(dims=(img=5,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=5,chan=32,y=14,x=14)),out_chans=(tn=uint32_t,v=32),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00190016 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=528,y=1,x=1,out_chan=32)),in=(dims=(chan=528,pel=984)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=5,y=14,x=14)),out=(dims=(chan=32,pel=1024)),out_pels=(tn=none,dims=(img=5,y=14,x=14)),out_ref=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=64,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000304832 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=528,y=1,x=1,out_chan=32)),in=(dims=(chan=528,pel=984)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=5,y=14,x=14)),out=(dims=(chan=32,pel=1024)),out_pels=(tn=none,dims=(img=5,y=14,x=14)),out_ref=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=64,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0023848 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=9,blk_bx=4,blk_in_chan=528,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=528,y=14,x=14)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=9,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00055504 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=9,blk_bx=4,blk_in_chan=528,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=528,y=14,x=14)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=9,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00237856 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=9,blk_bx=4,blk_in_chan=528,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=528,y=14,x=14)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=9,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000555104 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=9,blk_bx=4,blk_in_chan=528,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=528,y=14,x=14)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=9,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00217024 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000397152 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00215472 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00039584 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0043416 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=5,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000664928 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=5,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0029784 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=3,blk_bx=2,blk_in_chan=528,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=528,y=14,x=14)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=3,blk_bx=2,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000619872 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=3,blk_bx=2,blk_in_chan=528,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=528,y=14,x=14)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=3,blk_bx=2,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000130944 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000553952 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=9,blk_bx=4,blk_in_chan=528,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=528,y=14,x=14)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=9,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000551968 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=9,blk_bx=4,blk_in_chan=528,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=528,y=14,x=14)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=9,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000396832 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000394656 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000670688 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=5,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000631648 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=3,blk_bx=2,blk_in_chan=528,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=528,y=14,x=14)),out=(dims=(img=5,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=3,blk_bx=2,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),filts=(dims=(out_chan=32,in_chan=528,y=1,x=1)),in=(dims=(img=20,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=20,chan=32,y=14,x=14)),out_chans=(tn=uint32_t,v=32),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00185696 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=528,y=1,x=1,out_chan=32)),in=(dims=(chan=528,pel=3920)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=20,y=14,x=14)),out=(dims=(chan=32,pel=4096)),out_pels=(tn=none,dims=(img=20,y=14,x=14)),out_ref=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=64,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000358176 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=528,y=1,x=1,out_chan=32)),in=(dims=(chan=528,pel=3920)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=20,y=14,x=14)),out=(dims=(chan=32,pel=4096)),out_pels=(tn=none,dims=(img=20,y=14,x=14)),out_ref=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=64,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00269968 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=4,blk_in_chan=528,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=528,y=14,x=14)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=35,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000604192 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=4,blk_in_chan=528,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=528,y=14,x=14)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=35,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00163312 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=4,blk_in_chan=528,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=528,y=14,x=14)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=35,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000607648 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=4,blk_in_chan=528,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=528,y=14,x=14)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=35,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00254848 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000515168 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00246528 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000515328 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00444592 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=20,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.0008024 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=20,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00310608 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=9,blk_bx=2,blk_in_chan=528,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=528,y=14,x=14)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=9,blk_bx=2,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000660288 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=9,blk_bx=2,blk_in_chan=528,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=528,y=14,x=14)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=9,blk_bx=2,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000156064 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000613856 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=4,blk_in_chan=528,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=528,y=14,x=14)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=35,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000615648 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=4,blk_in_chan=528,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=528,y=14,x=14)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=35,blk_bx=4,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000515392 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000513088 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000808544 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=20,chan=528,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000666368 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=528,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=9,blk_bx=2,blk_in_chan=528,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=528,y=14,x=14)),out=(dims=(img=20,chan=32,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=9,blk_bx=2,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),filts=(dims=(out_chan=32,in_chan=832,y=1,x=1)),in=(dims=(img=1,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=1,chan=32,y=7,x=7)),out_chans=(tn=uint32_t,v=32),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00289264 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=832,y=1,x=1,out_chan=32)),in=(dims=(chan=832,pel=56)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=1,y=7,x=7)),out=(dims=(chan=32,pel=56)),out_pels=(tn=none,dims=(img=1,y=7,x=7)),out_ref=(dims=(img=1,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=7,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000326624 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=832,y=1,x=1,out_chan=32)),in=(dims=(chan=832,pel=56)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=1,y=7,x=7)),out=(dims=(chan=32,pel=56)),out_pels=(tn=none,dims=(img=1,y=7,x=7)),out_ref=(dims=(img=1,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=7,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00351296 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=2,blk_in_chan=832,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=832,y=7,x=7)),out=(dims=(img=1,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=2,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000759488 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=2,blk_in_chan=832,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=832,y=7,x=7)),out=(dims=(img=1,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=2,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00362032 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=2,blk_in_chan=832,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=832,y=7,x=7)),out=(dims=(img=1,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=2,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.0007592 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=2,blk_in_chan=832,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=832,y=7,x=7)),out=(dims=(img=1,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=2,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0030984 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000617184 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00308336 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000619712 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00643152 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=1,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=7,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000957184 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=1,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=7,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00480592 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=1,blk_in_chan=832,blk_y=7,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=832,y=7,x=7)),out=(dims=(img=1,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=1,out_chan_blk=1,blk_y=7,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00079904 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=1,blk_in_chan=832,blk_y=7,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=832,y=7,x=7)),out=(dims=(img=1,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=1,out_chan_blk=1,blk_y=7,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00016608 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00075936 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=2,blk_in_chan=832,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=832,y=7,x=7)),out=(dims=(img=1,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=2,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000754272 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=2,blk_in_chan=832,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=832,y=7,x=7)),out=(dims=(img=1,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=2,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000620384 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000615232 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=1,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000972256 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=1,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=7,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000754016 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=1,blk_in_chan=832,blk_y=7,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=832,y=7,x=7)),out=(dims=(img=1,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=1,out_chan_blk=1,blk_y=7,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),filts=(dims=(out_chan=32,in_chan=832,y=1,x=1)),in=(dims=(img=5,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=5,chan=32,y=7,x=7)),out_chans=(tn=uint32_t,v=32),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00291776 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=832,y=1,x=1,out_chan=32)),in=(dims=(chan=832,pel=248)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=5,y=7,x=7)),out=(dims=(chan=32,pel=248)),out_pels=(tn=none,dims=(img=5,y=7,x=7)),out_ref=(dims=(img=5,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=31,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000345056 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=832,y=1,x=1,out_chan=32)),in=(dims=(chan=832,pel=248)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=5,y=7,x=7)),out=(dims=(chan=32,pel=248)),out_pels=(tn=none,dims=(img=5,y=7,x=7)),out_ref=(dims=(img=5,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=31,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00354352 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=2,blk_in_chan=832,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=832,y=7,x=7)),out=(dims=(img=5,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=2,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000759232 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=2,blk_in_chan=832,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=832,y=7,x=7)),out=(dims=(img=5,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=2,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0034936 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=2,blk_in_chan=832,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=832,y=7,x=7)),out=(dims=(img=5,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=2,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00075808 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=2,blk_in_chan=832,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=832,y=7,x=7)),out=(dims=(img=5,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=2,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0031328 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00062368 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00312544 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000623648 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00623808 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=5,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=31,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00102118 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=5,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=31,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00433392 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=1,blk_in_chan=832,blk_y=31,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=832,y=7,x=7)),out=(dims=(img=5,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=1,out_chan_blk=1,blk_y=31,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000828096 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=1,blk_in_chan=832,blk_y=31,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=832,y=7,x=7)),out=(dims=(img=5,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=1,out_chan_blk=1,blk_y=31,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000170752 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000760864 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=2,blk_in_chan=832,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=832,y=7,x=7)),out=(dims=(img=5,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=2,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000754592 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=2,blk_in_chan=832,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=832,y=7,x=7)),out=(dims=(img=5,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=2,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.0006232 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000623392 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=5,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00103501 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=5,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=31,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000811584 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=1,blk_in_chan=832,blk_y=31,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=832,y=7,x=7)),out=(dims=(img=5,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=1,out_chan_blk=1,blk_y=31,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),filts=(dims=(out_chan=32,in_chan=832,y=1,x=1)),in=(dims=(img=20,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=20,chan=32,y=7,x=7)),out_chans=(tn=uint32_t,v=32),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00292 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=832,y=1,x=1,out_chan=32)),in=(dims=(chan=832,pel=984)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=20,y=7,x=7)),out=(dims=(chan=32,pel=1024)),out_pels=(tn=none,dims=(img=20,y=7,x=7)),out_ref=(dims=(img=20,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=64,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000487104 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=832,y=1,x=1,out_chan=32)),in=(dims=(chan=832,pel=984)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=20,y=7,x=7)),out=(dims=(chan=32,pel=1024)),out_pels=(tn=none,dims=(img=20,y=7,x=7)),out_ref=(dims=(img=20,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=64,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0037512 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=2,blk_in_chan=832,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=832,y=7,x=7)),out=(dims=(img=20,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=2,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000869696 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=2,blk_in_chan=832,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=832,y=7,x=7)),out=(dims=(img=20,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=2,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00328592 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=2,blk_in_chan=832,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=832,y=7,x=7)),out=(dims=(img=20,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=2,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000868 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=2,blk_in_chan=832,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=832,y=7,x=7)),out=(dims=(img=20,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=2,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0036536 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000725568 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00367792 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000725696 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00716624 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=20,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00117277 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=20,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00465904 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=1,blk_in_chan=832,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=832,y=7,x=7)),out=(dims=(img=20,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=1,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000979456 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=1,blk_in_chan=832,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=832,y=7,x=7)),out=(dims=(img=20,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=1,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000197312 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000868 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=2,blk_in_chan=832,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=832,y=7,x=7)),out=(dims=(img=20,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=2,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000866816 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=2,blk_in_chan=832,blk_y=8,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=832,y=7,x=7)),out=(dims=(img=20,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=2,out_chan_blk=1,blk_y=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000724512 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000719904 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=8)),in=(dims=(img=20,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=8,out_chan_tile=8,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00118534 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),in=(dims=(img=20,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00097584 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=32)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=4)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=1,blk_in_chan=832,blk_y=32,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=832,y=7,x=7)),out=(dims=(img=20,chan=32,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=1,out_chan_blk=1,blk_y=32,out_chan_tile=4,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),filts=(dims=(out_chan=48,in_chan=16,y=5,x=5)),in=(dims=(img=1,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),kern_sz=(tn=none,dims=(y=5,x=5)),out=(dims=(img=1,chan=48,y=14,x=14)),out_chans=(tn=uint32_t,v=48),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00164256 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=16,y=5,x=5,out_chan=48)),in=(dims=(chan=16,pel=328)),in_pad=(tn=none,dims=(y=2,x=2)),in_pels=(tn=none,dims=(img=1,y=16,x=16)),out=(dims=(chan=48,pel=256)),out_pels=(tn=none,dims=(img=1,y=16,x=16)),out_ref=(dims=(img=1,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=32,out_chan_tile=6,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000209472 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=16,y=5,x=5,out_chan=48)),in=(dims=(chan=16,pel=328)),in_pad=(tn=none,dims=(y=2,x=2)),in_pels=(tn=none,dims=(img=1,y=16,x=16)),out=(dims=(chan=48,pel=256)),out_pels=(tn=none,dims=(img=1,y=16,x=16)),out_ref=(dims=(img=1,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=32,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00062464 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=3,blk_bx=4,blk_in_chan=16,blk_y=9,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=16,y=14,x=14)),out=(dims=(img=1,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=3,blk_bx=4,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 6.9184e-05 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=3,blk_bx=4,blk_in_chan=16,blk_y=9,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=16,y=14,x=14)),out=(dims=(img=1,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=3,blk_bx=4,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0006168 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=3,blk_bx=4,blk_in_chan=16,blk_y=9,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=16,y=14,x=14)),out=(dims=(img=1,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=3,blk_bx=4,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 6.8256e-05 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=3,blk_bx=4,blk_in_chan=16,blk_y=9,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=16,y=14,x=14)),out=(dims=(img=1,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=3,blk_bx=4,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00188336 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=1,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=10,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000368256 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=1,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=10,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00186912 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=1,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=10,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000366464 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=1,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=10,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00320704 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=6)),in=(dims=(img=1,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00052864 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=6)),in=(dims=(img=1,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00126336 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=2,blk_in_chan=16,blk_y=25,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=16,y=14,x=14)),out=(dims=(img=1,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=2,out_chan_blk=1,blk_y=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000138592 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=2,blk_in_chan=16,blk_y=25,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=16,y=14,x=14)),out=(dims=(img=1,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=2,out_chan_blk=1,blk_y=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000119776 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 6.8448e-05 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=3,blk_bx=4,blk_in_chan=16,blk_y=9,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=16,y=14,x=14)),out=(dims=(img=1,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=3,blk_bx=4,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 6.6176e-05 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=3,blk_bx=4,blk_in_chan=16,blk_y=9,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=16,y=14,x=14)),out=(dims=(img=1,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=3,blk_bx=4,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.0003248 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=1,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=10,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000325568 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=1,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=10,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000441536 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=6)),in=(dims=(img=1,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00010608 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=2,blk_in_chan=16,blk_y=25,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=16,y=14,x=14)),out=(dims=(img=1,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=2,out_chan_blk=1,blk_y=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),filts=(dims=(out_chan=48,in_chan=16,y=5,x=5)),in=(dims=(img=5,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),kern_sz=(tn=none,dims=(y=5,x=5)),out=(dims=(img=5,chan=48,y=14,x=14)),out_chans=(tn=uint32_t,v=48),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00162256 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=16,y=5,x=5,out_chan=48)),in=(dims=(chan=16,pel=1352)),in_pad=(tn=none,dims=(y=2,x=2)),in_pels=(tn=none,dims=(img=5,y=16,x=16)),out=(dims=(chan=48,pel=1344)),out_pels=(tn=none,dims=(img=5,y=16,x=16)),out_ref=(dims=(img=5,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=42,out_chan_tile=6,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000224288 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=16,y=5,x=5,out_chan=48)),in=(dims=(chan=16,pel=1352)),in_pad=(tn=none,dims=(y=2,x=2)),in_pels=(tn=none,dims=(img=5,y=16,x=16)),out=(dims=(chan=48,pel=1344)),out_pels=(tn=none,dims=(img=5,y=16,x=16)),out_ref=(dims=(img=5,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=42,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0006448 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=14,blk_bx=4,blk_in_chan=16,blk_y=13,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=16,y=14,x=14)),out=(dims=(img=5,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=14,blk_bx=4,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 7.7088e-05 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=14,blk_bx=4,blk_in_chan=16,blk_y=13,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=16,y=14,x=14)),out=(dims=(img=5,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=14,blk_bx=4,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00063344 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=14,blk_bx=4,blk_in_chan=16,blk_y=13,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=16,y=14,x=14)),out=(dims=(img=5,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=14,blk_bx=4,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 8.0448e-05 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=14,blk_bx=4,blk_in_chan=16,blk_y=13,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=16,y=14,x=14)),out=(dims=(img=5,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=14,blk_bx=4,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0018816 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=5,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.0003696 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=5,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00186672 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=5,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00037104 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=5,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00321984 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=6)),in=(dims=(img=5,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=6,out_chan_blk=1,pels_tile=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000533184 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=6)),in=(dims=(img=5,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=6,out_chan_blk=1,pels_tile=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00125104 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=2,blk_in_chan=16,blk_y=29,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=16,y=14,x=14)),out=(dims=(img=5,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=2,out_chan_blk=1,blk_y=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000146176 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=2,blk_in_chan=16,blk_y=29,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=16,y=14,x=14)),out=(dims=(img=5,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=2,out_chan_blk=1,blk_y=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000128256 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 7.872e-05 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=14,blk_bx=4,blk_in_chan=16,blk_y=13,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=16,y=14,x=14)),out=(dims=(img=5,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=14,blk_bx=4,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 7.7152e-05 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=14,blk_bx=4,blk_in_chan=16,blk_y=13,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=16,y=14,x=14)),out=(dims=(img=5,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=14,blk_bx=4,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000336704 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=5,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000343168 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=5,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000446752 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=6)),in=(dims=(img=5,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=6,out_chan_blk=1,pels_tile=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000112576 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=2,blk_in_chan=16,blk_y=29,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=16,y=14,x=14)),out=(dims=(img=5,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=2,out_chan_blk=1,blk_y=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),filts=(dims=(out_chan=48,in_chan=16,y=5,x=5)),in=(dims=(img=20,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),kern_sz=(tn=none,dims=(y=5,x=5)),out=(dims=(img=20,chan=48,y=14,x=14)),out_chans=(tn=uint32_t,v=48),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00162688 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=16,y=5,x=5,out_chan=48)),in=(dims=(chan=16,pel=5192)),in_pad=(tn=none,dims=(y=2,x=2)),in_pels=(tn=none,dims=(img=20,y=16,x=16)),out=(dims=(chan=48,pel=5376)),out_pels=(tn=none,dims=(img=20,y=16,x=16)),out_ref=(dims=(img=20,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=42,out_chan_tile=6,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000311552 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=16,y=5,x=5,out_chan=48)),in=(dims=(chan=16,pel=5192)),in_pad=(tn=none,dims=(y=2,x=2)),in_pels=(tn=none,dims=(img=20,y=16,x=16)),out=(dims=(chan=48,pel=5376)),out_pels=(tn=none,dims=(img=20,y=16,x=16)),out_ref=(dims=(img=20,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=42,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00066528 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=56,blk_bx=4,blk_in_chan=16,blk_y=13,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=16,y=14,x=14)),out=(dims=(img=20,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=56,blk_bx=4,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00011792 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=56,blk_bx=4,blk_in_chan=16,blk_y=13,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=16,y=14,x=14)),out=(dims=(img=20,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=56,blk_bx=4,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0006656 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=56,blk_bx=4,blk_in_chan=16,blk_y=13,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=16,y=14,x=14)),out=(dims=(img=20,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=56,blk_bx=4,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000119232 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=56,blk_bx=4,blk_in_chan=16,blk_y=13,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=16,y=14,x=14)),out=(dims=(img=20,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=56,blk_bx=4,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.001892 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=20,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=196,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000406784 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=20,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=196,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00186736 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=20,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=196,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000407872 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=20,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=196,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.003212 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=6)),in=(dims=(img=20,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=24,out_chan_blk=1,pels_tile=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00053424 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=6)),in=(dims=(img=20,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=24,out_chan_blk=1,pels_tile=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00125888 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=14,blk_bx=2,blk_in_chan=16,blk_y=29,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=16,y=14,x=14)),out=(dims=(img=20,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=14,blk_bx=2,out_chan_blk=1,blk_y=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000174336 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=14,blk_bx=2,blk_in_chan=16,blk_y=29,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=16,y=14,x=14)),out=(dims=(img=20,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=14,blk_bx=2,out_chan_blk=1,blk_y=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000157728 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000111552 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=56,blk_bx=4,blk_in_chan=16,blk_y=13,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=16,y=14,x=14)),out=(dims=(img=20,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=56,blk_bx=4,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000107232 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=56,blk_bx=4,blk_in_chan=16,blk_y=13,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=16,y=14,x=14)),out=(dims=(img=20,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=56,blk_bx=4,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000370304 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=20,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=196,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000368192 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=20,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=196,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000452704 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=6)),in=(dims=(img=20,chan=16,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=24,out_chan_blk=1,pels_tile=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000141792 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=16,y=5,x=5,out_chan_reg=8,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=14,blk_bx=2,blk_in_chan=16,blk_y=29,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=16,y=14,x=14)),out=(dims=(img=20,chan=48,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=14,blk_bx=2,out_chan_blk=1,blk_y=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),filts=(dims=(out_chan=48,in_chan=832,y=1,x=1)),in=(dims=(img=1,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=1,chan=48,y=7,x=7)),out_chans=(tn=uint32_t,v=48),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00296608 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=832,y=1,x=1,out_chan=48)),in=(dims=(chan=832,pel=56)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=1,y=7,x=7)),out=(dims=(chan=48,pel=56)),out_pels=(tn=none,dims=(img=1,y=7,x=7)),out_ref=(dims=(img=1,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=7,out_chan_tile=6,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00032592 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=832,y=1,x=1,out_chan=48)),in=(dims=(chan=832,pel=56)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=1,y=7,x=7)),out=(dims=(chan=48,pel=56)),out_pels=(tn=none,dims=(img=1,y=7,x=7)),out_ref=(dims=(img=1,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=7,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00353664 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=2,blk_in_chan=832,blk_y=5,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=832,y=7,x=7)),out=(dims=(img=1,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=2,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.0007912 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=2,blk_in_chan=832,blk_y=5,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=832,y=7,x=7)),out=(dims=(img=1,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=2,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00358448 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=2,blk_in_chan=832,blk_y=5,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=832,y=7,x=7)),out=(dims=(img=1,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=2,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000790464 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=2,blk_in_chan=832,blk_y=5,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=832,y=7,x=7)),out=(dims=(img=1,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=2,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.002904 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=1,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=3,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000619072 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=1,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=3,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00287072 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=1,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=3,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000623264 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=1,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=3,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00595856 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=6)),in=(dims=(img=1,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=7,out_chan_tile=6,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000951072 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=6)),in=(dims=(img=1,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=7,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0053792 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=1,blk_in_chan=832,blk_y=7,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=832,y=7,x=7)),out=(dims=(img=1,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=1,out_chan_blk=1,blk_y=7,out_chan_tile=6,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000776 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=1,blk_in_chan=832,blk_y=7,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=832,y=7,x=7)),out=(dims=(img=1,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=1,out_chan_blk=1,blk_y=7,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000222656 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00076624 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=2,blk_in_chan=832,blk_y=5,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=832,y=7,x=7)),out=(dims=(img=1,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=2,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000763936 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=2,blk_in_chan=832,blk_y=5,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=832,y=7,x=7)),out=(dims=(img=1,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=2,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000624832 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=1,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=3,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00061888 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=1,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=3,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000945536 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=6)),in=(dims=(img=1,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=7,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000756224 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=1,blk_in_chan=832,blk_y=7,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=832,y=7,x=7)),out=(dims=(img=1,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=1,out_chan_blk=1,blk_y=7,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),filts=(dims=(out_chan=48,in_chan=832,y=1,x=1)),in=(dims=(img=5,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=5,chan=48,y=7,x=7)),out_chans=(tn=uint32_t,v=48),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00297008 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=832,y=1,x=1,out_chan=48)),in=(dims=(chan=832,pel=248)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=5,y=7,x=7)),out=(dims=(chan=48,pel=248)),out_pels=(tn=none,dims=(img=5,y=7,x=7)),out_ref=(dims=(img=5,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=31,out_chan_tile=6,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000336672 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=832,y=1,x=1,out_chan=48)),in=(dims=(chan=832,pel=248)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=5,y=7,x=7)),out=(dims=(chan=48,pel=248)),out_pels=(tn=none,dims=(img=5,y=7,x=7)),out_ref=(dims=(img=5,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=31,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00355568 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=2,blk_in_chan=832,blk_y=5,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=832,y=7,x=7)),out=(dims=(img=5,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=7,blk_bx=2,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000789952 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=2,blk_in_chan=832,blk_y=5,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=832,y=7,x=7)),out=(dims=(img=5,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=7,blk_bx=2,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00358208 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=2,blk_in_chan=832,blk_y=5,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=832,y=7,x=7)),out=(dims=(img=5,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=7,blk_bx=2,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000786528 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=2,blk_in_chan=832,blk_y=5,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=832,y=7,x=7)),out=(dims=(img=5,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=7,blk_bx=2,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.002968 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=5,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000624 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=5,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0029568 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=5,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00062304 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=5,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00583248 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=6)),in=(dims=(img=5,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00102198 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=6)),in=(dims=(img=5,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00504992 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=1,blk_in_chan=832,blk_y=21,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=832,y=7,x=7)),out=(dims=(img=5,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=1,out_chan_blk=1,blk_y=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000813472 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=1,blk_in_chan=832,blk_y=21,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=832,y=7,x=7)),out=(dims=(img=5,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=1,out_chan_blk=1,blk_y=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000197024 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000770272 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=2,blk_in_chan=832,blk_y=5,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=832,y=7,x=7)),out=(dims=(img=5,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=7,blk_bx=2,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000764096 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=2,blk_in_chan=832,blk_y=5,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=832,y=7,x=7)),out=(dims=(img=5,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=7,blk_bx=2,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.0006208 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=5,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000622912 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=5,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00101222 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=6)),in=(dims=(img=5,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000810048 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=2,blk_bx=1,blk_in_chan=832,blk_y=21,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=832,y=7,x=7)),out=(dims=(img=5,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=2,blk_bx=1,out_chan_blk=1,blk_y=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),filts=(dims=(out_chan=48,in_chan=832,y=1,x=1)),in=(dims=(img=20,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=20,chan=48,y=7,x=7)),out_chans=(tn=uint32_t,v=48),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0029696 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=832,y=1,x=1,out_chan=48)),in=(dims=(chan=832,pel=984)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=20,y=7,x=7)),out=(dims=(chan=48,pel=1008)),out_pels=(tn=none,dims=(img=20,y=7,x=7)),out_ref=(dims=(img=20,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=3,out_chan_blk=1,pels_tile=42,out_chan_tile=6,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000463136 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=832,y=1,x=1,out_chan=48)),in=(dims=(chan=832,pel=984)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=20,y=7,x=7)),out=(dims=(chan=48,pel=1008)),out_pels=(tn=none,dims=(img=20,y=7,x=7)),out_ref=(dims=(img=20,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=3,out_chan_blk=1,pels_tile=42,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0038816 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=28,blk_bx=2,blk_in_chan=832,blk_y=5,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=832,y=7,x=7)),out=(dims=(img=20,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=28,blk_bx=2,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000920544 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=28,blk_bx=2,blk_in_chan=832,blk_y=5,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=832,y=7,x=7)),out=(dims=(img=20,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=28,blk_bx=2,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00311696 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=28,blk_bx=2,blk_in_chan=832,blk_y=5,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=832,y=7,x=7)),out=(dims=(img=20,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=28,blk_bx=2,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000913472 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=28,blk_bx=2,blk_in_chan=832,blk_y=5,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=832,y=7,x=7)),out=(dims=(img=20,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=28,blk_bx=2,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00346864 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=20,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00072576 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=20,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00344944 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=20,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00072688 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=20,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.006632 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=6)),in=(dims=(img=20,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=6,out_chan_blk=1,pels_tile=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00117363 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=6)),in=(dims=(img=20,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=6,out_chan_blk=1,pels_tile=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0056096 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=1,blk_in_chan=832,blk_y=21,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=832,y=7,x=7)),out=(dims=(img=20,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=7,blk_bx=1,out_chan_blk=1,blk_y=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000948288 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=1,blk_in_chan=832,blk_y=21,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=832,y=7,x=7)),out=(dims=(img=20,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=7,blk_bx=1,out_chan_blk=1,blk_y=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00022576 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000884736 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=28,blk_bx=2,blk_in_chan=832,blk_y=5,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=832,y=7,x=7)),out=(dims=(img=20,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=28,blk_bx=2,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000885408 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=28,blk_bx=2,blk_in_chan=832,blk_y=5,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=832,y=7,x=7)),out=(dims=(img=20,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=28,blk_bx=2,out_chan_blk=1,blk_y=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00072432 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=20,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.0007232 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=4,out_chan_tile=12)),in=(dims=(img=20,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=5,out_chan_tile=12,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00117738 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=6)),in=(dims=(img=20,chan=832,y=7,x=7)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=6,out_chan_blk=1,pels_tile=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000939296 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=48)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=832,y=1,x=1,out_chan_reg=8,out_chan_tile=6)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=1,blk_in_chan=832,blk_y=21,blk_x=8)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=832,y=7,x=7)),out=(dims=(img=20,chan=48,y=7,x=7)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=7,blk_bx=1,out_chan_blk=1,blk_y=21,out_chan_tile=6,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),filts=(dims=(out_chan=64,in_chan=64,y=1,x=1)),in=(dims=(img=1,chan=64,y=56,x=56)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=1,chan=64,y=56,x=56)),out_chans=(tn=uint32_t,v=64),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00032 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=64,y=1,x=1,out_chan=64)),in=(dims=(chan=64,pel=3328)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=3328)),out_ref=(dims=(img=1,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 6.7584e-05 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=64,y=1,x=1,out_chan=64)),in=(dims=(chan=64,pel=3328)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=3328)),out_ref=(dims=(img=1,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00027712 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=196,blk_iter=8,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=56,x=56)),out_ref=(dims=(img=1,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=196,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 5.2e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=196,blk_iter=8,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=56,x=56)),out_ref=(dims=(img=1,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=196,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00037776 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=14,blk_bx=14,blk_in_chan=64,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=64,y=56,x=56)),out=(dims=(img=1,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=14,blk_bx=14,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 9.2064e-05 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=14,blk_bx=14,blk_in_chan=64,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=64,y=56,x=56)),out=(dims=(img=1,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=14,blk_bx=14,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0002784 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=196,blk_iter=8,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=56,x=56)),out_ref=(dims=(img=1,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=196,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 5.0272e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=196,blk_iter=8,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=56,x=56)),out_ref=(dims=(img=1,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=196,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0004736 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=64,y=56,x=56)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=196,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 8.3456e-05 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=64,y=56,x=56)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=196,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00055552 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=1,chan=64,y=56,x=56)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 8.8448e-05 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=1,chan=64,y=56,x=56)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00077248 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=25,blk_iter=8,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=56,x=56)),out_ref=(dims=(img=1,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 4.8352e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=25,blk_iter=8,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=56,x=56)),out_ref=(dims=(img=1,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 8.5408e-05 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 4.4544e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=196,blk_iter=8,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=56,x=56)),out_ref=(dims=(img=1,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=196,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 8.9792e-05 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=14,blk_bx=14,blk_in_chan=64,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=64,y=56,x=56)),out=(dims=(img=1,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=14,blk_bx=14,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 4.2912e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=196,blk_iter=8,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=56,x=56)),out_ref=(dims=(img=1,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=196,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 7.9136e-05 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=64,y=56,x=56)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=196,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 8.912e-05 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=1,chan=64,y=56,x=56)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 4.736e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=25,blk_iter=8,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=56,x=56)),out_ref=(dims=(img=1,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=25,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),filts=(dims=(out_chan=64,in_chan=64,y=1,x=1)),in=(dims=(img=5,chan=64,y=56,x=56)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=5,chan=64,y=56,x=56)),out_chans=(tn=uint32_t,v=64),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00035696 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=64,y=1,x=1,out_chan=64)),in=(dims=(chan=64,pel=15872)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=15872)),out_ref=(dims=(img=5,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00015296 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=64,y=1,x=1,out_chan=64)),in=(dims=(chan=64,pel=15872)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=15872)),out_ref=(dims=(img=5,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00041104 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=980,blk_iter=8,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=56,x=56)),out_ref=(dims=(img=5,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000198336 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=980,blk_iter=8,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=56,x=56)),out_ref=(dims=(img=5,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00055072 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=14,blk_in_chan=64,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=64,y=56,x=56)),out=(dims=(img=5,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=14,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00028944 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=14,blk_in_chan=64,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=64,y=56,x=56)),out=(dims=(img=5,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=14,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00041184 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=980,blk_iter=8,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=56,x=56)),out_ref=(dims=(img=5,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000198112 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=980,blk_iter=8,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=56,x=56)),out_ref=(dims=(img=5,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00058768 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=64,y=56,x=56)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000302464 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=64,y=56,x=56)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00072816 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=5,chan=64,y=56,x=56)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00029616 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=5,chan=64,y=56,x=56)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00086784 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=123,blk_iter=8,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=56,x=56)),out_ref=(dims=(img=5,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00017248 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=123,blk_iter=8,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=56,x=56)),out_ref=(dims=(img=5,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00017088 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.0001648 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=980,blk_iter=8,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=56,x=56)),out_ref=(dims=(img=5,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000279424 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=14,blk_in_chan=64,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=64,y=56,x=56)),out=(dims=(img=5,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=14,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.0001664 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=980,blk_iter=8,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=56,x=56)),out_ref=(dims=(img=5,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000244288 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=64,y=56,x=56)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000299808 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=5,chan=64,y=56,x=56)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000136864 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=123,blk_iter=8,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=56,x=56)),out_ref=(dims=(img=5,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),filts=(dims=(out_chan=64,in_chan=64,y=1,x=1)),in=(dims=(img=20,chan=64,y=56,x=56)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=20,chan=64,y=56,x=56)),out_chans=(tn=uint32_t,v=64),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00072752 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=64,y=1,x=1,out_chan=64)),in=(dims=(chan=64,pel=62720)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=62720)),out_ref=(dims=(img=20,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00039392 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=64,y=1,x=1,out_chan=64)),in=(dims=(chan=64,pel=62720)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=62720)),out_ref=(dims=(img=20,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00097344 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=3920,blk_iter=8,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=56,x=56)),out_ref=(dims=(img=20,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=3920,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00062768 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=3920,blk_iter=8,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=56,x=56)),out_ref=(dims=(img=20,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=3920,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00164304 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=280,blk_bx=14,blk_in_chan=64,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=64,y=56,x=56)),out=(dims=(img=20,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=280,blk_bx=14,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000953824 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=280,blk_bx=14,blk_in_chan=64,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=64,y=56,x=56)),out=(dims=(img=20,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=280,blk_bx=14,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00058464 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=3920,blk_iter=8,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=56,x=56)),out_ref=(dims=(img=20,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=3920,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000623104 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=3920,blk_iter=8,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=56,x=56)),out_ref=(dims=(img=20,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=3920,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00150768 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=64,y=56,x=56)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=3920,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00095552 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=64,y=56,x=56)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=3920,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00181104 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=20,chan=64,y=56,x=56)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=490,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.0008896 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=20,chan=64,y=56,x=56)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=490,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00314928 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=490,blk_iter=8,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=56,x=56)),out_ref=(dims=(img=20,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=490,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000441024 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=490,blk_iter=8,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=56,x=56)),out_ref=(dims=(img=20,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=490,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000486112 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00057264 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=3920,blk_iter=8,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=56,x=56)),out_ref=(dims=(img=20,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=3920,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00094256 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=280,blk_bx=14,blk_in_chan=64,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=64,y=56,x=56)),out=(dims=(img=20,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=280,blk_bx=14,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00056784 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=3920,blk_iter=8,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=56,x=56)),out_ref=(dims=(img=20,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=3920,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000853696 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=64,y=56,x=56)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=3920,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000923712 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=20,chan=64,y=56,x=56)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=490,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000412352 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=64,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=490,blk_iter=8,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=56,x=56)),out_ref=(dims=(img=20,chan=64,y=56,x=56)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=490,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),filts=(dims=(out_chan=64,in_chan=3,y=7,x=7)),in=(dims=(img=1,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),kern_sz=(tn=none,dims=(y=7,x=7)),out=(dims=(img=1,chan=64,y=112,x=112)),out_chans=(tn=uint32_t,v=64),stride=(tn=none,dims=(y=2,x=2)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00098288 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=3,y=7,x=7,out_chan=64)),in=(dims=(chan=3,pel=53136)),in_pad=(tn=none,dims=(y=3,x=3)),in_pels=(tn=none,dims=(img=1,y=228,x=228)),out=(dims=(chan=64,pel=13056)),out_pels=(tn=none,dims=(img=1,y=114,x=114)),out_ref=(dims=(img=1,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=51,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000325472 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=3,y=7,x=7,out_chan=64)),in=(dims=(chan=3,pel=53136)),in_pad=(tn=none,dims=(y=3,x=3)),in_pels=(tn=none,dims=(img=1,y=228,x=228)),out=(dims=(chan=64,pel=13056)),out_pels=(tn=none,dims=(img=1,y=114,x=114)),out_ref=(dims=(img=1,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=51,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00048448 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=28,blk_bx=28,blk_in_chan=3,blk_y=13,blk_x=13)),in_pad=(tn=none,dims=(y=3,x=3)),in_ref=(dims=(img=1,chan=3,y=224,x=224)),out=(dims=(img=1,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(blk_bline=28,blk_bx=28,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000205536 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=28,blk_bx=28,blk_in_chan=3,blk_y=13,blk_x=13)),in_pad=(tn=none,dims=(y=3,x=3)),in_ref=(dims=(img=1,chan=3,y=224,x=224)),out=(dims=(img=1,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(blk_bline=28,blk_bx=28,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00047936 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=28,blk_bx=28,blk_in_chan=3,blk_y=13,blk_x=13)),in_pad=(tn=none,dims=(y=3,x=3)),in_ref=(dims=(img=1,chan=3,y=224,x=224)),out=(dims=(img=1,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(blk_bline=28,blk_bx=28,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000203648 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=28,blk_bx=28,blk_in_chan=3,blk_y=13,blk_x=13)),in_pad=(tn=none,dims=(y=3,x=3)),in_ref=(dims=(img=1,chan=3,y=224,x=224)),out=(dims=(img=1,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(blk_bline=28,blk_bx=28,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00120672 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=1,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(pels_blk=784,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000400864 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=1,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(pels_blk=784,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00117984 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=1,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(pels_blk=784,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000403168 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=1,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(pels_blk=784,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00125296 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=1,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=1,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(pels_blk=98,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.0002584 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=1,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=1,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(pels_blk=98,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0006616 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=14,blk_in_chan=3,blk_y=37,blk_x=21)),in_pad=(tn=none,dims=(y=3,x=3)),in_ref=(dims=(img=1,chan=3,y=224,x=224)),out=(dims=(img=1,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(blk_bline=7,blk_bx=14,out_chan_blk=1,blk_y=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000177184 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=14,blk_in_chan=3,blk_y=37,blk_x=21)),in_pad=(tn=none,dims=(y=3,x=3)),in_ref=(dims=(img=1,chan=3,y=224,x=224)),out=(dims=(img=1,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(blk_bline=7,blk_bx=14,out_chan_blk=1,blk_y=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000367808 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=2,x=2)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000175712 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=28,blk_bx=28,blk_in_chan=3,blk_y=13,blk_x=13)),in_pad=(tn=none,dims=(y=3,x=3)),in_ref=(dims=(img=1,chan=3,y=224,x=224)),out=(dims=(img=1,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(blk_bline=28,blk_bx=28,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000175296 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=28,blk_bx=28,blk_in_chan=3,blk_y=13,blk_x=13)),in_pad=(tn=none,dims=(y=3,x=3)),in_ref=(dims=(img=1,chan=3,y=224,x=224)),out=(dims=(img=1,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(blk_bline=28,blk_bx=28,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000368672 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=1,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(pels_blk=784,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000370464 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=1,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(pels_blk=784,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000252832 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=1,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=1,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(pels_blk=98,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000162048 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=14,blk_in_chan=3,blk_y=37,blk_x=21)),in_pad=(tn=none,dims=(y=3,x=3)),in_ref=(dims=(img=1,chan=3,y=224,x=224)),out=(dims=(img=1,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(blk_bline=7,blk_bx=14,out_chan_blk=1,blk_y=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),filts=(dims=(out_chan=64,in_chan=3,y=7,x=7)),in=(dims=(img=5,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),kern_sz=(tn=none,dims=(y=7,x=7)),out=(dims=(img=5,chan=64,y=112,x=112)),out_chans=(tn=uint32_t,v=64),stride=(tn=none,dims=(y=2,x=2)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00258912 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=3,y=7,x=7,out_chan=64)),in=(dims=(chan=3,pel=261072)),in_pad=(tn=none,dims=(y=3,x=3)),in_pels=(tn=none,dims=(img=5,y=228,x=228)),out=(dims=(chan=64,pel=65024)),out_pels=(tn=none,dims=(img=5,y=114,x=114)),out_ref=(dims=(img=5,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=254,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00115741 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=3,y=7,x=7,out_chan=64)),in=(dims=(chan=3,pel=261072)),in_pad=(tn=none,dims=(y=3,x=3)),in_pels=(tn=none,dims=(img=5,y=228,x=228)),out=(dims=(chan=64,pel=65024)),out_pels=(tn=none,dims=(img=5,y=114,x=114)),out_ref=(dims=(img=5,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=254,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00165552 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=140,blk_bx=28,blk_in_chan=3,blk_y=13,blk_x=13)),in_pad=(tn=none,dims=(y=3,x=3)),in_ref=(dims=(img=5,chan=3,y=224,x=224)),out=(dims=(img=5,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(blk_bline=140,blk_bx=28,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000788576 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=140,blk_bx=28,blk_in_chan=3,blk_y=13,blk_x=13)),in_pad=(tn=none,dims=(y=3,x=3)),in_ref=(dims=(img=5,chan=3,y=224,x=224)),out=(dims=(img=5,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(blk_bline=140,blk_bx=28,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00117648 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=140,blk_bx=28,blk_in_chan=3,blk_y=13,blk_x=13)),in_pad=(tn=none,dims=(y=3,x=3)),in_ref=(dims=(img=5,chan=3,y=224,x=224)),out=(dims=(img=5,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(blk_bline=140,blk_bx=28,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000787424 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=140,blk_bx=28,blk_in_chan=3,blk_y=13,blk_x=13)),in_pad=(tn=none,dims=(y=3,x=3)),in_ref=(dims=(img=5,chan=3,y=224,x=224)),out=(dims=(img=5,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(blk_bline=140,blk_bx=28,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00396064 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=5,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(pels_blk=3920,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00173229 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=5,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(pels_blk=3920,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00371424 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=5,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(pels_blk=3920,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00173088 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=5,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(pels_blk=3920,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00287632 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=5,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=5,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(pels_blk=490,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00116208 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=5,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=5,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(pels_blk=490,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00219088 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=14,blk_in_chan=3,blk_y=37,blk_x=21)),in_pad=(tn=none,dims=(y=3,x=3)),in_ref=(dims=(img=5,chan=3,y=224,x=224)),out=(dims=(img=5,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(blk_bline=35,blk_bx=14,out_chan_blk=1,blk_y=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000623296 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=14,blk_in_chan=3,blk_y=37,blk_x=21)),in_pad=(tn=none,dims=(y=3,x=3)),in_ref=(dims=(img=5,chan=3,y=224,x=224)),out=(dims=(img=5,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(blk_bline=35,blk_bx=14,out_chan_blk=1,blk_y=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000672544 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=2,x=2)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000734656 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=140,blk_bx=28,blk_in_chan=3,blk_y=13,blk_x=13)),in_pad=(tn=none,dims=(y=3,x=3)),in_ref=(dims=(img=5,chan=3,y=224,x=224)),out=(dims=(img=5,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(blk_bline=140,blk_bx=28,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000735968 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=140,blk_bx=28,blk_in_chan=3,blk_y=13,blk_x=13)),in_pad=(tn=none,dims=(y=3,x=3)),in_ref=(dims=(img=5,chan=3,y=224,x=224)),out=(dims=(img=5,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(blk_bline=140,blk_bx=28,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00157357 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=5,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(pels_blk=3920,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00157898 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=5,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(pels_blk=3920,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00127565 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=5,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=5,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(pels_blk=490,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000604896 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=14,blk_in_chan=3,blk_y=37,blk_x=21)),in_pad=(tn=none,dims=(y=3,x=3)),in_ref=(dims=(img=5,chan=3,y=224,x=224)),out=(dims=(img=5,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(blk_bline=35,blk_bx=14,out_chan_blk=1,blk_y=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),filts=(dims=(out_chan=64,in_chan=3,y=7,x=7)),in=(dims=(img=20,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),kern_sz=(tn=none,dims=(y=7,x=7)),out=(dims=(img=20,chan=64,y=112,x=112)),out_chans=(tn=uint32_t,v=64),stride=(tn=none,dims=(y=2,x=2)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00573728 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=3,y=7,x=7,out_chan=64)),in=(dims=(chan=3,pel=1040832)),in_pad=(tn=none,dims=(y=3,x=3)),in_pels=(tn=none,dims=(img=20,y=228,x=228)),out=(dims=(chan=64,pel=260096)),out_pels=(tn=none,dims=(img=20,y=114,x=114)),out_ref=(dims=(img=20,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1016,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00420211 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=3,y=7,x=7,out_chan=64)),in=(dims=(chan=3,pel=1040832)),in_pad=(tn=none,dims=(y=3,x=3)),in_pels=(tn=none,dims=(img=20,y=228,x=228)),out=(dims=(chan=64,pel=260096)),out_pels=(tn=none,dims=(img=20,y=114,x=114)),out_ref=(dims=(img=20,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1016,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0057704 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=560,blk_bx=28,blk_in_chan=3,blk_y=13,blk_x=13)),in_pad=(tn=none,dims=(y=3,x=3)),in_ref=(dims=(img=20,chan=3,y=224,x=224)),out=(dims=(img=20,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(blk_bline=560,blk_bx=28,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00289299 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=560,blk_bx=28,blk_in_chan=3,blk_y=13,blk_x=13)),in_pad=(tn=none,dims=(y=3,x=3)),in_ref=(dims=(img=20,chan=3,y=224,x=224)),out=(dims=(img=20,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(blk_bline=560,blk_bx=28,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0024472 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=560,blk_bx=28,blk_in_chan=3,blk_y=13,blk_x=13)),in_pad=(tn=none,dims=(y=3,x=3)),in_ref=(dims=(img=20,chan=3,y=224,x=224)),out=(dims=(img=20,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(blk_bline=560,blk_bx=28,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00287933 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=560,blk_bx=28,blk_in_chan=3,blk_y=13,blk_x=13)),in_pad=(tn=none,dims=(y=3,x=3)),in_ref=(dims=(img=20,chan=3,y=224,x=224)),out=(dims=(img=20,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(blk_bline=560,blk_bx=28,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.012908 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=20,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(pels_blk=15680,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00661395 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=20,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(pels_blk=15680,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0067976 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=20,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(pels_blk=15680,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00660509 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=20,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(pels_blk=15680,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00993856 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=20,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=20,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(pels_blk=1960,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.0041327 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=20,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=20,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(pels_blk=1960,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0082784 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=140,blk_bx=14,blk_in_chan=3,blk_y=37,blk_x=21)),in_pad=(tn=none,dims=(y=3,x=3)),in_ref=(dims=(img=20,chan=3,y=224,x=224)),out=(dims=(img=20,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(blk_bline=140,blk_bx=14,out_chan_blk=1,blk_y=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00224339 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=140,blk_bx=14,blk_in_chan=3,blk_y=37,blk_x=21)),in_pad=(tn=none,dims=(y=3,x=3)),in_ref=(dims=(img=20,chan=3,y=224,x=224)),out=(dims=(img=20,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(blk_bline=140,blk_bx=14,out_chan_blk=1,blk_y=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00241059 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=2,x=2)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00275696 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=560,blk_bx=28,blk_in_chan=3,blk_y=13,blk_x=13)),in_pad=(tn=none,dims=(y=3,x=3)),in_ref=(dims=(img=20,chan=3,y=224,x=224)),out=(dims=(img=20,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(blk_bline=560,blk_bx=28,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00275802 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=560,blk_bx=28,blk_in_chan=3,blk_y=13,blk_x=13)),in_pad=(tn=none,dims=(y=3,x=3)),in_ref=(dims=(img=20,chan=3,y=224,x=224)),out=(dims=(img=20,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(blk_bline=560,blk_bx=28,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00610733 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=20,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(pels_blk=15680,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00610896 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=20,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(pels_blk=15680,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00454323 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=20,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=3,x=3)),out=(dims=(img=20,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(pels_blk=1960,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00222925 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=7,x=7,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=140,blk_bx=14,blk_in_chan=3,blk_y=37,blk_x=21)),in_pad=(tn=none,dims=(y=3,x=3)),in_ref=(dims=(img=20,chan=3,y=224,x=224)),out=(dims=(img=20,chan=64,y=112,x=112)),stride=(tn=none,dims=(y=2,x=2)),work=(tn=none,dims=(blk_bline=140,blk_bx=14,out_chan_blk=1,blk_y=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),filts=(dims=(out_chan=64,in_chan=192,y=1,x=1)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=1,chan=64,y=28,x=28)),out_chans=(tn=uint32_t,v=64),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0007184 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=192,y=1,x=1,out_chan=64)),in=(dims=(chan=192,pel=1024)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=1024)),out_ref=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 9.0368e-05 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=192,y=1,x=1,out_chan=64)),in=(dims=(chan=192,pel=1024)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=1024)),out_ref=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00066688 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=49,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),out_ref=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 3.6352e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=49,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),out_ref=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00097168 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=7,blk_in_chan=192,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=192,y=28,x=28)),out=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=7,blk_bx=7,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000178784 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=7,blk_in_chan=192,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=192,y=28,x=28)),out=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=7,blk_bx=7,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00066704 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=49,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),out_ref=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 3.504e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=49,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),out_ref=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00075872 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000148288 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00119056 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=7,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000182368 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=7,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0008624 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=7,blk_iter=24,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),out_ref=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=7,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 5.92e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=7,blk_iter=24,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),out_ref=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=7,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00010128 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 4.0448e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=49,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),out_ref=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000182656 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=7,blk_in_chan=192,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=192,y=28,x=28)),out=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=7,blk_bx=7,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 3.8912e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=49,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),out_ref=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00015008 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000183072 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=7,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 5.8912e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=7,blk_iter=24,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),out_ref=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=7,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),filts=(dims=(out_chan=64,in_chan=192,y=1,x=1)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=5,chan=64,y=28,x=28)),out_chans=(tn=uint32_t,v=64),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00072528 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=192,y=1,x=1,out_chan=64)),in=(dims=(chan=192,pel=4096)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=4096)),out_ref=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000132032 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=192,y=1,x=1,out_chan=64)),in=(dims=(chan=192,pel=4096)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=4096)),out_ref=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00068704 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),out_ref=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 8.7936e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),out_ref=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00097088 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=7,blk_in_chan=192,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=192,y=28,x=28)),out=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=35,blk_bx=7,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000242432 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=7,blk_in_chan=192,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=192,y=28,x=28)),out=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=35,blk_bx=7,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00068784 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),out_ref=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 8.6464e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),out_ref=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00089744 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00020544 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0013488 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00022288 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.000884 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=31,blk_iter=24,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),out_ref=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 9.1712e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=31,blk_iter=24,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),out_ref=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000143328 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 7.6832e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),out_ref=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000243616 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=7,blk_in_chan=192,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=192,y=28,x=28)),out=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=35,blk_bx=7,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 7.5424e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),out_ref=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000193312 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000215584 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 8.7008e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=31,blk_iter=24,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),out_ref=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),filts=(dims=(out_chan=64,in_chan=192,y=1,x=1)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=20,chan=64,y=28,x=28)),out_chans=(tn=uint32_t,v=64),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0007744 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=192,y=1,x=1,out_chan=64)),in=(dims=(chan=192,pel=15872)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=15872)),out_ref=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000326208 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=192,y=1,x=1,out_chan=64)),in=(dims=(chan=192,pel=15872)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=15872)),out_ref=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00116064 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=980,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),out_ref=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000285568 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=980,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),out_ref=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00120224 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=140,blk_bx=7,blk_in_chan=192,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=192,y=28,x=28)),out=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=140,blk_bx=7,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00060608 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=140,blk_bx=7,blk_in_chan=192,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=192,y=28,x=28)),out=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=140,blk_bx=7,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00103664 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=980,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),out_ref=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000289088 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=980,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),out_ref=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00115184 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000671392 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00156304 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000604672 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00107792 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=123,blk_iter=24,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),out_ref=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000205344 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=123,blk_iter=24,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),out_ref=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000209312 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000260544 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=980,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),out_ref=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00059552 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=140,blk_bx=7,blk_in_chan=192,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=192,y=28,x=28)),out=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=140,blk_bx=7,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000258912 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=980,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),out_ref=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000510592 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000594144 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000200416 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=123,blk_iter=24,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),out_ref=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),filts=(dims=(out_chan=64,in_chan=256,y=1,x=1)),in=(dims=(img=1,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=1,chan=64,y=28,x=28)),out_chans=(tn=uint32_t,v=64),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00087728 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=256,y=1,x=1,out_chan=64)),in=(dims=(chan=256,pel=1024)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=1024)),out_ref=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000115392 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=256,y=1,x=1,out_chan=64)),in=(dims=(chan=256,pel=1024)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=1024)),out_ref=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.000432 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=49,blk_iter=32,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),out_ref=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 4.1696e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=49,blk_iter=32,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),out_ref=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00115168 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=7,blk_in_chan=256,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=256,y=28,x=28)),out=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=7,blk_bx=7,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000236672 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=7,blk_in_chan=256,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=256,y=28,x=28)),out=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=7,blk_bx=7,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00043136 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=49,blk_iter=32,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),out_ref=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 4.2976e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=49,blk_iter=32,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),out_ref=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00029776 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000197376 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00149536 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=1,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=7,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000236672 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=1,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=7,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00106848 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=7,blk_iter=32,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),out_ref=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=7,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 7.0944e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=7,blk_iter=32,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),out_ref=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=7,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000109856 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 4.7072e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=49,blk_iter=32,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),out_ref=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000244416 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=7,blk_in_chan=256,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=256,y=28,x=28)),out=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=7,blk_bx=7,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 4.5696e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=49,blk_iter=32,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),out_ref=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000195584 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000237472 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=1,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=7,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 6.8576e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=7,blk_iter=32,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=28,x=28)),out_ref=(dims=(img=1,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=7,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),filts=(dims=(out_chan=64,in_chan=256,y=1,x=1)),in=(dims=(img=5,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=5,chan=64,y=28,x=28)),out_chans=(tn=uint32_t,v=64),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.000896 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=256,y=1,x=1,out_chan=64)),in=(dims=(chan=256,pel=4096)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=4096)),out_ref=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000167136 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=256,y=1,x=1,out_chan=64)),in=(dims=(chan=256,pel=4096)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=4096)),out_ref=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00043584 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=32,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),out_ref=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 9.9936e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=32,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),out_ref=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0012032 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=7,blk_in_chan=256,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=256,y=28,x=28)),out=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=35,blk_bx=7,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000320704 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=7,blk_in_chan=256,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=256,y=28,x=28)),out=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=35,blk_bx=7,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00043488 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=32,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),out_ref=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000105088 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=32,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),out_ref=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00113984 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000280704 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00170928 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=5,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000295232 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=5,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0011144 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=31,blk_iter=32,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),out_ref=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000107328 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=31,blk_iter=32,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),out_ref=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000161312 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 8.9184e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=32,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),out_ref=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000323552 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=7,blk_in_chan=256,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=256,y=28,x=28)),out=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=35,blk_bx=7,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 8.9696e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=32,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),out_ref=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000274656 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000294528 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=5,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000105664 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=31,blk_iter=32,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=28,x=28)),out_ref=(dims=(img=5,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),filts=(dims=(out_chan=64,in_chan=256,y=1,x=1)),in=(dims=(img=20,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=20,chan=64,y=28,x=28)),out_chans=(tn=uint32_t,v=64),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00097104 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=256,y=1,x=1,out_chan=64)),in=(dims=(chan=256,pel=15872)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=15872)),out_ref=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000410688 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=256,y=1,x=1,out_chan=64)),in=(dims=(chan=256,pel=15872)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=15872)),out_ref=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00078576 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=980,blk_iter=32,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),out_ref=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000330272 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=980,blk_iter=32,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),out_ref=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0016728 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=140,blk_bx=7,blk_in_chan=256,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=256,y=28,x=28)),out=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=140,blk_bx=7,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00076832 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=140,blk_bx=7,blk_in_chan=256,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=256,y=28,x=28)),out=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=140,blk_bx=7,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.000484 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=980,blk_iter=32,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),out_ref=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000323648 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=980,blk_iter=32,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),out_ref=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0014848 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00086448 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00193824 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=20,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000740768 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=20,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00131888 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=123,blk_iter=32,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),out_ref=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00024272 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=123,blk_iter=32,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),out_ref=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000244512 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000310048 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=980,blk_iter=32,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),out_ref=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000757248 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=140,blk_bx=7,blk_in_chan=256,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=256,y=28,x=28)),out=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=140,blk_bx=7,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000310112 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=980,blk_iter=32,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),out_ref=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000642592 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000754752 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=20,chan=256,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000237888 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=256,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=123,blk_iter=32,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=28,x=28)),out_ref=(dims=(img=20,chan=64,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=123,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),filts=(dims=(out_chan=64,in_chan=480,y=1,x=1)),in=(dims=(img=1,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=1,chan=64,y=14,x=14)),out_chans=(tn=uint32_t,v=64),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00161408 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=480,y=1,x=1,out_chan=64)),in=(dims=(chan=480,pel=200)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=200)),out_ref=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=25,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000184928 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=480,y=1,x=1,out_chan=64)),in=(dims=(chan=480,pel=200)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=200)),out_ref=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=25,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00062832 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=13,blk_iter=60,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),out_ref=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 5.7312e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=13,blk_iter=60,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),out_ref=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00214816 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=4,blk_in_chan=480,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=480,y=14,x=14)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000436352 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=4,blk_in_chan=480,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=480,y=14,x=14)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00062576 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=13,blk_iter=60,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),out_ref=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 5.9712e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=13,blk_iter=60,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),out_ref=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00163968 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000351552 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00264288 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=1,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000426464 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=1,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0018056 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=2,blk_iter=60,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),out_ref=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000107616 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=2,blk_iter=60,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),out_ref=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00015888 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 6.6912e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=13,blk_iter=60,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),out_ref=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00043184 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=4,blk_in_chan=480,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=480,y=14,x=14)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 6.6016e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=13,blk_iter=60,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),out_ref=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000350592 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000427808 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=1,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00010816 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=2,blk_iter=60,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),out_ref=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),filts=(dims=(out_chan=64,in_chan=480,y=1,x=1)),in=(dims=(img=5,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=5,chan=64,y=14,x=14)),out_chans=(tn=uint32_t,v=64),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
| 0105000000666C6F61740101DADA000000000000000004000000050000000031000003000000696D6740000000C4000000040000006368616E0E0000000E00000001000000790E00000001000000010000007805000000666C6F617400F5000000000000018D905F1F69F48806000000009F164F444200000058A6894A34020A4A25C1B7496D3CB549FCC957497A865B49F5111D49B1031D49483CCB48E14CC1488E78CB48F253A448793BAC48C06FAA48BF329E48215E994817EC9D48D11A8048B78982481AFB6D48845767486D4D6348D8DB6948B04543488F394848004E4148575D40488153144806671C482D461848603B174816C5A9462C22B346D98EAC46EE3FB046083FB5468252B5464C13CF46828F00439CEE9143B199E143ECFE7E42538A86447E968F42DAA1B343F8030744420B1943F9192F44C004FE4306200643E4BE544409D23D43CBF93843FDA91442000000000000000000000000E2A35642D1C0BE42E6A28D43000000009D5BE8423D86E842C665B14200000000795D5942 | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00160368 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=480,y=1,x=1,out_chan=64)),in=(dims=(chan=480,pel=1024)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=1024)),out_ref=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000269376 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=480,y=1,x=1,out_chan=64)),in=(dims=(chan=480,pel=1024)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=1024)),out_ref=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00065552 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=62,blk_iter=60,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),out_ref=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 7.9296e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=62,blk_iter=60,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),out_ref=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00218976 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=480,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=480,y=14,x=14)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000509056 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=480,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=480,y=14,x=14)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00065312 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=62,blk_iter=60,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),out_ref=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 7.92e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=62,blk_iter=60,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),out_ref=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00176032 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00035616 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00274912 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=5,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000430208 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=5,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00183168 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=8,blk_iter=60,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),out_ref=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00013184 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=8,blk_iter=60,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),out_ref=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000147136 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 8.4352e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=62,blk_iter=60,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),out_ref=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000493568 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=480,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=480,y=14,x=14)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 8.1408e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=62,blk_iter=60,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),out_ref=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000359968 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000430912 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=5,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000124064 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=8,blk_iter=60,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),out_ref=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),filts=(dims=(out_chan=64,in_chan=480,y=1,x=1)),in=(dims=(img=20,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=20,chan=64,y=14,x=14)),out_chans=(tn=uint32_t,v=64),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00161712 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=480,y=1,x=1,out_chan=64)),in=(dims=(chan=480,pel=4096)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=4096)),out_ref=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000287872 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=480,y=1,x=1,out_chan=64)),in=(dims=(chan=480,pel=4096)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=4096)),out_ref=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00065712 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=60,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),out_ref=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00014848 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=60,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),out_ref=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00253424 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=4,blk_in_chan=480,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=480,y=14,x=14)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000602816 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=4,blk_in_chan=480,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=480,y=14,x=14)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0004184 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=60,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),out_ref=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000147776 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=60,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),out_ref=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00221552 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.0005008 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00314736 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=20,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000552896 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=20,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00192736 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=31,blk_iter=60,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),out_ref=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000162144 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=31,blk_iter=60,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),out_ref=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00020832 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000134752 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=60,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),out_ref=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00058912 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=4,blk_in_chan=480,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=480,y=14,x=14)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000134144 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=60,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),out_ref=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000493568 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00054432 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=20,chan=480,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000153952 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=480,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=31,blk_iter=60,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),out_ref=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),filts=(dims=(out_chan=64,in_chan=512,y=1,x=1)),in=(dims=(img=1,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=1,chan=64,y=14,x=14)),out_chans=(tn=uint32_t,v=64),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0016456 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=512,y=1,x=1,out_chan=64)),in=(dims=(chan=512,pel=200)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=200)),out_ref=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=25,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000195424 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=512,y=1,x=1,out_chan=64)),in=(dims=(chan=512,pel=200)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=200)),out_ref=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=25,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00061968 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=13,blk_iter=64,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),out_ref=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 6.2304e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=13,blk_iter=64,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),out_ref=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00217152 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=4,blk_in_chan=512,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=512,y=14,x=14)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00045856 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=4,blk_in_chan=512,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=512,y=14,x=14)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00061264 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=13,blk_iter=64,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),out_ref=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 6.2208e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=13,blk_iter=64,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),out_ref=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.001732 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000372992 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00275168 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=1,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000453856 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=1,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0019088 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=2,blk_iter=64,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),out_ref=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000114688 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=2,blk_iter=64,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),out_ref=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00016192 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 7.0592e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=13,blk_iter=64,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),out_ref=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000458368 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=4,blk_in_chan=512,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=512,y=14,x=14)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 6.9856e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=13,blk_iter=64,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),out_ref=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000371872 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000454016 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=1,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000110976 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=2,blk_iter=64,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=64,y=14,x=14)),out_ref=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),filts=(dims=(out_chan=64,in_chan=512,y=1,x=1)),in=(dims=(img=5,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=5,chan=64,y=14,x=14)),out_chans=(tn=uint32_t,v=64),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
| 0105000000666C6F61740101DADA000000000000000004000000050000000031000003000000696D6740000000C4000000040000006368616E0E0000000E00000001000000790E00000001000000010000007805000000666C6F617400F5000000000000018D905F1F69F488060000000049EF53444200000031F8904A661C114AA5F3C14933F3C0498A196749B8D2664922A12649913122493325D74871A9D348238ECE4852ECBA48D9FBB248CE7CB2482D5FA7482A9AA548356BA548D871894892CC8848350E7B48E48873486EA6724854CF704859BE514840624948B36E43482C7A4B485A22224826001E482AE61B48B2791F480B34B1466F66C546CE75BE468ECDB946C431BC46B298C846B3CAC046D0432844D7326A43F00D94427EDD56442AE82A4398562B430B5BD443227A8D4355978A43F9EDF64390B18B439E744E431CD80F4400000000E7B70D4375CB4943E58D8342FBF1BD429E621743FE0A7B4348F1E4420000000000000000B229E942F66325430000000000000000CA3DC242 | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00163872 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=512,y=1,x=1,out_chan=64)),in=(dims=(chan=512,pel=1024)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=1024)),out_ref=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00028336 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=512,y=1,x=1,out_chan=64)),in=(dims=(chan=512,pel=1024)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=1024)),out_ref=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=4,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00063376 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=62,blk_iter=64,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),out_ref=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 8.4576e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=62,blk_iter=64,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),out_ref=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00220224 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=512,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=512,y=14,x=14)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000537248 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=512,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=512,y=14,x=14)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00063568 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=62,blk_iter=64,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),out_ref=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 8.3872e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=62,blk_iter=64,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),out_ref=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00187776 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000389184 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00287456 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=5,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00045728 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=5,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0019424 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=8,blk_iter=64,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),out_ref=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000139552 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=8,blk_iter=64,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),out_ref=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00015264 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 9.0208e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=62,blk_iter=64,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),out_ref=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000528288 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=512,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=512,y=14,x=14)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 8.7648e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=62,blk_iter=64,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),out_ref=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00038384 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00045824 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=5,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000127488 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=8,blk_iter=64,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=64,y=14,x=14)),out_ref=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),filts=(dims=(out_chan=64,in_chan=512,y=1,x=1)),in=(dims=(img=20,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=20,chan=64,y=14,x=14)),out_chans=(tn=uint32_t,v=64),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00168384 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=512,y=1,x=1,out_chan=64)),in=(dims=(chan=512,pel=4096)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=4096)),out_ref=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000301888 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=512,y=1,x=1,out_chan=64)),in=(dims=(chan=512,pel=4096)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=64,pel=4096)),out_ref=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=16,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00064 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=64,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),out_ref=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000152768 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=64,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),out_ref=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00252528 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=4,blk_in_chan=512,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=512,y=14,x=14)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000631968 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=4,blk_in_chan=512,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=512,y=14,x=14)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00039472 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=64,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),out_ref=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000154528 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=64,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),out_ref=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00233104 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000532384 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0033104 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=20,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000583392 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=20,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00205904 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=31,blk_iter=64,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),out_ref=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000166208 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=31,blk_iter=64,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),out_ref=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000211424 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000141312 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=64,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),out_ref=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000623648 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=4,blk_in_chan=512,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=512,y=14,x=14)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000141408 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=64,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),out_ref=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000527904 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000582848 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=20,chan=512,y=14,x=14)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000168928 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=512,y=1,x=1,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk=31,blk_iter=64,blk_iter_chan=8,blk_pel=128)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=64,y=14,x=14)),out_ref=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),filts=(dims=(out_chan=64,in_chan=24,y=5,x=5)),in=(dims=(img=1,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),kern_sz=(tn=none,dims=(y=5,x=5)),out=(dims=(img=1,chan=64,y=14,x=14)),out_chans=(tn=uint32_t,v=64),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00235536 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=24,y=5,x=5,out_chan=64)),in=(dims=(chan=24,pel=328)),in_pad=(tn=none,dims=(y=2,x=2)),in_pels=(tn=none,dims=(img=1,y=16,x=16)),out=(dims=(chan=64,pel=256)),out_pels=(tn=none,dims=(img=1,y=16,x=16)),out_ref=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000325408 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=24,y=5,x=5,out_chan=64)),in=(dims=(chan=24,pel=328)),in_pad=(tn=none,dims=(y=2,x=2)),in_pels=(tn=none,dims=(img=1,y=16,x=16)),out=(dims=(chan=64,pel=256)),out_pels=(tn=none,dims=(img=1,y=16,x=16)),out_ref=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00097968 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=4,blk_in_chan=24,blk_y=8,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=24,y=14,x=14)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 9.8688e-05 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=4,blk_in_chan=24,blk_y=8,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=24,y=14,x=14)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00097216 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=4,blk_in_chan=24,blk_y=8,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=24,y=14,x=14)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 9.7376e-05 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=4,blk_in_chan=24,blk_y=8,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=24,y=14,x=14)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00255536 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000524544 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00253472 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000526784 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00355664 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=1,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000614336 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=1,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00194352 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=2,blk_in_chan=24,blk_y=20,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=24,y=14,x=14)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=2,out_chan_blk=1,blk_y=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000153664 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=2,blk_in_chan=24,blk_y=20,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=24,y=14,x=14)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=2,out_chan_blk=1,blk_y=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00012848 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 9.8592e-05 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=4,blk_in_chan=24,blk_y=8,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=24,y=14,x=14)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 9.5776e-05 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=4,blk_in_chan=24,blk_y=8,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=24,y=14,x=14)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000468224 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000465408 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000553088 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=1,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000141952 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=2,blk_in_chan=24,blk_y=20,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=24,y=14,x=14)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=2,out_chan_blk=1,blk_y=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),filts=(dims=(out_chan=64,in_chan=24,y=5,x=5)),in=(dims=(img=5,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),kern_sz=(tn=none,dims=(y=5,x=5)),out=(dims=(img=5,chan=64,y=14,x=14)),out_chans=(tn=uint32_t,v=64),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00234848 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=24,y=5,x=5,out_chan=64)),in=(dims=(chan=24,pel=1352)),in_pad=(tn=none,dims=(y=2,x=2)),in_pels=(tn=none,dims=(img=5,y=16,x=16)),out=(dims=(chan=64,pel=1280)),out_pels=(tn=none,dims=(img=5,y=16,x=16)),out_ref=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=5,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000314784 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=24,y=5,x=5,out_chan=64)),in=(dims=(chan=24,pel=1352)),in_pad=(tn=none,dims=(y=2,x=2)),in_pels=(tn=none,dims=(img=5,y=16,x=16)),out=(dims=(chan=64,pel=1280)),out_pels=(tn=none,dims=(img=5,y=16,x=16)),out_ref=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=5,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0011024 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=24,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=24,y=14,x=14)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000109248 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=24,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=24,y=14,x=14)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00109104 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=24,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=24,y=14,x=14)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000106816 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=24,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=24,y=14,x=14)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00254528 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00054368 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00254304 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000541536 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00355632 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=5,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000620512 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=5,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00191872 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=2,blk_in_chan=24,blk_y=24,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=24,y=14,x=14)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=2,out_chan_blk=1,blk_y=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.0001576 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=2,blk_in_chan=24,blk_y=24,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=24,y=14,x=14)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=2,out_chan_blk=1,blk_y=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00012688 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000112576 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=24,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=24,y=14,x=14)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000109984 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=24,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=24,y=14,x=14)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000508448 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000508288 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00055872 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=5,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00014672 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=2,blk_in_chan=24,blk_y=24,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=24,y=14,x=14)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=2,out_chan_blk=1,blk_y=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),filts=(dims=(out_chan=64,in_chan=24,y=5,x=5)),in=(dims=(img=20,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),kern_sz=(tn=none,dims=(y=5,x=5)),out=(dims=(img=20,chan=64,y=14,x=14)),out_chans=(tn=uint32_t,v=64),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00235696 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=24,y=5,x=5,out_chan=64)),in=(dims=(chan=24,pel=5192)),in_pad=(tn=none,dims=(y=2,x=2)),in_pels=(tn=none,dims=(img=20,y=16,x=16)),out=(dims=(chan=64,pel=5120)),out_pels=(tn=none,dims=(img=20,y=16,x=16)),out_ref=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=20,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000441728 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=24,y=5,x=5,out_chan=64)),in=(dims=(chan=24,pel=5192)),in_pad=(tn=none,dims=(y=2,x=2)),in_pels=(tn=none,dims=(img=20,y=16,x=16)),out=(dims=(chan=64,pel=5120)),out_pels=(tn=none,dims=(img=20,y=16,x=16)),out_ref=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=20,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00115664 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=4,blk_in_chan=24,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=24,y=14,x=14)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000189984 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=4,blk_in_chan=24,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=24,y=14,x=14)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0011544 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=4,blk_in_chan=24,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=24,y=14,x=14)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000190208 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=4,blk_in_chan=24,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=24,y=14,x=14)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00258432 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000600096 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.002564 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00059968 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00354992 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=20,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000653952 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=20,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00195008 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=2,blk_in_chan=24,blk_y=24,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=24,y=14,x=14)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=2,out_chan_blk=1,blk_y=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000201248 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=2,blk_in_chan=24,blk_y=24,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=24,y=14,x=14)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=2,out_chan_blk=1,blk_y=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000206368 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000172192 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=4,blk_in_chan=24,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=24,y=14,x=14)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00017312 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=4,blk_in_chan=24,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=24,y=14,x=14)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000548192 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000540192 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000650752 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=20,chan=24,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000191456 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=24,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=2,blk_in_chan=24,blk_y=24,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=24,y=14,x=14)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=2,out_chan_blk=1,blk_y=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),filts=(dims=(out_chan=64,in_chan=32,y=5,x=5)),in=(dims=(img=1,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),kern_sz=(tn=none,dims=(y=5,x=5)),out=(dims=(img=1,chan=64,y=14,x=14)),out_chans=(tn=uint32_t,v=64),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00306384 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=32,y=5,x=5,out_chan=64)),in=(dims=(chan=32,pel=328)),in_pad=(tn=none,dims=(y=2,x=2)),in_pels=(tn=none,dims=(img=1,y=16,x=16)),out=(dims=(chan=64,pel=256)),out_pels=(tn=none,dims=(img=1,y=16,x=16)),out_ref=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00039936 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=32,y=5,x=5,out_chan=64)),in=(dims=(chan=32,pel=328)),in_pad=(tn=none,dims=(y=2,x=2)),in_pels=(tn=none,dims=(img=1,y=16,x=16)),out=(dims=(chan=64,pel=256)),out_pels=(tn=none,dims=(img=1,y=16,x=16)),out_ref=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=1,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00090384 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=4,blk_in_chan=32,blk_y=8,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=32,y=14,x=14)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00012736 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=4,blk_in_chan=32,blk_y=8,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=32,y=14,x=14)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0008936 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=4,blk_in_chan=32,blk_y=8,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=32,y=14,x=14)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000126464 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=4,blk_in_chan=32,blk_y=8,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=32,y=14,x=14)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00332032 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000696032 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00331648 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000699296 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00464768 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=1,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000816256 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=1,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0025304 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=2,blk_in_chan=32,blk_y=20,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=32,y=14,x=14)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=2,out_chan_blk=1,blk_y=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000197152 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=2,blk_in_chan=32,blk_y=20,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=32,y=14,x=14)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=2,out_chan_blk=1,blk_y=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000137312 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000125856 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=4,blk_in_chan=32,blk_y=8,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=32,y=14,x=14)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000123808 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=4,blk_bx=4,blk_in_chan=32,blk_y=8,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=32,y=14,x=14)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=4,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000619936 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000619456 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=13,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000730784 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=1,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=2,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.0001808 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=1,blk_bx=2,blk_in_chan=32,blk_y=20,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=1,chan=32,y=14,x=14)),out=(dims=(img=1,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=1,blk_bx=2,out_chan_blk=1,blk_y=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),filts=(dims=(out_chan=64,in_chan=32,y=5,x=5)),in=(dims=(img=5,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),kern_sz=(tn=none,dims=(y=5,x=5)),out=(dims=(img=5,chan=64,y=14,x=14)),out_chans=(tn=uint32_t,v=64),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0030416 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=32,y=5,x=5,out_chan=64)),in=(dims=(chan=32,pel=1352)),in_pad=(tn=none,dims=(y=2,x=2)),in_pels=(tn=none,dims=(img=5,y=16,x=16)),out=(dims=(chan=64,pel=1280)),out_pels=(tn=none,dims=(img=5,y=16,x=16)),out_ref=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=5,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000458432 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=32,y=5,x=5,out_chan=64)),in=(dims=(chan=32,pel=1352)),in_pad=(tn=none,dims=(y=2,x=2)),in_pels=(tn=none,dims=(img=5,y=16,x=16)),out=(dims=(chan=64,pel=1280)),out_pels=(tn=none,dims=(img=5,y=16,x=16)),out_ref=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=5,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00102336 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=32,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=32,y=14,x=14)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000138912 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=32,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=32,y=14,x=14)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00100688 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=32,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=32,y=14,x=14)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00014 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=32,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=32,y=14,x=14)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00331856 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000723296 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00330304 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000723392 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00460704 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=5,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00081968 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=5,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00249584 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=2,blk_in_chan=32,blk_y=24,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=32,y=14,x=14)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=2,out_chan_blk=1,blk_y=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000201696 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=2,blk_in_chan=32,blk_y=24,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=32,y=14,x=14)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=2,out_chan_blk=1,blk_y=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000137504 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00014416 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=32,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=32,y=14,x=14)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000141472 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=4,blk_in_chan=32,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=32,y=14,x=14)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000674176 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000675104 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=62,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000738752 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=5,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=8,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000186304 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=5,blk_bx=2,blk_in_chan=32,blk_y=24,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=5,chan=32,y=14,x=14)),out=(dims=(img=5,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=5,blk_bx=2,out_chan_blk=1,blk_y=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),filts=(dims=(out_chan=64,in_chan=32,y=5,x=5)),in=(dims=(img=20,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),kern_sz=(tn=none,dims=(y=5,x=5)),out=(dims=(img=20,chan=64,y=14,x=14)),out_chans=(tn=uint32_t,v=64),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00306704 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=32,y=5,x=5,out_chan=64)),in=(dims=(chan=32,pel=5192)),in_pad=(tn=none,dims=(y=2,x=2)),in_pels=(tn=none,dims=(img=20,y=16,x=16)),out=(dims=(chan=64,pel=5120)),out_pels=(tn=none,dims=(img=20,y=16,x=16)),out_ref=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=20,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000570784 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=32,y=5,x=5,out_chan=64)),in=(dims=(chan=32,pel=5192)),in_pad=(tn=none,dims=(y=2,x=2)),in_pels=(tn=none,dims=(img=20,y=16,x=16)),out=(dims=(chan=64,pel=5120)),out_pels=(tn=none,dims=(img=20,y=16,x=16)),out_ref=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=20,out_chan_blk=1,pels_tile=32,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00108896 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=4,blk_in_chan=32,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=32,y=14,x=14)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000236704 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=4,blk_in_chan=32,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=32,y=14,x=14)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00108384 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=4,blk_in_chan=32,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=32,y=14,x=14)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000232864 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=4,blk_in_chan=32,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=32,y=14,x=14)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00336 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000781472 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00341664 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000783328 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00460848 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=20,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000855872 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=20,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0025384 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=2,blk_in_chan=32,blk_y=24,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=32,y=14,x=14)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=2,out_chan_blk=1,blk_y=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000254944 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=2,blk_in_chan=32,blk_y=24,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=32,y=14,x=14)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=2,out_chan_blk=1,blk_y=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000251616 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000222656 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=4,blk_in_chan=32,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=32,y=14,x=14)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000222208 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=70,blk_bx=4,blk_in_chan=32,blk_y=12,blk_x=8)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=32,y=14,x=14)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=70,blk_bx=4,out_chan_blk=1,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000714784 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000714752 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=1,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000863936 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),in=(dims=(img=20,chan=32,y=14,x=14)),in_pad=(tn=none,dims=(y=2,x=2)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=31,out_chan_blk=1,pels_tile=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000251328 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=64)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=32,y=5,x=5,out_chan_reg=8,out_chan_tile=8)),flags=(tn=uint32_t),in=(dims=(blk_bline=18,blk_bx=2,blk_in_chan=32,blk_y=24,blk_x=12)),in_pad=(tn=none,dims=(y=2,x=2)),in_ref=(dims=(img=20,chan=32,y=14,x=14)),out=(dims=(img=20,chan=64,y=14,x=14)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=18,blk_bx=2,out_chan_blk=1,blk_y=16,out_chan_tile=8,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),filts=(dims=(out_chan=96,in_chan=96,y=1,x=1)),in=(dims=(img=1,chan=96,y=54,x=54)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=1,chan=96,y=54,x=54)),out_chans=(tn=uint32_t,v=96),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00042016 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=96,y=1,x=1,out_chan=96)),in=(dims=(chan=96,pel=3024)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=96,pel=3024)),out_ref=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=18,out_chan_blk=1,pels_tile=21,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 7.2832e-05 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=96,y=1,x=1,out_chan=96)),in=(dims=(chan=96,pel=3024)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=96,pel=3024)),out_ref=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=18,out_chan_blk=1,pels_tile=21,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.000372 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=183,blk_iter=12,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=54,x=54)),out_ref=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=183,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 8.464e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=183,blk_iter=12,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=54,x=54)),out_ref=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=183,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00063136 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=14,blk_bx=14,blk_in_chan=96,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=96,y=54,x=54)),out=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=14,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000144608 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=14,blk_bx=14,blk_in_chan=96,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=96,y=54,x=54)),out=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=14,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00037344 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=183,blk_iter=12,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=54,x=54)),out_ref=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=183,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 8.448e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=183,blk_iter=12,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=54,x=54)),out_ref=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=183,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0007184 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=96,y=54,x=54)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=183,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000131584 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=96,y=54,x=54)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=183,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00066352 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=96,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=1,chan=96,y=54,x=54)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=37,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000132736 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=96,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=1,chan=96,y=54,x=54)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=37,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00052048 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=96,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk=37,blk_iter=12,blk_iter_chan=8,blk_pel=80)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=54,x=54)),out_ref=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=37,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 6.9728e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=96,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk=37,blk_iter=12,blk_iter_chan=8,blk_pel=80)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=54,x=54)),out_ref=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=37,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 8.624e-05 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 7.6704e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=183,blk_iter=12,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=54,x=54)),out_ref=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=183,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000145568 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=14,blk_bx=14,blk_in_chan=96,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=96,y=54,x=54)),out=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=14,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 7.6352e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=183,blk_iter=12,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=54,x=54)),out_ref=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=183,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000122816 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=96,y=54,x=54)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=183,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000134912 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=96,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=1,chan=96,y=54,x=54)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=37,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 7.344e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=96,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk=37,blk_iter=12,blk_iter_chan=8,blk_pel=80)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=54,x=54)),out_ref=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=37,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),filts=(dims=(out_chan=96,in_chan=96,y=1,x=1)),in=(dims=(img=5,chan=96,y=54,x=54)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=5,chan=96,y=54,x=54)),out_chans=(tn=uint32_t,v=96),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0004784 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=96,y=1,x=1,out_chan=96)),in=(dims=(chan=96,pel=14616)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=96,pel=14616)),out_ref=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=87,out_chan_blk=1,pels_tile=21,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000240512 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=96,y=1,x=1,out_chan=96)),in=(dims=(chan=96,pel=14616)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=96,pel=14616)),out_ref=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=87,out_chan_blk=1,pels_tile=21,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00075248 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=912,blk_iter=12,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=54,x=54)),out_ref=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=912,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.0003192 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=912,blk_iter=12,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=54,x=54)),out_ref=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=912,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00137824 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=68,blk_bx=14,blk_in_chan=96,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=96,y=54,x=54)),out=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=68,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000583584 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=68,blk_bx=14,blk_in_chan=96,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=96,y=54,x=54)),out=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=68,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00075632 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=912,blk_iter=12,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=54,x=54)),out_ref=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=912,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000312 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=912,blk_iter=12,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=54,x=54)),out_ref=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=912,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00137296 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=96,y=54,x=54)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=912,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000529408 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=96,y=54,x=54)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=912,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00134896 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=96,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=5,chan=96,y=54,x=54)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=183,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000486016 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=96,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=5,chan=96,y=54,x=54)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=183,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00105616 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=96,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk=183,blk_iter=12,blk_iter_chan=8,blk_pel=80)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=54,x=54)),out_ref=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=183,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00023152 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=96,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk=183,blk_iter=12,blk_iter_chan=8,blk_pel=80)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=54,x=54)),out_ref=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=183,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000244128 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.0002688 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=912,blk_iter=12,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=54,x=54)),out_ref=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=912,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000577632 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=68,blk_bx=14,blk_in_chan=96,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=96,y=54,x=54)),out=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=68,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000270944 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=912,blk_iter=12,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=54,x=54)),out_ref=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=912,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000516448 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=96,y=54,x=54)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=912,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000446464 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=96,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=5,chan=96,y=54,x=54)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=183,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000226112 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=96,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk=183,blk_iter=12,blk_iter_chan=8,blk_pel=80)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=54,x=54)),out_ref=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=183,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),filts=(dims=(out_chan=96,in_chan=96,y=1,x=1)),in=(dims=(img=20,chan=96,y=54,x=54)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=20,chan=96,y=54,x=54)),out_chans=(tn=uint32_t,v=96),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00136016 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=96,y=1,x=1,out_chan=96)),in=(dims=(chan=96,pel=58464)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=96,pel=58464)),out_ref=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=348,out_chan_blk=1,pels_tile=21,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000749888 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=96,y=1,x=1,out_chan=96)),in=(dims=(chan=96,pel=58464)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=96,pel=58464)),out_ref=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=348,out_chan_blk=1,pels_tile=21,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00200656 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=3645,blk_iter=12,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=54,x=54)),out_ref=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=3645,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00104922 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=3645,blk_iter=12,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=54,x=54)),out_ref=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=3645,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0042976 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=270,blk_bx=14,blk_in_chan=96,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=96,y=54,x=54)),out=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=270,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00202688 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=270,blk_bx=14,blk_in_chan=96,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=96,y=54,x=54)),out=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=270,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00120816 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=3645,blk_iter=12,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=54,x=54)),out_ref=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=3645,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00102864 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=3645,blk_iter=12,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=54,x=54)),out_ref=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=3645,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00351632 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=96,y=54,x=54)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=3645,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00191894 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=96,y=54,x=54)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=3645,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00373088 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=96,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=20,chan=96,y=54,x=54)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=729,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00159552 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=96,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=20,chan=96,y=54,x=54)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=729,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00329808 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=96,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk=729,blk_iter=12,blk_iter_chan=8,blk_pel=80)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=54,x=54)),out_ref=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=729,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00077472 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=96,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk=729,blk_iter=12,blk_iter_chan=8,blk_pel=80)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=54,x=54)),out_ref=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=729,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00080288 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000893472 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=3645,blk_iter=12,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=54,x=54)),out_ref=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=3645,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00201584 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=270,blk_bx=14,blk_in_chan=96,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=96,y=54,x=54)),out=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=270,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000901248 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=3645,blk_iter=12,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=54,x=54)),out_ref=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=3645,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.0019057 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=96,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=96,y=54,x=54)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=3645,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00178493 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=96,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=20,chan=96,y=54,x=54)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=729,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00076608 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=96,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk=729,blk_iter=12,blk_iter_chan=8,blk_pel=80)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=54,x=54)),out_ref=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=729,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),filts=(dims=(out_chan=96,in_chan=192,y=1,x=1)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=1,chan=96,y=28,x=28)),out_chans=(tn=uint32_t,v=96),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00069344 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=192,y=1,x=1,out_chan=96)),in=(dims=(chan=192,pel=840)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=96,pel=840)),out_ref=(dims=(img=1,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=5,out_chan_blk=1,pels_tile=21,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 9.4752e-05 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=192,y=1,x=1,out_chan=96)),in=(dims=(chan=192,pel=840)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=96,pel=840)),out_ref=(dims=(img=1,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=5,out_chan_blk=1,pels_tile=21,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.000668 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=49,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=28,x=28)),out_ref=(dims=(img=1,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 4.8224e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=49,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=28,x=28)),out_ref=(dims=(img=1,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00094336 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=7,blk_in_chan=192,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=192,y=28,x=28)),out=(dims=(img=1,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=7,blk_bx=7,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000188832 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=7,blk_in_chan=192,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=192,y=28,x=28)),out=(dims=(img=1,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=7,blk_bx=7,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00066784 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=49,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=28,x=28)),out_ref=(dims=(img=1,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 4.7456e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=49,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=28,x=28)),out_ref=(dims=(img=1,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00087664 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.0001568 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00109984 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=10,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000194688 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=10,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0008056 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk=10,blk_iter=24,blk_iter_chan=8,blk_pel=80)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=28,x=28)),out_ref=(dims=(img=1,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=10,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 6.288e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk=10,blk_iter=24,blk_iter_chan=8,blk_pel=80)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=28,x=28)),out_ref=(dims=(img=1,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=10,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 9.2192e-05 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 4.7872e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=49,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=28,x=28)),out_ref=(dims=(img=1,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00019216 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=7,blk_bx=7,blk_in_chan=192,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=192,y=28,x=28)),out=(dims=(img=1,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=7,blk_bx=7,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 4.768e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=49,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=28,x=28)),out_ref=(dims=(img=1,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000157664 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000192416 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=1,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=10,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 6.6144e-05 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk=10,blk_iter=24,blk_iter_chan=8,blk_pel=80)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=28,x=28)),out_ref=(dims=(img=1,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=10,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),filts=(dims=(out_chan=96,in_chan=192,y=1,x=1)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=5,chan=96,y=28,x=28)),out_chans=(tn=uint32_t,v=96),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0007224 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=192,y=1,x=1,out_chan=96)),in=(dims=(chan=192,pel=4032)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=96,pel=4032)),out_ref=(dims=(img=5,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=24,out_chan_blk=1,pels_tile=21,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000149792 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=192,y=1,x=1,out_chan=96)),in=(dims=(chan=192,pel=4032)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=96,pel=4032)),out_ref=(dims=(img=5,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=24,out_chan_blk=1,pels_tile=21,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00070048 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=28,x=28)),out_ref=(dims=(img=5,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000157984 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=28,x=28)),out_ref=(dims=(img=5,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00106384 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=7,blk_in_chan=192,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=192,y=28,x=28)),out=(dims=(img=5,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=35,blk_bx=7,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00029504 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=7,blk_in_chan=192,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=192,y=28,x=28)),out=(dims=(img=5,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=35,blk_bx=7,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00069872 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=28,x=28)),out_ref=(dims=(img=5,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00015424 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=28,x=28)),out_ref=(dims=(img=5,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00090256 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000374528 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00124304 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000253984 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00081248 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk=49,blk_iter=24,blk_iter_chan=8,blk_pel=80)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=28,x=28)),out_ref=(dims=(img=5,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000125664 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk=49,blk_iter=24,blk_iter_chan=8,blk_pel=80)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=28,x=28)),out_ref=(dims=(img=5,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000113184 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000124896 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=28,x=28)),out_ref=(dims=(img=5,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000306592 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=35,blk_bx=7,blk_in_chan=192,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=192,y=28,x=28)),out=(dims=(img=5,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=35,blk_bx=7,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000124928 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=245,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=28,x=28)),out_ref=(dims=(img=5,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000381568 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=245,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000254272 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=5,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000124512 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk=49,blk_iter=24,blk_iter_chan=8,blk_pel=80)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=28,x=28)),out_ref=(dims=(img=5,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=49,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),filts=(dims=(out_chan=96,in_chan=192,y=1,x=1)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=1,x=1)),out=(dims=(img=20,chan=96,y=28,x=28)),out_chans=(tn=uint32_t,v=96),stride=(tn=none,dims=(y=1,x=1)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00082272 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=192,y=1,x=1,out_chan=96)),in=(dims=(chan=192,pel=15792)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=96,pel=15792)),out_ref=(dims=(img=20,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=94,out_chan_blk=1,pels_tile=21,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000395072 | |
| (str_vals=(func_name=k1conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=192,y=1,x=1,out_chan=96)),in=(dims=(chan=192,pel=15792)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(chan=96,pel=15792)),out_ref=(dims=(img=20,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=94,out_chan_blk=1,pels_tile=21,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00191024 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=980,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=28,x=28)),out_ref=(dims=(img=20,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00048992 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=980,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=28,x=28)),out_ref=(dims=(img=20,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00207168 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=140,blk_bx=7,blk_in_chan=192,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=192,y=28,x=28)),out=(dims=(img=20,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=140,blk_bx=7,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.0010568 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=140,blk_bx=7,blk_in_chan=192,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=192,y=28,x=28)),out=(dims=(img=20,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=140,blk_bx=7,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00118336 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=980,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=28,x=28)),out_ref=(dims=(img=20,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000489952 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=980,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=28,x=28)),out_ref=(dims=(img=20,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00183008 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00106765 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00203872 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=196,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00092576 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=196,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00135328 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk=196,blk_iter=24,blk_iter_chan=8,blk_pel=80)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=28,x=28)),out_ref=(dims=(img=20,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=196,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000322528 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk=196,blk_iter=24,blk_iter_chan=8,blk_pel=80)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=28,x=28)),out_ref=(dims=(img=20,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=196,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000340192 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=1,x=1)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000402528 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=980,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=28,x=28)),out_ref=(dims=(img=20,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00105379 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=140,blk_bx=7,blk_in_chan=192,blk_y=4,blk_x=4)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=192,y=28,x=28)),out=(dims=(img=20,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(blk_bline=140,blk_bx=7,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00040608 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk=980,blk_iter=24,blk_iter_chan=8,blk_pel=16)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=28,x=28)),out_ref=(dims=(img=20,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00104339 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=192,y=1,x=1,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=980,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00092464 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=20,chan=192,y=28,x=28)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=196,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000317184 | |
| (str_vals=(func_name=k1conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=192,y=1,x=1,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk=196,blk_iter=24,blk_iter_chan=8,blk_pel=80)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=28,x=28)),out_ref=(dims=(img=20,chan=96,y=28,x=28)),stride=(tn=none,dims=(y=1,x=1)),work=(tn=none,dims=(pels_blk=196,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),filts=(dims=(out_chan=96,in_chan=3,y=11,x=11)),in=(dims=(img=1,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=11,x=11)),out=(dims=(img=1,chan=96,y=54,x=54)),out_chans=(tn=uint32_t,v=96),stride=(tn=none,dims=(y=4,x=4)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00217008 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=3,y=11,x=11,out_chan=96)),in=(dims=(chan=3,pel=51752)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=1,y=224,x=224)),out=(dims=(chan=96,pel=3192)),out_pels=(tn=none,dims=(img=1,y=56,x=56)),out_ref=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=19,out_chan_blk=1,pels_tile=21,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000521664 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=3,y=11,x=11,out_chan=96)),in=(dims=(chan=3,pel=51752)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=1,y=224,x=224)),out=(dims=(chan=96,pel=3192)),out_pels=(tn=none,dims=(img=1,y=56,x=56)),out_ref=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=19,out_chan_blk=1,pels_tile=21,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00104064 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=14,blk_bx=14,blk_in_chan=3,blk_y=23,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=3,y=224,x=224)),out=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=14,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000152608 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=14,blk_bx=14,blk_in_chan=3,blk_y=23,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=3,y=224,x=224)),out=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=14,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00104224 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=14,blk_bx=14,blk_in_chan=3,blk_y=23,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=3,y=224,x=224)),out=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=14,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00015408 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=14,blk_bx=14,blk_in_chan=3,blk_y=23,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=3,y=224,x=224)),out=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=14,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.001952 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=183,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000426784 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=183,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00193808 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=183,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000425216 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=183,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0020632 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=1,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=37,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00043264 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=1,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=37,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00093696 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=6,blk_bx=7,blk_in_chan=3,blk_y=47,blk_x=39)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=3,y=224,x=224)),out=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=6,blk_bx=7,out_chan_blk=1,blk_y=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000150048 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=6,blk_bx=7,blk_in_chan=3,blk_y=47,blk_x=39)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=3,y=224,x=224)),out=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=6,blk_bx=7,out_chan_blk=1,blk_y=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000138304 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=4,x=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000154944 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=14,blk_bx=14,blk_in_chan=3,blk_y=23,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=3,y=224,x=224)),out=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=14,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000152384 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=14,blk_bx=14,blk_in_chan=3,blk_y=23,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=3,y=224,x=224)),out=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=14,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000410208 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=183,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000410112 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=183,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000419232 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=1,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=37,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000148864 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=6,blk_bx=7,blk_in_chan=3,blk_y=47,blk_x=39)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=3,y=224,x=224)),out=(dims=(img=1,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=6,blk_bx=7,out_chan_blk=1,blk_y=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),filts=(dims=(out_chan=96,in_chan=3,y=11,x=11)),in=(dims=(img=1,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=11,x=11)),out=(dims=(img=1,chan=96,y=55,x=55)),out_chans=(tn=uint32_t,v=96),stride=(tn=none,dims=(y=4,x=4)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00216352 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=3,y=11,x=11,out_chan=96)),in=(dims=(chan=3,pel=53592)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=1,y=228,x=228)),out=(dims=(chan=96,pel=3360)),out_pels=(tn=none,dims=(img=1,y=57,x=57)),out_ref=(dims=(img=1,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=20,out_chan_blk=1,pels_tile=21,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000545088 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=3,y=11,x=11,out_chan=96)),in=(dims=(chan=3,pel=53592)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=1,y=228,x=228)),out=(dims=(chan=96,pel=3360)),out_pels=(tn=none,dims=(img=1,y=57,x=57)),out_ref=(dims=(img=1,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=20,out_chan_blk=1,pels_tile=21,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00104048 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=14,blk_bx=14,blk_in_chan=3,blk_y=23,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=3,y=227,x=227)),out=(dims=(img=1,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=14,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000156064 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=14,blk_bx=14,blk_in_chan=3,blk_y=23,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=3,y=227,x=227)),out=(dims=(img=1,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=14,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00104016 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=14,blk_bx=14,blk_in_chan=3,blk_y=23,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=3,y=227,x=227)),out=(dims=(img=1,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=14,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000157536 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=14,blk_bx=14,blk_in_chan=3,blk_y=23,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=3,y=227,x=227)),out=(dims=(img=1,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=14,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00196208 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=190,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00042864 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=190,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0019584 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=190,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000429856 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=190,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00212944 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=1,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=38,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00042864 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=1,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=38,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00093968 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=6,blk_bx=7,blk_in_chan=3,blk_y=47,blk_x=39)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=3,y=227,x=227)),out=(dims=(img=1,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=6,blk_bx=7,out_chan_blk=1,blk_y=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000150304 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=6,blk_bx=7,blk_in_chan=3,blk_y=47,blk_x=39)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=3,y=227,x=227)),out=(dims=(img=1,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=6,blk_bx=7,out_chan_blk=1,blk_y=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000135264 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=4,x=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000156224 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=14,blk_bx=14,blk_in_chan=3,blk_y=23,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=3,y=227,x=227)),out=(dims=(img=1,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=14,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00015424 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=14,blk_bx=14,blk_in_chan=3,blk_y=23,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=3,y=227,x=227)),out=(dims=(img=1,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=14,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00041008 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=190,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00040752 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=1,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=190,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00040976 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=1,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=1,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=38,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000148832 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=6,blk_bx=7,blk_in_chan=3,blk_y=47,blk_x=39)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=1,chan=3,y=227,x=227)),out=(dims=(img=1,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=6,blk_bx=7,out_chan_blk=1,blk_y=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),filts=(dims=(out_chan=96,in_chan=3,y=11,x=11)),in=(dims=(img=5,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=11,x=11)),out=(dims=(img=5,chan=96,y=54,x=54)),out_chans=(tn=uint32_t,v=96),stride=(tn=none,dims=(y=4,x=4)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00309088 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=3,y=11,x=11,out_chan=96)),in=(dims=(chan=3,pel=252456)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=5,y=224,x=224)),out=(dims=(chan=96,pel=15792)),out_pels=(tn=none,dims=(img=5,y=56,x=56)),out_ref=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=94,out_chan_blk=1,pels_tile=21,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00108883 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=3,y=11,x=11,out_chan=96)),in=(dims=(chan=3,pel=252456)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=5,y=224,x=224)),out=(dims=(chan=96,pel=15792)),out_pels=(tn=none,dims=(img=5,y=56,x=56)),out_ref=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=94,out_chan_blk=1,pels_tile=21,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00389536 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=68,blk_bx=14,blk_in_chan=3,blk_y=30,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=3,y=224,x=224)),out=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=68,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000600224 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=68,blk_bx=14,blk_in_chan=3,blk_y=30,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=3,y=224,x=224)),out=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=68,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00369264 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=68,blk_bx=14,blk_in_chan=3,blk_y=30,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=3,y=224,x=224)),out=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=68,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000594848 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=68,blk_bx=14,blk_in_chan=3,blk_y=30,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=3,y=224,x=224)),out=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=68,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00518976 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=912,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00173382 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=912,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00518928 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=912,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00173798 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=912,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0026968 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=5,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=183,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00114266 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=5,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=183,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00205936 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=27,blk_bx=7,blk_in_chan=3,blk_y=54,blk_x=39)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=3,y=224,x=224)),out=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=27,blk_bx=7,out_chan_blk=1,blk_y=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000426016 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=27,blk_bx=7,blk_in_chan=3,blk_y=54,blk_x=39)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=3,y=224,x=224)),out=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=27,blk_bx=7,out_chan_blk=1,blk_y=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000448032 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=4,x=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000618976 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=68,blk_bx=14,blk_in_chan=3,blk_y=30,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=3,y=224,x=224)),out=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=68,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000615488 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=68,blk_bx=14,blk_in_chan=3,blk_y=30,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=3,y=224,x=224)),out=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=68,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00172608 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=912,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.0017152 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=912,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00098464 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=5,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=183,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000404288 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=27,blk_bx=7,blk_in_chan=3,blk_y=54,blk_x=39)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=3,y=224,x=224)),out=(dims=(img=5,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=27,blk_bx=7,out_chan_blk=1,blk_y=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),filts=(dims=(out_chan=96,in_chan=3,y=11,x=11)),in=(dims=(img=5,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=11,x=11)),out=(dims=(img=5,chan=96,y=55,x=55)),out_chans=(tn=uint32_t,v=96),stride=(tn=none,dims=(y=4,x=4)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0031016 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=3,y=11,x=11,out_chan=96)),in=(dims=(chan=3,pel=261528)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=5,y=228,x=228)),out=(dims=(chan=96,pel=16296)),out_pels=(tn=none,dims=(img=5,y=57,x=57)),out_ref=(dims=(img=5,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=97,out_chan_blk=1,pels_tile=21,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00138723 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=3,y=11,x=11,out_chan=96)),in=(dims=(chan=3,pel=261528)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=5,y=228,x=228)),out=(dims=(chan=96,pel=16296)),out_pels=(tn=none,dims=(img=5,y=57,x=57)),out_ref=(dims=(img=5,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=97,out_chan_blk=1,pels_tile=21,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00391616 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=69,blk_bx=14,blk_in_chan=3,blk_y=30,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=3,y=227,x=227)),out=(dims=(img=5,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=69,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000606048 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=69,blk_bx=14,blk_in_chan=3,blk_y=30,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=3,y=227,x=227)),out=(dims=(img=5,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=69,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00383792 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=69,blk_bx=14,blk_in_chan=3,blk_y=30,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=3,y=227,x=227)),out=(dims=(img=5,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=69,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.000601984 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=69,blk_bx=14,blk_in_chan=3,blk_y=30,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=3,y=227,x=227)),out=(dims=(img=5,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=69,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00522112 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=946,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00181402 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=946,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00522592 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=946,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00180112 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=946,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00260224 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=5,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=190,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00114813 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=5,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=190,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00210128 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=28,blk_bx=7,blk_in_chan=3,blk_y=54,blk_x=39)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=3,y=227,x=227)),out=(dims=(img=5,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=28,blk_bx=7,out_chan_blk=1,blk_y=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00045968 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=28,blk_bx=7,blk_in_chan=3,blk_y=54,blk_x=39)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=3,y=227,x=227)),out=(dims=(img=5,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=28,blk_bx=7,out_chan_blk=1,blk_y=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00046384 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=4,x=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00063104 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=69,blk_bx=14,blk_in_chan=3,blk_y=30,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=3,y=227,x=227)),out=(dims=(img=5,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=69,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000622592 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=69,blk_bx=14,blk_in_chan=3,blk_y=30,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=3,y=227,x=227)),out=(dims=(img=5,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=69,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00178147 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=946,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00179914 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=5,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=946,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00103597 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=5,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=5,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=190,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.000449408 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=28,blk_bx=7,blk_in_chan=3,blk_y=54,blk_x=39)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=5,chan=3,y=227,x=227)),out=(dims=(img=5,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=28,blk_bx=7,out_chan_blk=1,blk_y=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),filts=(dims=(out_chan=96,in_chan=3,y=11,x=11)),in=(dims=(img=20,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=11,x=11)),out=(dims=(img=20,chan=96,y=54,x=54)),out_chans=(tn=uint32_t,v=96),stride=(tn=none,dims=(y=4,x=4)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00873584 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=3,y=11,x=11,out_chan=96)),in=(dims=(chan=3,pel=1005096)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=20,y=224,x=224)),out=(dims=(chan=96,pel=62832)),out_pels=(tn=none,dims=(img=20,y=56,x=56)),out_ref=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=374,out_chan_blk=1,pels_tile=21,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00395987 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=3,y=11,x=11,out_chan=96)),in=(dims=(chan=3,pel=1005096)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=20,y=224,x=224)),out=(dims=(chan=96,pel=62832)),out_pels=(tn=none,dims=(img=20,y=56,x=56)),out_ref=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=374,out_chan_blk=1,pels_tile=21,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0115491 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=270,blk_bx=14,blk_in_chan=3,blk_y=30,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=3,y=224,x=224)),out=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=270,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00221453 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=270,blk_bx=14,blk_in_chan=3,blk_y=30,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=3,y=224,x=224)),out=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=270,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00715936 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=270,blk_bx=14,blk_in_chan=3,blk_y=30,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=3,y=224,x=224)),out=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=270,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00220803 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=270,blk_bx=14,blk_in_chan=3,blk_y=30,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=3,y=224,x=224)),out=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=270,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.014217 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=3645,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00663434 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=3645,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00992992 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=3645,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00664294 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=3645,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00799216 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=20,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=729,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00400829 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=20,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=729,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00568848 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=108,blk_bx=7,blk_in_chan=3,blk_y=54,blk_x=39)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=3,y=224,x=224)),out=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=108,blk_bx=7,out_chan_blk=1,blk_y=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00145904 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=108,blk_bx=7,blk_in_chan=3,blk_y=54,blk_x=39)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=3,y=224,x=224)),out=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=108,blk_bx=7,out_chan_blk=1,blk_y=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00152403 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=4,x=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00227334 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=270,blk_bx=14,blk_in_chan=3,blk_y=30,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=3,y=224,x=224)),out=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=270,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00226925 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=270,blk_bx=14,blk_in_chan=3,blk_y=30,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=3,y=224,x=224)),out=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=270,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00657238 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=3645,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00657059 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=3645,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00397744 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=20,chan=3,y=224,x=224)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=729,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00143402 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=108,blk_bx=7,blk_in_chan=3,blk_y=54,blk_x=39)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=3,y=224,x=224)),out=(dims=(img=20,chan=96,y=54,x=54)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=108,blk_bx=7,out_chan_blk=1,blk_y=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| /op_wisdom_t | |
| op_wisdom_t | |
| (str_vals=(type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),filts=(dims=(out_chan=96,in_chan=3,y=11,x=11)),in=(dims=(img=20,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),kern_sz=(tn=none,dims=(y=11,x=11)),out=(dims=(img=20,chan=96,y=55,x=55)),out_chans=(tn=uint32_t,v=96),stride=(tn=none,dims=(y=4,x=4)))) | |
| kg | |
| out | |
|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| |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=16 16,Kb=1,use_local_mem=2,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0101184 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=3,y=11,x=11,out_chan=96)),in=(dims=(chan=3,pel=1041288)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=20,y=228,x=228)),out=(dims=(chan=96,pel=65016)),out_pels=(tn=none,dims=(img=20,y=57,x=57)),out_ref=(dims=(img=20,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=387,out_chan_blk=1,pels_tile=21,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.0043041 | |
| (str_vals=(func_name=conv_simd,type=Convolution),nda_vals=(Kb=(tn=uint32_t,v=1),biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(in_chan=3,y=11,x=11,out_chan=96)),in=(dims=(chan=3,pel=1041288)),in_pad=(tn=none,dims=(y=0,x=0)),in_pels=(tn=none,dims=(img=20,y=228,x=228)),out=(dims=(chan=96,pel=65016)),out_pels=(tn=none,dims=(img=20,y=57,x=57)),out_ref=(dims=(img=20,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),vw=(tn=uint32_t,v=8),work=(tn=none,dims=(pels_blk=387,out_chan_blk=1,pels_tile=21,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0119286 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=275,blk_bx=14,blk_in_chan=3,blk_y=30,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=3,y=227,x=227)),out=(dims=(img=20,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=275,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00225645 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=275,blk_bx=14,blk_in_chan=3,blk_y=30,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=3,y=227,x=227)),out=(dims=(img=20,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=275,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00723392 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=275,blk_bx=14,blk_in_chan=3,blk_y=30,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=3,y=227,x=227)),out=(dims=(img=20,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=275,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00225731 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=275,blk_bx=14,blk_in_chan=3,blk_y=30,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=3,y=227,x=227)),out=(dims=(img=20,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=275,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0148922 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=3782,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.0068601 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=3782,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0101738 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=3782,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00686592 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=3782,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.0081416 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=20,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=757,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00407626 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=20,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=757,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=ocl,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| ocl:Fiji | |
| 0.00684688 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=110,blk_bx=7,blk_in_chan=3,blk_y=54,blk_x=39)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=3,y=227,x=227)),out=(dims=(img=20,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=110,blk_bx=7,out_chan_blk=1,blk_y=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| op_run_t | |
| ocl:GeForce GTX TITAN X | |
| 0.00151293 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),flags=(tn=uint32_t),in=(dims=(blk_bline=110,blk_bx=7,blk_in_chan=3,blk_y=54,blk_x=39)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=3,y=227,x=227)),out=(dims=(img=20,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=110,blk_bx=7,out_chan_blk=1,blk_y=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,use_culibs=1,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00158128 | |
| (str_vals=(func_name=cudnn_conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=0)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan=0,in_chan=0,y=0,x=0)),in=(dims=(img=0,chan=0,y=0,x=0)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=0,chan=0,y=0,x=0)),stride=(tn=none,dims=(y=4,x=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00231021 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=275,blk_bx=14,blk_in_chan=3,blk_y=30,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=3,y=227,x=227)),out=(dims=(img=20,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=275,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00231283 | |
| (str_vals=(func_name=tconv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),flags=(tn=uint32_t),in=(dims=(blk_bline=275,blk_bx=14,blk_in_chan=3,blk_y=30,blk_x=23)),in_pad=(tn=none,dims=(y=0,x=0)),in_ref=(dims=(img=20,chan=3,y=227,x=227)),out=(dims=(img=20,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(blk_bline=275,blk_bx=14,out_chan_blk=2,blk_y=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,k1conv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00681891 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=3782,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=4 4,MNb=8 8,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.0068105 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=2,in_chan=3,y=11,x=11,out_chan_reg=4,out_chan_tile=16)),in=(dims=(img=20,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=3782,out_chan_blk=2,pels_tile=4,out_chan_tile=16,pels=4,out_chan=4)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeForce GTX TITAN X | |
| 0.00401584 | |
| (str_vals=(func_name=conv,type=Convolution),nda_vals=(biases=(dims=(out_chan=96)),conv_has_relu=(tn=uint32_t,v=1),filts=(dims=(out_chan_blk=1,in_chan=3,y=11,x=11,out_chan_reg=8,out_chan_tile=12)),in=(dims=(img=20,chan=3,y=227,x=227)),in_pad=(tn=none,dims=(y=0,x=0)),out=(dims=(img=20,chan=96,y=55,x=55)),stride=(tn=none,dims=(y=4,x=4)),work=(tn=none,dims=(pels_blk=757,out_chan_blk=1,pels_tile=10,out_chan_tile=12,pels=8,out_chan=8)))) | |
| /op_tune_wisdom_t | |
| op_tune_wisdom_t | |
| (use_be=nvrtc,MNt=8 8,MNb=8 16,k1conv=1,tconv=1,tconv_max_ksz=11 11) | |
| op_run_t | |
| nvrtc:GeFor |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment