Skip to content

Instantly share code, notes, and snippets.

@inferrna

inferrna/bitonic_b.py

Last active August 29, 2015 14:24
Show Gist options
  • Save inferrna/2202fb8938fbc63e54dc to your computer and use it in GitHub Desktop.
Save inferrna/2202fb8938fbc63e54dc to your computer and use it in GitHub Desktop.
Download ZIP
opencl (pyopencl) bitonic sorter based on http://www.bealto.com/gpu-sorting_intro.html by Eric Bainville
Raw
bitonic_b.py
__copyright__ = """
Copyright (c) 2011, Eric Bainville
Copyright (c) 2015, Ilya Efimoff
All rights reserved.
"""
__license__ = """
Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
1. Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.
3. Neither the name of the copyright holder nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
"""
import pyopencl as cl
from pyopencl.tools import dtype_to_ctype
import bitonic_templates
from mako.template import Template
from operator import mul
from functools import reduce
class BitonicSort(object):
def __init__(self, context, shape, key_dtype, idx_dtype=None, axis=0):
x = shape[axis]
assert ((not x == 0) and not (x & (x - 1))), "Sorting dimension ({0}) size ({1}) is not power of 2.".format(axis, x)
self.cached_defs = {}
self.cached_progs = {'B2':{}, 'B4':{}, 'B8':{}, 'B16':{}, 'C4':{}, 'BLO':{}, 'BL':{}, 'PML':{}}
self.kernels_srcs = {'B2': bitonic_templates.ParallelBitonic_B2,
'B4': bitonic_templates.ParallelBitonic_B4,
'B8': bitonic_templates.ParallelBitonic_B8,
'B16':bitonic_templates.ParallelBitonic_B16,
'C4': bitonic_templates.ParallelBitonic_C4,
'BL': bitonic_templates.ParallelBitonic_Local,
'BLO':bitonic_templates.ParallelBitonic_Local_Optim,
'PML':bitonic_templates.ParallelMerge_Local}
self.dtype = dtype_to_ctype(key_dtype)
self.context = context
self.axis = axis
if idx_dtype==None:
self.argsort = 0
self.idx_t = 'uint' #Dummy
else:
self.argsort = 1
self.idx_t = dtype_to_ctype(idx_dtype)
self.defstpl = Template(bitonic_templates.defines)
self.rq = self.sort_b_prepare_wl(shape, self.axis)
def __call__(self, _arr, idx=None, mkcpy=True):
arr = _arr.copy() if mkcpy else _arr
rq = self.rq
p, nt, wg, aux = rq[0]
if self.argsort and not type(idx)==type(None):
if aux:
p.run(arr.queue, (nt,), wg, arr.data, idx.data, cl.LocalMemory(wg[0]*arr.dtype.itemsize),\
cl.LocalMemory(wg[0]*idx.dtype.itemsize))
for p, nt, wg,_ in rq[1:]:
p.run(arr.queue, (nt,), wg, arr.data, idx.data)
elif self.argsort==0:
if aux:
p.run(arr.queue, (nt,), wg, arr.data, cl.LocalMemory(wg[0]*arr.dtype.itemsize))
for p, nt, wg,_ in rq[1:]:
p.run(arr.queue, (nt,), wg, arr.data)
else:
raise ValueError("Array of indexes required for this sorter. If argsort is not needed,\
recreate sorter witout index datatype provided.")
return arr
def get_program(self, letter, params):
if params in self.cached_progs[letter].keys():
return self.cached_progs[letter][params]
else:
if params in self.cached_defs.keys():
defs = self.cached_defs[params]
else:
defs = self.defstpl.render(NS="\\", argsort=self.argsort, inc=params[0], dir=params[1],\
dtype=params[2], idxtype=params[3],\
dsize=params[4], nsize=params[5])
self.cached_defs[params] = defs
kid = Template(self.kernels_srcs[letter]).render(argsort=self.argsort)
prg = cl.Program(self.context, defs + kid).build()
self.cached_progs[letter][params] = prg
return prg
def sort_b_prepare_wl(self, shape, axis):
run_queue = []
ds = int(shape[axis])
size = reduce(mul, shape)
ndim = len(shape)
ns = reduce(mul, shape[(axis+1):]) if axis<ndim-1 else 1
ds = int(shape[axis])
allowb4 = True
allowb8 = True
allowb16 = True
wg = min(ds, self.context.devices[0].max_work_group_size)
length = wg>>1
prg = self.get_program('BLO', (1, 1, self.dtype, self.idx_t, ds, ns))
run_queue.append((prg, size, (wg,), True))
while length<ds:
inc = length;
while inc > 0:
ninc = 0;
direction = length<<1
if allowb16 and inc >= 8 and ninc == 0:
letter = 'B16'
ninc = 4;
elif allowb8 and inc >= 4 and ninc == 0:
letter = 'B8'
ninc = 3;
elif allowb4 and inc >= 2 and ninc == 0:
letter = 'B4'
ninc = 2;
elif inc >= 0:
letter = 'B2'
ninc = 1;
nThreads = size >> ninc;
prg = self.get_program(letter, (inc, direction, self.dtype, self.idx_t, ds, ns))
run_queue.append((prg, nThreads, None, False,))
inc >>= ninc;
length<<=1
return run_queue
Raw
bitonic_templates.py
__copyright__ = """
Copyright (c) 2011, Eric Bainville
Copyright (c) 2015, Ilya Efimoff
All rights reserved.
"""
__license__ = """
Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
1. Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.
3. Neither the name of the copyright holder nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
"""
defines = """
% if dtype=='double':
#pragma OPENCL EXTENSION cl_khr_fp64 : enable
% else:
#pragma OPENCL EXTENSION cl_khr_fp64 : disable
% endif
typedef ${dtype} data_t;
typedef ${idxtype} idx_t;
typedef ${idxtype}2 idx_t2;
#if CONFIG_USE_VALUE
#define getKey(a) ((a).x)
#define getValue(a) ((a).y)
#define makeData(k,v) ((${dtype}2)((k),(v)))
#else
#define getKey(a) (a)
#define getValue(a) (0)
#define makeData(k,v) (k)
#endif
#ifndef BLOCK_FACTOR
#define BLOCK_FACTOR 1
#endif
#define inc ${inc}
#define hinc ${inc>>1} //Half inc
#define qinc ${inc>>2} //Quarter inc
#define einc ${inc>>3} //Eighth of inc
#define dir ${dir}
% if argsort:
#define ORDER(a,b,ay,by) { bool swap = reverse ^ (getKey(a)<getKey(b));${NS}
data_t auxa = a; data_t auxb = b;${NS}
idx_t auya = ay; idx_t auyb = by;${NS}
a = (swap)?auxb:auxa; b = (swap)?auxa:auxb;${NS}
ay = (swap)?auyb:auya; by = (swap)?auya:auyb;}
#define ORDERV(x,y,a,b) { bool swap = reverse ^ (getKey(x[a])<getKey(x[b]));${NS}
data_t auxa = x[a]; data_t auxb = x[b];${NS}
idx_t auya = y[a]; idx_t auyb = y[b];${NS}
x[a] = (swap)?auxb:auxa; x[b] = (swap)?auxa:auxb;${NS}
y[a] = (swap)?auyb:auya; y[b] = (swap)?auya:auyb;}
#define B2V(x,y,a) { ORDERV(x,y,a,a+1) }
#define B4V(x,y,a) { for (int i4=0;i4<2;i4++) { ORDERV(x,y,a+i4,a+i4+2) } B2V(x,y,a) B2V(x,y,a+2) }
#define B8V(x,y,a) { for (int i8=0;i8<4;i8++) { ORDERV(x,y,a+i8,a+i8+4) } B4V(x,y,a) B4V(x,y,a+4) }
#define B16V(x,y,a) { for (int i16=0;i16<8;i16++) { ORDERV(x,y,a+i16,a+i16+8) } B8V(x,y,a) B8V(x,y,a+8) }
% else:
#define ORDER(a,b) { bool swap = reverse ^ (getKey(a)<getKey(b)); data_t auxa = a; data_t auxb = b; a = (swap)?auxb:auxa; b = (swap)?auxa:auxb; }
#define ORDERV(x,a,b) { bool swap = reverse ^ (getKey(x[a])<getKey(x[b]));${NS}
data_t auxa = x[a]; data_t auxb = x[b];${NS}
x[a] = (swap)?auxb:auxa; x[b] = (swap)?auxa:auxb; }
#define B2V(x,a) { ORDERV(x,a,a+1) }
#define B4V(x,a) { for (int i4=0;i4<2;i4++) { ORDERV(x,a+i4,a+i4+2) } B2V(x,a) B2V(x,a+2) }
#define B8V(x,a) { for (int i8=0;i8<4;i8++) { ORDERV(x,a+i8,a+i8+4) } B4V(x,a) B4V(x,a+4) }
#define B16V(x,a) { for (int i16=0;i16<8;i16++) { ORDERV(x,a+i16,a+i16+8) } B8V(x,a) B8V(x,a+8) }
% endif
#define nsize ${nsize} //Total next dimensions sizes sum. (Block size)
#define dsize ${dsize} //Dimension size
"""
ParallelBitonic_B2 = """
// N/2 threads
//ParallelBitonic_B2
__kernel void run(__global data_t * data\\
% if argsort:
, __global idx_t * index)
% else:
)
% endif
{
int t = get_global_id(0) % (dsize>>1); // thread index
int gt = get_global_id(0) / (dsize>>1);
int low = t & (inc - 1); // low order bits (below INC)
int i = (t<<1) - low; // insert 0 at position INC
int gi = i/dsize; // block index
bool reverse = ((dir & i) == 0);// ^ (gi%2); // asc/desc order
int offset = (gt/nsize)*nsize*dsize+(gt%nsize);
data += i*nsize + offset; // translate to first value
% if argsort:
index += i*nsize + offset; // translate to first value
% endif
// Load data
data_t x0 = data[ 0];
data_t x1 = data[inc*nsize];
% if argsort:
// Load index
idx_t i0 = index[ 0];
idx_t i1 = index[inc*nsize];
% endif
// Sort
% if argsort:
ORDER(x0,x1,i0,i1)
% else:
ORDER(x0,x1)
% endif
// Store data
data[0 ] = x0;
data[inc*nsize] = x1;
% if argsort:
// Store index
index[ 0] = i0;
index[inc*nsize] = i1;
% endif
}
"""
ParallelBitonic_B4 = """
// N/4 threads
//ParallelBitonic_B4
__kernel void run(__global data_t * data\\
% if argsort:
, __global idx_t * index)
% else:
)
% endif
{
int t = get_global_id(0) % (dsize>>2); // thread index
int gt = get_global_id(0) / (dsize>>2);
int low = t & (hinc - 1); // low order bits (below INC)
int i = ((t - low) << 2) + low; // insert 00 at position INC
bool reverse = ((dir & i) == 0); // asc/desc order
int offset = (gt/nsize)*nsize*dsize+(gt%nsize);
data += i*nsize + offset; // translate to first value
% if argsort:
index += i*nsize + offset; // translate to first value
% endif
// Load data
data_t x0 = data[ 0];
data_t x1 = data[ hinc*nsize];
data_t x2 = data[2*hinc*nsize];
data_t x3 = data[3*hinc*nsize];
% if argsort:
// Load index
idx_t i0 = index[ 0];
idx_t i1 = index[ hinc*nsize];
idx_t i2 = index[2*hinc*nsize];
idx_t i3 = index[3*hinc*nsize];
% endif
// Sort
% if argsort:
ORDER(x0,x2,i0,i2)
ORDER(x1,x3,i1,i3)
ORDER(x0,x1,i0,i1)
ORDER(x2,x3,i2,i3)
% else:
ORDER(x0,x2)
ORDER(x1,x3)
ORDER(x0,x1)
ORDER(x2,x3)
% endif
// Store data
data[ 0] = x0;
data[ hinc*nsize] = x1;
data[2*hinc*nsize] = x2;
data[3*hinc*nsize] = x3;
% if argsort:
// Store index
index[ 0] = i0;
index[ hinc*nsize] = i1;
index[2*hinc*nsize] = i2;
index[3*hinc*nsize] = i3;
% endif
}
"""
ParallelBitonic_B8 = """
// N/8 threads
//ParallelBitonic_B8
__kernel void run(__global data_t * data\\
% if argsort:
, __global idx_t * index)
% else:
)
% endif
{
int t = get_global_id(0) % (dsize>>3); // thread index
int gt = get_global_id(0) / (dsize>>3);
int low = t & (qinc - 1); // low order bits (below INC)
int i = ((t - low) << 3) + low; // insert 000 at position INC
bool reverse = ((dir & i) == 0); // asc/desc order
int offset = (gt/nsize)*nsize*dsize+(gt%nsize);
data += i*nsize + offset; // translate to first value
% if argsort:
index += i*nsize + offset; // translate to first value
% endif
// Load
data_t x[8];
% if argsort:
idx_t y[8];
% endif
for (int k=0;k<8;k++) x[k] = data[k*qinc*nsize];
% if argsort:
for (int k=0;k<8;k++) y[k] = index[k*qinc*nsize];
% endif
// Sort
% if argsort:
B8V(x,y,0)
% else:
B8V(x,0)
% endif
// Store
for (int k=0;k<8;k++) data[k*qinc*nsize] = x[k];
% if argsort:
for (int k=0;k<8;k++) index[k*qinc*nsize] = y[k];
% endif
}
"""
ParallelBitonic_B16 = """
// N/16 threads
//ParallelBitonic_B16
__kernel void run(__global data_t * data\\
% if argsort:
, __global idx_t * index)
% else:
)
% endif
{
int t = get_global_id(0) % (dsize>>4); // thread index
int gt = get_global_id(0) / (dsize>>4);
int low = t & (einc - 1); // low order bits (below INC)
int i = ((t - low) << 4) + low; // insert 0000 at position INC
bool reverse = ((dir & i) == 0); // asc/desc order
int offset = (gt/nsize)*nsize*dsize+(gt%nsize);
data += i*nsize + offset; // translate to first value
% if argsort:
index += i*nsize + offset; // translate to first value
% endif
// Load
data_t x[16];
% if argsort:
idx_t y[16];
% endif
for (int k=0;k<16;k++) x[k] = data[k*einc*nsize];
% if argsort:
for (int k=0;k<16;k++) y[k] = index[k*einc*nsize];
% endif
// Sort
% if argsort:
B16V(x,y,0)
% else:
B16V(x,0)
% endif
// Store
for (int k=0;k<16;k++) data[k*einc*nsize] = x[k];
% if argsort:
for (int k=0;k<16;k++) index[k*einc*nsize] = y[k];
% endif
}
"""
ParallelBitonic_C4 = """
//ParallelBitonic_C4
__kernel void run\\
% if argsort:
(__global data_t * data, __global idx_t * index, __local data_t * aux, __local idx_t * auy)
% else:
(__global data_t * data, __local data_t * aux)
% endif
{
int t = get_global_id(0); // thread index
int wgBits = 4*get_local_size(0) - 1; // bit mask to get index in local memory AUX (size is 4*WG)
int linc,low,i;
bool reverse;
data_t x[4];
% if argsort:
idx_t y[4];
% endif
// First iteration, global input, local output
linc = hinc;
low = t & (linc - 1); // low order bits (below INC)
i = ((t - low) << 2) + low; // insert 00 at position INC
reverse = ((dir & i) == 0); // asc/desc order
for (int k=0;k<4;k++) x[k] = data[i+k*linc];
% if argsort:
for (int k=0;k<4;k++) y[k] = index[i+k*linc];
B4V(x,y,0);
for (int k=0;k<4;k++) auy[(i+k*linc) & wgBits] = y[k];
% else:
B4V(x,0);
% endif
for (int k=0;k<4;k++) aux[(i+k*linc) & wgBits] = x[k];
barrier(CLK_LOCAL_MEM_FENCE);
// Internal iterations, local input and output
for ( ;linc>1;linc>>=2)
{
low = t & (linc - 1); // low order bits (below INC)
i = ((t - low) << 2) + low; // insert 00 at position INC
reverse = ((dir & i) == 0); // asc/desc order
for (int k=0;k<4;k++) x[k] = aux[(i+k*linc) & wgBits];
% if argsort:
for (int k=0;k<4;k++) y[k] = auy[(i+k*linc) & wgBits];
B4V(x,y,0);
barrier(CLK_LOCAL_MEM_FENCE);
for (int k=0;k<4;k++) auy[(i+k*linc) & wgBits] = y[k];
% else:
B4V(x,0);
barrier(CLK_LOCAL_MEM_FENCE);
% endif
for (int k=0;k<4;k++) aux[(i+k*linc) & wgBits] = x[k];
barrier(CLK_LOCAL_MEM_FENCE);
}
// Final iteration, local input, global output, INC=1
i = t << 2;
reverse = ((dir & i) == 0); // asc/desc order
for (int k=0;k<4;k++) x[k] = aux[(i+k) & wgBits];
% if argsort:
for (int k=0;k<4;k++) y[k] = auy[(i+k) & wgBits];
B4V(x,y,0);
for (int k=0;k<4;k++) index[i+k] = y[k];
% else:
B4V(x,0);
% endif
for (int k=0;k<4;k++) data[i+k] = x[k];
}
"""
ParallelMerge_Local = """
// N threads, WG is workgroup size. Sort WG input blocks in each workgroup.
__kernel void run(__global const data_t * in,__global data_t * out,__local data_t * aux)
{
int i = get_local_id(0); // index in workgroup
int wg = get_local_size(0); // workgroup size = block size, power of 2
// Move IN, OUT to block start
int offset = get_group_id(0) * wg;
in += offset; out += offset;
// Load block in AUX[WG]
aux[i] = in[i];
barrier(CLK_LOCAL_MEM_FENCE); // make sure AUX is entirely up to date
// Now we will merge sub-sequences of length 1,2,...,WG/2
for (int length=1;length<wg;length<<=1)
{
data_t iData = aux[i];
data_t iKey = getKey(iData);
int ii = i & (length-1); // index in our sequence in 0..length-1
int sibling = (i - ii) ^ length; // beginning of the sibling sequence
int pos = 0;
for (int pinc=length;pinc>0;pinc>>=1) // increment for dichotomic search
{
int j = sibling+pos+pinc-1;
data_t jKey = getKey(aux[j]);
bool smaller = (jKey < iKey) || ( jKey == iKey && j < i );
pos += (smaller)?pinc:0;
pos = min(pos,length);
}
int bits = 2*length-1; // mask for destination
int dest = ((ii + pos) & bits) | (i & ~bits); // destination index in merged sequence
barrier(CLK_LOCAL_MEM_FENCE);
aux[dest] = iData;
barrier(CLK_LOCAL_MEM_FENCE);
}
// Write output
out[i] = aux[i];
}
"""
ParallelBitonic_Local = """
// N threads, WG is workgroup size. Sort WG input blocks in each workgroup.
__kernel void run(__global const data_t * in,__global data_t * out,__local data_t * aux)
{
int i = get_local_id(0); // index in workgroup
int wg = get_local_size(0); // workgroup size = block size, power of 2
// Move IN, OUT to block start
int offset = get_group_id(0) * wg;
in += offset; out += offset;
// Load block in AUX[WG]
aux[i] = in[i];
barrier(CLK_LOCAL_MEM_FENCE); // make sure AUX is entirely up to date
// Loop on sorted sequence length
for (int length=1;length<wg;length<<=1)
{
bool direction = ((i & (length<<1)) != 0); // direction of sort: 0=asc, 1=desc
// Loop on comparison distance (between keys)
for (int pinc=length;pinc>0;pinc>>=1)
{
int j = i + pinc; // sibling to compare
data_t iData = aux[i];
uint iKey = getKey(iData);
data_t jData = aux[j];
uint jKey = getKey(jData);
bool smaller = (jKey < iKey) || ( jKey == iKey && j < i );
bool swap = smaller ^ (j < i) ^ direction;
barrier(CLK_LOCAL_MEM_FENCE);
aux[i] = (swap)?jData:iData;
barrier(CLK_LOCAL_MEM_FENCE);
}
}
// Write output
out[i] = aux[i];
}
"""
ParallelBitonic_A = """
__kernel void ParallelBitonic_A(__global const data_t * in)
{
int i = get_global_id(0); // thread index
int j = i ^ inc; // sibling to compare
// Load values at I and J
data_t iData = in[i];
uint iKey = getKey(iData);
data_t jData = in[j];
uint jKey = getKey(jData);
// Compare
bool smaller = (jKey < iKey) || ( jKey == iKey && j < i );
bool swap = smaller ^ (j < i) ^ ((dir & i) != 0);
// Store
in[i] = (swap)?jData:iData;
}
"""
ParallelBitonic_Local_Optim = """
__kernel void run\\
% if argsort:
(__global data_t * data, __global idx_t * index, __local data_t * aux, __local idx_t * auy)
% else:
(__global data_t * data, __local data_t * aux)
% endif
{
int t = get_global_id(0) % dsize; // thread index
int gt = get_global_id(0) / dsize;
int offset = (gt/nsize)*nsize*dsize+(gt%nsize);
int i = get_local_id(0); // index in workgroup
int wg = get_local_size(0); // workgroup size = block size, power of 2
// Move IN, OUT to block start
//int offset = get_group_id(0) * wg;
data += offset;
// Load block in AUX[WG]
data_t iData = data[t*nsize];
aux[i] = iData;
% if argsort:
index += offset;
// Load block in AUY[WG]
idx_t iidx = index[t*nsize];
auy[i] = iidx;
% endif
barrier(CLK_LOCAL_MEM_FENCE); // make sure AUX is entirely up to date
// Loop on sorted sequence length
for (int pwg=1;pwg<=wg;pwg<<=1){
int loffset = pwg*(i/pwg);
int ii = i%pwg;
for (int length=1;length<pwg;length<<=1){
bool direction = ii & (length<<1); // direction of sort: 0=asc, 1=desc
// Loop on comparison distance (between keys)
for (int pinc=length;pinc>0;pinc>>=1){
int j = ii ^ pinc; // sibling to compare
data_t jData = aux[loffset+j];
% if argsort:
idx_t jidx = auy[loffset+j];
% endif
data_t iKey = getKey(iData);
data_t jKey = getKey(jData);
bool smaller = (jKey < iKey) || ( jKey == iKey && j < ii );
bool swap = smaller ^ (ii>j) ^ direction;
iData = (swap)?jData:iData; // update iData
% if argsort:
iidx = (swap)?jidx:iidx; // update iidx
% endif
barrier(CLK_LOCAL_MEM_FENCE);
aux[loffset+ii] = iData;
% if argsort:
auy[loffset+ii] = iidx;
% endif
barrier(CLK_LOCAL_MEM_FENCE);
}
}
}
// Write output
data[t*nsize] = iData;
% if argsort:
index[t*nsize] = iidx;
% endif
}
"""
Raw
c4.py
def sort_c4_prepare(shape, axis):
run_queue = []
size = shape[0]
n = size
ds = n
ns = 1
allowb4 = True
allowb8 = True
allowb16 = True
length = 1
while length<size:
inc = length
strategy = []
ii = inc
while ii > 0:
if ii in [256, 128, 32, 8]:
strategy.append(-1)
break
d = 1
if ii==256: d = 1;
elif ii==512 and allowb4: d = 2;
elif ii==1024 and allowb8: d = 3;
elif ii==2048 and allowb16: d = 4;
elif ii>=8 and allowb16: d = 4;
elif ii>=4 and allowb8: d = 3;
elif ii>=2 and allowb4: d = 2;
else: d = 1;
strategy.append(d);
ii >>= d
while inc > 0:
ninc = 0;
kid = -1;
doLocal = 0;
nThreads = 0;
d = strategy.pop(0)
direction = length<<1
#defs = defstpl.render(inc=inc, dir=direction) #Define custom parameters
if d == -1:
letter = 'C4' #PARALLEL_BITONIC_C4_KERNEL;
ninc = -1; # reduce all bits
doLocal = 4;
nThreads = n >> 2;
elif d == 4:
letter = 'B16' #PARALLEL_BITONIC_B16_KERNEL;
ninc = 4;
nThreads = n >> ninc;
elif d == 3:
letter = 'B8' #PARALLEL_BITONIC_B8_KERNEL;
ninc = 3;
nThreads = n >> ninc;
elif d == 2:
letter = 'B4' #PARALLEL_BITONIC_B4_KERNEL;
ninc = 2;
nThreads = n >> ninc;
elif d == 1:
letter = 'B2' #PARALLEL_BITONIC_B2_KERNEL;
ninc = 1;
nThreads = n >> ninc;
else:
print("Strategy error!");
prg = get_program(letter, (inc, direction, 'float', 'uint', ds, ns))
run_queue.append((prg, nThreads, doLocal>0))
if ninc < 0: break
inc >>= ninc
length<<=1
return {'q': run_queue, 'mems': (np.cl.LocalMemory(mwg*4*4), np.cl.LocalMemory(mwg*4*4),)}
#print("length =", length)
def sort_c4_run(arr, rqm, idx=None):
rq = rqm['q']
lx, ly = rqm['mems']
if argsort:
for p, nt, dl in rq:
if dl:
p.run(np.queue, (nt,), (512,), arr.data, idx.data, lx, ly)
else:
p.run(np.queue, (nt,), (512,), arr.data, idx.data)
else:
for p, nt, dl in rq:
if dl:
p.run(np.queue, (nt,), (512,), arr.data, lx)
else:
p.run(np.queue, (nt,), (512,), arr.data)
Raw
test.py
import bitonic_b
import pyopencl as cl
from pyopencl import array
from pyopencl import clrandom
import numpy as np
ctx = cl.create_some_context()
queue = cl.CommandQueue(ctx)
s = clrandom.rand(queue, (4, 512, 5,), np.float32, luxury=None, a=0, b=1.0)
sorter = bitonic_b.BitonicSort(ctx, s.shape, s.dtype, axis=1)
sgs = sorter(s)
r1 = (np.sort(s.get(), axis=1) == sgs.get()).all()
size = pow(2, 20)
index = array.arange(queue, 0, size, 1, dtype=np.int32)
m = clrandom.rand(queue, (size,), np.float32, luxury=None, a=0, b=1.0)
sorterm = bitonic_b.BitonicSort(ctx, m.shape, m.dtype, idx_dtype=index.dtype, axis=0)
ms = sorterm(m, idx=index)
r2 = (np.sort(m.get()) == ms.get()).all()
r3 = (np.argsort(m.get()) == index.get()).all() #May be False just because of identical values in array.
r4 = (m.get()[np.argsort(m.get())] == m.get()[index.get()]).all() #Check values by indices. Must always be True even if r3 is False
print("Sorting over axis finished with {0}, sorting with argsort finished with {1}, {2} and {3}.".format(r1, r2, r3, r4))
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment