opencl-header.tar.gz

RK3399.md

RK3399

Setup OS

I bought the basic Firefly RK3399 from Amazon for $130. It has 2GB RAM and 16 GB eMMC. There is an advanced version with 4GB RAM, but cost more $199.

The default OS doesn't have OpenCL support. The easy way is downloading an image with OpenCL, namely libOpenCL.so pre-installed and then flash into the board.

1. Download a Ubuntu 16.04 image from Google Drive.
2. Extract the download 7z file to get a 2.2 GB image Firefly-RK3399_xubuntu1604_201711301130.img.
3. Flash the image by following this document. (Username firefly, password: firefly)

In particular, if you have a Linux desktop, then we can use the upgrade_tool_v1_26.tar provided in this repo. (Note: : the other two linux tools rkflashkit and rkflashtool do not support flash the whole image. )

1. Cut off the power of the board

2. Connect it to the linux host machine with provided usb cable, and use the provided type-c converter to connect to the board

3. Press the RECOVERY key, next power on, and then release the key after around two second.

4. Install upgrade_tool by:

   tar -zxvf upgrade_tool_v1_26.tar
   sudo chown root:root upgrade_tool
   sudo chmod +x upgrade_tool
   sudo cp upgrade_tool /usr/local/bin

   Then run upgrade_tool you will see the board is connected

5. Flash the image by

   sudo upgrade_tool uf Firefly-RK3399_xubuntu1604_201711301130.img

Reboot the board then you will login into the Ubuntu automatically.

But note that this image only ships a libOpenCL.so, if you want to compile OpenCL program on the board, you need to install the header files, which is provided in opencl-header.tar.gz, just extract it into a proper directory such as

sudo tar -zxvf opencl-header.tar.gz -C /usr/include/

Sanity Check with clpeak

We can check if the opencl driver works properly by using the benchmark tool clpeak

sudo apt-get update && sudo apt-get install cmake git
git clone https://github.com/krrishnarraj/clpeak && cd clpeak
cmake . -DCMAKE_CXX_COMPILER=g++ && make
./clpeak

If everything works well, then you probably will see the following outputs:

Platform: ARM Platform
  Device: Mali-T860
    Driver version  : 1.2 (Linux ARM)
    Compute units   : 4
    Clock frequency : 200 MHz

    Global memory bandwidth (GBPS)
      float   : 3.17
      float2  : 6.07
      float4  : 7.88
      float8  : 6.55
      float16 : 6.26

    Single-precision compute (GFLOPS)
      float   : 25.09
      float2  : 45.51
      float4  : 46.22
      float8  : 41.67
      float16 : 46.40

    half-precision compute (GFLOPS)
      half   : 23.11
      half2  : 50.19
      half4  : 98.30
      half8  : 93.48
      half16 : 93.94

    Double-precision compute (GFLOPS)
      double   : 3.59
      double2  : 3.30
      double4  : 20.97
      double8  : 20.65
      double16 : 20.39

    Integer compute (GIOPS)
      int   : 20.15
      int2  : 49.64
      int4  : 47.12
      int8  : 49.17
      int16 : 41.47

    Transfer bandwidth (GBPS)
      enqueueWriteBuffer         : 4.61
      enqueueReadBuffer          : 2.60
      enqueueMapBuffer(for read) : 475.11
        memcpy from mapped ptr   : 2.50
      enqueueUnmap(after write)  : 2790.39
        memcpy to mapped ptr     : 1.92

    Kernel launch latency : 190.64 us

Run NNVM Compiler

Follow the cross complication tutorial to compile the runtime on RK3399, and local TVM.

Follow deploy pretrained model on Mali to run Resnet-18 on RK3399. To benchmark the performance, we can append the following code block:

import time
ntimes = 10
tic = time.time()
for _ in range(ntimes):
    module.run()
ctx.sync()
print((time.time()-tic)/ntimes)

upgrade_tool_v1_26.tar