Build applications written in NVIDIA® CUDA™ code for OpenCL™ 1.2 devices.

• Compile using cocl
• link using -lcocl -lOpenCL
• at runtime, loads libOpenCL.so

Here is a screenshot of running on a Mac:

• write a CUDA sourcecode file, or find an existing one
• here's a simple example: cuda_sample.cu
• Use cocl to compile cuda_sample.cu:

$ cocl cuda_sample.cu
   ...
   ... (bunch of compily stuff) ...
   ...

    ./cuda_sample.cu compiled into ./cuda_sample

Run:

$ ./cuda_sample
Using Intel , OpenCL platform: Intel Gen OCL Driver
Using OpenCL device: Intel(R) HD Graphics 5500 BroadWell U-Processor GT2
hostFloats[2] 123
hostFloats[2] 222
hostFloats[2] 444

If you want, you can compile in two steps:

cocl -c teststream.cu
g++ -o teststream teststream.o -lcocl -lclblast -leasycl -lclew

Result is the same:

$ ./cuda_sample
Using Intel , OpenCL platform: Intel Gen OCL Driver
Using OpenCL device: Intel(R) HD Graphics 5500 BroadWell U-Processor GT2
hostFloats[2] 123
hostFloats[2] 222
hostFloats[2] 444

The options provided to -devicell-opt are passed through to opt-3.8, https://llvm.org/docs/Passes.html

opt-3.8 fits in as follows:

• clang-3.8 -x cuda --device-only converts the incoming .cu file to LLVM IR
• opt-3.8 optimizes the IR. -devicell-opt control this
• ir-to-opencl writes the IR as OpenCL

Recommended generation options:

• -devicell-opt inline -devicell-opt mem2reg -devicell-opt instcombine --devicell-opt O2

You can open the -device.cl file to look at the OpenCL generated, and compare the effects of different options.

Behind the scenes, there are a few parts:

• Device-side, cocl converts the CUDA kernels into OpenCL kernels
• Host-side, cocl:
  • converts the cuda kernel launch code into opencl kernel launch code, and
  • bakes in the OpenCL code

More detail

New!

• the device-IR to OpenCL step happens at runtime now
  • surprisingly, this actually is faster than doing it offline
  • thats because the GPU driver only needs to compile the small amount of OpenCL needed for a specific kernel, rather than an entire IR file
• in addition, address-space deduction is significantly facilitated

• compiler for host-side code, including memory allocation, copy, streams, kernel launches
• compiler for device-side code, handling templated C++ code, converting it into bog-standard OpenCL 1.2 code
• BLAS (using Cedric Nugteren's CLBlast)

• Operating system:
  • Tested/developed on Ubuntu 16.04
  • Ubuntu 14.04 does seem to work ok too (not tested very much though...)
  • Mac OS X, see https://travis-ci.org/hughperkins/cuda-on-cl
  • Other operating systems, and clang/llvm versions, might work too, but untested. Your mileage may vary :-)
• OpenCL-enabled GPU, and appropriate OpenCL drivers installed for the GPU

cd ~
wget https://llvm.org/releases/3.8.0/clang+llvm-3.8.0-x86_64-apple-darwin.tar.xz
tar -xf clang+llvm-3.8.0-x86_64-apple-darwin.tar.xz

set CLANG_HOME to the root of this directory, ie export CLANG_HOME=$HOME/clang+llvm-3.8.0-x86_64-apple-darwin

sudo apt-get install llvm-3.8 llvm-3.8-dev clang-3.8

set CLANG_HOME to /usr/lib/llvm-3.8

• download from https://llvm.org/releases/download.html , and decompress
• set CLANG_HOME to point to the resulting directory, the one that contains bin, and lib

On Ubuntu 16.04:

sudo apt-get install git cmake cmake-curses-gui libc6-dev-i386 make gcc g++

On other systems:

• somehow install similar things as for Ubuntu 16.04 section

git clone --recursive https://github.com/hughperkins/cuda-on-cl
cd cuda-on-cl
mkdir build
cd build
cmake ..
make -j 4
sudo make install

Note that you'll need to continue to export CLANG_HOME environment variable when using cocl.

There are the following tests:

• low-level tests, using gtest
• mid-level tests, using py.test
• end-to-end tests
• Eigen tests, in Eigen repo, https://bitbucket.org/hughperkins/eigen/src/eigen-cl/unsupported/test/cuda-on-cl/?at=eigen-cl

cd build
make -j 4
./cocl_unittests

No dependencies on graphics card etc. It simply takes some hand-crafted IR, and writes it to OpenCL. It never actually tries to run the OpenCL, so it validates:

• can cocl handle the IR without choking/crashing?
• do the hand-crafted OpenCL expected results match up with the actual cocl outputs?

pip install -r test/requirements.txt

OFFSET_32BIT=1 \
COCL_OPTIONS='--devicell-opt inline --devicell-opt mem2reg --devicell-opt instcombine --devicell-opt O2' \
py.test -v

• python tests are at test

You can modify the options in COCL_OPTIONS. There are passed to the cocl command, see section #Options above.

If you set OFFSET_32BIT to off in your cmake options, you should remove the OFFSET_32BIT=1 optio nwhen running py.test

Run:

cd build
ccmake ..

turn on BUILD_TESTS, and run the build.

Now you can do, from build directory:

make run-tests

You can run a test by name, eg:

make run-offsetkernelargs

Result:

################################
# running:
################################
LD_LIBRARY_PATH=build: build/test-cocl-offsetkernelargs
Using Intel , OpenCL platform: Intel Gen OCL Driver
Using OpenCL device: Intel(R) HD Graphics 5500 BroadWell U-Processor GT2
126.456

• end-to-end tests are at test/cocl

From ccmake .., there are various options you can choose, that affect hte OpenCL code produced. These options will affect how well the OpenCL generation works, and how acceptable it is to your GPU driver. If you're reading the OpenCL code ,they will affect readability too.

You can see the section Options above for more details.

See docker. Docker images run ok on beignet and NVIDIA :-)

• Eigen-CL: Minimally-tweaked fork of Eigen, which can be compiled/run using cuda-on-cl, on an OpenCL device, https://bitbucket.org/hughperkins/eigen/commits/branch/eigen-cl
• Tensorflow-CL: Fork of Tensorflow, that can be built and run on an OpenCL-1.2 enabled GPU, using cuda-on-cl, https://github.com/hughperkins/tensorflow-cl

Apache 2.0

• Nov 25:
  • created release 4.0.4:
    • all current Eigen tests, https://bitbucket.org/hughperkins/eigen/src/eigen-cl/unsupported/test/cuda-on-cl/?at=eigen-cl , pass for me currently, using this release, on both beignet 1.2.1, on hd5500, and on NVIDIA 940M, using driver 367.57
    • fixes some issues with walk-dependency order during cl walk/generation
    • fixed an issue with un-initialized pointers, in structs passed as kernel parameters, into GPU kernels
• Nov 24:
  • merge from branch clwriter:
    • lots of refactorization under-the-hood
    • can handle determining the address-space of functions returning pointers
    • opencl generation is at runtime now => facilitates determining address-space; and counter-intuitively is actually faster, because less OpenCL to compile by the GPU driver
• Nov 18:
  • Mac build working :-) https://travis-ci.org/hughperkins/cuda-on-cl/builds/176997220#L1356
• Nov 17:
  • merged runtime-compile branch into master branch. This brings a few changes:
    • opencl generation is now at runtime, rather than at compile time
      • this lets us build only the one specific kernel we need
      • means more information is available at generation time, facilitating the generation process
    • build on Mac OS X is more or less working, eg https://travis-ci.org/hughperkins/cuda-on-cl/builds/176580716
    • code radically refactorized underneath
    • remove --run_branch_transforms, --branches_as_switch, for now
• Nov 8:
  • exposed generation options as cocl options, eg --run_branching_transforms, --branches_as_switch, and the --devicell-opt [opt] options
• Nov 6:
  • created dockerfiles for Beignet and NVIDIA docker
• Nov 5:
  • switched from Makefile to CMakeLists.txt => build/install instructions have changed, see above
  • added a cmake file, so you can easily add cocl to your cmakelists file, eg see https://bitbucket.org/hughperkins/eigen/src/d84b9f44f924e36a8527e66a46a189395f046d21/unsupported/test/cuda-on-cl/CMakeLists.txt?at=eigen-cl&fileviewer=file-view-default for an example
• Older news