Run Tensorflow on OpenCL™ 1.2 devices

This repo was created from the original Tensorflow repository at:

• https://github.com/tensorflow/tensorflow

Please see the main repository for full Tensorflow documentation. This readme will only focus on the OpenCL porting aspects of Tensorflow.

• Ubuntu 16.04 is used for development
• Ubuntu 14.04 will probably work, but might need some small tweaking (which I can help with)
• Mac OS X builds ok :-) Doesnt run yet though... wheel at https://s3.amazonaws.com/hughperkinstravis/cache/tensorflow-cl/travis/tensorflowpkg.tar.gz Log of failed importing https://travis-ci.org/hughperkins/tensorflow-cl#L2419
  • if you want to help debug the import issue, some ideas at hughperkins#11
• You need:
  • the tensorflow non-gpu installation pre-requisites,
  • an OpenCL 1.2-enabled GPU, and OpenCL 1.2-enabled drivers
  • python 3
• Simply download https://github.com/hughperkins/tensorflow-cl/releases/download/v0.14.0/tensorflow-0.11.0rc0-py3-none-any.whl , and
• Install using pip:

pip install --upgrade tensorflow-0.11.0rc0-py3-none-any.whl

If you want, you can build from source

pip install -r tensorflow/stream_executor/cl/test/requirements.txt

py.test -v

• tensorflow code stays 100% NVIDIA® CUDA™
• cuda-on-cl compiles the CUDA code into OpenCL
• Cedric Nugteren's CLBlast provides BLAS (matrix multiplications)

• OpenCL Torch
• DeepCL

• CLBlast BLAS for OpenCL
• cuda-on-cl Compile CUDA apps for OpenCL
• EasyCL Handles running kernels, passing in arguments etc, on OpenCL

• Dec 3:
  • BUILT A MAC WHEEL!!! This is entirely untested. But the wheel is here: https://s3.amazonaws.com/hughperkinstravis/cache/tensorflow-cl/travis/tensorflowpkg.tar.gz (Simply untar it, and pip install it)
    • corresponding travis log is at https://travis-ci.org/hughperkins/tensorflow-cl/builds/180917138 and https://travis-ci.org/hughperkins/tensorflow-cl/builds/180410593
    • note that I had to built this in several stages, since it's a 3 hour build, and the logs for this are at https://s3.amazonaws.com/hughperkinstravis/cache/tensorflow-cl/travis/90-c520cc1-log.txt and https://s3.amazonaws.com/hughperkinstravis/cache/tensorflow-cl/travis/91-c55079d-log.txt
    • hmmm, doesnt seem to import yet though. https://travis-ci.org/hughperkins/tensorflow-cl#L2419
• Nov 29:
  • Mac build ran to completion! On Travis. Build output https://travis-ci.org/hughperkins/tensorflow-cl/builds/179727517 Yes, it didnt run, didnt create the wheel. But the build_pip_package target built to completion. which is a huge step forward :-) Travis script here: .travis.yml
• Nov 25:
  • release wheel v0.14.0
    • this fixes argmin, argmax, and softmax
    • tons of changes under-the-hood
• Nov 10:
  • released wheel v0.13.0
    • beignet test results fairly solidly match K520 results now
    • fixed the regression on not_equal operator
    • removed the spam from memory copy
• Nov 9:
  • fixed unary and binary operators on beignet
  • note that the tools/bazel.rc.templ has changed. Please make sure to copy the new value into tools/bazel.rc, or re-run configure (probably need to do bazel clean anyway, so might as well do ./configure)