Saturday, 3 March 2018

Compiling Tensorflow with OpenCL SYCL on Intel GPU/CPU

Useful resources:
https://arrayfire.com/opencl-on-intel-hd-iris-graphics-on-linux/

https://developer.codeplay.com/computecppce/v0.6.0/supported-platforms

https://askubuntu.com/questions/850281/opencl-on-ubuntu-16-04-intel-sandy-bridge-cpu

https://software.intel.com/en-us/articles/opencl-drivers#latest_linux_driver

https://github.com/benoitsteiner/tensorflow-opencl/issues/37

I had the same problem but solved the import error as following

I have the computecpp folder stored at /usr/local/computecpp.
By default, libComputeCpp.so is under computercpp/lib.
So do > export LD_LIBRARY_PATH=/usr/local/computecpp/lib to add the path
Now import should work.


http://deep-beta.co.uk/setting-up-tensorflow-with-opencl-using-sycl/


https://github.com/tensorflow/tensorflow/issues/15204#issuecomment-370069712

https://github.com/tensorflow/tensorflow/issues/2109

https://github.com/lukeiwanski/tensorflow/issues/215

https://github.com/tensorflow/tensorflow/issues/2109

https://github.com/benoitsteiner/tensorflow-opencl

https://github.com/lukeiwanski/tensorflow

https://github.com/hughperkins/tf-coriander


What has worked out-of-the-box:
https://github.com/lukeiwanski/tensorflow/tree/dev/eigen_mehdi

The only distribution that compiles fine  - without any modification

The issue is with Intel drivers - when running tensorflow training, intel GPU device is not detected - while CPU is detected but the code hangs - might require compiling of linux header with a patch.


Till now no success in getting the device and running it