Does TensorFlow plan to support OpenCL?

Question

I am under the impression that TF only supports CUDA. From my research, the only way to get it to support OpenCL and AMD GPUs is to get a separate "adapter" module such as TF Coriander and suffer through the performance failure:

From the GitHub:

for the recurrent networks, Coriander is around 4 times slower than using NVIDIA® CUDA™ directly.

The module does seem to be under near constant development, checking the insights tab, but I'm not sure how optimistic to be. So should I be optimistic about this and expect a near-native experience in under a few years? If not, is support being planned for an actually native solution? If they aren't will AMD's upcoming 7nm architecture change the dev's minds?

I realize this may be a little off-topic. If it is, please tell me if I can put anywhere else.

Answer 1

As part of contrib, you can build Tensorflow with SYCL support.

SYCL is "single source OpenCL", a new standard from Khronos that allows one to write high level C++ code that can be compiled to run on OpenCL devices.

The folks at CodePlay software have been heavily involved in it, and you can see one their blogposts on the topic here.

So in short, you're not going to get a pip package of it, you'll need to build it yourself. And the performance might not be as good, since the project is still early days.

You can find a tutorial on how to get started here. Bear in mind, this uses CodePlay's propitiatory version of SYCL, but maybe you can get it working with an open implementation such as triSYCL.