For me, the number one gripe with TF on Mac is the lack of OpenCL support. With the new iMac Pro coming out with an AMD Vega GPU, and the 2016 MBP having Intel graphics, I feel like it should at least be on their list of things to do. But then again, it's probably not a good idea to train on a Mac.
I see a couple of issues with the Mac / TF OpenCL situation.
Apple is pushing metal and performance shaders,
I would not bet on strong support from them keeping up to date drivers with performant ocl.
I asked at wwdc17 about vulkan and they clearly do not want it on their platform.
Last but not least there is an issue with the ocl support.
It is (was?) done on linux using computeCPP, compiled to SPIR (Standard Portable Intermediate Representation). MacOS does not support SPIR at the moment.
My best guess would be to use IOMMU magic and hardware virtualization to hook a GPU to a linux VM with MacOS as a host.
ERROR: /private/var/tmp/_bazel_qixiang/5571543a3fba611a0b0ca8efd02f3a37/external/protobuf_archive/BUILD:265:1: C++ compilation of rule '@protobuf_archive//:js_embed' failed (Exit 1).
gcc-7: error: unrecognized command line option '-fcolor-diagnostics'
gcc-7: error: unrecognized command line option '-Wthread-safety'; did you mean '-fthread-jumps'?
gcc-7: error: unrecognized command line option '-Wself-assign'; did you mean '-Wcast-align'?
Target //tensorflow/tools/pip_package:build_pip_package failed to build
Bazel recognizes it's compiling on macOS, so it's using clang-compatible flags, which are incompatible with gcc. You should use the default host compiler, as stated in the [configure output][1] in the post. /usr/bin/gcc is for all intents and purposes clang on macOS.
gcc-7 is only installed to provide libgomp which is otherwise missing from a standard macOS installation.
I'll update the post to make it clearer. Thanks for pointing that out!
It's worth noting that any TF user (including those on NVidia Macs) can get significant speedups (20%+) by recompiling to use all the features of your CPU.
On OSX (as the linked post shows) it's pretty rare to use the GPU.
Not all TF components are GPU enabled anyway. The default LSTM isn't and you need to use CuLSTM. That is nice but doesn't support the GoBackwards operation so you can't use it in BiLSTMs.
GPUs are great but there is lots done on the CPU too.
Edit: To clarify, the comments on the LSTM implementations are actually on the Keras/TF LSTM. Not sure what the native TF one is like.
For me, the number one gripe with TF on Mac is the lack of OpenCL support. With the new iMac Pro coming out with an AMD Vega GPU, and the 2016 MBP having Intel graphics, I feel like it should at least be on their list of things to do. But then again, it's probably not a good idea to train on a Mac.
I see a couple of issues with the Mac / TF OpenCL situation.
Apple is pushing metal and performance shaders, I would not bet on strong support from them keeping up to date drivers with performant ocl. I asked at wwdc17 about vulkan and they clearly do not want it on their platform.
The TF OpenCL issue is vocal on the repo. Yet to see progress for macos. https://github.com/tensorflow/tensorflow/issues/22
Last but not least there is an issue with the ocl support. It is (was?) done on linux using computeCPP, compiled to SPIR (Standard Portable Intermediate Representation). MacOS does not support SPIR at the moment.
My best guess would be to use IOMMU magic and hardware virtualization to hook a GPU to a linux VM with MacOS as a host.
I run bazel build and I get this:
ERROR: /private/var/tmp/_bazel_qixiang/5571543a3fba611a0b0ca8efd02f3a37/external/protobuf_archive/BUILD:265:1: C++ compilation of rule '@protobuf_archive//:js_embed' failed (Exit 1). gcc-7: error: unrecognized command line option '-fcolor-diagnostics' gcc-7: error: unrecognized command line option '-Wthread-safety'; did you mean '-fthread-jumps'? gcc-7: error: unrecognized command line option '-Wself-assign'; did you mean '-Wcast-align'? Target //tensorflow/tools/pip_package:build_pip_package failed to build
any help?
Author of the post here.
Bazel recognizes it's compiling on macOS, so it's using clang-compatible flags, which are incompatible with gcc. You should use the default host compiler, as stated in the [configure output][1] in the post. /usr/bin/gcc is for all intents and purposes clang on macOS.
gcc-7 is only installed to provide libgomp which is otherwise missing from a standard macOS installation.
I'll update the post to make it clearer. Thanks for pointing that out!
[1]: https://gist.github.com/crmne/a22858264d68dd9fe935fc4ee48e04...
wow...it works! THX！
Should also publish the benchmarks of best of class machines running TensorFlow - for Linux, Windows and Mac HWs.
Show Apple why they need to release better Mac Desktop/Laptop HW or risk loss the heart and mind of next generation of developers.
It's worth noting that any TF user (including those on NVidia Macs) can get significant speedups (20%+) by recompiling to use all the features of your CPU.
TF even warns you about this when you start up.
if tf runs most of the workload GPU side, SSE enabled binary is less compatible and no longer a bottleneck.
On OSX (as the linked post shows) it's pretty rare to use the GPU.
Not all TF components are GPU enabled anyway. The default LSTM isn't and you need to use CuLSTM. That is nice but doesn't support the GoBackwards operation so you can't use it in BiLSTMs.
GPUs are great but there is lots done on the CPU too.
Edit: To clarify, the comments on the LSTM implementations are actually on the Keras/TF LSTM. Not sure what the native TF one is like.
Are you talking about LSTM implementation in OSX version of TF? Because it definitely runs on GPU in Windows/Linux.