GPU device not found on OSX

[maxgillett]

Hello,

I'm trying to install TensorFlow with GPU support for a CUDA Capability 3.0 device. I have followed the instructions found here: https://gist.github.com/Mistobaan/dd32287eeb6859c6668d, and everything compiles without error, but TensorFlow doesn't recognize my GPU device. The steps that I'm taking are below, as is the error output.

I've modified bazel to use gcc-4.8, so these are the commands I'm running:

bazel build -c opt --config=cuda //tensorflow/cc:tutorials_example_trainer ---crosstool_top=//tools/cpp:toolchain
bazel build -c opt --config=cuda //tensorflow/tools/pip_package:build_pip_package --crosstool_top=//tools/cpp:toolchain

Running the example with:

bazel-bin/tensorflow/cc/tutorials_example_trainer --use_gpu

produces the following output:

F tensorflow/cc/tutorials/example_trainer.cc:118] Check failed: ::tensorflow::Status::OK() == (session->Run({{"x", x}}, {"y:0", "y_normalized:0"}, {}, &outputs)) (OK vs. Invalid argument: Cannot assign a device to node 'Const/_2': Could not satisfy explicit device specification '/gpu:0'
     [[Node: Const/_2 = Const[dtype=DT_INT32, value=Tensor<type: int32 shape: [] values: 0>, _device="/gpu:0"]()]])
F tensorflow/cc/tutorials/example_trainer.cc:118] Check failed: ::tensorflow::Status::OK() == (session->Run({{"x", x}}, {"y:0", "y_normalized:0"}, {}, &outputs)) (OK vs. Invalid argument: Cannot assign a device to node 'Const/_2': Could not satisfy explicit device specification '/gpu:0'
     [[Node: Const/_2 = Const[dtype=DT_INT32, value=Tensor<type: in

I can also confirm that after installing the pip package and running a test script, no devices are found.

A few things to note:

• Running deviceQuery confirms that a CUDA device (Device 0: GeForce GT 650M) is found.
• I've made sure to run bazel clean before compiling, and have tried compiling using sudo just to see if this was a permissions issue.

Thanks in advance for your help.

[maxgillett]

EDIT: I can successfully compile also with Clang 7.0.2, which I can confirm is supported by NVCC. I'm still getting the same error as above.

It turns out that I was actually compiling with gcc-5, not gcc-4.8. From what I can tell CUDA doesn't support version 5, so this might be the problem.

I'm trying now to compile with 4.8, but I'm getting this error:

/var/tmp//cccfsYPu.s:2774:expecting string instruction after `rep'~~

From what I've read this seems to be a problem with binutils (at least on Linux), so I installed binutils 2.25.1 with Homebrew, but it didn't help. I've also tried including the compiler flag `-march=corei7'

[vrv]

De-duping with #491