About 2 minor bug fixes on CUDA macOSX 10.13.6

[llv22]

🐛 Bug

During building pytorch 1.7 on macOSX 10.13.6, cuda 10.1(update2), Xcode 10.1(clang 1000.11.45.5), I found the following two bugs and adjusted code accordingly on my local mac. Just update information to you in case that you need to fix as well.

To Reproduce

Using "MAGMA_HOME="/usr/local/lib/magma-cu101" MACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ USE_OPENMP=OFF USE_FBGEMM=OFF python setup.py bdist_wheel" to build pytorch on 1.7 branch

Steps to reproduce the behavior:

1. For issue during building nnpack

"ModuleNotFoundError: No module named 'peachpy.x86_64.avx’" 

1. Error when building ../torch/csrc/jit/codegen/cuda/kernel_cache.cpp

../torch/csrc/jit/codegen/cuda/kernel_cache.cpp:29:25: error: format specifies type 'long' but the argument has type 'int64_t' (aka 'long long') [-Werror,-Wformat]
        printf("%ld, ", shape_symbol.static_size());
                ~~~     ^~~~~~~~~~~~~~~~~~~~~~~~~~
                %lld
../torch/csrc/jit/codegen/cuda/kernel_cache.cpp:31:28: error: format specifies type 'long' but the argument has type 'int64_t' (aka 'long long') [-Werror,-Wformat]
        printf("s(%ld), ", *reinterpret_cast<const int64_t*>(&shape_symbol));
                  ~~~      ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
                  %lld

My code changes

1. For issue during building nnpack: have to generate bytecode ahead of time
   https://github.com/pytorch/pytorch/blob/master/cmake/External/nnpack.cmake#51, add following codes:

  # Orlando: check if avx bytecode has been generated or not
  if(NOT EXISTS "${CAFFE2_THIRD_PARTY_ROOT}/python-peachpy/peachpy/x86_64/avx.py")
    execute_process(COMMAND python setup.by develop
      WORKING_DIRECTORY "${CAFFE2_THIRD_PARTY_ROOT}/python-peachpy")
  endif()

2. Error when building ../torch/csrc/jit/codegen/cuda/kernel_cache.cpp
   https://github.com/pytorch/pytorch/blob/master/torch/csrc/jit/codegen/cuda/kernel_cache.cpp#L48
   int64_t refers to third_party/eigen/Eigen/src/Core/util/Meta.h

      const auto& shape_symbol = sizes.value()[i];
      if (shape_symbol.is_static()) {
        printf("%lld, ", shape_symbol.static_size());
      } else {
        printf("s(%lld), ", *reinterpret_cast<const int64_t*>(&shape_symbol));
      }

Environment

• PyTorch Version (e.g., 1.0): 1.7
• OS (e.g., Linux): macOS 10.13.6
• How you installed PyTorch (conda, pip, source): source
• Build command you used (if compiling from source): MAGMA_HOME="/usr/local/lib/magma-cu101" MACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ USE_OPENMP=OFF USE_FBGEMM=OFF python setup.py bdist_wheel
• Python version: 3.7
• CUDA/cuDNN version: 10.1update 2, cudnn 7.6.5
• GPU models and configuration:
• Any other relevant information:

cc @malfet @seemethere @walterddr @ngimel

[llv22]

@osalpekar I tried to figure out how to submit a minor merge for the bug. Anyway, submit patch here.
orlando-for-patch-torch1.8-mac.patch.txt
this also apply for 1.5-1.7

[ezyang]

@llv22 would you mind opening a pull request for your patch?

[ezyang]

By the way, CUDA on OS X is an unsupported configuration (we're happy to take your fixes, but we won't be testing these continuously)

[llv22]

@ezyang actually I tried with create new branch from master and can't push this branch to torch to create pull request. Is it caused by some missing rights? or we can have another process to make pull request ready?

[ezyang]

Make a fork of PyTorch, and then push your branch to that fork and then open a PR from that branch

[llv22]

@ezyang OK, refer to #46968.