pytorch

66dc1aba - [Inductor][MacOS] resolve macos openmp problem and provide a holistic instruction (#107111)

[Inductor][MacOS] resolve macos openmp problem and provide a holistic instruction (#107111) There has been several reports of difficulty in using OpenMP in MacOS, e.g.: https://github.com/pytorch/pytorch/issues/95708 . And there are several PRs to fix it, e.g.: https://github.com/pytorch/pytorch/pull/93895 and https://github.com/pytorch/pytorch/pull/105136 . This PR tries to explain the root cause, and provide a holistic and systematic way to fix the problem. For the OpenMP program below to run, the compiler must: - Be able to process macros like `#pragma omp parallel` - Be able to find header files like `<omp.h>` - Be able to link to a library file like `libomp` ```C++ #include <omp.h> int main() { omp_set_num_threads(4); #pragma omp parallel { int id = omp_get_thread_num(); int nthrds = omp_get_num_threads(); int y = id * nthrds; } } ``` In MacOS, there might be different compiler tools: - Apple builtin `clang++`, installed with `xcode commandline tools`. The default `g++` and `clang++` commands both point to the Apple version, as can be confirmed by `g++ --version` - Public `clang++`, can be installed via `brew install llvm`. - Public GNU compiler `g++`, can be installed via `brew install gcc`. Among these compilers, public `clang++` from LLVM and `g++` from GNU both support OpenMP with the flag `-fopenmp`. They have shipped with `<omp.h>` and `libomp` support. The only problem is that Apple builtin `clang++` does not contain `<omp.h>` or `libomp`. Therefore, users can follow the steps to enable OpenMP support: - Use a compiler other than Apple builtin clang++ by specifying the `CXX` environment variable - Use `conda install llvm-openmp` to place the header files and lib files inside conda environments (and can be discovered by `CONDA_PREFIX`) - Use `brew install libomp` to place the header files and lib files inside brew control (and can be discovered by `brew --prefix libomp`) - Use a custom install of OpenMP by specifying an `OMP_PREFIX` where header files and lib files can be found. This PR reflects the above logic, and might serve as a final solution for resolving OpenMP issues in MacOS. This PR also resolves the discussion raised in https://dev-discuss.pytorch.org/t/can-we-add-a-default-backend-when-openmp-is-not-available/1382/5 with @jansel , and provide a way for brew users to automatically find the installation via `brew --prefix libomp`, and provide instructions to switch to another compiler by `CXX` environment variable. I have tested the following code in different conditions: - Use `CXX` to point to an LLVM-clang++, works fine. - Use `CXX` to point to a GNU g++, not working because the compiler flag `-Xclang`. Manually removing the code `base_flags += " -Xclang"` works. - Use default compiler and `conda install llvm-openmp`, works fine - Use default compiler and `brew install libomp`, works fine - Do nothing, compiler complains `omp.h` not found. ```python import torch @torch.compile def f(x): return x + 1 f(torch.randn(5, 5)) ``` If we want the code to be more portable, we can also deal with the `-Xclang` issue. Pull Request resolved: https://github.com/pytorch/pytorch/pull/107111 Approved by: https://github.com/jgong5, https://github.com/jansel