[all-commits] [llvm/llvm-project] 85974a: [flang-rt] Add experimental support for GPU build ...

[Joseph Huber via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 85974a0537b59268288108d531b0681fa35afc4a
      https://github.com/llvm/llvm-project/commit/85974a0537b59268288108d531b0681fa35afc4a
  Author: Joseph Huber <huberjn at outlook.com>
  Date:   2025-03-24 (Mon, 24 Mar 2025)

  Changed paths:
    M flang-rt/CMakeLists.txt
    M flang-rt/cmake/modules/AddFlangRT.cmake
    M flang-rt/cmake/modules/HandleLibs.cmake
    M flang-rt/lib/runtime/CMakeLists.txt
    M flang/cmake/modules/FlangCommon.cmake

  Log Message:
  -----------
  [flang-rt] Add experimental support for GPU build (#131826)

Summary:
This patch adds initial support for compiling `flang-rt` directly for
the GPU. The method used here matches what's already done for `libc` and
`libc++` for the GPU and builds off of those projects.

Mainly this requires setting up some flags and setting the sources that
currently work. This will deposit the resulting library in the
appropriate directory. These files are then intended to be linked via
`-Xoffload-linker` support in the offloading driver.
```
lib/clang/21/lib/nvptx64-nvidia-cuda/libflang_rt.runtime.a
lib/clang/21/lib/amdgcn-amd-amdhsa/libflang_rt.runtime.a
```

This is obviously missing a lot of functions, mainly the `io` support.
Most of what we cannot support is due to using POSIX things that just
don't make sense on the GPU. Stuff like `pthreads` or `sema`.

Getting unit tests to run on this will also be a challenge. We could run
tests the same way we do with `libc`, but the problem there is that the
`libc` test suite is freestanding while `gtest` currently doesn't
compile on the GPU bcause it uses a lot of weird stuff. If the unit
tests were simply `int main` then it would work.

I don't understand the actual runtime code very well, I'd appreciate
some guidance on how to actually support Fortran IO from this interface.
As I understand it, Fortran IO requires a stack-like operation, which
conflicts with the SIMT model GPUs use. Worst case scenario we could
burn some LDS to keep a stack, or serialize it somehow since we can
always just iterate over all the active lanes.

Building this right now looks like this, which depends on the arguments
added in https://github.com/llvm/llvm-project/pull/131695.
```
    -DRUNTIMES_nvptx64-nvidia-cuda_LLVM_ENABLE_RUNTIMES=compiler-rt;libc;libcxx;libcxxabi;flang-rt \
    -DRUNTIMES_amdgcn-amd-amdhsa_LLVM_ENABLE_RUNTIMES=compiler-rt;libc;libcxx;libcxxabi;flang-rt \
    -DRUNTIMES_nvptx64-nvidia-cuda_FLANG_RT_LIBC_PROVIDER=llvm \
    -DRUNTIMES_nvptx64-nvidia-cuda_FLANG_RT_LIBCXX_PROVIDER=llvm \
    -DRUNTIMES_amdgcn-amd-amdhsa_FLANG_RT_LIBC_PROVIDER=llvm \
    -DRUNTIMES_amdgcn-amd-amdhsa_FLANG_RT_LIBCXX_PROVIDER=llvm
```



To unsubscribe from these emails, change your notification settings at https://github.com/llvm/llvm-project/settings/notifications