[all-commits] [llvm/llvm-project] 83ff0f: [Clang][OpenMP] Allow unified_shared_memory for Pa...

[Michael Kruse via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 83ff0ff46337422171fb36f934bd56c2bc1be15c
      https://github.com/llvm/llvm-project/commit/83ff0ff46337422171fb36f934bd56c2bc1be15c
  Author: Michael Kruse <llvm-project at meinersbur.de>
  Date:   2021-05-13 (Thu, 13 May 2021)

  Changed paths:
    M clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp
    M clang/test/OpenMP/requires_codegen.cpp

  Log Message:
  -----------
  [Clang][OpenMP] Allow unified_shared_memory for Pascal-generation GPUs.

The Pascal architecture supports the page migration engine required for
unified_shared_memory, as indicated by NVIDIA:
 * https://developer.nvidia.com/blog/unified-memory-cuda-beginners/
 * https://developer.nvidia.com/blog/beyond-gpu-memory-limits-unified-memory-pascal/
 * https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#um-requirements

The limitation was introduced in D54493 which justified the cut-off by
the requirement for unified addressing. However, Unified Virtual
Addressing (UVA) is already available with sm20 (Fermi, Kepler,
Maxwell):
 * https://docs.nvidia.com/cuda/gpudirect-rdma/index.html#basics-of-uva-cuda-memory-management

Unified shared memory might even be possible with these, but with
migration of entire allocations on kernel startup.

Reviewed By: JonChesterfield

Differential Revision: https://reviews.llvm.org/D101595