[Openmp-commits] [PATCH] D132660: [openmp][amdgpu] Implement target_alloc_host as fine grain HSA memory

[Jon Chesterfield via Phabricator via Openmp-commits]

JonChesterfield added inline comments.


================
Comment at: openmp/libomptarget/plugins/amdgpu/src/rtl.cpp:2693
   DP("Tgt free data (tgt:%016llx).\n", (long long unsigned)(Elf64_Addr)TgtPtr);
   Err = core::Runtime::Memfree(TgtPtr);
   if (Err != HSA_STATUS_SUCCESS) {
----------------
Curious aside - cuda rtl.cpp has a hashtable of pointers to determine what sort of allocation was involved, which is ugly, and also broken if a given address can have different kind across multiple allocations (maybe? cuda implementation property) as it grows without bound on multiple allocations. I was postponing implementing this patch until that could be removed in favour of a dedicated free call, but it turns out HSA does the same sort of map-pointer-to-deallocator under the hood anyway.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D132660/new/

https://reviews.llvm.org/D132660