[all-commits] [llvm/llvm-project] df6b80: [offload][OpenMP] Improve cross-team reduction gri...

[Robert Imschweiler via All-commits]

  Branch: refs/heads/users/ro-i/xteam-auto-grid
  Home:   https://github.com/llvm/llvm-project
  Commit: df6b80d3e4fd1a24c35e73bc4f4f58ca71fcec4a
      https://github.com/llvm/llvm-project/commit/df6b80d3e4fd1a24c35e73bc4f4f58ca71fcec4a
  Author: Robert Imschweiler <robert.imschweiler at amd.com>
  Date:   2026-07-08 (Wed, 08 Jul 2026)

  Changed paths:
    M clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp
    M offload/plugins-nextgen/common/include/PluginInterface.h
    M offload/plugins-nextgen/common/src/PluginInterface.cpp

  Log Message:
  -----------
  [offload][OpenMP] Improve cross-team reduction grid selection

The default cross-team reduction algorithm benefits from larger and
fewer teams. Implement that by using 2 x the default number of threads
and 1/2 x the default number of teams for reduction kernels.  This
doesn't change the default total number of threads, it just
redistributes them.

This is a first, rather simple heuristic, derived from (a subset of)
what AOMP does.
The performance benefits I observed for the reduction tests in
https://github.com/ro-i/xteam-test on a gfx942
(c71339705091500f731e2a39f247d2660bacbdce) are up to a few percent, with
no regressions.

Claude assisted with this patch.



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