Skip to content

WIP: tu: Enable concurrent resolves

Danylo Piliaiev requested to merge Danil/mesa:turnip/feature/concurrent-resolves into main

What we know about concurrent resolves:

  • RB_CCU_CNTL::CONCURRENT_RESOLVE controls the resolve modes, I named them:
    • DISABLED
    • AUTO - I don't have evidence that it works, doesn't do anything on a630 or a660, though blob doesn't enable concurrent resolve on a630 and uses MANUAL mode on a660. Presumably HW itself determines when to synchronize blits.
    • MANUAL - we have to manually specify the last blit in the sequence via RB_BLIT_INFO::LAST. I could guess that blits execution is deferred until LAST is met or something else forces synchronization.
  • RB_CCU_CNTL::UNK1 - slows down the resolves.
  • RB_BLIT_INFO::BUFFER_ID - if two blits have the same buffer id, they cannot be executed in parallel. blob uses it in both GL and VK contrary to the old comment.
  • RB_BLIT_INFO::LAST - see CONCURRENT_RESOLVE::MANUAL.

Testing on a660, enabling MANUAL concurrent resolve mode I've got 10-20% perf improvement of gmem stores. For testing I had 8 color and 1 D/S attachment with size of 2048x2048, and one triangle draw call.

Tested with v631 blob driver:

GPU Resolve Mode RB_CCU_CNTL::UNK1
616/618/619 disabled true
620 auto false
630 disabled true
640 disabled true
650 auto false
660 manual false

Merge request reports