Skip to content

implement clear_buffer for A6xx and A7xx

David Heidelberg requested to merge dh/mesa:a6xx-fast-clear_buffer into main

Increase performance of

  • OpenGL ARB_clear_buffer_object impl.
  • OpenCL clEnqueueFillBuffer implementation

Dependency for: !25840 (closed)

Benchmarks (mesa 24.1.2 + Freedreno Rusticl patches, Tinygrad LLM):

  • 950ms per token without
  • 550ms per token with this clear_buffer done by GPU

Relevant tests:

  • GL-CTS: KHR-GL46.direct_state_access.buffers_clear
  • piglit GL: arb_clear_buffer_object-.*
  • piglit CL (with rusticl): cl-api-enqueue-fill-buffer
Edited by David Heidelberg

Merge request reports

Loading