Skip to content

ac: improve LS->HS LDS allocation, tune clear/copy_buffer performance, replace "TC" with "L2", minor tweaks

Marek Olšák requested to merge mareko/mesa:ac-nir-lshs-copybuffer-perf into main
  • WAVES_PER_SH programming is fixed for gfx12
  • LDS for LS->HS varyings is not allocated for any varyings that are passed via VGPRs. Previously, all varyings had to be passed via VGPRs for LDS not to be allocated.
  • Add a DCC image store support failure condition into ac_prepare_compute_blit.
  • ac_prepare_cs_clear_copy_buffer is tuned to have optimal performance for the top GPUs of all generations. The difference in performance compared to what we had before is tremendous.
  • Change most of "TC" to "L2" in comments.

Merge request reports