Skip to content

aco: optimize ddx/ddy

Georg Lehmann requested to merge DadSchoorse/mesa:aco-interp-ddxy into main

When using DPP for derivatives, reorder the operations to have v_mov_b32_dpp as the last instruction. This allows fusing it with the use instruction. On GFX11.5, use VINTERP instead of DPP when possible, because it can be dual issued.

Edited by Georg Lehmann

Merge request reports