Skip to content

intel/nir: add a pass to shrink const block loads

Lionel Landwerlin requested to merge llandwerlin/mesa:review/intel-nir-block-load-trim into main

The NIR helpers for ray tracing sometimes load more than necessary. This simple optimization trims the loads to what is needed.

Some numbers out of shader-db on Q2RTX :

Instructions in all programs: 85137 -> 84366 (-0.9%)
SENDs in all programs: 3438 -> 3452 (+0.4%)
Loops in all programs: 40 -> 40 (+0.0%)
Cycles in all programs: 2462650 -> 2415668 (-1.9%)
Spills in all programs: 401 -> 368 (-8.2%)
Fills in all programs: 693 -> 625 (-9.8%)

Signed-off-by: Lionel Landwerlin lionel.g.landwerlin@intel.com

Edited by Lionel Landwerlin

Merge request reports