Skip to content

nir/loop_analyze: adjust force unrolling to only include interesting modes

Rhys Perry requested to merge pendingchaos/mesa:nir_loop_analyze_force_unroll into master

Instead of force-unrolling any loop which reads an entire array, only do it for arrays which might be faster to access with constant indices.

Significantly improves compile-time for these CTS tests, which could previously timeout:

  • dEQP-VK.spirv_assembly.instruction.graphics.16bit_storage.struct_mixed_types.uniform_buffer_block_geom
  • dEQP-VK.spirv_assembly.instruction.graphics.16bit_storage.struct_mixed_types.uniform_geom
  • dEQP-VK.spirv_assembly.instruction.graphics.8bit_storage.struct_mixed_types.storage_buffer_geom
  • dEQP-VK.spirv_assembly.instruction.graphics.spirv_ids_abuse.lots_ids_geom

fossil-db (Navi):

Totals from 19 (0.01% of 137413) affected shaders:
SGPRs: 1728 -> 1688 (-2.31%)
VGPRs: 1176 -> 1168 (-0.68%)
CodeSize: 198496 -> 136580 (-31.19%)
MaxWaves: 154 -> 156 (+1.30%)
Instrs: 38889 -> 26029 (-33.07%)
Cycles: 446108 -> 1059924 (+137.59%); split: -0.91%, +138.51%
VMEM: 3245 -> 2926 (-9.83%)
SMEM: 850 -> 828 (-2.59%); split: +4.71%, -7.29%
VClause: 549 -> 533 (-2.91%)
SClause: 1810 -> 1522 (-15.91%)
Copies: 2209 -> 1705 (-22.82%); split: -22.95%, +0.14%
Branches: 854 -> 603 (-29.39%); split: -29.86%, +0.47%
PreSGPRs: 1512 -> 1506 (-0.40%); split: -0.53%, +0.13%

Merge request reports