• Marek Olšák's avatar
    radeonsi: fuse or lower ffma optimally on all chips · f1284505
    Marek Olšák authored
    
    
    LLVM is going to support the legacy instructions soon.
    
    This change switches FMA to MAD for gfx10.
    
    54793 shaders in 33659 tests
    Totals:
    SGPRS: 2632554 -> 2629570 (-0.11 %)
    VGPRS: 1536364 -> 1535312 (-0.07 %)
    Spilled SGPRs: 3602 -> 3562 (-1.11 %)
    Spilled VGPRs: 44 -> 40 (-9.09 %)
    Private memory VGPRs: 256 -> 256 (0.00 %)
    Scratch size: 312 -> 308 (-1.28 %) dwords per thread
    Code Size: 55422660 -> 55345408 (-0.14 %) bytes
    Max Waves: 963983 -> 964200 (0.02 %)
    
    Acked-by: Pierre-Eric Pelloux-Prayer's avatarPierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
    Part-of: <!6756>
    f1284505