nir,radeonsi: move ffma fusing to late optimizations for better codegen
The freedreno trace changes were suggested by Rob Clark.
ALU performance is higher, because ffma is used more often, but so is
register usage, because trinary opcodes (such as ffma) usually need
at least 3 live registers.
54793 shaders in 33659 tests
Totals:
SGPRS: 2639746 -> 2642938 (0.12 %)
VGPRS: 1534120 -> 1536392 (0.15 %)
Spilled SGPRs: 3541 -> 3618 (2.17 %)
Spilled VGPRs: 33 -> 44 (33.33 %)
Scratch size: 292 -> 312 (6.85 %) dwords per thread
Code Size: 55639836 -> 55620116 (-0.04 %) bytes
Max Waves: 964785 -> 963977 (-0.08 %)
Totals from affected shaders:
SGPRS: 1105800 -> 1108992 (0.29 %)
VGPRS: 635292 -> 637564 (0.36 %)
Spilled SGPRs: 3193 -> 3270 (2.41 %)
Spilled VGPRs: 33 -> 44 (33.33 %)
Scratch size: 36 -> 56 (55.56 %) dwords per thread
Code Size: 31568708 -> 31548988 (-0.06 %) bytes
Max Waves: 319991 -> 319183 (-0.25 %)
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6596>
4 files changed