NeroReflex/mesa
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

aco: use v_cvt_f32_ubyte for signed casts too

Browse source 
The extract is always positive, so signed vs unsigned conversion doesn't matter.

Foz-DB GFX11:
Totals from 167 (0.13% of 133461) affected shaders:
Instrs: 401631 -> 401225 (-0.10%)
CodeSize: 2107256 -> 2104344 (-0.14%)
VGPRs: 13320 -> 13332 (+0.09%)
Latency: 6468063 -> 6467241 (-0.01%)
InvThroughput: 801854 -> 801653 (-0.03%)
Copies: 13926 -> 13927 (+0.01%); split: -0.08%, +0.09%

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24893>
This commit is contained in:
Georg Lehmann 2023-08-25 19:23:14 +02:00 • committed by Marge Bot
parent 144c2d4e4a
commit cda5784eb3
1 changed files with 6 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

8
src/amd/compiler/aco_optimizer.cpp
View file

@ -1065,7 +1065,9 @@ can_apply_extract(opt_ctx& ctx, aco_ptr<Instruction>& instr, unsigned idx, ssa_i
return false;
} else if (sel.size() == 4) {
return true;
} else if (instr->opcode == aco_opcode::v_cvt_f32_u32 && sel.size() == 1 && !sel.sign_extend()) {
} else if ((instr->opcode == aco_opcode::v_cvt_f32_u32 ||
instr->opcode == aco_opcode::v_cvt_f32_i32) &&
sel.size() == 1 && !sel.sign_extend()) {
return true;
} else if (instr->opcode == aco_opcode::v_lshlrev_b32 && instr->operands[0].isConstant() &&
sel.offset() == 0 &&
@ -1119,7 +1121,9 @@ apply_extract(opt_ctx& ctx, aco_ptr<Instruction>& instr, unsigned idx, ssa_info&
if (sel.size() == 4) {
/* full dword selection */
} else if (instr->opcode == aco_opcode::v_cvt_f32_u32 && sel.size() == 1 && !sel.sign_extend()) {
} else if ((instr->opcode == aco_opcode::v_cvt_f32_u32 ||
instr->opcode == aco_opcode::v_cvt_f32_i32) &&
sel.size() == 1 && !sel.sign_extend()) {
switch (sel.offset()) {
case 0: instr->opcode = aco_opcode::v_cvt_f32_ubyte0; break;
case 1: instr->opcode = aco_opcode::v_cvt_f32_ubyte1; break;