fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2025-12-24 08:50:13 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

aco/optimizer: generalize p_create_vector of split vector opt

Browse source 
Foz-DB Navi48:
Totals from 116 (0.14% of 80251) affected shaders:
MaxWaves: 2965 -> 2972 (+0.24%)
Instrs: 145933 -> 144632 (-0.89%); split: -0.91%, +0.02%
CodeSize: 815968 -> 806512 (-1.16%); split: -1.20%, +0.04%
VGPRs: 7240 -> 7144 (-1.33%); split: -1.66%, +0.33%
Latency: 3065858 -> 3063802 (-0.07%); split: -0.11%, +0.05%
InvThroughput: 745395 -> 743506 (-0.25%); split: -0.26%, +0.01%
VClause: 3702 -> 3694 (-0.22%); split: -0.65%, +0.43%
SClause: 3187 -> 3191 (+0.13%)
Copies: 12716 -> 11804 (-7.17%); split: -7.42%, +0.25%
Branches: 3501 -> 3503 (+0.06%)
PreVGPRs: 5400 -> 5327 (-1.35%); split: -1.41%, +0.06%
VALU: 76455 -> 75492 (-1.26%); split: -1.30%, +0.04%
SALU: 23594 -> 23595 (+0.00%); split: -0.00%, +0.01%
VOPD: 1478 -> 1527 (+3.32%); split: +4.67%, -1.35%

Mostly helps FSR4.

Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35674>
This commit is contained in:
Georg Lehmann 2025-06-22 15:07:20 +02:00 • committed by Marge Bot
parent 5fe1b95b33
commit 01d20680e2
1 changed files with 43 additions and 24 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

67
src/amd/compiler/aco_optimizer.cpp
View file

@ -1609,37 +1609,56 @@ label_instruction(opt_ctx& ctx, aco_ptr<Instruction>& instr)
}
}
offset = 0;
for (unsigned i = 0; i < ops.size(); i++) {
if (ops[i].isTemp()) {
if (ctx.info[ops[i].tempId()].is_temp() &&
ops[i].regClass() == ctx.info[ops[i].tempId()].temp.regClass()) {
ops[i].setTemp(ctx.info[ops[i].tempId()].temp);
}
/* If this and the following operands make up all definitions of a `p_split_vector`,
* replace them with the operand of the `p_split_vector` instruction.
*/
Instruction* parent = ctx.info[ops[i].tempId()].parent_instr;
if (parent->opcode == aco_opcode::p_split_vector &&
(offset % 4 == 0 || parent->operands[0].bytes() < 4) &&
parent->definitions.size() <= ops.size() - i) {
copy_prop = true;
for (unsigned j = 0; copy_prop && j < parent->definitions.size(); j++) {
copy_prop &= ops[i + j].isTemp() &&
ops[i + j].getTemp() == parent->definitions[j].getTemp();
}
if (copy_prop) {
ops.erase(ops.begin() + i + 1, ops.begin() + i + parent->definitions.size());
ops[i] = parent->operands[0];
}
}
}
offset += ops[i].bytes();
}
/* combine expanded operands to new vector */
if (ops.size() != instr->operands.size()) {
assert(ops.size() > instr->operands.size());
if (ops.size() <= instr->operands.size()) {
while (instr->operands.size() > ops.size())
instr->operands.pop_back();
if (ops.size() == 1) {
instr->opcode = aco_opcode::p_parallelcopy;
if (ops[0].isTemp())
ctx.info[instr->definitions[0].tempId()].set_temp(ops[0].getTemp());
}
} else {
Definition def = instr->definitions[0];
instr.reset(
create_instruction(aco_opcode::p_create_vector, Format::PSEUDO, ops.size(), 1));
for (unsigned i = 0; i < ops.size(); i++) {
if (ops[i].isTemp() && ctx.info[ops[i].tempId()].is_temp() &&
ops[i].regClass() == ctx.info[ops[i].tempId()].temp.regClass())
ops[i].setTemp(ctx.info[ops[i].tempId()].temp);
instr->operands[i] = ops[i];
}
instr->definitions[0] = def;
} else {
for (unsigned i = 0; i < ops.size(); i++) {
assert(instr->operands[i] == ops[i]);
}
}
if (instr->operands.size() == 2 && instr->operands[1].isTemp()) {
/* check if this is created from split_vector */
ssa_info& info = ctx.info[instr->operands[1].tempId()];
if (info.parent_instr->opcode == aco_opcode::p_split_vector) {
Instruction* split = info.parent_instr;
if (instr->operands[0].isTemp() &&
instr->operands[0].getTemp() == split->definitions[0].getTemp() &&
instr->operands[1].getTemp() == split->definitions[1].getTemp() &&
instr->definitions[0].regClass() == split->operands[0].regClass())
ctx.info[instr->definitions[0].tempId()].set_temp(split->operands[0].getTemp());
}
}
for (unsigned i = 0; i < ops.size(); i++)
instr->operands[i] = ops[i];
break;
}
case aco_opcode::p_split_vector: {