fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-04-30 23:28:06 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 5

aco: add affinities for possible sopk optimizations

Browse source 
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/29512>
This commit is contained in:
Georg Lehmann 2024-06-01 20:28:56 +02:00 • committed by Marge Bot
parent fac475bc25
commit 39380d475a
1 changed files with 40 additions and 30 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

70
src/amd/compiler/aco_register_allocation.cpp
View file

@ -2586,6 +2586,32 @@ vop3_can_use_vop2acc(ra_ctx& ctx, Instruction* instr)
return true; return true;
} }
bool
sop2_can_use_sopk(ra_ctx& ctx, Instruction* instr)
{
if (instr->opcode != aco_opcode::s_add_i32 && instr->opcode != aco_opcode::s_mul_i32 &&
instr->opcode != aco_opcode::s_cselect_b32)
return false;
uint32_t literal_idx = 0;
if (instr->opcode != aco_opcode::s_cselect_b32 && instr->operands[1].isLiteral())
literal_idx = 1;
if (!instr->operands[!literal_idx].isTemp() || !instr->operands[!literal_idx].isKillBeforeDef())
return false;
if (!instr->operands[literal_idx].isLiteral())
return false;
const uint32_t i16_mask = 0xffff8000u;
uint32_t value = instr->operands[literal_idx].constantValue();
if ((value & i16_mask) && (value & i16_mask) != i16_mask)
return false;
return true;
}
void void
get_affinities(ra_ctx& ctx, std::vector<IDSet>& live_out_per_block) get_affinities(ra_ctx& ctx, std::vector<IDSet>& live_out_per_block)
{ {
@ -2667,6 +2693,8 @@ get_affinities(ra_ctx& ctx, std::vector<IDSet>& live_out_per_block)
op = instr->operands[op_fixed_to_def0]; op = instr->operands[op_fixed_to_def0];
} else if (vop3_can_use_vop2acc(ctx, instr.get())) { } else if (vop3_can_use_vop2acc(ctx, instr.get())) {
op = instr->operands[2]; op = instr->operands[2];
} else if (sop2_can_use_sopk(ctx, instr.get())) {
op = instr->operands[instr->operands[0].isLiteral()];
} else { } else {
continue; continue;
} }
@ -2794,27 +2822,11 @@ void
optimize_encoding_sopk(ra_ctx& ctx, RegisterFile& register_file, aco_ptr<Instruction>& instr) optimize_encoding_sopk(ra_ctx& ctx, RegisterFile& register_file, aco_ptr<Instruction>& instr)
{ {
/* try to optimize sop2 with literal source to sopk */ /* try to optimize sop2 with literal source to sopk */
if (instr->opcode != aco_opcode::s_add_i32 && instr->opcode != aco_opcode::s_mul_i32 && if (!sop2_can_use_sopk(ctx, instr.get()))
instr->opcode != aco_opcode::s_cselect_b32)
return; return;
unsigned literal_idx = instr->operands[1].isLiteral();
uint32_t literal_idx = 0; if (instr->operands[!literal_idx].physReg() >= 128)
if (instr->opcode != aco_opcode::s_cselect_b32 && instr->operands[1].isLiteral())
literal_idx = 1;
if (!instr->operands[!literal_idx].isTemp() ||
!instr->operands[!literal_idx].isKillBeforeDef() ||
instr->operands[!literal_idx].getTemp().type() != RegType::sgpr ||
instr->operands[!literal_idx].physReg() >= 128)
return;
if (!instr->operands[literal_idx].isLiteral())
return;
const uint32_t i16_mask = 0xffff8000u;
uint32_t value = instr->operands[literal_idx].constantValue();
if ((value & i16_mask) && (value & i16_mask) != i16_mask)
return; return;
unsigned def_id = instr->definitions[0].tempId(); unsigned def_id = instr->definitions[0].tempId();
@ -2826,19 +2838,17 @@ optimize_encoding_sopk(ra_ctx& ctx, RegisterFile& register_file, aco_ptr<Instruc
} }
instr->format = Format::SOPK; instr->format = Format::SOPK;
SALU_instruction* instr_sopk = &instr->salu(); instr->salu().imm = instr->operands[literal_idx].constantValue() & 0xffff;
instr_sopk->imm = instr_sopk->operands[literal_idx].constantValue() & 0xffff;
if (literal_idx == 0) if (literal_idx == 0)
std::swap(instr_sopk->operands[0], instr_sopk->operands[1]); std::swap(instr->operands[0], instr->operands[1]);
if (instr_sopk->operands.size() > 2) if (instr->operands.size() > 2)
std::swap(instr_sopk->operands[1], instr_sopk->operands[2]); std::swap(instr->operands[1], instr->operands[2]);
instr_sopk->operands.pop_back(); instr->operands.pop_back();
switch (instr_sopk->opcode) { switch (instr->opcode) {
case aco_opcode::s_add_i32: instr_sopk->opcode = aco_opcode::s_addk_i32; break; case aco_opcode::s_add_i32: instr->opcode = aco_opcode::s_addk_i32; break;
case aco_opcode::s_mul_i32: instr_sopk->opcode = aco_opcode::s_mulk_i32; break; case aco_opcode::s_mul_i32: instr->opcode = aco_opcode::s_mulk_i32; break;
case aco_opcode::s_cselect_b32: instr_sopk->opcode = aco_opcode::s_cmovk_i32; break; case aco_opcode::s_cselect_b32: instr->opcode = aco_opcode::s_cmovk_i32; break;
default: unreachable("illegal instruction"); default: unreachable("illegal instruction");
} }
} }