AlexIndustrial/mesa
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

aco/optimizer: move med3 -> add_clamp opt later

Browse Source 
Soon we will apply omod later,
when the combine_instruction reaches the multiplication with constant.

Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38658>
This commit is contained in:
Georg Lehmann
2025-11-25 14:24:31 +01:00
committed by Marge Bot
parent 39a61502e5
commit fc29821d3b
1 changed files with 19 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/amd/compiler/aco_optimizer.cpp
+19 -19
View File
@@ -2015,7 +2015,7 @@ bool
detect_clamp(Instruction* instr, unsigned* clamped_idx)
{
VALU_instruction& valu = instr->valu();
if (valu.omod != 0 || valu.opsel != 0)
if (valu.omod != 0 || valu.opsel != 0 || instr->isDPP())
return false;
unsigned idx = 0;
@@ -4585,24 +4585,6 @@ combine_instruction(opt_ctx& ctx, aco_ptr<Instruction>& instr)
ctx.info[instr->definitions[0].tempId()].set_combined(ctx.pre_combine_instrs.size() - 1);
}
#undef add_opt
if (instr->opcode == aco_opcode::v_med3_f32 || instr->opcode == aco_opcode::v_med3_f16) {
/* Optimize v_med3 to v_add so that it can be dual issued on GFX11. We start with v_med3 in
* case omod can be applied.
*/
unsigned idx;
if (detect_clamp(instr.get(), &idx)) {
instr->format = asVOP3(Format::VOP2);
instr->operands[0] = instr->operands[idx];
instr->operands[1] = Operand::zero();
instr->opcode =
instr->opcode == aco_opcode::v_med3_f32 ? aco_opcode::v_add_f32 : aco_opcode::v_add_f16;
instr->valu().clamp = true;
instr->valu().abs = (uint8_t)instr->valu().abs[idx];
instr->valu().neg = (uint8_t)instr->valu().neg[idx];
instr->operands.pop_back();
}
}
}
struct remat_entry {
@@ -4772,6 +4754,24 @@ select_instruction(opt_ctx& ctx, aco_ptr<Instruction>& instr)
return;
}
if (instr->opcode == aco_opcode::v_med3_f32 || instr->opcode == aco_opcode::v_med3_f16) {
/* Optimize v_med3 to v_add so that it can be dual issued on GFX11. We start with v_med3 in
* case omod can be applied.
*/
unsigned idx;
if (detect_clamp(instr.get(), &idx)) {
instr->format = asVOP3(Format::VOP2);
instr->operands[0] = instr->operands[idx];
instr->operands[1] = Operand::zero();
instr->opcode =
instr->opcode == aco_opcode::v_med3_f32 ? aco_opcode::v_add_f32 : aco_opcode::v_add_f16;
instr->valu().clamp = true;
instr->valu().abs = (uint8_t)instr->valu().abs[idx];
instr->valu().neg = (uint8_t)instr->valu().neg[idx];
instr->operands.pop_back();
}
}
/* convert split_vector into a copy or extract_vector if only one definition is ever used */
if (instr->opcode == aco_opcode::p_split_vector) {
unsigned num_used = 0;