summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorRhys Perry <pendingchaos02@gmail.com>2019-12-17 14:55:24 +0000
committerMarge Bot <eric+marge@anholt.net>2020-09-03 19:06:52 +0000
commitdc79152be975f69235a20a268b68e54a68562649 (patch)
treeec2db73f33a18373ca3c003a4662864fc9f95c12
parentfdadbdb513ebb71f9bb4e2784acc24148ca0defb (diff)
aco: fix mad splitting after applying output modifiers
Previously, this wasn't done because the mad label wasn't passed to the new definition. fossil-db (Navi): Totals from 5770 (4.24% of 135946) affected shaders: SGPRs: 391920 -> 391872 (-0.01%) VGPRs: 349084 -> 348424 (-0.19%); split: -0.20%, +0.01% CodeSize: 34639636 -> 34637496 (-0.01%); split: -0.02%, +0.01% MaxWaves: 58828 -> 58862 (+0.06%) Instrs: 6723436 -> 6723297 (-0.00%); split: -0.02%, +0.02% Cycles: 197594168 -> 197591968 (-0.00%); split: -0.02%, +0.02% Signed-off-by: Rhys Perry <pendingchaos02@gmail.com> Reviewed-by: Daniel Schürmann <daniel@schuermann.dev> Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6293>
-rw-r--r--src/amd/compiler/aco_optimizer.cpp34
1 files changed, 24 insertions, 10 deletions
diff --git a/src/amd/compiler/aco_optimizer.cpp b/src/amd/compiler/aco_optimizer.cpp
index 0a43964457a..a7cc0cdf6c8 100644
--- a/src/amd/compiler/aco_optimizer.cpp
+++ b/src/amd/compiler/aco_optimizer.cpp
@@ -2468,22 +2468,32 @@ bool apply_omod_clamp(opt_ctx &ctx, Block& block, aco_ptr<Instruction>& instr)
bool op1 = instr->operands[1].isTemp() && ctx.info[instr->operands[1].tempId()].is_omod_success();
if (op0 || op1) {
unsigned idx = op0 ? 0 : 1;
+ Instruction* omod_instr = ctx.info[instr->operands[idx].tempId()].instr;
+
/* omod was successfully applied */
- /* if the omod instruction is v_mad, we also have to change the original add */
- if (ctx.info[instr->operands[idx].tempId()].is_mad()) {
- Instruction* add_instr = ctx.mad_infos[ctx.info[instr->operands[idx].tempId()].instr->pass_flags].add_instr.get();
- if (ctx.info[instr->definitions[0].tempId()].is_clamp())
- static_cast<VOP3A_instruction*>(add_instr)->clamp = true;
- add_instr->definitions[0] = instr->definitions[0];
- }
- Instruction* omod_instr = ctx.info[instr->operands[idx].tempId()].instr;
/* check if we have an additional clamp modifier */
if (ctx.info[instr->definitions[0].tempId()].is_clamp() && ctx.uses[instr->definitions[0].tempId()] == 1 &&
ctx.uses[ctx.info[instr->definitions[0].tempId()].temp.id()]) {
+ /* if the omod instruction is v_mad, we also have to change the original add */
+ if (ctx.info[instr->operands[idx].tempId()].is_mad()) {
+ uint32_t mad_info_idx = ctx.info[instr->operands[idx].tempId()].instr->pass_flags;
+ Instruction* add_instr = ctx.mad_infos[mad_info_idx].add_instr.get();
+ static_cast<VOP3A_instruction*>(add_instr)->clamp = true;
+ }
+
static_cast<VOP3A_instruction*>(omod_instr)->clamp = true;
ctx.info[instr->definitions[0].tempId()].set_clamp_success(omod_instr);
}
+
+ /* if the omod instruction is v_mad, we also have to change the original add */
+ if (ctx.info[instr->operands[idx].tempId()].is_mad()) {
+ uint32_t mad_info_idx = ctx.info[instr->operands[idx].tempId()].instr->pass_flags;
+ Instruction* add_instr = ctx.mad_infos[mad_info_idx].add_instr.get();
+ add_instr->definitions[0] = instr->definitions[0];
+ ctx.info[instr->definitions[0].tempId()].set_mad(omod_instr, mad_info_idx);
+ }
+
/* change definition ssa-id of modified instruction */
omod_instr->definitions[0] = instr->definitions[0];
@@ -2514,13 +2524,17 @@ bool apply_omod_clamp(opt_ctx &ctx, Block& block, aco_ptr<Instruction>& instr)
}
if (found_zero && found_one && instr->operands[idx].isTemp() &&
ctx.info[instr->operands[idx].tempId()].is_clamp_success()) {
+ Instruction* clamp_instr = ctx.info[instr->operands[idx].tempId()].instr;
+
/* clamp was successfully applied */
/* if the clamp instruction is v_mad, we also have to change the original add */
if (ctx.info[instr->operands[idx].tempId()].is_mad()) {
- Instruction* add_instr = ctx.mad_infos[ctx.info[instr->operands[idx].tempId()].instr->pass_flags].add_instr.get();
+ uint32_t mad_info_idx = ctx.info[instr->operands[idx].tempId()].instr->pass_flags;
+ Instruction* add_instr = ctx.mad_infos[mad_info_idx].add_instr.get();
add_instr->definitions[0] = instr->definitions[0];
+
+ ctx.info[instr->definitions[0].tempId()].set_mad(clamp_instr, mad_info_idx);
}
- Instruction* clamp_instr = ctx.info[instr->operands[idx].tempId()].instr;
/* change definition ssa-id of modified instruction */
clamp_instr->definitions[0] = instr->definitions[0];