fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2025-12-21 18:00:13 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

aco: optimize boolean phis with uniform selections

Browse source 
Even though the boolean can be divergent, the control flow can be (at
least partially) uniform. For example, we don't have to create any
s_andn2_b64/s_and_b64/s_or_b64 instructions with this code:
a = ...
loop {
    b = bool_phi a, c
    if (uniform)
        break
    c = ...
}
d = phi c

fossil-db (Navi):
Totals from 5506 (4.05% of 135946) affected shaders:
SGPRs: 605720 -> 604024 (-0.28%)
SpillSGPRs: 52025 -> 51733 (-0.56%)
CodeSize: 65221188 -> 64957808 (-0.40%); split: -0.41%, +0.00%
Instrs: 12637881 -> 12584610 (-0.42%); split: -0.42%, +0.00%

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/3388>
This commit is contained in:
Rhys Perry 2020-01-09 16:51:34 +00:00 • committed by Marge Bot
parent f622e80494
commit 9a089baff1
1 changed files with 27 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

28
src/amd/compiler/aco_lower_phis.cpp
View file

@ -35,6 +35,9 @@
namespace aco { namespace aco {
struct ssa_state { struct ssa_state {
bool checked_preds_for_uniform;
bool all_preds_uniform;
bool needs_init; bool needs_init;
uint64_t cur_undef_operands; uint64_t cur_undef_operands;
@ -152,6 +155,19 @@ void lower_divergent_bool_phi(Program *program, ssa_state *state, Block *block,
{ {
Builder bld(program); Builder bld(program);
if (!state->checked_preds_for_uniform) {
state->all_preds_uniform = !(block->kind & block_kind_merge);
for (unsigned pred : block->logical_preds)
state->all_preds_uniform = state->all_preds_uniform && (program->blocks[pred].kind & block_kind_uniform);
state->checked_preds_for_uniform = true;
}
if (state->all_preds_uniform) {
assert(block->logical_preds.size() == block->linear_preds.size());
phi->opcode = aco_opcode::p_linear_phi;
return;
}
state->latest.resize(program->blocks.size()); state->latest.resize(program->blocks.size());
uint64_t undef_operands = 0; uint64_t undef_operands = 0;
@ -180,14 +196,23 @@ void lower_divergent_bool_phi(Program *program, ssa_state *state, Block *block,
state->writes[block->logical_preds[i]] = program->allocateId(); state->writes[block->logical_preds[i]] = program->allocateId();
} }
bool uniform_merge = block->kind & block_kind_loop_header;
for (unsigned i = 0; i < phi->operands.size(); i++) { for (unsigned i = 0; i < phi->operands.size(); i++) {
Block *pred = &program->blocks[block->logical_preds[i]]; Block *pred = &program->blocks[block->logical_preds[i]];
bool need_get_ssa = !uniform_merge;
if (block->kind & block_kind_loop_header && !(pred->kind & block_kind_uniform))
uniform_merge = false;
if (phi->operands[i].isUndefined()) if (phi->operands[i].isUndefined())
continue; continue;
Operand cur = get_ssa(program, pred->index, state, true); Operand cur(bld.lm);
if (need_get_ssa)
cur = get_ssa(program, pred->index, state, true);
assert(cur.regClass() == bld.lm); assert(cur.regClass() == bld.lm);
Temp new_cur = {state->writes.at(pred->index), program->lane_mask}; Temp new_cur = {state->writes.at(pred->index), program->lane_mask};
assert(new_cur.regClass() == bld.lm); assert(new_cur.regClass() == bld.lm);
@ -241,6 +266,7 @@ void lower_phis(Program* program)
ssa_state state; ssa_state state;
for (Block& block : program->blocks) { for (Block& block : program->blocks) {
state.checked_preds_for_uniform = false;
state.needs_init = true; state.needs_init = true;
for (aco_ptr<Instruction>& phi : block.instructions) { for (aco_ptr<Instruction>& phi : block.instructions) {
if (phi->opcode == aco_opcode::p_phi) { if (phi->opcode == aco_opcode::p_phi) {