shader: lower wave32 whole-wave UMin and UMax reductions to subgroup reductions

Co-authored-by: Senaxx <michelbeld@gmail.com>
Co-authored-by: Mac <me@itsmemac.com>
This commit is contained in:
nmzik
2026-10-26 01:12:50 +01:00
parent be1c84f4d7
commit 6357740fa1
7 changed files with 102 additions and 60 deletions
+30 -18
View File
@@ -628,17 +628,18 @@ void TestReadLaneElimination() {
"dynamic-lane read was rewritten unsafely");
}
void TestCompleteUMinReduction() {
void TestCompleteWaveReduction() {
enum class Case {
Masked, Subset, Literal, LiteralMasked, LiteralMove, Wave32,
UnrelatedSeed, UnmaskedSymbolic, WrongNeutral, RowMask, BankMask,
Masked, Subset, Literal, LiteralMasked, LiteralMove, Wave32, Wave32Rows,
Wave32Maximum, UnrelatedSeed, UnmaskedSymbolic, WrongNeutral, RowMask, BankMask,
Shift, FetchInactive, BoundControl, Dpp8, MoveExec, UpdateExec,
Previous, DppSource, CrossExec, CrossRow, CrossSelectorLow,
CrossSelectorHigh, CrossFetch, CrossBound, CrossSource, FinalExec,
Endpoint, EndpointSource, DifferentWitness, CrossBlock, WrongOperation
};
for (const auto test : {Case::Masked, Case::Subset, Case::Literal,
Case::LiteralMasked, Case::LiteralMove, Case::Wave32, Case::UnrelatedSeed,
Case::LiteralMasked, Case::LiteralMove, Case::Wave32, Case::Wave32Rows,
Case::Wave32Maximum, Case::UnrelatedSeed,
Case::UnmaskedSymbolic, Case::WrongNeutral, Case::RowMask, Case::BankMask,
Case::Shift, Case::FetchInactive, Case::BoundControl, Case::Dpp8,
Case::MoveExec, Case::UpdateExec, Case::Previous, Case::DppSource,
@@ -657,13 +658,15 @@ void TestCompleteUMinReduction() {
auto seed_mask = test == Case::UnrelatedSeed ? other_mask : mask;
if (test == Case::Subset)
seed_mask = fixture.Emit(ValueOpcode::LogicalAnd, {other_mask, mask});
// A UMax scan seeds lanes outside EXEC with zero.
const bool maximum = test == Case::Wave32Maximum;
const uint32_t neutral = test == Case::WrongNeutral || maximum ? 0u : UINT32_MAX;
const auto seed = test == Case::Literal || test == Case::UnmaskedSymbolic ? source :
fixture.Emit(ValueOpcode::SelectU32,
{seed_mask, source, Value(test == Case::WrongNeutral ? 0u : UINT32_MAX)});
fixture.Emit(ValueOpcode::SelectU32, {seed_mask, source, Value(neutral)});
auto prefix = fixture.Emit(ValueOpcode::SelectU32, {enabled, seed, source});
Value preserved;
const auto operation = test == Case::WrongOperation ? ValueOpcode::BitwiseOr32 :
ValueOpcode::UMin32;
auto operation = maximum ? ValueOpcode::UMax32 : ValueOpcode::UMin32;
if (test == Case::WrongOperation) operation = ValueOpcode::BitwiseOr32;
for (const uint32_t shift : {1u, 2u, 4u, 8u}) {
DppMoveFlags flags{.control = 0x110u + shift};
if (shift == 2) {
@@ -704,10 +707,16 @@ void TestCompleteUMinReduction() {
const auto final = fixture.Emit(ValueOpcode::SelectU32,
{test == Case::FinalExec ? mask : enabled, combined, prefix});
const uint32_t result_block = test == Case::CrossBlock ? 1u : 0u;
// Wave32 reads each half's row endpoint from the scan itself.
const bool rows = test == Case::Wave32Rows || maximum;
const auto endpoints = rows ? prefix : final;
uint32_t low_lane = test == Case::Endpoint ? 30u : 31u;
if (rows) low_lane = 15u;
const auto low = fixture.Emit(ValueOpcode::ReadLane,
{final, Value(test == Case::Endpoint ? 30u : 31u)}, 0, result_block);
{endpoints, Value(low_lane)}, 0, result_block);
const auto high = fixture.Emit(ValueOpcode::ReadLane,
{test == Case::EndpointSource ? prefix : final, Value(63u)}, 0, result_block);
{test == Case::EndpointSource ? prefix : endpoints, Value(rows ? 31u : 63u)}, 0,
result_block);
const auto result = fixture.Emit(operation, {high, low}, 0, result_block);
const auto use = fixture.Emit(ValueOpcode::ReferenceU32, {result}, 0, result_block);
using Terminator = Libs::Graphics::ShaderRecompiler::CFG::TerminatorKind;
@@ -718,16 +727,19 @@ void TestCompleteUMinReduction() {
fixture.BlockAt().terminator.true_block = &fixture.BlockAt(1);
}
ValidateProgram(fixture.program, true);
const auto stats = EliminateReadLane(fixture.program, test == Case::Wave32 ? 32u : 64u);
const auto stats =
EliminateReadLane(fixture.program, test == Case::Wave32 || rows ? 32u : 64u);
const bool matches = test == Case::Masked || test == Case::Subset ||
test == Case::Literal || test == Case::LiteralMasked || test == Case::LiteralMove;
test == Case::Literal || test == Case::LiteralMasked || test == Case::LiteralMove ||
rows;
Check(stats.rewritten_reductions == (matches ? 1u : 0u) && stats.rewritten_reads == 0,
("UMin reduction proof failed case " + std::to_string(static_cast<unsigned>(test))).c_str());
("wave reduction proof failed case " + std::to_string(static_cast<unsigned>(test))).c_str());
if (matches) {
const auto *reduction = use.Instruction()->Arg(0).ResolveInstruction();
Check(reduction->GetOpcode() == ValueOpcode::SubgroupUMinU32 &&
Check(reduction->GetOpcode() ==
(maximum ? ValueOpcode::SubgroupUMaxU32 : ValueOpcode::SubgroupUMinU32) &&
reduction->Arg(0).Resolve() == seed.Resolve(),
"UMin reduction changed the neutral-masked input");
"wave reduction changed the neutral-masked input");
RemoveIdentities(fixture.program);
EliminateDeadCode(fixture.program.blocks);
uint32_t updates = 0;
@@ -735,10 +747,10 @@ void TestCompleteUMinReduction() {
updates += inst.GetOpcode() == ValueOpcode::DppUpdateU32;
Check(inst.GetOpcode() != ValueOpcode::ReadLane &&
inst.GetOpcode() != ValueOpcode::Permlane16U32,
"UMin reduction retained unused endpoint or cross-row operations");
"wave reduction retained unused endpoint or cross-row operations");
}
Check(updates == 2 && preserved.Instruction()->Arg(0).Instruction()->HasUses(),
"UMin reduction destroyed an independently consumed prefix or kept its dead suffix");
"wave reduction destroyed an independently consumed prefix or kept its dead suffix");
ValidateProgram(fixture.program, true);
} else {
Check(use.Instruction()->Arg(0).Resolve() == result,
@@ -884,7 +896,7 @@ int main() {
TestSharedIntegerRuntimeDependencies();
TestConstantBufferBounds();
TestReadLaneElimination();
TestCompleteUMinReduction();
TestCompleteWaveReduction();
TestOptimizationPipeline();
TestControlFlowValueSurvivesReadLaneFolding();
TestDeadPhiCyclesAndPlanningRoots();