raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-19 18:36:53 +02:00
Code Issues Packages Projects Releases Wiki Activity

shader: fold full wave64 masks before register predication

Browse Source
This commit is contained in:
nmzik
2026-10-12 17:20:42 +02:00
parent fd66f580b4
commit fa68e104ea
2 changed files with 36 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/shaderCfgTests.cpp
+25 -2
View File
@@ -12390,7 +12390,7 @@ void TestValuePhiValidation() {
#endif
}
void TestWave32MaskProjection() {
void TestLaneMaskProjection() {
using namespace ShaderRecompiler::IR;
for (const uint32_t wave_size : {32u, 64u}) {
Program program;
@@ -12429,6 +12429,19 @@ void TestWave32MaskProjection() {
: U32(Value(3u));
return ir.INotEqual(ir.BitwiseAnd(ir.ShiftRightLogical(word, bit), U32(Value(1u))), zero);
};
const auto low_half = e.ULessThan(U32(e.Emit(ValueOpcode::LaneId)), U32(Value(32u)));
const auto mask_bit = [&](uint32_t low, uint32_t high) {
const auto word = wave_size == 64u ? e.Select(low_half, U32(Value(low)), U32(Value(high)))
: U32(Value(low));
return thread_bit(e, word);
};
const auto full = mask_bit(UINT32_MAX, UINT32_MAX);
const auto empty = mask_bit(0u, 0u);
const auto low_only = mask_bit(UINT32_MAX, 0u);
const auto high_only = mask_bit(0u, UINT32_MAX);
const auto partial = mask_bit(0x80000000u, 0x80000000u);
const auto full_write = e.Select(full, U32(Value(42u)), raw);
const auto empty_write = e.Select(empty, U32(Value(42u)), raw);
const auto active_mask = ballot(e, active);
const auto direct = thread_bit(e, active_mask);
const auto arbitrary = thread_bit(e, raw);
@@ -12471,7 +12484,14 @@ void TestWave32MaskProjection() {
program.blocks[1]->condition = remaining_predicate;
ConstantPropagationPass(program.blocks, wave_size);
Check(full.Resolve() == Value(true) && empty.Resolve() == Value(false) &&
full_write.Resolve() == Value(42u) && empty_write.Resolve() == raw.Resolve(),
"constant EXEC mask retained lane extraction or register predication");
Check(partial.Resolve().TryInstruction() != nullptr,
"partial mask was incorrectly folded to a constant predicate");
if (wave_size == 32u) {
Check(low_only.Resolve() == Value(true) && high_only.Resolve() == Value(false),
"wave32 constant mask incorrectly consumed the high half");
Check(direct.Resolve() == active.Resolve(), "wave32 ballot round trip lost its predicate");
Check(EquivalentValue(program, combined, combined_predicate),
"wave32 OR mask projection changed a lane predicate");
@@ -12491,6 +12511,9 @@ void TestWave32MaskProjection() {
projected->PhiBlock(1) == loop,
"mask projection lost loop Phi type or predecessor order");
} else {
Check(low_only.Resolve().TryInstruction() != nullptr &&
high_only.Resolve().TryInstruction() != nullptr,
"wave64 mixed halves were incorrectly folded to a constant predicate");
Check(direct.Resolve().TryInstruction()->GetOpcode() == ValueOpcode::INotEqual32 &&
live.Resolve().TryInstruction()->GetOpcode() == ValueOpcode::INotEqual32 &&
combined.Resolve().TryInstruction()->GetOpcode() == ValueOpcode::INotEqual32 &&
@@ -15698,7 +15721,7 @@ int main() {
TestFinalSsaRejectsRegisterStatePseudos();
#endif
TestValuePhiValidation();
TestWave32MaskProjection();
TestLaneMaskProjection();
TestU64ShiftConstantPropagation();
TestU64RegisterPairRoundTrip();
#if KYTY_PLATFORM != KYTY_PLATFORM_WINDOWS