raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-20 02:47:02 +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
src/graphics/shader/recompiler/ir/passes/ConstantPropagation.cpp
+11 -5
View File
@@ -706,14 +706,14 @@ void FoldInstruction(Block& block, Block::iterator instruction,
}
}
// In wave32, extracting this invocation's bit from a ballot recovers its predicate.
// Keep that identity through scalar EXEC operations and their loop-carried mask Phis.
// Full and empty masks have a constant predicate in either wave size. In wave32,
// ballot projection also survives scalar EXEC operations and loop-carried mask Phis.
class LaneMaskProjection {
public:
explicit LaneMaskProjection(std::unordered_set<Inst*>& lowered_ancillary)
: m_lowered_ancillary(lowered_ancillary) {}
void Fold(Inst& inst) {
void Fold(Inst& inst, uint32_t wave_size) {
if (inst.GetOpcode() != ValueOpcode::INotEqual32 || !Immediate(Arg(inst, 1), 0u)) return;
const auto* bit = Arg(inst, 0).TryInstruction();
if (bit == nullptr || bit->GetOpcode() != ValueOpcode::BitwiseAnd32 ||
@@ -725,9 +725,15 @@ public:
!Immediate(Arg(*index, 1), 31u)) return;
const auto* lane = Arg(*index, 0).TryInstruction();
if (lane == nullptr || lane->GetOpcode() != ValueOpcode::LaneId) return;
const auto mask = Arg(*shift, 0);
if (Immediate(mask, 0u) || Immediate(mask, UINT32_MAX)) {
Replace(inst, Value(mask.U32() != 0u));
return;
}
if (wave_size != 32u) return;
m_visited.clear();
m_grounded = false;
if (CanProject(Arg(*shift, 0)) && m_grounded) Replace(inst, Project(Arg(*shift, 0)));
if (CanProject(mask) && m_grounded) Replace(inst, Project(mask));
}
private:
@@ -809,7 +815,7 @@ void ConstantPropagationPass(const BlockList& blocks, uint32_t wave_size) {
LaneMaskProjection mask_projection(lowered_ancillary);
for (auto* block: blocks) {
for (auto inst = block->begin(); inst != block->end(); ++inst) {
if (wave_size == 32u) mask_projection.Fold(*inst);
mask_projection.Fold(*inst, wave_size);
FoldInstruction(*block, inst, lowered_ancillary);
}
}