raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-27 13:26:52 +01:00
Code Issues Packages Projects Releases Wiki Activity

shader: fold complementary masks so wave reductions see full EXEC

Browse Source 
Co-authored-by: Senaxx <michelbeld@gmail.com>
Co-authored-by: Mac <me@itsmemac.com>
This commit is contained in:
nmzik
2026-10-26 01:37:18 +01:00
parent 6357740fa1
commit ce5e57aa36
3 changed files with 66 additions and 92 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/shader/recompiler/ir/passes/ReadLaneElimination.cpp
+24 -58
View File
@@ -104,11 +104,10 @@ Value GetRealValue(PhiMap& phi_map, Value source, uint32_t lane, uint32_t wave_s
// Fragment subgroups may have no physical invocation at either endpoint.
class WaveReductionMatcher {
public:
WaveReductionMatcher(Block* block, ValueOpcode operation)
: block(block), operation(operation),
identity(operation == ValueOpcode::UMin32 ? UINT32_MAX : 0u) {}
WaveReductionMatcher(const Block* block, ValueOpcode operation)
: block(block), operation(operation) {}
Value Match(const Inst& reduction, uint32_t wave_size) {
Value Match(const Inst& reduction, uint32_t wave_size) const {
const auto* low = Get(reduction.Arg(0), ValueOpcode::ReadLane);
const auto* high = Get(reduction.Arg(1), ValueOpcode::ReadLane);
if (low == nullptr || high == nullptr || low->Arg(0).Resolve() != high->Arg(0).Resolve()) {
@@ -124,45 +123,32 @@ public:
(lane(low->Arg(1), wave_size - 1u) && lane(high->Arg(1), half - 1u)))) {
return {};
}
if (wave_size == 32u) return MatchScan(Unwrap(low->Arg(0)));
const auto* joined = Get(Unwrap(low->Arg(0)), operation);
if (wave_size == 32u) return MatchScan(low->Arg(0));
const auto* joined = Get(low->Arg(0), operation);
if (joined == nullptr) return {};
for (size_t index = 0; index < 2; index++) {
auto attempt = *this;
const auto* cross = attempt.Get(attempt.Unwrap(joined->Arg(index)), ValueOpcode::Permlane16U32);
const auto* cross = Get(joined->Arg(index), ValueOpcode::Permlane16U32);
if (cross == nullptr) continue;
const auto flags = cross->Flags<PermlaneFlags>();
auto prefix = attempt.Unwrap(joined->Arg(1 - index));
const auto flags = cross->Flags<PermlaneFlags>();
const auto prefix = joined->Arg(1 - index).Resolve();
if (!flags.x16 || flags.fetch_inactive || !flags.bound_control ||
cross->Arg(1).Resolve() != Value(UINT32_MAX) ||
cross->Arg(2).Resolve() != Value(UINT32_MAX) ||
!attempt.FullExec(cross->Arg(3)) || attempt.Unwrap(cross->Arg(0)) != prefix) {
cross->Arg(2).Resolve() != Value(UINT32_MAX) || !FullExec(cross->Arg(3)) ||
cross->Arg(0).Resolve() != prefix) {
continue;
}
if (const auto seed = attempt.MatchScan(prefix); !seed.IsEmpty()) return seed;
if (const auto seed = MatchScan(prefix); !seed.IsEmpty()) return seed;
}
return {};
}
private:
// The row scan's input, whose lanes outside the saved EXEC hold the operation's identity.
Value MatchScan(Value prefix) {
// The row scan's input.
Value MatchScan(Value prefix) const {
for (const uint32_t shift: {8u, 4u, 2u, 1u}) {
prefix = MatchRow(prefix, shift);
}
prefix = Unwrap(prefix);
if (prefix.IsEmpty() || saved_exec.IsEmpty()) return prefix;
const auto* seed = Get(prefix, ValueOpcode::SelectU32);
if (seed != nullptr && seed->Arg(2).Resolve() == Value(identity)) {
const auto predicate = seed->Arg(0).Resolve();
const auto* subset = Get(predicate, ValueOpcode::LogicalAnd);
if (predicate == saved_exec ||
(subset != nullptr && (subset->Arg(0).Resolve() == saved_exec ||
subset->Arg(1).Resolve() == saved_exec))) {
return prefix;
}
}
return {};
return prefix;
}
Inst* Get(Value value, ValueOpcode opcode) const {
@@ -171,28 +157,8 @@ private:
? inst : nullptr;
}
bool FullExec(Value predicate) {
predicate = predicate.Resolve();
if (predicate == Value(true)) return true;
const auto* either = Get(predicate, ValueOpcode::LogicalOr);
if (either == nullptr) return false;
for (size_t index = 0; index < 2; index++) {
const auto* negate = Get(either->Arg(index), ValueOpcode::LogicalNot);
if (negate == nullptr || negate->Arg(0).Resolve() != either->Arg(1 - index).Resolve()) continue;
const auto original = negate->Arg(0).Resolve();
if (saved_exec.IsEmpty()) saved_exec = original;
return saved_exec == original;
}
return false;
}
Value Unwrap(Value value) {
while (const auto* select = Get(value, ValueOpcode::SelectU32)) {
if (!FullExec(select->Arg(0))) break;
value = select->Arg(1);
}
return value.Resolve();
}
// Constant propagation folds the x | ~x EXEC of s_orn2_saveexec to true.
static bool FullExec(Value predicate) { return predicate.Resolve() == Value(true); }
static bool IsRowShift(const Inst& inst, uint32_t shift) {
const auto flags = inst.Flags<DppMoveFlags>();
@@ -200,24 +166,24 @@ private:
flags.bank_mask == 0xf && !flags.fetch_inactive && !flags.bound_control && !flags.dpp8;
}
Value MatchRow(Value value, uint32_t shift) {
const auto* update = Get(Unwrap(value), ValueOpcode::DppUpdateU32);
Value MatchRow(Value value, uint32_t shift) const {
const auto* update = Get(value, ValueOpcode::DppUpdateU32);
if (update == nullptr || !IsRowShift(*update, shift) || !FullExec(update->Arg(2))) return {};
const auto* combined = Get(update->Arg(0), operation);
if (combined == nullptr) return {};
for (size_t index = 0; index < 2; index++) {
const auto* move = Get(combined->Arg(index), ValueOpcode::DppMoveU32);
if (move == nullptr || !IsRowShift(*move, shift) || !FullExec(move->Arg(1))) continue;
const auto previous = Unwrap(combined->Arg(1 - index));
if (Unwrap(move->Arg(0)) == previous && Unwrap(update->Arg(1)) == previous) return previous;
const auto previous = combined->Arg(1 - index).Resolve();
if (move->Arg(0).Resolve() == previous && update->Arg(1).Resolve() == previous) {
return previous;
}
}
return {};
}
Block* block;
ValueOpcode operation;
uint32_t identity;
Value saved_exec;
const Block* block;
ValueOpcode operation;
};
} // namespace