raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-22 11:56:58 +02:00
Code Issues Packages Projects Releases Wiki Activity

shader: prove guarded descriptor counts for interleaved image records

Browse Source
This commit is contained in:
nmzik
2026-10-13 00:07:06 +02:00
parent 56738aa4e1
commit a287a54f26
3 changed files with 261 additions and 54 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/shader/recompiler/ir/passes/ResourceTracking.cpp
+50 -29
View File
@@ -1030,34 +1030,37 @@ private:
LaneQuantifier lanes = LaneQuantifier::All;
};
bool NonzeroOnEntry(Value value, const Block* block) const {
// Each unique predecessor must execute before this use. Stop at joins: an
// unrelated comparison is not a bound on FindILsb's zero-input sentinel.
bool HasEntryGuard(const Block* block, const auto& accepts, const Block* definition = nullptr) const {
// Each unique predecessor must execute before this use. Stop at joins and
// definitions so a guard from an earlier iteration cannot bound a new value.
for (size_t depth = 0; block != nullptr && depth < m_program.blocks.size(); ++depth) {
if (block->ImmPredecessors().size() != 1u) {
if (block == definition || block->ImmPredecessors().size() != 1u) {
return false;
}
const auto* previous = block->ImmPredecessors()[0];
const auto edge = ConditionalEdge(previous, block);
if (edge && edge->lanes == LaneQuantifier::All) {
const auto* test = edge->condition.TryInstruction();
if (test != nullptr && test->NumArgs() == 2u &&
((test->GetOpcode() == ValueOpcode::INotEqual32 && edge->positive) ||
(test->GetOpcode() == ValueOpcode::IEqual32 && !edge->positive))) {
for (uint32_t arg = 0; arg < 2u; ++arg) {
uint32_t immediate;
if (ImmediateU32(test->Arg(arg), immediate) && immediate == 0u &&
EquivalentValue(m_program, test->Arg(arg ^ 1u), value)) {
return true;
}
}
}
}
if (edge && edge->lanes == LaneQuantifier::All && accepts(*edge)) return true;
block = previous;
}
return false;
}
bool NonzeroOnEntry(Value value, const Block* block) const {
return HasEntryGuard(block, [&](const EdgePredicate& edge) {
const auto* test = edge.condition.TryInstruction();
if (test != nullptr && test->NumArgs() == 2u &&
((test->GetOpcode() == ValueOpcode::INotEqual32 && edge.positive) ||
(test->GetOpcode() == ValueOpcode::IEqual32 && !edge.positive))) {
for (uint32_t arg = 0; arg < 2u; ++arg) {
uint32_t immediate;
if (ImmediateU32(test->Arg(arg), immediate) && immediate == 0u &&
EquivalentValue(m_program, test->Arg(arg ^ 1u), value)) return true;
}
}
return false;
});
}
uint32_t WorkgroupAxis(Value key) const {
const auto* builtin = key.Resolve().TryInstruction();
uint32_t axis = UINT32_MAX;
@@ -1656,23 +1659,33 @@ private:
return false;
}
Value UnsignedBound(Value guard, Value value) const {
Value UnsignedBound(Value guard, Value value, bool positive = true) const {
const auto* test = SimplifyGuard(guard).TryInstruction();
if (test == nullptr) return {};
if (test->GetOpcode() == ValueOpcode::LogicalAnd) {
const auto op = test->GetOpcode();
if (op == ValueOpcode::LogicalNot)
return UnsignedBound(test->Arg(0), value, !positive);
if (op == ValueOpcode::LogicalAnd && positive) {
const auto left = UnsignedBound(test->Arg(0), value);
return left.IsEmpty() ? UnsignedBound(test->Arg(1), value) : left;
}
if (test->GetOpcode() == ValueOpcode::ULessThan32 &&
if ((positive ? op == ValueOpcode::ULessThan32 : op == ValueOpcode::UGreaterThanEqual32) &&
EquivalentValue(m_program, test->Arg(0), value)) return test->Arg(1);
if (test->GetOpcode() == ValueOpcode::LogicalNot) {
const auto* inverse = test->Arg(0).Resolve().TryInstruction();
if (inverse != nullptr && inverse->GetOpcode() == ValueOpcode::ULessThanEqual32 &&
EquivalentValue(m_program, inverse->Arg(1), value)) return inverse->Arg(0);
}
if (!positive && op == ValueOpcode::ULessThanEqual32 &&
EquivalentValue(m_program, test->Arg(1), value)) return test->Arg(0);
return {};
}
Value UnsignedBoundOnEntry(Value key, const Block* block) const {
const auto* definition = key.Resolve().TryInstruction();
Value bound;
const bool found = HasEntryGuard(block, [&](const EdgePredicate& edge) {
bound = UnsignedBound(edge.condition, key, edge.positive);
return !bound.IsEmpty() && ValidateRuntimeValue(m_program, bound, RuntimeValueType::Integer);
}, definition != nullptr ? definition->Parent() : nullptr);
return found ? bound : Value {};
}
Value SharedLoadBound(Value value, Value active) {
const auto* read = ActiveValue(value, active).TryInstruction();
if (read == nullptr || read->GetOpcode() != ValueOpcode::LoadSharedU32 ||
@@ -2406,12 +2419,20 @@ private:
// the GPU key using the same plan as other indirect image reads.
plan.retain_reads |= table_source.dword_count == 4u;
} else if (table_source.dword_count == 2u) {
if (table_stride != 32u) return false;
const auto guarded_count = UnsignedBoundOnEntry(key, handle.Parent());
const auto* selector = key.Resolve().TryInstruction();
const bool bitscan = selector != nullptr && selector->GetOpcode() == ValueOpcode::FindILsb32 &&
const bool bitscan = guarded_count.IsEmpty() && selector != nullptr &&
selector->GetOpcode() == ValueOpcode::FindILsb32 &&
selector->NumArgs() == 1u && !m_shader_writes &&
NonzeroOnEntry(selector->Arg(0), handle.Parent());
if (bitscan) {
if (!guarded_count.IsEmpty()) {
if (table_stride < 32u || (table_stride & 3u) != 0u) return false;
// Materialization captures the count and descriptor reads; renderer alias
// checks reject writes to those ranges while unrelated stores remain valid.
indirect.key_count = guarded_count;
} else if (table_stride != 32u) {
return false;
} else if (bitscan) {
indirect.key_count = Value(32u);
} else if (!m_shader_writes) {
if (const auto* bound = BoundedLoop(key, handle.Parent())) {