gremling/KytyPS5
1
0 Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-28 00:12:40 +00:00
Code Issues Packages Projects Releases Wiki Activity

shader: bound indexed material image selectors before heap expansion

Browse Source
This commit is contained in:
nmzik committed 2026-10-19 02:34:25 +02:00
1 parent ac3711ce7c
commit de81300b28
4 files changed
+397 -47
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

src/graphics/shader/recompiler/ir/passes/ResourceTracking.cpp
+54 -21
View File
@@ -1558,14 +1558,49 @@ private:
return read;
}
const Inst* IndexedMaterialRead(const Inst* read, Value active,
DescriptorSource::IndirectDescriptor::SelectorRead& selector) const {
uint32_t lane = 0;
if (read != nullptr && read->GetOpcode() == ValueOpcode::CompositeExtractU32x2) {
if (!ImmediateU32(read->Arg(1), lane) || lane >= 2u) return nullptr;
read = read->Arg(0).Resolve().TryInstruction();
if (read == nullptr || read->GetOpcode() != ValueOpcode::LoadBufferU32x2) return nullptr;
} else if (read == nullptr || read->GetOpcode() != ValueOpcode::LoadBufferU32) {
return nullptr;
}
const auto& memory = m_program.memory_info[read->Flags<MemoryFlags>().index];
if (memory.kind != ResourceKind::Buffer || memory.typed || memory.formatted ||
!memory.idxen || memory.offen || memory.offset > UINT32_MAX - lane * 4u ||
read->Arg(2).Resolve() != Value(0u) || read->Arg(3).Resolve() != Value(0u) ||
!EquivalentValue(m_program, active, read->Arg(4))) return nullptr;
selector.offset = memory.offset + lane * 4u;
return read;
}
bool MatchUniformizedMaterialKey(Value key, const Inst& image,
DescriptorSource::IndirectDescriptor& indirect,
DescriptorSource& material_source) {
DescriptorSource& material_source, bool indexed) {
const auto* selected = UniformizedMaterialValue(key, image);
if (selected == nullptr) return false;
const auto active = selected->Arg(0).Resolve();
DescriptorSource::IndirectDescriptor::SelectorRead selector;
const auto* read = SelectorReadValue(selected->Arg(1), active, selector);
if (indexed) {
read = IndexedMaterialRead(read, active, selector);
// This optional domain is checked against the dispatch's writable buffers.
// Other stages and unbounded/image writes need the original heap domain.
if (read == nullptr || m_program.stage != ShaderType::Compute ||
std::ranges::any_of(m_program.blocks, [](const Block* block) {
return std::ranges::any_of(*block, [](const Inst& inst) {
const auto op = inst.GetOpcode();
return ImageOpcodeInfoOf(op).resource_class == ImageResourceClass::Storage ||
AddressOpcodeInfoOf(op).access == AddressAccess::Write;
});
}) || !MakeRuntimeTableSource(*read, material_source)) return false;
selector.source = InternSource(material_source);
indirect.selector = std::move(selector);
return true;
}
if (read == nullptr || read->GetOpcode() != ValueOpcode::LoadAddressU32 ||
read->NumArgs() != 4u || !Implies(active, read->Arg(3)))
return false;
@@ -2971,6 +3006,13 @@ private:
// Keep live descriptor payloads, but project descriptor-only SMEM onto
// the GPU key using the same plan as other indirect image reads.
plan.retain_reads |= table_source.dword_count == 4u;
} else if (!sampler && !plan.table_indexed && table_source.dword_count == 4u &&
MatchUniformizedMaterialKey(key, handle, indirect, material_source, true)) {
// Preserve the heap's byte-offset key so runtime geometry can choose the
// tighter selector domain without changing the GPU's descriptor reads.
indirect.table_scalar = true;
plan.table_read = plan.reads[0];
plan.retain_reads = true;
} else if (sampler && table_source.dword_count == 4u && !plan.table_indexed &&
SetExactTableMaximum(key, indirect, &handle)) {
indirect.table_scalar = true;
@@ -3020,7 +3062,7 @@ private:
}
}
if (indirect.key_count.IsEmpty() &&
!MatchUniformizedMaterialKey(key, handle, indirect, material_source) &&
!MatchUniformizedMaterialKey(key, handle, indirect, material_source, false) &&
!MatchReversedImageRange(key, indirect)) {
const auto maximum = SelectorMaximum(key);
if (maximum >= MaxIndirectDescriptorProbes) return false;
@@ -3064,6 +3106,10 @@ private:
indirect.selector_mask = selected.selector_mask;
indirect.signed_count = selected.signed_count;
indirect.workgroup_axis = selected.workgroup_axis;
if (selected.table_scalar) {
indirect.table_scalar = true;
plan.table_read = plan.reads[0];
}
material_source = m_sources[selected.selector->source];
shared_selector = true;
break;
@@ -3264,26 +3310,13 @@ private:
if (value == nullptr || (selected != nullptr && selected != value)) return false;
selected = value;
}
const auto* read = selected->Arg(1).Resolve().TryInstruction();
uint32_t lane = 0;
if (read != nullptr && read->GetOpcode() == ValueOpcode::CompositeExtractU32x2) {
if (!ImmediateU32(read->Arg(1), lane) || lane >= 2u) return false;
read = read->Arg(0).Resolve().TryInstruction();
if (read == nullptr || read->GetOpcode() != ValueOpcode::LoadBufferU32x2) return false;
} else if (read == nullptr || read->GetOpcode() != ValueOpcode::LoadBufferU32) {
return false;
}
const auto& memory = m_program.memory_info[read->Flags<MemoryFlags>().index];
uint32_t offset = 1u, scalar = 1u;
if (memory.kind != ResourceKind::Buffer || memory.typed || memory.formatted ||
!memory.idxen || memory.offen || memory.offset > UINT32_MAX - lane * 4u ||
!ImmediateU32(read->Arg(2), offset) || offset != 0u ||
!ImmediateU32(read->Arg(3), scalar) || scalar != 0u ||
!EquivalentValue(m_program, selected->Arg(0), read->Arg(4))) return false;
if (!MakeRuntimeTableSource(*read, material_source) ||
DescriptorSource::IndirectDescriptor::SelectorRead selector;
const auto* read = IndexedMaterialRead(selected->Arg(1).Resolve().TryInstruction(),
selected->Arg(0), selector);
if (read == nullptr || !MakeRuntimeTableSource(*read, material_source) ||
!BoundedSelectedIndex(read->Arg(1), read->Arg(4), indirect)) return false;
indirect.selector.emplace(DescriptorSource::IndirectDescriptor::SelectorRead{.source = InternSource(material_source),
.offset = memory.offset + lane * 4u});
selector.source = InternSource(material_source);
indirect.selector = std::move(selector);
return finish();
}