raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-23 04:16:57 +02:00
Code Issues Packages Projects Releases Wiki Activity

shader: share bounded record descriptors for packed loads and byte stores

Browse Source 
Co-authored-by: fxpw <fxpw@outlook.com>
This commit is contained in:
nmzik
2026-10-14 05:07:24 +02:00
parent 4cf5a96b79
commit 26d326d1b0
12 changed files with 581 additions and 142 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/shader/recompiler/ir/passes/ResourceTracking.cpp
+80 -26
View File
@@ -394,6 +394,7 @@ public:
memory.sampler = patch.sampler;
}
}
std::vector<std::pair<const IndirectDescriptorPlan*, Value>> buffer_selections;
for (const auto& plan: m_indirect_descriptors) {
auto key = plan.key;
if (plan.table_read != nullptr) {
@@ -452,8 +453,32 @@ public:
plan.handle->SetArg(word, Value(0u));
continue;
}
if (plan.handle->NumArgs() == 4u) {
plan.handle->SetArg(0, plan.key);
if (plan.handle->GetOpcode() == ValueOpcode::GetBufferResource) {
Value selected;
for (const auto& [previous, value]: buffer_selections) {
if (plan.selection_anchor != nullptr && previous->reads == plan.reads &&
previous->source == plan.source &&
previous->handle->Flags<uint32_t>() == plan.handle->Flags<uint32_t>() &&
EquivalentValue(m_program, previous->key, plan.key) &&
AvailableBefore(*value.Instruction(), *plan.handle)) {
selected = value;
break;
}
}
if (selected.IsEmpty()) {
const auto* anchor = plan.selection_anchor != nullptr ? plan.selection_anchor : plan.handle;
auto* block = anchor->Parent();
const auto where = std::ranges::find_if(block->Instructions(),
[&](const Inst& inst) { return &inst == anchor; });
auto& selection = *block->PrependNewInst(where, ValueOpcode::GetBufferResourceIndex, {plan.key});
selection.SetFlags(plan.handle->Flags<uint32_t>());
selected = Value(&selection);
buffer_selections.emplace_back(&plan, selected);
}
plan.handle->SetArg(0, selected);
for (uint32_t word = 1; word < 4u; ++word) plan.handle->SetArg(word, Value(0u));
} else if (plan.handle->NumArgs() == 4u) {
plan.handle->SetArg(0, key);
for (uint32_t word = 1; word < 4u; ++word) plan.handle->SetArg(word, Value(0u));
} else {
if (plan.reads[0] == nullptr) plan.handle->SetArg(0, key);
@@ -519,6 +544,7 @@ private:
bool retain_reads = false;
bool table_indexed = false;
const Inst* table_read = nullptr;
const Inst* selection_anchor = nullptr;
};
[[noreturn]] void Fail(uint32_t pc, const std::string& reason) const {
@@ -931,7 +957,7 @@ private:
const auto& b = *descriptor.indirect_descriptor;
if (a.selector != b.selector || a.table_source != b.table_source ||
a.table_offset != b.table_offset || a.table_immediate != b.table_immediate ||
a.table_stride != b.table_stride || a.indexed_store != b.indexed_store ||
a.table_stride != b.table_stride || a.indexed != b.indexed ||
a.workgroup_axis != b.workgroup_axis || a.sources != b.sources ||
a.signed_count != b.signed_count ||
!EquivalentValue(m_program, a.key_count, b.key_count) ||
@@ -2522,6 +2548,25 @@ private:
return true;
}
bool IndirectBufferAccess(const Inst& inst) const {
const auto opcode = inst.GetOpcode();
if (opcode != ValueOpcode::StoreBufferU32 && opcode != ValueOpcode::StoreBufferU8 &&
opcode != ValueOpcode::LoadBufferU32x2) return false;
const auto index = inst.Flags<MemoryFlags>().index;
if (index >= m_program.memory_info.size()) return false;
const auto& memory = m_program.memory_info[index];
if (memory.kind != ResourceKind::Buffer || memory.typed || memory.idxen == memory.offen ||
(memory.idxen && (memory.offset != 0u || inst.Arg(2).Resolve() != Value(0u))) ||
inst.Arg(3).Resolve() != Value(0u)) return false;
if (opcode == ValueOpcode::StoreBufferU32)
return !memory.formatted || (memory.idxen && memory.data_bits == 32u && memory.data_dwords == 1u);
if (!memory.idxen) return false;
if (opcode == ValueOpcode::StoreBufferU8)
return !memory.formatted && memory.data_bits == 8u;
return memory.formatted && memory.data_bits == 16u && memory.data_dwords == 2u &&
memory.component_count == 4u;
}
bool MatchDescriptorTable(Inst& handle, const DescriptorSource& descriptor,
IndirectDescriptorPlan& plan,
DescriptorSource& table_source, Value& key, uint32_t& table_offset,
@@ -2599,10 +2644,10 @@ private:
if (handle.NumArgs() == 8u) {
if (memory->kind == ResourceKind::ScalarAddress)
plan.retain_reads |= !UsesOnlyImageDescriptors(*read);
} else if (std::ranges::any_of(read->Uses(), [](const Use& use) {
} else if (std::ranges::any_of(read->Uses(), [&](const Use& use) {
return use.user->GetOpcode() != ValueOpcode::GetBufferResource ||
std::ranges::any_of(use.user->Uses(), [](const Use& consumer) {
return consumer.user->GetOpcode() != ValueOpcode::StoreBufferU32;
std::ranges::any_of(use.user->Uses(), [&](const Use& consumer) {
return !IndirectBufferAccess(*consumer.user);
});
})) {
return false;
@@ -2780,19 +2825,14 @@ private:
IndirectDescriptorPlan& plan) {
if (handle.GetOpcode() != ValueOpcode::GetBufferResource || handle.NumArgs() != 4u ||
handle.Uses().empty()) return false;
if (handle.Uses().front().user->GetOpcode() != ValueOpcode::StoreBufferU32) return false;
if (!IndirectBufferAccess(*handle.Uses().front().user)) return false;
const bool legacy_store = handle.Uses().front().user->GetOpcode() == ValueOpcode::StoreBufferU32;
const auto indexed = m_program.memory_info[handle.Uses().front().user->Flags<MemoryFlags>().index].idxen;
for (const auto& use: handle.Uses()) {
const auto& inst = *use.user;
if (inst.GetOpcode() != ValueOpcode::StoreBufferU32) return false;
if (!IndirectBufferAccess(inst)) return false;
const auto& memory = m_program.memory_info[inst.Flags<MemoryFlags>().index];
uint32_t scalar = 1u;
if (memory.kind != ResourceKind::Buffer || memory.typed ||
(memory.formatted && (!indexed || memory.data_bits != 32u || memory.data_dwords != 1u)) ||
memory.idxen != indexed || memory.offen == indexed ||
(indexed && (memory.offset != 0u || inst.Arg(2).Resolve() != Value(0u))) ||
!ImmediateU32(inst.Arg(3), scalar) || scalar != 0u)
return false;
if (memory.idxen != indexed) return false;
}
DescriptorSource table_source;
Value key;
@@ -2800,12 +2840,25 @@ private:
uint32_t table_stride = 0;
uint32_t table_immediate = 0;
if (!MatchDescriptorTable(handle, descriptor, plan, table_source, key, table_offset,
table_stride, table_immediate) || table_stride != 16u) return false;
table_stride, table_immediate) ||
table_stride < 16u || (table_stride & 3u) != 0u) return false;
DescriptorSource::IndirectDescriptor indirect;
indirect.table_stride = table_stride;
indirect.table_offset = table_offset;
indirect.table_immediate = table_immediate;
indirect.indexed_store = indexed;
indirect.indexed = indexed;
for (uint32_t word = 0; word < 4u; ++word) {
const auto* anchor = plan.reads[word];
const auto* key_inst = key.Resolve().TryInstruction();
if (AvailableBefore(*anchor, handle) &&
(key_inst == nullptr || (key_inst != anchor && AvailableBefore(*key_inst, *anchor))) &&
std::all_of(plan.reads.begin(), plan.reads.begin() + 4u,
[&](const Inst* read) { return AvailableBefore(*read, *anchor); })) {
plan.selection_anchor = anchor;
break;
}
}
if (!legacy_store && plan.selection_anchor == nullptr) return false;
const auto finish = [&] {
indirect.table_source = InternSource(table_source);
DescriptorSource source;
@@ -2817,19 +2870,20 @@ private:
plan.key = key;
return true;
};
if (table_source.dword_count == 2u && MatchPackedScalarKey(key, indirect)) return finish();
if (const auto* bound = BoundedLoop(key, handle.Parent()); bound != nullptr) {
indirect.key_count = bound->Arg(1);
indirect.signed_count = bound->GetOpcode() == ValueOpcode::SLessThan32;
return finish();
}
if (!legacy_store || table_stride != 16u) return false;
if (table_source.dword_count == 2u) {
const auto* bound = BoundedLoop(key, handle.Parent());
if (bound != nullptr) {
indirect.key_count = bound->Arg(1);
indirect.signed_count = bound->GetOpcode() == ValueOpcode::SLessThan32;
} else {
if (!MatchPackedScalarKey(key, indirect)) {
// EXEC implications must describe the same iteration as the selected value.
if (m_native_cfg.irreducible || !m_native_cfg.back_edges.empty()) return false;
const auto* lane = key.Resolve().TryInstruction();
if (lane == nullptr || lane->GetOpcode() != ValueOpcode::ReadFirstLane ||
!std::ranges::all_of(handle.Uses(), [&](const Use& use) {
return Implies(use.user->Arg(5), lane->Arg(1));
return Implies(use.user->Arg(use.user->NumArgs() - 1u), lane->Arg(1));
})) return false;
const auto maximum = SelectorMaximum(lane->Arg(0), {}, 0u, lane->Arg(1));
if (maximum >= MaxIndirectDescriptorProbes) return false;
@@ -3112,7 +3166,7 @@ private:
for (auto* block: m_program.blocks) {
for (auto& inst: *block) {
if ((ImageOpcodeInfoOf(inst.GetOpcode()).access == ImageAccess::None &&
inst.GetOpcode() != ValueOpcode::StoreBufferU32) ||
!IndirectBufferAccess(inst)) ||
inst.NumArgs() == 0u) {
continue;
}
@@ -3138,7 +3192,7 @@ private:
m_indirect_descriptors.push_back(std::move(plan));
}
};
plan_handle(inst.Arg(0), inst.GetOpcode() == ValueOpcode::StoreBufferU32 ? 4u : 8u,
plan_handle(inst.Arg(0), IndirectBufferAccess(inst) ? 4u : 8u,
memory.resource * 4u, false);
if (inst.GetOpcode() == ValueOpcode::ImageSampleRaw &&
(memory.image_sample_flags & Decoder::ImageSampleFlagAdjust) == 0u)