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: enumerate set ranks of eight-word light bitmaps

Browse Source
This commit is contained in:
Claxten
2026-10-27 00:11:27 +01:00
committed by nmzik
parent ffb6ad6cab
commit 914a3bb4b8
3 changed files with 122 additions and 57 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/shader/recompiler/ir/passes/ResourceMaterialization.cpp
+7 -5
View File
@@ -377,19 +377,21 @@ bool ResolveSelectorKeys(const ResourcePlan& program, const DescriptorSource::In
}
} else if (selector->bitset) {
ShaderBufferResource material;
std::array<uint32_t, 4> masks {};
std::array<uint32_t, 8> storage {};
const auto record_words = std::min<size_t>(selector->stride / 4u, storage.size());
const auto masks = std::span(storage).first(record_words);
uint32_t first = 0, count = 0;
if (!DecodeBufferDescriptor(material_value, material) ||
!WalkedValue(program, indirect.selector_first, first) ||
!EvaluateIndirectCount(program, indirect, runtime, count) || count != masks.size() * 32u ||
selector->stride != sizeof(masks) || selector->offset != 0u ||
material.Type() != 0u || material.Stride() != sizeof(masks) ||
selector->stride != masks.size_bytes() || selector->offset != 0u ||
material.Type() != 0u || material.Stride() != selector->stride ||
material.SwizzleEnabled() || material.AddTid() || material.OutOfBounds() != 0u ||
(material.Base48() & 3u) != 0u) return false;
if (material.Format() != Prospero::BufferFormat::kInvalid) {
const auto words = material.GetSize() / sizeof(uint32_t);
// Bitmap storage can grow with resolution while its candidate domain stays
// at 128 keys. Bound scratch memory, preserving each coherent read's capture.
// Bitmap storage can grow with resolution while its candidate domain stays at 32 keys
// per record word. Bound scratch memory, preserving each coherent read's capture.
for (uint64_t first_word = 0; first_word < words; first_word += keys.size()) {
keys.resize(std::min<uint64_t>(words - first_word, MaxIndirectDescriptorProbes));
if (!ReadScalarTable(material.Base48(), material.GetSize(), first_word * 4u,
src/graphics/shader/recompiler/ir/passes/ResourceTracking.cpp
+40 -24
View File
@@ -2567,7 +2567,8 @@ private:
// Each equation is an inductive subset obligation. Phis require every incoming
// value; AND needs either operand. Grounding rejects cycles without an initial
// value, then the greatest fixed point checks every initial and carried edge.
bool BitsetMaskSubset(Value mask, Value active, Value word, const Inst& read) const {
bool BitsetMaskSubset(Value mask, Value active, Value word,
std::span<const Inst* const> reads) const {
struct State {
Value value;
Value guard;
@@ -2579,7 +2580,8 @@ private:
bool all = true;
std::vector<size_t> terms;
};
const auto load_guard = read.Arg(4).Resolve();
const auto load_guard = reads[0]->Arg(4).Resolve();
const auto words = static_cast<uint32_t>(reads.size()) * 4u;
std::vector<State> states(2);
std::vector<Equation> equations(2);
const auto add = [&](State state) {
@@ -2590,9 +2592,9 @@ private:
equations.emplace_back();
return states.size() - 1u;
};
std::array<size_t, 5> roots {add({.value = active})};
for (uint32_t component = 0; component < 4u; ++component)
roots[component + 1u] = add({mask, active, component, component});
std::vector<size_t> roots {add({.value = active})};
for (uint32_t component = 0; component < words; ++component)
roots.push_back(add({mask, active, component, component}));
for (size_t index = 2; index < states.size(); ++index) {
const auto state = states[index];
const auto* inst = state.value.TryInstruction();
@@ -2642,8 +2644,8 @@ private:
switch (inst->GetOpcode()) {
case ValueOpcode::CompositeExtractU32x4: {
uint32_t component;
if (inst->Arg(0).Resolve().TryInstruction() == &read &&
ImmediateU32(inst->Arg(1), component) && component == state.component)
if (inst->Arg(0).Resolve().TryInstruction() == reads[state.component / 4u] &&
ImmediateU32(inst->Arg(1), component) && component == state.component % 4u)
equation.terms = {1u};
break;
}
@@ -2667,17 +2669,17 @@ private:
const auto* masked = key.TryInstruction();
uint32_t bound = 0;
if (masked != nullptr && masked->GetOpcode() == ValueOpcode::BitwiseAnd32 &&
DescendingLoop(word, masked->Parent(), bound) && bound == 3u) {
DescendingLoop(word, masked->Parent(), bound) && bound == words - 1u) {
for (uint32_t side = 0; side < 2u; ++side) {
uint32_t bits;
if (ImmediateU32(masked->Arg(side), bits) && (bits & 3u) == 3u)
if (ImmediateU32(masked->Arg(side), bits) && (bits & bound) == bound)
key = masked->Arg(side ^ 1u).Resolve();
}
}
return EquivalentValue(m_program, key, word);
});
uint32_t literal;
if (ImmediateU32(equality, literal) && literal < 4u) selected_word = literal;
if (ImmediateU32(equality, literal) && literal < words) selected_word = literal;
equation.terms.push_back(term(inst->Arg(1), inst->Arg(0), selected_word));
}
if (!selected || !*selected)
@@ -2715,7 +2717,7 @@ private:
return std::ranges::all_of(roots, [&](size_t root) { return valid[root]; });
}
bool MatchBitsetSelector(Value rank, Value active, SelectorBounds bounds,
bool MatchBitsetSelector(Value rank, Value active, SelectorBounds bounds, uint32_t words,
DescriptorSource::IndirectDescriptor& indirect) {
const auto selected_value = [&](Value value) {
for (;;) {
@@ -2739,7 +2741,7 @@ private:
}
const auto* word_inst = word.TryInstruction();
if (word_inst == nullptr || std::ranges::none_of(bounds, [&](const auto& bound) {
return bound.first == word_inst && bound.second == 3u;
return bound.first == word_inst && bound.second == words - 1u;
})) return false;
// FFBH's leading-zero count is inverted again to recover the set-bit rank.
for (uint32_t inverse = 0; inverse < 2u; ++inverse) {
@@ -2751,21 +2753,34 @@ private:
const auto* scan = bit.TryInstruction();
if (scan == nullptr || scan->GetOpcode() != ValueOpcode::FindUMsb32 ||
!ImpliesNonzero(active, scan->Arg(0))) return false;
const auto bitmap_load = [&](const Inst& read, uint32_t offset) {
const auto index = read.Flags<MemoryFlags>().index;
if (read.GetOpcode() != ValueOpcode::LoadBufferU32x4 || InCycle(read.Parent()) ||
index >= m_program.memory_info.size()) return false;
const auto& memory = m_program.memory_info[index];
return memory.kind == ResourceKind::Buffer && !memory.typed && !memory.formatted &&
memory.idxen && !memory.offen && memory.offset == offset &&
memory.data_bits == 32u && memory.data_dwords == 4u &&
read.Arg(2).Resolve() == Value(0u) && read.Arg(3).Resolve() == Value(0u);
};
for (const auto* block: m_program.blocks) {
for (const auto& read: *block) {
if (read.GetOpcode() != ValueOpcode::LoadBufferU32x4 || InCycle(block)) continue;
const auto index = read.Flags<MemoryFlags>().index;
if (index >= m_program.memory_info.size()) continue;
const auto& memory = m_program.memory_info[index];
if (memory.kind != ResourceKind::Buffer || memory.typed || memory.formatted ||
!memory.idxen || memory.offen || memory.offset != 0u || memory.data_bits != 32u ||
memory.data_dwords != 4u || read.Arg(2).Resolve() != Value(0u) ||
read.Arg(3).Resolve() != Value(0u) ||
!BitsetMaskSubset(scan->Arg(0), active, word, read)) continue;
if (!bitmap_load(read, 0u)) continue;
// An eight-word record continues in a second four-word load of the same buffer and EXEC.
std::array<const Inst*, 2> reads {&read};
if (words == 8u)
for (const auto* other: m_program.blocks)
for (const auto& load: *other)
if (reads[1] == nullptr && bitmap_load(load, 16u) &&
EquivalentValue(m_program, load.Arg(0), read.Arg(0)) &&
EquivalentValue(m_program, load.Arg(4), read.Arg(4))) reads[1] = &load;
const auto parts = std::span(reads).first(words / 4u);
DescriptorSource source;
if (!MakeRuntimeTableSource(read, source)) continue;
if (std::ranges::find(parts, nullptr) != parts.end() ||
!BitsetMaskSubset(scan->Arg(0), active, word, parts) ||
!MakeRuntimeTableSource(read, source)) continue;
indirect.selector.emplace(DescriptorSource::IndirectDescriptor::SelectorRead {
.source = InternSource(source), .stride = 16u, .offset = 0u, .bitset = true});
.source = InternSource(source), .stride = words * 4u, .offset = 0u, .bitset = true});
return true;
}
}
@@ -2795,7 +2810,8 @@ private:
first.SetArg(1, Value(maximum));
indirect.selector_first = Value(&first);
indirect.key_count = Value(maximum + 1u);
if (maximum == 127u) MatchBitsetSelector(lane->Arg(0), lane->Arg(1), bounds, indirect);
if (maximum == 127u || maximum == 255u)
MatchBitsetSelector(lane->Arg(0), lane->Arg(1), bounds, (maximum + 1u) / 32u, indirect);
return true;
}