raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-24 12:56:53 +02:00
Code Issues Packages Projects Releases Wiki Activity

shader: share finite texture selection through masked first-lane reads

Browse Source 
Co-authored-by: fxpw <fxpw@outlook.com>
This commit is contained in:
nmzik
2026-10-14 17:29:49 +02:00
parent 6d6da57a95
commit 7c9da0e56f
3 changed files with 292 additions and 23 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/shader/recompiler/ir/passes/ResourceTracking.cpp
+46 -23
View File
@@ -3196,9 +3196,6 @@ private:
return sampler && (m_program.memory_info[use.user->Flags<MemoryFlags>().index]
.image_sample_flags & Decoder::ImageSampleFlagAdjust) != 0u;
})) return false;
bool has_phi = false;
for (size_t word = 0; word < handle.NumArgs(); ++word) has_phi |= handle.Arg(word).Resolve().IsPhi();
if (!has_phi) return false;
// Normalize synchronized descriptor words to one GPU ordinal.
// Build the complete graph before changing IR so an unsupported leaf is transactional.
struct Choice {
@@ -3225,32 +3222,51 @@ private:
return index;
}
const auto* branch = descriptor.dwords[bad].TryInstruction();
if (branch == nullptr || branch->Parent() == nullptr ||
branch->GetOpcode() != ValueOpcode::Phi ||
branch->NumPhiBlocks() != branch->NumArgs()) return UINT32_MAX;
choices[index].branch = branch;
const auto count = branch->NumArgs();
if (branch == nullptr || branch->Parent() == nullptr) return UINT32_MAX;
const auto op = branch->GetOpcode();
const bool phi = op == ValueOpcode::Phi;
const bool lane = op == ValueOpcode::ReadFirstLane;
if ((!phi && !lane && op != ValueOpcode::SelectU32) ||
(phi && branch->NumPhiBlocks() != branch->NumArgs())) return UINT32_MAX;
if (lane && !std::ranges::all_of(handle.Uses(), [&](const Use& use) {
return Implies(use.user->Arg(use.user->NumArgs() - 1u), branch->Arg(1));
})) return UINT32_MAX;
const auto count = phi ? branch->NumArgs() : lane ? 1u : 2u;
std::array<const Inst*, 8> selected_words {};
const auto* anchor = branch;
for (uint32_t word = 0; word < descriptor.dword_count; ++word) {
const auto* selected = descriptor.dwords[word].TryInstruction();
if (selected == nullptr || selected->GetOpcode() != op ||
selected->Parent() != branch->Parent()) {
if (!ValidateRuntimeValue(m_program, descriptor.dwords[word])) return UINT32_MAX;
continue;
}
if (phi ? selected->NumArgs() != count || selected->NumPhiBlocks() != count
: !EquivalentValue(m_program, selected->Arg(lane ? 1u : 0u),
branch->Arg(lane ? 1u : 0u))) return UINT32_MAX;
selected_words[word] = selected;
if (!phi && AvailableBefore(*anchor, *selected)) anchor = selected;
}
// Later words may depend on choices computed after the first descriptor word.
choices[index].branch = anchor;
for (uint32_t arm = 0; arm < count; ++arm) {
// A nonzero bit scan cannot choose its zero-input switch sentinel.
// Preserve the Phi edge; its ordinal is never observed at runtime.
if (ImpossibleSwitchEdge(branch->PhiBlock(arm), branch->Parent())) {
if (phi && ImpossibleSwitchEdge(anchor->PhiBlock(arm), anchor->Parent())) {
choices[index].children.push_back(UINT32_MAX);
continue;
}
auto child = descriptor;
for (uint32_t word = 0; word < child.dword_count; ++word) {
const auto* selected = descriptor.dwords[word].TryInstruction();
if (selected == nullptr || selected->GetOpcode() != branch->GetOpcode() ||
selected->Parent() != branch->Parent()) {
if (!ValidateRuntimeValue(m_program, descriptor.dwords[word])) return UINT32_MAX;
continue;
const auto* selected = selected_words[word];
if (selected == nullptr) continue;
uint32_t incoming = lane ? 0u : arm + 1u;
if (phi) {
incoming = 0u;
while (incoming < count && selected->PhiBlock(incoming) != anchor->PhiBlock(arm))
++incoming;
if (incoming == count) return UINT32_MAX;
}
if (selected->NumArgs() != count || selected->NumPhiBlocks() != count)
return UINT32_MAX;
uint32_t incoming = 0;
while (incoming < count && selected->PhiBlock(incoming) != branch->PhiBlock(arm))
++incoming;
if (incoming == count) return UINT32_MAX;
child.dwords[word] = selected->Arg(incoming);
}
const auto next = self(self, child);
@@ -3282,17 +3298,24 @@ private:
const auto where = std::ranges::find_if(block->Instructions(), [&](const Inst& inst) {
return &inst == choice.branch;
});
auto& key = *block->PrependNewInst(where, choice.branch->GetOpcode());
key.SetFlags(Type::U32);
const auto op = choice.branch->GetOpcode();
auto& key = *(op == ValueOpcode::Phi ? block->PrependNewInst(where, op)
: op == ValueOpcode::ReadFirstLane
? block->PrependNewInst(where, op, {Value(0u), choice.branch->Arg(1)})
: block->PrependNewInst(where, op, {choice.branch->Arg(0), Value(0u), Value(0u)}));
if (op == ValueOpcode::Phi) key.SetFlags(Type::U32);
choice.key = Value(&key);
}
for (const auto& choice: choices) {
if (choice.branch == nullptr) continue;
auto* key = choice.key.Instruction();
const bool phi = choice.branch->GetOpcode() == ValueOpcode::Phi;
const bool lane = choice.branch->GetOpcode() == ValueOpcode::ReadFirstLane;
for (uint32_t arm = 0; arm < choice.children.size(); ++arm) {
const auto child = choice.children[arm];
const auto value = child == UINT32_MAX ? Value(0u) : choices[child].key;
key->AddPhiOperand(choice.branch->PhiBlock(arm), value);
if (phi) key->AddPhiOperand(choice.branch->PhiBlock(arm), value);
else key->SetArg(lane ? 0u : arm + 1u, value);
}
}
plan.handle = &handle;