raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-20 02:47:02 +02:00
Code Issues Packages Projects Releases Wiki Activity

shader: prove narrowing loop masks through shared implication analysis

Browse Source 
Co-authored-by: fxpw <fxpw@outlook.com>
This commit is contained in:
nmzik
2026-10-15 14:05:41 +02:00
parent a50e4de389
commit 2285a25182
3 changed files with 114 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/ResourceTrackingTests.cpp
+46 -6
View File
@@ -5387,7 +5387,8 @@ void TestGuardedWaterfallBufferRecords() {
using Libs::Graphics::DstSel;
enum class Guard {
Exact, Reversed, Greater, GreaterEqual, WrongGreater, SignedGreater,
NoRange, NoEquality, OtherLocal, UnguardedUse, SeparateVotes
NoRange, NoEquality, OtherLocal, UnguardedUse, SeparateVotes,
Loop, LoopReversed, LoopEntries, LoopBadEntry, LoopWiden, LoopUnanchored, LoopVarying
};
const auto make = [](Guard guard) {
auto f = std::make_unique<Fixture>();
@@ -5398,7 +5399,6 @@ void TestGuardedWaterfallBufferRecords() {
const auto local = f->Emit(ValueOpcode::LoadBufferU32,
{input, lane, Value(0u), Value(0u), Value(true)},
f->AddMemory({.kind = ResourceKind::Buffer, .idxen = true}, 0x200u));
const auto key = f->Emit(ValueOpcode::ReadLane, {local, Value(0u)});
const auto initial = f->Emit(ValueOpcode::ULessThan32, {lane, Value(32u)});
const auto range_local = guard == Guard::OtherLocal
? f->Emit(ValueOpcode::IAdd32, {local, Value(1u)}) : local;
@@ -5413,11 +5413,48 @@ void TestGuardedWaterfallBufferRecords() {
range = f->Emit(ValueOpcode::LogicalAnd, {initial, compare});
} else {
// Match ANDN1_SAVEEXEC, including its initial-EXEC witness.
auto executing = initial;
auto comparison_mask = initial;
if (guard >= Guard::Loop) {
auto* entry = f->block;
auto* header = f->AddBlock();
f->block = header;
auto& mask = header->AppendNewInst(ValueOpcode::Phi, {}, uint64_t(Type::U1));
if (guard != Guard::LoopUnanchored) entry->AddBranch(header);
header->AddBranch(header);
if (guard == Guard::LoopVarying) {
auto& predicate = header->AppendNewInst(ValueOpcode::Phi, {}, uint64_t(Type::U1));
predicate.AddPhiOperand(entry, initial);
predicate.AddPhiOperand(header, f->Emit(ValueOpcode::LogicalNot, {Value(&predicate)}));
comparison_mask = Value(&predicate);
}
const auto next = guard == Guard::LoopWiden
? f->Emit(ValueOpcode::LogicalNot, {initial})
: guard == Guard::LoopVarying ? comparison_mask
: f->Emit(ValueOpcode::ULessThan32, {lane, Value(16u)});
const auto carried = f->Emit(guard == Guard::LoopWiden
? ValueOpcode::LogicalOr : ValueOpcode::LogicalAnd, {Value(&mask), next});
if (guard == Guard::LoopReversed) mask.AddPhiOperand(header, carried);
if (guard != Guard::LoopUnanchored) mask.AddPhiOperand(entry, initial);
if (guard != Guard::LoopReversed) mask.AddPhiOperand(header, carried);
if (guard == Guard::LoopEntries || guard == Guard::LoopBadEntry ||
guard == Guard::LoopUnanchored) {
auto* alternate = f->AddBlock();
(guard == Guard::LoopUnanchored ? header : entry)->AddBranch(alternate);
alternate->AddBranch(header);
const auto incoming = guard == Guard::LoopUnanchored
? f->Emit(ValueOpcode::LogicalAnd, {Value(&mask), initial}, 0, alternate)
: guard == Guard::LoopBadEntry ? Value(true) : initial;
mask.AddPhiOperand(alternate, incoming);
}
executing = Value(&mask);
}
const auto excluded = f->Emit(ValueOpcode::LogicalAnd,
{initial, f->Emit(ValueOpcode::ULessThanEqual32, {count, range_local})});
{comparison_mask, f->Emit(ValueOpcode::ULessThanEqual32, {count, range_local})});
range = f->Emit(ValueOpcode::LogicalAnd,
{initial, f->Emit(ValueOpcode::LogicalNot, {excluded})});
{executing, f->Emit(ValueOpcode::LogicalNot, {excluded})});
}
const auto key = f->Emit(ValueOpcode::ReadLane, {local, Value(0u)});
const auto equal = f->Emit(ValueOpcode::IEqual32,
guard == Guard::Reversed ? std::initializer_list<Value>{local, key}
: std::initializer_list<Value>{key, local});
@@ -5466,14 +5503,17 @@ void TestGuardedWaterfallBufferRecords() {
ResourceSpecialization specialization;
for (const auto guard : {Guard::Exact, Guard::Reversed, Guard::Greater, Guard::GreaterEqual,
Guard::WrongGreater, Guard::SignedGreater, Guard::NoRange, Guard::NoEquality,
Guard::OtherLocal, Guard::UnguardedUse, Guard::SeparateVotes}) {
Guard::OtherLocal, Guard::UnguardedUse, Guard::SeparateVotes,
Guard::Loop, Guard::LoopReversed, Guard::LoopEntries, Guard::LoopBadEntry,
Guard::LoopWiden, Guard::LoopUnanchored, Guard::LoopVarying}) {
auto fixture = make(guard);
const auto plan = ExtractResourcePlan(fixture->program);
const auto root = std::ranges::find_if(plan.info.buffers, [&](const auto& buffer) {
return plan.descriptor_sources[buffer.source].indirect_descriptor.has_value();
});
Check(root != plan.info.buffers.end(), "waterfall buffer lost its indirect resource");
const bool proven = guard == Guard::Exact || guard == Guard::Reversed || guard == Guard::Greater;
const bool proven = guard == Guard::Exact || guard == Guard::Reversed || guard == Guard::Greater ||
guard == Guard::Loop || guard == Guard::LoopReversed || guard == Guard::LoopEntries;
const auto& domain = *plan.descriptor_sources[root->source].indirect_descriptor;
Check(domain.key_count.IsEmpty() != proven && !domain.selector &&
domain.table_stride == 488u && domain.table_immediate == 216u,