shader: bound descriptor worklists by their finite bit masks

This commit is contained in:
nmzik
2026-10-13 00:37:14 +02:00
parent 9bcf80bf67
commit d70cbd4184
3 changed files with 193 additions and 73 deletions
+93 -14
View File
@@ -565,26 +565,28 @@ void TestGuardedDirectImageTable() {
Nonzero, SccNonZero, Plain, Zero, Unrelated, Bypass, ExecZero, VccZero,
Bounded, BoundMissing, BoundUnrelated, BoundBypass, BoundAnyLane,
BoundDiamond, BoundBranchSpecific, BoundMismatched,
BoundExec, BoundExecUnguarded, BoundExecMismatched, BoundExecAnd, BoundExecUnprovenActive
BoundExec, BoundExecUnguarded, BoundExecMismatched, BoundExecAnd, BoundExecUnprovenActive,
Finite, FiniteZero, FiniteMissing, FiniteWrong, FiniteSigned, FiniteAdd, FiniteXor
};
const auto make_plan = [](Guard guard) {
const bool bounded = guard >= Guard::Bounded;
const bool exec_bound = guard >= Guard::BoundExec;
const bool finite = guard >= Guard::Finite;
const bool bounded = guard >= Guard::Bounded && !finite;
const bool exec_bound = bounded && guard >= Guard::BoundExec;
const bool bypass = guard == Guard::Bypass || guard == Guard::BoundBypass;
const uint32_t table_offset = exec_bound ? 672u : bounded ? 3696u : 344u;
const uint32_t table_offset = finite ? 3344u : exec_bound ? 672u : bounded ? 3696u : 344u;
const uint32_t table_stride = exec_bound ? 784u : bounded ? 272u : 32u;
Fixture fixture(bounded ? ShaderType::Compute : ShaderType::Pixel);
Fixture fixture(bounded || finite ? ShaderType::Compute : ShaderType::Pixel);
fixture.program.wave_size = 64u;
auto *entry = fixture.block;
auto *middle = fixture.AddBlock();
auto *before_sample = fixture.AddBlock();
auto *sample = fixture.AddBlock();
auto *exit = fixture.AddBlock();
const auto varying = bounded
const auto varying = bounded || finite
? fixture.Emit(ValueOpcode::GetBuiltin,
{Value(static_cast<uint32_t>(StageInputKind::LocalInvocationId)), Value(0u)})
: fixture.Emit(ValueOpcode::GetAttribute, {Value(0u), Value(0u)});
const auto mask = fixture.Emit(ValueOpcode::ReadFirstLane, {varying, Value(true)});
auto mask = fixture.Emit(ValueOpcode::ReadFirstLane, {varying, Value(true)});
const auto srt = fixture.Address(fixture.UserData(0), fixture.UserData(1));
Value key;
Value limit;
@@ -596,13 +598,44 @@ void TestGuardedDirectImageTable() {
limit = fixture.Emit(ValueOpcode::LoadAddressU32,
{srt, Value(0u), Value(0u), Value(true)}, fixture.AddMemory(memory, 0x10));
}
Block *latch = nullptr;
if (finite) {
latch = fixture.AddBlock();
MemoryInfo memory;
memory.kind = ResourceKind::ScalarBuffer;
const auto input = fixture.Buffer({Value(0x900000u), Value(0u), Value(4096u), Value(0u)});
const auto loaded = fixture.Emit(ValueOpcode::ReadConstBuffer, {input, mask},
fixture.AddMemory(memory, 0x2ecu));
const auto initial = fixture.Emit(ValueOpcode::BitFieldUExtract,
{loaded, Value(0u), Value(guard == Guard::FiniteZero ? 0u : 12u)});
auto &phi = middle->AppendNewInst(ValueOpcode::Phi, {}, static_cast<uint64_t>(Type::U32));
mask = Value(&phi);
key = fixture.Emit(ValueOpcode::FindILsb32, {mask}, 0, middle);
const auto bit = fixture.Emit(ValueOpcode::ShiftLeftLogical32,
{Value(1u), fixture.Emit(ValueOpcode::BitwiseAnd32, {key, Value(31u)}, 0, latch)}, 0, latch);
const auto next = guard == Guard::FiniteAdd
? fixture.Emit(ValueOpcode::BitwiseOr32, {mask, Value(1u << 31u)}, 0, latch)
: guard == Guard::FiniteXor
? fixture.Emit(ValueOpcode::BitwiseXor32, {mask, bit}, 0, latch)
: fixture.Emit(ValueOpcode::BitwiseAnd32,
{mask, fixture.Emit(ValueOpcode::BitwiseNot32, {bit}, 0, latch)}, 0, latch);
phi.AddPhiOperand(entry, initial);
phi.AddPhiOperand(latch, next);
}
const auto kind = guard == Guard::ExecZero || guard == Guard::BoundAnyLane || exec_bound
? CFG::BranchCondition::ExecZero
: guard == Guard::VccZero ? CFG::BranchCondition::VccZero
: CFG::BranchCondition::SccZero;
Value condition;
Value active(true);
if (bounded) {
if (finite) {
condition = guard == Guard::FiniteMissing ? Value(false)
: guard == Guard::FiniteSigned
? fixture.Emit(ValueOpcode::LogicalNot,
{fixture.Emit(ValueOpcode::SLessThan32, {key, Value(12u)}, 0, middle)}, 0, middle)
: fixture.Emit(ValueOpcode::IEqual32,
{key, Value(guard == Guard::FiniteWrong ? 31u : UINT32_MAX)}, 0, middle);
} else if (bounded) {
condition = guard == Guard::BoundMissing ? Value(false) : fixture.Emit(
ValueOpcode::UGreaterThanEqual32,
{guard == Guard::BoundUnrelated ? fixture.UserData(2) : key, limit});
@@ -630,7 +663,7 @@ void TestGuardedDirectImageTable() {
condition = fixture.Emit(ValueOpcode::LogicalNot, {condition});
}
if (guard != Guard::Plain && guard != Guard::SccNonZero) {
condition = fixture.Emit(ValueOpcode::ConditionRef, {condition}, kind);
condition = fixture.Emit(ValueOpcode::ConditionRef, {condition}, kind, finite ? middle : entry);
}
fixture.program.blocks[0]->condition = condition;
fixture.program.blocks[0]->terminator = {
@@ -645,6 +678,18 @@ void TestGuardedDirectImageTable() {
fixture.program.blocks[4]->terminator = {
.kind = bypass ? CFG::TerminatorKind::Branch : CFG::TerminatorKind::Return,
.true_block = fixture.program.blocks[3u]};
if (finite) {
entry->condition = {};
entry->terminator = {.kind = CFG::TerminatorKind::Branch, .true_block = middle};
middle->condition = condition;
middle->terminator = {.kind = CFG::TerminatorKind::ConditionalBranch,
.true_block = guard == Guard::FiniteXor ? latch : exit, .false_block = before_sample};
before_sample->condition = fixture.Emit(ValueOpcode::IEqual32, {varying, Value(0u)}, 0, before_sample);
before_sample->terminator = {.kind = CFG::TerminatorKind::ConditionalBranch,
.true_block = sample, .false_block = latch};
sample->terminator = {.kind = CFG::TerminatorKind::Branch, .true_block = latch};
latch->terminator = {.kind = CFG::TerminatorKind::Branch, .true_block = middle};
}
if (guard == Guard::BoundDiamond || exec_bound) {
middle->condition = fixture.Emit(ValueOpcode::IEqual32,
{mask, Value(1u)});
@@ -673,7 +718,7 @@ void TestGuardedDirectImageTable() {
}
std::array<Value, 2> pointer;
for (uint32_t word = 0; word < pointer.size(); ++word) {
if (bounded) {
if (bounded || finite) {
pointer[word] = srt.Instruction()->Arg(word);
continue;
}
@@ -685,7 +730,7 @@ void TestGuardedDirectImageTable() {
fixture.AddMemory(memory, 0x20));
}
fixture.block = sample;
if (!bounded) key = fixture.Emit(ValueOpcode::FindILsb32, {mask});
if (!bounded && !finite) key = fixture.Emit(ValueOpcode::FindILsb32, {mask});
const auto offset = fixture.Emit(ValueOpcode::IAdd32,
{bounded ? fixture.Emit(ValueOpcode::IMul32, {key, Value(table_stride)})
: fixture.Emit(ValueOpcode::ShiftLeftLogical32, {key, Value(5u)}),
@@ -719,7 +764,7 @@ void TestGuardedDirectImageTable() {
fixture.Emit(ValueOpcode::ImageSampleRaw,
{image, sampler, fixture.ImageAddress(), other_active}, fixture.AddMemory(memory, 0x12c));
}
if (bounded) {
if (bounded || finite) {
const auto output = fixture.Image({Value(0x8000u),
Value(static_cast<uint32_t>(Libs::Graphics::Prospero::BufferFormat::k32UInt) << 20u),
Value(0u), Value((static_cast<uint32_t>(Libs::Graphics::Prospero::ImageType::kColor2D) << 28u) |
@@ -751,7 +796,8 @@ void TestGuardedDirectImageTable() {
Check(indirect && !indirect->selector.has_value() &&
indirect->table_offset == table_offset && indirect->table_stride == table_stride &&
(bounded ? !indirect->key_count.IsEmpty()
: indirect->key_count.Resolve() == Value(32u)) &&
: indirect->key_count.Resolve() == Value(guard == Guard::FiniteZero ? 0u :
finite && guard != Guard::FiniteAdd && guard != Guard::FiniteXor ? 12u : 32u)) &&
fixture.program.descriptor_sources[indirect->table_source].dword_count == 2u,
"guarded direct image table lost its pointer or proven selector range");
return ExtractResourcePlan(fixture.program);
@@ -765,7 +811,8 @@ void TestGuardedDirectImageTable() {
Guard::BoundUnrelated, Guard::BoundBypass, Guard::BoundAnyLane,
Guard::BoundBranchSpecific, Guard::BoundMismatched,
Guard::BoundExecUnguarded, Guard::BoundExecMismatched,
Guard::BoundExecAnd, Guard::BoundExecUnprovenActive}) {
Guard::BoundExecAnd, Guard::BoundExecUnprovenActive,
Guard::FiniteMissing, Guard::FiniteWrong, Guard::FiniteSigned}) {
CheckFatal([&] { make_plan(guard); }, "not a valid runtime value",
"direct table accepted a selector without a dominating range guard");
}
@@ -892,6 +939,38 @@ void TestGuardedDirectImageTable() {
records.descriptor_reads == 0u,
"unrepresentable table span was read before validating its count");
}
for (const auto guard : {Guard::Finite, Guard::FiniteZero, Guard::FiniteAdd, Guard::FiniteXor}) {
const auto finite = make_plan(guard);
Check(finite.capture_specialization_reads && finite.info.images[1].written,
"finite bit table with image writes omitted specialization read capture");
LinearTestMemory records;
const uint32_t count = guard == Guard::FiniteZero ? 0u : guard == Guard::Finite ? 12u : 32u;
const uint64_t first = records.base + 3344u;
fill_table(records, first);
// The GPU may select no bits or a sparse subset. Null rows remain valid candidates.
std::fill_n(records.words.begin() + (first - records.base) / 4u + 5u * 8u, 8u, 0u);
records.fail_address = first + count * 32u;
user_data = {static_cast<uint32_t>(records.base), 0u};
runtime.userdata = &records;
Check(MaterializeResources(finite, runtime, snapshot, specialization) &&
records.descriptor_reads == count && snapshot.images.size() == std::max(count, 1u) + 1u,
"finite bit table probed beyond its mask bound or lost a null row");
if (count == 0u) {
Check(std::ranges::all_of(snapshot.images[0].dwords, [](uint32_t word) { return word == 0u; }),
"empty bit loop retained a reachable image");
continue;
}
Check(snapshot.flattened_srt[specialization.images[0].indirect_mapping_offset] == count,
"finite bit table lost its candidate count");
for (uint32_t key = 0; key < count; ++key) {
Check(std::ranges::any_of(snapshot.specialization_reads, [&](const auto &read) {
return read.first == first + key * 32u && read.second == 32u;
}), "finite bit table omitted a descriptor from writable-alias validation");
}
records.fail_address = first + (count - 1u) * 32u + 28u;
Check(!MaterializeResources(finite, runtime, snapshot, specialization),
"finite bit table accepted an unreadable last candidate");
}
}
void TestBoundedComputeImageLoop() {