mirror of
https://github.com/KytyPS5/KytyPS5.git
synced 2026-10-20 02:47:02 +02:00
shader: prove inactive resources across masked traversal merges
This commit is contained in:
@@ -673,7 +673,7 @@ void TestSharedControlPredicates() {
|
||||
program.blocks[i]->terminator.false_block = program.blocks[2];
|
||||
}
|
||||
auto plan = ExtractResourcePlan(program);
|
||||
Check(plan.value_storage.size() == depth + 5 &&
|
||||
Check(plan.value_storage.size() <= 2u * depth + 12u &&
|
||||
plan.control_flow.size() == 4 &&
|
||||
plan.control_flow[0].condition == plan.control_flow[1].condition &&
|
||||
!plan.capture_specialization_reads,
|
||||
@@ -793,6 +793,156 @@ void TestFiniteImageRefreshReusesScalarReads() {
|
||||
"finite image refresh grew reusable resource storage after warmup");
|
||||
}
|
||||
|
||||
void TestEmptyTraversalSkipsUnmappedDescriptor(bool independent_mask) {
|
||||
using namespace Libs::Graphics::ShaderRecompiler::IR;
|
||||
using Libs::Graphics::ShaderRecompiler::CFG::TerminatorKind;
|
||||
Program program;
|
||||
program.stage = Libs::Graphics::ShaderType::Compute;
|
||||
program.srt_plan_complete = program.resource_tracking_complete = true;
|
||||
for (uint32_t i = 0; i < 6; ++i) AddValueBlock(program);
|
||||
auto &entry = *program.blocks[0];
|
||||
auto &loop = *program.blocks[1];
|
||||
auto &body = *program.blocks[2];
|
||||
auto &merge = *program.blocks[3];
|
||||
auto &use = *program.blocks[4];
|
||||
auto &exit = *program.blocks[5];
|
||||
const auto emit = [](Block &block, ValueOpcode opcode, std::initializer_list<Value> args) {
|
||||
return Value(&block.AppendNewInst(opcode, args));
|
||||
};
|
||||
const auto branch = [](Block &block, Block &target, Value condition = {}, Block *other = nullptr) {
|
||||
block.terminator.kind = other ? TerminatorKind::ConditionalBranch : TerminatorKind::Branch;
|
||||
block.terminator.true_block = ⌖
|
||||
block.terminator.false_block = other;
|
||||
block.condition = condition;
|
||||
block.AddBranch(&target);
|
||||
if (other) block.AddBranch(other);
|
||||
};
|
||||
const auto lane = emit(entry, ValueOpcode::GetBuiltin,
|
||||
{Value(static_cast<uint32_t>(StageInputKind::LocalInvocationId)), Value(0u)});
|
||||
const auto active = emit(entry, ValueOpcode::INotEqual32, {lane, Value(0u)});
|
||||
const auto initial_mask = independent_mask
|
||||
? emit(entry, ValueOpcode::IEqual32, {lane, Value(0u)}) : active;
|
||||
const auto initial = emit(entry, ValueOpcode::SelectU32,
|
||||
{initial_mask, Value(UINT32_MAX), Value(0u)});
|
||||
program.memory_info.push_back({.kind = ResourceKind::ScalarAddress, .planning_only = true});
|
||||
const auto address = emit(entry, ValueOpcode::GetAddressResource, {Value(0x1000u), Value(0u)});
|
||||
const auto count = emit(entry, ValueOpcode::LoadAddressU32,
|
||||
{address, Value(0u), Value(0u), Value(true)});
|
||||
count.Instruction()->SetFlags(MemoryFlags{.index = 0});
|
||||
const auto rejected = emit(entry, ValueOpcode::ULessThanEqual32, {count, lane});
|
||||
const auto accepted = emit(entry, ValueOpcode::LogicalNot, {rejected});
|
||||
branch(entry, loop, emit(entry, ValueOpcode::LogicalAnd, {active, accepted}), &merge);
|
||||
|
||||
// A nonempty traversal contains a varying loop-carried value. Its cycle must
|
||||
// remain unknown; only the proven empty traversal can select the bypass arm.
|
||||
auto &carried = loop.AppendNewInst(ValueOpcode::Phi);
|
||||
carried.SetFlags(Type::U32);
|
||||
const auto next = emit(body, ValueOpcode::IAdd32, {Value(&carried), Value(1u)});
|
||||
carried.AddPhiOperand(&entry, lane);
|
||||
carried.AddPhiOperand(&body, next);
|
||||
branch(loop, body, emit(loop, ValueOpcode::INotEqual32, {Value(&carried), Value(UINT32_MAX)}), &merge);
|
||||
loop.terminator.loop_header = true;
|
||||
loop.terminator.continue_block = &body;
|
||||
loop.terminator.merge_block = &merge;
|
||||
branch(body, loop);
|
||||
auto &selected = merge.AppendNewInst(ValueOpcode::Phi);
|
||||
selected.SetFlags(Type::U32);
|
||||
selected.AddPhiOperand(&entry, initial);
|
||||
selected.AddPhiOperand(&loop, Value(&carried));
|
||||
const auto valid = emit(merge, ValueOpcode::INotEqual32, {Value(&selected), Value(UINT32_MAX)});
|
||||
branch(merge, use, emit(merge, ValueOpcode::LogicalAnd, {active, valid}), &exit);
|
||||
branch(use, exit);
|
||||
DescriptorSource source;
|
||||
source.dwords = {Value(0x8000u), Value(1u << 16u), Value(16u), Value(3u << 28u)};
|
||||
source.dword_count = 4;
|
||||
program.descriptor_sources.push_back(source);
|
||||
program.info.buffers.push_back({.source = 0, .read = true});
|
||||
program.memory_info.push_back({.kind = ResourceKind::Buffer, .resource = 0});
|
||||
const auto buffer = emit(use, ValueOpcode::GetBufferResource,
|
||||
{source.dwords[0], source.dwords[1], source.dwords[2], source.dwords[3]});
|
||||
emit(use, ValueOpcode::LoadBufferU32, {buffer, Value(0u), Value(0u), Value(0u), Value(true)})
|
||||
.Instruction()->SetFlags(MemoryFlags{.index = 1});
|
||||
auto plan = ExtractResourcePlan(program);
|
||||
struct Reads { uint32_t count = 0; bool clean = true; } reads;
|
||||
const SrtRuntime runtime{
|
||||
.userdata = &reads,
|
||||
.read_specialization_memory = +[](void *data, uint64_t address, std::span<uint32_t> words) {
|
||||
const auto &reads = *static_cast<Reads *>(data);
|
||||
if (!reads.clean || address != 0x1000u || words.size() != 1) return false;
|
||||
words[0] = reads.count;
|
||||
return true;
|
||||
}};
|
||||
ResourceSnapshot snapshot;
|
||||
ResourceSpecialization specialization;
|
||||
for (const auto state : {Reads{0u, true}, Reads{4u, true}, Reads{0u, false}, Reads{0u, true}}) {
|
||||
reads = state;
|
||||
uint32_t condition = 99;
|
||||
Check(!SrtWalker(plan, runtime).Evaluate(plan.control_flow[3].condition, condition),
|
||||
"direct merge evaluation used missing or previous runtime reachability");
|
||||
Check(MaterializeResources(plan, runtime, snapshot, specialization),
|
||||
"partial traversal proof rejected a valid resource snapshot");
|
||||
const bool inactive = !independent_mask && reads.clean && reads.count == 0u;
|
||||
Check(snapshot.buffers[0].dwords[0] == (inactive ? 0u : 0x8000u),
|
||||
"empty traversal kept its descriptor or an uncertain traversal lost it");
|
||||
Check(specialization.buffers[0].zero_stride_oob == inactive,
|
||||
"inactive traversal did not expose dead vector reads to specialization");
|
||||
if (inactive) {
|
||||
Check(!snapshot.specialization_reads.empty(), "empty traversal lost its clean scalar read");
|
||||
for (const auto &read : snapshot.specialization_reads)
|
||||
Check(read == std::pair<uint64_t, uint64_t>{0x1000u, 4u},
|
||||
"empty traversal proof captured an unrelated memory range");
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
void TestCyclicControlPredicatesRemainUnknown() {
|
||||
using namespace Libs::Graphics::ShaderRecompiler::IR;
|
||||
using Libs::Graphics::ShaderRecompiler::CFG::TerminatorKind;
|
||||
Program program;
|
||||
program.stage = Libs::Graphics::ShaderType::Compute;
|
||||
program.srt_plan_complete = program.resource_tracking_complete = true;
|
||||
for (uint32_t i = 0; i < 4; ++i) AddValueBlock(program);
|
||||
auto &entry = *program.blocks[0];
|
||||
auto &left = *program.blocks[1];
|
||||
auto &right = *program.blocks[2];
|
||||
auto &exit = *program.blocks[3];
|
||||
const auto lane = Value(&entry.AppendNewInst(ValueOpcode::GetBuiltin,
|
||||
{Value(static_cast<uint32_t>(StageInputKind::LocalInvocationId)), Value(0u)}));
|
||||
const auto u = Value(&entry.AppendNewInst(ValueOpcode::INotEqual32, {lane, Value(0u)}));
|
||||
const auto v = Value(&entry.AppendNewInst(ValueOpcode::INotEqual32, {lane, Value(1u)}));
|
||||
auto &p = left.AppendNewInst(ValueOpcode::Phi);
|
||||
auto &q = right.AppendNewInst(ValueOpcode::Phi);
|
||||
p.SetFlags(Type::U1);
|
||||
q.SetFlags(Type::U1);
|
||||
const auto a = Value(&left.AppendNewInst(ValueOpcode::LogicalAnd, {u, Value(&p)}));
|
||||
const auto b = Value(&right.AppendNewInst(ValueOpcode::LogicalAnd, {v, Value(&q)}));
|
||||
p.AddPhiOperand(&entry, Value(true));
|
||||
p.AddPhiOperand(&right, b);
|
||||
q.AddPhiOperand(&entry, Value(true));
|
||||
q.AddPhiOperand(&left, a);
|
||||
const auto branch = [](Block &block, Value condition, Block &yes, Block &no) {
|
||||
block.condition = condition;
|
||||
block.terminator.kind = TerminatorKind::ConditionalBranch;
|
||||
block.terminator.true_block = &yes;
|
||||
block.terminator.false_block = &no;
|
||||
block.AddBranch(&yes);
|
||||
block.AddBranch(&no);
|
||||
};
|
||||
// Two loop entries have no single header whose Phi can be discarded up front.
|
||||
branch(entry, u, left, right);
|
||||
branch(left, a, right, exit);
|
||||
branch(right, b, left, exit);
|
||||
auto plan = ExtractResourcePlan(program);
|
||||
SrtWalker walker(plan, {});
|
||||
uint32_t value = 99;
|
||||
Check(!walker.Evaluate(plan.control_flow[1].condition, value) &&
|
||||
!walker.Evaluate(plan.control_flow[2].condition, value),
|
||||
"cyclic varying control acquired an unsupported uniform value");
|
||||
std::vector<uint32_t> flat;
|
||||
Check(walker.RefreshFlatBuffer(flat) && plan.visited_blocks == std::vector<uint8_t>({1, 1, 1, 1}),
|
||||
"cyclic control lost a possible branch during reachability");
|
||||
}
|
||||
|
||||
void TestMixedSamplerVariantsShareRuntimeDescriptor() {
|
||||
using namespace Libs::Graphics::ShaderRecompiler::IR;
|
||||
auto program = MixedSamplerProgram();
|
||||
@@ -853,6 +1003,9 @@ int main() {
|
||||
TestFailedMaterializationRejectsStage();
|
||||
TestSharedControlPredicates();
|
||||
TestFiniteImageRefreshReusesScalarReads();
|
||||
for (const bool independent_mask : {false, true})
|
||||
TestEmptyTraversalSkipsUnmappedDescriptor(independent_mask);
|
||||
TestCyclicControlPredicatesRemainUnknown();
|
||||
TestMixedSamplerVariantsShareRuntimeDescriptor();
|
||||
std::puts("ResourceMaterializationTests: all cases passed");
|
||||
return 0;
|
||||
|
||||
Reference in New Issue
Block a user