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: admit bounded multi-component formatted buffer loads

Browse Source 
Co-authored-by: fxpw <fxpw@outlook.com>
This commit is contained in:
nmzik
2026-10-14 22:04:55 +02:00
parent dcc642551f
commit 6455546f26
3 changed files with 89 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/ResourceTrackingTests.cpp
+26 -11
View File
@@ -4791,7 +4791,8 @@ void TestBoundedFormattedBufferStore() {
std::unique_ptr<Fixture> MakeBufferRecordFixture(uint32_t step = 1u, bool bypass = false,
bool gpu_selected = false, uint32_t bias = 0u,
bool split_offsets = false,
bool workgroup_selected = false) {
bool workgroup_selected = false,
uint32_t d32_components = 0u) {
namespace CFG = Libs::Graphics::ShaderRecompiler::CFG;
auto fixture = std::make_unique<Fixture>();
auto& f = *fixture;
@@ -4851,12 +4852,21 @@ std::unique_ptr<Fixture> MakeBufferRecordFixture(uint32_t step = 1u, bool bypass
for (uint32_t use = 0; use < (gpu_selected ? 2u : formatted ? 112u : 16u); ++use) {
const auto handle = f.Buffer(words);
if (formatted) {
const auto loaded = f.Emit(ValueOpcode::LoadBufferU32x2,
const auto opcode = d32_components == 3u ? ValueOpcode::LoadBufferU32x3
: d32_components == 4u ? ValueOpcode::LoadBufferU32x4
: ValueOpcode::LoadBufferU32x2;
const auto extract = d32_components == 3u ? ValueOpcode::CompositeExtractU32x3
: d32_components == 4u ? ValueOpcode::CompositeExtractU32x4
: ValueOpcode::CompositeExtractU32x2;
const auto loaded = f.Emit(opcode,
{handle, lane, Value(0u), Value(0u), Value(true)},
f.AddMemory({.kind = ResourceKind::Buffer, .data_dwords = 2u, .data_bits = 16u,
.component_count = 4u, .formatted = true, .idxen = true}, 0x1914u));
f.AddMemory({.kind = ResourceKind::Buffer,
.data_dwords = d32_components == 0u ? 2u : d32_components,
.data_bits = d32_components == 0u ? 16u : 32u,
.component_count = d32_components == 0u ? 4u : d32_components,
.formatted = true, .idxen = true}, 0x1914u));
f.Emit(ValueOpcode::ReferenceU32,
{f.Emit(ValueOpcode::CompositeExtractU32x2, {loaded, Value(0u)})});
{f.Emit(extract, {loaded, Value(0u)})});
} else {
f.Emit(ValueOpcode::StoreBufferU8,
{handle, lane, Value(0u), Value(0u), Value(uint8_t{42u}), Value(true)},
@@ -4868,10 +4878,13 @@ std::unique_ptr<Fixture> MakeBufferRecordFixture(uint32_t step = 1u, bool bypass
return fixture;
}
void TestBoundedBufferRecords() {
void TestBoundedBufferRecords(uint32_t d32_components) {
namespace Prospero = Libs::Graphics::Prospero;
using Libs::Graphics::DstSel;
auto fixture = MakeBufferRecordFixture();
auto fixture = MakeBufferRecordFixture(1u, false, false, 0u, false, false, d32_components);
const auto load_opcode = d32_components == 3u ? ValueOpcode::LoadBufferU32x3
: d32_components == 4u ? ValueOpcode::LoadBufferU32x4
: ValueOpcode::LoadBufferU32x2;
const auto plan = ExtractResourcePlan(fixture->program);
Check(plan.info.buffers.size() == 2u, "record descriptors did not share their resource roots");
uint32_t selections = 0u;
@@ -4884,7 +4897,7 @@ void TestBoundedBufferRecords() {
"record selection moved outside its loop or lost its buffer root");
selected[inst.Flags<uint32_t>()] = &inst;
}
if (inst.GetOpcode() == ValueOpcode::LoadBufferU32x2 || inst.GetOpcode() == ValueOpcode::StoreBufferU8) {
if (inst.GetOpcode() == load_opcode || inst.GetOpcode() == ValueOpcode::StoreBufferU8) {
const auto root = fixture->program.memory_info[inst.Flags<MemoryFlags>().index].resource;
Check(inst.Arg(0).Instruction()->Arg(0).Instruction() == selected[root],
"record operation did not use its dominating shared selection");
@@ -4957,7 +4970,7 @@ void TestBoundedBufferRecords() {
ApplyResourceSpecialization(fixture->program, specialization);
uint32_t loads = 0u;
for (const auto* block : fixture->program.blocks)
for (const auto& inst : *block) loads += inst.GetOpcode() == ValueOpcode::LoadBufferU32x2;
for (const auto& inst : *block) loads += inst.GetOpcode() == load_opcode;
Check(loads == 112u, "empty root incorrectly folded nonempty formatted candidates");
memory.words[first] = first_address;
memory.words[first + 2u] = 16u;
@@ -4973,7 +4986,7 @@ void TestBoundedBufferRecords() {
memory.words[0] = MaxIndirectDescriptorProbes + 1u;
Check(!materialize(), "record table accepted an unbounded candidate count");
for (const auto [step, bypass] : {std::pair{2u, false}, std::pair{1u, true}}) {
const auto unbounded = MakeBufferRecordFixture(step, bypass);
const auto unbounded = MakeBufferRecordFixture(step, bypass, false, 0u, false, false, d32_components);
for (const auto& buffer : unbounded->program.info.buffers) {
const auto& source = unbounded->program.descriptor_sources[buffer.source];
if (source.indirect_descriptor)
@@ -7276,7 +7289,9 @@ int main() {
Run("dynamic SRT", TestDynamicSrtReadRemainsExplicit);
Run("writable descriptor phi", TestWritableDescriptorPhi);
Run("bounded formatted buffer store", TestBoundedFormattedBufferStore);
Run("bounded buffer records", TestBoundedBufferRecords);
Run("bounded buffer records", [] {
for (const uint32_t components : {0u, 2u, 3u, 4u}) TestBoundedBufferRecords(components);
});
Run("GPU-selected buffer records", TestGpuSelectedBufferRecords);
Run("workgroup-selected buffer records", TestWorkgroupSelectedBufferRecords);
Run("GPU-selected formatted record store", TestGpuSelectedFormattedRecordStore);