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: bound workgroup selector prefixes from runtime starting indices

Browse Source
This commit is contained in:
nmzik
2026-10-14 20:39:25 +02:00
parent ceccd746d8
commit 5313181ecf
4 changed files with 252 additions and 114 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/ResourceTrackingTests.cpp
+106 -2
View File
@@ -5236,8 +5236,9 @@ void TestWorkgroupSelectedBufferRecords() {
void TestGpuSelectedFormattedRecordStore() {
namespace Prospero = Libs::Graphics::Prospero;
enum class Prefix { None, Forward, Reversed, Lane, TwoAxes };
const auto make = [](uint32_t stride = 592u, uint32_t shift = 16u, bool split = false,
bool wide = false) {
bool wide = false, Prefix prefix = Prefix::None) {
auto fixture = std::make_unique<Fixture>();
auto& f = *fixture;
std::array<Value, 4> table_words, selector_words;
@@ -5247,8 +5248,20 @@ void TestGpuSelectedFormattedRecordStore() {
}
const auto table = wide ? f.Address(table_words[0], table_words[1]) : f.Buffer(table_words);
const auto lane = f.Emit(ValueOpcode::LaneId);
auto selector_index = lane;
if (prefix != Prefix::None) {
const auto group = f.Emit(ValueOpcode::GetBuiltin,
{Value(static_cast<uint32_t>(StageInputKind::WorkgroupId)), Value(0u)});
const auto first = prefix == Prefix::Lane ? lane : prefix == Prefix::TwoAxes
? f.Emit(ValueOpcode::GetBuiltin,
{Value(static_cast<uint32_t>(StageInputKind::WorkgroupId)), Value(1u)})
: f.Emit(ValueOpcode::LoadAddressU32, {table, Value(0u), Value(0u), Value(true)},
f.AddMemory({.kind = ResourceKind::ScalarAddress, .offset = 0x234u}, 0x1088u));
selector_index = f.Emit(ValueOpcode::IAdd32,
{prefix == Prefix::Reversed ? group : first, prefix == Prefix::Reversed ? first : group});
}
const auto packed = wide ? f.Emit(ValueOpcode::ReadConstBuffer,
{f.Buffer(selector_words), f.Emit(ValueOpcode::IMul32, {lane, Value(8u)})},
{f.Buffer(selector_words), f.Emit(ValueOpcode::IMul32, {selector_index, Value(8u)})},
f.AddMemory({.kind = ResourceKind::ScalarBuffer, .offset = 4u}, 0x1c4cu)) :
f.Emit(ValueOpcode::LoadBufferU32,
{f.Buffer(selector_words), lane, Value(0u), Value(0u), Value(true)},
@@ -5514,6 +5527,97 @@ void TestGpuSelectedFormattedRecordStore() {
for (const auto [shift, split] : {std::pair{0u, false}, std::pair{16u, true}})
CheckFatal([&] { make(16u, shift, split, true); }, "not a valid runtime value",
"finite RGBA16 store accepted an unbounded key or independently wrapping DWORDs");
struct PrefixMemory : LinearTestMemory { uint32_t read_attempts = 0u; } prefix_memory;
const auto read_prefix = [](void* userdata, uint64_t address, std::span<uint32_t> words) {
auto* memory = static_cast<PrefixMemory*>(userdata);
++memory->read_attempts;
return ReadLinearTestMemory(memory, address, words);
};
std::ranges::fill(prefix_memory.words, 0xdeadbeefu);
constexpr uint32_t selector_offset = 0x1000u;
constexpr uint32_t first_offset = table_offset - 48u + 0x234u;
for (uint32_t row = 0; row < 2u; ++row) {
const std::array descriptor{0x100000u + row * 256u, 8u << 16u, 20u, rgba};
std::copy(descriptor.begin(), descriptor.end(),
prefix_memory.words.begin() + table_offset / 4u + row * 4u);
}
const std::array numeric_descriptor{0xbf8d214eu, 0xbf78904bu, 0x3e930d39u, 0x33d6bf95u};
std::copy(numeric_descriptor.begin(), numeric_descriptor.end(),
prefix_memory.words.begin() + table_offset / 4u + 4u * 4u);
for (uint32_t index = 0; index < 12u; ++index)
prefix_memory.words[selector_offset / 4u + index * 2u + 1u] =
((index < 4u ? index & 1u : index < 6u ? 1u : 4u) << 16u) | 0xbeefu;
std::array<uint32_t, 8> prefix_data{uint32_t(prefix_memory.base) + table_offset - 48u, 0u, 0u, 0u,
uint32_t(prefix_memory.base) + selector_offset, 8u << 16u, 12u, 0x5204u};
std::array counts{2u, 1u, 1u};
SrtRuntime prefix_runtime{.user_data = prefix_data, .read_memory = read_prefix,
.userdata = &prefix_memory, .read_specialization_memory = read_prefix,
.workgroup_counts = counts};
prefix_memory.watched_address = prefix_memory.base + selector_offset + 4u;
for (const auto prefix : {Prefix::Forward, Prefix::Reversed}) {
const auto selected = make(16u, 16u, false, true, prefix);
auto selected_plan = ExtractResourcePlan(selected->program);
auto& indirect = *selected_plan.descriptor_sources[
selected_plan.info.buffers[wide_root].source].indirect_descriptor;
Check(indirect.workgroup_axis == 0u && !indirect.selector_first.IsImmediate() &&
indirect.key_count.IsEmpty() && indirect.selector && indirect.selector->stride == 8u &&
indirect.selector->offset == 4u && indirect.table_immediate == 48u,
"runtime-first selector lost its workgroup prefix or captured table geometry");
const auto refresh = [&] {
prefix_memory.watched_reads = 0u;
prefix_memory.read_attempts = 0u;
return MaterializeResources(selected_plan, prefix_runtime, snapshot, specialization);
};
for (const uint32_t first : {0u, 2u, 4u, 0u}) {
prefix_memory.words[first_offset / 4u] = first;
Check(refresh() && snapshot.buffers.size() == selected_plan.info.buffers.size() + (first == 4u ? 0u : 1u) &&
snapshot.buffers[wide_root].dwords[0] == 0x100000u + (first == 4u ? 256u : 0u),
"runtime-first refresh froze a prior selector range or enumerated unrelated numeric rows");
for (const auto read : {std::pair{prefix_memory.base + first_offset, uint64_t{4u}},
std::pair{prefix_memory.base + selector_offset + first * 8u + 4u, uint64_t{12u}}})
Check(std::ranges::find(snapshot.specialization_reads, read) != snapshot.specialization_reads.end(),
"runtime first or selected prefix escaped specialization read capture");
}
prefix_memory.fail_address = prefix_memory.base + first_offset;
Check(!refresh() && prefix_memory.read_attempts == 1u,
"unreadable runtime first reached the selector payload");
prefix_memory.fail_address = UINT64_MAX;
prefix_memory.words[first_offset / 4u] = 6u;
Check(!refresh(), "an actually selected invalid numeric descriptor was silently omitted");
for (const uint32_t first : {UINT32_MAX, 0x20000000u}) {
prefix_memory.words[first_offset / 4u] = first;
Check(!refresh() && prefix_memory.read_attempts == 1u,
"wrapped first-plus-group or scaled prefix read selector payload");
}
prefix_memory.words[first_offset / 4u] = UINT32_MAX;
counts[0] = 7u;
indirect.selector->stride = 0xfffffffcu;
Check(!refresh() && prefix_memory.read_attempts == 1u,
"wrapped first-plus-group overflowed the host prefix product before rejection");
indirect.selector->stride = 8u;
counts[0] = 1u;
prefix_memory.words[first_offset / 4u] = 0x1fffffffu;
Check(refresh() && prefix_memory.watched_reads == 0u && snapshot.buffers[wide_root].dwords[0] == 0x100000u,
"last nonwrapping selector DWORD lost scalar OOB key zero");
prefix_memory.words[first_offset / 4u] = UINT32_MAX;
for (const auto empty : {std::array{0u, 1u, 1u}, std::array{2u, 0u, 1u}, std::array{2u, 1u, 0u}}) {
counts = empty;
Check(refresh() && prefix_memory.watched_reads == 0u &&
snapshot.buffers.size() == selected_plan.info.buffers.size() &&
snapshot.buffers[wide_root].dwords[0] == 0u,
"empty dispatch evaluated an unreachable wrapped selector prefix");
}
counts = {2u, 1u, 1u};
}
for (const auto prefix : {Prefix::Lane, Prefix::TwoAxes}) {
const auto rejected = make(16u, 16u, false, true, prefix);
const auto& indirect = rejected->program.descriptor_sources[
rejected->program.info.buffers[wide_root].source].indirect_descriptor;
Check(indirect && indirect->workgroup_axis == UINT32_MAX && !indirect->selector &&
indirect->key_count == Value(32u),
"lane-varying or multiple-axis first incorrectly narrowed the finite descriptor domain");
}
}
void TestFiniteBufferSelection() {