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: resolve workgroup-selected descriptor prefixes from dispatch arguments

Browse Source 
Co-authored-by: fxpw <fxpw@outlook.com>
This commit is contained in:
nmzik
2026-10-14 20:29:48 +02:00
parent 82470c7181
commit ceccd746d8
9 changed files with 358 additions and 36 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/ResourceTrackingTests.cpp
+160 -15
View File
@@ -222,7 +222,8 @@ MakeIndirectImageFixture(bool malformed, uint32_t table_stride = 32, uint32_t ta
ValueOpcode::GetBuiltin,
{Value(static_cast<uint32_t>(finite_key ? StageInputKind::WorkgroupId
: StageInputKind::GlobalInvocationId)), Value(0u)});
const auto selector = finite_key ? invocation
const auto selector = finite_key
? fixture->Emit(ValueOpcode::BitwiseXor32, {invocation, Value(1u)})
: fixture->Emit(ValueOpcode::ReadFirstLane, {invocation, Value(true)});
Value member;
if (nested) {
@@ -2699,7 +2700,7 @@ void TestUniformizedMaterialImageKeys() {
memory.fail_address = base + 16u * 16u + 8u;
memory.reads = memory.descriptor_reads = 0u;
Check(MaterializeResources(packed, runtime, snapshot, specialization) &&
memory.reads == 19u && memory.descriptor_reads == 3u && snapshot.images.size() == 3u &&
memory.reads == 4u && memory.descriptor_reads == 3u && snapshot.images.size() == 3u &&
snapshot.images[0].dwords[0] == 0x200u && snapshot.images[1].dwords[0] == 0x400u &&
snapshot.images[2].dwords[0] == 0x600u,
"packed selector did not extract before clamping or read outside its sixteen records");
@@ -2719,7 +2720,7 @@ void TestUniformizedMaterialImageKeys() {
snapshot.images.size() == 3u && snapshot.images[0].dwords[0] == 0x200u &&
snapshot.images[1].dwords[0] == 0x400u && snapshot.images[2].dwords[0] == 0x600u &&
std::ranges::find(snapshot.specialization_reads,
std::pair<uint64_t, uint64_t>{base + base_offset + immediate, 4u}) !=
std::pair<uint64_t, uint64_t>{base + base_offset + immediate, 244u}) !=
snapshot.specialization_reads.end(),
"global selector did not preserve exact base-plus-offset while scalar descriptor reads aligned the base");
}
@@ -2731,7 +2732,7 @@ void TestUniformizedMaterialImageKeys() {
memory.fail_address = base + 16u * 16u + 8u;
memory.reads = memory.descriptor_reads = 0u;
Check(MaterializeResources(merged, runtime, snapshot, specialization) &&
memory.reads == 19u && memory.descriptor_reads == 3u && snapshot.images.size() == 3u &&
memory.reads == 4u && memory.descriptor_reads == 3u && snapshot.images.size() == 3u &&
snapshot.images[0].dwords[0] == 0x200u && snapshot.images[1].dwords[0] == 0x400u &&
snapshot.images[2].dwords[0] == 0x600u,
"merged nibble selector did not preserve its sixteen-record domain and lower-half projection");
@@ -2978,12 +2979,14 @@ void TestScalarBufferSelectorDomain() {
ResourceSnapshot snapshot;
ResourceSpecialization specialization;
memory.watched_address = 0x1000u + 144u;
memory.fail_address = 0x1000u + 160u;
Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
snapshot.images.size() == 3u && memory.watched_reads == 1u &&
memory.watched_dwords == 1u && snapshot.images[0].dwords[0] == 0x100u &&
memory.watched_dwords == 9u && snapshot.images[0].dwords[0] == 0x100u &&
snapshot.images[2].dwords[0] == 0x102u,
"scalar selector read outside its wrapped domain or omitted the OOB zero key");
"scalar selector interpreted batched padding or omitted the OOB zero key");
Check(std::ranges::find(snapshot.specialization_reads,
std::pair<uint64_t, uint64_t>{0x1090u, 36u}) != snapshot.specialization_reads.end(),
"scalar selector batch escaped specialization overlap tracking");
const auto scaled_lane = make_plan(true, false, true);
memory.words[176u / 4u] = UINT32_MAX;
Check(MaterializeResources(scaled_lane, runtime, snapshot, specialization) &&
@@ -3723,7 +3726,7 @@ void TestSharedAppendSelectorBounds() {
Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
snapshot.images.size() == 1u && memory.descriptor_reads == 0u && memory.watched_reads == 0u,
"empty shared append lists probed material keys or image descriptors");
memory.fail_address = 0x10b0u;
memory.fail_address = UINT64_MAX;
for (const auto counts : {std::array{1u, 2u}, std::array{3u, 1u}, std::array{1u, 1u}, std::array{0u, 0u}}) {
memory.words[0x800u / 4u] = counts[0];
memory.words[0x804u / 4u] = counts[1];
@@ -3778,10 +3781,13 @@ void TestSharedAppendSelectorBounds() {
words[0] = memory.counts[(address - 0x40000000u) / 4u];
return true;
}
if (words.size() == 1u && address >= 0x1090u &&
(address - 0x1090u) % 160u == 0u && (address - 0x1090u) / 160u < 65536u) {
words[0] = 1u;
++memory.keys_read;
if (address >= 0x1090u && (address - 0x1090u) % 160u == 0u &&
address - 0x1090u + words.size_bytes() <= 65535u * 160u + 4u) {
std::ranges::fill(words, 0xdeadbeefu);
for (size_t word = 0; word < words.size(); word += 40u) {
words[word] = 1u;
++memory.keys_read;
}
return true;
}
if (words.size() != 8u || address != 0x50000020u) return false;
@@ -4784,7 +4790,8 @@ 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 split_offsets = false,
bool workgroup_selected = false) {
namespace CFG = Libs::Graphics::ShaderRecompiler::CFG;
auto fixture = std::make_unique<Fixture>();
auto& f = *fixture;
@@ -4796,8 +4803,8 @@ std::unique_ptr<Fixture> MakeBufferRecordFixture(uint32_t step = 1u, bool bypass
if (gpu_selected) {
std::array<Value, 4> selector_words;
for (uint32_t word = 0; word < 4u; ++word) selector_words[word] = f.UserData(4u + word);
const auto group = f.Emit(ValueOpcode::GetBuiltin,
{Value(static_cast<uint32_t>(StageInputKind::WorkgroupId)), Value(0u)});
const auto group = workgroup_selected ? f.Emit(ValueOpcode::GetBuiltin,
{Value(static_cast<uint32_t>(StageInputKind::WorkgroupId)), Value(0u)}) : lane;
key = f.Emit(ValueOpcode::ReadConstBuffer,
{f.Buffer(selector_words), f.Emit(ValueOpcode::IMul32, {group, Value(12u)})},
f.AddMemory({.kind = ResourceKind::ScalarBuffer}, 0x1bcu));
@@ -5090,6 +5097,143 @@ void TestGpuSelectedBufferRecords() {
"not a valid runtime value", "split dynamic descriptor offsets lost independent U32 wrapping");
}
void TestWorkgroupSelectedBufferRecords() {
namespace Prospero = Libs::Graphics::Prospero;
auto fixture = MakeBufferRecordFixture(1u, false, true, 0u, false, true);
auto plan = ExtractResourcePlan(fixture->program);
std::vector<uint32_t> roots;
for (uint32_t root = 0; root < plan.info.buffers.size(); ++root) {
const auto& source = plan.descriptor_sources[plan.info.buffers[root].source];
if (!source.indirect_descriptor) continue;
const auto& indirect = *source.indirect_descriptor;
Check(indirect.workgroup_axis == 0u && indirect.selector_first == Value(0u) &&
indirect.selector && indirect.selector->stride == 12u &&
indirect.table_stride == 192u && !indirect.table_scalar,
"workgroup selector lost its exact prefix or descriptor geometry");
roots.push_back(root);
}
Check(roots.size() == 2u, "workgroup selector did not retain both buffer operations");
LinearTestMemory memory;
constexpr uint32_t table = 256u, selector = 0x1000u;
memory.words.resize((selector + MaxIndirectDescriptorProbes * 12u) / 4u);
std::fill_n(memory.words.begin() + table / 4u, 384u / 4u, 0xdeadbeefu);
for (uint32_t row = 0; row < 2u; ++row) {
for (uint32_t member = 0; member < 2u; ++member) {
const auto first = (table + row * 192u + 160u + member * 16u) / 4u;
memory.words[first] = 0x100000u + row * 0x1000u + member * 0x10000u;
memory.words[first + 1u] = (member == 0u ? 8u : 1u) << 16u;
memory.words[first + 2u] = 16u;
memory.words[first + 3u] = member == 0u
? (uint32_t(Prospero::BufferFormat::k16_16_16_16Float) << 12u) | Libs::Graphics::DstSel(4, 5, 6, 7)
: 0x5204u;
}
}
const std::array<uint32_t, 4> numeric_descriptor{0x640u, 3u, 0x3ba3d70au, 0x3a23d70au};
std::copy(numeric_descriptor.begin(), numeric_descriptor.end(), memory.words.begin() + (table + 304u) / 4u);
std::array<uint32_t, 8> user_data{uint32_t(memory.base) + table, 192u << 16u, 2u, 0x5204u,
uint32_t(memory.base) + selector, 12u << 16u, MaxIndirectDescriptorProbes, 0x5204u};
memory.watched_address = memory.base;
SrtRuntime runtime{.user_data = user_data, .read_memory = ReadLinearTestMemory,
.userdata = &memory, .read_specialization_memory = ReadLinearTestMemory,
.indirect_args_address = memory.base};
ResourceSnapshot snapshot;
ResourceSpecialization specialization;
const auto materialize = [&] {
memory.reads = memory.watched_reads = 0u;
return MaterializeResources(plan, runtime, snapshot, specialization);
};
for (const auto dimensions : {std::array{2u, 1u, 1u}, std::array{0u, 1u, 1u},
std::array{2u, 0u, 1u}, std::array{2u, 1u, 0u},
std::array{1u, 1u, 1u}, std::array{2u, 1u, 1u}}) {
std::copy(dimensions.begin(), dimensions.end(), memory.words.begin());
memory.words[selector / 4u] = dimensions[0] == 1u ? 0u : 1u;
memory.words[selector / 4u + 3u] = 0u;
memory.words[selector / 4u + 6u] = 22369622u; // Would select the numeric fields at +288/+304.
const bool empty = std::ranges::find(dimensions, 0u) != dimensions.end();
Check(materialize() && memory.watched_reads == 1u && memory.watched_dwords == 3u &&
snapshot.buffers.size() == plan.info.buffers.size() + (empty ? 0u : (dimensions[0] - 1u) * 2u) &&
memory.reads == (empty ? 1u : 3u + dimensions[0] * 2u),
"workgroup prefix reread arguments, visited an inactive key or retained stale candidates");
Check(std::ranges::find(snapshot.specialization_reads, std::pair{memory.base, uint64_t{12u}}) !=
snapshot.specialization_reads.end(),
"indirect dispatch arguments escaped writable-alias tracking");
if (!empty)
Check(std::ranges::find(snapshot.specialization_reads,
std::pair{memory.base + selector, uint64_t{(dimensions[0] - 1u) * 12u + 4u}}) !=
snapshot.specialization_reads.end(),
"batched selector prefix escaped writable-alias tracking");
}
const auto valid = specialization;
memory.words[selector / 4u] = 0x40000001u; // key*192 wraps to the valid second row.
Check(materialize() && specialization == valid,
"GPU descriptor key lost its native U32 multiplication wrap");
memory.words[selector / 4u] = 22369622u;
Check(!materialize(), "an actually selected numeric descriptor was silently filtered");
memory.words[selector / 4u] = 1u;
user_data[5] = 0u;
user_data[6] = 14u; // Second key has fewer than four in-bounds bytes: SMEM returns zero.
Check(materialize() && snapshot.buffers.size() == plan.info.buffers.size() + 2u,
"partial scalar selector did not zero-fill its OOB DWORD");
user_data[6] = 0u;
Check(materialize() && snapshot.buffers.size() == plan.info.buffers.size() &&
memory.reads == 3u,
"empty selector V# did not select key zero without reading payload");
// Existing vector selectors use the V# stride; each scalar snapshot read
// independently rounds its resulting byte offset down to a DWORD.
for (const uint32_t stride : {5u, 6u}) {
for (const auto root : roots) {
auto& indirect = *plan.descriptor_sources[plan.info.buffers[root].source].indirect_descriptor;
indirect.selector->stride = 0u;
indirect.selector_first = Value(1u);
}
user_data[5] = stride << 16u;
user_data[6] = 5u;
memory.words[0] = 4u;
std::fill_n(memory.words.begin() + selector / 4u, 8u, 0xdeadbeefu);
for (uint32_t index = 0; index < 4u; ++index)
memory.words[selector / 4u + ((index + 1u) * stride) / 4u] = index & 1u;
Check(materialize() && memory.reads == 7u &&
snapshot.buffers.size() == plan.info.buffers.size() + 2u,
"batched vector selector changed independent scalar DWORD alignment");
}
for (const auto root : roots) {
auto& indirect = *plan.descriptor_sources[plan.info.buffers[root].source].indirect_descriptor;
indirect.selector->stride = 12u;
indirect.selector_first = Value(0u);
}
user_data[5] = 12u << 16u;
user_data[6] = MaxIndirectDescriptorProbes;
std::fill(memory.words.begin() + selector / 4u, memory.words.end(), 0u);
memory.words[0] = MaxIndirectDescriptorProbes;
Check(materialize() && memory.reads == 9u,
"strided selector prefix was not read in bounded contiguous batches");
memory.words[0] = MaxIndirectDescriptorProbes + 1u;
Check(!materialize() && memory.reads == 1u,
"selector cap was checked after reading its GPU payload");
memory.words[0] = 2u;
for (const auto root : roots)
plan.descriptor_sources[plan.info.buffers[root].source].indirect_descriptor->selector->offset = UINT32_MAX - 3u;
Check(!materialize() && memory.reads == 1u,
"selector offset arithmetic accepted a wrapped prefix span");
for (const auto root : roots)
plan.descriptor_sources[plan.info.buffers[root].source].indirect_descriptor->selector->offset = 0u;
runtime.indirect_args_address = 0u;
const std::array direct_counts{1u, 1u, 1u};
runtime.workgroup_counts = std::span(direct_counts).first(1u);
Check(!materialize() && memory.reads == 0u,
"workgroup prefix accepted an incomplete dispatch extent");
runtime.workgroup_counts = direct_counts;
memory.fail_address = memory.base;
Check(materialize() && memory.watched_reads == 0u,
"direct workgroup counts unnecessarily read indirect arguments");
const auto ordinary = MakeBufferRecordFixture(1u, false, true);
const auto ordinary_plan = ExtractResourcePlan(ordinary->program);
runtime.indirect_args_address = memory.base;
user_data[2] = 0u;
Check(MaterializeResources(ordinary_plan, runtime, snapshot, specialization),
"a plan without workgroup counts read the indirect argument pointer");
}
void TestGpuSelectedFormattedRecordStore() {
namespace Prospero = Libs::Graphics::Prospero;
const auto make = [](uint32_t stride = 592u, uint32_t shift = 16u, bool split = false,
@@ -7024,6 +7168,7 @@ int main() {
Run("bounded formatted buffer store", TestBoundedFormattedBufferStore);
Run("bounded buffer records", TestBoundedBufferRecords);
Run("GPU-selected buffer records", TestGpuSelectedBufferRecords);
Run("workgroup-selected buffer records", TestWorkgroupSelectedBufferRecords);
Run("GPU-selected formatted record store", TestGpuSelectedFormattedRecordStore);
Run("finite buffer selection", TestFiniteBufferSelection);
Run("packed scalar buffer selection", TestPackedScalarBufferSelection);