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 waterfall buffer descriptors with same-lane guards

Browse Source 
Co-authored-by: fxpw <fxpw@outlook.com>
This commit is contained in:
nmzik
2026-10-15 12:14:28 +02:00
parent b4406b0467
commit 232e2f2593
3 changed files with 255 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/ShaderRecompilerComputeTests.cpp
+110
View File
@@ -37982,6 +37982,114 @@ void CheckGpuSelectedRecordBufferAccess(VulkanHarness &vulkan) {
std::printf("[compute] %-32s ok\n", "GpuSelectedRecordBufferAccess");
}
void CheckGuardedWaterfallBufferRecords(VulkanHarness &vulkan) {
const u32 alignment = vulkan.RuntimeContext().GetPhysicalDeviceProperties()
.limits.minStorageBufferOffsetAlignment;
const u32 slot = (4096u + alignment - 1u) / alignment * alignment;
constexpr u32 selectors = 256u, table = 512u;
for (const auto [wave, upper_only] : {std::pair{32u, false}, std::pair{64u, false},
std::pair{64u, true}}) {
TestCase test;
test.name = upper_only ? "GuardedWaterfallRecordsUpper64" : wave == 32u
? "GuardedWaterfallRecords32" : "GuardedWaterfallRecords64";
test.initial.assign((slot * 3u + wave * 16u + 16u) / 4u, 0xdeadbeefu);
std::fill_n(test.initial.begin() + table / 4u, 976u / 4u, 0u);
constexpr std::array poison{0x012b6890u, 0x501502f9u, 0x493ba30eu, 0x3ea2f352u};
std::copy(poison.begin(), poison.end(), test.initial.begin() + (table + 480u) / 4u);
test.has_user_data = test.has_compute_info = true;
const std::array<u32, 8> descriptors{table, 488u << 16u, 2u, 0x5204u,
selectors, 4u << 16u, wave, 0x5204u};
std::copy(descriptors.begin(), descriptors.end(), test.user_data.begin());
const std::array output{slot * 3u, 16u << 16u, wave, 0x5204u};
std::copy(output.begin(), output.end(), test.user_data.begin() + 16u);
for (u32 row = 0; row < 2u; ++row) {
const std::array descriptor{slot * (row + 1u), 12u << 16u, wave,
(u32(Prospero::BufferFormat::k32_32_32Float) << 12u) | DstSel(4, 5, 6, 1)};
std::copy(descriptor.begin(), descriptor.end(),
test.initial.begin() + (table + row * 488u + 216u) / 4u);
for (u32 lane = 0; lane < wave; ++lane)
for (u32 component = 0; component < 3u; ++component)
test.initial[slot * (row + 1u) / 4u + lane * 3u + component] =
0x3f800000u + (row << 20u) + (lane << 12u) + (component << 8u);
}
test.compute_info.wave_size = wave;
test.compute_info.threads_num[0] = wave;
test.compute_info.threads_num[1] = test.compute_info.threads_num[2] = 1u;
test.compute_info.thread_ids_num = 1;
test.compute_info.lds_size_dwords = 0;
auto &code = test.code;
code.push_back(EncodeVop1(0x01, 30, Vgpr(0)));
if (upper_only) {
AppendSMovLiteral(&code, 126, 0u);
AppendSMovLiteral(&code, 127, UINT32_MAX);
}
code.insert(code.end(), {EncodeSop1(0x04, 20, 126),
EncodeMubuf0(0x0c, 0, true, false), EncodeMubuf1(5, 1, 30), EncodeSopp(0x0c, 0),
EncodeVopc(0xc3, 2, 5), EncodeSop1(0x37, 52, 106)}); // Exclude N <= local key.
const auto empty = code.size();
code.push_back(0u);
code.push_back(EncodeSop1(0x04, 54, 126));
const auto loop = code.size();
code.push_back(EncodeSop1(0x14, 34, 54));
AppendVop3(&code, 0x360, 32, Vgpr(5), 34);
code.insert(code.end(), {EncodeVopc(0xc2, 32, 5), EncodeSop1(0x24, 58, 106)});
const auto skip = code.size();
code.push_back(0u);
code.insert(code.end(), {EncodeSop2(0x26, 33, 32, 255), 488u,
EncodeSmem0(0x0a, 12, 0), EncodeSmem1(216u, 33), EncodeSopp(0x0c, 0),
EncodeMubuf0(0x02, 0, true, false, true), EncodeMubuf1(6, 3, 30),
EncodeSopp(0x0c, 0), EncodeMubuf0(0x1f, 0, true, false), EncodeMubuf1(6, 4, 30)});
code[skip] = EncodeSopp(0x08, code.size() - skip - 1u);
code.insert(code.end(), {EncodeSop2(0x15, 54, 54, 106), EncodeSop1(0x04, 126, 58)});
code.push_back(EncodeSopp(0x05, static_cast<int16_t>(loop - code.size() - 1u)));
code[empty] = EncodeSopp(0x08, code.size() - empty - 1u);
code.push_back(EncodeSop1(0x04, 126, 20));
AppendVMovLiteral(&code, 9, 0x76543210u);
code.insert(code.end(), {EncodeMubuf0(0x1c, 12u, true, false), EncodeMubuf1(9, 4, 30)});
AppendEnd(&code);
test.required_spirv = {"OpLoopMerge", "OpGroupNonUniformShuffle", "Volatile"};
test.forbidden_spirv = {"load_indirect_buffer_", "indirect_buffer_format_geometry"};
const auto compiled = CompileCase(test, vulkan.SubgroupSize());
for (const auto &descriptor : compiled.resources.buffers) {
ShaderBufferResource resource{};
std::copy_n(descriptor.dwords.begin(), 4u, resource.fields);
test.storage_buffer_ranges.emplace_back(resource.Base48(), std::max(4u, u32(resource.GetSize())));
}
TestCase producer;
producer.name = "WaterfallRecordKeyProducer";
producer.initial = test.initial;
producer.has_compute_info = true;
producer.compute_info.wave_size = 32u;
producer.compute_info.threads_num[0] = producer.compute_info.threads_num[1] =
producer.compute_info.threads_num[2] = 1u;
producer.compute_info.lds_size_dwords = 0;
test.expected = test.initial;
for (u32 lane = 0; lane < wave; ++lane) {
// This excluded key would wrap to the unsupported descriptor at byte480.
const u32 key = lane % 3u == 2u ? 0x029f79b5u : lane % 3u;
AppendVMovLiteral(&producer.code, 1, key);
AppendVMovLiteral(&producer.code, 2, selectors + lane * 4u);
AppendBufferStoreDword(&producer.code, 1, 2);
test.expected[selectors / 4u + lane] = key;
if (upper_only && lane < 32u) continue;
const u32 out = slot * 3u / 4u + lane * 4u;
if (key < 2u)
std::copy_n(test.initial.begin() + slot * (key + 1u) / 4u + lane * 3u, 3u,
test.expected.begin() + out);
test.expected[out + 3u] = 0x76543210u;
}
AppendEnd(&producer.code);
const auto producer_shader = CompileCase(producer, vulkan.SubgroupSize());
auto buffer = vulkan.CreateStorageBuffer(test.name, test.initial, test.initial.size());
vulkan.Dispatch(producer, producer_shader, buffer);
vulkan.Dispatch(test, compiled, buffer);
CompareWords(test, "same-lane bound, GPU keys, wrapped poison, EXEC restoration and canaries",
test.expected, vulkan.ReadBuffer(test.name, buffer, test.expected.size()));
vulkan.DestroyBuffer(&buffer);
}
std::printf("[compute] %-32s ok\n", "GuardedWaterfallBufferRecords");
}
void CheckGpuSelectedFormattedRecordStore(VulkanHarness &vulkan) {
constexpr u32 selectors = 128, table = 256, stride = 592, member = 448;
const u32 alignment = vulkan.RuntimeContext().GetPhysicalDeviceProperties()
@@ -47491,6 +47599,7 @@ int main(int argc, char **argv) {
CheckIndirectBufferStore(vulkan);
CheckBoundedRecordBufferAccess(vulkan);
CheckGpuSelectedRecordBufferAccess(vulkan);
CheckGuardedWaterfallBufferRecords(vulkan);
CheckGpuSelectedFormattedRecordStore(vulkan);
CheckRuntimeBufferRecords(vulkan);
CheckSharedIndirectBufferLoads(vulkan);
@@ -48198,6 +48307,7 @@ int main(int argc, char **argv) {
CheckIndirectBufferStore(vulkan);
CheckBoundedRecordBufferAccess(vulkan);
CheckGpuSelectedRecordBufferAccess(vulkan);
CheckGuardedWaterfallBufferRecords(vulkan);
CheckGpuSelectedFormattedRecordStore(vulkan);
CheckFiniteBufferSelection(vulkan);
CheckRuntimeBufferRecords(vulkan);