raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-20 10:56:58 +02:00
Code Issues Packages Projects Releases Wiki Activity

shader: bound indirect image selectors by counted scalar rows

Browse Source
This commit is contained in:
nmzik
2026-10-12 22:05:31 +02:00
parent 9a7c54da58
commit 8ee31abd9d
5 changed files with 483 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/ShaderRecompilerComputeTests.cpp
+116
View File
@@ -36176,6 +36176,120 @@ void CheckFiniteInlineSamplerPhi(VulkanHarness &vulkan) {
std::printf("[compute] %-32s ok\n", test.name);
}
void CheckCountedRowImageSelector(VulkanHarness &vulkan) {
using O = ShaderOpcode;
using F = Prospero::BufferFormat;
constexpr u32 poison = 0x345d1746u;
const auto alignment = static_cast<u32>(vulkan.RuntimeContext().GetPhysicalDeviceProperties()
.limits.minStorageBufferOffsetAlignment);
const auto slot = std::max(256u, alignment);
const u32 srt = slot, counts = slot * 2u, keys = slot * 3u, material = slot * 4u;
for (const u32 wave : {32u, 64u}) {
TestCase test;
test.name = wave == 32u ? "CountedRowImageSelector32" : "CountedRowImageSelector64";
test.initial.resize((material + 396u * 4u) / 4u, 0u);
std::fill_n(test.initial.begin(), 64u, 0xdeadbeefu);
const std::array<u32, 12> descriptors{counts, 4u << 16u, 3u, 0x14004u,
keys, 4u << 16u, 48u, 0x14004u, material, 396u << 16u, 4u, 0x14004u};
std::copy(descriptors.begin(), descriptors.end(), test.initial.begin() + srt / 4u);
test.initial[counts / 4u] = 3u;
test.initial[counts / 4u + 2u] = 1u;
std::fill_n(test.initial.begin() + keys / 4u, 48u, poison);
test.initial[keys / 4u] = 2u;
test.initial[keys / 4u + 1u] = 0x80000001u;
test.initial[keys / 4u + 2u] = 0x1890cedeu; // Dynamic +184 wraps to material byte32.
test.initial[keys / 4u + 32u] = 3u;
const auto put_image = [&](u32 offset, u32 identity, F format) {
auto at = test.initial.begin() + (material + offset) / 4u;
at[0] = identity;
at[1] = static_cast<u32>(format) << 20u;
at[3] = DstSel(4, 5, 6, 7) |
(static_cast<u32>(Prospero::ImageType::kColor2D) << 28u);
};
for (u32 key = 0; key < 4u; ++key)
put_image(key * 396u + 184u, 0x100u + key, F::k32_32_32_32Float);
put_image(32u, 0x104u, F::k32_32_32_32Float);
put_image(256u, 0x200u, F::k8SScaled);
test.expected = test.initial;
test.expected[0] = 0x3f800002u;
test.expected[1] = 0x3f800001u;
test.expected[2] = 0x3f800004u;
test.expected[32] = 0x3f800003u;
test.user_data[0] = srt;
test.user_data[50] = 256u;
test.user_data[51] = 3u << 28u;
test.has_user_data = true;
test.has_compute_info = true;
test.compute_info.wave_size = wave;
test.compute_info.threads_num[0] = test.compute_info.threads_num[1] =
test.compute_info.threads_num[2] = 1u;
test.compute_info.group_id[0] = true;
test.compute_info.workgroup_register = 44u;
test.dispatch_x = 3u;
auto &code = test.code;
code.insert(code.end(), {EncodeSmem0(0x03, 8), EncodeSmem1(0, 125),
EncodeSmem0(0x02, 16), EncodeSmem1(32, 125)});
code.push_back(EncodeSop2(0x1e, 22, 44, InlineU32(2)));
code.insert(code.end(), {EncodeSmem0(0x08, 21, 4), EncodeSmem1(0, 22)});
AppendSMovLiteral(&code, 20, 0u);
AppendVMovU32(&code, 0, 0x3f000000u);
AppendVMovU32(&code, 1, 0x3f000000u);
const auto header = code.size();
code.push_back(EncodeSopc(0x0a, 20, 21));
const auto exit_branch = code.size();
code.push_back(0u);
code.push_back(EncodeSop2(0x1e, 22, 44, InlineU32(4)));
code.push_back(EncodeSop2(0x02, 22, 22, 20));
code.push_back(EncodeSop2(0x1e, 22, 22, InlineU32(2)));
code.insert(code.end(), {EncodeSmem0(0x08, 23, 6), EncodeSmem1(0, 22)});
code.push_back(EncodeSop2(0x26, 24, 23, 255)); code.push_back(396u);
code.push_back(EncodeSop2(0x02, 24, 24, 255)); code.push_back(184u);
code.insert(code.end(), {EncodeSmem0(0x0b, 32, 8), EncodeSmem1(0, 24),
EncodeMimg0(0x27, 1), EncodeMimg1(4, 0, 8, 10)});
code.push_back(EncodeVop1(0x01, 30, 22));
AppendBufferStoreDword(&code, 4, 30);
code.push_back(EncodeSop2(0x02, 20, 20, InlineU32(1)));
code.push_back(EncodeSopp(0x02, header - code.size() - 1u));
code[exit_branch] = EncodeSopp(0x04, code.size() - exit_branch - 1u);
AppendEnd(&code);
test.opcodes = {O::S_LOAD_DWORDX8, O::S_LOAD_DWORDX4, O::S_BUFFER_LOAD_DWORD,
O::S_BUFFER_LOAD_DWORDX8, O::S_MOV_B32, O::S_LSHL_B32, O::S_ADD_I32,
O::S_MUL_I32, O::S_CMP_LT_U32, O::S_CBRANCH_SCC0, O::S_BRANCH,
O::V_MOV_B32, O::IMAGE_SAMPLE, O::BUFFER_STORE_DWORD, O::S_ENDPGM};
const auto compiled = CompileCase(test, vulkan.SubgroupSize());
Require(test.name, "exact active images", compiled.resources.images.size() == 4u &&
std::ranges::all_of(compiled.resources.images, [](const auto &image) {
return image.dwords[0] >= 0x101u && image.dwords[0] <= 0x104u;
}), "unused row entries selected malformed material fields");
for (u32 word = 0; word < 4u; ++word)
Require(test.name, "count descriptor capture",
std::ranges::any_of(compiled.resources.specialization_reads, [&](const auto &read) {
return read.first <= srt + word * 4u &&
read.first + read.second >= srt + (word + 1u) * 4u;
}), "count V# loaded from SRT escaped writable-alias capture");
for (const auto &buffer : compiled.resources.buffers) {
ShaderBufferResource descriptor{};
std::copy_n(buffer.dwords.begin(), 4u, descriptor.fields);
test.storage_buffer_ranges.emplace_back(static_cast<u32>(descriptor.Base48()),
static_cast<u32>(descriptor.GetSize()));
}
std::vector<VulkanHarness::Image> images;
for (const auto &image : compiled.resources.images)
images.push_back(vulkan.CreateImageMips(test.name, 1, 1, vk::Format::eR32G32B32A32Sfloat,
vk::ImageUsageFlagBits::eSampled, {{0x3f800000u + image.dwords[0] - 0x100u, 0u, 0u, 0u}},
4u, vk::ImageLayout::eShaderReadOnlyOptimal, vk::ImageType::e2D, vk::ImageViewType::e2D, 1));
auto sampler = vulkan.CreateSampler(test.name, 1);
auto output = vulkan.CreateStorageBuffer(test.name, test.initial, test.initial.size());
vulkan.Dispatch(test, compiled, output, nullptr, nullptr, nullptr, nullptr, sampler, images);
CompareWords(test, "active row images, wrapped keys and untouched unused slots", test.expected,
vulkan.ReadBuffer(test.name, output, test.expected.size()));
vulkan.DestroyBuffer(&output);
vulkan.Device().destroySampler(sampler);
for (auto &image : images) vulkan.DestroyImage(&image);
std::printf("[compute] %-32s ok\n", test.name);
}
}
void CheckIndirectImageKeySwitch(VulkanHarness &vulkan) {
constexpr const char *name = "IndirectImageKeySwitch";
constexpr uint32_t mapping_capacity = 1793u;
@@ -44568,6 +44682,7 @@ int main(int argc, char **argv) {
if (argc == 2 && std::strcmp(argv[1], "--indirect-image-only") == 0) {
CheckImageSamplerSpecialization();
VulkanHarness vulkan;
CheckCountedRowImageSelector(vulkan);
CheckIndirectImageKeySwitch(vulkan);
CheckFiniteInlineSamplerPhi(vulkan);
RunCase(&vulkan, ImageCubeGradientsPreserveDerivatives());
@@ -44733,6 +44848,7 @@ int main(int argc, char **argv) {
CheckIndirectBufferStore(vulkan);
CheckRuntimeBufferRecords(vulkan);
CheckComputeThreadDimensions(vulkan);
CheckCountedRowImageSelector(vulkan);
CheckIndirectImageKeySwitch(vulkan);
CheckDynamicSampleOffsets(vulkan, 32u);
CheckDynamicSampleOffsets(vulkan, 64u);