shader: specialize bounded pointer and packed image tables

This commit is contained in:
nmzik
2026-10-13 01:12:19 +02:00
parent c9ec82e32e
commit 1fa79d5cac
5 changed files with 219 additions and 24 deletions
+109
View File
@@ -36505,6 +36505,113 @@ void CheckCountedRowImageSelector(VulkanHarness &vulkan) {
}
}
void CheckPointerImageTable(VulkanHarness &vulkan) {
using O = ShaderOpcode;
constexpr u32 base = 256u, table_offset = 24364u;
for (const u32 variant : {0u, 1u, 2u})
for (const u32 wave : {32u, 64u}) {
const bool packed = variant == 2u;
const u32 count = variant * 2u;
TestCase test;
test.name = packed ? "PackedPointerImageTable" : "CountedPointerImageTable";
test.dispatch_x = packed ? 4u : 1u;
test.bda_mappings = {{0u, 0u}};
test.initial.resize((base + table_offset + 32u * 32u) / 4u);
std::fill_n(test.initial.begin(), test.dispatch_x, 0xdeadbeefu);
test.initial[(base + 492u) / 4u] = count;
for (u32 row = 0; row < 32u; ++row) {
auto image = MakeSampledTextureData(Prospero::BufferFormat::k32Float);
image[0] = row + 1u;
image[3] = DstSel(4, 0, 0, 1) |
(static_cast<u32>(Prospero::ImageType::kColor2D) << 28u);
std::copy_n(image.begin(), 8,
test.initial.begin() + (base + table_offset + row * 32u) / 4u);
}
test.initial[(base + 2756u) / 4u] = 1u;
test.initial[(base + 2756u + 48u) / 4u] = 31u;
constexpr std::array fields{31u, 0u, 7u, 1u};
for (u32 row = 0; row < fields.size(); ++row)
test.initial[(base + 5008u + row * 60u) / 4u] = 0xffe00abcu | (fields[row] << 12u);
test.user_data[0] = base;
const auto output = MakeStructuredStorageBufferData(0u, test.dispatch_x * 4u);
std::copy_n(output.begin(), 4, test.user_data.begin() + 48);
test.has_user_data = 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;
auto &code = test.code;
AppendVMovLiteral(&code, 0, 0u);
AppendVMovLiteral(&code, 20, 0x3f000000u);
AppendVMovLiteral(&code, 21, 0x3f000000u);
size_t loop = 0, exit = 0, skip = 0;
if (packed) {
code.insert(code.end(), {EncodeSop2(0x26, 23, 44, InlineU32(60)),
EncodeSop2(0x02, 23, 23, 255), 5008u,
EncodeSmem0(0x00, 24), EncodeSmem1(0, 23),
EncodeSop2(0x20, 24, 24, InlineU32(12)),
EncodeSop2(0x1e, 23, 24, InlineU32(5)),
EncodeSop2(0x0e, 23, 23, 255), 992u});
} else {
code.insert(code.end(), {EncodeSmem0(0x00, 21), EncodeSmem1(492, 125),
EncodeSMovB32(20, InlineU32(0))});
loop = code.size();
code.push_back(EncodeSopc(0x0a, 20, 21));
exit = code.size(); code.push_back(0u);
code.insert(code.end(), {EncodeSop2(0x26, 23, 20, InlineU32(48)),
EncodeSop2(0x02, 23, 23, 255), 2756u,
EncodeSmem0(0x00, 24), EncodeSmem1(0, 23),
EncodeSopc(0x03, 24, InlineU32(0))});
skip = code.size(); code.push_back(0u);
code.push_back(EncodeSop2(0x1e, 23, 24, InlineU32(5)));
}
code.insert(code.end(), {EncodeSop2(0x02, 23, 23, 255), table_offset,
EncodeSmem0(0x03, 8), EncodeSmem1(0, 23),
EncodeMimg0(0x27, 1), EncodeMimg1(4, 20, 2, 10),
EncodeVop2(0x03, 0, Vgpr(4), 0)});
if (!packed) {
code[skip] = EncodeSopp(0x04, code.size() - skip - 1u);
code.push_back(EncodeSop2(0x02, 20, 20, InlineU32(1)));
code.push_back(EncodeSopp(0x02, loop - code.size() - 1u));
code[exit] = EncodeSopp(0x04, code.size() - exit - 1u);
}
code.push_back(EncodeSop2(0x1e, 25, 44, InlineU32(2)));
code.push_back(EncodeVop1(0x01, 30, 25));
AppendBufferStoreDword(&code, 0, 30);
AppendEnd(&code);
test.opcodes = {O::S_LOAD_DWORD, O::S_MUL_I32, O::S_ADD_I32, O::S_LSHL_B32,
O::S_LOAD_DWORDX8, O::IMAGE_SAMPLE, O::V_ADD_F32, O::V_MOV_B32,
O::BUFFER_STORE_DWORD, O::S_ENDPGM};
if (packed) test.opcodes.insert(test.opcodes.end(), {O::S_LSHR_B32, O::S_AND_B32});
else test.opcodes.insert(test.opcodes.end(), {O::S_MOV_B32, O::S_CMP_LT_U32,
O::S_CMP_GE_I32, O::S_CBRANCH_SCC0, O::S_BRANCH});
const auto compiled = CompileCase(test, vulkan.SubgroupSize());
Require(test.name, "bounded image candidates",
packed ? compiled.resources.images.size() == 32u
: compiled.resources.images.size() == std::max(count, 1u),
"descriptor planning escaped its selector domain");
std::vector<VulkanHarness::Image> images;
for (const auto &descriptor : compiled.resources.images)
images.push_back(vulkan.CreateImage2D(test.name, 1, 1, vk::Format::eR32Sfloat,
vk::ImageUsageFlagBits::eSampled,
{std::bit_cast<u32>(static_cast<float>(descriptor.dwords[0]))}, 1,
vk::ImageLayout::eShaderReadOnlyOptimal));
auto buffer = vulkan.CreateStorageBuffer(test.name, test.initial, test.initial.size(), true);
auto sampler = vulkan.CreateSampler(test.name, 1);
vulkan.Dispatch(test, compiled, buffer, nullptr, nullptr, nullptr, nullptr, sampler, images);
std::vector<u32> expected(test.dispatch_x, std::bit_cast<u32>(count == 0u ? 0.0f : 34.0f));
if (packed) for (u32 row = 0; row < fields.size(); ++row)
expected[row] = std::bit_cast<u32>(static_cast<float>(fields[row] + 1u));
CompareWords(test, "bounded pointer selectors and packed fields", expected,
vulkan.ReadBuffer(test.name, buffer, expected.size()));
vulkan.DestroyBuffer(&buffer);
vulkan.Device().destroySampler(sampler);
for (auto &image : images) vulkan.DestroyImage(&image);
std::printf("[compute] %s wave=%u count=%u ok\n", test.name, wave, count);
}
}
void CheckGuardedRecordImageTable(VulkanHarness &vulkan) {
using O = ShaderOpcode;
constexpr u32 table = 256u;
@@ -45134,6 +45241,7 @@ int main(int argc, char **argv) {
CheckImageSamplerSpecialization();
VulkanHarness vulkan;
CheckCountedRowImageSelector(vulkan);
CheckPointerImageTable(vulkan);
CheckGuardedRecordImageTable(vulkan);
CheckIndirectImageKeySwitch(vulkan);
CheckFiniteInlineSamplerPhi(vulkan);
@@ -45301,6 +45409,7 @@ int main(int argc, char **argv) {
CheckRuntimeBufferRecords(vulkan);
CheckComputeThreadDimensions(vulkan);
CheckCountedRowImageSelector(vulkan);
CheckPointerImageTable(vulkan);
CheckGuardedRecordImageTable(vulkan);
CheckIndirectImageKeySwitch(vulkan);
CheckDynamicSampleOffsets(vulkan, 32u);