shader: share nested pointer-table image and sampler selection

Co-authored-by: fxpw <fxpw@outlook.com>
This commit is contained in:
nmzik
2026-10-14 16:40:15 +02:00
parent 3a925efddf
commit 487bbf95fd
7 changed files with 484 additions and 27 deletions
+111
View File
@@ -37549,6 +37549,115 @@ void CheckPointerImageTable(VulkanHarness &vulkan) {
}
}
void CheckNestedPointerImageTable(VulkanHarness &vulkan) {
using O = ShaderOpcode;
constexpr u32 keys_address = 128, table = 256, stride = 592, objects = 4096;
constexpr std::array keys{5u, 0u, 3u, 1u, 4u, 2u};
for (const u32 wave : {32u, 64u}) {
TestCase test;
test.name = "NestedPointerImageSamplerTable";
test.initial.resize((objects + keys.size() * 0x430u) / 4u, 0);
std::fill_n(test.initial.begin(), keys.size(), 0xdeadbeefu);
std::fill_n(test.initial.begin() + keys_address / 4u, keys.size(), UINT32_MAX);
for (u32 row = 0; row < keys.size(); ++row) {
const u32 object = objects + row * 0x430u;
test.initial[(table + row * stride) / 4u] = object;
auto image = MakeSampledTextureData(Prospero::BufferFormat::k32Float);
image[0] = 10u + row;
image[1] |= 1u << 30u; // Two texels: zero, then the row's value.
image[3] = DstSel(4, 0, 0, 1) |
(static_cast<u32>(Prospero::ImageType::kColor2D) << 28u);
std::copy_n(image.begin(), 8, test.initial.begin() + (object + 0x3d0u) / 4u);
const std::array sampler{0u, 0x00fff000u, (row & 1u) ? 0x0a500000u : 0u, 0u};
std::copy(sampler.begin(), sampler.end(),
test.initial.begin() + (object + 0x3f8u) / 4u);
}
auto records = MakeStructuredStorageBufferData(stride, keys.size());
records[0] = table;
std::copy_n(records.begin(), 4, test.user_data.begin() + 32);
auto key_buffer = MakeStructuredStorageBufferData(0, keys.size() * 4u);
key_buffer[0] = keys_address;
std::copy_n(key_buffer.begin(), 4, test.user_data.begin() + 4);
const auto output = MakeStructuredStorageBufferData(0, keys.size() * 4u);
std::copy_n(output.begin(), 4, test.user_data.begin() + 48);
test.bda_mappings = {{0u, 0u}};
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] = 1;
test.compute_info.group_id[0] = true;
test.compute_info.workgroup_register = 44;
test.dispatch_x = keys.size();
auto &code = test.code;
code.push_back(EncodeSop2(0x1e, 23, 44, InlineU32(2)));
code.push_back(EncodeVop1(0x01, 30, 23));
code.insert(code.end(), {EncodeMubuf0(0x0c), EncodeMubuf1(4, 1, 30),
EncodeVop2(0x16, 6, InlineU32(16), 4)});
AppendVop3(&code, 0x360, 22, Vgpr(6), InlineU32(0));
code.insert(code.end(), {EncodeSop2(0x26, 23, 22, 255), stride,
EncodeSmem0(0x09, 20, 16), EncodeSmem1(0, 23),
EncodeSmem0(0x03, 8, 10), EncodeSmem1(0x3d0, 125),
EncodeSmem0(0x02, 16, 10), EncodeSmem1(0x3f8, 125)});
AppendVMovLiteral(&code, 20, std::bit_cast<u32>(0.5f));
AppendVMovLiteral(&code, 21, std::bit_cast<u32>(0.5f));
code.insert(code.end(), {EncodeMimg0(0x27, 1), EncodeMimg1(5, 20, 2, 4)});
AppendBufferStoreDword(&code, 5, 30);
AppendEnd(&code);
test.opcodes = {O::S_LSHL_B32, O::V_MOV_B32, O::BUFFER_LOAD_DWORD,
O::V_LSHRREV_B32, O::V_READLANE_B32, O::S_MUL_I32,
O::S_BUFFER_LOAD_DWORDX2, O::S_LOAD_DWORDX8, O::S_LOAD_DWORDX4,
O::IMAGE_SAMPLE, O::BUFFER_STORE_DWORD, O::S_ENDPGM};
const auto compiled = CompileCase(test, vulkan.SubgroupSize());
for (const auto &value : compiled.resources.buffers) {
ShaderBufferResource descriptor{};
std::copy_n(value.dwords.begin(), 4, descriptor.fields);
test.storage_buffer_ranges.emplace_back(descriptor.Base48(),
std::max(4u, u32(descriptor.GetSize())));
}
std::vector<VulkanHarness::Image> images;
for (const auto &descriptor : compiled.resources.images)
images.push_back(vulkan.CreateImage2D(test.name, 2, 1, vk::Format::eR32Sfloat,
vk::ImageUsageFlagBits::eSampled,
{0u, std::bit_cast<u32>(float(descriptor.dwords[0]))}, 1,
vk::ImageLayout::eShaderReadOnlyOptimal));
std::vector<vk::Sampler> samplers;
for (const auto &resource : compiled.program.info.samplers) {
ShaderSamplerResource descriptor{};
std::copy_n(compiled.resources.samplers[resource.snapshot_index].dwords.begin(),
4, descriptor.fields);
samplers.push_back(vulkan.CreateSampler(test.name, 1,
descriptor.XyMagFilter() == 1 ? vk::Filter::eLinear : vk::Filter::eNearest));
}
TestCase producer;
producer.name = "NestedPointerSelectorProducer";
producer.initial = test.initial;
producer.has_compute_info = true;
producer.compute_info.threads_num[0] = producer.compute_info.threads_num[1] =
producer.compute_info.threads_num[2] = 1;
producer.compute_info.wave_size = 32;
for (u32 group = 0; group < keys.size(); ++group) {
AppendVMovLiteral(&producer.code, 1, (keys[group] << 16u) | 0xbeefu);
AppendVMovLiteral(&producer.code, 2, keys_address + group * 4u);
AppendBufferStoreDword(&producer.code, 1, 2);
}
AppendEnd(&producer.code);
const auto producer_shader = CompileCase(producer, vulkan.SubgroupSize());
auto buffer = vulkan.CreateStorageBuffer(test.name, test.initial, test.initial.size(), true);
vulkan.Dispatch(producer, producer_shader, buffer);
vulkan.Dispatch(test, compiled, buffer, nullptr, nullptr, nullptr, nullptr,
nullptr, images, samplers);
std::vector<u32> expected;
for (const auto key : keys)
expected.push_back(std::bit_cast<u32>(float(10u + key) * ((key & 1u) ? 0.5f : 1.0f)));
CompareWords(test, "GPU-selected nested image and sampler", expected,
vulkan.ReadBuffer(test.name, buffer, expected.size()));
vulkan.DestroyBuffer(&buffer);
for (const auto sampler : samplers) vulkan.Device().destroySampler(sampler);
for (auto &image : images) vulkan.DestroyImage(&image);
std::printf("[compute] %s wave=%u ok\n", test.name, wave);
}
}
void CheckGuardedRecordImageTable(VulkanHarness &vulkan) {
using O = ShaderOpcode;
constexpr u32 table = 256u;
@@ -46364,6 +46473,7 @@ int main(int argc, char **argv) {
VulkanHarness vulkan;
CheckCountedRowImageSelector(vulkan);
CheckPointerImageTable(vulkan);
CheckNestedPointerImageTable(vulkan);
CheckGuardedRecordImageTable(vulkan);
CheckIndirectImageKeySwitch(vulkan);
CheckFiniteInlineSamplerPhi(vulkan);
@@ -46536,6 +46646,7 @@ int main(int argc, char **argv) {
CheckComputeThreadDimensions(vulkan);
CheckCountedRowImageSelector(vulkan);
CheckPointerImageTable(vulkan);
CheckNestedPointerImageTable(vulkan);
CheckGuardedRecordImageTable(vulkan);
CheckIndirectImageKeySwitch(vulkan);
CheckDynamicSampleOffsets(vulkan, 32u);