mirror of
https://github.com/KytyPS5/KytyPS5.git
synced 2026-10-19 18:36:53 +02:00
shader: remove unreachable mixed image specialization paths
This commit is contained in:
nmzik
@@ -61,10 +61,10 @@ uint32_t AddressF32(ValueEmitContext& ctx, const IR::MemoryInfo& mem, const IR::
|
||||
: Unary(ctx.state, spv::OpBitcast, TypeF32(ctx.state), value);
|
||||
}
|
||||
|
||||
ImageSampleLayout Layout(const IR::MemoryInfo& mem, ImageDimension dimension) {
|
||||
ImageSampleLayout Layout(const IR::MemoryInfo& mem) {
|
||||
ImageSampleLayout layout;
|
||||
uint32_t cursor = 0;
|
||||
const auto& info = ImageDimensionInfoFor(dimension);
|
||||
const auto& info = ImageDimensionInfoFor(mem.image_dimension);
|
||||
if (HasFlag(mem, Decoder::ImageSampleFlagOffset)) layout.offset = cursor++;
|
||||
if (HasFlag(mem, Decoder::ImageSampleFlagBias)) layout.bias = cursor++;
|
||||
if (HasFlag(mem, Decoder::ImageSampleFlagCompare)) layout.dref = cursor++;
|
||||
@@ -80,48 +80,6 @@ ImageSampleLayout Layout(const IR::MemoryInfo& mem, ImageDimension dimension) {
|
||||
return layout;
|
||||
}
|
||||
|
||||
void ValidateIndirectImageOperands(ValueEmitContext& ctx, const IR::Inst& inst) {
|
||||
const auto op = inst.GetOpcode();
|
||||
const auto& memory = ctx.Memory(inst);
|
||||
const auto& images = ctx.state.program.info.images;
|
||||
const auto& root = images.at(memory.resource);
|
||||
if (root.indirect_root != memory.resource) return;
|
||||
if (op != IR::ValueOpcode::ImageSampleRaw &&
|
||||
op != IR::ValueOpcode::ImageGatherRaw && op != IR::ValueOpcode::ImageRead) {
|
||||
ctx.Fail(inst, "requires an unsupported indirect image operation");
|
||||
}
|
||||
const auto& native = ImageDimensionInfoFor(memory.image_dimension);
|
||||
const auto native_layout = Layout(memory, memory.image_dimension);
|
||||
const bool sampling = op == IR::ValueOpcode::ImageSampleRaw ||
|
||||
op == IR::ValueOpcode::ImageGatherRaw;
|
||||
const auto validate = [&](const IR::ImageResource& image) {
|
||||
const auto& candidate = ImageDimensionInfoFor(image.dimension);
|
||||
// Sample/fetch ignore the native array slice for a non-array T# while
|
||||
// keeping the original positions of later operands.
|
||||
const bool projected_2d = memory.image_dimension == ImageDimension::Dim2DArray &&
|
||||
image.dimension == ImageDimension::Dim2D;
|
||||
bool compatible = projected_2d ||
|
||||
(candidate.coordinate_components == native.coordinate_components &&
|
||||
candidate.multisampled == native.multisampled);
|
||||
if (sampling) {
|
||||
compatible &= candidate.multisampled == 0u && native.multisampled == 0u;
|
||||
compatible &= projected_2d || Layout(memory, image.dimension) == native_layout;
|
||||
if (HasFlag(memory, Decoder::ImageSampleFlagDerivative)) {
|
||||
compatible &= candidate.spatial_components == native.spatial_components;
|
||||
}
|
||||
if (op == IR::ValueOpcode::ImageGatherRaw ||
|
||||
HasFlag(memory, Decoder::ImageSampleFlagCompare)) {
|
||||
compatible &= image.dimension != ImageDimension::Dim3D;
|
||||
}
|
||||
}
|
||||
if (!compatible) ctx.Fail(inst, "has incompatible image operand layout");
|
||||
};
|
||||
validate(root);
|
||||
for (const auto resource: root.indirect_resources) {
|
||||
if (resource != memory.resource) validate(images.at(resource));
|
||||
}
|
||||
}
|
||||
|
||||
uint32_t ZeroF32(EmitterState& state) {
|
||||
return ConstantF32(state, 0);
|
||||
}
|
||||
@@ -646,45 +604,6 @@ uint32_t StoreTexel(ValueEmitContext& ctx, const IR::MemoryInfo& mem, uint32_t d
|
||||
return PackImageTexel(ctx, mem, texel);
|
||||
}
|
||||
|
||||
struct ImageSampler {
|
||||
uint32_t id = 0;
|
||||
bool dynamic = false;
|
||||
};
|
||||
|
||||
auto EmitImageSamplers(ValueEmitContext& ctx, const IR::Inst& inst) {
|
||||
auto& state = ctx.state;
|
||||
const auto& mem = ctx.Memory(inst);
|
||||
const auto& images = state.program.info.images;
|
||||
const auto& samplers = state.program.info.samplers;
|
||||
std::array<ImageSampler, static_cast<size_t>(Prospero::TextureNumericClass::Sint) + 1u> result {};
|
||||
const auto add = [&](uint32_t image, uint32_t sampler) {
|
||||
auto& entry = result[static_cast<size_t>(images[image].numeric_class)];
|
||||
if (entry.id != 0u) return;
|
||||
const auto& choices = samplers.at(sampler).indirect_resources;
|
||||
uint32_t index = 0;
|
||||
if (!choices.empty()) {
|
||||
const auto* handle = inst.Arg(1).ResolveInstruction();
|
||||
if (handle == nullptr || handle->GetOpcode() != IR::ValueOpcode::GetSamplerResource)
|
||||
ctx.Fail(inst, "has invalid finite sampler key provenance");
|
||||
index = EmitIndexSwitch(
|
||||
state, ctx.Def(handle->Arg(0)), static_cast<uint32_t>(choices.size()),
|
||||
TypeU32(state), [&](uint32_t ordinal) { return ConstantU32(state, choices[ordinal]); });
|
||||
}
|
||||
entry = {LoadSamplerDescriptor(state, sampler, index), index != 0u};
|
||||
};
|
||||
add(mem.resource, mem.sampler);
|
||||
if (images[mem.resource].indirect_root == mem.resource) {
|
||||
for (const auto& pair: state.program.info.sampled_pairs)
|
||||
if (images[pair.image].indirect_root == mem.resource &&
|
||||
samplers[pair.sampler].snapshot_index == samplers[mem.sampler].snapshot_index)
|
||||
add(pair.image, pair.sampler);
|
||||
for (const auto candidate: images[mem.resource].indirect_resources)
|
||||
if (result[static_cast<size_t>(images[candidate].numeric_class)].id == 0u)
|
||||
ctx.Fail(inst, "has no sampler variant for an indirect image candidate");
|
||||
}
|
||||
return result;
|
||||
}
|
||||
|
||||
template <typename Emit>
|
||||
uint32_t EmitImageAccess(ValueEmitContext& ctx, const IR::Inst& inst, uint32_t result_type,
|
||||
const Emit& emit) {
|
||||
@@ -778,7 +697,6 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
|
||||
const auto& mem = ctx.Memory(inst);
|
||||
const auto image_arg = inst.Arg(0);
|
||||
ctx.ResourceIndex(image_arg, IR::ValueOpcode::GetImageResource);
|
||||
ValidateIndirectImageOperands(ctx, inst);
|
||||
const auto& image = state.program.info.images.at(mem.resource);
|
||||
const auto* address = ctx.ImageAddress(inst.Arg(image_info.needs_sampler ? 2 : 1));
|
||||
if (op == IR::ValueOpcode::ImageQueryDimensions) {
|
||||
@@ -811,14 +729,15 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
|
||||
return;
|
||||
}
|
||||
if (op == IR::ValueOpcode::ImageRead) {
|
||||
const auto numeric_class = image.numeric_class;
|
||||
const auto condition = ctx.Arg(inst, 2);
|
||||
ctx.Define(
|
||||
inst,
|
||||
EmitValueOrDefaultIfCondition(
|
||||
state, condition, TypeU32Vector(state, 4), ConstantU32CompositeZero(state, 4),
|
||||
[&]() {
|
||||
return EmitImageAccess(
|
||||
ctx, inst, TypeU32Vector(state, 4),
|
||||
const auto color = EmitImageAccess(
|
||||
ctx, inst, ImageVectorType(state, numeric_class, 4),
|
||||
[&](uint32_t resource, uint32_t array_index) {
|
||||
const auto& candidate = state.program.info.images[resource];
|
||||
const auto dimension = candidate.dimension;
|
||||
@@ -837,9 +756,10 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
|
||||
descriptor, coord, spv::ImageOperandsLodMask,
|
||||
LodU32(ctx, mem, *address, mem.image_dimension));
|
||||
}
|
||||
return ResultVector(ctx, UnpackImageTexel(ctx, mem, fetched),
|
||||
candidate.numeric_class, false, mem);
|
||||
return fetched;
|
||||
});
|
||||
return ResultVector(ctx, UnpackImageTexel(ctx, mem, color), numeric_class,
|
||||
false, mem);
|
||||
}));
|
||||
return;
|
||||
}
|
||||
@@ -860,7 +780,7 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
|
||||
if (op == IR::ValueOpcode::ImageSampleRaw || op == IR::ValueOpcode::ImageGatherRaw) {
|
||||
const auto dimension = image.dimension;
|
||||
const auto& dimension_info = ImageDimensionInfoFor(dimension);
|
||||
const auto layout = Layout(mem, mem.image_dimension);
|
||||
const auto layout = Layout(mem);
|
||||
const auto numeric_class = image.numeric_class;
|
||||
const bool dref = HasFlag(mem, Decoder::ImageSampleFlagCompare);
|
||||
IR::Value offset_value;
|
||||
@@ -894,6 +814,9 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
|
||||
ctx.Fail(inst, "has an unsupported 1D-array gather");
|
||||
return;
|
||||
}
|
||||
const auto result_numeric_class = dref ? Prospero::TextureNumericClass::Float
|
||||
: numeric_class;
|
||||
const auto result_type = ImageVectorType(state, result_numeric_class, 4);
|
||||
uint32_t component_or_dref;
|
||||
if (dref) {
|
||||
component_or_dref = layout.dref != NoImageComponent
|
||||
@@ -902,13 +825,9 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
|
||||
} else {
|
||||
component_or_dref = ConstantU32(state, ImageGatherSource(state, mem));
|
||||
}
|
||||
const auto samplers = EmitImageSamplers(ctx, inst);
|
||||
const auto sampler_id = LoadSamplerDescriptor(state, mem.sampler);
|
||||
const auto EmitGather = [&](uint32_t resource, uint32_t array_index) {
|
||||
const auto& candidate = state.program.info.images[resource];
|
||||
const auto numeric_class = dref ? Prospero::TextureNumericClass::Float
|
||||
: candidate.numeric_class;
|
||||
const auto result_type = ImageVectorType(state, numeric_class, 4);
|
||||
const auto sampler = samplers[static_cast<size_t>(candidate.numeric_class)];
|
||||
const auto dimension = candidate.dimension;
|
||||
const auto coord = CoordF32(
|
||||
ctx, mem, *address, layout.coord,
|
||||
@@ -922,8 +841,7 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
|
||||
offset = PackedOffset(ctx, offset_value, dimension, operand_mask);
|
||||
}
|
||||
const auto EmitMip = [&](uint32_t mip) {
|
||||
const auto sampled = MakeSampledImage(state, resource, sampler.id, mip, array_index,
|
||||
sampler.dynamic);
|
||||
const auto sampled = MakeSampledImage(state, resource, sampler_id, mip, array_index);
|
||||
const auto sample = state.builder.AllocateId();
|
||||
const std::array operands {operand_mask, offset};
|
||||
state.builder.AddFunction(
|
||||
@@ -932,15 +850,15 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
|
||||
std::span(operands).first(operand_mask != 0u ? 2u : 0u));
|
||||
return sample;
|
||||
};
|
||||
const auto gathered = candidate.mip_mode == IR::ImageMipMode::Dynamic && candidate.mip_count > 1u
|
||||
return candidate.mip_mode == IR::ImageMipMode::Dynamic && candidate.mip_count > 1u
|
||||
? EmitIndexSwitch(
|
||||
state, GatherMip(ctx, inst, mem, *address, layout, candidate.mip_count),
|
||||
candidate.mip_count, result_type, EmitMip)
|
||||
: EmitMip(0u);
|
||||
return ResultVector(ctx, UnpackImageGather(ctx, mem, gathered), numeric_class,
|
||||
false, mem, true);
|
||||
};
|
||||
ctx.Define(inst, EmitImageAccess(ctx, inst, TypeU32Vector(state, 4), EmitGather));
|
||||
const auto sample = EmitImageAccess(ctx, inst, result_type, EmitGather);
|
||||
ctx.Define(inst, ResultVector(ctx, UnpackImageGather(ctx, mem, sample),
|
||||
result_numeric_class, false, mem, true));
|
||||
return;
|
||||
}
|
||||
const bool explicit_lod = HasFlag(mem, Decoder::ImageSampleFlagDerivative) ||
|
||||
@@ -953,8 +871,10 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
|
||||
} else if (dref) {
|
||||
opcode = spv::OpImageSampleDrefImplicitLod;
|
||||
}
|
||||
uint32_t result_type = ImageVectorType(state, numeric_class, 4);
|
||||
uint32_t dref_value = 0;
|
||||
if (dref) {
|
||||
result_type = TypeF32(state);
|
||||
dref_value = ZeroF32(state);
|
||||
if (layout.dref != NoImageComponent) {
|
||||
dref_value = AddressF32(ctx, mem, *address, layout.dref);
|
||||
@@ -980,16 +900,23 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
|
||||
operands[1] = AddressF32(ctx, mem, *address, layout.bias);
|
||||
operand_count = 2;
|
||||
}
|
||||
const auto& samplers = state.program.info.samplers.at(mem.sampler).indirect_resources;
|
||||
uint32_t sampler_index = 0;
|
||||
if (!samplers.empty()) {
|
||||
const auto* handle = inst.Arg(1).ResolveInstruction();
|
||||
if (handle == nullptr || handle->GetOpcode() != IR::ValueOpcode::GetSamplerResource)
|
||||
ctx.Fail(inst, "has invalid finite sampler key provenance");
|
||||
sampler_index = EmitIndexSwitch(
|
||||
state, ctx.Def(handle->Arg(0)), static_cast<uint32_t>(samplers.size()),
|
||||
TypeU32(state), [&](uint32_t ordinal) { return ConstantU32(state, samplers[ordinal]); });
|
||||
}
|
||||
if (layout.offset != NoImageComponent &&
|
||||
!offset_value.IsImmediate() && !state.maintenance8_enabled) {
|
||||
ctx.Fail(inst, "requires maintenance8 for a dynamic sample texel offset");
|
||||
}
|
||||
const auto samplers = EmitImageSamplers(ctx, inst);
|
||||
const auto sampler_id = LoadSamplerDescriptor(state, mem.sampler, sampler_index);
|
||||
const auto EmitSample = [&](uint32_t resource, uint32_t array_index) {
|
||||
const auto& candidate = state.program.info.images[resource];
|
||||
const auto result_type = dref ? TypeF32(state)
|
||||
: ImageVectorType(state, candidate.numeric_class, 4);
|
||||
const auto sampler = samplers[static_cast<size_t>(candidate.numeric_class)];
|
||||
auto sample_operands = operands;
|
||||
auto sample_operand_count = operand_count;
|
||||
if (layout.offset != NoImageComponent) {
|
||||
@@ -1001,16 +928,19 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
|
||||
CoordF32(ctx, mem, *address, layout.coord,
|
||||
ImageDimensionInfoFor(candidate.dimension).coordinate_components,
|
||||
candidate.cube);
|
||||
const auto sampled = MakeSampledImage(state, resource, sampler.id, 0u, array_index,
|
||||
sampler.dynamic);
|
||||
const auto sampled = MakeSampledImage(state, resource, sampler_id, 0u, array_index,
|
||||
sampler_index != 0u);
|
||||
const auto sample = state.builder.AllocateId();
|
||||
state.builder.AddFunction(opcode, result_type, sample, sampled, coord,
|
||||
std::span(&dref_value, dref ? 1u : 0u),
|
||||
std::span<const uint32_t>(sample_operands).first(sample_operand_count));
|
||||
return ResultVector(ctx, dref ? sample : UnpackImageTexel(ctx, mem, sample),
|
||||
candidate.numeric_class, dref, mem);
|
||||
return sample;
|
||||
};
|
||||
ctx.Define(inst, EmitImageAccess(ctx, inst, TypeU32Vector(state, 4), EmitSample));
|
||||
auto result = EmitImageAccess(ctx, inst, result_type, EmitSample);
|
||||
if (!dref) {
|
||||
result = UnpackImageTexel(ctx, mem, result);
|
||||
}
|
||||
ctx.Define(inst, ResultVector(ctx, result, numeric_class, dref, mem));
|
||||
return;
|
||||
}
|
||||
if (image_info.access == IR::ImageAccess::Atomic) {
|
||||
|
||||
@@ -272,8 +272,6 @@ struct ImageSampleLayout {
|
||||
uint32_t lod = NoImageComponent;
|
||||
uint32_t grad_x = NoImageComponent;
|
||||
uint32_t grad_y = NoImageComponent;
|
||||
|
||||
bool operator==(const ImageSampleLayout&) const = default;
|
||||
};
|
||||
|
||||
struct F32Class {
|
||||
|
||||
@@ -648,6 +648,10 @@ static bool BuildResourceSpecialization(const ResourcePlan& program, ResourceSna
|
||||
return SpecializationFail("indirect image specialization has no typed candidate");
|
||||
}
|
||||
const auto& image_class = specialization.images[exemplar];
|
||||
const auto is_2d = [](Decoder::ImageDimension dimension) {
|
||||
return dimension == Decoder::ImageDimension::Dim2D ||
|
||||
dimension == Decoder::ImageDimension::Dim2DArray;
|
||||
};
|
||||
for (uint32_t candidate = 0; candidate < specialization.images.size(); candidate++) {
|
||||
auto& image = specialization.images[candidate];
|
||||
if (image.indirect_root != root_index) {
|
||||
@@ -661,7 +665,11 @@ static bool BuildResourceSpecialization(const ResourcePlan& program, ResourceSna
|
||||
image.shader_swizzle = image_class.shader_swizzle;
|
||||
image.cube = image_class.cube;
|
||||
}
|
||||
if (image.mip_count != image_class.mip_count ||
|
||||
const bool same_coordinates = image.dimension == image_class.dimension &&
|
||||
image.cube == image_class.cube;
|
||||
if (image.numeric_class != image_class.numeric_class ||
|
||||
(!same_coordinates && !(is_2d(image.dimension) && is_2d(image_class.dimension))) ||
|
||||
image.mip_count != image_class.mip_count ||
|
||||
image.conversion_format != image_class.conversion_format ||
|
||||
image.shader_swizzle != image_class.shader_swizzle) {
|
||||
return SpecializationFail(
|
||||
@@ -1309,20 +1317,9 @@ void ApplyResourceSpecialization(Program& program, const ResourceSpecialization&
|
||||
}
|
||||
|
||||
SamplerPlan sampler_plan;
|
||||
EXIT_IF(!BuildSamplerPlan(program.info, sampler_plan));
|
||||
auto& samplers = program.info.samplers;
|
||||
auto& sampled_pairs = program.info.sampled_pairs;
|
||||
const auto original_pair_count = sampled_pairs.size();
|
||||
for (size_t index = 0; index < original_pair_count; ++index) {
|
||||
const auto pair = sampled_pairs[index];
|
||||
uint32_t classes = 1u << static_cast<uint32_t>(images[pair.image].numeric_class);
|
||||
for (const auto candidate: images[pair.image].indirect_resources) {
|
||||
const auto bit = 1u << static_cast<uint32_t>(images[candidate].numeric_class);
|
||||
if ((classes & bit) != 0u) continue;
|
||||
classes |= bit;
|
||||
sampled_pairs.push_back({candidate, pair.sampler, pair.first_use_pc});
|
||||
}
|
||||
}
|
||||
EXIT_IF(!BuildSamplerPlan(program.info, sampler_plan));
|
||||
const auto original_sampler_count = samplers.size();
|
||||
samplers.reserve(sampler_plan.sampler_count);
|
||||
for (uint32_t index = 0; index < sampler_plan.sampler_count; index++) {
|
||||
@@ -1434,6 +1431,10 @@ void ApplyResourceSpecialization(Program& program, const ResourceSpecialization&
|
||||
memory.sampler = sampler_plan.mapping[memory.sampler][type];
|
||||
EXIT_IF(memory.sampler == UINT32_MAX);
|
||||
}
|
||||
EXIT_IF(image.indirect_root == memory.resource &&
|
||||
inst.GetOpcode() != ValueOpcode::ImageSampleRaw &&
|
||||
inst.GetOpcode() != ValueOpcode::ImageGatherRaw &&
|
||||
inst.GetOpcode() != ValueOpcode::ImageRead);
|
||||
}
|
||||
}
|
||||
for (auto& memory: memory_info) {
|
||||
|
||||
@@ -434,18 +434,8 @@ void TestBoundedImageViewEligibility() {
|
||||
plan.descriptor_sources.push_back(source);
|
||||
}
|
||||
plan.descriptor_sources[image_source].indirect_descriptor->sources = std::move(selected);
|
||||
Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
|
||||
snapshot.images.size() == 2u &&
|
||||
snapshot.images[0].dwords == descriptor(0x30u, Type::kColor2D) &&
|
||||
snapshot.images[1].dwords == descriptor(0x30u, Type::kColor3D) &&
|
||||
specialization.images[0].dimension == Decoder::ImageDimension::Dim2D &&
|
||||
specialization.images[1].dimension == Decoder::ImageDimension::Dim3D,
|
||||
"explicitly selected image dimensions were discarded before operand validation");
|
||||
const auto converted = plan.descriptor_sources[image_source].indirect_descriptor->sources.back();
|
||||
plan.descriptor_sources[converted].dwords[1] = Value(
|
||||
static_cast<uint32_t>(Libs::Graphics::Prospero::BufferFormat::k11_11_10UInt) << 20u);
|
||||
Check(!MaterializeResources(plan, runtime, snapshot, specialization),
|
||||
"explicitly selected incompatible conversion was silently normalized to null");
|
||||
"explicitly selected incompatible descriptor was silently normalized to null");
|
||||
}
|
||||
|
||||
void TestWaterfallImageTable() {
|
||||
@@ -1405,52 +1395,6 @@ void TestDescendingImageRange() {
|
||||
check_bitmap({0u, 127u}, 0u);
|
||||
}
|
||||
|
||||
void TestMixedIndirectImageSamplers() {
|
||||
auto fixture = MakeIndirectImageFixture(false, 32u, 0u, 0u);
|
||||
fixture->PlanAndTrack();
|
||||
auto& samplers = fixture->program.info.samplers;
|
||||
samplers.resize(3u, samplers[0]);
|
||||
samplers[0].indirect_resources = {1u, 2u};
|
||||
const auto plan = ExtractResourcePlan(fixture->program);
|
||||
std::array<uint32_t, 8> data {0x3000u, 16u << 16u, 1u, 0u,
|
||||
0x2000u, 32u << 16u, 3u, 0u};
|
||||
LinearTestMemory memory;
|
||||
using F = Libs::Graphics::Prospero::BufferFormat;
|
||||
for (uint32_t record = 0; record < 3u; ++record) {
|
||||
const auto offset = (0x1000u + record * 32u) / 4u;
|
||||
memory.words[offset] = 0x100u + record;
|
||||
memory.words[offset + 1u] = static_cast<uint32_t>(
|
||||
record == 0u ? F::k32_32_32_32Float : F::k32_32_32_32UInt) << 20u;
|
||||
memory.words[offset + 3u] = Libs::Graphics::DstSel(4, 5, 6, 7) |
|
||||
(uint32_t(Libs::Graphics::Prospero::ImageType::kColor2D) << 28u);
|
||||
}
|
||||
const SrtRuntime runtime {.user_data = data, .userdata = &memory,
|
||||
.read_specialization_memory = ReadLinearTestMemory};
|
||||
ResourceSnapshot snapshot;
|
||||
ResourceSpecialization specialization;
|
||||
Check(MaterializeResources(plan, runtime, snapshot, specialization) && snapshot.images.size() == 4u,
|
||||
"mixed float/uint image table failed materialization");
|
||||
ApplyResourceSpecialization(fixture->program, specialization);
|
||||
const auto& info = fixture->program.info;
|
||||
Check(info.sampled_pairs.size() == 2u && info.samplers.size() == 6u,
|
||||
"indirect image candidates duplicated sampler pairs or lost finite class variants");
|
||||
for (const auto& pair : info.sampled_pairs) {
|
||||
const auto integer = info.images[pair.image].numeric_class ==
|
||||
Libs::Graphics::Prospero::TextureNumericClass::Uint;
|
||||
const auto& sampler = info.samplers[pair.sampler];
|
||||
Check(sampler.snapshot_index == 0u && sampler.integer_border == integer &&
|
||||
sampler.force_point_filtering == integer && sampler.indirect_resources.size() == 2u &&
|
||||
pair.first_use_pc == 0x10f0u,
|
||||
"mixed image class changed native sampler identity or integer-border semantics");
|
||||
for (uint32_t ordinal = 0; ordinal < sampler.indirect_resources.size(); ++ordinal) {
|
||||
const auto& candidate = info.samplers[sampler.indirect_resources[ordinal]];
|
||||
Check(candidate.snapshot_index == ordinal + 1u &&
|
||||
candidate.integer_border == integer && candidate.force_point_filtering == integer,
|
||||
"mixed image class mapped a finite sampler choice to the wrong variant");
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
void TestUniformizedMaterialImageKeys() {
|
||||
namespace CFG = Libs::Graphics::ShaderRecompiler::CFG;
|
||||
enum class Variant {
|
||||
@@ -4607,7 +4551,6 @@ int main() {
|
||||
Run("scalar buffer selector domain", TestScalarBufferSelectorDomain);
|
||||
Run("shared append selector bounds", TestSharedAppendSelectorBounds);
|
||||
Run("descending image range", TestDescendingImageRange);
|
||||
Run("mixed indirect image samplers", TestMixedIndirectImageSamplers);
|
||||
Run("SRT runtime", TestSrtFlatteningAndRuntimeMemoization);
|
||||
Run("dynamic SRT", TestDynamicSrtReadRemainsExplicit);
|
||||
Run("writable descriptor phi", TestWritableDescriptorPhi);
|
||||
|
||||
@@ -1987,7 +1987,6 @@ public:
|
||||
vk::ImageLayout layout = vk::ImageLayout::eUndefined;
|
||||
u32 width = 0;
|
||||
u32 height = 0;
|
||||
u32 depth = 1;
|
||||
u32 layers = 1;
|
||||
u32 mip_levels = 1;
|
||||
u32 dwords_per_pixel = 0;
|
||||
@@ -15452,12 +15451,11 @@ public:
|
||||
u32 view_layers = 0,
|
||||
vk::SampleCountFlagBits samples = vk::SampleCountFlagBits::e1,
|
||||
vk::Format view_format = vk::Format::eUndefined,
|
||||
u32 view_base_mip = 0, u32 depth = 1) {
|
||||
u32 view_base_mip = 0) {
|
||||
Image ret;
|
||||
ret.format = format;
|
||||
ret.width = width;
|
||||
ret.height = height;
|
||||
ret.depth = depth;
|
||||
ret.layers = layers;
|
||||
ret.mip_levels = std::max<u32>(static_cast<u32>(initial_mips.size()), 1u);
|
||||
ret.dwords_per_pixel = dwords_per_pixel;
|
||||
@@ -15468,7 +15466,7 @@ public:
|
||||
image_info.format = format;
|
||||
image_info.extent.width = width;
|
||||
image_info.extent.height = height;
|
||||
image_info.extent.depth = depth;
|
||||
image_info.extent.depth = 1;
|
||||
image_info.mipLevels = ret.mip_levels;
|
||||
image_info.arrayLayers = layers;
|
||||
image_info.samples = samples;
|
||||
@@ -15533,13 +15531,13 @@ public:
|
||||
size_t total_dwords = 0;
|
||||
for (u32 level = 0; level < ret.mip_levels; level++) {
|
||||
total_dwords +=
|
||||
ImageMipDwordCount(width, height, dwords_per_pixel, level, layers) * MipExtent(depth, level);
|
||||
ImageMipDwordCount(width, height, dwords_per_pixel, level, layers);
|
||||
}
|
||||
std::vector<u32> contents(total_dwords, 0);
|
||||
size_t offset = 0;
|
||||
for (u32 level = 0; level < ret.mip_levels; level++) {
|
||||
const auto level_dwords =
|
||||
ImageMipDwordCount(width, height, dwords_per_pixel, level, layers) * MipExtent(depth, level);
|
||||
ImageMipDwordCount(width, height, dwords_per_pixel, level, layers);
|
||||
const auto &src = initial_mips[level];
|
||||
for (size_t i = 0; i < src.size() && i < level_dwords; i++) {
|
||||
contents[offset + i] = src[i];
|
||||
@@ -15683,8 +15681,7 @@ public:
|
||||
if (image.resource_class == ShaderRecompiler::IR::ImageResourceClass::Sampled) {
|
||||
supported = supported ||
|
||||
image.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim1D ||
|
||||
image.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim1DArray ||
|
||||
image.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim3D;
|
||||
image.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim1DArray;
|
||||
}
|
||||
Require(test.name, "dispatch", supported,
|
||||
"unsupported image dimension needs a matching Vulkan test view");
|
||||
@@ -19205,11 +19202,11 @@ private:
|
||||
copy.imageSubresource.layerCount = image->layers;
|
||||
copy.imageExtent.width = MipExtent(image->width, level);
|
||||
copy.imageExtent.height = MipExtent(image->height, level);
|
||||
copy.imageExtent.depth = MipExtent(image->depth, level);
|
||||
copy.imageExtent.depth = 1;
|
||||
copies.push_back(copy);
|
||||
offset += static_cast<vk::DeviceSize>(
|
||||
ImageMipDwordCount(image->width, image->height,
|
||||
image->dwords_per_pixel, level, image->layers) * MipExtent(image->depth, level) *
|
||||
image->dwords_per_pixel, level, image->layers) *
|
||||
sizeof(u32));
|
||||
}
|
||||
cmd.copyBufferToImage(staging, image->image,
|
||||
@@ -36137,9 +36134,6 @@ void CheckIndirectImageKeySwitch(VulkanHarness &vulkan) {
|
||||
program.info.images[root.indirect_resources[103]].dimension =
|
||||
ShaderRecompiler::Decoder::ImageDimension::Dim2DArray;
|
||||
program.info.images[root.indirect_resources[103]].cube = true;
|
||||
program.memory_info[0].image_dimension =
|
||||
ShaderRecompiler::Decoder::ImageDimension::Dim2DArray;
|
||||
program.memory_info[0].image_address_components = 4;
|
||||
program.shader_info_complete = false;
|
||||
CollectShaderInfo(program, {.compute = &compute});
|
||||
spirv = ShaderRecompiler::Spirv::EmitProgram(program, {.compute = &compute});
|
||||
@@ -36148,6 +36142,9 @@ void CheckIndirectImageKeySwitch(VulkanHarness &vulkan) {
|
||||
CountText(text, "OpImageSampleExplicitLod") == 3,
|
||||
"one cube split the native arrays into per-image samples");
|
||||
|
||||
program.memory_info[0].image_dimension =
|
||||
ShaderRecompiler::Decoder::ImageDimension::Dim2DArray;
|
||||
program.memory_info[0].image_address_components = 4;
|
||||
for (u32 component = 0; component < 4u; component++) {
|
||||
constexpr std::array values{1.375f, 1.625f, 10.0f, 2.0f};
|
||||
address.SetArg(component, Value(std::bit_cast<u32>(values[component])));
|
||||
@@ -36201,83 +36198,6 @@ void CheckIndirectImageKeySwitch(VulkanHarness &vulkan) {
|
||||
Require(name, "mixed sample count", samples == 3u,
|
||||
"mixed candidate switch did not retain every image");
|
||||
}
|
||||
#if KYTY_PLATFORM != KYTY_PLATFORM_WINDOWS
|
||||
enum class RejectedAccess { LodPosition, Derivatives, Gather, Depth, Multisampled, Query, Write, Atomic };
|
||||
for (const auto [rejected, label] : {
|
||||
std::pair{RejectedAccess::LodPosition, "IndirectImageLodPositionRejected"},
|
||||
std::pair{RejectedAccess::Derivatives, "IndirectImageDerivativeLayoutRejected"},
|
||||
std::pair{RejectedAccess::Gather, "IndirectImageVolumeGatherRejected"},
|
||||
std::pair{RejectedAccess::Depth, "IndirectImageVolumeDrefRejected"},
|
||||
std::pair{RejectedAccess::Multisampled, "IndirectImageMsaaSampleRejected"},
|
||||
std::pair{RejectedAccess::Query, "IndirectImageQueryRejected"},
|
||||
std::pair{RejectedAccess::Write, "IndirectImageWriteRejected"},
|
||||
std::pair{RejectedAccess::Atomic, "IndirectImageAtomicRejected"}}) {
|
||||
ExpectFatal(label, [&] {
|
||||
using Dimension = ShaderRecompiler::Decoder::ImageDimension;
|
||||
for (auto &resource : program.info.images) {
|
||||
resource.dimension = Dimension::Dim2DArray;
|
||||
resource.cube = false;
|
||||
}
|
||||
program.info.images[2].dimension = Dimension::Dim3D;
|
||||
auto &mem = program.memory_info[0];
|
||||
const auto position = std::ranges::find_if(*block, [&](const Inst &inst) { return &inst == &sample; });
|
||||
if (rejected == RejectedAccess::LodPosition) {
|
||||
// Four available operands do not make native x,y,LOD compatible with x,y,z,LOD.
|
||||
mem.image_dimension = Dimension::Dim2D;
|
||||
program.info.images[0].dimension = program.info.images[1].dimension = Dimension::Dim2D;
|
||||
} else if (rejected == RejectedAccess::Derivatives) {
|
||||
mem.image_sample_flags = ShaderRecompiler::Decoder::ImageSampleFlagDerivative;
|
||||
mem.image_address_components = 9u; // Capacity alone must not change the native gradient layout.
|
||||
} else if (rejected == RejectedAccess::Gather) {
|
||||
mem.image_sample_flags = ShaderRecompiler::Decoder::ImageSampleFlagLevelZero;
|
||||
mem.image_address_components = 3u;
|
||||
const auto gather = block->PrependNewInst(position, ValueOpcode::ImageGatherRaw,
|
||||
{Value(&image), Value(&sampler), Value(&address), Value(0u), Value(0u),
|
||||
Value(0u), Value(0u), Value(true)}, memory_flag_bits);
|
||||
sample.ReplaceUsesWith(Value(&*gather), false);
|
||||
} else if (rejected == RejectedAccess::Depth) {
|
||||
mem.image_sample_flags |= ShaderRecompiler::Decoder::ImageSampleFlagCompare;
|
||||
mem.image_address_components = 5u;
|
||||
for (auto &resource : program.info.images) resource.depth_compare = true;
|
||||
program.info.samplers[0].depth_compare = true;
|
||||
} else if (rejected == RejectedAccess::Multisampled) {
|
||||
program.info.images[2].dimension = Dimension::Dim2DMsaaArray;
|
||||
} else if (rejected == RejectedAccess::Query) {
|
||||
const auto query = block->PrependNewInst(position, ValueOpcode::ImageQueryDimensions,
|
||||
{Value(&image), Value(&address)}, memory_flag_bits);
|
||||
sample.ReplaceUsesWith(Value(&*query), false);
|
||||
} else {
|
||||
sample_x.ReplaceUsesWith(Value(0u), false);
|
||||
sample.Invalidate();
|
||||
sampler.Invalidate();
|
||||
program.info.samplers.clear();
|
||||
program.info.sampled_pairs.clear();
|
||||
mem.image_sample_flags = 0u;
|
||||
mem.image_address_components = 3u;
|
||||
const bool atomic = rejected == RejectedAccess::Atomic;
|
||||
for (auto &resource : program.info.images) {
|
||||
resource.resource_class = ImageResourceClass::Storage;
|
||||
resource.numeric_class = atomic ? Prospero::TextureNumericClass::Uint
|
||||
: Prospero::TextureNumericClass::Float;
|
||||
resource.written = true;
|
||||
resource.atomic = atomic;
|
||||
}
|
||||
if (atomic) {
|
||||
block->PrependNewInst(position, ValueOpcode::ImageAtomicIAdd32,
|
||||
{Value(&image), Value(&address), Value(1u), Value(true)}, memory_flag_bits);
|
||||
} else {
|
||||
const auto texel = block->PrependNewInst(position, ValueOpcode::CompositeConstructU32x4,
|
||||
{Value(0u), Value(0u), Value(0u), Value(0u)});
|
||||
block->PrependNewInst(position, ValueOpcode::ImageWrite,
|
||||
{Value(&image), Value(&address), Value(&*texel), Value(true)}, memory_flag_bits);
|
||||
}
|
||||
}
|
||||
program.shader_info_complete = false;
|
||||
CollectShaderInfo(program, {.compute = &compute});
|
||||
ShaderRecompiler::Spirv::EmitProgram(program, {.compute = &compute});
|
||||
});
|
||||
}
|
||||
#endif
|
||||
// Keep an unrelated native root between the null image and appended children,
|
||||
// then select both sides of a cube boundary and the unmapped suffix on the GPU.
|
||||
root.indirect_resources = {0u, 2u, 3u, 4u, 5u, 6u};
|
||||
@@ -36328,148 +36248,49 @@ void CheckIndirectImageKeySwitch(VulkanHarness &vulkan) {
|
||||
resource == 4u ? vk::ImageViewType::e2DArray : vk::ImageViewType::e2D, layers));
|
||||
}
|
||||
const auto native_sampler = vulkan.CreateSampler(test.name);
|
||||
const auto mip_sampler = vulkan.CreateSampler(test.name, 2u);
|
||||
constexpr u32 volume_mip_bits = 0x43800001u;
|
||||
constexpr u32 array_mip_bits = 0x42800001u;
|
||||
const auto mip_pixels = [](u32 texels, u32 first) {
|
||||
std::vector<u32> pixels(texels * 4u);
|
||||
for (u32 texel = 0; texel < texels; ++texel)
|
||||
std::fill_n(pixels.begin() + texel * 4u, 4u, first + texel);
|
||||
return pixels;
|
||||
};
|
||||
auto volume_texture = vulkan.CreateImageMips(
|
||||
test.name, 4, 4, vk::Format::eR32G32B32A32Sfloat, vk::ImageUsageFlagBits::eSampled,
|
||||
{mip_pixels(4u * 4u * 8u, 0x44000001u), mip_pixels(2u * 2u * 4u, volume_mip_bits)}, 4u,
|
||||
vk::ImageLayout::eShaderReadOnlyOptimal, vk::ImageType::e3D, vk::ImageViewType::e3D, 1u,
|
||||
0u, 0u, vk::SampleCountFlagBits::e1, vk::Format::eUndefined, 0u, 8u);
|
||||
auto array_texture = vulkan.CreateImageMips(
|
||||
test.name, 4, 4, vk::Format::eR32G32B32A32Sfloat, vk::ImageUsageFlagBits::eSampled,
|
||||
{mip_pixels(4u * 4u * 2u, 0x43000001u), mip_pixels(2u * 2u * 2u, array_mip_bits)}, 4u,
|
||||
vk::ImageLayout::eShaderReadOnlyOptimal, vk::ImageType::e2D, vk::ImageViewType::e2DArray, 2u);
|
||||
constexpr u32 uint_texel = 0x80000106u;
|
||||
auto uint_texture = vulkan.CreateImageMips(
|
||||
test.name, 2, 2, vk::Format::eR32G32B32A32Uint, vk::ImageUsageFlagBits::eSampled,
|
||||
{std::vector<u32>(16u, uint_texel), std::vector<u32>(4u, uint_texel + 10u)}, 4u,
|
||||
vk::ImageLayout::eShaderReadOnlyOptimal, vk::ImageType::e2D, vk::ImageViewType::e2D, 1u);
|
||||
vk::SamplerCreateInfo integer_sampler_info{};
|
||||
integer_sampler_info.addressModeU = integer_sampler_info.addressModeV =
|
||||
integer_sampler_info.addressModeW = vk::SamplerAddressMode::eClampToEdge;
|
||||
integer_sampler_info.borderColor = vk::BorderColor::eIntOpaqueWhite;
|
||||
vk::Sampler integer_sampler;
|
||||
RequireVk(test.name, "integer sampler", vulkan.Device().createSampler(
|
||||
&integer_sampler_info, nullptr, &integer_sampler), "vkCreateSampler");
|
||||
integer_sampler_info.addressModeU = integer_sampler_info.addressModeV =
|
||||
integer_sampler_info.addressModeW = vk::SamplerAddressMode::eClampToBorder;
|
||||
vk::Sampler integer_border_sampler;
|
||||
RequireVk(test.name, "integer border sampler", vulkan.Device().createSampler(
|
||||
&integer_sampler_info, nullptr, &integer_border_sampler), "vkCreateSampler");
|
||||
const std::array native_samplers{native_sampler, native_sampler, native_sampler,
|
||||
integer_sampler, integer_sampler, integer_border_sampler};
|
||||
enum class Access { Sample, MixedDimensions, Gather, GatherHomogeneous, MixedSample, MixedGather, MixedFetch, Fetch };
|
||||
enum class Access { Sample, Gather, GatherHomogeneous, Fetch };
|
||||
auto *access = &sample;
|
||||
bool uint_bound = false;
|
||||
bool volume_bound = false;
|
||||
for (const auto mode : {Access::Sample, Access::MixedDimensions, Access::Gather,
|
||||
Access::GatherHomogeneous, Access::MixedSample,
|
||||
Access::MixedGather, Access::MixedFetch, Access::Fetch}) {
|
||||
const bool mixed = mode == Access::MixedSample || mode == Access::MixedGather || mode == Access::MixedFetch;
|
||||
const bool dimensions = mode == Access::MixedDimensions;
|
||||
const bool fetch = mode == Access::Fetch || mode == Access::MixedFetch;
|
||||
const bool gather = mode == Access::Gather || mode == Access::GatherHomogeneous || mode == Access::MixedGather;
|
||||
for (const auto mode : {Access::Sample, Access::Gather,
|
||||
Access::GatherHomogeneous, Access::Fetch}) {
|
||||
const bool fetch = mode == Access::Fetch;
|
||||
const bool gather = mode == Access::Gather || mode == Access::GatherHomogeneous;
|
||||
const bool homogeneous = mode == Access::GatherHomogeneous;
|
||||
test.name = dimensions ? "IndirectImageArrayCubeVolumeLod"
|
||||
: mixed ? (fetch ? "IndirectImageMixedFetch" : gather ? "IndirectImageMixedGather"
|
||||
: "IndirectImageMixedSample")
|
||||
: fetch ? "IndirectImageFetchContiguousRuns"
|
||||
: homogeneous ? "IndirectImageGatherHomogeneous"
|
||||
: gather ? "IndirectImageGatherContiguousRuns" : "IndirectImageContiguousRuns";
|
||||
if (volume_bound != dimensions) {
|
||||
std::swap(textures[6], volume_texture);
|
||||
std::swap(textures[3], array_texture);
|
||||
compiled.program.info.images[6].dimension = dimensions
|
||||
? ShaderRecompiler::Decoder::ImageDimension::Dim3D
|
||||
: ShaderRecompiler::Decoder::ImageDimension::Dim2D;
|
||||
compiled.program.info.images[3].dimension = dimensions
|
||||
? ShaderRecompiler::Decoder::ImageDimension::Dim2DArray
|
||||
: ShaderRecompiler::Decoder::ImageDimension::Dim2D;
|
||||
volume_bound = dimensions;
|
||||
}
|
||||
if (uint_bound != mixed) {
|
||||
std::swap(textures[6], uint_texture);
|
||||
compiled.program.info.images[6].numeric_class = mixed ? Prospero::TextureNumericClass::Uint
|
||||
: Prospero::TextureNumericClass::Float;
|
||||
uint_bound = mixed;
|
||||
}
|
||||
if (mode == Access::MixedSample) {
|
||||
const auto position = std::ranges::find_if(*block, [&](const Inst &inst) { return &inst == &sampler; });
|
||||
const auto sampler_key = block->PrependNewInst(position, ValueOpcode::BitwiseAnd32,
|
||||
{Value(&key), Value(1u)});
|
||||
sampler.SetArg(0, Value(&*sampler_key));
|
||||
compiled.program.info.samplers.resize(6);
|
||||
for (u32 index = 0; index < 6u; ++index) {
|
||||
auto &resource = compiled.program.info.samplers[index];
|
||||
resource.snapshot_index = index % 3u;
|
||||
resource.force_point_filtering = index >= 3u;
|
||||
resource.integer_border = index >= 3u;
|
||||
}
|
||||
compiled.program.info.samplers[0].indirect_resources = {1u, 2u};
|
||||
compiled.program.info.samplers[3].indirect_resources = {4u, 5u};
|
||||
compiled.program.info.sampled_pairs.push_back({6u, 3u, 0x10f0u});
|
||||
}
|
||||
const auto opcode = fetch ? ValueOpcode::ImageRead : gather ? ValueOpcode::ImageGatherRaw
|
||||
: ValueOpcode::ImageSampleRaw;
|
||||
if (access->GetOpcode() != opcode) {
|
||||
if (mode == Access::Gather) {
|
||||
test.name = "IndirectImageGatherContiguousRuns";
|
||||
const auto position = std::ranges::find_if(*block, [&](const Inst &inst) {
|
||||
return &inst == access;
|
||||
});
|
||||
const auto read_at = fetch ? block->PrependNewInst(position, opcode,
|
||||
{Value(&image), Value(&address), Value(true)}, memory_flag_bits)
|
||||
: gather ? block->PrependNewInst(position, opcode,
|
||||
{Value(&image), Value(&sampler), Value(&address), Value(0u), Value(0u),
|
||||
Value(0u), Value(0u), Value(true)}, memory_flag_bits)
|
||||
: block->PrependNewInst(position, opcode,
|
||||
{Value(&image), Value(&sampler), Value(&address), Value(true)}, memory_flag_bits);
|
||||
auto &read = *read_at;
|
||||
auto &read = *block->PrependNewInst(
|
||||
position, ValueOpcode::ImageGatherRaw,
|
||||
{Value(&image), Value(&sampler), Value(&address), Value(0u), Value(0u),
|
||||
Value(0u), Value(0u), Value(true)}, memory_flag_bits);
|
||||
access->ReplaceUsesWith(Value(&read), false);
|
||||
access = &read;
|
||||
}
|
||||
if (dimensions) {
|
||||
const auto position = std::ranges::find_if(*block, [&](const Inst &inst) { return &inst == &address; });
|
||||
const auto parity = block->PrependNewInst(position, ValueOpcode::BitwiseAnd32, {Value(&key), Value(1u)});
|
||||
const auto odd = block->PrependNewInst(position, ValueOpcode::INotEqual32, {Value(&*parity), Value(0u)});
|
||||
const auto z = block->PrependNewInst(position, ValueOpcode::SelectU32,
|
||||
{Value(&*odd), Value(std::bit_cast<u32>(0.75f)), Value(std::bit_cast<u32>(0.25f))});
|
||||
const auto cube_lane = block->PrependNewInst(position, ValueOpcode::IEqual32, {Value(&key), Value(3u)});
|
||||
const auto coordinate_z = block->PrependNewInst(position, ValueOpcode::SelectU32,
|
||||
{Value(&*cube_lane), Value(0u), Value(&*z)});
|
||||
address.SetArg(0, Value(std::bit_cast<u32>(1.25f)));
|
||||
address.SetArg(1, Value(std::bit_cast<u32>(1.75f)));
|
||||
address.SetArg(2, Value(&*coordinate_z));
|
||||
address.SetArg(3, Value(std::bit_cast<u32>(1.0f)));
|
||||
} else if (gather) {
|
||||
compiled.program.info.images[4].cube = false;
|
||||
auto &gather_memory = compiled.program.memory_info[0];
|
||||
gather_memory.dmask = 1;
|
||||
gather_memory.image_sample_flags = ShaderRecompiler::Decoder::ImageSampleFlagLevelZero |
|
||||
(mixed ? 0u : ShaderRecompiler::Decoder::ImageSampleFlagOffset);
|
||||
// Nonmixed GATHER4_LZ_O places its packed signed offset before the coordinates.
|
||||
address.SetArg(0, Value(mixed ? std::bit_cast<u32>(1.5f) : 0x3fu));
|
||||
address.SetArg(1, Value(std::bit_cast<u32>(mixed ? 1.5f : 0.75f)));
|
||||
address.SetArg(2, Value(mixed ? 0u : std::bit_cast<u32>(0.25f)));
|
||||
ShaderRecompiler::Decoder::ImageSampleFlagOffset;
|
||||
// GATHER4_LZ_O places its packed signed offset before the coordinates.
|
||||
address.SetArg(0, Value(0x3fu));
|
||||
address.SetArg(1, Value(std::bit_cast<u32>(0.75f)));
|
||||
address.SetArg(2, Value(std::bit_cast<u32>(0.25f)));
|
||||
address.SetArg(3, Value(0u));
|
||||
} else if (mode == Access::MixedSample) {
|
||||
compiled.program.memory_info[0].dmask = 0xf;
|
||||
compiled.program.memory_info[0].image_sample_flags = ShaderRecompiler::Decoder::ImageSampleFlagLod;
|
||||
for (u32 component = 0; component < 4u; ++component)
|
||||
address.SetArg(component, Value(component < 2u ? std::bit_cast<u32>(1.5f) : 0u));
|
||||
}
|
||||
auto &selected_resources = compiled.program.info.images[0].indirect_resources;
|
||||
selected_resources = homogeneous ? std::vector<u32>{0u, 2u, 3u, 5u, 6u}
|
||||
: root.indirect_resources;
|
||||
if (homogeneous) test.name = "IndirectImageGatherHomogeneous";
|
||||
compiled.resources.flattened_srt[6] = homogeneous ? 4u : ordinals[5];
|
||||
compiled.resources.flattened_srt[7] = mixed || dimensions ? 5u : ordinals[6];
|
||||
compiled.resources.flattened_srt[8] = dimensions ? 2u : ordinals[7];
|
||||
if (fetch) {
|
||||
test.name = "IndirectImageFetchContiguousRuns";
|
||||
const auto position = std::ranges::find_if(*block, [&](const Inst &inst) {
|
||||
return &inst == access;
|
||||
});
|
||||
auto &read = *block->PrependNewInst(
|
||||
position, ValueOpcode::ImageRead,
|
||||
{Value(&image), Value(&address), Value(true)}, memory_flag_bits);
|
||||
access->ReplaceUsesWith(Value(&read), false);
|
||||
sampler.Invalidate();
|
||||
compiled.program.info.samplers.clear();
|
||||
compiled.program.info.sampled_pairs.clear();
|
||||
@@ -36497,7 +36318,7 @@ void CheckIndirectImageKeySwitch(VulkanHarness &vulkan) {
|
||||
const auto image_opcode = fetch ? "OpImageFetch" : gather ? "OpImageGather"
|
||||
: "OpImageSampleExplicitLod";
|
||||
Require(test.name, "native array image runs", tools.Disassemble(compiled.spirv, &text) &&
|
||||
CountText(text, image_opcode) == (homogeneous ? 1u : dimensions ? 5u : mixed ? 4u : 3u) * halves,
|
||||
CountText(text, image_opcode) == (homogeneous ? 1u : 3u) * halves,
|
||||
"mixed runs expanded into per-image accesses");
|
||||
if (homogeneous) {
|
||||
Require(test.name, "homogeneous gather array", text.find("OpSwitch") == std::string::npos,
|
||||
@@ -36519,37 +36340,25 @@ void CheckIndirectImageKeySwitch(VulkanHarness &vulkan) {
|
||||
Require(test.name, "nonuniform image operands",
|
||||
std::ranges::count_if(image_operands, [&](u32 operand) {
|
||||
return std::ranges::find(decorated, operand) != decorated.end();
|
||||
}) == (homogeneous ? 1u : mixed ? (fetch ? 1u : 4u) : 2u) * halves,
|
||||
}) == (homogeneous ? 1u : 2u) * halves,
|
||||
"a dynamically selected image operand lacks its nonuniform decoration");
|
||||
}
|
||||
test.initial.assign(wave_size, 0xdeadbeefu);
|
||||
test.expected.assign(wave_size, 0u);
|
||||
for (u32 lane = 0; lane < ordinals.size(); ++lane) {
|
||||
const auto resource = selected_resources[compiled.resources.flattened_srt[1u + lane]];
|
||||
test.expected[lane] = dimensions && resource == 6u
|
||||
? volume_mip_bits + ((lane & 1u) ? 3u : 1u) * 4u + 3u
|
||||
: dimensions && resource == 3u ? array_mip_bits + (lane & 1u) * 4u + 3u
|
||||
: mixed && resource == 6u
|
||||
? (fetch ? uint_texel + 10u : (lane & 1u) ? 1u : uint_texel)
|
||||
: std::bit_cast<u32>(float(resource + ((fetch || dimensions) && resource != 0u ? 10u : 0u)));
|
||||
test.expected[lane] = std::bit_cast<u32>(float(resource + (fetch && resource != 0u ? 10u : 0u)));
|
||||
}
|
||||
auto buffer = vulkan.CreateStorageBuffer(test.name, test.initial, test.initial.size());
|
||||
vulkan.Dispatch(test, compiled, buffer, nullptr, nullptr, nullptr, nullptr,
|
||||
fetch ? vk::Sampler{} : dimensions ? mip_sampler : native_sampler, textures,
|
||||
mixed && !fetch ? std::span<const vk::Sampler>(native_samplers) : std::span<const vk::Sampler>{});
|
||||
fetch ? vk::Sampler{} : native_sampler, textures);
|
||||
const auto actual = vulkan.ReadBuffer(test.name, buffer, test.expected.size());
|
||||
vulkan.DestroyBuffer(&buffer);
|
||||
CompareWords(test, "root, native slots, array runs and mapping bounds", test.expected, actual);
|
||||
}
|
||||
std::printf("[compute] %-32s ok\n", test.name);
|
||||
}
|
||||
vulkan.Device().destroySampler(integer_sampler);
|
||||
vulkan.Device().destroySampler(integer_border_sampler);
|
||||
vulkan.Device().destroySampler(native_sampler);
|
||||
vulkan.Device().destroySampler(mip_sampler);
|
||||
vulkan.DestroyImage(&uint_texture);
|
||||
vulkan.DestroyImage(&volume_texture);
|
||||
vulkan.DestroyImage(&array_texture);
|
||||
for (auto &texture : textures) vulkan.DestroyImage(&texture);
|
||||
}
|
||||
|
||||
|
||||
@@ -14257,8 +14257,6 @@ void TestBoundedScalarMaterialImageKeys() {
|
||||
uint64_t unavailable = 0;
|
||||
bool clean = true;
|
||||
bool mixed = false;
|
||||
bool mixed_dimension = false;
|
||||
bool mixed_conversion = false;
|
||||
bool same_key = false;
|
||||
uint32_t keys = 0;
|
||||
uint32_t images = 0;
|
||||
@@ -14287,13 +14285,10 @@ void TestBoundedScalarMaterialImageKeys() {
|
||||
std::ranges::fill(words, 0u);
|
||||
words[0] = key;
|
||||
const bool integer = key == 5573u || key == 5588u;
|
||||
words[1] = static_cast<uint32_t>(m.mixed_conversion && key == 5573u
|
||||
? Prospero::BufferFormat::k11_11_10UInt
|
||||
: integer && !(m.mixed && key == 5573u)
|
||||
? Prospero::BufferFormat::k32UInt : Prospero::BufferFormat::k32Float) << 20u;
|
||||
words[1] = static_cast<uint32_t>(integer && !(m.mixed && key == 5573u)
|
||||
? Prospero::BufferFormat::k32UInt : Prospero::BufferFormat::k32Float) << 20u;
|
||||
words[3] = DstSel(4, 5, 6, 7) |
|
||||
(static_cast<uint32_t>(m.mixed_dimension && key == 5573u
|
||||
? Prospero::ImageType::kColor2D : Prospero::ImageType::kColor3D) << 28u);
|
||||
(static_cast<uint32_t>(Prospero::ImageType::kColor3D) << 28u);
|
||||
return true;
|
||||
}
|
||||
return false;
|
||||
@@ -14340,22 +14335,8 @@ void TestBoundedScalarMaterialImageKeys() {
|
||||
Check(!MaterializeResources(plan, runtime, snapshot, specialization) && memory.keys == 0,
|
||||
"scalar material byte range escaped V# bounds");
|
||||
memory = {.mixed = true};
|
||||
Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
|
||||
snapshot.images.size() == 4u &&
|
||||
std::ranges::count_if(specialization.images, [](const auto& image) {
|
||||
return image.numeric_class == Prospero::TextureNumericClass::Uint;
|
||||
}) == 1,
|
||||
"selected material fields lost their mixed float/uint image classes");
|
||||
memory = {.mixed_dimension = true};
|
||||
Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
|
||||
snapshot.images.size() == 4u &&
|
||||
std::ranges::count_if(specialization.images, [](const auto& image) {
|
||||
return image.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim2D;
|
||||
}) == 1,
|
||||
"selected material fields lost their per-candidate image dimensions");
|
||||
memory = {.mixed_conversion = true};
|
||||
Check(!MaterializeResources(plan, runtime, snapshot, specialization),
|
||||
"selected material fields bypassed image conversion compatibility");
|
||||
"selected material fields bypassed image compatibility");
|
||||
memory = {};
|
||||
Check(MaterializeResources(plan, runtime, snapshot, specialization),
|
||||
"native scalar material image selection did not recover after rejected candidates");
|
||||
|
||||
Reference in New Issue
Block a user