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: merge typed image candidates as raw vectors with shared sampler variants

Browse Source
This commit is contained in:
nmzik
2026-10-12 18:54:54 +02:00
parent 49ebbfd627
commit 652154aead
5 changed files with 238 additions and 99 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/shader/recompiler/backend/spirv/spirvEmitterImage.cpp
+68 -48
View File
@@ -269,20 +269,13 @@ uint32_t ResultVector(ValueEmitContext& ctx, uint32_t value,
result, packed[0], packed[1], packed[2], packed[3]);
return result;
}
uint32_t component[4] {};
for (uint32_t index = 0; index < 4u; index++) {
if (dref) {
component[index] = index == 0u ? FloatBits(ctx, value) : ConstantU32(ctx.state, 0);
continue;
}
const auto scalar = ctx.state.builder.AllocateId();
ctx.state.builder.AddFunction(
spv::OpCompositeExtract, ImageScalarType(ctx.state, value_class), scalar, value, index);
component[index] = SampledComponentBits(ctx, scalar, value_class);
}
if (!dref)
return value_class == Prospero::TextureNumericClass::Uint
? value : Unary(ctx.state, spv::OpBitcast, TypeU32Vector(ctx.state, 4), value);
const auto result = ctx.state.builder.AllocateId();
const auto zero = ConstantU32(ctx.state, 0);
ctx.state.builder.AddFunction(spv::OpCompositeConstruct, TypeU32Vector(ctx.state, 4), result,
component[0], component[1], component[2], component[3]);
FloatBits(ctx, value), zero, zero, zero);
return result;
}
@@ -610,6 +603,45 @@ 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) {
@@ -735,15 +767,14 @@ 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),
[&]() {
const auto color = EmitImageAccess(
ctx, inst, ImageVectorType(state, numeric_class, 4),
return EmitImageAccess(
ctx, inst, TypeU32Vector(state, 4),
[&](uint32_t resource, uint32_t array_index) {
const auto& candidate = state.program.info.images[resource];
const auto dimension = candidate.dimension;
@@ -762,10 +793,9 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
descriptor, coord, spv::ImageOperandsLodMask,
LodU32(ctx, mem, *address, mem.image_dimension));
}
return fetched;
return ResultVector(ctx, UnpackImageTexel(ctx, mem, fetched),
candidate.numeric_class, false, mem);
});
return ResultVector(ctx, UnpackImageTexel(ctx, mem, color), numeric_class,
false, mem);
}));
return;
}
@@ -820,9 +850,6 @@ 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
@@ -831,9 +858,13 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
} else {
component_or_dref = ConstantU32(state, ImageGatherSource(state, mem));
}
const auto sampler_id = LoadSamplerDescriptor(state, mem.sampler);
const auto samplers = EmitImageSamplers(ctx, inst);
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,
@@ -847,7 +878,8 @@ 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);
const auto sampled = MakeSampledImage(state, resource, sampler.id, mip, array_index,
sampler.dynamic);
const auto sample = state.builder.AllocateId();
const std::array operands {operand_mask, offset};
state.builder.AddFunction(
@@ -856,15 +888,15 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
std::span(operands).first(operand_mask != 0u ? 2u : 0u));
return sample;
};
return candidate.mip_mode == IR::ImageMipMode::Dynamic && candidate.mip_count > 1u
const auto gathered = 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);
};
const auto sample = EmitImageAccess(ctx, inst, result_type, EmitGather);
ctx.Define(inst, ResultVector(ctx, UnpackImageGather(ctx, mem, sample),
result_numeric_class, false, mem, true));
ctx.Define(inst, EmitImageAccess(ctx, inst, TypeU32Vector(state, 4), EmitGather));
return;
}
const bool explicit_lod = HasFlag(mem, Decoder::ImageSampleFlagDerivative) ||
@@ -877,10 +909,8 @@ 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);
@@ -906,23 +936,16 @@ 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()) {
ctx.Fail(inst, "has a nonconstant sample texel offset");
}
const auto sampler_id = LoadSamplerDescriptor(state, mem.sampler, sampler_index);
const auto samplers = EmitImageSamplers(ctx, inst);
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) {
@@ -934,19 +957,16 @@ 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_index != 0u);
const auto sampled = MakeSampledImage(state, resource, sampler.id, 0u, array_index,
sampler.dynamic);
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 sample;
return ResultVector(ctx, dref ? sample : UnpackImageTexel(ctx, mem, sample),
candidate.numeric_class, dref, mem);
};
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));
ctx.Define(inst, EmitImageAccess(ctx, inst, TypeU32Vector(state, 4), EmitSample));
return;
}
if (image_info.access == IR::ImageAccess::Atomic) {
src/graphics/shader/recompiler/ir/passes/ResourceMaterialization.cpp
+19 -13
View File
@@ -106,16 +106,12 @@ Prospero::BufferFormat ImageConversionFormat(Prospero::BufferFormat format) {
: Prospero::BufferFormat::kInvalid;
}
enum class SamplerClass : uint8_t { Float, Integer, PointInteger };
enum class SamplerClass : uint8_t { Float, Integer };
SamplerClass ClassifySampler(const ImageResource& image) {
if (image.numeric_class == Prospero::TextureNumericClass::Sint ||
(image.numeric_class == Prospero::TextureNumericClass::Uint &&
image.conversion_format != Prospero::BufferFormat::kInvalid)) {
return SamplerClass::PointInteger;
}
return image.numeric_class == Prospero::TextureNumericClass::Uint ? SamplerClass::Integer
: SamplerClass::Float;
return image.numeric_class == Prospero::TextureNumericClass::Uint ||
image.numeric_class == Prospero::TextureNumericClass::Sint
? SamplerClass::Integer : SamplerClass::Float;
}
bool DescriptorIsCube(const DescriptorValue& descriptor) {
@@ -469,7 +465,7 @@ struct SamplerPlan {
uint32_t source;
SamplerClass type;
};
std::array<std::array<uint32_t, 3>, ShaderInfo::MaxSamplers> mapping;
std::array<std::array<uint32_t, 2>, ShaderInfo::MaxSamplers> mapping;
std::array<Binding, ShaderInfo::MaxSamplers> bindings;
uint32_t sampler_count = 0;
};
@@ -645,8 +641,7 @@ static bool BuildResourceSpecialization(const ResourcePlan& program, ResourceSna
}
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))) ||
if ((!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) {
@@ -1334,9 +1329,20 @@ 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++) {
@@ -1345,7 +1351,7 @@ void ApplyResourceSpecialization(Program& program, const ResourceSpecialization&
samplers.push_back(samplers[binding.source]);
}
samplers[index].snapshot_index = binding.source;
samplers[index].force_point_filtering = binding.type == SamplerClass::PointInteger;
samplers[index].force_point_filtering = binding.type == SamplerClass::Integer;
samplers[index].integer_border = binding.type != SamplerClass::Float;
}
for (uint32_t index = 0; index < sampler_plan.sampler_count; index++) {