N-archive/suyu
1
0
Fork 0
mirror of https://git.suyu.dev/suyu/suyu.git synced 2024-11-10 00:42:45 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions

shader: Implement SULD and SUST

Browse source
This commit is contained in:
ReinUsesLisp 2021-04-09 01:45:39 -03:00 committed by ameerj
parent 094da34456
commit 7cb2ab3585
31 changed files with 739 additions and 209 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

1
src/shader_recompiler/CMakeLists.txt
View file

@ -133,6 +133,7 @@ add_library(shader_recompiler STATIC
frontend/maxwell/translate/impl/predicate_set_predicate.cpp
frontend/maxwell/translate/impl/predicate_set_register.cpp
frontend/maxwell/translate/impl/select_source_with_predicate.cpp
frontend/maxwell/translate/impl/surface_load_store.cpp
frontend/maxwell/translate/impl/texture_fetch.cpp
frontend/maxwell/translate/impl/texture_fetch_swizzled.cpp
frontend/maxwell/translate/impl/texture_gather_swizzled.cpp

97
src/shader_recompiler/backend/spirv/emit_context.cpp
View file

@ -18,41 +18,70 @@ namespace {
Id ImageType(EmitContext& ctx, const TextureDescriptor& desc) {
const spv::ImageFormat format{spv::ImageFormat::Unknown};
const Id type{ctx.F32[1]};
const bool depth{desc.is_depth};
switch (desc.type) {
case TextureType::Color1D:
return ctx.TypeImage(type, spv::Dim::Dim1D, false, false, false, 1, format);
return ctx.TypeImage(type, spv::Dim::Dim1D, depth, false, false, 1, format);
case TextureType::ColorArray1D:
return ctx.TypeImage(type, spv::Dim::Dim1D, false, true, false, 1, format);
return ctx.TypeImage(type, spv::Dim::Dim1D, depth, true, false, 1, format);
case TextureType::Color2D:
return ctx.TypeImage(type, spv::Dim::Dim2D, false, false, false, 1, format);
return ctx.TypeImage(type, spv::Dim::Dim2D, depth, false, false, 1, format);
case TextureType::ColorArray2D:
return ctx.TypeImage(type, spv::Dim::Dim2D, false, true, false, 1, format);
return ctx.TypeImage(type, spv::Dim::Dim2D, depth, true, false, 1, format);
case TextureType::Color3D:
return ctx.TypeImage(type, spv::Dim::Dim3D, false, false, false, 1, format);
return ctx.TypeImage(type, spv::Dim::Dim3D, depth, false, false, 1, format);
case TextureType::ColorCube:
return ctx.TypeImage(type, spv::Dim::Cube, false, false, false, 1, format);
return ctx.TypeImage(type, spv::Dim::Cube, depth, false, false, 1, format);
case TextureType::ColorArrayCube:
return ctx.TypeImage(type, spv::Dim::Cube, false, true, false, 1, format);
case TextureType::Shadow1D:
return ctx.TypeImage(type, spv::Dim::Dim1D, true, false, false, 1, format);
case TextureType::ShadowArray1D:
return ctx.TypeImage(type, spv::Dim::Dim1D, true, true, false, 1, format);
case TextureType::Shadow2D:
return ctx.TypeImage(type, spv::Dim::Dim2D, true, false, false, 1, format);
case TextureType::ShadowArray2D:
return ctx.TypeImage(type, spv::Dim::Dim2D, true, true, false, 1, format);
case TextureType::Shadow3D:
return ctx.TypeImage(type, spv::Dim::Dim3D, true, false, false, 1, format);
case TextureType::ShadowCube:
return ctx.TypeImage(type, spv::Dim::Cube, true, false, false, 1, format);
case TextureType::ShadowArrayCube:
return ctx.TypeImage(type, spv::Dim::Cube, true, true, false, 1, format);
return ctx.TypeImage(type, spv::Dim::Cube, depth, true, false, 1, format);
case TextureType::Buffer:
break;
}
throw InvalidArgument("Invalid texture type {}", desc.type);
}
Id ImageType(EmitContext& ctx, const ImageDescriptor& desc) {
const spv::ImageFormat format{[&] {
switch (desc.format) {
case ImageFormat::Typeless:
return spv::ImageFormat::Unknown;
case ImageFormat::R8_UINT:
return spv::ImageFormat::R8ui;
case ImageFormat::R8_SINT:
return spv::ImageFormat::R8i;
case ImageFormat::R16_UINT:
return spv::ImageFormat::R16ui;
case ImageFormat::R16_SINT:
return spv::ImageFormat::R16i;
case ImageFormat::R32_UINT:
return spv::ImageFormat::R32ui;
case ImageFormat::R32G32_UINT:
return spv::ImageFormat::Rg32ui;
case ImageFormat::R32G32B32A32_UINT:
return spv::ImageFormat::Rgba32ui;
}
throw InvalidArgument("Invalid image format {}", desc.format);
}()};
const Id type{ctx.U32[1]};
switch (desc.type) {
case TextureType::Color1D:
return ctx.TypeImage(type, spv::Dim::Dim1D, false, false, false, 2, format);
case TextureType::ColorArray1D:
return ctx.TypeImage(type, spv::Dim::Dim1D, false, true, false, 2, format);
case TextureType::Color2D:
return ctx.TypeImage(type, spv::Dim::Dim2D, false, false, false, 2, format);
case TextureType::ColorArray2D:
return ctx.TypeImage(type, spv::Dim::Dim2D, false, true, false, 2, format);
case TextureType::Color3D:
return ctx.TypeImage(type, spv::Dim::Dim3D, false, false, false, 2, format);
case TextureType::Buffer:
throw NotImplementedException("Image buffer");
default:
break;
}
throw InvalidArgument("Invalid texture type {}", desc.type);
}
Id DefineVariable(EmitContext& ctx, Id type, std::optional<spv::BuiltIn> builtin,
spv::StorageClass storage_class) {
const Id pointer_type{ctx.TypePointer(storage_class, type)};
@ -134,6 +163,7 @@ EmitContext::EmitContext(const Profile& profile_, IR::Program& program, u32& bin
DefineStorageBuffers(program.info, binding);
DefineTextureBuffers(program.info, binding);
DefineTextures(program.info, binding);
DefineImages(program.info, binding);
DefineAttributeMemAccess(program.info);
DefineLabels(program);
}
@ -572,6 +602,31 @@ void EmitContext::DefineTextures(const Info& info, u32& binding) {
}
}
void EmitContext::DefineImages(const Info& info, u32& binding) {
images.reserve(info.image_descriptors.size());
for (const ImageDescriptor& desc : info.image_descriptors) {
if (desc.count != 1) {
throw NotImplementedException("Array of textures");
}
const Id image_type{ImageType(*this, desc)};
const Id pointer_type{TypePointer(spv::StorageClass::UniformConstant, image_type)};
const Id id{AddGlobalVariable(pointer_type, spv::StorageClass::UniformConstant)};
Decorate(id, spv::Decoration::Binding, binding);
Decorate(id, spv::Decoration::DescriptorSet, 0U);
Name(id, fmt::format("img{}_{:02x}", desc.cbuf_index, desc.cbuf_offset));
for (u32 index = 0; index < desc.count; ++index) {
images.push_back(ImageDefinition{
.id{id},
.image_type{image_type},
});
}
if (profile.supported_spirv >= 0x00010400) {
interfaces.push_back(id);
}
binding += desc.count;
}
}
void EmitContext::DefineLabels(IR::Program& program) {
for (IR::Block* const block : program.blocks) {
block->SetDefinition(OpLabel());

9
src/shader_recompiler/backend/spirv/emit_context.h
View file

@ -35,6 +35,11 @@ struct TextureDefinition {
Id image_type;
};
struct ImageDefinition {
Id id;
Id image_type;
};
struct UniformDefinitions {
Id U8{};
Id S8{};
@ -95,8 +100,9 @@ public:
std::array<UniformDefinitions, Info::MAX_CBUFS> cbufs{};
std::array<Id, Info::MAX_SSBOS> ssbos{};
std::vector<TextureDefinition> textures;
std::vector<Id> texture_buffers;
std::vector<TextureDefinition> textures;
std::vector<ImageDefinition> images;
Id workgroup_id{};
Id local_invocation_id{};
@ -156,6 +162,7 @@ private:
void DefineStorageBuffers(const Info& info, u32& binding);
void DefineTextureBuffers(const Info& info, u32& binding);
void DefineTextures(const Info& info, u32& binding);
void DefineImages(const Info& info, u32& binding);
void DefineAttributeMemAccess(const Info& info);
void DefineLabels(IR::Program& program);

1
src/shader_recompiler/backend/spirv/emit_spirv.cpp
View file

@ -253,6 +253,7 @@ void SetupCapabilities(const Profile& profile, const Info& info, EmitContext& ct
ctx.AddCapability(spv::Capability::ImageGatherExtended);
ctx.AddCapability(spv::Capability::ImageQuery);
ctx.AddCapability(spv::Capability::SampledBuffer);
ctx.AddCapability(spv::Capability::StorageImageReadWithoutFormat);
}
Id PhiArgDef(EmitContext& ctx, IR::Inst* inst, size_t index) {

6
src/shader_recompiler/backend/spirv/emit_spirv.h
View file

@ -369,6 +369,8 @@ Id EmitBindlessImageFetch(EmitContext&);
Id EmitBindlessImageQueryDimensions(EmitContext&);
Id EmitBindlessImageQueryLod(EmitContext&);
Id EmitBindlessImageGradient(EmitContext&);
Id EmitBindlessImageRead(EmitContext&);
Id EmitBindlessImageWrite(EmitContext&);
Id EmitBoundImageSampleImplicitLod(EmitContext&);
Id EmitBoundImageSampleExplicitLod(EmitContext&);
Id EmitBoundImageSampleDrefImplicitLod(EmitContext&);
@ -379,6 +381,8 @@ Id EmitBoundImageFetch(EmitContext&);
Id EmitBoundImageQueryDimensions(EmitContext&);
Id EmitBoundImageQueryLod(EmitContext&);
Id EmitBoundImageGradient(EmitContext&);
Id EmitBoundImageRead(EmitContext&);
Id EmitBoundImageWrite(EmitContext&);
Id EmitImageSampleImplicitLod(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords,
Id bias_lc, Id offset);
Id EmitImageSampleExplicitLod(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords,
@ -397,6 +401,8 @@ Id EmitImageQueryDimensions(EmitContext& ctx, IR::Inst* inst, const IR::Value& i
Id EmitImageQueryLod(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords);
Id EmitImageGradient(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords,
Id derivates, Id offset, Id lod_clamp);
Id EmitImageRead(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords);
void EmitImageWrite(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords, Id color);
Id EmitVoteAll(EmitContext& ctx, Id pred);
Id EmitVoteAny(EmitContext& ctx, Id pred);
Id EmitVoteEqual(EmitContext& ctx, Id pred);

46
src/shader_recompiler/backend/spirv/emit_spirv_image.cpp
View file

@ -144,6 +144,18 @@ Id TextureImage(EmitContext& ctx, const IR::Value& index, IR::TextureInstInfo in
}
}
Id Image(EmitContext& ctx, const IR::Value& index, IR::TextureInstInfo info) {
if (!index.IsImmediate()) {
throw NotImplementedException("Indirect image indexing");
}
if (info.type == TextureType::Buffer) {
throw NotImplementedException("Image buffer");
} else {
const ImageDefinition def{ctx.images.at(index.U32())};
return ctx.OpLoad(def.image_type, def.id);
}
}
Id Decorate(EmitContext& ctx, IR::Inst* inst, Id sample) {
const auto info{inst->Flags<IR::TextureInstInfo>()};
if (info.relaxed_precision != 0) {
@ -209,6 +221,14 @@ Id EmitBindlessImageGradient(EmitContext&) {
throw LogicError("Unreachable instruction");
}
Id EmitBindlessImageRead(EmitContext&) {
throw LogicError("Unreachable instruction");
}
Id EmitBindlessImageWrite(EmitContext&) {
throw LogicError("Unreachable instruction");
}
Id EmitBoundImageSampleImplicitLod(EmitContext&) {
throw LogicError("Unreachable instruction");
}
@ -249,6 +269,14 @@ Id EmitBoundImageGradient(EmitContext&) {
throw LogicError("Unreachable instruction");
}
Id EmitBoundImageRead(EmitContext&) {
throw LogicError("Unreachable instruction");
}
Id EmitBoundImageWrite(EmitContext&) {
throw LogicError("Unreachable instruction");
}
Id EmitImageSampleImplicitLod(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords,
Id bias_lc, Id offset) {
const auto info{inst->Flags<IR::TextureInstInfo>()};
@ -322,23 +350,16 @@ Id EmitImageQueryDimensions(EmitContext& ctx, IR::Inst* inst, const IR::Value& i
const auto mips{[&] { return ctx.OpImageQueryLevels(ctx.U32[1], image); }};
switch (info.type) {
case TextureType::Color1D:
case TextureType::Shadow1D:
return ctx.OpCompositeConstruct(ctx.U32[4], ctx.OpImageQuerySizeLod(ctx.U32[1], image, lod),
zero, zero, mips());
case TextureType::ColorArray1D:
case TextureType::Color2D:
case TextureType::ColorCube:
case TextureType::ShadowArray1D:
case TextureType::Shadow2D:
case TextureType::ShadowCube:
return ctx.OpCompositeConstruct(ctx.U32[4], ctx.OpImageQuerySizeLod(ctx.U32[2], image, lod),
zero, mips());
case TextureType::ColorArray2D:
case TextureType::Color3D:
case TextureType::ColorArrayCube:
case TextureType::ShadowArray2D:
case TextureType::Shadow3D:
case TextureType::ShadowArrayCube:
return ctx.OpCompositeConstruct(ctx.U32[4], ctx.OpImageQuerySizeLod(ctx.U32[3], image, lod),
mips());
case TextureType::Buffer:
@ -365,4 +386,15 @@ Id EmitImageGradient(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, I
coords, operands.Mask(), operands.Span());
}
Id EmitImageRead(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords) {
const auto info{inst->Flags<IR::TextureInstInfo>()};
return Emit(&EmitContext::OpImageSparseRead, &EmitContext::OpImageRead, ctx, inst, ctx.U32[4],
Image(ctx, index, info), coords, std::nullopt, std::span<const Id>{});
}
void EmitImageWrite(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords, Id color) {
const auto info{inst->Flags<IR::TextureInstInfo>()};
ctx.OpImageWrite(Image(ctx, index, info), coords, color);
}
} // namespace Shader::Backend::SPIRV

11
src/shader_recompiler/frontend/ir/ir_emitter.cpp
View file

@ -1620,6 +1620,17 @@ Value IREmitter::ImageGradient(const Value& handle, const Value& coords, const V
return Inst(op, Flags{info}, handle, coords, derivates, offset, lod_clamp);
}
Value IREmitter::ImageRead(const Value& handle, const Value& coords, TextureInstInfo info) {
const Opcode op{handle.IsImmediate() ? Opcode::BoundImageRead : Opcode::BindlessImageRead};
return Inst(op, Flags{info}, handle, coords);
}
void IREmitter::ImageWrite(const Value& handle, const Value& coords, const Value& color,
TextureInstInfo info) {
const Opcode op{handle.IsImmediate() ? Opcode::BoundImageWrite : Opcode::BindlessImageWrite};
Inst(op, Flags{info}, handle, coords, color);
}
U1 IREmitter::VoteAll(const U1& value) {
return Inst<U1>(Opcode::VoteAll, value);
}

7
src/shader_recompiler/frontend/ir/ir_emitter.h
View file

@ -265,20 +265,19 @@ public:
[[nodiscard]] Value ImageQueryLod(const Value& handle, const Value& coords,
TextureInstInfo info);
[[nodiscard]] Value ImageGather(const Value& handle, const Value& coords, const Value& offset,
const Value& offset2, TextureInstInfo info);
[[nodiscard]] Value ImageGatherDref(const Value& handle, const Value& coords,
const Value& offset, const Value& offset2, const F32& dref,
TextureInstInfo info);
[[nodiscard]] Value ImageFetch(const Value& handle, const Value& coords, const Value& offset,
const U32& lod, const U32& multisampling, TextureInstInfo info);
[[nodiscard]] Value ImageGradient(const Value& handle, const Value& coords,
const Value& derivates, const Value& offset,
const F32& lod_clamp, TextureInstInfo info);
[[nodiscard]] Value ImageRead(const Value& handle, const Value& coords, TextureInstInfo info);
[[nodiscard]] void ImageWrite(const Value& handle, const Value& coords, const Value& color,
TextureInstInfo info);
[[nodiscard]] U1 VoteAll(const U1& value);
[[nodiscard]] U1 VoteAny(const U1& value);

12
src/shader_recompiler/frontend/ir/modifiers.h
View file

@ -43,11 +43,13 @@ static_assert(sizeof(FpControl) <= sizeof(u32));
union TextureInstInfo {
u32 raw;
BitField<0, 8, TextureType> type;
BitField<8, 1, u32> has_bias;
BitField<9, 1, u32> has_lod_clamp;
BitField<10, 1, u32> relaxed_precision;
BitField<11, 2, u32> gather_component;
BitField<13, 2, u32> num_derivates;
BitField<8, 1, u32> is_depth;
BitField<9, 1, u32> has_bias;
BitField<10, 1, u32> has_lod_clamp;
BitField<11, 1, u32> relaxed_precision;
BitField<12, 2, u32> gather_component;
BitField<14, 2, u32> num_derivates;
BitField<16, 3, ImageFormat> image_format;
};
static_assert(sizeof(TextureInstInfo) <= sizeof(u32));

6
src/shader_recompiler/frontend/ir/opcodes.inc
View file

@ -389,6 +389,8 @@ OPCODE(BindlessImageFetch, F32x4, U32,
OPCODE(BindlessImageQueryDimensions, U32x4, U32, U32, )
OPCODE(BindlessImageQueryLod, F32x4, U32, Opaque, )
OPCODE(BindlessImageGradient, F32x4, U32, Opaque, Opaque, Opaque, Opaque, )
OPCODE(BindlessImageRead, U32x4, U32, Opaque, )
OPCODE(BindlessImageWrite, Void, U32, Opaque, U32x4, )
OPCODE(BoundImageSampleImplicitLod, F32x4, U32, Opaque, Opaque, Opaque, )
OPCODE(BoundImageSampleExplicitLod, F32x4, U32, Opaque, Opaque, Opaque, )
@ -400,6 +402,8 @@ OPCODE(BoundImageFetch, F32x4, U32,
OPCODE(BoundImageQueryDimensions, U32x4, U32, U32, )
OPCODE(BoundImageQueryLod, F32x4, U32, Opaque, )
OPCODE(BoundImageGradient, F32x4, U32, Opaque, Opaque, Opaque, Opaque, )
OPCODE(BoundImageRead, U32x4, U32, Opaque, )
OPCODE(BoundImageWrite, Void, U32, Opaque, U32x4, )
OPCODE(ImageSampleImplicitLod, F32x4, U32, Opaque, Opaque, Opaque, )
OPCODE(ImageSampleExplicitLod, F32x4, U32, Opaque, Opaque, Opaque, )
@ -411,6 +415,8 @@ OPCODE(ImageFetch, F32x4, U32,
OPCODE(ImageQueryDimensions, U32x4, U32, U32, )
OPCODE(ImageQueryLod, F32x4, U32, Opaque, )
OPCODE(ImageGradient, F32x4, U32, Opaque, Opaque, Opaque, Opaque, )
OPCODE(ImageRead, U32x4, U32, Opaque, )
OPCODE(ImageWrite, Void, U32, Opaque, U32x4, )
// Warp operations
OPCODE(VoteAll, U1, U1, )

8
src/shader_recompiler/frontend/maxwell/translate/impl/not_implemented.cpp
View file

@ -281,18 +281,10 @@ void TranslatorVisitor::SUATOM_cas(u64) {
ThrowNotImplemented(Opcode::SUATOM_cas);
}
void TranslatorVisitor::SULD(u64) {
ThrowNotImplemented(Opcode::SULD);
}
void TranslatorVisitor::SURED(u64) {
ThrowNotImplemented(Opcode::SURED);
}
void TranslatorVisitor::SUST(u64) {
ThrowNotImplemented(Opcode::SUST);
}
void TranslatorVisitor::SYNC(u64) {
ThrowNotImplemented(Opcode::SYNC);
}

280
src/shader_recompiler/frontend/maxwell/translate/impl/surface_load_store.cpp Normal file
View file

@ -0,0 +1,280 @@
// Copyright 2021 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#include <array>
#include <bit>
#include "common/bit_field.h"
#include "common/common_types.h"
#include "shader_recompiler/frontend/ir/modifiers.h"
#include "shader_recompiler/frontend/maxwell/translate/impl/impl.h"
namespace Shader::Maxwell {
namespace {
enum class Type : u64 {
_1D,
BUFFER_1D,
ARRAY_1D,
_2D,
ARRAY_2D,
_3D,
};
constexpr unsigned R = 1 << 0;
constexpr unsigned G = 1 << 1;
constexpr unsigned B = 1 << 2;
constexpr unsigned A = 1 << 3;
constexpr std::array MASK{
0U, //
R, //
G, //
R | G, //
B, //
R | B, //
G | B, //
R | G | B, //
A, //
R | A, //
G | A, //
R | G | A, //
B | A, //
R | B | A, //
G | B | A, //
R | G | B | A, //
};
enum class Size : u64 {
U8,
S8,
U16,
S16,
B32,
B64,
B128,
};
enum class Clamp : u64 {
IGN,
Default,
TRAP,
};
enum class LoadCache : u64 {
Default,
CG,
CI,
CV,
};
enum class StoreCache : u64 {
Default,
CG,
CS,
WT,
};
ImageFormat Format(Size size) {
switch (size) {
case Size::U8:
return ImageFormat::R8_UINT;
case Size::S8:
return ImageFormat::R8_SINT;
case Size::U16:
return ImageFormat::R16_UINT;
case Size::S16:
return ImageFormat::R16_SINT;
case Size::B32:
return ImageFormat::R32_UINT;
case Size::B64:
return ImageFormat::R32G32_UINT;
case Size::B128:
return ImageFormat::R32G32B32A32_UINT;
}
throw NotImplementedException("Invalid size {}", size);
}
int SizeInRegs(Size size) {
switch (size) {
case Size::U8:
case Size::S8:
case Size::U16:
case Size::S16:
case Size::B32:
return 1;
case Size::B64:
return 2;
case Size::B128:
return 4;
}
throw NotImplementedException("Invalid size {}", size);
}
TextureType GetType(Type type) {
switch (type) {
case Type::_1D:
return TextureType::Color1D;
case Type::BUFFER_1D:
return TextureType::Buffer;
case Type::ARRAY_1D:
return TextureType::ColorArray1D;
case Type::_2D:
return TextureType::Color2D;
case Type::ARRAY_2D:
return TextureType::ColorArray2D;
case Type::_3D:
return TextureType::Color3D;
}
throw NotImplementedException("Invalid type {}", type);
}
IR::Value MakeCoords(TranslatorVisitor& v, IR::Reg reg, Type type) {
const auto array{[&](int index) {
return v.ir.BitFieldExtract(v.X(reg + index), v.ir.Imm32(0), v.ir.Imm32(16));
}};
switch (type) {
case Type::_1D:
case Type::BUFFER_1D:
return v.X(reg);
case Type::ARRAY_1D:
return v.ir.CompositeConstruct(v.X(reg), array(1));
case Type::_2D:
return v.ir.CompositeConstruct(v.X(reg), v.X(reg + 1));
case Type::ARRAY_2D:
return v.ir.CompositeConstruct(v.X(reg), v.X(reg + 1), array(2));
case Type::_3D:
return v.ir.CompositeConstruct(v.X(reg), v.X(reg + 1), v.X(reg + 3));
}
throw NotImplementedException("Invalid type {}", type);
}
unsigned SwizzleMask(u64 swizzle) {
if (swizzle == 0 || swizzle >= MASK.size()) {
throw NotImplementedException("Invalid swizzle {}", swizzle);
}
return MASK[swizzle];
}
IR::Value MakeColor(IR::IREmitter& ir, IR::Reg reg, int num_regs) {
std::array<IR::U32, 4> colors;
for (int i = 0; i < num_regs; ++i) {
colors[i] = ir.GetReg(reg + i);
}
for (int i = num_regs; i < 4; ++i) {
colors[i] = ir.Imm32(0);
}
return ir.CompositeConstruct(colors[0], colors[1], colors[2], colors[3]);
}
} // Anonymous namespace
void TranslatorVisitor::SULD(u64 insn) {
union {
u64 raw;
BitField<51, 1, u64> is_bound;
BitField<52, 1, u64> d;
BitField<23, 1, u64> ba;
BitField<33, 3, Type> type;
BitField<24, 2, LoadCache> cache;
BitField<20, 3, Size> size; // .D
BitField<20, 4, u64> swizzle; // .P
BitField<49, 2, Clamp> clamp;
BitField<0, 8, IR::Reg> dest_reg;
BitField<8, 8, IR::Reg> coord_reg;
BitField<36, 13, u64> bound_offset; // is_bound
BitField<39, 8, IR::Reg> bindless_reg; // !is_bound
} const suld{insn};
if (suld.clamp != Clamp::IGN) {
throw NotImplementedException("Clamp {}", suld.clamp.Value());
}
if (suld.cache != LoadCache::Default) {
throw NotImplementedException("Cache {}", suld.cache.Value());
}
const bool is_typed{suld.d != 0};
if (is_typed && suld.ba != 0) {
throw NotImplementedException("BA");
}
const ImageFormat format{is_typed ? Format(suld.size) : ImageFormat::Typeless};
const TextureType type{GetType(suld.type)};
const IR::Value coords{MakeCoords(*this, suld.coord_reg, suld.type)};
const IR::U32 handle{suld.is_bound != 0 ? ir.Imm32(static_cast<u32>(suld.bound_offset * 4))
: X(suld.bindless_reg)};
IR::TextureInstInfo info{};
info.type.Assign(type);
info.image_format.Assign(format);
const IR::Value result{ir.ImageRead(handle, coords, info)};
IR::Reg dest_reg{suld.dest_reg};
if (is_typed) {
const int num_regs{SizeInRegs(suld.size)};
for (int i = 0; i < num_regs; ++i) {
X(dest_reg + i, IR::U32{ir.CompositeExtract(result, i)});
}
} else {
const unsigned mask{SwizzleMask(suld.swizzle)};
const int bits{std::popcount(mask)};
if (!IR::IsAligned(dest_reg, bits == 3 ? 4 : bits)) {
throw NotImplementedException("Unaligned destination register");
}
for (unsigned component = 0; component < 4; ++component) {
if (((mask >> component) & 1) == 0) {
continue;
}
X(dest_reg, IR::U32{ir.CompositeExtract(result, component)});
++dest_reg;
}
}
}
void TranslatorVisitor::SUST(u64 insn) {
union {
u64 raw;
BitField<51, 1, u64> is_bound;
BitField<52, 1, u64> d;
BitField<23, 1, u64> ba;
BitField<33, 3, Type> type;
BitField<24, 2, StoreCache> cache;
BitField<20, 3, Size> size; // .D
BitField<20, 4, u64> swizzle; // .P
BitField<49, 2, Clamp> clamp;
BitField<0, 8, IR::Reg> data_reg;
BitField<8, 8, IR::Reg> coord_reg;
BitField<36, 13, u64> bound_offset; // is_bound
BitField<39, 8, IR::Reg> bindless_reg; // !is_bound
} const sust{insn};
if (sust.clamp != Clamp::IGN) {
throw NotImplementedException("Clamp {}", sust.clamp.Value());
}
if (sust.cache != StoreCache::Default) {
throw NotImplementedException("Cache {}", sust.cache.Value());
}
const bool is_typed{sust.d != 0};
if (is_typed && sust.ba != 0) {
throw NotImplementedException("BA");
}
const ImageFormat format{is_typed ? Format(sust.size) : ImageFormat::Typeless};
const TextureType type{GetType(sust.type)};
const IR::Value coords{MakeCoords(*this, sust.coord_reg, sust.type)};
const IR::U32 handle{sust.is_bound != 0 ? ir.Imm32(static_cast<u32>(sust.bound_offset * 4))
: X(sust.bindless_reg)};
IR::TextureInstInfo info{};
info.type.Assign(type);
info.image_format.Assign(format);
IR::Value color;
if (is_typed) {
color = MakeColor(ir, sust.data_reg, SizeInRegs(sust.size));
} else {
const unsigned mask{SwizzleMask(sust.swizzle)};
if (mask != 0xf) {
throw NotImplementedException("Non-full mask");
}
color = MakeColor(ir, sust.data_reg, 4);
}
ir.ImageWrite(handle, coords, color, info);
}
} // namespace Shader::Maxwell

19
src/shader_recompiler/frontend/maxwell/translate/impl/texture_fetch.cpp
View file

@ -33,24 +33,24 @@ enum class TextureType : u64 {
ARRAY_CUBE,
};
Shader::TextureType GetType(TextureType type, bool dc) {
Shader::TextureType GetType(TextureType type) {
switch (type) {
case TextureType::_1D:
return dc ? Shader::TextureType::Shadow1D : Shader::TextureType::Color1D;
return Shader::TextureType::Color1D;
case TextureType::ARRAY_1D:
return dc ? Shader::TextureType::ShadowArray1D : Shader::TextureType::ColorArray1D;
return Shader::TextureType::ColorArray1D;
case TextureType::_2D:
return dc ? Shader::TextureType::Shadow2D : Shader::TextureType::Color2D;
return Shader::TextureType::Color2D;
case TextureType::ARRAY_2D:
return dc ? Shader::TextureType::ShadowArray2D : Shader::TextureType::ColorArray2D;
return Shader::TextureType::ColorArray2D;
case TextureType::_3D:
return dc ? Shader::TextureType::Shadow3D : Shader::TextureType::Color3D;
return Shader::TextureType::Color3D;
case TextureType::ARRAY_3D:
throw NotImplementedException("3D array texture type");
case TextureType::CUBE:
return dc ? Shader::TextureType::ShadowCube : Shader::TextureType::ColorCube;
return Shader::TextureType::ColorCube;
case TextureType::ARRAY_CUBE:
return dc ? Shader::TextureType::ShadowArrayCube : Shader::TextureType::ColorArrayCube;
return Shader::TextureType::ColorArrayCube;
}
throw NotImplementedException("Invalid texture type {}", type);
}
@ -169,7 +169,8 @@ void Impl(TranslatorVisitor& v, u64 insn, bool aoffi, Blod blod, bool lc,
dref = v.F(meta_reg++);
}
IR::TextureInstInfo info{};
info.type.Assign(GetType(tex.type, tex.dc != 0));
info.type.Assign(GetType(tex.type));
info.is_depth.Assign(tex.dc != 0 ? 1 : 0);
info.has_bias.Assign(blod == Blod::LB || blod == Blod::LBA ? 1 : 0);
info.has_lod_clamp.Assign(lc ? 1 : 0);

12
src/shader_recompiler/frontend/maxwell/translate/impl/texture_fetch_swizzled.cpp
View file

@ -95,18 +95,21 @@ IR::Value Sample(TranslatorVisitor& v, u64 insn) {
{}, info);
case 4: // 2D.DC
CheckAlignment(reg_a, 2);
info.type.Assign(TextureType::Shadow2D);
info.type.Assign(TextureType::Color2D);
info.is_depth.Assign(1);
return v.ir.ImageSampleDrefImplicitLod(handle, Composite(v, reg_a, reg_a + 1), v.F(reg_b),
{}, {}, {}, info);
case 5: // 2D.LL.DC
CheckAlignment(reg_a, 2);
CheckAlignment(reg_b, 2);
info.type.Assign(TextureType::Shadow2D);
info.type.Assign(TextureType::Color2D);
info.is_depth.Assign(1);
return v.ir.ImageSampleDrefExplicitLod(handle, Composite(v, reg_a, reg_a + 1),
v.F(reg_b + 1), v.F(reg_b), {}, {}, info);
case 6: // 2D.LZ.DC
CheckAlignment(reg_a, 2);
info.type.Assign(TextureType::Shadow2D);
info.type.Assign(TextureType::Color2D);
info.is_depth.Assign(1);
return v.ir.ImageSampleDrefExplicitLod(handle, Composite(v, reg_a, reg_a + 1), v.F(reg_b),
zero, {}, {}, info);
case 7: // ARRAY_2D
@ -124,7 +127,8 @@ IR::Value Sample(TranslatorVisitor& v, u64 insn) {
case 9: // ARRAY_2D.LZ.DC
CheckAlignment(reg_a, 2);
CheckAlignment(reg_b, 2);
info.type.Assign(TextureType::ShadowArray2D);
info.type.Assign(TextureType::ColorArray2D);
info.is_depth.Assign(1);
return v.ir.ImageSampleDrefExplicitLod(
handle, v.ir.CompositeConstruct(v.F(reg_a + 1), v.F(reg_b), ReadArray(v, v.X(reg_a))),
v.F(reg_b + 1), zero, {}, {}, info);

19
src/shader_recompiler/frontend/maxwell/translate/impl/texture_gather.cpp
View file

@ -37,24 +37,24 @@ enum class ComponentType : u64 {
A = 3,
};
Shader::TextureType GetType(TextureType type, bool dc) {
Shader::TextureType GetType(TextureType type) {
switch (type) {
case TextureType::_1D:
return dc ? Shader::TextureType::Shadow1D : Shader::TextureType::Color1D;
return Shader::TextureType::Color1D;
case TextureType::ARRAY_1D:
return dc ? Shader::TextureType::ShadowArray1D : Shader::TextureType::ColorArray1D;
return Shader::TextureType::ColorArray1D;
case TextureType::_2D:
return dc ? Shader::TextureType::Shadow2D : Shader::TextureType::Color2D;
return Shader::TextureType::Color2D;
case TextureType::ARRAY_2D:
return dc ? Shader::TextureType::ShadowArray2D : Shader::TextureType::ColorArray2D;
return Shader::TextureType::ColorArray2D;
case TextureType::_3D:
return dc ? Shader::TextureType::Shadow3D : Shader::TextureType::Color3D;
return Shader::TextureType::Color3D;
case TextureType::ARRAY_3D:
throw NotImplementedException("3D array texture type");
case TextureType::CUBE:
return dc ? Shader::TextureType::ShadowCube : Shader::TextureType::ColorCube;
return Shader::TextureType::ColorCube;
case TextureType::ARRAY_CUBE:
return dc ? Shader::TextureType::ShadowArrayCube : Shader::TextureType::ColorArrayCube;
return Shader::TextureType::ColorArrayCube;
}
throw NotImplementedException("Invalid texture type {}", type);
}
@ -163,7 +163,8 @@ void Impl(TranslatorVisitor& v, u64 insn, ComponentType component_type, OffsetTy
dref = v.F(meta_reg++);
}
IR::TextureInstInfo info{};
info.type.Assign(GetType(tld4.type, tld4.dc != 0));
info.type.Assign(GetType(tld4.type));
info.is_depth.Assign(tld4.dc != 0 ? 1 : 0);
info.gather_component.Assign(static_cast<u32>(component_type));
const IR::Value sample{[&] {
if (tld4.dc == 0) {

3
src/shader_recompiler/frontend/maxwell/translate/impl/texture_gather_swizzled.cpp
View file

@ -59,7 +59,8 @@ IR::Value Sample(TranslatorVisitor& v, u64 insn) {
info.relaxed_precision.Assign(1);
}
info.gather_component.Assign(static_cast<u32>(tld4s.component_type.Value()));
info.type.Assign(tld4s.dc != 0 ? Shader::TextureType::Shadow2D : Shader::TextureType::Color2D);
info.type.Assign(Shader::TextureType::Color2D);
info.is_depth.Assign(tld4s.dc != 0 ? 1 : 0);
IR::Value coords;
if (tld4s.aoffi != 0) {
CheckAlignment(reg_a, 2);

18
src/shader_recompiler/frontend/maxwell/translate/impl/texture_gradient.cpp
View file

@ -23,24 +23,24 @@ enum class TextureType : u64 {
ARRAY_CUBE,
};
Shader::TextureType GetType(TextureType type, bool dc) {
Shader::TextureType GetType(TextureType type) {
switch (type) {
case TextureType::_1D:
return dc ? Shader::TextureType::Shadow1D : Shader::TextureType::Color1D;
return Shader::TextureType::Color1D;
case TextureType::ARRAY_1D:
return dc ? Shader::TextureType::ShadowArray1D : Shader::TextureType::ColorArray1D;
return Shader::TextureType::ColorArray1D;
case TextureType::_2D:
return dc ? Shader::TextureType::Shadow2D : Shader::TextureType::Color2D;
return Shader::TextureType::Color2D;
case TextureType::ARRAY_2D:
return dc ? Shader::TextureType::ShadowArray2D : Shader::TextureType::ColorArray2D;
return Shader::TextureType::ColorArray2D;
case TextureType::_3D:
return dc ? Shader::TextureType::Shadow3D : Shader::TextureType::Color3D;
return Shader::TextureType::Color3D;
case TextureType::ARRAY_3D:
throw NotImplementedException("3D array texture type");
case TextureType::CUBE:
return dc ? Shader::TextureType::ShadowCube : Shader::TextureType::ColorCube;
return Shader::TextureType::ColorCube;
case TextureType::ARRAY_CUBE:
return dc ? Shader::TextureType::ShadowArrayCube : Shader::TextureType::ColorArrayCube;
return Shader::TextureType::ColorArrayCube;
}
throw NotImplementedException("Invalid texture type {}", type);
}
@ -152,7 +152,7 @@ void Impl(TranslatorVisitor& v, u64 insn, bool is_bindless) {
}
IR::TextureInstInfo info{};
info.type.Assign(GetType(txd.type, false));
info.type.Assign(GetType(txd.type));
info.num_derivates.Assign(num_derivates);
info.has_lod_clamp.Assign(has_lod_clamp ? 1 : 0);
const IR::Value sample{v.ir.ImageGradient(handle, coords, derivates, offset, lod_clamp, info)};

18
src/shader_recompiler/frontend/maxwell/translate/impl/texture_load.cpp
View file

@ -23,24 +23,24 @@ enum class TextureType : u64 {
ARRAY_CUBE,
};
Shader::TextureType GetType(TextureType type, bool dc) {
Shader::TextureType GetType(TextureType type) {
switch (type) {
case TextureType::_1D:
return dc ? Shader::TextureType::Shadow1D : Shader::TextureType::Color1D;
return Shader::TextureType::Color1D;
case TextureType::ARRAY_1D:
return dc ? Shader::TextureType::ShadowArray1D : Shader::TextureType::ColorArray1D;
return Shader::TextureType::ColorArray1D;
case TextureType::_2D:
return dc ? Shader::TextureType::Shadow2D : Shader::TextureType::Color2D;
return Shader::TextureType::Color2D;
case TextureType::ARRAY_2D:
return dc ? Shader::TextureType::ShadowArray2D : Shader::TextureType::ColorArray2D;
return Shader::TextureType::ColorArray2D;
case TextureType::_3D:
return dc ? Shader::TextureType::Shadow3D : Shader::TextureType::Color3D;
return Shader::TextureType::Color3D;
case TextureType::ARRAY_3D:
throw NotImplementedException("3D array texture type");
case TextureType::CUBE:
return dc ? Shader::TextureType::ShadowCube : Shader::TextureType::ColorCube;
return Shader::TextureType::ColorCube;
case TextureType::ARRAY_CUBE:
return dc ? Shader::TextureType::ShadowArrayCube : Shader::TextureType::ColorArrayCube;
return Shader::TextureType::ColorArrayCube;
}
throw NotImplementedException("Invalid texture type {}", type);
}
@ -137,7 +137,7 @@ void Impl(TranslatorVisitor& v, u64 insn, bool is_bindless) {
throw NotImplementedException("TLD.CL - CLAMP is not implmented");
}
IR::TextureInstInfo info{};
info.type.Assign(GetType(tld.type, false));
info.type.Assign(GetType(tld.type));
const IR::Value sample{v.ir.ImageFetch(handle, coords, offset, lod, multisample, info)};
IR::Reg dest_reg{tld.dest_reg};

2
src/shader_recompiler/frontend/maxwell/translate/impl/texture_load_swizzled.cpp
View file

@ -2,7 +2,7 @@
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#include <utility>
#include <array>
#include "common/bit_field.h"
#include "common/common_types.h"

18
src/shader_recompiler/frontend/maxwell/translate/impl/texture_mipmap_level.cpp
View file

@ -23,24 +23,24 @@ enum class TextureType : u64 {
ARRAY_CUBE,
};
Shader::TextureType GetType(TextureType type, bool dc) {
Shader::TextureType GetType(TextureType type) {
switch (type) {
case TextureType::_1D:
return dc ? Shader::TextureType::Shadow1D : Shader::TextureType::Color1D;
return Shader::TextureType::Color1D;
case TextureType::ARRAY_1D:
return dc ? Shader::TextureType::ShadowArray1D : Shader::TextureType::ColorArray1D;
return Shader::TextureType::ColorArray1D;
case TextureType::_2D:
return dc ? Shader::TextureType::Shadow2D : Shader::TextureType::Color2D;
return Shader::TextureType::Color2D;
case TextureType::ARRAY_2D:
return dc ? Shader::TextureType::ShadowArray2D : Shader::TextureType::ColorArray2D;
return Shader::TextureType::ColorArray2D;
case TextureType::_3D:
return dc ? Shader::TextureType::Shadow3D : Shader::TextureType::Color3D;
return Shader::TextureType::Color3D;
case TextureType::ARRAY_3D:
throw NotImplementedException("3D array texture type");
case TextureType::CUBE:
return dc ? Shader::TextureType::ShadowCube : Shader::TextureType::ColorCube;
return Shader::TextureType::ColorCube;
case TextureType::ARRAY_CUBE:
return dc ? Shader::TextureType::ShadowArrayCube : Shader::TextureType::ColorArrayCube;
return Shader::TextureType::ColorArrayCube;
}
throw NotImplementedException("Invalid texture type {}", type);
}
@ -97,7 +97,7 @@ void Impl(TranslatorVisitor& v, u64 insn, bool is_bindless) {
handle = v.ir.Imm32(static_cast<u32>(tmml.cbuf_offset.Value() * 4));
}
IR::TextureInstInfo info{};
info.type.Assign(GetType(tmml.type, false));
info.type.Assign(GetType(tmml.type));
const IR::Value sample{v.ir.ImageQueryLod(handle, coords, info)};
IR::Reg dest_reg{tmml.dest_reg};

3
src/shader_recompiler/ir_opt/collect_shader_info_pass.cpp
View file

@ -416,8 +416,7 @@ void VisitUsages(Info& info, IR::Inst& inst) {
case IR::Opcode::ImageQueryLod:
case IR::Opcode::ImageGradient: {
const TextureType type{inst.Flags<IR::TextureInstInfo>().type};
info.uses_sampled_1d |= type == TextureType::Color1D || type == TextureType::ColorArray1D ||
type == TextureType::Shadow1D || type == TextureType::ShadowArray1D;
info.uses_sampled_1d |= type == TextureType::Color1D || type == TextureType::ColorArray1D;
info.uses_sparse_residency |=
inst.GetAssociatedPseudoOperation(IR::Opcode::GetSparseFromOp) != nullptr;
break;

91
src/shader_recompiler/ir_opt/texture_pass.cpp
View file

@ -61,6 +61,12 @@ IR::Opcode IndexedInstruction(const IR::Inst& inst) {
case IR::Opcode::BoundImageGradient:
case IR::Opcode::BindlessImageGradient:
return IR::Opcode::ImageGradient;
case IR::Opcode::BoundImageRead:
case IR::Opcode::BindlessImageRead:
return IR::Opcode::ImageRead;
case IR::Opcode::BoundImageWrite:
case IR::Opcode::BindlessImageWrite:
return IR::Opcode::ImageWrite;
default:
return IR::Opcode::Void;
}
@ -78,6 +84,8 @@ bool IsBindless(const IR::Inst& inst) {
case IR::Opcode::BindlessImageQueryDimensions:
case IR::Opcode::BindlessImageQueryLod:
case IR::Opcode::BindlessImageGradient:
case IR::Opcode::BindlessImageRead:
case IR::Opcode::BindlessImageWrite:
return true;
case IR::Opcode::BoundImageSampleImplicitLod:
case IR::Opcode::BoundImageSampleExplicitLod:
@ -89,6 +97,8 @@ bool IsBindless(const IR::Inst& inst) {
case IR::Opcode::BoundImageQueryDimensions:
case IR::Opcode::BoundImageQueryLod:
case IR::Opcode::BoundImageGradient:
case IR::Opcode::BoundImageRead:
case IR::Opcode::BoundImageWrite:
return false;
default:
throw InvalidArgument("Invalid opcode {}", inst.GetOpcode());
@ -147,10 +157,18 @@ TextureInst MakeInst(Environment& env, IR::Block* block, IR::Inst& inst) {
class Descriptors {
public:
explicit Descriptors(TextureDescriptors& texture_descriptors_,
TextureBufferDescriptors& texture_buffer_descriptors_)
: texture_descriptors{texture_descriptors_}, texture_buffer_descriptors{
texture_buffer_descriptors_} {}
explicit Descriptors(TextureBufferDescriptors& texture_buffer_descriptors_,
TextureDescriptors& texture_descriptors_,
ImageDescriptors& image_descriptors_)
: texture_buffer_descriptors{texture_buffer_descriptors_},
texture_descriptors{texture_descriptors_}, image_descriptors{image_descriptors_} {}
u32 Add(const TextureBufferDescriptor& desc) {
return Add(texture_buffer_descriptors, desc, [&desc](const auto& existing) {
return desc.cbuf_index == existing.cbuf_index &&
desc.cbuf_offset == existing.cbuf_offset;
});
}
u32 Add(const TextureDescriptor& desc) {
return Add(texture_descriptors, desc, [&desc](const auto& existing) {
@ -159,11 +177,14 @@ public:
});
}
u32 Add(const TextureBufferDescriptor& desc) {
return Add(texture_buffer_descriptors, desc, [&desc](const auto& existing) {
return desc.cbuf_index == existing.cbuf_index &&
u32 Add(const ImageDescriptor& desc) {
const u32 index{Add(image_descriptors, desc, [&desc](const auto& existing) {
return desc.type == existing.type && desc.format == existing.format &&
desc.cbuf_index == existing.cbuf_index &&
desc.cbuf_offset == existing.cbuf_offset;
});
})};
image_descriptors[index].is_written |= desc.is_written;
return index;
}
private:
@ -178,8 +199,9 @@ private:
return static_cast<u32>(descriptors.size()) - 1;
}
TextureDescriptors& texture_descriptors;
TextureBufferDescriptors& texture_buffer_descriptors;
TextureDescriptors& texture_descriptors;
ImageDescriptors& image_descriptors;
};
} // Anonymous namespace
@ -201,8 +223,9 @@ void TexturePass(Environment& env, IR::Program& program) {
return lhs.cbuf.index < rhs.cbuf.index;
});
Descriptors descriptors{
program.info.texture_descriptors,
program.info.texture_buffer_descriptors,
program.info.texture_descriptors,
program.info.image_descriptors,
};
for (TextureInst& texture_inst : to_replace) {
// TODO: Handle arrays
@ -233,19 +256,41 @@ void TexturePass(Environment& env, IR::Program& program) {
break;
}
u32 index;
if (flags.type == TextureType::Buffer) {
index = descriptors.Add(TextureBufferDescriptor{
.cbuf_index = cbuf.index,
.cbuf_offset = cbuf.offset,
.count = 1,
});
} else {
index = descriptors.Add(TextureDescriptor{
.type = flags.type,
.cbuf_index = cbuf.index,
.cbuf_offset = cbuf.offset,
.count = 1,
});
switch (inst->GetOpcode()) {
case IR::Opcode::ImageRead:
case IR::Opcode::ImageWrite: {
const bool is_written{inst->GetOpcode() == IR::Opcode::ImageWrite};
if (flags.type == TextureType::Buffer) {
throw NotImplementedException("Image buffer");
} else {
index = descriptors.Add(ImageDescriptor{
.type = flags.type,
.format = flags.image_format,
.is_written = is_written,
.cbuf_index = cbuf.index,
.cbuf_offset = cbuf.offset,
.count = 1,
});
}
break;
}
default:
if (flags.type == TextureType::Buffer) {
index = descriptors.Add(TextureBufferDescriptor{
.cbuf_index = cbuf.index,
.cbuf_offset = cbuf.offset,
.count = 1,
});
} else {
index = descriptors.Add(TextureDescriptor{
.type = flags.type,
.is_depth = flags.is_depth != 0,
.cbuf_index = cbuf.index,
.cbuf_offset = cbuf.offset,
.count = 1,
});
}
break;
}
inst->SetArg(0, IR::Value{index});
}

61
src/shader_recompiler/shader_info.h
View file

@ -22,15 +22,20 @@ enum class TextureType : u32 {
Color3D,
ColorCube,
ColorArrayCube,
Shadow1D,
ShadowArray1D,
Shadow2D,
ShadowArray2D,
Shadow3D,
ShadowCube,
ShadowArrayCube,
Buffer,
};
constexpr u32 NUM_TEXTURE_TYPES = 8;
enum class ImageFormat : u32 {
Typeless,
R8_UINT,
R8_SINT,
R16_UINT,
R16_SINT,
R32_UINT,
R32G32_UINT,
R32G32B32A32_UINT,
};
enum class Interpolation {
Smooth,
@ -43,21 +48,6 @@ struct InputVarying {
bool used{false};
};
struct TextureDescriptor {
TextureType type;
u32 cbuf_index;
u32 cbuf_offset;
u32 count;
};
using TextureDescriptors = boost::container::small_vector<TextureDescriptor, 12>;
struct TextureBufferDescriptor {
u32 cbuf_index;
u32 cbuf_offset;
u32 count;
};
using TextureBufferDescriptors = boost::container::small_vector<TextureBufferDescriptor, 2>;
struct ConstantBufferDescriptor {
u32 index;
u32 count;
@ -70,6 +60,32 @@ struct StorageBufferDescriptor {
bool is_written;
};
struct TextureBufferDescriptor {
u32 cbuf_index;
u32 cbuf_offset;
u32 count;
};
using TextureBufferDescriptors = boost::container::small_vector<TextureBufferDescriptor, 6>;
struct TextureDescriptor {
TextureType type;
bool is_depth;
u32 cbuf_index;
u32 cbuf_offset;
u32 count;
};
using TextureDescriptors = boost::container::small_vector<TextureDescriptor, 12>;
struct ImageDescriptor {
TextureType type;
ImageFormat format;
bool is_written;
u32 cbuf_index;
u32 cbuf_offset;
u32 count;
};
using ImageDescriptors = boost::container::small_vector<ImageDescriptor, 4>;
struct Info {
static constexpr size_t MAX_CBUFS{18};
static constexpr size_t MAX_SSBOS{16};
@ -121,6 +137,7 @@ struct Info {
boost::container::static_vector<StorageBufferDescriptor, MAX_SSBOS> storage_buffers_descriptors;
TextureBufferDescriptors texture_buffer_descriptors;
TextureDescriptors texture_descriptors;
ImageDescriptors image_descriptors;
};
} // namespace Shader

4
src/video_core/renderer_vulkan/blit_image.cpp
View file

@ -361,7 +361,7 @@ void BlitImageHelper::BlitColor(const Framebuffer* dst_framebuffer, const ImageV
.operation = operation,
};
const VkPipelineLayout layout = *one_texture_pipeline_layout;
const VkImageView src_view = src_image_view.Handle(ImageViewType::e2D);
const VkImageView src_view = src_image_view.Handle(Shader::TextureType::Color2D);
const VkSampler sampler = is_linear ? *linear_sampler : *nearest_sampler;
const VkPipeline pipeline = FindOrEmplacePipeline(key);
const VkDescriptorSet descriptor_set = one_texture_descriptor_allocator.Commit();
@ -435,7 +435,7 @@ void BlitImageHelper::ConvertR16ToD16(const Framebuffer* dst_framebuffer,
void BlitImageHelper::Convert(VkPipeline pipeline, const Framebuffer* dst_framebuffer,
const ImageView& src_image_view) {
const VkPipelineLayout layout = *one_texture_pipeline_layout;
const VkImageView src_view = src_image_view.Handle(ImageViewType::e2D);
const VkImageView src_view = src_image_view.Handle(Shader::TextureType::Color2D);
const VkSampler sampler = *nearest_sampler;
const VkDescriptorSet descriptor_set = one_texture_descriptor_allocator.Commit();
const VkExtent2D extent{

43
src/video_core/renderer_vulkan/pipeline_helper.h
View file

@ -97,6 +97,9 @@ public:
for ([[maybe_unused]] const auto& desc : info.texture_descriptors) {
Add(VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, stage);
}
for (const auto& desc : info.image_descriptors) {
Add(VK_DESCRIPTOR_TYPE_STORAGE_IMAGE, stage);
}
}
private:
@ -127,36 +130,6 @@ private:
size_t offset{};
};
inline VideoCommon::ImageViewType CastType(Shader::TextureType type) {
switch (type) {
case Shader::TextureType::Color1D:
case Shader::TextureType::Shadow1D:
return VideoCommon::ImageViewType::e1D;
case Shader::TextureType::ColorArray1D:
case Shader::TextureType::ShadowArray1D:
return VideoCommon::ImageViewType::e1DArray;
case Shader::TextureType::Color2D:
case Shader::TextureType::Shadow2D:
return VideoCommon::ImageViewType::e2D;
case Shader::TextureType::ColorArray2D:
case Shader::TextureType::ShadowArray2D:
return VideoCommon::ImageViewType::e2DArray;
case Shader::TextureType::Color3D:
case Shader::TextureType::Shadow3D:
return VideoCommon::ImageViewType::e3D;
case Shader::TextureType::ColorCube:
case Shader::TextureType::ShadowCube:
return VideoCommon::ImageViewType::Cube;
case Shader::TextureType::ColorArrayCube:
case Shader::TextureType::ShadowArrayCube:
return VideoCommon::ImageViewType::CubeArray;
case Shader::TextureType::Buffer:
break;
}
UNREACHABLE_MSG("Invalid texture type {}", type);
return {};
}
inline void PushImageDescriptors(const Shader::Info& info, const VkSampler*& samplers,
const ImageId*& image_view_ids, TextureCache& texture_cache,
VKUpdateDescriptorQueue& update_descriptor_queue) {
@ -164,9 +137,17 @@ inline void PushImageDescriptors(const Shader::Info& info, const VkSampler*& sam
for (const auto& desc : info.texture_descriptors) {
const VkSampler sampler{*(samplers++)};
ImageView& image_view{texture_cache.GetImageView(*(image_view_ids++))};
const VkImageView vk_image_view{image_view.Handle(CastType(desc.type))};
const VkImageView vk_image_view{image_view.Handle(desc.type)};
update_descriptor_queue.AddSampledImage(vk_image_view, sampler);
}
for (const auto& desc : info.image_descriptors) {
ImageView& image_view{texture_cache.GetImageView(*(image_view_ids++))};
if (desc.is_written) {
texture_cache.MarkModification(image_view.image_id);
}
const VkImageView vk_image_view{image_view.StorageView(desc.type, desc.format)};
update_descriptor_queue.AddImage(vk_image_view);
}
}
} // namespace Vulkan

4
src/video_core/renderer_vulkan/vk_compute_pipeline.cpp
View file

@ -108,6 +108,10 @@ void ComputePipeline::Configure(Tegra::Engines::KeplerCompute& kepler_compute,
Sampler* const sampler = texture_cache.GetComputeSampler(handle.sampler);
samplers.push_back(sampler->Handle());
}
for (const auto& desc : info.image_descriptors) {
const TextureHandle handle{read_handle(desc.cbuf_index, desc.cbuf_offset)};
image_view_indices.push_back(handle.image);
}
const std::span indices_span(image_view_indices.data(), image_view_indices.size());
texture_cache.FillComputeImageViews(indices_span, image_view_ids);

4
src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp
View file

@ -186,6 +186,10 @@ void GraphicsPipeline::Configure(bool is_indexed) {
Sampler* const sampler{texture_cache.GetGraphicsSampler(handle.sampler)};
samplers.push_back(sampler->Handle());
}
for (const auto& desc : info.image_descriptors) {
const TextureHandle handle{read_handle(desc.cbuf_index, desc.cbuf_offset)};
image_view_indices.push_back(handle.image);
}
}
const std::span indices_span(image_view_indices.data(), image_view_indices.size());
texture_cache.FillGraphicsImageViews(indices_span, image_view_ids);

2
src/video_core/renderer_vulkan/vk_rasterizer.cpp
View file

@ -494,7 +494,7 @@ bool RasterizerVulkan::AccelerateDisplay(const Tegra::FramebufferConfig& config,
if (!image_view) {
return false;
}
screen_info.image_view = image_view->Handle(VideoCommon::ImageViewType::e2D);
screen_info.image_view = image_view->Handle(Shader::TextureType::Color2D);
screen_info.width = image_view->size.width;
screen_info.height = image_view->size.height;
screen_info.is_srgb = VideoCore::Surface::IsPixelFormatSRGB(image_view->format);

112
src/video_core/renderer_vulkan/vk_texture_cache.cpp
View file

@ -215,6 +215,30 @@ constexpr VkBorderColor ConvertBorderColor(const std::array<float, 4>& color) {
return VK_COMPONENT_SWIZZLE_ZERO;
}
[[nodiscard]] VkImageViewType ImageViewType(Shader::TextureType type) {
switch (type) {
case Shader::TextureType::Color1D:
return VK_IMAGE_VIEW_TYPE_1D;
case Shader::TextureType::Color2D:
return VK_IMAGE_VIEW_TYPE_2D;
case Shader::TextureType::ColorCube:
return VK_IMAGE_VIEW_TYPE_CUBE;
case Shader::TextureType::Color3D:
return VK_IMAGE_VIEW_TYPE_3D;
case Shader::TextureType::ColorArray1D:
return VK_IMAGE_VIEW_TYPE_1D_ARRAY;
case Shader::TextureType::ColorArray2D:
return VK_IMAGE_VIEW_TYPE_2D_ARRAY;
case Shader::TextureType::ColorArrayCube:
return VK_IMAGE_VIEW_TYPE_CUBE_ARRAY;
case Shader::TextureType::Buffer:
UNREACHABLE_MSG("Texture buffers can't be image views");
return VK_IMAGE_VIEW_TYPE_1D;
}
UNREACHABLE_MSG("Invalid image view type={}", type);
return VK_IMAGE_VIEW_TYPE_2D;
}
[[nodiscard]] VkImageViewType ImageViewType(VideoCommon::ImageViewType type) {
switch (type) {
case VideoCommon::ImageViewType::e1D:
@ -232,7 +256,7 @@ constexpr VkBorderColor ConvertBorderColor(const std::array<float, 4>& color) {
case VideoCommon::ImageViewType::CubeArray:
return VK_IMAGE_VIEW_TYPE_CUBE_ARRAY;
case VideoCommon::ImageViewType::Rect:
LOG_WARNING(Render_Vulkan, "Unnormalized image view type not supported");
UNIMPLEMENTED_MSG("Rect image view");
return VK_IMAGE_VIEW_TYPE_2D;
case VideoCommon::ImageViewType::Buffer:
UNREACHABLE_MSG("Texture buffers can't be image views");
@ -539,6 +563,28 @@ struct RangedBarrierRange {
}
};
[[nodiscard]] VkFormat Format(Shader::ImageFormat format) {
switch (format) {
case Shader::ImageFormat::Typeless:
break;
case Shader::ImageFormat::R8_SINT:
return VK_FORMAT_R8_SINT;
case Shader::ImageFormat::R8_UINT:
return VK_FORMAT_R8_UINT;
case Shader::ImageFormat::R16_UINT:
return VK_FORMAT_R16_UINT;
case Shader::ImageFormat::R16_SINT:
return VK_FORMAT_R16_SINT;
case Shader::ImageFormat::R32_UINT:
return VK_FORMAT_R32_UINT;
case Shader::ImageFormat::R32G32_UINT:
return VK_FORMAT_R32G32_UINT;
case Shader::ImageFormat::R32G32B32A32_UINT:
return VK_FORMAT_R32G32B32A32_UINT;
}
UNREACHABLE_MSG("Invalid image format={}", format);
return VK_FORMAT_R32_UINT;
}
} // Anonymous namespace
void TextureCacheRuntime::Finish() {
@ -577,7 +623,7 @@ void TextureCacheRuntime::BlitImage(Framebuffer* dst_framebuffer, ImageView& dst
return;
}
}
ASSERT(src.ImageFormat() == dst.ImageFormat());
ASSERT(src.format == dst.format);
ASSERT(!(is_dst_msaa && !is_src_msaa));
ASSERT(operation == Fermi2D::Operation::SrcCopy);
@ -915,8 +961,9 @@ void Image::DownloadMemory(const StagingBufferRef& map, std::span<const BufferIm
ImageView::ImageView(TextureCacheRuntime& runtime, const VideoCommon::ImageViewInfo& info,
ImageId image_id_, Image& image)
: VideoCommon::ImageViewBase{info, image.info, image_id_}, device{&runtime.device},
image_handle{image.Handle()}, image_format{image.info.format}, samples{ConvertSampleCount(
image.info.num_samples)} {
image_handle{image.Handle()}, samples{ConvertSampleCount(image.info.num_samples)} {
using Shader::TextureType;
const VkImageAspectFlags aspect_mask = ImageViewAspectMask(info);
std::array<SwizzleSource, 4> swizzle{
SwizzleSource::R,
@ -954,39 +1001,39 @@ ImageView::ImageView(TextureCacheRuntime& runtime, const VideoCommon::ImageViewI
},
.subresourceRange = MakeSubresourceRange(aspect_mask, info.range),
};
const auto create = [&](VideoCommon::ImageViewType view_type, std::optional<u32> num_layers) {
const auto create = [&](TextureType tex_type, std::optional<u32> num_layers) {
VkImageViewCreateInfo ci{create_info};
ci.viewType = ImageViewType(view_type);
ci.viewType = ImageViewType(tex_type);
if (num_layers) {
ci.subresourceRange.layerCount = *num_layers;
}
vk::ImageView handle = device->GetLogical().CreateImageView(ci);
if (device->HasDebuggingToolAttached()) {
handle.SetObjectNameEXT(VideoCommon::Name(*this, view_type).c_str());
handle.SetObjectNameEXT(VideoCommon::Name(*this).c_str());
}
image_views[static_cast<size_t>(view_type)] = std::move(handle);
image_views[static_cast<size_t>(tex_type)] = std::move(handle);
};
switch (info.type) {
case VideoCommon::ImageViewType::e1D:
case VideoCommon::ImageViewType::e1DArray:
create(VideoCommon::ImageViewType::e1D, 1);
create(VideoCommon::ImageViewType::e1DArray, std::nullopt);
render_target = Handle(VideoCommon::ImageViewType::e1DArray);
create(TextureType::Color1D, 1);
create(TextureType::ColorArray1D, std::nullopt);
render_target = Handle(TextureType::ColorArray1D);
break;
case VideoCommon::ImageViewType::e2D:
case VideoCommon::ImageViewType::e2DArray:
create(VideoCommon::ImageViewType::e2D, 1);
create(VideoCommon::ImageViewType::e2DArray, std::nullopt);
render_target = Handle(VideoCommon::ImageViewType::e2DArray);
create(TextureType::Color2D, 1);
create(TextureType::ColorArray2D, std::nullopt);
render_target = Handle(Shader::TextureType::ColorArray2D);
break;
case VideoCommon::ImageViewType::e3D:
create(VideoCommon::ImageViewType::e3D, std::nullopt);
render_target = Handle(VideoCommon::ImageViewType::e3D);
create(TextureType::Color3D, std::nullopt);
render_target = Handle(Shader::TextureType::Color3D);
break;
case VideoCommon::ImageViewType::Cube:
case VideoCommon::ImageViewType::CubeArray:
create(VideoCommon::ImageViewType::Cube, 6);
create(VideoCommon::ImageViewType::CubeArray, std::nullopt);
create(TextureType::ColorCube, 6);
create(TextureType::ColorArrayCube, std::nullopt);
break;
case VideoCommon::ImageViewType::Rect:
UNIMPLEMENTED();
@ -1009,7 +1056,8 @@ VkImageView ImageView::DepthView() {
if (depth_view) {
return *depth_view;
}
depth_view = MakeDepthStencilView(VK_IMAGE_ASPECT_DEPTH_BIT);
const auto& info = MaxwellToVK::SurfaceFormat(*device, FormatType::Optimal, true, format);
depth_view = MakeView(info.format, VK_IMAGE_ASPECT_DEPTH_BIT);
return *depth_view;
}
@ -1017,18 +1065,38 @@ VkImageView ImageView::StencilView() {
if (stencil_view) {
return *stencil_view;
}
stencil_view = MakeDepthStencilView(VK_IMAGE_ASPECT_STENCIL_BIT);
const auto& info = MaxwellToVK::SurfaceFormat(*device, FormatType::Optimal, true, format);
stencil_view = MakeView(info.format, VK_IMAGE_ASPECT_STENCIL_BIT);
return *stencil_view;
}
vk::ImageView ImageView::MakeDepthStencilView(VkImageAspectFlags aspect_mask) {
VkImageView ImageView::StorageView(Shader::TextureType texture_type,
Shader::ImageFormat image_format) {
if (image_format == Shader::ImageFormat::Typeless) {
return Handle(texture_type);
}
const bool is_signed{image_format == Shader::ImageFormat::R8_SINT ||
image_format == Shader::ImageFormat::R16_SINT};
if (!storage_views) {
storage_views = std::make_unique<StorageViews>();
}
auto& views{is_signed ? storage_views->signeds : storage_views->unsigneds};
auto& view{views[static_cast<size_t>(texture_type)]};
if (view) {
return *view;
}
view = MakeView(Format(image_format), VK_IMAGE_ASPECT_COLOR_BIT);
return *view;
}
vk::ImageView ImageView::MakeView(VkFormat vk_format, VkImageAspectFlags aspect_mask) {
return device->GetLogical().CreateImageView({
.sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
.pNext = nullptr,
.flags = 0,
.image = image_handle,
.viewType = ImageViewType(type),
.format = MaxwellToVK::SurfaceFormat(*device, FormatType::Optimal, true, format).format,
.format = vk_format,
.components{
.r = VK_COMPONENT_SWIZZLE_IDENTITY,
.g = VK_COMPONENT_SWIZZLE_IDENTITY,

23
src/video_core/renderer_vulkan/vk_texture_cache.h
View file

@ -7,6 +7,7 @@
#include <compare>
#include <span>
#include "shader_recompiler/shader_info.h"
#include "video_core/renderer_vulkan/vk_staging_buffer_pool.h"
#include "video_core/texture_cache/texture_cache.h"
#include "video_core/vulkan_common/vulkan_memory_allocator.h"
@ -133,8 +134,11 @@ public:
[[nodiscard]] VkImageView StencilView();
[[nodiscard]] VkImageView Handle(VideoCommon::ImageViewType query_type) const noexcept {
return *image_views[static_cast<size_t>(query_type)];
[[nodiscard]] VkImageView StorageView(Shader::TextureType texture_type,
Shader::ImageFormat image_format);
[[nodiscard]] VkImageView Handle(Shader::TextureType texture_type) const noexcept {
return *image_views[static_cast<size_t>(texture_type)];
}
[[nodiscard]] VkImage ImageHandle() const noexcept {
@ -145,10 +149,6 @@ public:
return render_target;
}
[[nodiscard]] PixelFormat ImageFormat() const noexcept {
return image_format;
}
[[nodiscard]] VkSampleCountFlagBits Samples() const noexcept {
return samples;
}
@ -162,15 +162,20 @@ public:
}
private:
[[nodiscard]] vk::ImageView MakeDepthStencilView(VkImageAspectFlags aspect_mask);
struct StorageViews {
std::array<vk::ImageView, Shader::NUM_TEXTURE_TYPES> signeds;
std::array<vk::ImageView, Shader::NUM_TEXTURE_TYPES> unsigneds;
};
[[nodiscard]] vk::ImageView MakeView(VkFormat vk_format, VkImageAspectFlags aspect_mask);
const Device* device = nullptr;
std::array<vk::ImageView, VideoCommon::NUM_IMAGE_VIEW_TYPES> image_views;
std::array<vk::ImageView, Shader::NUM_TEXTURE_TYPES> image_views;
std::unique_ptr<StorageViews> storage_views;
vk::ImageView depth_view;
vk::ImageView stencil_view;
VkImage image_handle = VK_NULL_HANDLE;
VkImageView render_target = VK_NULL_HANDLE;
PixelFormat image_format = PixelFormat::Invalid;
VkSampleCountFlagBits samples = VK_SAMPLE_COUNT_1_BIT;
GPUVAddr gpu_addr = 0;
u32 buffer_size = 0;

8
src/video_core/texture_cache/texture_cache.h
View file

@ -117,6 +117,9 @@ public:
/// Return a reference to the given image view id
[[nodiscard]] ImageView& GetImageView(ImageViewId id) noexcept;
/// Mark an image as modified from the GPU
void MarkModification(ImageId id) noexcept;
/// Fill image_view_ids with the graphics images in indices
void FillGraphicsImageViews(std::span<const u32> indices,
std::span<ImageViewId> image_view_ids);
@ -526,6 +529,11 @@ typename P::ImageView& TextureCache<P>::GetImageView(ImageViewId id) noexcept {
return slot_image_views[id];
}
template <class P>
void TextureCache<P>::MarkModification(ImageId id) noexcept {
MarkModification(slot_images[id]);
}
template <class P>
void TextureCache<P>::FillGraphicsImageViews(std::span<const u32> indices,
std::span<ImageViewId> image_view_ids) {