diff options
author | bunnei <bunneidev@gmail.com> | 2021-11-17 03:52:11 +0100 |
---|---|---|
committer | GitHub <noreply@github.com> | 2021-11-17 03:52:11 +0100 |
commit | 71313509f75aeafe425e531824d1faa9e7c0a40b (patch) | |
tree | cb1df371d288677fcede6a3409eb079e0d278163 /src/shader_recompiler | |
parent | Merge pull request #7347 from lioncash/catch (diff) | |
parent | TextureCache: Fix Automatic Anisotropic. (diff) | |
download | yuzu-71313509f75aeafe425e531824d1faa9e7c0a40b.tar yuzu-71313509f75aeafe425e531824d1faa9e7c0a40b.tar.gz yuzu-71313509f75aeafe425e531824d1faa9e7c0a40b.tar.bz2 yuzu-71313509f75aeafe425e531824d1faa9e7c0a40b.tar.lz yuzu-71313509f75aeafe425e531824d1faa9e7c0a40b.tar.xz yuzu-71313509f75aeafe425e531824d1faa9e7c0a40b.tar.zst yuzu-71313509f75aeafe425e531824d1faa9e7c0a40b.zip |
Diffstat (limited to '')
33 files changed, 665 insertions, 10 deletions
diff --git a/src/shader_recompiler/CMakeLists.txt b/src/shader_recompiler/CMakeLists.txt index b5b7e5e83..bc3df80c8 100644 --- a/src/shader_recompiler/CMakeLists.txt +++ b/src/shader_recompiler/CMakeLists.txt @@ -221,6 +221,7 @@ add_library(shader_recompiler STATIC ir_opt/lower_fp16_to_fp32.cpp ir_opt/lower_int64_to_int32.cpp ir_opt/passes.h + ir_opt/rescaling_pass.cpp ir_opt/ssa_rewrite_pass.cpp ir_opt/texture_pass.cpp ir_opt/verification_pass.cpp diff --git a/src/shader_recompiler/backend/bindings.h b/src/shader_recompiler/backend/bindings.h index 35503000c..669702553 100644 --- a/src/shader_recompiler/backend/bindings.h +++ b/src/shader_recompiler/backend/bindings.h @@ -14,6 +14,8 @@ struct Bindings { u32 storage_buffer{}; u32 texture{}; u32 image{}; + u32 texture_scaling_index{}; + u32 image_scaling_index{}; }; } // namespace Shader::Backend diff --git a/src/shader_recompiler/backend/glasm/emit_context.cpp b/src/shader_recompiler/backend/glasm/emit_context.cpp index 069c019ad..8fd459dfe 100644 --- a/src/shader_recompiler/backend/glasm/emit_context.cpp +++ b/src/shader_recompiler/backend/glasm/emit_context.cpp @@ -6,6 +6,7 @@ #include "shader_recompiler/backend/bindings.h" #include "shader_recompiler/backend/glasm/emit_context.h" +#include "shader_recompiler/backend/glasm/emit_glasm.h" #include "shader_recompiler/frontend/ir/program.h" #include "shader_recompiler/profile.h" #include "shader_recompiler/runtime_info.h" @@ -55,7 +56,8 @@ EmitContext::EmitContext(IR::Program& program, Bindings& bindings, const Profile } if (!runtime_info.glasm_use_storage_buffers) { if (const size_t num = info.storage_buffers_descriptors.size(); num > 0) { - Add("PARAM c[{}]={{program.local[0..{}]}};", num, num - 1); + const size_t index{num + PROGRAM_LOCAL_PARAMETER_STORAGE_BUFFER_BASE}; + Add("PARAM c[{}]={{program.local[0..{}]}};", index, index - 1); } } stage = program.stage; diff --git a/src/shader_recompiler/backend/glasm/emit_glasm.cpp b/src/shader_recompiler/backend/glasm/emit_glasm.cpp index 4ce1c4f54..004658546 100644 --- a/src/shader_recompiler/backend/glasm/emit_glasm.cpp +++ b/src/shader_recompiler/backend/glasm/emit_glasm.cpp @@ -448,6 +448,9 @@ std::string EmitGLASM(const Profile& profile, const RuntimeInfo& runtime_info, I header += fmt::format("SHARED_MEMORY {};", program.shared_memory_size); header += fmt::format("SHARED shared_mem[]={{program.sharedmem}};"); } + if (program.info.uses_rescaling_uniform) { + header += "PARAM scaling[1]={program.local[0..0]};"; + } header += "TEMP "; for (size_t index = 0; index < ctx.reg_alloc.NumUsedRegisters(); ++index) { header += fmt::format("R{},", index); diff --git a/src/shader_recompiler/backend/glasm/emit_glasm.h b/src/shader_recompiler/backend/glasm/emit_glasm.h index bcb55f062..292655acb 100644 --- a/src/shader_recompiler/backend/glasm/emit_glasm.h +++ b/src/shader_recompiler/backend/glasm/emit_glasm.h @@ -13,6 +13,8 @@ namespace Shader::Backend::GLASM { +constexpr u32 PROGRAM_LOCAL_PARAMETER_STORAGE_BUFFER_BASE = 1; + [[nodiscard]] std::string EmitGLASM(const Profile& profile, const RuntimeInfo& runtime_info, IR::Program& program, Bindings& bindings); diff --git a/src/shader_recompiler/backend/glasm/emit_glasm_image.cpp b/src/shader_recompiler/backend/glasm/emit_glasm_image.cpp index 09e3a9b82..d325d31c7 100644 --- a/src/shader_recompiler/backend/glasm/emit_glasm_image.cpp +++ b/src/shader_recompiler/backend/glasm/emit_glasm_image.cpp @@ -608,6 +608,24 @@ void EmitImageWrite(EmitContext& ctx, IR::Inst& inst, const IR::Value& index, Re ctx.Add("STOREIM.{} {},{},{},{};", format, image, color, coord, type); } +void EmitIsTextureScaled(EmitContext& ctx, IR::Inst& inst, const IR::Value& index) { + if (!index.IsImmediate()) { + throw NotImplementedException("Non-constant texture rescaling"); + } + ctx.Add("AND.U RC.x,scaling[0].x,{};" + "SNE.S {},RC.x,0;", + 1u << index.U32(), ctx.reg_alloc.Define(inst)); +} + +void EmitIsImageScaled(EmitContext& ctx, IR::Inst& inst, const IR::Value& index) { + if (!index.IsImmediate()) { + throw NotImplementedException("Non-constant texture rescaling"); + } + ctx.Add("AND.U RC.x,scaling[0].y,{};" + "SNE.S {},RC.x,0;", + 1u << index.U32(), ctx.reg_alloc.Define(inst)); +} + void EmitImageAtomicIAdd32(EmitContext& ctx, IR::Inst& inst, const IR::Value& index, Register coord, ScalarU32 value) { ImageAtomic(ctx, inst, index, coord, value, "ADD.U32"); diff --git a/src/shader_recompiler/backend/glasm/emit_glasm_instructions.h b/src/shader_recompiler/backend/glasm/emit_glasm_instructions.h index 12afda43b..1f343bff5 100644 --- a/src/shader_recompiler/backend/glasm/emit_glasm_instructions.h +++ b/src/shader_recompiler/backend/glasm/emit_glasm_instructions.h @@ -72,6 +72,7 @@ void EmitInvocationId(EmitContext& ctx, IR::Inst& inst); void EmitSampleId(EmitContext& ctx, IR::Inst& inst); void EmitIsHelperInvocation(EmitContext& ctx, IR::Inst& inst); void EmitYDirection(EmitContext& ctx, IR::Inst& inst); +void EmitResolutionDownFactor(EmitContext& ctx, IR::Inst& inst); void EmitLoadLocal(EmitContext& ctx, IR::Inst& inst, ScalarU32 word_offset); void EmitWriteLocal(EmitContext& ctx, ScalarU32 word_offset, ScalarU32 value); void EmitUndefU1(EmitContext& ctx, IR::Inst& inst); @@ -303,6 +304,8 @@ void EmitIAdd64(EmitContext& ctx, IR::Inst& inst, Register a, Register b); void EmitISub32(EmitContext& ctx, IR::Inst& inst, ScalarS32 a, ScalarS32 b); void EmitISub64(EmitContext& ctx, IR::Inst& inst, Register a, Register b); void EmitIMul32(EmitContext& ctx, IR::Inst& inst, ScalarS32 a, ScalarS32 b); +void EmitSDiv32(EmitContext& ctx, IR::Inst& inst, ScalarS32 a, ScalarS32 b); +void EmitUDiv32(EmitContext& ctx, IR::Inst& inst, ScalarU32 a, ScalarU32 b); void EmitINeg32(EmitContext& ctx, IR::Inst& inst, ScalarS32 value); void EmitINeg64(EmitContext& ctx, IR::Inst& inst, Register value); void EmitIAbs32(EmitContext& ctx, IR::Inst& inst, ScalarS32 value); @@ -553,6 +556,8 @@ void EmitImageGradient(EmitContext& ctx, IR::Inst& inst, const IR::Value& index, void EmitImageRead(EmitContext& ctx, IR::Inst& inst, const IR::Value& index, Register coord); void EmitImageWrite(EmitContext& ctx, IR::Inst& inst, const IR::Value& index, Register coord, Register color); +void EmitIsTextureScaled(EmitContext& ctx, IR::Inst& inst, const IR::Value& index); +void EmitIsImageScaled(EmitContext& ctx, IR::Inst& inst, const IR::Value& index); void EmitBindlessImageAtomicIAdd32(EmitContext&); void EmitBindlessImageAtomicSMin32(EmitContext&); void EmitBindlessImageAtomicUMin32(EmitContext&); diff --git a/src/shader_recompiler/backend/glasm/emit_glasm_integer.cpp b/src/shader_recompiler/backend/glasm/emit_glasm_integer.cpp index f55c26b76..8aa494a4d 100644 --- a/src/shader_recompiler/backend/glasm/emit_glasm_integer.cpp +++ b/src/shader_recompiler/backend/glasm/emit_glasm_integer.cpp @@ -90,6 +90,14 @@ void EmitIMul32(EmitContext& ctx, IR::Inst& inst, ScalarS32 a, ScalarS32 b) { ctx.Add("MUL.S {}.x,{},{};", inst, a, b); } +void EmitSDiv32(EmitContext& ctx, IR::Inst& inst, ScalarS32 a, ScalarS32 b) { + ctx.Add("DIV.S {}.x,{},{};", inst, a, b); +} + +void EmitUDiv32(EmitContext& ctx, IR::Inst& inst, ScalarU32 a, ScalarU32 b) { + ctx.Add("DIV.U {}.x,{},{};", inst, a, b); +} + void EmitINeg32(EmitContext& ctx, IR::Inst& inst, ScalarS32 value) { if (value.type != Type::Register && static_cast<s32>(value.imm_u32) < 0) { ctx.Add("MOV.S {},{};", inst, -static_cast<s32>(value.imm_u32)); diff --git a/src/shader_recompiler/backend/glasm/emit_glasm_not_implemented.cpp b/src/shader_recompiler/backend/glasm/emit_glasm_not_implemented.cpp index e537f6073..681aeda8d 100644 --- a/src/shader_recompiler/backend/glasm/emit_glasm_not_implemented.cpp +++ b/src/shader_recompiler/backend/glasm/emit_glasm_not_implemented.cpp @@ -210,6 +210,10 @@ void EmitYDirection(EmitContext& ctx, IR::Inst& inst) { ctx.Add("MOV.F {}.x,y_direction[0].w;", inst); } +void EmitResolutionDownFactor(EmitContext& ctx, IR::Inst& inst) { + ctx.Add("MOV.F {}.x,scaling[0].z;", inst); +} + void EmitUndefU1(EmitContext& ctx, IR::Inst& inst) { ctx.Add("MOV.S {}.x,0;", inst); } diff --git a/src/shader_recompiler/backend/glsl/emit_context.cpp b/src/shader_recompiler/backend/glsl/emit_context.cpp index 4e6f2c0fe..97bd59302 100644 --- a/src/shader_recompiler/backend/glsl/emit_context.cpp +++ b/src/shader_recompiler/backend/glsl/emit_context.cpp @@ -393,6 +393,9 @@ EmitContext::EmitContext(IR::Program& program, Bindings& bindings, const Profile DefineGenericOutput(index, program.invocations); } } + if (info.uses_rescaling_uniform) { + header += "layout(location=0) uniform vec4 scaling;"; + } DefineConstantBuffers(bindings); DefineStorageBuffers(bindings); SetupImages(bindings); diff --git a/src/shader_recompiler/backend/glsl/emit_glsl_context_get_set.cpp b/src/shader_recompiler/backend/glsl/emit_glsl_context_get_set.cpp index 170db269a..4c26f3829 100644 --- a/src/shader_recompiler/backend/glsl/emit_glsl_context_get_set.cpp +++ b/src/shader_recompiler/backend/glsl/emit_glsl_context_get_set.cpp @@ -445,6 +445,10 @@ void EmitYDirection(EmitContext& ctx, IR::Inst& inst) { ctx.AddF32("{}=gl_FrontMaterial.ambient.a;", inst); } +void EmitResolutionDownFactor(EmitContext& ctx, IR::Inst& inst) { + ctx.AddF32("{}=scaling.z;", inst); +} + void EmitLoadLocal(EmitContext& ctx, IR::Inst& inst, std::string_view word_offset) { ctx.AddU32("{}=lmem[{}];", inst, word_offset); } diff --git a/src/shader_recompiler/backend/glsl/emit_glsl_image.cpp b/src/shader_recompiler/backend/glsl/emit_glsl_image.cpp index 447eb8e0a..2f78d0267 100644 --- a/src/shader_recompiler/backend/glsl/emit_glsl_image.cpp +++ b/src/shader_recompiler/backend/glsl/emit_glsl_image.cpp @@ -612,6 +612,22 @@ void EmitImageAtomicExchange32(EmitContext& ctx, IR::Inst& inst, const IR::Value value); } +void EmitIsTextureScaled(EmitContext& ctx, IR::Inst& inst, const IR::Value& index) { + if (!index.IsImmediate()) { + throw NotImplementedException("Non-constant texture rescaling"); + } + const u32 image_index{index.U32()}; + ctx.AddU1("{}=(ftou(scaling.x)&{})!=0;", inst, 1u << image_index); +} + +void EmitIsImageScaled(EmitContext& ctx, IR::Inst& inst, const IR::Value& index) { + if (!index.IsImmediate()) { + throw NotImplementedException("Non-constant texture rescaling"); + } + const u32 image_index{index.U32()}; + ctx.AddU1("{}=(ftou(scaling.y)&{})!=0;", inst, 1u << image_index); +} + void EmitBindlessImageSampleImplicitLod(EmitContext&) { NotImplemented(); } diff --git a/src/shader_recompiler/backend/glsl/emit_glsl_instructions.h b/src/shader_recompiler/backend/glsl/emit_glsl_instructions.h index 5936d086f..f86502e4c 100644 --- a/src/shader_recompiler/backend/glsl/emit_glsl_instructions.h +++ b/src/shader_recompiler/backend/glsl/emit_glsl_instructions.h @@ -85,6 +85,7 @@ void EmitInvocationId(EmitContext& ctx, IR::Inst& inst); void EmitSampleId(EmitContext& ctx, IR::Inst& inst); void EmitIsHelperInvocation(EmitContext& ctx, IR::Inst& inst); void EmitYDirection(EmitContext& ctx, IR::Inst& inst); +void EmitResolutionDownFactor(EmitContext& ctx, IR::Inst& inst); void EmitLoadLocal(EmitContext& ctx, IR::Inst& inst, std::string_view word_offset); void EmitWriteLocal(EmitContext& ctx, std::string_view word_offset, std::string_view value); void EmitUndefU1(EmitContext& ctx, IR::Inst& inst); @@ -362,6 +363,8 @@ void EmitIAdd64(EmitContext& ctx, IR::Inst& inst, std::string_view a, std::strin void EmitISub32(EmitContext& ctx, IR::Inst& inst, std::string_view a, std::string_view b); void EmitISub64(EmitContext& ctx, IR::Inst& inst, std::string_view a, std::string_view b); void EmitIMul32(EmitContext& ctx, IR::Inst& inst, std::string_view a, std::string_view b); +void EmitSDiv32(EmitContext& ctx, IR::Inst& inst, std::string_view a, std::string_view b); +void EmitUDiv32(EmitContext& ctx, IR::Inst& inst, std::string_view a, std::string_view b); void EmitINeg32(EmitContext& ctx, IR::Inst& inst, std::string_view value); void EmitINeg64(EmitContext& ctx, IR::Inst& inst, std::string_view value); void EmitIAbs32(EmitContext& ctx, IR::Inst& inst, std::string_view value); @@ -627,6 +630,8 @@ void EmitImageRead(EmitContext& ctx, IR::Inst& inst, const IR::Value& index, std::string_view coords); void EmitImageWrite(EmitContext& ctx, IR::Inst& inst, const IR::Value& index, std::string_view coords, std::string_view color); +void EmitIsTextureScaled(EmitContext& ctx, IR::Inst& inst, const IR::Value& index); +void EmitIsImageScaled(EmitContext& ctx, IR::Inst& inst, const IR::Value& index); void EmitBindlessImageAtomicIAdd32(EmitContext&); void EmitBindlessImageAtomicSMin32(EmitContext&); void EmitBindlessImageAtomicUMin32(EmitContext&); diff --git a/src/shader_recompiler/backend/glsl/emit_glsl_integer.cpp b/src/shader_recompiler/backend/glsl/emit_glsl_integer.cpp index 38419f88f..88c1d4c5e 100644 --- a/src/shader_recompiler/backend/glsl/emit_glsl_integer.cpp +++ b/src/shader_recompiler/backend/glsl/emit_glsl_integer.cpp @@ -78,6 +78,14 @@ void EmitIMul32(EmitContext& ctx, IR::Inst& inst, std::string_view a, std::strin ctx.AddU32("{}=uint({}*{});", inst, a, b); } +void EmitSDiv32(EmitContext& ctx, IR::Inst& inst, std::string_view a, std::string_view b) { + ctx.AddU32("{}=uint(int({})/int({}));", inst, a, b); +} + +void EmitUDiv32(EmitContext& ctx, IR::Inst& inst, std::string_view a, std::string_view b) { + ctx.AddU32("{}={}/{};", inst, a, b); +} + void EmitINeg32(EmitContext& ctx, IR::Inst& inst, std::string_view value) { ctx.AddU32("{}=uint(-({}));", inst, value); } diff --git a/src/shader_recompiler/backend/spirv/emit_context.cpp b/src/shader_recompiler/backend/spirv/emit_context.cpp index 3c84e6466..723455462 100644 --- a/src/shader_recompiler/backend/spirv/emit_context.cpp +++ b/src/shader_recompiler/backend/spirv/emit_context.cpp @@ -7,11 +7,14 @@ #include <climits> #include <string_view> +#include <boost/container/static_vector.hpp> + #include <fmt/format.h> #include "common/common_types.h" #include "common/div_ceil.h" #include "shader_recompiler/backend/spirv/emit_context.h" +#include "shader_recompiler/backend/spirv/emit_spirv.h" namespace Shader::Backend::SPIRV { namespace { @@ -474,8 +477,9 @@ void VectorTypes::Define(Sirit::Module& sirit_ctx, Id base_type, std::string_vie EmitContext::EmitContext(const Profile& profile_, const RuntimeInfo& runtime_info_, IR::Program& program, Bindings& bindings) - : Sirit::Module(profile_.supported_spirv), profile{profile_}, - runtime_info{runtime_info_}, stage{program.stage} { + : Sirit::Module(profile_.supported_spirv), profile{profile_}, runtime_info{runtime_info_}, + stage{program.stage}, texture_rescaling_index{bindings.texture_scaling_index}, + image_rescaling_index{bindings.image_scaling_index} { const bool is_unified{profile.unified_descriptor_binding}; u32& uniform_binding{is_unified ? bindings.unified : bindings.uniform_buffer}; u32& storage_binding{is_unified ? bindings.unified : bindings.storage_buffer}; @@ -492,10 +496,11 @@ EmitContext::EmitContext(const Profile& profile_, const RuntimeInfo& runtime_inf DefineStorageBuffers(program.info, storage_binding); DefineTextureBuffers(program.info, texture_binding); DefineImageBuffers(program.info, image_binding); - DefineTextures(program.info, texture_binding); - DefineImages(program.info, image_binding); + DefineTextures(program.info, texture_binding, bindings.texture_scaling_index); + DefineImages(program.info, image_binding, bindings.image_scaling_index); DefineAttributeMemAccess(program.info); DefineGlobalMemoryFunctions(program.info); + DefineRescalingInput(program.info); } EmitContext::~EmitContext() = default; @@ -996,6 +1001,73 @@ void EmitContext::DefineGlobalMemoryFunctions(const Info& info) { define(&StorageDefinitions::U32x4, storage_types.U32x4, U32[4], sizeof(u32[4])); } +void EmitContext::DefineRescalingInput(const Info& info) { + if (!info.uses_rescaling_uniform) { + return; + } + if (profile.unified_descriptor_binding) { + DefineRescalingInputPushConstant(); + } else { + DefineRescalingInputUniformConstant(); + } +} + +void EmitContext::DefineRescalingInputPushConstant() { + boost::container::static_vector<Id, 3> members{}; + u32 member_index{0}; + + rescaling_textures_type = TypeArray(U32[1], Const(4u)); + Decorate(rescaling_textures_type, spv::Decoration::ArrayStride, 4u); + members.push_back(rescaling_textures_type); + rescaling_textures_member_index = member_index++; + + rescaling_images_type = TypeArray(U32[1], Const(NUM_IMAGE_SCALING_WORDS)); + Decorate(rescaling_images_type, spv::Decoration::ArrayStride, 4u); + members.push_back(rescaling_images_type); + rescaling_images_member_index = member_index++; + + if (stage != Stage::Compute) { + members.push_back(F32[1]); + rescaling_downfactor_member_index = member_index++; + } + const Id push_constant_struct{TypeStruct(std::span(members.data(), members.size()))}; + Decorate(push_constant_struct, spv::Decoration::Block); + Name(push_constant_struct, "ResolutionInfo"); + + MemberDecorate(push_constant_struct, rescaling_textures_member_index, spv::Decoration::Offset, + static_cast<u32>(offsetof(RescalingLayout, rescaling_textures))); + MemberName(push_constant_struct, rescaling_textures_member_index, "rescaling_textures"); + + MemberDecorate(push_constant_struct, rescaling_images_member_index, spv::Decoration::Offset, + static_cast<u32>(offsetof(RescalingLayout, rescaling_images))); + MemberName(push_constant_struct, rescaling_images_member_index, "rescaling_images"); + + if (stage != Stage::Compute) { + MemberDecorate(push_constant_struct, rescaling_downfactor_member_index, + spv::Decoration::Offset, + static_cast<u32>(offsetof(RescalingLayout, down_factor))); + MemberName(push_constant_struct, rescaling_downfactor_member_index, "down_factor"); + } + const Id pointer_type{TypePointer(spv::StorageClass::PushConstant, push_constant_struct)}; + rescaling_push_constants = AddGlobalVariable(pointer_type, spv::StorageClass::PushConstant); + Name(rescaling_push_constants, "rescaling_push_constants"); + + if (profile.supported_spirv >= 0x00010400) { + interfaces.push_back(rescaling_push_constants); + } +} + +void EmitContext::DefineRescalingInputUniformConstant() { + const Id pointer_type{TypePointer(spv::StorageClass::UniformConstant, F32[4])}; + rescaling_uniform_constant = + AddGlobalVariable(pointer_type, spv::StorageClass::UniformConstant); + Decorate(rescaling_uniform_constant, spv::Decoration::Location, 0u); + + if (profile.supported_spirv >= 0x00010400) { + interfaces.push_back(rescaling_uniform_constant); + } +} + void EmitContext::DefineConstantBuffers(const Info& info, u32& binding) { if (info.constant_buffer_descriptors.empty()) { return; @@ -1184,7 +1256,7 @@ void EmitContext::DefineImageBuffers(const Info& info, u32& binding) { } } -void EmitContext::DefineTextures(const Info& info, u32& binding) { +void EmitContext::DefineTextures(const Info& info, u32& binding, u32& scaling_index) { textures.reserve(info.texture_descriptors.size()); for (const TextureDescriptor& desc : info.texture_descriptors) { const Id image_type{ImageType(*this, desc)}; @@ -1206,13 +1278,14 @@ void EmitContext::DefineTextures(const Info& info, u32& binding) { interfaces.push_back(id); } ++binding; + ++scaling_index; } if (info.uses_atomic_image_u32) { image_u32 = TypePointer(spv::StorageClass::Image, U32[1]); } } -void EmitContext::DefineImages(const Info& info, u32& binding) { +void EmitContext::DefineImages(const Info& info, u32& binding, u32& scaling_index) { images.reserve(info.image_descriptors.size()); for (const ImageDescriptor& desc : info.image_descriptors) { if (desc.count != 1) { @@ -1233,6 +1306,7 @@ void EmitContext::DefineImages(const Info& info, u32& binding) { interfaces.push_back(id); } ++binding; + ++scaling_index; } } diff --git a/src/shader_recompiler/backend/spirv/emit_context.h b/src/shader_recompiler/backend/spirv/emit_context.h index 112c52382..63f8185d9 100644 --- a/src/shader_recompiler/backend/spirv/emit_context.h +++ b/src/shader_recompiler/backend/spirv/emit_context.h @@ -238,6 +238,16 @@ public: Id indexed_load_func{}; Id indexed_store_func{}; + Id rescaling_uniform_constant{}; + Id rescaling_push_constants{}; + Id rescaling_textures_type{}; + Id rescaling_images_type{}; + u32 rescaling_textures_member_index{}; + u32 rescaling_images_member_index{}; + u32 rescaling_downfactor_member_index{}; + u32 texture_rescaling_index{}; + u32 image_rescaling_index{}; + Id local_memory{}; Id shared_memory_u8{}; @@ -310,10 +320,13 @@ private: void DefineStorageBuffers(const Info& info, u32& binding); void DefineTextureBuffers(const Info& info, u32& binding); void DefineImageBuffers(const Info& info, u32& binding); - void DefineTextures(const Info& info, u32& binding); - void DefineImages(const Info& info, u32& binding); + void DefineTextures(const Info& info, u32& binding, u32& scaling_index); + void DefineImages(const Info& info, u32& binding, u32& scaling_index); void DefineAttributeMemAccess(const Info& info); void DefineGlobalMemoryFunctions(const Info& info); + void DefineRescalingInput(const Info& info); + void DefineRescalingInputPushConstant(); + void DefineRescalingInputUniformConstant(); void DefineInputs(const IR::Program& program); void DefineOutputs(const IR::Program& program); diff --git a/src/shader_recompiler/backend/spirv/emit_spirv.h b/src/shader_recompiler/backend/spirv/emit_spirv.h index db0c935fe..4b25534ce 100644 --- a/src/shader_recompiler/backend/spirv/emit_spirv.h +++ b/src/shader_recompiler/backend/spirv/emit_spirv.h @@ -16,6 +16,19 @@ namespace Shader::Backend::SPIRV { +constexpr u32 NUM_TEXTURE_SCALING_WORDS = 4; +constexpr u32 NUM_IMAGE_SCALING_WORDS = 2; +constexpr u32 NUM_TEXTURE_AND_IMAGE_SCALING_WORDS = + NUM_TEXTURE_SCALING_WORDS + NUM_IMAGE_SCALING_WORDS; + +struct RescalingLayout { + alignas(16) std::array<u32, NUM_TEXTURE_SCALING_WORDS> rescaling_textures; + alignas(16) std::array<u32, NUM_IMAGE_SCALING_WORDS> rescaling_images; + alignas(16) u32 down_factor; +}; +constexpr u32 RESCALING_LAYOUT_WORDS_OFFSET = offsetof(RescalingLayout, rescaling_textures); +constexpr u32 RESCALING_LAYOUT_DOWN_FACTOR_OFFSET = offsetof(RescalingLayout, down_factor); + [[nodiscard]] std::vector<u32> EmitSPIRV(const Profile& profile, const RuntimeInfo& runtime_info, IR::Program& program, Bindings& bindings); diff --git a/src/shader_recompiler/backend/spirv/emit_spirv_context_get_set.cpp b/src/shader_recompiler/backend/spirv/emit_spirv_context_get_set.cpp index d3a93d5f4..bac683ae1 100644 --- a/src/shader_recompiler/backend/spirv/emit_spirv_context_get_set.cpp +++ b/src/shader_recompiler/backend/spirv/emit_spirv_context_get_set.cpp @@ -526,6 +526,18 @@ Id EmitYDirection(EmitContext& ctx) { return ctx.Const(ctx.runtime_info.y_negate ? -1.0f : 1.0f); } +Id EmitResolutionDownFactor(EmitContext& ctx) { + if (ctx.profile.unified_descriptor_binding) { + const Id pointer_type{ctx.TypePointer(spv::StorageClass::PushConstant, ctx.F32[1])}; + const Id index{ctx.Const(ctx.rescaling_downfactor_member_index)}; + const Id pointer{ctx.OpAccessChain(pointer_type, ctx.rescaling_push_constants, index)}; + return ctx.OpLoad(ctx.F32[1], pointer); + } else { + const Id composite{ctx.OpLoad(ctx.F32[4], ctx.rescaling_uniform_constant)}; + return ctx.OpCompositeExtract(ctx.F32[1], composite, 2u); + } +} + Id EmitLoadLocal(EmitContext& ctx, Id word_offset) { const Id pointer{ctx.OpAccessChain(ctx.private_u32, ctx.local_memory, word_offset)}; return ctx.OpLoad(ctx.U32[1], pointer); diff --git a/src/shader_recompiler/backend/spirv/emit_spirv_image.cpp b/src/shader_recompiler/backend/spirv/emit_spirv_image.cpp index 1d5364309..4d168a96d 100644 --- a/src/shader_recompiler/backend/spirv/emit_spirv_image.cpp +++ b/src/shader_recompiler/backend/spirv/emit_spirv_image.cpp @@ -224,6 +224,36 @@ Id Emit(MethodPtrType sparse_ptr, MethodPtrType non_sparse_ptr, EmitContext& ctx Decorate(ctx, inst, sample); return ctx.OpCompositeExtract(result_type, sample, 1U); } + +Id IsScaled(EmitContext& ctx, const IR::Value& index, Id member_index, u32 base_index) { + const Id push_constant_u32{ctx.TypePointer(spv::StorageClass::PushConstant, ctx.U32[1])}; + Id bit{}; + if (index.IsImmediate()) { + // Use BitwiseAnd instead of BitfieldExtract for better codegen on Nvidia OpenGL. + // LOP32I.NZ is used to set the predicate rather than BFE+ISETP. + const u32 index_value{index.U32() + base_index}; + const Id word_index{ctx.Const(index_value / 32)}; + const Id bit_index_mask{ctx.Const(1u << (index_value % 32))}; + const Id pointer{ctx.OpAccessChain(push_constant_u32, ctx.rescaling_push_constants, + member_index, word_index)}; + const Id word{ctx.OpLoad(ctx.U32[1], pointer)}; + bit = ctx.OpBitwiseAnd(ctx.U32[1], word, bit_index_mask); + } else { + Id index_value{ctx.Def(index)}; + if (base_index != 0) { + index_value = ctx.OpIAdd(ctx.U32[1], index_value, ctx.Const(base_index)); + } + const Id bit_index{ctx.OpBitwiseAnd(ctx.U32[1], index_value, ctx.Const(31u))}; + bit = ctx.OpBitFieldUExtract(ctx.U32[1], index_value, bit_index, ctx.Const(1u)); + } + return ctx.OpINotEqual(ctx.U1, bit, ctx.u32_zero_value); +} + +Id BitTest(EmitContext& ctx, Id mask, Id bit) { + const Id shifted{ctx.OpShiftRightLogical(ctx.U32[1], mask, bit)}; + const Id bit_value{ctx.OpBitwiseAnd(ctx.U32[1], shifted, ctx.Const(1u))}; + return ctx.OpINotEqual(ctx.U1, bit_value, ctx.u32_zero_value); +} } // Anonymous namespace Id EmitBindlessImageSampleImplicitLod(EmitContext&) { @@ -470,4 +500,28 @@ void EmitImageWrite(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id ctx.OpImageWrite(Image(ctx, index, info), coords, color); } +Id EmitIsTextureScaled(EmitContext& ctx, const IR::Value& index) { + if (ctx.profile.unified_descriptor_binding) { + const Id member_index{ctx.Const(ctx.rescaling_textures_member_index)}; + return IsScaled(ctx, index, member_index, ctx.texture_rescaling_index); + } else { + const Id composite{ctx.OpLoad(ctx.F32[4], ctx.rescaling_uniform_constant)}; + const Id mask_f32{ctx.OpCompositeExtract(ctx.F32[1], composite, 0u)}; + const Id mask{ctx.OpBitcast(ctx.U32[1], mask_f32)}; + return BitTest(ctx, mask, ctx.Def(index)); + } +} + +Id EmitIsImageScaled(EmitContext& ctx, const IR::Value& index) { + if (ctx.profile.unified_descriptor_binding) { + const Id member_index{ctx.Const(ctx.rescaling_images_member_index)}; + return IsScaled(ctx, index, member_index, ctx.image_rescaling_index); + } else { + const Id composite{ctx.OpLoad(ctx.F32[4], ctx.rescaling_uniform_constant)}; + const Id mask_f32{ctx.OpCompositeExtract(ctx.F32[1], composite, 1u)}; + const Id mask{ctx.OpBitcast(ctx.U32[1], mask_f32)}; + return BitTest(ctx, mask, ctx.Def(index)); + } +} + } // namespace Shader::Backend::SPIRV diff --git a/src/shader_recompiler/backend/spirv/emit_spirv_instructions.h b/src/shader_recompiler/backend/spirv/emit_spirv_instructions.h index c9db1c164..6cd22dd3e 100644 --- a/src/shader_recompiler/backend/spirv/emit_spirv_instructions.h +++ b/src/shader_recompiler/backend/spirv/emit_spirv_instructions.h @@ -75,6 +75,7 @@ Id EmitInvocationId(EmitContext& ctx); Id EmitSampleId(EmitContext& ctx); Id EmitIsHelperInvocation(EmitContext& ctx); Id EmitYDirection(EmitContext& ctx); +Id EmitResolutionDownFactor(EmitContext& ctx); Id EmitLoadLocal(EmitContext& ctx, Id word_offset); void EmitWriteLocal(EmitContext& ctx, Id word_offset, Id value); Id EmitUndefU1(EmitContext& ctx); @@ -283,6 +284,8 @@ Id EmitIAdd64(EmitContext& ctx, Id a, Id b); Id EmitISub32(EmitContext& ctx, Id a, Id b); Id EmitISub64(EmitContext& ctx, Id a, Id b); Id EmitIMul32(EmitContext& ctx, Id a, Id b); +Id EmitSDiv32(EmitContext& ctx, Id a, Id b); +Id EmitUDiv32(EmitContext& ctx, Id a, Id b); Id EmitINeg32(EmitContext& ctx, Id value); Id EmitINeg64(EmitContext& ctx, Id value); Id EmitIAbs32(EmitContext& ctx, Id value); @@ -510,6 +513,8 @@ Id EmitImageGradient(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, I 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 EmitIsTextureScaled(EmitContext& ctx, const IR::Value& index); +Id EmitIsImageScaled(EmitContext& ctx, const IR::Value& index); Id EmitBindlessImageAtomicIAdd32(EmitContext&); Id EmitBindlessImageAtomicSMin32(EmitContext&); Id EmitBindlessImageAtomicUMin32(EmitContext&); diff --git a/src/shader_recompiler/backend/spirv/emit_spirv_integer.cpp b/src/shader_recompiler/backend/spirv/emit_spirv_integer.cpp index 3501d7495..50277eec3 100644 --- a/src/shader_recompiler/backend/spirv/emit_spirv_integer.cpp +++ b/src/shader_recompiler/backend/spirv/emit_spirv_integer.cpp @@ -72,6 +72,14 @@ Id EmitIMul32(EmitContext& ctx, Id a, Id b) { return ctx.OpIMul(ctx.U32[1], a, b); } +Id EmitSDiv32(EmitContext& ctx, Id a, Id b) { + return ctx.OpSDiv(ctx.U32[1], a, b); +} + +Id EmitUDiv32(EmitContext& ctx, Id a, Id b) { + return ctx.OpUDiv(ctx.U32[1], a, b); +} + Id EmitINeg32(EmitContext& ctx, Id value) { return ctx.OpSNegate(ctx.U32[1], value); } diff --git a/src/shader_recompiler/frontend/ir/basic_block.cpp b/src/shader_recompiler/frontend/ir/basic_block.cpp index 7c08b25ce..974efa4a0 100644 --- a/src/shader_recompiler/frontend/ir/basic_block.cpp +++ b/src/shader_recompiler/frontend/ir/basic_block.cpp @@ -22,6 +22,11 @@ void Block::AppendNewInst(Opcode op, std::initializer_list<Value> args) { PrependNewInst(end(), op, args); } +Block::iterator Block::PrependNewInst(iterator insertion_point, const Inst& base_inst) { + Inst* const inst{inst_pool->Create(base_inst)}; + return instructions.insert(insertion_point, *inst); +} + Block::iterator Block::PrependNewInst(iterator insertion_point, Opcode op, std::initializer_list<Value> args, u32 flags) { Inst* const inst{inst_pool->Create(op, flags)}; diff --git a/src/shader_recompiler/frontend/ir/basic_block.h b/src/shader_recompiler/frontend/ir/basic_block.h index 9ce1ed07e..fbfe98266 100644 --- a/src/shader_recompiler/frontend/ir/basic_block.h +++ b/src/shader_recompiler/frontend/ir/basic_block.h @@ -40,6 +40,9 @@ public: /// Appends a new instruction to the end of this basic block. void AppendNewInst(Opcode op, std::initializer_list<Value> args); + /// Prepends a copy of an instruction to this basic block before the insertion point. + iterator PrependNewInst(iterator insertion_point, const Inst& base_inst); + /// Prepends a new instruction to this basic block before the insertion point. iterator PrependNewInst(iterator insertion_point, Opcode op, std::initializer_list<Value> args = {}, u32 flags = 0); diff --git a/src/shader_recompiler/frontend/ir/ir_emitter.cpp b/src/shader_recompiler/frontend/ir/ir_emitter.cpp index 13159a68d..356f889ac 100644 --- a/src/shader_recompiler/frontend/ir/ir_emitter.cpp +++ b/src/shader_recompiler/frontend/ir/ir_emitter.cpp @@ -375,6 +375,10 @@ F32 IREmitter::YDirection() { return Inst<F32>(Opcode::YDirection); } +F32 IREmitter::ResolutionDownFactor() { + return Inst<F32>(Opcode::ResolutionDownFactor); +} + U32 IREmitter::LaneId() { return Inst<U32>(Opcode::LaneId); } @@ -1141,6 +1145,10 @@ U32 IREmitter::IMul(const U32& a, const U32& b) { return Inst<U32>(Opcode::IMul32, a, b); } +U32 IREmitter::IDiv(const U32& a, const U32& b, bool is_signed) { + return Inst<U32>(is_signed ? Opcode::SDiv32 : Opcode::UDiv32, a, b); +} + U32U64 IREmitter::INeg(const U32U64& value) { switch (value.Type()) { case Type::U32: @@ -1938,6 +1946,14 @@ Value IREmitter::ImageAtomicExchange(const Value& handle, const Value& coords, c return Inst(op, Flags{info}, handle, coords, value); } +U1 IREmitter::IsTextureScaled(const U32& index) { + return Inst<U1>(Opcode::IsTextureScaled, index); +} + +U1 IREmitter::IsImageScaled(const U32& index) { + return Inst<U1>(Opcode::IsImageScaled, index); +} + U1 IREmitter::VoteAll(const U1& value) { return Inst<U1>(Opcode::VoteAll, value); } diff --git a/src/shader_recompiler/frontend/ir/ir_emitter.h b/src/shader_recompiler/frontend/ir/ir_emitter.h index 1b89ca5a0..13eefa88b 100644 --- a/src/shader_recompiler/frontend/ir/ir_emitter.h +++ b/src/shader_recompiler/frontend/ir/ir_emitter.h @@ -102,6 +102,8 @@ public: [[nodiscard]] U1 IsHelperInvocation(); [[nodiscard]] F32 YDirection(); + [[nodiscard]] F32 ResolutionDownFactor(); + [[nodiscard]] U32 LaneId(); [[nodiscard]] U32 LoadGlobalU8(const U64& address); @@ -207,6 +209,7 @@ public: [[nodiscard]] U32U64 IAdd(const U32U64& a, const U32U64& b); [[nodiscard]] U32U64 ISub(const U32U64& a, const U32U64& b); [[nodiscard]] U32 IMul(const U32& a, const U32& b); + [[nodiscard]] U32 IDiv(const U32& a, const U32& b, bool is_signed = false); [[nodiscard]] U32U64 INeg(const U32U64& value); [[nodiscard]] U32 IAbs(const U32& value); [[nodiscard]] U32U64 ShiftLeftLogical(const U32U64& base, const U32& shift); @@ -356,6 +359,10 @@ public: TextureInstInfo info); [[nodiscard]] Value ImageAtomicExchange(const Value& handle, const Value& coords, const Value& value, TextureInstInfo info); + + [[nodiscard]] U1 IsTextureScaled(const U32& index); + [[nodiscard]] U1 IsImageScaled(const U32& index); + [[nodiscard]] U1 VoteAll(const U1& value); [[nodiscard]] U1 VoteAny(const U1& value); [[nodiscard]] U1 VoteEqual(const U1& value); diff --git a/src/shader_recompiler/frontend/ir/microinstruction.cpp b/src/shader_recompiler/frontend/ir/microinstruction.cpp index 30b470bdd..97e2bf6af 100644 --- a/src/shader_recompiler/frontend/ir/microinstruction.cpp +++ b/src/shader_recompiler/frontend/ir/microinstruction.cpp @@ -47,6 +47,17 @@ Inst::Inst(IR::Opcode op_, u32 flags_) noexcept : op{op_}, flags{flags_} { } } +Inst::Inst(const Inst& base) : op{base.op}, flags{base.flags} { + if (base.op == Opcode::Phi) { + throw NotImplementedException("Copying phi node"); + } + std::construct_at(&args); + const size_t num_args{base.NumArgs()}; + for (size_t index = 0; index < num_args; ++index) { + SetArg(index, base.Arg(index)); + } +} + Inst::~Inst() { if (op == Opcode::Phi) { std::destroy_at(&phi_args); diff --git a/src/shader_recompiler/frontend/ir/opcodes.inc b/src/shader_recompiler/frontend/ir/opcodes.inc index d91098c80..6929919df 100644 --- a/src/shader_recompiler/frontend/ir/opcodes.inc +++ b/src/shader_recompiler/frontend/ir/opcodes.inc @@ -62,6 +62,7 @@ OPCODE(InvocationId, U32, OPCODE(SampleId, U32, ) OPCODE(IsHelperInvocation, U1, ) OPCODE(YDirection, F32, ) +OPCODE(ResolutionDownFactor, F32, ) // Undefined OPCODE(UndefU1, U1, ) @@ -286,6 +287,8 @@ OPCODE(IAdd64, U64, U64, OPCODE(ISub32, U32, U32, U32, ) OPCODE(ISub64, U64, U64, U64, ) OPCODE(IMul32, U32, U32, U32, ) +OPCODE(SDiv32, U32, U32, U32, ) +OPCODE(UDiv32, U32, U32, U32, ) OPCODE(INeg32, U32, U32, ) OPCODE(INeg64, U64, U64, ) OPCODE(IAbs32, U32, U32, ) @@ -490,6 +493,9 @@ OPCODE(ImageGradient, F32x4, Opaq OPCODE(ImageRead, U32x4, Opaque, Opaque, ) OPCODE(ImageWrite, Void, Opaque, Opaque, U32x4, ) +OPCODE(IsTextureScaled, U1, U32, ) +OPCODE(IsImageScaled, U1, U32, ) + // Atomic Image operations OPCODE(BindlessImageAtomicIAdd32, U32, U32, Opaque, U32, ) diff --git a/src/shader_recompiler/frontend/ir/value.h b/src/shader_recompiler/frontend/ir/value.h index 6c9ef6bdd..947579852 100644 --- a/src/shader_recompiler/frontend/ir/value.h +++ b/src/shader_recompiler/frontend/ir/value.h @@ -116,10 +116,10 @@ public: class Inst : public boost::intrusive::list_base_hook<> { public: explicit Inst(IR::Opcode op_, u32 flags_) noexcept; + explicit Inst(const Inst& base); ~Inst(); Inst& operator=(const Inst&) = delete; - Inst(const Inst&) = delete; Inst& operator=(Inst&&) = delete; Inst(Inst&&) = delete; diff --git a/src/shader_recompiler/frontend/maxwell/translate_program.cpp b/src/shader_recompiler/frontend/maxwell/translate_program.cpp index 2fc542f0e..267ebe4af 100644 --- a/src/shader_recompiler/frontend/maxwell/translate_program.cpp +++ b/src/shader_recompiler/frontend/maxwell/translate_program.cpp @@ -179,6 +179,10 @@ IR::Program TranslateProgram(ObjectPool<IR::Inst>& inst_pool, ObjectPool<IR::Blo Optimization::TexturePass(env, program); Optimization::ConstantPropagationPass(program); + + if (Settings::values.resolution_info.active) { + Optimization::RescalingPass(program); + } Optimization::DeadCodeEliminationPass(program); if (Settings::values.renderer_debug) { Optimization::VerificationPass(program); diff --git a/src/shader_recompiler/ir_opt/collect_shader_info_pass.cpp b/src/shader_recompiler/ir_opt/collect_shader_info_pass.cpp index f69e1c9cc..1e476d83d 100644 --- a/src/shader_recompiler/ir_opt/collect_shader_info_pass.cpp +++ b/src/shader_recompiler/ir_opt/collect_shader_info_pass.cpp @@ -430,6 +430,11 @@ void VisitUsages(Info& info, IR::Inst& inst) { case IR::Opcode::IsHelperInvocation: info.uses_is_helper_invocation = true; break; + case IR::Opcode::ResolutionDownFactor: + case IR::Opcode::IsTextureScaled: + case IR::Opcode::IsImageScaled: + info.uses_rescaling_uniform = true; + break; case IR::Opcode::LaneId: info.uses_subgroup_invocation_id = true; break; diff --git a/src/shader_recompiler/ir_opt/passes.h b/src/shader_recompiler/ir_opt/passes.h index 2f89b1ea0..f877c7ba0 100644 --- a/src/shader_recompiler/ir_opt/passes.h +++ b/src/shader_recompiler/ir_opt/passes.h @@ -19,6 +19,7 @@ void GlobalMemoryToStorageBufferPass(IR::Program& program); void IdentityRemovalPass(IR::Program& program); void LowerFp16ToFp32(IR::Program& program); void LowerInt64ToInt32(IR::Program& program); +void RescalingPass(IR::Program& program); void SsaRewritePass(IR::Program& program); void TexturePass(Environment& env, IR::Program& program); void VerificationPass(const IR::Program& program); diff --git a/src/shader_recompiler/ir_opt/rescaling_pass.cpp b/src/shader_recompiler/ir_opt/rescaling_pass.cpp new file mode 100644 index 000000000..c28500dd1 --- /dev/null +++ b/src/shader_recompiler/ir_opt/rescaling_pass.cpp @@ -0,0 +1,327 @@ +// Copyright 2021 yuzu Emulator Project +// Licensed under GPLv2 or any later version +// Refer to the license.txt file included. + +#include "common/alignment.h" +#include "common/settings.h" +#include "shader_recompiler/environment.h" +#include "shader_recompiler/frontend/ir/ir_emitter.h" +#include "shader_recompiler/frontend/ir/modifiers.h" +#include "shader_recompiler/frontend/ir/program.h" +#include "shader_recompiler/frontend/ir/value.h" +#include "shader_recompiler/ir_opt/passes.h" +#include "shader_recompiler/shader_info.h" + +namespace Shader::Optimization { +namespace { +[[nodiscard]] bool IsTextureTypeRescalable(TextureType type) { + switch (type) { + case TextureType::Color2D: + case TextureType::ColorArray2D: + return true; + case TextureType::Color1D: + case TextureType::ColorArray1D: + case TextureType::Color3D: + case TextureType::ColorCube: + case TextureType::ColorArrayCube: + case TextureType::Buffer: + break; + } + return false; +} + +void VisitMark(IR::Block& block, IR::Inst& inst) { + switch (inst.GetOpcode()) { + case IR::Opcode::ShuffleIndex: + case IR::Opcode::ShuffleUp: + case IR::Opcode::ShuffleDown: + case IR::Opcode::ShuffleButterfly: { + const IR::Value shfl_arg{inst.Arg(0)}; + if (shfl_arg.IsImmediate()) { + break; + } + const IR::Inst* const arg_inst{shfl_arg.InstRecursive()}; + if (arg_inst->GetOpcode() != IR::Opcode::BitCastU32F32) { + break; + } + const IR::Value bitcast_arg{arg_inst->Arg(0)}; + if (bitcast_arg.IsImmediate()) { + break; + } + IR::Inst* const bitcast_inst{bitcast_arg.InstRecursive()}; + bool must_patch_outside = false; + if (bitcast_inst->GetOpcode() == IR::Opcode::GetAttribute) { + const IR::Attribute attr{bitcast_inst->Arg(0).Attribute()}; + switch (attr) { + case IR::Attribute::PositionX: + case IR::Attribute::PositionY: + bitcast_inst->SetFlags<u32>(0xDEADBEEF); + must_patch_outside = true; + break; + default: + break; + } + } + if (must_patch_outside) { + const auto it{IR::Block::InstructionList::s_iterator_to(inst)}; + IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)}; + const IR::F32 new_inst{&*block.PrependNewInst(it, inst)}; + const IR::F32 up_factor{ir.FPRecip(ir.ResolutionDownFactor())}; + const IR::Value converted{ir.FPMul(new_inst, up_factor)}; + inst.ReplaceUsesWith(converted); + } + break; + } + + default: + break; + } +} + +void PatchFragCoord(IR::Block& block, IR::Inst& inst) { + IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)}; + const IR::F32 down_factor{ir.ResolutionDownFactor()}; + const IR::F32 frag_coord{ir.GetAttribute(inst.Arg(0).Attribute())}; + const IR::F32 downscaled_frag_coord{ir.FPMul(frag_coord, down_factor)}; + inst.ReplaceUsesWith(downscaled_frag_coord); +} + +void PatchPointSize(IR::Block& block, IR::Inst& inst) { + IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)}; + const IR::F32 point_value{inst.Arg(1)}; + const IR::F32 up_factor{ir.FPRecip(ir.ResolutionDownFactor())}; + const IR::F32 upscaled_point_value{ir.FPMul(point_value, up_factor)}; + inst.SetArg(1, upscaled_point_value); +} + +[[nodiscard]] IR::U32 Scale(IR::IREmitter& ir, const IR::U1& is_scaled, const IR::U32& value) { + IR::U32 scaled_value{value}; + if (const u32 up_scale = Settings::values.resolution_info.up_scale; up_scale != 1) { + scaled_value = ir.IMul(scaled_value, ir.Imm32(up_scale)); + } + if (const u32 down_shift = Settings::values.resolution_info.down_shift; down_shift != 0) { + scaled_value = ir.ShiftRightArithmetic(scaled_value, ir.Imm32(down_shift)); + } + return IR::U32{ir.Select(is_scaled, scaled_value, value)}; +} + +[[nodiscard]] IR::U32 SubScale(IR::IREmitter& ir, const IR::U1& is_scaled, const IR::U32& value, + const IR::Attribute attrib) { + const IR::F32 up_factor{ir.Imm32(Settings::values.resolution_info.up_factor)}; + const IR::F32 base{ir.FPMul(ir.ConvertUToF(32, 32, value), up_factor)}; + const IR::F32 frag_coord{ir.GetAttribute(attrib)}; + const IR::F32 down_factor{ir.Imm32(Settings::values.resolution_info.down_factor)}; + const IR::F32 floor{ir.FPMul(up_factor, ir.FPFloor(ir.FPMul(frag_coord, down_factor)))}; + const IR::F16F32F64 deviation{ir.FPAdd(base, ir.FPAdd(frag_coord, ir.FPNeg(floor)))}; + return IR::U32{ir.Select(is_scaled, ir.ConvertFToU(32, deviation), value)}; +} + +[[nodiscard]] IR::U32 DownScale(IR::IREmitter& ir, const IR::U1& is_scaled, const IR::U32& value) { + IR::U32 scaled_value{value}; + if (const u32 down_shift = Settings::values.resolution_info.down_shift; down_shift != 0) { + scaled_value = ir.ShiftLeftLogical(scaled_value, ir.Imm32(down_shift)); + } + if (const u32 up_scale = Settings::values.resolution_info.up_scale; up_scale != 1) { + scaled_value = ir.IDiv(scaled_value, ir.Imm32(up_scale)); + } + return IR::U32{ir.Select(is_scaled, scaled_value, value)}; +} + +void PatchImageQueryDimensions(IR::Block& block, IR::Inst& inst) { + const auto it{IR::Block::InstructionList::s_iterator_to(inst)}; + IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)}; + const auto info{inst.Flags<IR::TextureInstInfo>()}; + const IR::U1 is_scaled{ir.IsTextureScaled(ir.Imm32(info.descriptor_index))}; + switch (info.type) { + case TextureType::Color2D: + case TextureType::ColorArray2D: { + const IR::Value new_inst{&*block.PrependNewInst(it, inst)}; + const IR::U32 width{DownScale(ir, is_scaled, IR::U32{ir.CompositeExtract(new_inst, 0)})}; + const IR::U32 height{DownScale(ir, is_scaled, IR::U32{ir.CompositeExtract(new_inst, 1)})}; + const IR::Value replacement{ir.CompositeConstruct( + width, height, ir.CompositeExtract(new_inst, 2), ir.CompositeExtract(new_inst, 3))}; + inst.ReplaceUsesWith(replacement); + break; + } + case TextureType::Color1D: + case TextureType::ColorArray1D: + case TextureType::Color3D: + case TextureType::ColorCube: + case TextureType::ColorArrayCube: + case TextureType::Buffer: + // Nothing to patch here + break; + } +} + +void ScaleIntegerComposite(IR::IREmitter& ir, IR::Inst& inst, const IR::U1& is_scaled, + size_t index) { + const IR::Value composite{inst.Arg(index)}; + if (composite.IsEmpty()) { + return; + } + const auto info{inst.Flags<IR::TextureInstInfo>()}; + const IR::U32 x{Scale(ir, is_scaled, IR::U32{ir.CompositeExtract(composite, 0)})}; + const IR::U32 y{Scale(ir, is_scaled, IR::U32{ir.CompositeExtract(composite, 1)})}; + switch (info.type) { + case TextureType::Color2D: + inst.SetArg(index, ir.CompositeConstruct(x, y)); + break; + case TextureType::ColorArray2D: { + const IR::U32 z{ir.CompositeExtract(composite, 2)}; + inst.SetArg(index, ir.CompositeConstruct(x, y, z)); + break; + } + case TextureType::Color1D: + case TextureType::ColorArray1D: + case TextureType::Color3D: + case TextureType::ColorCube: + case TextureType::ColorArrayCube: + case TextureType::Buffer: + // Nothing to patch here + break; + } +} + +void SubScaleCoord(IR::IREmitter& ir, IR::Inst& inst, const IR::U1& is_scaled) { + const auto info{inst.Flags<IR::TextureInstInfo>()}; + const IR::Value coord{inst.Arg(1)}; + const IR::U32 coord_x{ir.CompositeExtract(coord, 0)}; + const IR::U32 coord_y{ir.CompositeExtract(coord, 1)}; + + const IR::U32 scaled_x{SubScale(ir, is_scaled, coord_x, IR::Attribute::PositionX)}; + const IR::U32 scaled_y{SubScale(ir, is_scaled, coord_y, IR::Attribute::PositionY)}; + switch (info.type) { + case TextureType::Color2D: + inst.SetArg(1, ir.CompositeConstruct(scaled_x, scaled_y)); + break; + case TextureType::ColorArray2D: { + const IR::U32 z{ir.CompositeExtract(coord, 2)}; + inst.SetArg(1, ir.CompositeConstruct(scaled_x, scaled_y, z)); + break; + } + case TextureType::Color1D: + case TextureType::ColorArray1D: + case TextureType::Color3D: + case TextureType::ColorCube: + case TextureType::ColorArrayCube: + case TextureType::Buffer: + // Nothing to patch here + break; + } +} + +void SubScaleImageFetch(IR::Block& block, IR::Inst& inst) { + IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)}; + const auto info{inst.Flags<IR::TextureInstInfo>()}; + if (!IsTextureTypeRescalable(info.type)) { + return; + } + const IR::U1 is_scaled{ir.IsTextureScaled(ir.Imm32(info.descriptor_index))}; + SubScaleCoord(ir, inst, is_scaled); + // Scale ImageFetch offset + ScaleIntegerComposite(ir, inst, is_scaled, 2); +} + +void SubScaleImageRead(IR::Block& block, IR::Inst& inst) { + IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)}; + const auto info{inst.Flags<IR::TextureInstInfo>()}; + if (!IsTextureTypeRescalable(info.type)) { + return; + } + const IR::U1 is_scaled{ir.IsImageScaled(ir.Imm32(info.descriptor_index))}; + SubScaleCoord(ir, inst, is_scaled); +} + +void PatchImageFetch(IR::Block& block, IR::Inst& inst) { + IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)}; + const auto info{inst.Flags<IR::TextureInstInfo>()}; + if (!IsTextureTypeRescalable(info.type)) { + return; + } + const IR::U1 is_scaled{ir.IsTextureScaled(ir.Imm32(info.descriptor_index))}; + ScaleIntegerComposite(ir, inst, is_scaled, 1); + // Scale ImageFetch offset + ScaleIntegerComposite(ir, inst, is_scaled, 2); +} + +void PatchImageRead(IR::Block& block, IR::Inst& inst) { + IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)}; + const auto info{inst.Flags<IR::TextureInstInfo>()}; + if (!IsTextureTypeRescalable(info.type)) { + return; + } + const IR::U1 is_scaled{ir.IsImageScaled(ir.Imm32(info.descriptor_index))}; + ScaleIntegerComposite(ir, inst, is_scaled, 1); +} + +void Visit(const IR::Program& program, IR::Block& block, IR::Inst& inst) { + const bool is_fragment_shader{program.stage == Stage::Fragment}; + switch (inst.GetOpcode()) { + case IR::Opcode::GetAttribute: { + const IR::Attribute attr{inst.Arg(0).Attribute()}; + switch (attr) { + case IR::Attribute::PositionX: + case IR::Attribute::PositionY: + if (is_fragment_shader && inst.Flags<u32>() != 0xDEADBEEF) { + PatchFragCoord(block, inst); + } + break; + default: + break; + } + break; + } + case IR::Opcode::SetAttribute: { + const IR::Attribute attr{inst.Arg(0).Attribute()}; + switch (attr) { + case IR::Attribute::PointSize: + if (inst.Flags<u32>() != 0xDEADBEEF) { + PatchPointSize(block, inst); + } + break; + default: + break; + } + break; + } + case IR::Opcode::ImageQueryDimensions: + PatchImageQueryDimensions(block, inst); + break; + case IR::Opcode::ImageFetch: + if (is_fragment_shader) { + SubScaleImageFetch(block, inst); + } else { + PatchImageFetch(block, inst); + } + break; + case IR::Opcode::ImageRead: + if (is_fragment_shader) { + SubScaleImageRead(block, inst); + } else { + PatchImageRead(block, inst); + } + break; + default: + break; + } +} +} // Anonymous namespace + +void RescalingPass(IR::Program& program) { + const bool is_fragment_shader{program.stage == Stage::Fragment}; + if (is_fragment_shader) { + for (IR::Block* const block : program.post_order_blocks) { + for (IR::Inst& inst : block->Instructions()) { + VisitMark(*block, inst); + } + } + } + for (IR::Block* const block : program.post_order_blocks) { + for (IR::Inst& inst : block->Instructions()) { + Visit(program, *block, inst); + } + } +} + +} // namespace Shader::Optimization diff --git a/src/shader_recompiler/shader_info.h b/src/shader_recompiler/shader_info.h index 4ef4dbd40..9f375c30e 100644 --- a/src/shader_recompiler/shader_info.h +++ b/src/shader_recompiler/shader_info.h @@ -172,6 +172,7 @@ struct Info { bool uses_global_memory{}; bool uses_atomic_image_u32{}; bool uses_shadow_lod{}; + bool uses_rescaling_uniform{}; IR::Type used_constant_buffer_types{}; IR::Type used_storage_buffer_types{}; @@ -190,4 +191,13 @@ struct Info { ImageDescriptors image_descriptors; }; +template <typename Descriptors> +u32 NumDescriptors(const Descriptors& descriptors) { + u32 num{}; + for (const auto& desc : descriptors) { + num += desc.count; + } + return num; +} + } // namespace Shader |