| 123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560 |
- // Copyright 2021 yuzu Emulator Project
- // Licensed under GPLv2 or any later version
- // Refer to the license.txt file included.
- #include "shader_recompiler/environment.h"
- #include "shader_recompiler/frontend/ir/microinstruction.h"
- #include "shader_recompiler/frontend/ir/modifiers.h"
- #include "shader_recompiler/frontend/ir/program.h"
- #include "shader_recompiler/shader_info.h"
- namespace Shader::Optimization {
- namespace {
- void AddConstantBufferDescriptor(Info& info, u32 index, u32 count) {
- if (count != 1) {
- throw NotImplementedException("Constant buffer descriptor indexing");
- }
- if ((info.constant_buffer_mask & (1U << index)) != 0) {
- return;
- }
- info.constant_buffer_mask |= 1U << index;
- auto& cbufs{info.constant_buffer_descriptors};
- cbufs.insert(std::ranges::lower_bound(cbufs, index, {}, &ConstantBufferDescriptor::index),
- ConstantBufferDescriptor{
- .index{index},
- .count{1},
- });
- }
- void GetAttribute(Info& info, IR::Attribute attribute) {
- if (IR::IsGeneric(attribute)) {
- info.input_generics.at(IR::GenericAttributeIndex(attribute)).used = true;
- return;
- }
- switch (attribute) {
- case IR::Attribute::PositionX:
- case IR::Attribute::PositionY:
- case IR::Attribute::PositionZ:
- case IR::Attribute::PositionW:
- info.loads_position = true;
- break;
- case IR::Attribute::InstanceId:
- info.loads_instance_id = true;
- break;
- case IR::Attribute::VertexId:
- info.loads_vertex_id = true;
- break;
- case IR::Attribute::FrontFace:
- info.loads_front_face = true;
- break;
- case IR::Attribute::PointSpriteS:
- case IR::Attribute::PointSpriteT:
- info.loads_point_coord = true;
- break;
- default:
- throw NotImplementedException("Get attribute {}", attribute);
- }
- }
- void SetAttribute(Info& info, IR::Attribute attribute) {
- if (IR::IsGeneric(attribute)) {
- info.stores_generics.at(IR::GenericAttributeIndex(attribute)) = true;
- return;
- }
- switch (attribute) {
- case IR::Attribute::PointSize:
- info.stores_point_size = true;
- break;
- case IR::Attribute::PositionX:
- case IR::Attribute::PositionY:
- case IR::Attribute::PositionZ:
- case IR::Attribute::PositionW:
- info.stores_position = true;
- break;
- case IR::Attribute::ClipDistance0:
- case IR::Attribute::ClipDistance1:
- case IR::Attribute::ClipDistance2:
- case IR::Attribute::ClipDistance3:
- case IR::Attribute::ClipDistance4:
- case IR::Attribute::ClipDistance5:
- case IR::Attribute::ClipDistance6:
- case IR::Attribute::ClipDistance7:
- info.stores_clip_distance = true;
- break;
- case IR::Attribute::ViewportIndex:
- info.stores_viewport_index = true;
- break;
- default:
- throw NotImplementedException("Set attribute {}", attribute);
- }
- }
- void VisitUsages(Info& info, IR::Inst& inst) {
- switch (inst.Opcode()) {
- case IR::Opcode::CompositeConstructF16x2:
- case IR::Opcode::CompositeConstructF16x3:
- case IR::Opcode::CompositeConstructF16x4:
- case IR::Opcode::CompositeExtractF16x2:
- case IR::Opcode::CompositeExtractF16x3:
- case IR::Opcode::CompositeExtractF16x4:
- case IR::Opcode::CompositeInsertF16x2:
- case IR::Opcode::CompositeInsertF16x3:
- case IR::Opcode::CompositeInsertF16x4:
- case IR::Opcode::SelectF16:
- case IR::Opcode::BitCastU16F16:
- case IR::Opcode::BitCastF16U16:
- case IR::Opcode::PackFloat2x16:
- case IR::Opcode::UnpackFloat2x16:
- case IR::Opcode::ConvertS16F16:
- case IR::Opcode::ConvertS32F16:
- case IR::Opcode::ConvertS64F16:
- case IR::Opcode::ConvertU16F16:
- case IR::Opcode::ConvertU32F16:
- case IR::Opcode::ConvertU64F16:
- case IR::Opcode::ConvertF16S8:
- case IR::Opcode::ConvertF16S16:
- case IR::Opcode::ConvertF16S32:
- case IR::Opcode::ConvertF16S64:
- case IR::Opcode::ConvertF16U8:
- case IR::Opcode::ConvertF16U16:
- case IR::Opcode::ConvertF16U32:
- case IR::Opcode::ConvertF16U64:
- case IR::Opcode::FPAbs16:
- case IR::Opcode::FPAdd16:
- case IR::Opcode::FPCeil16:
- case IR::Opcode::FPFloor16:
- case IR::Opcode::FPFma16:
- case IR::Opcode::FPMul16:
- case IR::Opcode::FPNeg16:
- case IR::Opcode::FPRoundEven16:
- case IR::Opcode::FPSaturate16:
- case IR::Opcode::FPClamp16:
- case IR::Opcode::FPTrunc16:
- case IR::Opcode::FPOrdEqual16:
- case IR::Opcode::FPUnordEqual16:
- case IR::Opcode::FPOrdNotEqual16:
- case IR::Opcode::FPUnordNotEqual16:
- case IR::Opcode::FPOrdLessThan16:
- case IR::Opcode::FPUnordLessThan16:
- case IR::Opcode::FPOrdGreaterThan16:
- case IR::Opcode::FPUnordGreaterThan16:
- case IR::Opcode::FPOrdLessThanEqual16:
- case IR::Opcode::FPUnordLessThanEqual16:
- case IR::Opcode::FPOrdGreaterThanEqual16:
- case IR::Opcode::FPUnordGreaterThanEqual16:
- case IR::Opcode::FPIsNan16:
- info.uses_fp16 = true;
- break;
- case IR::Opcode::CompositeConstructF64x2:
- case IR::Opcode::CompositeConstructF64x3:
- case IR::Opcode::CompositeConstructF64x4:
- case IR::Opcode::CompositeExtractF64x2:
- case IR::Opcode::CompositeExtractF64x3:
- case IR::Opcode::CompositeExtractF64x4:
- case IR::Opcode::CompositeInsertF64x2:
- case IR::Opcode::CompositeInsertF64x3:
- case IR::Opcode::CompositeInsertF64x4:
- case IR::Opcode::SelectF64:
- case IR::Opcode::BitCastU64F64:
- case IR::Opcode::BitCastF64U64:
- case IR::Opcode::PackDouble2x32:
- case IR::Opcode::UnpackDouble2x32:
- case IR::Opcode::FPAbs64:
- case IR::Opcode::FPAdd64:
- case IR::Opcode::FPCeil64:
- case IR::Opcode::FPFloor64:
- case IR::Opcode::FPFma64:
- case IR::Opcode::FPMax64:
- case IR::Opcode::FPMin64:
- case IR::Opcode::FPMul64:
- case IR::Opcode::FPNeg64:
- case IR::Opcode::FPRecip64:
- case IR::Opcode::FPRecipSqrt64:
- case IR::Opcode::FPRoundEven64:
- case IR::Opcode::FPSaturate64:
- case IR::Opcode::FPClamp64:
- case IR::Opcode::FPTrunc64:
- case IR::Opcode::FPOrdEqual64:
- case IR::Opcode::FPUnordEqual64:
- case IR::Opcode::FPOrdNotEqual64:
- case IR::Opcode::FPUnordNotEqual64:
- case IR::Opcode::FPOrdLessThan64:
- case IR::Opcode::FPUnordLessThan64:
- case IR::Opcode::FPOrdGreaterThan64:
- case IR::Opcode::FPUnordGreaterThan64:
- case IR::Opcode::FPOrdLessThanEqual64:
- case IR::Opcode::FPUnordLessThanEqual64:
- case IR::Opcode::FPOrdGreaterThanEqual64:
- case IR::Opcode::FPUnordGreaterThanEqual64:
- case IR::Opcode::FPIsNan64:
- case IR::Opcode::ConvertS16F64:
- case IR::Opcode::ConvertS32F64:
- case IR::Opcode::ConvertS64F64:
- case IR::Opcode::ConvertU16F64:
- case IR::Opcode::ConvertU32F64:
- case IR::Opcode::ConvertU64F64:
- case IR::Opcode::ConvertF32F64:
- case IR::Opcode::ConvertF64F32:
- case IR::Opcode::ConvertF64S8:
- case IR::Opcode::ConvertF64S16:
- case IR::Opcode::ConvertF64S32:
- case IR::Opcode::ConvertF64S64:
- case IR::Opcode::ConvertF64U8:
- case IR::Opcode::ConvertF64U16:
- case IR::Opcode::ConvertF64U32:
- case IR::Opcode::ConvertF64U64:
- info.uses_fp64 = true;
- break;
- default:
- break;
- }
- switch (inst.Opcode()) {
- case IR::Opcode::GetCbufU8:
- case IR::Opcode::GetCbufS8:
- case IR::Opcode::UndefU8:
- case IR::Opcode::LoadGlobalU8:
- case IR::Opcode::LoadGlobalS8:
- case IR::Opcode::WriteGlobalU8:
- case IR::Opcode::WriteGlobalS8:
- case IR::Opcode::LoadStorageU8:
- case IR::Opcode::LoadStorageS8:
- case IR::Opcode::WriteStorageU8:
- case IR::Opcode::WriteStorageS8:
- case IR::Opcode::LoadSharedU8:
- case IR::Opcode::LoadSharedS8:
- case IR::Opcode::WriteSharedU8:
- case IR::Opcode::SelectU8:
- case IR::Opcode::ConvertF16S8:
- case IR::Opcode::ConvertF16U8:
- case IR::Opcode::ConvertF32S8:
- case IR::Opcode::ConvertF32U8:
- case IR::Opcode::ConvertF64S8:
- case IR::Opcode::ConvertF64U8:
- info.uses_int8 = true;
- break;
- default:
- break;
- }
- switch (inst.Opcode()) {
- case IR::Opcode::GetCbufU16:
- case IR::Opcode::GetCbufS16:
- case IR::Opcode::UndefU16:
- case IR::Opcode::LoadGlobalU16:
- case IR::Opcode::LoadGlobalS16:
- case IR::Opcode::WriteGlobalU16:
- case IR::Opcode::WriteGlobalS16:
- case IR::Opcode::LoadStorageU16:
- case IR::Opcode::LoadStorageS16:
- case IR::Opcode::WriteStorageU16:
- case IR::Opcode::WriteStorageS16:
- case IR::Opcode::LoadSharedU16:
- case IR::Opcode::LoadSharedS16:
- case IR::Opcode::WriteSharedU16:
- case IR::Opcode::SelectU16:
- case IR::Opcode::BitCastU16F16:
- case IR::Opcode::BitCastF16U16:
- case IR::Opcode::ConvertS16F16:
- case IR::Opcode::ConvertS16F32:
- case IR::Opcode::ConvertS16F64:
- case IR::Opcode::ConvertU16F16:
- case IR::Opcode::ConvertU16F32:
- case IR::Opcode::ConvertU16F64:
- case IR::Opcode::ConvertF16S16:
- case IR::Opcode::ConvertF16U16:
- case IR::Opcode::ConvertF32S16:
- case IR::Opcode::ConvertF32U16:
- case IR::Opcode::ConvertF64S16:
- case IR::Opcode::ConvertF64U16:
- info.uses_int16 = true;
- break;
- default:
- break;
- }
- switch (inst.Opcode()) {
- case IR::Opcode::UndefU64:
- case IR::Opcode::LoadGlobalU8:
- case IR::Opcode::LoadGlobalS8:
- case IR::Opcode::LoadGlobalU16:
- case IR::Opcode::LoadGlobalS16:
- case IR::Opcode::LoadGlobal32:
- case IR::Opcode::LoadGlobal64:
- case IR::Opcode::LoadGlobal128:
- case IR::Opcode::WriteGlobalU8:
- case IR::Opcode::WriteGlobalS8:
- case IR::Opcode::WriteGlobalU16:
- case IR::Opcode::WriteGlobalS16:
- case IR::Opcode::WriteGlobal32:
- case IR::Opcode::WriteGlobal64:
- case IR::Opcode::WriteGlobal128:
- case IR::Opcode::SelectU64:
- case IR::Opcode::BitCastU64F64:
- case IR::Opcode::BitCastF64U64:
- case IR::Opcode::PackUint2x32:
- case IR::Opcode::UnpackUint2x32:
- case IR::Opcode::IAdd64:
- case IR::Opcode::ISub64:
- case IR::Opcode::INeg64:
- case IR::Opcode::ShiftLeftLogical64:
- case IR::Opcode::ShiftRightLogical64:
- case IR::Opcode::ShiftRightArithmetic64:
- case IR::Opcode::ConvertS64F16:
- case IR::Opcode::ConvertS64F32:
- case IR::Opcode::ConvertS64F64:
- case IR::Opcode::ConvertU64F16:
- case IR::Opcode::ConvertU64F32:
- case IR::Opcode::ConvertU64F64:
- case IR::Opcode::ConvertU64U32:
- case IR::Opcode::ConvertU32U64:
- case IR::Opcode::ConvertF16U64:
- case IR::Opcode::ConvertF32U64:
- case IR::Opcode::ConvertF64U64:
- info.uses_int64 = true;
- break;
- default:
- break;
- }
- switch (inst.Opcode()) {
- case IR::Opcode::DemoteToHelperInvocation:
- info.uses_demote_to_helper_invocation = true;
- break;
- case IR::Opcode::GetAttribute:
- GetAttribute(info, inst.Arg(0).Attribute());
- break;
- case IR::Opcode::SetAttribute:
- SetAttribute(info, inst.Arg(0).Attribute());
- break;
- case IR::Opcode::GetAttributeIndexed:
- info.loads_indexed_attributes = true;
- break;
- case IR::Opcode::SetAttributeIndexed:
- info.stores_indexed_attributes = true;
- break;
- case IR::Opcode::SetFragColor:
- info.stores_frag_color[inst.Arg(0).U32()] = true;
- break;
- case IR::Opcode::SetFragDepth:
- info.stores_frag_depth = true;
- break;
- case IR::Opcode::WorkgroupId:
- info.uses_workgroup_id = true;
- break;
- case IR::Opcode::LocalInvocationId:
- info.uses_local_invocation_id = true;
- break;
- case IR::Opcode::ShuffleIndex:
- case IR::Opcode::ShuffleUp:
- case IR::Opcode::ShuffleDown:
- case IR::Opcode::ShuffleButterfly:
- case IR::Opcode::LaneId:
- info.uses_subgroup_invocation_id = true;
- break;
- case IR::Opcode::GetCbufU8:
- case IR::Opcode::GetCbufS8:
- case IR::Opcode::GetCbufU16:
- case IR::Opcode::GetCbufS16:
- case IR::Opcode::GetCbufU32:
- case IR::Opcode::GetCbufF32:
- case IR::Opcode::GetCbufU32x2: {
- if (const IR::Value index{inst.Arg(0)}; index.IsImmediate()) {
- AddConstantBufferDescriptor(info, index.U32(), 1);
- } else {
- throw NotImplementedException("Constant buffer with non-immediate index");
- }
- switch (inst.Opcode()) {
- case IR::Opcode::GetCbufU8:
- case IR::Opcode::GetCbufS8:
- info.used_constant_buffer_types |= IR::Type::U8;
- break;
- case IR::Opcode::GetCbufU16:
- case IR::Opcode::GetCbufS16:
- info.used_constant_buffer_types |= IR::Type::U16;
- break;
- case IR::Opcode::GetCbufU32:
- info.used_constant_buffer_types |= IR::Type::U32;
- break;
- case IR::Opcode::GetCbufF32:
- info.used_constant_buffer_types |= IR::Type::F32;
- break;
- case IR::Opcode::GetCbufU32x2:
- info.used_constant_buffer_types |= IR::Type::U32x2;
- break;
- default:
- break;
- }
- break;
- }
- case IR::Opcode::BindlessImageSampleImplicitLod:
- case IR::Opcode::BindlessImageSampleExplicitLod:
- case IR::Opcode::BindlessImageSampleDrefImplicitLod:
- case IR::Opcode::BindlessImageSampleDrefExplicitLod:
- case IR::Opcode::BindlessImageGather:
- case IR::Opcode::BindlessImageGatherDref:
- case IR::Opcode::BindlessImageFetch:
- case IR::Opcode::BindlessImageQueryDimensions:
- case IR::Opcode::BindlessImageQueryLod:
- case IR::Opcode::BindlessImageGradient:
- case IR::Opcode::BoundImageSampleImplicitLod:
- case IR::Opcode::BoundImageSampleExplicitLod:
- case IR::Opcode::BoundImageSampleDrefImplicitLod:
- case IR::Opcode::BoundImageSampleDrefExplicitLod:
- case IR::Opcode::BoundImageGather:
- case IR::Opcode::BoundImageGatherDref:
- case IR::Opcode::BoundImageFetch:
- case IR::Opcode::BoundImageQueryDimensions:
- case IR::Opcode::BoundImageQueryLod:
- case IR::Opcode::BoundImageGradient:
- case IR::Opcode::ImageSampleImplicitLod:
- case IR::Opcode::ImageSampleExplicitLod:
- case IR::Opcode::ImageSampleDrefImplicitLod:
- case IR::Opcode::ImageSampleDrefExplicitLod:
- case IR::Opcode::ImageGather:
- case IR::Opcode::ImageGatherDref:
- case IR::Opcode::ImageFetch:
- case IR::Opcode::ImageQueryDimensions:
- 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_sparse_residency |=
- inst.GetAssociatedPseudoOperation(IR::Opcode::GetSparseFromOp) != nullptr;
- break;
- }
- case IR::Opcode::SubgroupEqMask:
- case IR::Opcode::SubgroupLtMask:
- case IR::Opcode::SubgroupLeMask:
- case IR::Opcode::SubgroupGtMask:
- case IR::Opcode::SubgroupGeMask:
- info.uses_subgroup_mask = true;
- break;
- case IR::Opcode::VoteAll:
- case IR::Opcode::VoteAny:
- case IR::Opcode::VoteEqual:
- case IR::Opcode::SubgroupBallot:
- info.uses_subgroup_vote = true;
- break;
- case IR::Opcode::FSwizzleAdd:
- info.uses_fswzadd = true;
- break;
- default:
- break;
- }
- }
- void VisitFpModifiers(Info& info, IR::Inst& inst) {
- switch (inst.Opcode()) {
- case IR::Opcode::FPAdd16:
- case IR::Opcode::FPFma16:
- case IR::Opcode::FPMul16:
- case IR::Opcode::FPRoundEven16:
- case IR::Opcode::FPFloor16:
- case IR::Opcode::FPCeil16:
- case IR::Opcode::FPTrunc16: {
- const auto control{inst.Flags<IR::FpControl>()};
- switch (control.fmz_mode) {
- case IR::FmzMode::DontCare:
- break;
- case IR::FmzMode::FTZ:
- case IR::FmzMode::FMZ:
- info.uses_fp16_denorms_flush = true;
- break;
- case IR::FmzMode::None:
- info.uses_fp16_denorms_preserve = true;
- break;
- }
- break;
- }
- case IR::Opcode::FPAdd32:
- case IR::Opcode::FPFma32:
- case IR::Opcode::FPMul32:
- case IR::Opcode::FPRoundEven32:
- case IR::Opcode::FPFloor32:
- case IR::Opcode::FPCeil32:
- case IR::Opcode::FPTrunc32:
- case IR::Opcode::FPOrdEqual32:
- case IR::Opcode::FPUnordEqual32:
- case IR::Opcode::FPOrdNotEqual32:
- case IR::Opcode::FPUnordNotEqual32:
- case IR::Opcode::FPOrdLessThan32:
- case IR::Opcode::FPUnordLessThan32:
- case IR::Opcode::FPOrdGreaterThan32:
- case IR::Opcode::FPUnordGreaterThan32:
- case IR::Opcode::FPOrdLessThanEqual32:
- case IR::Opcode::FPUnordLessThanEqual32:
- case IR::Opcode::FPOrdGreaterThanEqual32:
- case IR::Opcode::FPUnordGreaterThanEqual32:
- case IR::Opcode::ConvertF16F32:
- case IR::Opcode::ConvertF64F32: {
- const auto control{inst.Flags<IR::FpControl>()};
- switch (control.fmz_mode) {
- case IR::FmzMode::DontCare:
- break;
- case IR::FmzMode::FTZ:
- case IR::FmzMode::FMZ:
- info.uses_fp32_denorms_flush = true;
- break;
- case IR::FmzMode::None:
- info.uses_fp32_denorms_preserve = true;
- break;
- }
- break;
- }
- default:
- break;
- }
- }
- void Visit(Info& info, IR::Inst& inst) {
- VisitUsages(info, inst);
- VisitFpModifiers(info, inst);
- }
- void GatherInfoFromHeader(Environment& env, Info& info) {
- auto stage = env.ShaderStage();
- if (stage == Stage::Compute) {
- return;
- }
- const auto& header = env.SPH();
- if (stage == Stage::Fragment) {
- if (!info.loads_indexed_attributes) {
- return;
- }
- for (size_t i = 0; i < info.input_generics.size(); i++) {
- info.input_generics[i].used =
- info.input_generics[i].used || header.ps.IsGenericVectorActive(i);
- }
- info.loads_position = info.loads_position || header.ps.imap_systemb.position != 0;
- return;
- }
- if (info.loads_indexed_attributes) {
- for (size_t i = 0; i < info.input_generics.size(); i++) {
- info.input_generics[i].used =
- info.input_generics[i].used || header.vtg.IsInputGenericVectorActive(i);
- }
- }
- if (info.stores_indexed_attributes) {
- info.loads_position = info.loads_position || header.vtg.imap_systemb.position != 0;
- for (size_t i = 0; i < info.stores_generics.size(); i++) {
- info.stores_generics[i] =
- info.stores_generics[i] || header.vtg.IsOutputGenericVectorActive(i);
- }
- info.stores_clip_distance =
- info.stores_clip_distance || header.vtg.omap_systemc.clip_distances != 0;
- info.stores_position = info.stores_position || header.vtg.omap_systemb.position != 0;
- }
- }
- } // Anonymous namespace
- void CollectShaderInfoPass(Environment& env, IR::Program& program) {
- Info& info{program.info};
- for (IR::Block* const block : program.post_order_blocks) {
- for (IR::Inst& inst : block->Instructions()) {
- Visit(info, inst);
- }
- }
- GatherInfoFromHeader(env, info);
- }
- } // namespace Shader::Optimization
|