From ec6fc5fe78c9038fc9ad7259b7b3a7be751ecef6 Mon Sep 17 00:00:00 2001 From: ReinUsesLisp Date: Mon, 17 May 2021 02:52:01 -0300 Subject: glasm: Implement TEX and TEXS instructions Remove lod clamp from texture instructions with lod, as this is not needed (nor supported). --- src/shader_recompiler/backend/glasm/emit_glasm.cpp | 3 + .../backend/glasm/emit_glasm_image.cpp | 282 ++++++++++++++++++--- .../backend/glasm/emit_glasm_instructions.h | 6 +- .../backend/glasm/emit_glasm_memory.cpp | 2 +- .../backend/spirv/emit_spirv_image.cpp | 8 +- .../backend/spirv/emit_spirv_instructions.h | 4 +- 6 files changed, 258 insertions(+), 47 deletions(-) (limited to 'src/shader_recompiler/backend') diff --git a/src/shader_recompiler/backend/glasm/emit_glasm.cpp b/src/shader_recompiler/backend/glasm/emit_glasm.cpp index 0c591f73c..d7a08e4b3 100644 --- a/src/shader_recompiler/backend/glasm/emit_glasm.cpp +++ b/src/shader_recompiler/backend/glasm/emit_glasm.cpp @@ -283,6 +283,9 @@ void SetupOptions(std::string& header, Info info) { if (info.uses_subgroup_shuffles) { header += "OPTION NV_shader_thread_shuffle;"; } + if (info.uses_sparse_residency) { + header += "OPTION EXT_sparse_texture2;"; + } const auto non_zero_frag_colors{info.stores_frag_color | std::views::drop(1)}; if (std::ranges::find(non_zero_frag_colors, true) != non_zero_frag_colors.end()) { header += "OPTION ARB_draw_buffers;"; diff --git a/src/shader_recompiler/backend/glasm/emit_glasm_image.cpp b/src/shader_recompiler/backend/glasm/emit_glasm_image.cpp index 2af5483d9..3d76ab315 100644 --- a/src/shader_recompiler/backend/glasm/emit_glasm_image.cpp +++ b/src/shader_recompiler/backend/glasm/emit_glasm_image.cpp @@ -9,6 +9,34 @@ namespace Shader::Backend::GLASM { namespace { +struct ScopedRegister { + ScopedRegister() = default; + ScopedRegister(RegAlloc& reg_alloc_) : reg_alloc{®_alloc_}, reg{reg_alloc->AllocReg()} {} + + ~ScopedRegister() { + if (reg_alloc) { + reg_alloc->FreeReg(reg); + } + } + + ScopedRegister& operator=(ScopedRegister&& rhs) noexcept { + if (reg_alloc) { + reg_alloc->FreeReg(reg); + } + reg_alloc = std::exchange(rhs.reg_alloc, nullptr); + reg = rhs.reg; + } + + ScopedRegister(ScopedRegister&& rhs) noexcept + : reg_alloc{std::exchange(rhs.reg_alloc, nullptr)}, reg{rhs.reg} {} + + ScopedRegister& operator=(const ScopedRegister&) = delete; + ScopedRegister(const ScopedRegister&) = delete; + + RegAlloc* reg_alloc{}; + Register reg; +}; + std::string Texture([[maybe_unused]] EmitContext& ctx, IR::TextureInstInfo info, [[maybe_unused]] const IR::Value& index) { // FIXME @@ -36,6 +64,61 @@ std::string_view TextureType(IR::TextureInstInfo info) { } throw InvalidArgument("Invalid texture type {}", info.type.Value()); } + +std::string_view ShadowTextureType(IR::TextureInstInfo info) { + switch (info.type) { + case TextureType::Color1D: + return "SHADOW1D"; + case TextureType::ColorArray1D: + return "SHADOWARRAY1D"; + case TextureType::Color2D: + return "SHADOW2D"; + case TextureType::ColorArray2D: + return "SHADOWARRAY2D"; + case TextureType::Color3D: + return "SHADOW3D"; + case TextureType::ColorCube: + return "SHADOWCUBE"; + case TextureType::ColorArrayCube: + return "SHADOWARRAYCUBE"; + case TextureType::Buffer: + return "SHADOWBUFFER"; + } + throw InvalidArgument("Invalid texture type {}", info.type.Value()); +} + +std::string Offset(EmitContext& ctx, const IR::Value& offset) { + if (offset.IsEmpty()) { + return ""; + } + return fmt::format(",offset({})", Register{ctx.reg_alloc.Consume(offset)}); +} + +std::pair Coord(EmitContext& ctx, const IR::Value& coord) { + if (coord.IsImmediate()) { + ScopedRegister scoped_reg(ctx.reg_alloc); + return {fmt::to_string(scoped_reg.reg), std::move(scoped_reg)}; + } + std::string coord_vec{fmt::to_string(Register{ctx.reg_alloc.Consume(coord)})}; + if (coord.InstRecursive()->HasUses()) { + // Move non-dead coords to a separate register, although this should never happen because + // vectors are only assembled for immediate texture instructions + ctx.Add("MOV.F RC,{};", coord_vec); + coord_vec = "RC"; + } + return {std::move(coord_vec), ScopedRegister{}}; +} + +void StoreSparse(EmitContext& ctx, IR::Inst* sparse_inst) { + if (!sparse_inst) { + return; + } + const Register sparse_ret{ctx.reg_alloc.Define(*sparse_inst)}; + ctx.Add("MOV.S {},-1;" + "MOV.S {}(NONRESIDENT),0;", + sparse_ret, sparse_ret); + sparse_inst->Invalidate(); +} } // Anonymous namespace void EmitImageSampleImplicitLod(EmitContext& ctx, IR::Inst& inst, const IR::Value& index, @@ -46,17 +129,8 @@ void EmitImageSampleImplicitLod(EmitContext& ctx, IR::Inst& inst, const IR::Valu const std::string_view lod_clamp_mod{info.has_lod_clamp ? ".LODCLAMP" : ""}; const std::string_view type{TextureType(info)}; const std::string texture{Texture(ctx, info, index)}; - std::string offset_vec; - if (!offset.IsEmpty()) { - offset_vec = fmt::format(",offset({})", Register{ctx.reg_alloc.Consume(offset)}); - } - std::string coord_vec{fmt::to_string(Register{ctx.reg_alloc.Consume(coord)})}; - if (coord.InstRecursive()->HasUses()) { - // Move non-dead coords to a separate register, although this should never happen because - // vectors are only assembled for immediate texture instructions - ctx.Add("MOV.F RC,{};", coord_vec); - coord_vec = "RC"; - } + const std::string offset_vec{Offset(ctx, offset)}; + const auto [coord_vec, coord_alloc]{Coord(ctx, coord)}; const Register ret{ctx.reg_alloc.Define(inst)}; if (info.has_bias) { if (info.type == TextureType::ColorArrayCube) { @@ -83,38 +157,172 @@ void EmitImageSampleImplicitLod(EmitContext& ctx, IR::Inst& inst, const IR::Valu type, offset_vec); } } - if (sparse_inst) { - const Register sparse_ret{ctx.reg_alloc.Define(*sparse_inst)}; - ctx.Add("MOV.S {},-1;" - "MOV.S {}(NONRESIDENT),0;", - sparse_ret, sparse_ret); - sparse_inst->Invalidate(); - } + StoreSparse(ctx, sparse_inst); } -void EmitImageSampleExplicitLod([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst, - [[maybe_unused]] const IR::Value& index, - [[maybe_unused]] Register coord, [[maybe_unused]] Register lod_lc, - [[maybe_unused]] const IR::Value& offset) { - throw NotImplementedException("GLASM instruction"); +void EmitImageSampleExplicitLod(EmitContext& ctx, IR::Inst& inst, const IR::Value& index, + const IR::Value& coord, ScalarF32 lod, const IR::Value& offset) { + const auto info{inst.Flags()}; + const auto sparse_inst{inst.GetAssociatedPseudoOperation(IR::Opcode::GetSparseFromOp)}; + const std::string_view sparse_mod{sparse_inst ? ".SPARSE" : ""}; + const std::string_view type{TextureType(info)}; + const std::string texture{Texture(ctx, info, index)}; + const std::string offset_vec{Offset(ctx, offset)}; + const auto [coord_vec, coord_alloc]{Coord(ctx, coord)}; + const Register ret{ctx.reg_alloc.Define(inst)}; + if (info.type == TextureType::ColorArrayCube) { + ctx.Add("TXL.F{} {},{},{},{},ARRAYCUBE{};", sparse_mod, ret, coord_vec, lod, texture, + offset_vec); + } else { + ctx.Add("MOV.F {}.w,{};" + "TXL.F{} {},{},{},{}{};", + coord_vec, lod, sparse_mod, ret, coord_vec, texture, type, offset_vec); + } + StoreSparse(ctx, sparse_inst); } -void EmitImageSampleDrefImplicitLod([[maybe_unused]] EmitContext& ctx, - [[maybe_unused]] IR::Inst& inst, - [[maybe_unused]] const IR::Value& index, - [[maybe_unused]] Register coord, [[maybe_unused]] Register dref, - [[maybe_unused]] Register bias_lc, - [[maybe_unused]] const IR::Value& offset) { - throw NotImplementedException("GLASM instruction"); +void EmitImageSampleDrefImplicitLod(EmitContext& ctx, IR::Inst& inst, const IR::Value& index, + const IR::Value& coord, ScalarF32 dref, Register bias_lc, + const IR::Value& offset) { + const auto info{inst.Flags()}; + const auto sparse_inst{inst.GetAssociatedPseudoOperation(IR::Opcode::GetSparseFromOp)}; + const std::string_view sparse_mod{sparse_inst ? ".SPARSE" : ""}; + const std::string_view type{ShadowTextureType(info)}; + const std::string texture{Texture(ctx, info, index)}; + const std::string offset_vec{Offset(ctx, offset)}; + const auto [coord_vec, coord_alloc]{Coord(ctx, coord)}; + const Register ret{ctx.reg_alloc.Define(inst)}; + if (info.has_bias) { + if (info.has_lod_clamp) { + switch (info.type) { + case TextureType::Color1D: + case TextureType::ColorArray1D: + case TextureType::Color2D: + ctx.Add("MOV.F {}.z,{};" + "MOV.F {}.w,{}.x;" + "TXB.F.LODCLAMP{} {},{},{}.y,{},{}{};", + coord_vec, dref, coord_vec, bias_lc, sparse_mod, ret, coord_vec, bias_lc, + texture, type, offset_vec); + break; + case TextureType::ColorArray2D: + case TextureType::ColorCube: + ctx.Add("MOV.F {}.w,{};" + "TXB.F.LODCLAMP{} {},{},{},{},{}{};", + coord_vec, dref, sparse_mod, ret, coord_vec, bias_lc, texture, type, + offset_vec); + break; + default: + throw NotImplementedException("Invalid type {} with bias and lod clamp", + info.type.Value()); + } + } else { + switch (info.type) { + case TextureType::Color1D: + case TextureType::ColorArray1D: + case TextureType::Color2D: + ctx.Add("MOV.F {}.z,{};" + "MOV.F {}.w,{}.x;" + "TXB.F{} {},{},{},{}{};", + coord_vec, dref, coord_vec, bias_lc, sparse_mod, ret, coord_vec, texture, + type, offset_vec); + break; + case TextureType::ColorArray2D: + case TextureType::ColorCube: + ctx.Add("MOV.F {}.w,{};" + "TXB.F{} {},{},{},{},{}{};", + coord_vec, dref, sparse_mod, ret, coord_vec, bias_lc, texture, type, + offset_vec); + break; + case TextureType::ColorArrayCube: { + const ScopedRegister pair{ctx.reg_alloc}; + ctx.Add("MOV.F {}.x,{};" + "MOV.F {}.y,{}.x;" + "TXB.F{} {},{},{},{},{}{};", + pair.reg, dref, pair.reg, bias_lc, sparse_mod, ret, coord_vec, pair.reg, + texture, type, offset_vec); + break; + } + default: + throw NotImplementedException("Invalid type {}", info.type.Value()); + } + } + } else { + if (info.has_lod_clamp) { + if (info.type != TextureType::ColorArrayCube) { + const bool w_swizzle{info.type == TextureType::ColorArray2D || + info.type == TextureType::ColorCube}; + const char dref_swizzle{w_swizzle ? 'w' : 'z'}; + ctx.Add("MOV.F {}.{},{};" + "TEX.F.LODCLAMP{} {},{},{},{},{}{};", + coord_vec, dref_swizzle, dref, sparse_mod, ret, coord_vec, bias_lc, texture, + type, offset_vec); + } else { + const ScopedRegister pair{ctx.reg_alloc}; + ctx.Add("MOV.F {}.x,{};" + "MOV.F {}.y,{};" + "TEX.F.LODCLAMP{} {},{},{},{},{}{};", + pair.reg, dref, pair.reg, bias_lc, sparse_mod, ret, coord_vec, pair.reg, + texture, type, offset_vec); + } + } else { + if (info.type != TextureType::ColorArrayCube) { + const bool w_swizzle{info.type == TextureType::ColorArray2D || + info.type == TextureType::ColorCube}; + const char dref_swizzle{w_swizzle ? 'w' : 'z'}; + ctx.Add("MOV.F {}.{},{};" + "TEX.F{} {},{},{},{}{};", + coord_vec, dref_swizzle, dref, sparse_mod, ret, coord_vec, texture, type, + offset_vec); + } else { + const ScopedRegister pair{ctx.reg_alloc}; + ctx.Add("TEX.F{} {},{},{},{},{}{};", sparse_mod, ret, coord_vec, dref, texture, + type, offset_vec); + } + } + } + StoreSparse(ctx, sparse_inst); } -void EmitImageSampleDrefExplicitLod([[maybe_unused]] EmitContext& ctx, - [[maybe_unused]] IR::Inst& inst, - [[maybe_unused]] const IR::Value& index, - [[maybe_unused]] Register coord, [[maybe_unused]] Register dref, - [[maybe_unused]] Register lod_lc, - [[maybe_unused]] const IR::Value& offset) { - throw NotImplementedException("GLASM instruction"); +void EmitImageSampleDrefExplicitLod(EmitContext& ctx, IR::Inst& inst, const IR::Value& index, + const IR::Value& coord, ScalarF32 dref, ScalarF32 lod, + const IR::Value& offset) { + const auto info{inst.Flags()}; + const auto sparse_inst{inst.GetAssociatedPseudoOperation(IR::Opcode::GetSparseFromOp)}; + const std::string_view sparse_mod{sparse_inst ? ".SPARSE" : ""}; + const std::string_view type{ShadowTextureType(info)}; + const std::string texture{Texture(ctx, info, index)}; + const std::string offset_vec{Offset(ctx, offset)}; + const auto [coord_vec, coord_alloc]{Coord(ctx, coord)}; + const Register ret{ctx.reg_alloc.Define(inst)}; + switch (info.type) { + case TextureType::Color1D: + case TextureType::ColorArray1D: + case TextureType::Color2D: + ctx.Add("MOV.F {}.z,{};" + "MOV.F {}.w,{};" + "TXL.F{} {},{},{},{}{};", + coord_vec, dref, coord_vec, lod, sparse_mod, ret, coord_vec, texture, type, + offset_vec); + break; + case TextureType::ColorArray2D: + case TextureType::ColorCube: + ctx.Add("MOV.F {}.w,{};" + "TXL.F{} {},{},{},{},{}{};", + coord_vec, dref, sparse_mod, ret, coord_vec, lod, texture, type, offset_vec); + break; + case TextureType::ColorArrayCube: { + const ScopedRegister pair{ctx.reg_alloc}; + ctx.Add("MOV.F {}.x,{};" + "MOV.F {}.y,{};" + "TXL.F{} {},{},{},{},{}{};", + pair.reg, dref, pair.reg, lod, sparse_mod, ret, coord_vec, pair.reg, texture, type, + offset_vec); + break; + } + default: + throw NotImplementedException("Invalid type {}", info.type.Value()); + } + StoreSparse(ctx, sparse_inst); } void EmitImageGather([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst, diff --git a/src/shader_recompiler/backend/glasm/emit_glasm_instructions.h b/src/shader_recompiler/backend/glasm/emit_glasm_instructions.h index a128f9ac4..54e7fab3c 100644 --- a/src/shader_recompiler/backend/glasm/emit_glasm_instructions.h +++ b/src/shader_recompiler/backend/glasm/emit_glasm_instructions.h @@ -527,12 +527,12 @@ void EmitBoundImageWrite(EmitContext&); void EmitImageSampleImplicitLod(EmitContext& ctx, IR::Inst& inst, const IR::Value& index, const IR::Value& coord, Register bias_lc, const IR::Value& offset); void EmitImageSampleExplicitLod(EmitContext& ctx, IR::Inst& inst, const IR::Value& index, - Register coord, Register lod_lc, const IR::Value& offset); + const IR::Value& coord, ScalarF32 lod, const IR::Value& offset); void EmitImageSampleDrefImplicitLod(EmitContext& ctx, IR::Inst& inst, const IR::Value& index, - Register coord, Register dref, Register bias_lc, + const IR::Value& coord, ScalarF32 dref, Register bias_lc, const IR::Value& offset); void EmitImageSampleDrefExplicitLod(EmitContext& ctx, IR::Inst& inst, const IR::Value& index, - Register coord, Register dref, Register lod_lc, + const IR::Value& coord, ScalarF32 dref, ScalarF32 lod, const IR::Value& offset); void EmitImageGather(EmitContext& ctx, IR::Inst& inst, const IR::Value& index, Register coord, const IR::Value& offset, const IR::Value& offset2); diff --git a/src/shader_recompiler/backend/glasm/emit_glasm_memory.cpp b/src/shader_recompiler/backend/glasm/emit_glasm_memory.cpp index 0c6a6e1c8..dd307a9a3 100644 --- a/src/shader_recompiler/backend/glasm/emit_glasm_memory.cpp +++ b/src/shader_recompiler/backend/glasm/emit_glasm_memory.cpp @@ -38,7 +38,7 @@ void Store(EmitContext& ctx, const IR::Value& binding, ScalarU32 offset, ValueTy void Load(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, ScalarU32 offset, std::string_view size) { const Register ret{ctx.reg_alloc.Define(inst)}; - StorageOp(ctx, binding, offset, fmt::format("STORE.{} {},DC.x;", size, ret), + StorageOp(ctx, binding, offset, fmt::format("LOAD.{} {},DC.x;", size, ret), fmt::format("MOV.U {},{{0,0,0,0}};", ret)); } } // Anonymous namespace diff --git a/src/shader_recompiler/backend/spirv/emit_spirv_image.cpp b/src/shader_recompiler/backend/spirv/emit_spirv_image.cpp index 5832104df..99b883746 100644 --- a/src/shader_recompiler/backend/spirv/emit_spirv_image.cpp +++ b/src/shader_recompiler/backend/spirv/emit_spirv_image.cpp @@ -337,9 +337,9 @@ Id EmitImageSampleImplicitLod(EmitContext& ctx, IR::Inst* inst, const IR::Value& } Id EmitImageSampleExplicitLod(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords, - Id lod_lc, const IR::Value& offset) { + Id lod, const IR::Value& offset) { const auto info{inst->Flags()}; - const ImageOperands operands(ctx, false, true, info.has_lod_clamp != 0, lod_lc, offset); + const ImageOperands operands(ctx, false, true, false, lod, offset); return Emit(&EmitContext::OpImageSparseSampleExplicitLod, &EmitContext::OpImageSampleExplicitLod, ctx, inst, ctx.F32[4], Texture(ctx, info, index), coords, operands.Mask(), operands.Span()); @@ -356,9 +356,9 @@ Id EmitImageSampleDrefImplicitLod(EmitContext& ctx, IR::Inst* inst, const IR::Va } Id EmitImageSampleDrefExplicitLod(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, - Id coords, Id dref, Id lod_lc, const IR::Value& offset) { + Id coords, Id dref, Id lod, const IR::Value& offset) { const auto info{inst->Flags()}; - const ImageOperands operands(ctx, false, true, info.has_lod_clamp != 0, lod_lc, offset); + const ImageOperands operands(ctx, false, true, false, lod, offset); return Emit(&EmitContext::OpImageSparseSampleDrefExplicitLod, &EmitContext::OpImageSampleDrefExplicitLod, ctx, inst, ctx.F32[1], Texture(ctx, info, index), coords, dref, operands.Mask(), operands.Span()); diff --git a/src/shader_recompiler/backend/spirv/emit_spirv_instructions.h b/src/shader_recompiler/backend/spirv/emit_spirv_instructions.h index 0a2b31772..22260d2a9 100644 --- a/src/shader_recompiler/backend/spirv/emit_spirv_instructions.h +++ b/src/shader_recompiler/backend/spirv/emit_spirv_instructions.h @@ -491,11 +491,11 @@ Id EmitBoundImageWrite(EmitContext&); Id EmitImageSampleImplicitLod(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords, Id bias_lc, const IR::Value& offset); Id EmitImageSampleExplicitLod(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords, - Id lod_lc, const IR::Value& offset); + Id lod, const IR::Value& offset); Id EmitImageSampleDrefImplicitLod(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords, Id dref, Id bias_lc, const IR::Value& offset); Id EmitImageSampleDrefExplicitLod(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, - Id coords, Id dref, Id lod_lc, const IR::Value& offset); + Id coords, Id dref, Id lod, const IR::Value& offset); Id EmitImageGather(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords, const IR::Value& offset, const IR::Value& offset2); Id EmitImageGatherDref(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords, -- cgit v1.2.3