diff options
| author | 2021-07-25 11:39:04 -0700 | |
|---|---|---|
| committer | 2021-07-25 11:39:04 -0700 | |
| commit | 98b26b6e126d4775fdf3f773fe8a8ac808a8ff8f (patch) | |
| tree | 816faa96c2c4d291825063433331a8ea4b3d08f1 /src/shader_recompiler/backend/glsl/emit_glsl.cpp | |
| parent | Merge pull request #6699 from lat9nq/common-threads (diff) | |
| parent | shader: Support out of bound local memory reads and immediate writes (diff) | |
| download | yuzu-98b26b6e126d4775fdf3f773fe8a8ac808a8ff8f.tar.gz yuzu-98b26b6e126d4775fdf3f773fe8a8ac808a8ff8f.tar.xz yuzu-98b26b6e126d4775fdf3f773fe8a8ac808a8ff8f.zip | |
Merge pull request #6585 from ameerj/hades
Shader Decompiler Rewrite
Diffstat (limited to 'src/shader_recompiler/backend/glsl/emit_glsl.cpp')
| -rw-r--r-- | src/shader_recompiler/backend/glsl/emit_glsl.cpp | 252 |
1 files changed, 252 insertions, 0 deletions
diff --git a/src/shader_recompiler/backend/glsl/emit_glsl.cpp b/src/shader_recompiler/backend/glsl/emit_glsl.cpp new file mode 100644 index 000000000..8a430d573 --- /dev/null +++ b/src/shader_recompiler/backend/glsl/emit_glsl.cpp | |||
| @@ -0,0 +1,252 @@ | |||
| 1 | // Copyright 2021 yuzu Emulator Project | ||
| 2 | // Licensed under GPLv2 or any later version | ||
| 3 | // Refer to the license.txt file included. | ||
| 4 | |||
| 5 | #include <algorithm> | ||
| 6 | #include <string> | ||
| 7 | #include <tuple> | ||
| 8 | #include <type_traits> | ||
| 9 | |||
| 10 | #include "common/div_ceil.h" | ||
| 11 | #include "common/settings.h" | ||
| 12 | #include "shader_recompiler/backend/glsl/emit_context.h" | ||
| 13 | #include "shader_recompiler/backend/glsl/emit_glsl.h" | ||
| 14 | #include "shader_recompiler/backend/glsl/emit_glsl_instructions.h" | ||
| 15 | #include "shader_recompiler/frontend/ir/ir_emitter.h" | ||
| 16 | |||
| 17 | namespace Shader::Backend::GLSL { | ||
| 18 | namespace { | ||
| 19 | template <class Func> | ||
| 20 | struct FuncTraits {}; | ||
| 21 | |||
| 22 | template <class ReturnType_, class... Args> | ||
| 23 | struct FuncTraits<ReturnType_ (*)(Args...)> { | ||
| 24 | using ReturnType = ReturnType_; | ||
| 25 | |||
| 26 | static constexpr size_t NUM_ARGS = sizeof...(Args); | ||
| 27 | |||
| 28 | template <size_t I> | ||
| 29 | using ArgType = std::tuple_element_t<I, std::tuple<Args...>>; | ||
| 30 | }; | ||
| 31 | |||
| 32 | template <auto func, typename... Args> | ||
| 33 | void SetDefinition(EmitContext& ctx, IR::Inst* inst, Args... args) { | ||
| 34 | inst->SetDefinition<Id>(func(ctx, std::forward<Args>(args)...)); | ||
| 35 | } | ||
| 36 | |||
| 37 | template <typename ArgType> | ||
| 38 | auto Arg(EmitContext& ctx, const IR::Value& arg) { | ||
| 39 | if constexpr (std::is_same_v<ArgType, std::string_view>) { | ||
| 40 | return ctx.var_alloc.Consume(arg); | ||
| 41 | } else if constexpr (std::is_same_v<ArgType, const IR::Value&>) { | ||
| 42 | return arg; | ||
| 43 | } else if constexpr (std::is_same_v<ArgType, u32>) { | ||
| 44 | return arg.U32(); | ||
| 45 | } else if constexpr (std::is_same_v<ArgType, IR::Attribute>) { | ||
| 46 | return arg.Attribute(); | ||
| 47 | } else if constexpr (std::is_same_v<ArgType, IR::Patch>) { | ||
| 48 | return arg.Patch(); | ||
| 49 | } else if constexpr (std::is_same_v<ArgType, IR::Reg>) { | ||
| 50 | return arg.Reg(); | ||
| 51 | } | ||
| 52 | } | ||
| 53 | |||
| 54 | template <auto func, bool is_first_arg_inst, size_t... I> | ||
| 55 | void Invoke(EmitContext& ctx, IR::Inst* inst, std::index_sequence<I...>) { | ||
| 56 | using Traits = FuncTraits<decltype(func)>; | ||
| 57 | if constexpr (std::is_same_v<typename Traits::ReturnType, Id>) { | ||
| 58 | if constexpr (is_first_arg_inst) { | ||
| 59 | SetDefinition<func>( | ||
| 60 | ctx, inst, *inst, | ||
| 61 | Arg<typename Traits::template ArgType<I + 2>>(ctx, inst->Arg(I))...); | ||
| 62 | } else { | ||
| 63 | SetDefinition<func>( | ||
| 64 | ctx, inst, Arg<typename Traits::template ArgType<I + 1>>(ctx, inst->Arg(I))...); | ||
| 65 | } | ||
| 66 | } else { | ||
| 67 | if constexpr (is_first_arg_inst) { | ||
| 68 | func(ctx, *inst, Arg<typename Traits::template ArgType<I + 2>>(ctx, inst->Arg(I))...); | ||
| 69 | } else { | ||
| 70 | func(ctx, Arg<typename Traits::template ArgType<I + 1>>(ctx, inst->Arg(I))...); | ||
| 71 | } | ||
| 72 | } | ||
| 73 | } | ||
| 74 | |||
| 75 | template <auto func> | ||
| 76 | void Invoke(EmitContext& ctx, IR::Inst* inst) { | ||
| 77 | using Traits = FuncTraits<decltype(func)>; | ||
| 78 | static_assert(Traits::NUM_ARGS >= 1, "Insufficient arguments"); | ||
| 79 | if constexpr (Traits::NUM_ARGS == 1) { | ||
| 80 | Invoke<func, false>(ctx, inst, std::make_index_sequence<0>{}); | ||
| 81 | } else { | ||
| 82 | using FirstArgType = typename Traits::template ArgType<1>; | ||
| 83 | static constexpr bool is_first_arg_inst = std::is_same_v<FirstArgType, IR::Inst&>; | ||
| 84 | using Indices = std::make_index_sequence<Traits::NUM_ARGS - (is_first_arg_inst ? 2 : 1)>; | ||
| 85 | Invoke<func, is_first_arg_inst>(ctx, inst, Indices{}); | ||
| 86 | } | ||
| 87 | } | ||
| 88 | |||
| 89 | void EmitInst(EmitContext& ctx, IR::Inst* inst) { | ||
| 90 | switch (inst->GetOpcode()) { | ||
| 91 | #define OPCODE(name, result_type, ...) \ | ||
| 92 | case IR::Opcode::name: \ | ||
| 93 | return Invoke<&Emit##name>(ctx, inst); | ||
| 94 | #include "shader_recompiler/frontend/ir/opcodes.inc" | ||
| 95 | #undef OPCODE | ||
| 96 | } | ||
| 97 | throw LogicError("Invalid opcode {}", inst->GetOpcode()); | ||
| 98 | } | ||
| 99 | |||
| 100 | bool IsReference(IR::Inst& inst) { | ||
| 101 | return inst.GetOpcode() == IR::Opcode::Reference; | ||
| 102 | } | ||
| 103 | |||
| 104 | void PrecolorInst(IR::Inst& phi) { | ||
| 105 | // Insert phi moves before references to avoid overwritting other phis | ||
| 106 | const size_t num_args{phi.NumArgs()}; | ||
| 107 | for (size_t i = 0; i < num_args; ++i) { | ||
| 108 | IR::Block& phi_block{*phi.PhiBlock(i)}; | ||
| 109 | auto it{std::find_if_not(phi_block.rbegin(), phi_block.rend(), IsReference).base()}; | ||
| 110 | IR::IREmitter ir{phi_block, it}; | ||
| 111 | const IR::Value arg{phi.Arg(i)}; | ||
| 112 | if (arg.IsImmediate()) { | ||
| 113 | ir.PhiMove(phi, arg); | ||
| 114 | } else { | ||
| 115 | ir.PhiMove(phi, IR::Value{arg.InstRecursive()}); | ||
| 116 | } | ||
| 117 | } | ||
| 118 | for (size_t i = 0; i < num_args; ++i) { | ||
| 119 | IR::IREmitter{*phi.PhiBlock(i)}.Reference(IR::Value{&phi}); | ||
| 120 | } | ||
| 121 | } | ||
| 122 | |||
| 123 | void Precolor(const IR::Program& program) { | ||
| 124 | for (IR::Block* const block : program.blocks) { | ||
| 125 | for (IR::Inst& phi : block->Instructions()) { | ||
| 126 | if (!IR::IsPhi(phi)) { | ||
| 127 | break; | ||
| 128 | } | ||
| 129 | PrecolorInst(phi); | ||
| 130 | } | ||
| 131 | } | ||
| 132 | } | ||
| 133 | |||
| 134 | void EmitCode(EmitContext& ctx, const IR::Program& program) { | ||
| 135 | for (const IR::AbstractSyntaxNode& node : program.syntax_list) { | ||
| 136 | switch (node.type) { | ||
| 137 | case IR::AbstractSyntaxNode::Type::Block: | ||
| 138 | for (IR::Inst& inst : node.data.block->Instructions()) { | ||
| 139 | EmitInst(ctx, &inst); | ||
| 140 | } | ||
| 141 | break; | ||
| 142 | case IR::AbstractSyntaxNode::Type::If: | ||
| 143 | ctx.Add("if({}){{", ctx.var_alloc.Consume(node.data.if_node.cond)); | ||
| 144 | break; | ||
| 145 | case IR::AbstractSyntaxNode::Type::EndIf: | ||
| 146 | ctx.Add("}}"); | ||
| 147 | break; | ||
| 148 | case IR::AbstractSyntaxNode::Type::Break: | ||
| 149 | if (node.data.break_node.cond.IsImmediate()) { | ||
| 150 | if (node.data.break_node.cond.U1()) { | ||
| 151 | ctx.Add("break;"); | ||
| 152 | } | ||
| 153 | } else { | ||
| 154 | ctx.Add("if({}){{break;}}", ctx.var_alloc.Consume(node.data.break_node.cond)); | ||
| 155 | } | ||
| 156 | break; | ||
| 157 | case IR::AbstractSyntaxNode::Type::Return: | ||
| 158 | case IR::AbstractSyntaxNode::Type::Unreachable: | ||
| 159 | ctx.Add("return;"); | ||
| 160 | break; | ||
| 161 | case IR::AbstractSyntaxNode::Type::Loop: | ||
| 162 | ctx.Add("for(;;){{"); | ||
| 163 | break; | ||
| 164 | case IR::AbstractSyntaxNode::Type::Repeat: | ||
| 165 | if (Settings::values.disable_shader_loop_safety_checks) { | ||
| 166 | ctx.Add("if(!{}){{break;}}}}", ctx.var_alloc.Consume(node.data.repeat.cond)); | ||
| 167 | } else { | ||
| 168 | ctx.Add("if(--loop{}<0 || !{}){{break;}}}}", ctx.num_safety_loop_vars++, | ||
| 169 | ctx.var_alloc.Consume(node.data.repeat.cond)); | ||
| 170 | } | ||
| 171 | break; | ||
| 172 | default: | ||
| 173 | throw NotImplementedException("AbstractSyntaxNode Type {}", node.type); | ||
| 174 | } | ||
| 175 | } | ||
| 176 | } | ||
| 177 | |||
| 178 | std::string GlslVersionSpecifier(const EmitContext& ctx) { | ||
| 179 | if (ctx.uses_y_direction || ctx.info.stores.Legacy() || ctx.info.loads.Legacy()) { | ||
| 180 | return " compatibility"; | ||
| 181 | } | ||
| 182 | return ""; | ||
| 183 | } | ||
| 184 | |||
| 185 | bool IsPreciseType(GlslVarType type) { | ||
| 186 | switch (type) { | ||
| 187 | case GlslVarType::PrecF32: | ||
| 188 | case GlslVarType::PrecF64: | ||
| 189 | return true; | ||
| 190 | default: | ||
| 191 | return false; | ||
| 192 | } | ||
| 193 | } | ||
| 194 | |||
| 195 | void DefineVariables(const EmitContext& ctx, std::string& header) { | ||
| 196 | for (u32 i = 0; i < static_cast<u32>(GlslVarType::Void); ++i) { | ||
| 197 | const auto type{static_cast<GlslVarType>(i)}; | ||
| 198 | const auto& tracker{ctx.var_alloc.GetUseTracker(type)}; | ||
| 199 | const auto type_name{ctx.var_alloc.GetGlslType(type)}; | ||
| 200 | const bool has_precise_bug{ctx.stage == Stage::Fragment && ctx.profile.has_gl_precise_bug}; | ||
| 201 | const auto precise{!has_precise_bug && IsPreciseType(type) ? "precise " : ""}; | ||
| 202 | // Temps/return types that are never used are stored at index 0 | ||
| 203 | if (tracker.uses_temp) { | ||
| 204 | header += fmt::format("{}{} t{}={}(0);", precise, type_name, | ||
| 205 | ctx.var_alloc.Representation(0, type), type_name); | ||
| 206 | } | ||
| 207 | for (u32 index = 0; index < tracker.num_used; ++index) { | ||
| 208 | header += fmt::format("{}{} {}={}(0);", precise, type_name, | ||
| 209 | ctx.var_alloc.Representation(index, type), type_name); | ||
| 210 | } | ||
| 211 | } | ||
| 212 | for (u32 i = 0; i < ctx.num_safety_loop_vars; ++i) { | ||
| 213 | header += fmt::format("int loop{}=0x2000;", i); | ||
| 214 | } | ||
| 215 | } | ||
| 216 | } // Anonymous namespace | ||
| 217 | |||
| 218 | std::string EmitGLSL(const Profile& profile, const RuntimeInfo& runtime_info, IR::Program& program, | ||
| 219 | Bindings& bindings) { | ||
| 220 | EmitContext ctx{program, bindings, profile, runtime_info}; | ||
| 221 | Precolor(program); | ||
| 222 | EmitCode(ctx, program); | ||
| 223 | const std::string version{fmt::format("#version 450{}\n", GlslVersionSpecifier(ctx))}; | ||
| 224 | ctx.header.insert(0, version); | ||
| 225 | if (program.shared_memory_size > 0) { | ||
| 226 | const auto requested_size{program.shared_memory_size}; | ||
| 227 | const auto max_size{profile.gl_max_compute_smem_size}; | ||
| 228 | const bool needs_clamp{requested_size > max_size}; | ||
| 229 | if (needs_clamp) { | ||
| 230 | LOG_WARNING(Shader_GLSL, "Requested shared memory size ({}) exceeds device limit ({})", | ||
| 231 | requested_size, max_size); | ||
| 232 | } | ||
| 233 | const auto smem_size{needs_clamp ? max_size : requested_size}; | ||
| 234 | ctx.header += fmt::format("shared uint smem[{}];", Common::DivCeil(smem_size, 4U)); | ||
| 235 | } | ||
| 236 | ctx.header += "void main(){\n"; | ||
| 237 | if (program.local_memory_size > 0) { | ||
| 238 | ctx.header += fmt::format("uint lmem[{}];", Common::DivCeil(program.local_memory_size, 4U)); | ||
| 239 | } | ||
| 240 | DefineVariables(ctx, ctx.header); | ||
| 241 | if (ctx.uses_cc_carry) { | ||
| 242 | ctx.header += "uint carry;"; | ||
| 243 | } | ||
| 244 | if (program.info.uses_subgroup_shuffles) { | ||
| 245 | ctx.header += "bool shfl_in_bounds;"; | ||
| 246 | } | ||
| 247 | ctx.code.insert(0, ctx.header); | ||
| 248 | ctx.code += '}'; | ||
| 249 | return ctx.code; | ||
| 250 | } | ||
| 251 | |||
| 252 | } // namespace Shader::Backend::GLSL | ||