Merge pull request #2409 from ReinUsesLisp/half-floats

shader_ir/decode: Miscellaneous fixes to half-float decompilation

src/video_core/renderer_opengl/gl_shader_decompiler.cpp

@@ -119,14 +119,10 @@ std::string GetTopologyName(Tegra::Shader::OutputTopology topology) {
 
 /// Returns true if an object has to be treated as precise
 bool IsPrecise(Operation operand) {
-    const auto& meta = operand.GetMeta();
-
+    const auto& meta{operand.GetMeta()};
     if (const auto arithmetic = std::get_if<MetaArithmetic>(&meta)) {
         return arithmetic->precise;
     }
-    if (const auto half_arithmetic = std::get_if<MetaHalfArithmetic>(&meta)) {
-        return half_arithmetic->precise;
-    }
     return false;
 }
 
@@ -627,28 +623,7 @@ private:
     }
 
     std::string VisitOperand(Operation operation, std::size_t operand_index, Type type) {
-        std::string value = VisitOperand(operation, operand_index);
-        switch (type) {
-        case Type::HalfFloat: {
-            const auto half_meta = std::get_if<MetaHalfArithmetic>(&operation.GetMeta());
-            if (!half_meta) {
-                value = "toHalf2(" + value + ')';
-            }
-
-            switch (half_meta->types.at(operand_index)) {
-            case Tegra::Shader::HalfType::H0_H1:
-                return "toHalf2(" + value + ')';
-            case Tegra::Shader::HalfType::F32:
-                return "vec2(" + value + ')';
-            case Tegra::Shader::HalfType::H0_H0:
-                return "vec2(toHalf2(" + value + ")[0])";
-            case Tegra::Shader::HalfType::H1_H1:
-                return "vec2(toHalf2(" + value + ")[1])";
-            }
-        }
-        default:
-            return CastOperand(value, type);
-        }
+        return CastOperand(VisitOperand(operation, operand_index), type);
     }
 
     std::string CastOperand(const std::string& value, Type type) const {
@@ -662,9 +637,7 @@ private:
         case Type::Uint:
             return "ftou(" + value + ')';
         case Type::HalfFloat:
-            // Can't be handled as a stand-alone value
-            UNREACHABLE();
-            return value;
+            return "toHalf2(" + value + ')';
         }
         UNREACHABLE();
         return value;
@@ -1083,13 +1056,40 @@ private:
         return BitwiseCastResult(value, Type::HalfFloat);
     }
 
+    std::string HClamp(Operation operation) {
+        const std::string value = VisitOperand(operation, 0, Type::HalfFloat);
+        const std::string min = VisitOperand(operation, 1, Type::Float);
+        const std::string max = VisitOperand(operation, 2, Type::Float);
+        const std::string clamped = "clamp(" + value + ", vec2(" + min + "), vec2(" + max + "))";
+        return ApplyPrecise(operation, BitwiseCastResult(clamped, Type::HalfFloat));
+    }
+
+    std::string HUnpack(Operation operation) {
+        const std::string operand{VisitOperand(operation, 0, Type::HalfFloat)};
+        const auto value = [&]() -> std::string {
+            switch (std::get<Tegra::Shader::HalfType>(operation.GetMeta())) {
+            case Tegra::Shader::HalfType::H0_H1:
+                return operand;
+            case Tegra::Shader::HalfType::F32:
+                return "vec2(fromHalf2(" + operand + "))";
+            case Tegra::Shader::HalfType::H0_H0:
+                return "vec2(" + operand + "[0])";
+            case Tegra::Shader::HalfType::H1_H1:
+                return "vec2(" + operand + "[1])";
+            }
+            UNREACHABLE();
+            return "0";
+        }();
+        return "fromHalf2(" + value + ')';
+    }
+
     std::string HMergeF32(Operation operation) {
         return "float(toHalf2(" + Visit(operation[0]) + ")[0])";
     }
 
     std::string HMergeH0(Operation operation) {
-        return "fromHalf2(vec2(toHalf2(" + Visit(operation[0]) + ")[1], toHalf2(" +
-               Visit(operation[1]) + ")[0]))";
+        return "fromHalf2(vec2(toHalf2(" + Visit(operation[1]) + ")[0], toHalf2(" +
+               Visit(operation[0]) + ")[1]))";
     }
 
     std::string HMergeH1(Operation operation) {
@@ -1189,34 +1189,46 @@ private:
         return GenerateUnary(operation, "any", Type::Bool, Type::Bool2);
     }
 
+    template <bool with_nan>
+    std::string GenerateHalfComparison(Operation operation, std::string compare_op) {
+        std::string comparison{GenerateBinaryCall(operation, compare_op, Type::Bool2,
+                                                  Type::HalfFloat, Type::HalfFloat)};
+        if constexpr (!with_nan) {
+            return comparison;
+        }
+        return "halfFloatNanComparison(" + comparison + ", " +
+               VisitOperand(operation, 0, Type::HalfFloat) + ", " +
+               VisitOperand(operation, 1, Type::HalfFloat) + ')';
+    }
+
+    template <bool with_nan>
     std::string Logical2HLessThan(Operation operation) {
-        return GenerateBinaryCall(operation, "lessThan", Type::Bool2, Type::HalfFloat,
-                                  Type::HalfFloat);
+        return GenerateHalfComparison<with_nan>(operation, "lessThan");
     }
 
+    template <bool with_nan>
     std::string Logical2HEqual(Operation operation) {
-        return GenerateBinaryCall(operation, "equal", Type::Bool2, Type::HalfFloat,
-                                  Type::HalfFloat);
+        return GenerateHalfComparison<with_nan>(operation, "equal");
     }
 
+    template <bool with_nan>
     std::string Logical2HLessEqual(Operation operation) {
-        return GenerateBinaryCall(operation, "lessThanEqual", Type::Bool2, Type::HalfFloat,
-                                  Type::HalfFloat);
+        return GenerateHalfComparison<with_nan>(operation, "lessThanEqual");
     }
 
+    template <bool with_nan>
     std::string Logical2HGreaterThan(Operation operation) {
-        return GenerateBinaryCall(operation, "greaterThan", Type::Bool2, Type::HalfFloat,
-                                  Type::HalfFloat);
+        return GenerateHalfComparison<with_nan>(operation, "greaterThan");
     }
 
+    template <bool with_nan>
     std::string Logical2HNotEqual(Operation operation) {
-        return GenerateBinaryCall(operation, "notEqual", Type::Bool2, Type::HalfFloat,
-                                  Type::HalfFloat);
+        return GenerateHalfComparison<with_nan>(operation, "notEqual");
     }
 
+    template <bool with_nan>
     std::string Logical2HGreaterEqual(Operation operation) {
-        return GenerateBinaryCall(operation, "greaterThanEqual", Type::Bool2, Type::HalfFloat,
-                                  Type::HalfFloat);
+        return GenerateHalfComparison<with_nan>(operation, "greaterThanEqual");
     }
 
     std::string Texture(Operation operation) {
@@ -1505,6 +1517,8 @@ private:
         &GLSLDecompiler::Fma<Type::HalfFloat>,
         &GLSLDecompiler::Absolute<Type::HalfFloat>,
         &GLSLDecompiler::HNegate,
+        &GLSLDecompiler::HClamp,
+        &GLSLDecompiler::HUnpack,
         &GLSLDecompiler::HMergeF32,
         &GLSLDecompiler::HMergeH0,
         &GLSLDecompiler::HMergeH1,
@@ -1541,12 +1555,18 @@ private:
         &GLSLDecompiler::LogicalNotEqual<Type::Uint>,
         &GLSLDecompiler::LogicalGreaterEqual<Type::Uint>,
 
-        &GLSLDecompiler::Logical2HLessThan,
-        &GLSLDecompiler::Logical2HEqual,
-        &GLSLDecompiler::Logical2HLessEqual,
-        &GLSLDecompiler::Logical2HGreaterThan,
-        &GLSLDecompiler::Logical2HNotEqual,
-        &GLSLDecompiler::Logical2HGreaterEqual,
+        &GLSLDecompiler::Logical2HLessThan<false>,
+        &GLSLDecompiler::Logical2HEqual<false>,
+        &GLSLDecompiler::Logical2HLessEqual<false>,
+        &GLSLDecompiler::Logical2HGreaterThan<false>,
+        &GLSLDecompiler::Logical2HNotEqual<false>,
+        &GLSLDecompiler::Logical2HGreaterEqual<false>,
+        &GLSLDecompiler::Logical2HLessThan<true>,
+        &GLSLDecompiler::Logical2HEqual<true>,
+        &GLSLDecompiler::Logical2HLessEqual<true>,
+        &GLSLDecompiler::Logical2HGreaterThan<true>,
+        &GLSLDecompiler::Logical2HNotEqual<true>,
+        &GLSLDecompiler::Logical2HGreaterEqual<true>,
 
         &GLSLDecompiler::Texture,
         &GLSLDecompiler::TextureLod,
@@ -1647,6 +1667,12 @@ std::string GetCommonDeclarations() {
            "}\n\n"
            "vec2 toHalf2(float value) {\n"
            "    return unpackHalf2x16(ftou(value));\n"
+           "}\n\n"
+           "bvec2 halfFloatNanComparison(bvec2 comparison, vec2 pair1, vec2 pair2) {\n"
+           "    bvec2 is_nan1 = isnan(pair1);\n"
+           "    bvec2 is_nan2 = isnan(pair2);\n"
+           "    return bvec2(comparison.x || is_nan1.x || is_nan2.x, comparison.y || is_nan1.y || "
+           "is_nan2.y);\n"
            "}\n";
 }
 

src/video_core/renderer_vulkan/vk_shader_decompiler.cpp

@@ -76,14 +76,10 @@ constexpr u32 GetGenericAttributeLocation(Attribute::Index attribute) {
 
 /// Returns true if an object has to be treated as precise
 bool IsPrecise(Operation operand) {
-    const auto& meta = operand.GetMeta();
-
+    const auto& meta{operand.GetMeta()};
     if (std::holds_alternative<MetaArithmetic>(meta)) {
         return std::get<MetaArithmetic>(meta).precise;
     }
-    if (std::holds_alternative<MetaHalfArithmetic>(meta)) {
-        return std::get<MetaHalfArithmetic>(meta).precise;
-    }
     return false;
 }
 
@@ -746,6 +742,16 @@ private:
         return {};
     }
 
+    Id HClamp(Operation operation) {
+        UNIMPLEMENTED();
+        return {};
+    }
+
+    Id HUnpack(Operation operation) {
+        UNIMPLEMENTED();
+        return {};
+    }
+
     Id HMergeF32(Operation operation) {
         UNIMPLEMENTED();
         return {};
@@ -1218,6 +1224,8 @@ private:
         &SPIRVDecompiler::Ternary<&Module::OpFma, Type::HalfFloat>,
         &SPIRVDecompiler::Unary<&Module::OpFAbs, Type::HalfFloat>,
         &SPIRVDecompiler::HNegate,
+        &SPIRVDecompiler::HClamp,
+        &SPIRVDecompiler::HUnpack,
         &SPIRVDecompiler::HMergeF32,
         &SPIRVDecompiler::HMergeH0,
         &SPIRVDecompiler::HMergeH1,
@@ -1254,6 +1262,13 @@ private:
         &SPIRVDecompiler::Binary<&Module::OpINotEqual, Type::Bool, Type::Uint>,
         &SPIRVDecompiler::Binary<&Module::OpUGreaterThanEqual, Type::Bool, Type::Uint>,
 
+        &SPIRVDecompiler::Binary<&Module::OpFOrdLessThan, Type::Bool, Type::HalfFloat>,
+        &SPIRVDecompiler::Binary<&Module::OpFOrdEqual, Type::Bool, Type::HalfFloat>,
+        &SPIRVDecompiler::Binary<&Module::OpFOrdLessThanEqual, Type::Bool, Type::HalfFloat>,
+        &SPIRVDecompiler::Binary<&Module::OpFOrdGreaterThan, Type::Bool, Type::HalfFloat>,
+        &SPIRVDecompiler::Binary<&Module::OpFOrdNotEqual, Type::Bool, Type::HalfFloat>,
+        &SPIRVDecompiler::Binary<&Module::OpFOrdGreaterThanEqual, Type::Bool, Type::HalfFloat>,
+        // TODO(Rodrigo): Should these use the OpFUnord* variants?
         &SPIRVDecompiler::Binary<&Module::OpFOrdLessThan, Type::Bool, Type::HalfFloat>,
         &SPIRVDecompiler::Binary<&Module::OpFOrdEqual, Type::Bool, Type::HalfFloat>,
         &SPIRVDecompiler::Binary<&Module::OpFOrdLessThanEqual, Type::Bool, Type::HalfFloat>,

src/video_core/shader/decode/arithmetic_half.cpp

@@ -18,7 +18,9 @@ u32 ShaderIR::DecodeArithmeticHalf(NodeBlock& bb, u32 pc) {
 
     if (opcode->get().GetId() == OpCode::Id::HADD2_C ||
         opcode->get().GetId() == OpCode::Id::HADD2_R) {
-        UNIMPLEMENTED_IF(instr.alu_half.ftz != 0);
+        if (instr.alu_half.ftz != 0) {
+            LOG_WARNING(HW_GPU, "{} FTZ not implemented", opcode->get().GetName());
+        }
     }
     UNIMPLEMENTED_IF_MSG(instr.alu_half.saturate != 0, "Half float saturation not implemented");
 
@@ -27,9 +29,8 @@ u32 ShaderIR::DecodeArithmeticHalf(NodeBlock& bb, u32 pc) {
     const bool negate_b =
         opcode->get().GetId() != OpCode::Id::HMUL2_C && instr.alu_half.negate_b != 0;
 
-    const Node op_a = GetOperandAbsNegHalf(GetRegister(instr.gpr8), instr.alu_half.abs_a, negate_a);
-
-    // instr.alu_half.type_a
+    Node op_a = UnpackHalfFloat(GetRegister(instr.gpr8), instr.alu_half.type_a);
+    op_a = GetOperandAbsNegHalf(op_a, instr.alu_half.abs_a, negate_a);
 
     Node op_b = [&]() {
         switch (opcode->get().GetId()) {
@@ -44,17 +45,17 @@ u32 ShaderIR::DecodeArithmeticHalf(NodeBlock& bb, u32 pc) {
             return Immediate(0);
         }
     }();
+    op_b = UnpackHalfFloat(op_b, instr.alu_half.type_b);
     op_b = GetOperandAbsNegHalf(op_b, instr.alu_half.abs_b, negate_b);
 
     Node value = [&]() {
-        MetaHalfArithmetic meta{true, {instr.alu_half_imm.type_a, instr.alu_half.type_b}};
         switch (opcode->get().GetId()) {
         case OpCode::Id::HADD2_C:
         case OpCode::Id::HADD2_R:
-            return Operation(OperationCode::HAdd, meta, op_a, op_b);
+            return Operation(OperationCode::HAdd, PRECISE, op_a, op_b);
         case OpCode::Id::HMUL2_C:
         case OpCode::Id::HMUL2_R:
-            return Operation(OperationCode::HMul, meta, op_a, op_b);
+            return Operation(OperationCode::HMul, PRECISE, op_a, op_b);
         default:
             UNIMPLEMENTED_MSG("Unhandled half float instruction: {}", opcode->get().GetName());
             return Immediate(0);

src/video_core/shader/decode/arithmetic_half_immediate.cpp

@@ -17,34 +17,33 @@ u32 ShaderIR::DecodeArithmeticHalfImmediate(NodeBlock& bb, u32 pc) {
     const auto opcode = OpCode::Decode(instr);
 
     if (opcode->get().GetId() == OpCode::Id::HADD2_IMM) {
-        UNIMPLEMENTED_IF(instr.alu_half_imm.ftz != 0);
+        if (instr.alu_half_imm.ftz != 0) {
+            LOG_WARNING(HW_GPU, "{} FTZ not implemented", opcode->get().GetName());
+        }
     } else {
         UNIMPLEMENTED_IF(instr.alu_half_imm.precision != Tegra::Shader::HalfPrecision::None);
     }
-    UNIMPLEMENTED_IF_MSG(instr.alu_half_imm.saturate != 0,
-                         "Half float immediate saturation not implemented");
 
-    Node op_a = GetRegister(instr.gpr8);
+    Node op_a = UnpackHalfFloat(GetRegister(instr.gpr8), instr.alu_half_imm.type_a);
     op_a = GetOperandAbsNegHalf(op_a, instr.alu_half_imm.abs_a, instr.alu_half_imm.negate_a);
 
     const Node op_b = UnpackHalfImmediate(instr, true);
 
     Node value = [&]() {
-        MetaHalfArithmetic meta{true, {instr.alu_half_imm.type_a}};
         switch (opcode->get().GetId()) {
         case OpCode::Id::HADD2_IMM:
-            return Operation(OperationCode::HAdd, meta, op_a, op_b);
+            return Operation(OperationCode::HAdd, PRECISE, op_a, op_b);
         case OpCode::Id::HMUL2_IMM:
-            return Operation(OperationCode::HMul, meta, op_a, op_b);
+            return Operation(OperationCode::HMul, PRECISE, op_a, op_b);
         default:
             UNREACHABLE();
             return Immediate(0);
         }
     }();
-    value = HalfMerge(GetRegister(instr.gpr0), value, instr.alu_half_imm.merge);
 
+    value = GetSaturatedHalfFloat(value, instr.alu_half_imm.saturate);
+    value = HalfMerge(GetRegister(instr.gpr0), value, instr.alu_half_imm.merge);
     SetRegister(bb, instr.gpr0, value);
-
     return pc;
 }
 

src/video_core/shader/decode/half_set.cpp

@@ -18,11 +18,13 @@ u32 ShaderIR::DecodeHalfSet(NodeBlock& bb, u32 pc) {
     const Instruction instr = {program_code[pc]};
     const auto opcode = OpCode::Decode(instr);
 
-    UNIMPLEMENTED_IF(instr.hset2.ftz != 0);
+    if (instr.hset2.ftz != 0) {
+        LOG_WARNING(HW_GPU, "{} FTZ not implemented", opcode->get().GetName());
+    }
+
+    Node op_a = UnpackHalfFloat(GetRegister(instr.gpr8), instr.hset2.type_a);
+    op_a = GetOperandAbsNegHalf(op_a, instr.hset2.abs_a, instr.hset2.negate_a);
 
-    // instr.hset2.type_a
-    // instr.hset2.type_b
-    Node op_a = GetRegister(instr.gpr8);
     Node op_b = [&]() {
         switch (opcode->get().GetId()) {
         case OpCode::Id::HSET2_R:
@@ -32,14 +34,12 @@ u32 ShaderIR::DecodeHalfSet(NodeBlock& bb, u32 pc) {
             return Immediate(0);
         }
     }();
-
-    op_a = GetOperandAbsNegHalf(op_a, instr.hset2.abs_a, instr.hset2.negate_a);
+    op_b = UnpackHalfFloat(op_b, instr.hset2.type_b);
     op_b = GetOperandAbsNegHalf(op_b, instr.hset2.abs_b, instr.hset2.negate_b);
 
     const Node second_pred = GetPredicate(instr.hset2.pred39, instr.hset2.neg_pred);
 
-    MetaHalfArithmetic meta{false, {instr.hset2.type_a, instr.hset2.type_b}};
-    const Node comparison_pair = GetPredicateComparisonHalf(instr.hset2.cond, meta, op_a, op_b);
+    const Node comparison_pair = GetPredicateComparisonHalf(instr.hset2.cond, op_a, op_b);
 
     const OperationCode combiner = GetPredicateCombiner(instr.hset2.op);
 

src/video_core/shader/decode/half_set_predicate.cpp

@@ -19,10 +19,10 @@ u32 ShaderIR::DecodeHalfSetPredicate(NodeBlock& bb, u32 pc) {
 
     UNIMPLEMENTED_IF(instr.hsetp2.ftz != 0);
 
-    Node op_a = GetRegister(instr.gpr8);
+    Node op_a = UnpackHalfFloat(GetRegister(instr.gpr8), instr.hsetp2.type_a);
     op_a = GetOperandAbsNegHalf(op_a, instr.hsetp2.abs_a, instr.hsetp2.negate_a);
 
-    const Node op_b = [&]() {
+    Node op_b = [&]() {
         switch (opcode->get().GetId()) {
         case OpCode::Id::HSETP2_R:
             return GetOperandAbsNegHalf(GetRegister(instr.gpr20), instr.hsetp2.abs_a,
@@ -32,6 +32,7 @@ u32 ShaderIR::DecodeHalfSetPredicate(NodeBlock& bb, u32 pc) {
             return Immediate(0);
         }
     }();
+    op_b = UnpackHalfFloat(op_b, instr.hsetp2.type_b);
 
     // We can't use the constant predicate as destination.
     ASSERT(instr.hsetp2.pred3 != static_cast<u64>(Pred::UnusedIndex));
@@ -42,8 +43,7 @@ u32 ShaderIR::DecodeHalfSetPredicate(NodeBlock& bb, u32 pc) {
     const OperationCode pair_combiner =
         instr.hsetp2.h_and ? OperationCode::LogicalAll2 : OperationCode::LogicalAny2;
 
-    MetaHalfArithmetic meta = {false, {instr.hsetp2.type_a, instr.hsetp2.type_b}};
-    const Node comparison = GetPredicateComparisonHalf(instr.hsetp2.cond, meta, op_a, op_b);
+    const Node comparison = GetPredicateComparisonHalf(instr.hsetp2.cond, op_a, op_b);
     const Node first_pred = Operation(pair_combiner, comparison);
 
     // Set the primary predicate to the result of Predicate OP SecondPredicate

src/video_core/shader/decode/hfma2.cpp

@@ -27,10 +27,6 @@ u32 ShaderIR::DecodeHfma2(NodeBlock& bb, u32 pc) {
     }
 
     constexpr auto identity = HalfType::H0_H1;
-
-    const HalfType type_a = instr.hfma2.type_a;
-    const Node op_a = GetRegister(instr.gpr8);
-
     bool neg_b{}, neg_c{};
     auto [saturate, type_b, op_b, type_c,
           op_c] = [&]() -> std::tuple<bool, HalfType, Node, HalfType, Node> {
@@ -62,11 +58,11 @@ u32 ShaderIR::DecodeHfma2(NodeBlock& bb, u32 pc) {
     }();
     UNIMPLEMENTED_IF_MSG(saturate, "HFMA2 saturation is not implemented");
 
-    op_b = GetOperandAbsNegHalf(op_b, false, neg_b);
-    op_c = GetOperandAbsNegHalf(op_c, false, neg_c);
+    const Node op_a = UnpackHalfFloat(GetRegister(instr.gpr8), instr.hfma2.type_a);
+    op_b = GetOperandAbsNegHalf(UnpackHalfFloat(op_b, type_b), false, neg_b);
+    op_c = GetOperandAbsNegHalf(UnpackHalfFloat(op_c, type_c), false, neg_c);
 
-    MetaHalfArithmetic meta{true, {type_a, type_b, type_c}};
-    Node value = Operation(OperationCode::HFma, meta, op_a, op_b, op_c);
+    Node value = Operation(OperationCode::HFma, PRECISE, op_a, op_b, op_c);
     value = HalfMerge(GetRegister(instr.gpr0), value, instr.hfma2.merge);
 
     SetRegister(bb, instr.gpr0, value);

src/video_core/shader/shader_ir.cpp

@@ -189,7 +189,11 @@ Node ShaderIR::UnpackHalfImmediate(Instruction instr, bool has_negation) {
     const Node first_negate = GetPredicate(instr.half_imm.first_negate != 0);
     const Node second_negate = GetPredicate(instr.half_imm.second_negate != 0);
 
-    return Operation(OperationCode::HNegate, HALF_NO_PRECISE, value, first_negate, second_negate);
+    return Operation(OperationCode::HNegate, NO_PRECISE, value, first_negate, second_negate);
+}
+
+Node ShaderIR::UnpackHalfFloat(Node value, Tegra::Shader::HalfType type) {
+    return Operation(OperationCode::HUnpack, type, value);
 }
 
 Node ShaderIR::HalfMerge(Node dest, Node src, Tegra::Shader::HalfMerge merge) {
@@ -209,17 +213,26 @@ Node ShaderIR::HalfMerge(Node dest, Node src, Tegra::Shader::HalfMerge merge) {
 
 Node ShaderIR::GetOperandAbsNegHalf(Node value, bool absolute, bool negate) {
     if (absolute) {
-        value = Operation(OperationCode::HAbsolute, HALF_NO_PRECISE, value);
+        value = Operation(OperationCode::HAbsolute, NO_PRECISE, value);
     }
     if (negate) {
-        value = Operation(OperationCode::HNegate, HALF_NO_PRECISE, value, GetPredicate(true),
+        value = Operation(OperationCode::HNegate, NO_PRECISE, value, GetPredicate(true),
                           GetPredicate(true));
     }
     return value;
 }
 
+Node ShaderIR::GetSaturatedHalfFloat(Node value, bool saturate) {
+    if (!saturate) {
+        return value;
+    }
+    const Node positive_zero = Immediate(std::copysignf(0, 1));
+    const Node positive_one = Immediate(1.0f);
+    return Operation(OperationCode::HClamp, NO_PRECISE, value, positive_zero, positive_one);
+}
+
 Node ShaderIR::GetPredicateComparisonFloat(PredCondition condition, Node op_a, Node op_b) {
-    static const std::unordered_map<PredCondition, OperationCode> PredicateComparisonTable = {
+    const std::unordered_map<PredCondition, OperationCode> PredicateComparisonTable = {
         {PredCondition::LessThan, OperationCode::LogicalFLessThan},
         {PredCondition::Equal, OperationCode::LogicalFEqual},
         {PredCondition::LessEqual, OperationCode::LogicalFLessEqual},
@@ -255,7 +268,7 @@ Node ShaderIR::GetPredicateComparisonFloat(PredCondition condition, Node op_a, N
 
 Node ShaderIR::GetPredicateComparisonInteger(PredCondition condition, bool is_signed, Node op_a,
                                              Node op_b) {
-    static const std::unordered_map<PredCondition, OperationCode> PredicateComparisonTable = {
+    const std::unordered_map<PredCondition, OperationCode> PredicateComparisonTable = {
         {PredCondition::LessThan, OperationCode::LogicalILessThan},
         {PredCondition::Equal, OperationCode::LogicalIEqual},
         {PredCondition::LessEqual, OperationCode::LogicalILessEqual},
@@ -283,40 +296,32 @@ Node ShaderIR::GetPredicateComparisonInteger(PredCondition condition, bool is_si
     return predicate;
 }
 
-Node ShaderIR::GetPredicateComparisonHalf(Tegra::Shader::PredCondition condition,
-                                          const MetaHalfArithmetic& meta, Node op_a, Node op_b) {
-
-    UNIMPLEMENTED_IF_MSG(condition == PredCondition::LessThanWithNan ||
-                             condition == PredCondition::NotEqualWithNan ||
-                             condition == PredCondition::LessEqualWithNan ||
-                             condition == PredCondition::GreaterThanWithNan ||
-                             condition == PredCondition::GreaterEqualWithNan,
-                         "Unimplemented NaN comparison for half floats");
-
-    static const std::unordered_map<PredCondition, OperationCode> PredicateComparisonTable = {
+Node ShaderIR::GetPredicateComparisonHalf(Tegra::Shader::PredCondition condition, Node op_a,
+                                          Node op_b) {
+    const std::unordered_map<PredCondition, OperationCode> PredicateComparisonTable = {
         {PredCondition::LessThan, OperationCode::Logical2HLessThan},
         {PredCondition::Equal, OperationCode::Logical2HEqual},
         {PredCondition::LessEqual, OperationCode::Logical2HLessEqual},
         {PredCondition::GreaterThan, OperationCode::Logical2HGreaterThan},
         {PredCondition::NotEqual, OperationCode::Logical2HNotEqual},
         {PredCondition::GreaterEqual, OperationCode::Logical2HGreaterEqual},
-        {PredCondition::LessThanWithNan, OperationCode::Logical2HLessThan},
-        {PredCondition::NotEqualWithNan, OperationCode::Logical2HNotEqual},
-        {PredCondition::LessEqualWithNan, OperationCode::Logical2HLessEqual},
-        {PredCondition::GreaterThanWithNan, OperationCode::Logical2HGreaterThan},
-        {PredCondition::GreaterEqualWithNan, OperationCode::Logical2HGreaterEqual}};
+        {PredCondition::LessThanWithNan, OperationCode::Logical2HLessThanWithNan},
+        {PredCondition::NotEqualWithNan, OperationCode::Logical2HNotEqualWithNan},
+        {PredCondition::LessEqualWithNan, OperationCode::Logical2HLessEqualWithNan},
+        {PredCondition::GreaterThanWithNan, OperationCode::Logical2HGreaterThanWithNan},
+        {PredCondition::GreaterEqualWithNan, OperationCode::Logical2HGreaterEqualWithNan}};
 
     const auto comparison{PredicateComparisonTable.find(condition)};
     UNIMPLEMENTED_IF_MSG(comparison == PredicateComparisonTable.end(),
                          "Unknown predicate comparison operation");
 
-    const Node predicate = Operation(comparison->second, meta, op_a, op_b);
+    const Node predicate = Operation(comparison->second, NO_PRECISE, op_a, op_b);
 
     return predicate;
 }
 
 OperationCode ShaderIR::GetPredicateCombiner(PredOperation operation) {
-    static const std::unordered_map<PredOperation, OperationCode> PredicateOperationTable = {
+    const std::unordered_map<PredOperation, OperationCode> PredicateOperationTable = {
         {PredOperation::And, OperationCode::LogicalAnd},
         {PredOperation::Or, OperationCode::LogicalOr},
         {PredOperation::Xor, OperationCode::LogicalXor},

src/video_core/shader/shader_ir.h

@@ -109,11 +109,13 @@ enum class OperationCode {
     UBitfieldExtract, /// (MetaArithmetic, uint value, int offset, int offset) -> uint
     UBitCount,        /// (MetaArithmetic, uint) -> uint
 
-    HAdd,      /// (MetaHalfArithmetic, f16vec2 a, f16vec2 b) -> f16vec2
-    HMul,      /// (MetaHalfArithmetic, f16vec2 a, f16vec2 b) -> f16vec2
-    HFma,      /// (MetaHalfArithmetic, f16vec2 a, f16vec2 b, f16vec2 c) -> f16vec2
+    HAdd,      /// (MetaArithmetic, f16vec2 a, f16vec2 b) -> f16vec2
+    HMul,      /// (MetaArithmetic, f16vec2 a, f16vec2 b) -> f16vec2
+    HFma,      /// (MetaArithmetic, f16vec2 a, f16vec2 b, f16vec2 c) -> f16vec2
     HAbsolute, /// (f16vec2 a) -> f16vec2
     HNegate,   /// (f16vec2 a, bool first, bool second) -> f16vec2
+    HClamp,    /// (f16vec2 src, float min, float max) -> f16vec2
+    HUnpack,   /// (Tegra::Shader::HalfType, T value) -> f16vec2
     HMergeF32, /// (f16vec2 src) -> float
     HMergeH0,  /// (f16vec2 dest, f16vec2 src) -> f16vec2
     HMergeH1,  /// (f16vec2 dest, f16vec2 src) -> f16vec2
@@ -150,12 +152,18 @@ enum class OperationCode {
     LogicalUNotEqual,     /// (uint a, uint b) -> bool
     LogicalUGreaterEqual, /// (uint a, uint b) -> bool
 
-    Logical2HLessThan,     /// (MetaHalfArithmetic, f16vec2 a, f16vec2) -> bool2
-    Logical2HEqual,        /// (MetaHalfArithmetic, f16vec2 a, f16vec2) -> bool2
-    Logical2HLessEqual,    /// (MetaHalfArithmetic, f16vec2 a, f16vec2) -> bool2
-    Logical2HGreaterThan,  /// (MetaHalfArithmetic, f16vec2 a, f16vec2) -> bool2
-    Logical2HNotEqual,     /// (MetaHalfArithmetic, f16vec2 a, f16vec2) -> bool2
-    Logical2HGreaterEqual, /// (MetaHalfArithmetic, f16vec2 a, f16vec2) -> bool2
+    Logical2HLessThan,            /// (MetaHalfArithmetic, f16vec2 a, f16vec2) -> bool2
+    Logical2HEqual,               /// (MetaHalfArithmetic, f16vec2 a, f16vec2) -> bool2
+    Logical2HLessEqual,           /// (MetaHalfArithmetic, f16vec2 a, f16vec2) -> bool2
+    Logical2HGreaterThan,         /// (MetaHalfArithmetic, f16vec2 a, f16vec2) -> bool2
+    Logical2HNotEqual,            /// (MetaHalfArithmetic, f16vec2 a, f16vec2) -> bool2
+    Logical2HGreaterEqual,        /// (MetaHalfArithmetic, f16vec2 a, f16vec2) -> bool2
+    Logical2HLessThanWithNan,     /// (MetaHalfArithmetic, f16vec2 a, f16vec2) -> bool2
+    Logical2HEqualWithNan,        /// (MetaHalfArithmetic, f16vec2 a, f16vec2) -> bool2
+    Logical2HLessEqualWithNan,    /// (MetaHalfArithmetic, f16vec2 a, f16vec2) -> bool2
+    Logical2HGreaterThanWithNan,  /// (MetaHalfArithmetic, f16vec2 a, f16vec2) -> bool2
+    Logical2HNotEqualWithNan,     /// (MetaHalfArithmetic, f16vec2 a, f16vec2) -> bool2
+    Logical2HGreaterEqualWithNan, /// (MetaHalfArithmetic, f16vec2 a, f16vec2) -> bool2
 
     Texture,                /// (MetaTexture, float[N] coords) -> float4
     TextureLod,             /// (MetaTexture, float[N] coords) -> float4
@@ -308,13 +316,6 @@ struct MetaArithmetic {
     bool precise{};
 };
 
-struct MetaHalfArithmetic {
-    bool precise{};
-    std::array<Tegra::Shader::HalfType, 3> types = {Tegra::Shader::HalfType::H0_H1,
-                                                    Tegra::Shader::HalfType::H0_H1,
-                                                    Tegra::Shader::HalfType::H0_H1};
-};
-
 struct MetaTexture {
     const Sampler& sampler;
     Node array{};
@@ -326,11 +327,10 @@ struct MetaTexture {
     u32 element{};
 };
 
-constexpr MetaArithmetic PRECISE = {true};
-constexpr MetaArithmetic NO_PRECISE = {false};
-constexpr MetaHalfArithmetic HALF_NO_PRECISE = {false};
+inline constexpr MetaArithmetic PRECISE = {true};
+inline constexpr MetaArithmetic NO_PRECISE = {false};
 
-using Meta = std::variant<MetaArithmetic, MetaHalfArithmetic, MetaTexture>;
+using Meta = std::variant<MetaArithmetic, MetaTexture, Tegra::Shader::HalfType>;
 
 /// Holds any kind of operation that can be done in the IR
 class OperationNode final {
@@ -734,10 +734,14 @@ private:
 
     /// Unpacks a half immediate from an instruction
     Node UnpackHalfImmediate(Tegra::Shader::Instruction instr, bool has_negation);
+    /// Unpacks a binary value into a half float pair with a type format
+    Node UnpackHalfFloat(Node value, Tegra::Shader::HalfType type);
     /// Merges a half pair into another value
     Node HalfMerge(Node dest, Node src, Tegra::Shader::HalfMerge merge);
     /// Conditionally absolute/negated half float pair. Absolute is applied first
     Node GetOperandAbsNegHalf(Node value, bool absolute, bool negate);
+    /// Conditionally saturates a half float pair
+    Node GetSaturatedHalfFloat(Node value, bool saturate = true);
 
     /// Returns a predicate comparing two floats
     Node GetPredicateComparisonFloat(Tegra::Shader::PredCondition condition, Node op_a, Node op_b);
@@ -745,8 +749,7 @@ private:
     Node GetPredicateComparisonInteger(Tegra::Shader::PredCondition condition, bool is_signed,
                                        Node op_a, Node op_b);
     /// Returns a predicate comparing two half floats. meta consumes how both pairs will be compared
-    Node GetPredicateComparisonHalf(Tegra::Shader::PredCondition condition,
-                                    const MetaHalfArithmetic& meta, Node op_a, Node op_b);
+    Node GetPredicateComparisonHalf(Tegra::Shader::PredCondition condition, Node op_a, Node op_b);
 
     /// Returns a predicate combiner operation
     OperationCode GetPredicateCombiner(Tegra::Shader::PredOperation operation);