From 4e57f9d5cfc32b37fe7b6a1563ca2101ec59887c Mon Sep 17 00:00:00 2001 From: ReinUsesLisp Date: Sat, 9 May 2020 04:55:15 -0300 Subject: [PATCH 1/2] shader_ir: Separate float-point comparisons in ordered and unordered This allows us to use native SPIR-V instructions without having to manually check for NAN. --- src/video_core/engines/shader_bytecode.h | 28 +++-- .../renderer_opengl/gl_shader_decompiler.cpp | 97 ++++++++------- .../renderer_vulkan/vk_shader_decompiler.cpp | 27 ++++- src/video_core/shader/decode.cpp | 2 +- src/video_core/shader/decode/xmad.cpp | 12 +- src/video_core/shader/node.h | 21 ++-- src/video_core/shader/shader_ir.cpp | 111 ++++++++---------- 7 files changed, 163 insertions(+), 135 deletions(-) diff --git a/src/video_core/engines/shader_bytecode.h b/src/video_core/engines/shader_bytecode.h index 8dae754d49..e7cb87589a 100644 --- a/src/video_core/engines/shader_bytecode.h +++ b/src/video_core/engines/shader_bytecode.h @@ -168,18 +168,22 @@ enum class Pred : u64 { }; enum class PredCondition : u64 { - LessThan = 1, - Equal = 2, - LessEqual = 3, - GreaterThan = 4, - NotEqual = 5, - GreaterEqual = 6, - LessThanWithNan = 9, - LessEqualWithNan = 11, - GreaterThanWithNan = 12, - NotEqualWithNan = 13, - GreaterEqualWithNan = 14, - // TODO(Subv): Other condition types + F = 0, // Always false + LT = 1, // Ordered less than + EQ = 2, // Ordered equal + LE = 3, // Ordered less than or equal + GT = 4, // Ordered greater than + NE = 5, // Ordered not equal + GE = 6, // Ordered greater than or equal + NUM = 7, // Ordered + NAN_ = 8, // Unordered + LTU = 9, // Unordered less than + EQU = 10, // Unordered equal + LEU = 11, // Unordered less than or equal + GTU = 12, // Unordered greater than + NEU = 13, // Unordered not equal + GEU = 14, // Unordered greater than or equal + T = 15, // Always true }; enum class PredOperation : u64 { diff --git a/src/video_core/renderer_opengl/gl_shader_decompiler.cpp b/src/video_core/renderer_opengl/gl_shader_decompiler.cpp index 99fd4ae2c0..d071abd84c 100644 --- a/src/video_core/renderer_opengl/gl_shader_decompiler.cpp +++ b/src/video_core/renderer_opengl/gl_shader_decompiler.cpp @@ -1840,34 +1840,31 @@ private: Type::HalfFloat}; } - template - Expression LogicalLessThan(Operation operation) { - return GenerateBinaryInfix(operation, "<", Type::Bool, type, type); + template + Expression Comparison(Operation operation) { + static_assert(!unordered || type == Type::Float); + + const Expression expr = GenerateBinaryInfix(operation, op, Type::Bool, type, type); + if constexpr (!unordered) { + return expr; + } + // Unordered comparisons are always true for NaN operands. + return {fmt::format("({} || isnan({}) || isnan({}))", expr.AsBool(), + VisitOperand(operation, 0).AsFloat(), + VisitOperand(operation, 1).AsFloat()), + Type::Bool}; } - template - Expression LogicalEqual(Operation operation) { - return GenerateBinaryInfix(operation, "==", Type::Bool, type, type); + Expression FOrdered(Operation operation) { + return {fmt::format("(!isnan({}) && !isnan({}))", VisitOperand(operation, 0).AsFloat(), + VisitOperand(operation, 1).AsFloat()), + Type::Bool}; } - template - Expression LogicalLessEqual(Operation operation) { - return GenerateBinaryInfix(operation, "<=", Type::Bool, type, type); - } - - template - Expression LogicalGreaterThan(Operation operation) { - return GenerateBinaryInfix(operation, ">", Type::Bool, type, type); - } - - template - Expression LogicalNotEqual(Operation operation) { - return GenerateBinaryInfix(operation, "!=", Type::Bool, type, type); - } - - template - Expression LogicalGreaterEqual(Operation operation) { - return GenerateBinaryInfix(operation, ">=", Type::Bool, type, type); + Expression FUnordered(Operation operation) { + return {fmt::format("(isnan({}) || isnan({}))", VisitOperand(operation, 0).AsFloat(), + VisitOperand(operation, 1).AsFloat()), + Type::Bool}; } Expression LogicalAddCarry(Operation operation) { @@ -2324,6 +2321,13 @@ private: Func() = delete; ~Func() = delete; + static constexpr std::string_view LessThan = "<"; + static constexpr std::string_view Equal = "=="; + static constexpr std::string_view LessEqual = "<="; + static constexpr std::string_view GreaterThan = ">"; + static constexpr std::string_view NotEqual = "!="; + static constexpr std::string_view GreaterEqual = ">="; + static constexpr std::string_view Add = "Add"; static constexpr std::string_view Min = "Min"; static constexpr std::string_view Max = "Max"; @@ -2425,27 +2429,34 @@ private: &GLSLDecompiler::LogicalPick2, &GLSLDecompiler::LogicalAnd2, - &GLSLDecompiler::LogicalLessThan, - &GLSLDecompiler::LogicalEqual, - &GLSLDecompiler::LogicalLessEqual, - &GLSLDecompiler::LogicalGreaterThan, - &GLSLDecompiler::LogicalNotEqual, - &GLSLDecompiler::LogicalGreaterEqual, - &GLSLDecompiler::LogicalFIsNan, + &GLSLDecompiler::Comparison, + &GLSLDecompiler::Comparison, + &GLSLDecompiler::Comparison, + &GLSLDecompiler::Comparison, + &GLSLDecompiler::Comparison, + &GLSLDecompiler::Comparison, + &GLSLDecompiler::FOrdered, + &GLSLDecompiler::FUnordered, + &GLSLDecompiler::Comparison, + &GLSLDecompiler::Comparison, + &GLSLDecompiler::Comparison, + &GLSLDecompiler::Comparison, + &GLSLDecompiler::Comparison, + &GLSLDecompiler::Comparison, - &GLSLDecompiler::LogicalLessThan, - &GLSLDecompiler::LogicalEqual, - &GLSLDecompiler::LogicalLessEqual, - &GLSLDecompiler::LogicalGreaterThan, - &GLSLDecompiler::LogicalNotEqual, - &GLSLDecompiler::LogicalGreaterEqual, + &GLSLDecompiler::Comparison, + &GLSLDecompiler::Comparison, + &GLSLDecompiler::Comparison, + &GLSLDecompiler::Comparison, + &GLSLDecompiler::Comparison, + &GLSLDecompiler::Comparison, - &GLSLDecompiler::LogicalLessThan, - &GLSLDecompiler::LogicalEqual, - &GLSLDecompiler::LogicalLessEqual, - &GLSLDecompiler::LogicalGreaterThan, - &GLSLDecompiler::LogicalNotEqual, - &GLSLDecompiler::LogicalGreaterEqual, + &GLSLDecompiler::Comparison, + &GLSLDecompiler::Comparison, + &GLSLDecompiler::Comparison, + &GLSLDecompiler::Comparison, + &GLSLDecompiler::Comparison, + &GLSLDecompiler::Comparison, &GLSLDecompiler::LogicalAddCarry, diff --git a/src/video_core/renderer_vulkan/vk_shader_decompiler.cpp b/src/video_core/renderer_vulkan/vk_shader_decompiler.cpp index 18678968cb..167e20e919 100644 --- a/src/video_core/renderer_vulkan/vk_shader_decompiler.cpp +++ b/src/video_core/renderer_vulkan/vk_shader_decompiler.cpp @@ -1618,6 +1618,24 @@ private: return {}; } + Expression LogicalFOrdered(Operation operation) { + // Emulate SPIR-V's OpOrdered + const Id op_a = AsFloat(Visit(operation[0])); + const Id op_b = AsFloat(Visit(operation[1])); + const Id is_num_a = OpFOrdEqual(t_bool, op_a, op_a); + const Id is_num_b = OpFOrdEqual(t_bool, op_b, op_b); + return {OpLogicalAnd(t_bool, is_num_a, is_num_b), Type::Bool}; + } + + Expression LogicalFUnordered(Operation operation) { + // Emulate SPIR-V's OpUnordered + const Id op_a = AsFloat(Visit(operation[0])); + const Id op_b = AsFloat(Visit(operation[1])); + const Id is_nan_a = OpIsNan(t_bool, op_a); + const Id is_nan_b = OpIsNan(t_bool, op_b); + return {OpLogicalOr(t_bool, is_nan_a, is_nan_b), Type::Bool}; + } + Id GetTextureSampler(Operation operation) { const auto& meta = std::get(operation.GetMeta()); ASSERT(!meta.sampler.is_buffer); @@ -2511,7 +2529,14 @@ private: &SPIRVDecompiler::Binary<&Module::OpFOrdGreaterThan, Type::Bool, Type::Float>, &SPIRVDecompiler::Binary<&Module::OpFOrdNotEqual, Type::Bool, Type::Float>, &SPIRVDecompiler::Binary<&Module::OpFOrdGreaterThanEqual, Type::Bool, Type::Float>, - &SPIRVDecompiler::Unary<&Module::OpIsNan, Type::Bool, Type::Float>, + &SPIRVDecompiler::LogicalFOrdered, + &SPIRVDecompiler::LogicalFUnordered, + &SPIRVDecompiler::Binary<&Module::OpFUnordLessThan, Type::Bool, Type::Float>, + &SPIRVDecompiler::Binary<&Module::OpFUnordEqual, Type::Bool, Type::Float>, + &SPIRVDecompiler::Binary<&Module::OpFUnordLessThanEqual, Type::Bool, Type::Float>, + &SPIRVDecompiler::Binary<&Module::OpFUnordGreaterThan, Type::Bool, Type::Float>, + &SPIRVDecompiler::Binary<&Module::OpFUnordNotEqual, Type::Bool, Type::Float>, + &SPIRVDecompiler::Binary<&Module::OpFUnordGreaterThanEqual, Type::Bool, Type::Float>, &SPIRVDecompiler::Binary<&Module::OpSLessThan, Type::Bool, Type::Int>, &SPIRVDecompiler::Binary<&Module::OpIEqual, Type::Bool, Type::Int>, diff --git a/src/video_core/shader/decode.cpp b/src/video_core/shader/decode.cpp index a75a5cc638..eeac328a6a 100644 --- a/src/video_core/shader/decode.cpp +++ b/src/video_core/shader/decode.cpp @@ -255,7 +255,7 @@ void ShaderIR::InsertControlFlow(NodeBlock& bb, const ShaderBlock& block) { Node n = Operation(OperationCode::Branch, Immediate(branch_case.address)); Node op_b = Immediate(branch_case.cmp_value); Node condition = - GetPredicateComparisonInteger(Tegra::Shader::PredCondition::Equal, false, op_a, op_b); + GetPredicateComparisonInteger(Tegra::Shader::PredCondition::EQ, false, op_a, op_b); auto result = Conditional(condition, {n}); bb.push_back(result); global_code.push_back(result); diff --git a/src/video_core/shader/decode/xmad.cpp b/src/video_core/shader/decode/xmad.cpp index 6191ffba1b..c83dc6615a 100644 --- a/src/video_core/shader/decode/xmad.cpp +++ b/src/video_core/shader/decode/xmad.cpp @@ -97,19 +97,19 @@ u32 ShaderIR::DecodeXmad(NodeBlock& bb, u32 pc) { return SignedOperation(OperationCode::IAdd, is_signed_c, original_c, shifted_b); } case Tegra::Shader::XmadMode::CSfu: { - const Node comp_a = GetPredicateComparisonInteger(PredCondition::Equal, is_signed_a, - op_a, Immediate(0)); - const Node comp_b = GetPredicateComparisonInteger(PredCondition::Equal, is_signed_b, - op_b, Immediate(0)); + const Node comp_a = + GetPredicateComparisonInteger(PredCondition::EQ, is_signed_a, op_a, Immediate(0)); + const Node comp_b = + GetPredicateComparisonInteger(PredCondition::EQ, is_signed_b, op_b, Immediate(0)); const Node comp = Operation(OperationCode::LogicalOr, comp_a, comp_b); const Node comp_minus_a = GetPredicateComparisonInteger( - PredCondition::NotEqual, is_signed_a, + PredCondition::NE, is_signed_a, SignedOperation(OperationCode::IBitwiseAnd, is_signed_a, op_a, Immediate(0x80000000)), Immediate(0)); const Node comp_minus_b = GetPredicateComparisonInteger( - PredCondition::NotEqual, is_signed_b, + PredCondition::NE, is_signed_b, SignedOperation(OperationCode::IBitwiseAnd, is_signed_b, op_b, Immediate(0x80000000)), Immediate(0)); diff --git a/src/video_core/shader/node.h b/src/video_core/shader/node.h index 601c822d29..f75b622407 100644 --- a/src/video_core/shader/node.h +++ b/src/video_core/shader/node.h @@ -110,13 +110,20 @@ enum class OperationCode { LogicalPick2, /// (bool2 pair, uint index) -> bool LogicalAnd2, /// (bool2 a) -> bool - LogicalFLessThan, /// (float a, float b) -> bool - LogicalFEqual, /// (float a, float b) -> bool - LogicalFLessEqual, /// (float a, float b) -> bool - LogicalFGreaterThan, /// (float a, float b) -> bool - LogicalFNotEqual, /// (float a, float b) -> bool - LogicalFGreaterEqual, /// (float a, float b) -> bool - LogicalFIsNan, /// (float a) -> bool + LogicalFOrdLessThan, /// (float a, float b) -> bool + LogicalFOrdEqual, /// (float a, float b) -> bool + LogicalFOrdLessEqual, /// (float a, float b) -> bool + LogicalFOrdGreaterThan, /// (float a, float b) -> bool + LogicalFOrdNotEqual, /// (float a, float b) -> bool + LogicalFOrdGreaterEqual, /// (float a, float b) -> bool + LogicalFOrdered, /// (float a, float b) -> bool + LogicalFUnordered, /// (float a, float b) -> bool + LogicalFUnordLessThan, /// (float a, float b) -> bool + LogicalFUnordEqual, /// (float a, float b) -> bool + LogicalFUnordLessEqual, /// (float a, float b) -> bool + LogicalFUnordGreaterThan, /// (float a, float b) -> bool + LogicalFUnordNotEqual, /// (float a, float b) -> bool + LogicalFUnordGreaterEqual, /// (float a, float b) -> bool LogicalILessThan, /// (int a, int b) -> bool LogicalIEqual, /// (int a, int b) -> bool diff --git a/src/video_core/shader/shader_ir.cpp b/src/video_core/shader/shader_ir.cpp index 8226749262..e322c3402c 100644 --- a/src/video_core/shader/shader_ir.cpp +++ b/src/video_core/shader/shader_ir.cpp @@ -10,6 +10,7 @@ #include "common/common_types.h" #include "common/logging/log.h" #include "video_core/engines/shader_bytecode.h" +#include "video_core/shader/node.h" #include "video_core/shader/node_helper.h" #include "video_core/shader/registry.h" #include "video_core/shader/shader_ir.h" @@ -243,56 +244,44 @@ Node ShaderIR::GetSaturatedHalfFloat(Node value, bool saturate) { } Node ShaderIR::GetPredicateComparisonFloat(PredCondition condition, Node op_a, Node op_b) { - static constexpr std::array comparison_table{ - std::pair{PredCondition::LessThan, OperationCode::LogicalFLessThan}, - std::pair{PredCondition::Equal, OperationCode::LogicalFEqual}, - std::pair{PredCondition::LessEqual, OperationCode::LogicalFLessEqual}, - std::pair{PredCondition::GreaterThan, OperationCode::LogicalFGreaterThan}, - std::pair{PredCondition::NotEqual, OperationCode::LogicalFNotEqual}, - std::pair{PredCondition::GreaterEqual, OperationCode::LogicalFGreaterEqual}, - std::pair{PredCondition::LessThanWithNan, OperationCode::LogicalFLessThan}, - std::pair{PredCondition::NotEqualWithNan, OperationCode::LogicalFNotEqual}, - std::pair{PredCondition::LessEqualWithNan, OperationCode::LogicalFLessEqual}, - std::pair{PredCondition::GreaterThanWithNan, OperationCode::LogicalFGreaterThan}, - std::pair{PredCondition::GreaterEqualWithNan, OperationCode::LogicalFGreaterEqual}, - }; - - const auto comparison = - std::find_if(comparison_table.cbegin(), comparison_table.cend(), - [condition](const auto entry) { return condition == entry.first; }); - UNIMPLEMENTED_IF_MSG(comparison == comparison_table.cend(), - "Unknown predicate comparison operation"); - - Node predicate = Operation(comparison->second, NO_PRECISE, op_a, op_b); - - if (condition == PredCondition::LessThanWithNan || - condition == PredCondition::NotEqualWithNan || - condition == PredCondition::LessEqualWithNan || - condition == PredCondition::GreaterThanWithNan || - condition == PredCondition::GreaterEqualWithNan) { - predicate = Operation(OperationCode::LogicalOr, predicate, - Operation(OperationCode::LogicalFIsNan, op_a)); - predicate = Operation(OperationCode::LogicalOr, predicate, - Operation(OperationCode::LogicalFIsNan, op_b)); + if (condition == PredCondition::T) { + return GetPredicate(true); + } else if (condition == PredCondition::F) { + return GetPredicate(false); } - return predicate; + static constexpr std::array comparison_table{ + OperationCode(0), + OperationCode::LogicalFOrdLessThan, // LT + OperationCode::LogicalFOrdEqual, // EQ + OperationCode::LogicalFOrdLessEqual, // LE + OperationCode::LogicalFOrdGreaterThan, // GT + OperationCode::LogicalFOrdNotEqual, // NE + OperationCode::LogicalFOrdGreaterEqual, // GE + OperationCode::LogicalFOrdered, // NUM + OperationCode::LogicalFUnordered, // NAN + OperationCode::LogicalFUnordLessThan, // LTU + OperationCode::LogicalFUnordEqual, // EQU + OperationCode::LogicalFUnordLessEqual, // LEU + OperationCode::LogicalFUnordGreaterThan, // GTU + OperationCode::LogicalFUnordNotEqual, // NEU + OperationCode::LogicalFUnordGreaterEqual, // GEU + }; + const std::size_t index = static_cast(condition); + ASSERT_MSG(index < std::size(comparison_table), "Invalid condition={}", index); + + return Operation(comparison_table[index], op_a, op_b); } Node ShaderIR::GetPredicateComparisonInteger(PredCondition condition, bool is_signed, Node op_a, Node op_b) { static constexpr std::array comparison_table{ - std::pair{PredCondition::LessThan, OperationCode::LogicalILessThan}, - std::pair{PredCondition::Equal, OperationCode::LogicalIEqual}, - std::pair{PredCondition::LessEqual, OperationCode::LogicalILessEqual}, - std::pair{PredCondition::GreaterThan, OperationCode::LogicalIGreaterThan}, - std::pair{PredCondition::NotEqual, OperationCode::LogicalINotEqual}, - std::pair{PredCondition::GreaterEqual, OperationCode::LogicalIGreaterEqual}, - std::pair{PredCondition::LessThanWithNan, OperationCode::LogicalILessThan}, - std::pair{PredCondition::NotEqualWithNan, OperationCode::LogicalINotEqual}, - std::pair{PredCondition::LessEqualWithNan, OperationCode::LogicalILessEqual}, - std::pair{PredCondition::GreaterThanWithNan, OperationCode::LogicalIGreaterThan}, - std::pair{PredCondition::GreaterEqualWithNan, OperationCode::LogicalIGreaterEqual}, + std::pair{PredCondition::LT, OperationCode::LogicalILessThan}, + std::pair{PredCondition::EQ, OperationCode::LogicalIEqual}, + std::pair{PredCondition::LE, OperationCode::LogicalILessEqual}, + std::pair{PredCondition::GT, OperationCode::LogicalIGreaterThan}, + std::pair{PredCondition::NE, OperationCode::LogicalINotEqual}, + std::pair{PredCondition::GE, OperationCode::LogicalIGreaterEqual}, }; const auto comparison = @@ -301,32 +290,24 @@ Node ShaderIR::GetPredicateComparisonInteger(PredCondition condition, bool is_si UNIMPLEMENTED_IF_MSG(comparison == comparison_table.cend(), "Unknown predicate comparison operation"); - Node predicate = SignedOperation(comparison->second, is_signed, NO_PRECISE, std::move(op_a), - std::move(op_b)); - - UNIMPLEMENTED_IF_MSG(condition == PredCondition::LessThanWithNan || - condition == PredCondition::NotEqualWithNan || - condition == PredCondition::LessEqualWithNan || - condition == PredCondition::GreaterThanWithNan || - condition == PredCondition::GreaterEqualWithNan, - "NaN comparisons for integers are not implemented"); - return predicate; + return SignedOperation(comparison->second, is_signed, NO_PRECISE, std::move(op_a), + std::move(op_b)); } Node ShaderIR::GetPredicateComparisonHalf(Tegra::Shader::PredCondition condition, Node op_a, Node op_b) { static constexpr std::array comparison_table{ - std::pair{PredCondition::LessThan, OperationCode::Logical2HLessThan}, - std::pair{PredCondition::Equal, OperationCode::Logical2HEqual}, - std::pair{PredCondition::LessEqual, OperationCode::Logical2HLessEqual}, - std::pair{PredCondition::GreaterThan, OperationCode::Logical2HGreaterThan}, - std::pair{PredCondition::NotEqual, OperationCode::Logical2HNotEqual}, - std::pair{PredCondition::GreaterEqual, OperationCode::Logical2HGreaterEqual}, - std::pair{PredCondition::LessThanWithNan, OperationCode::Logical2HLessThanWithNan}, - std::pair{PredCondition::NotEqualWithNan, OperationCode::Logical2HNotEqualWithNan}, - std::pair{PredCondition::LessEqualWithNan, OperationCode::Logical2HLessEqualWithNan}, - std::pair{PredCondition::GreaterThanWithNan, OperationCode::Logical2HGreaterThanWithNan}, - std::pair{PredCondition::GreaterEqualWithNan, OperationCode::Logical2HGreaterEqualWithNan}, + std::pair{PredCondition::LT, OperationCode::Logical2HLessThan}, + std::pair{PredCondition::EQ, OperationCode::Logical2HEqual}, + std::pair{PredCondition::LE, OperationCode::Logical2HLessEqual}, + std::pair{PredCondition::GT, OperationCode::Logical2HGreaterThan}, + std::pair{PredCondition::NE, OperationCode::Logical2HNotEqual}, + std::pair{PredCondition::GE, OperationCode::Logical2HGreaterEqual}, + std::pair{PredCondition::LTU, OperationCode::Logical2HLessThanWithNan}, + std::pair{PredCondition::LEU, OperationCode::Logical2HLessEqualWithNan}, + std::pair{PredCondition::GTU, OperationCode::Logical2HGreaterThanWithNan}, + std::pair{PredCondition::NEU, OperationCode::Logical2HNotEqualWithNan}, + std::pair{PredCondition::GEU, OperationCode::Logical2HGreaterEqualWithNan}, }; const auto comparison = @@ -397,7 +378,7 @@ void ShaderIR::SetInternalFlagsFromFloat(NodeBlock& bb, Node value, bool sets_cc if (!sets_cc) { return; } - Node zerop = Operation(OperationCode::LogicalFEqual, std::move(value), Immediate(0.0f)); + Node zerop = Operation(OperationCode::LogicalFOrdEqual, std::move(value), Immediate(0.0f)); SetInternalFlag(bb, InternalFlag::Zero, std::move(zerop)); LOG_WARNING(HW_GPU, "Condition codes implementation is incomplete"); } From 8b329ddcc9b39353b9545289b3bd653a77db0103 Mon Sep 17 00:00:00 2001 From: ReinUsesLisp Date: Sat, 9 May 2020 04:56:50 -0300 Subject: [PATCH 2/2] gl_shader_decompiler: Properly emulate NaN behaviour on NE "Not equal" operators on GLSL seem to behave as unordered when we expect an ordered comparison. Manually emulate this checking for LGE values (numbers, not-NaNs). --- src/video_core/renderer_opengl/gl_shader_decompiler.cpp | 9 +++++++++ 1 file changed, 9 insertions(+) diff --git a/src/video_core/renderer_opengl/gl_shader_decompiler.cpp b/src/video_core/renderer_opengl/gl_shader_decompiler.cpp index d071abd84c..960ebf1a12 100644 --- a/src/video_core/renderer_opengl/gl_shader_decompiler.cpp +++ b/src/video_core/renderer_opengl/gl_shader_decompiler.cpp @@ -1845,6 +1845,15 @@ private: static_assert(!unordered || type == Type::Float); const Expression expr = GenerateBinaryInfix(operation, op, Type::Bool, type, type); + + if constexpr (op.compare("!=") == 0 && type == Type::Float && !unordered) { + // GLSL's operator!=(float, float) doesn't seem be ordered. This happens on both AMD's + // and Nvidia's proprietary stacks. Manually force an ordered comparison. + return {fmt::format("({} && !isnan({}) && !isnan({}))", expr.AsBool(), + VisitOperand(operation, 0).AsFloat(), + VisitOperand(operation, 1).AsFloat()), + Type::Bool}; + } if constexpr (!unordered) { return expr; }