From a77e764726938a26803fa90a9c69ccdd32ab09cd Mon Sep 17 00:00:00 2001 From: ReinUsesLisp Date: Sun, 21 Mar 2021 00:42:56 -0300 Subject: [PATCH] shader: Add support for fp16 comparisons and misc fixes --- .../backend/spirv/emit_spirv.h | 2 ++ .../spirv/emit_spirv_floating_point.cpp | 8 +++++++ .../frontend/ir/ir_emitter.cpp | 23 +++++++++++++++---- .../frontend/ir/ir_emitter.h | 6 ++--- src/shader_recompiler/frontend/ir/opcodes.inc | 2 ++ .../impl/half_floating_point_add.cpp | 3 +-- ...half_floating_point_fused_multiply_add.cpp | 4 +--- .../impl/half_floating_point_multiply.cpp | 3 +-- .../impl/half_floating_point_set.cpp | 1 + .../ir_opt/collect_shader_info_pass.cpp | 16 +++++++++++++ .../ir_opt/lower_fp16_to_fp32.cpp | 2 ++ 11 files changed, 56 insertions(+), 14 deletions(-) diff --git a/src/shader_recompiler/backend/spirv/emit_spirv.h b/src/shader_recompiler/backend/spirv/emit_spirv.h index e297a0e20..486ef10a7 100644 --- a/src/shader_recompiler/backend/spirv/emit_spirv.h +++ b/src/shader_recompiler/backend/spirv/emit_spirv.h @@ -234,7 +234,9 @@ Id EmitFPOrdGreaterThanEqual64(EmitContext& ctx, Id lhs, Id rhs); Id EmitFPUnordGreaterThanEqual16(EmitContext& ctx, Id lhs, Id rhs); Id EmitFPUnordGreaterThanEqual32(EmitContext& ctx, Id lhs, Id rhs); Id EmitFPUnordGreaterThanEqual64(EmitContext& ctx, Id lhs, Id rhs); +Id EmitFPIsNan16(EmitContext& ctx, Id value); Id EmitFPIsNan32(EmitContext& ctx, Id value); +Id EmitFPIsNan64(EmitContext& ctx, Id value); Id EmitIAdd32(EmitContext& ctx, IR::Inst* inst, Id a, Id b); void EmitIAdd64(EmitContext& ctx); Id EmitISub32(EmitContext& ctx, Id a, Id b); diff --git a/src/shader_recompiler/backend/spirv/emit_spirv_floating_point.cpp b/src/shader_recompiler/backend/spirv/emit_spirv_floating_point.cpp index e635b1ffb..1fdf66cb6 100644 --- a/src/shader_recompiler/backend/spirv/emit_spirv_floating_point.cpp +++ b/src/shader_recompiler/backend/spirv/emit_spirv_floating_point.cpp @@ -346,8 +346,16 @@ Id EmitFPUnordGreaterThanEqual64(EmitContext& ctx, Id lhs, Id rhs) { return ctx.OpFUnordGreaterThanEqual(ctx.U1, lhs, rhs); } +Id EmitFPIsNan16(EmitContext& ctx, Id value) { + return ctx.OpIsNan(ctx.U1, value); +} + Id EmitFPIsNan32(EmitContext& ctx, Id value) { return ctx.OpIsNan(ctx.U1, value); } +Id EmitFPIsNan64(EmitContext& ctx, Id value) { + return ctx.OpIsNan(ctx.U1, value); +} + } // namespace Shader::Backend::SPIRV diff --git a/src/shader_recompiler/frontend/ir/ir_emitter.cpp b/src/shader_recompiler/frontend/ir/ir_emitter.cpp index 652f6949e..1eda95071 100644 --- a/src/shader_recompiler/frontend/ir/ir_emitter.cpp +++ b/src/shader_recompiler/frontend/ir/ir_emitter.cpp @@ -895,15 +895,30 @@ U1 IREmitter::FPGreaterThanEqual(const F16F32F64& lhs, const F16F32F64& rhs, FpC } } -U1 IREmitter::FPIsNan(const F32& value) { - return Inst(Opcode::FPIsNan32, value); +U1 IREmitter::FPIsNan(const F16F32F64& value) { + switch (value.Type()) { + case Type::F16: + return Inst(Opcode::FPIsNan16, value); + case Type::F32: + return Inst(Opcode::FPIsNan32, value); + case Type::F64: + return Inst(Opcode::FPIsNan64, value); + default: + ThrowInvalidType(value.Type()); + } } -U1 IREmitter::FPOrdered(const F32& lhs, const F32& rhs) { +U1 IREmitter::FPOrdered(const F16F32F64& lhs, const F16F32F64& rhs) { + if (lhs.Type() != rhs.Type()) { + throw InvalidArgument("Mismatching types {} and {}", lhs.Type(), rhs.Type()); + } return LogicalAnd(LogicalNot(FPIsNan(lhs)), LogicalNot(FPIsNan(rhs))); } -U1 IREmitter::FPUnordered(const F32& lhs, const F32& rhs) { +U1 IREmitter::FPUnordered(const F16F32F64& lhs, const F16F32F64& rhs) { + if (lhs.Type() != rhs.Type()) { + throw InvalidArgument("Mismatching types {} and {}", lhs.Type(), rhs.Type()); + } return LogicalOr(FPIsNan(lhs), FPIsNan(rhs)); } diff --git a/src/shader_recompiler/frontend/ir/ir_emitter.h b/src/shader_recompiler/frontend/ir/ir_emitter.h index 8edb11154..ab4537d88 100644 --- a/src/shader_recompiler/frontend/ir/ir_emitter.h +++ b/src/shader_recompiler/frontend/ir/ir_emitter.h @@ -161,9 +161,9 @@ public: FpControl control = {}, bool ordered = true); [[nodiscard]] U1 FPGreaterThanEqual(const F16F32F64& lhs, const F16F32F64& rhs, FpControl control = {}, bool ordered = true); - [[nodiscard]] U1 FPIsNan(const F32& value); - [[nodiscard]] U1 FPOrdered(const F32& lhs, const F32& rhs); - [[nodiscard]] U1 FPUnordered(const F32& lhs, const F32& rhs); + [[nodiscard]] U1 FPIsNan(const F16F32F64& value); + [[nodiscard]] U1 FPOrdered(const F16F32F64& lhs, const F16F32F64& rhs); + [[nodiscard]] U1 FPUnordered(const F16F32F64& lhs, const F16F32F64& rhs); [[nodiscard]] F32F64 FPMax(const F32F64& lhs, const F32F64& rhs, FpControl control = {}); [[nodiscard]] F32F64 FPMin(const F32F64& lhs, const F32F64& rhs, FpControl control = {}); diff --git a/src/shader_recompiler/frontend/ir/opcodes.inc b/src/shader_recompiler/frontend/ir/opcodes.inc index 8471db7b9..884eea7a8 100644 --- a/src/shader_recompiler/frontend/ir/opcodes.inc +++ b/src/shader_recompiler/frontend/ir/opcodes.inc @@ -236,7 +236,9 @@ OPCODE(FPOrdGreaterThanEqual64, U1, F64, OPCODE(FPUnordGreaterThanEqual16, U1, F16, F16, ) OPCODE(FPUnordGreaterThanEqual32, U1, F32, F32, ) OPCODE(FPUnordGreaterThanEqual64, U1, F64, F64, ) +OPCODE(FPIsNan16, U1, F16, ) OPCODE(FPIsNan32, U1, F32, ) +OPCODE(FPIsNan64, U1, F64, ) // Integer operations OPCODE(IAdd32, U32, U32, U32, ) diff --git a/src/shader_recompiler/frontend/maxwell/translate/impl/half_floating_point_add.cpp b/src/shader_recompiler/frontend/maxwell/translate/impl/half_floating_point_add.cpp index 19e3401ca..03e7bf047 100644 --- a/src/shader_recompiler/frontend/maxwell/translate/impl/half_floating_point_add.cpp +++ b/src/shader_recompiler/frontend/maxwell/translate/impl/half_floating_point_add.cpp @@ -6,7 +6,6 @@ namespace Shader::Maxwell { namespace { - void HADD2(TranslatorVisitor& v, u64 insn, Merge merge, bool ftz, bool sat, bool abs_a, bool neg_a, Swizzle swizzle_a, bool abs_b, bool neg_b, Swizzle swizzle_b, const IR::U32& src_b) { union { @@ -66,7 +65,7 @@ void HADD2(TranslatorVisitor& v, u64 insn, bool sat, bool abs_b, bool neg_b, Swi HADD2(v, insn, hadd2.merge, hadd2.ftz != 0, sat, hadd2.abs_a != 0, hadd2.neg_a != 0, hadd2.swizzle_a, abs_b, neg_b, swizzle_b, src_b); } -} // namespace +} // Anonymous namespace void TranslatorVisitor::HADD2_reg(u64 insn) { union { diff --git a/src/shader_recompiler/frontend/maxwell/translate/impl/half_floating_point_fused_multiply_add.cpp b/src/shader_recompiler/frontend/maxwell/translate/impl/half_floating_point_fused_multiply_add.cpp index 2f3996274..8b234bd6a 100644 --- a/src/shader_recompiler/frontend/maxwell/translate/impl/half_floating_point_fused_multiply_add.cpp +++ b/src/shader_recompiler/frontend/maxwell/translate/impl/half_floating_point_fused_multiply_add.cpp @@ -6,7 +6,6 @@ namespace Shader::Maxwell { namespace { - void HFMA2(TranslatorVisitor& v, u64 insn, Merge merge, Swizzle swizzle_a, bool neg_b, bool neg_c, Swizzle swizzle_b, Swizzle swizzle_c, const IR::U32& src_b, const IR::U32& src_c, bool sat, HalfPrecision precision) { @@ -85,8 +84,7 @@ void HFMA2(TranslatorVisitor& v, u64 insn, bool neg_b, bool neg_c, Swizzle swizz HFMA2(v, insn, hfma2.merge, hfma2.swizzle_a, neg_b, neg_c, swizzle_b, swizzle_c, src_b, src_c, sat, precision); } - -} // namespace +} // Anonymous namespace void TranslatorVisitor::HFMA2_reg(u64 insn) { union { diff --git a/src/shader_recompiler/frontend/maxwell/translate/impl/half_floating_point_multiply.cpp b/src/shader_recompiler/frontend/maxwell/translate/impl/half_floating_point_multiply.cpp index ff34a8c8f..2451a6ef6 100644 --- a/src/shader_recompiler/frontend/maxwell/translate/impl/half_floating_point_multiply.cpp +++ b/src/shader_recompiler/frontend/maxwell/translate/impl/half_floating_point_multiply.cpp @@ -6,7 +6,6 @@ namespace Shader::Maxwell { namespace { - void HMUL2(TranslatorVisitor& v, u64 insn, Merge merge, bool sat, bool abs_a, bool neg_a, Swizzle swizzle_a, bool abs_b, bool neg_b, Swizzle swizzle_b, const IR::U32& src_b, HalfPrecision precision) { @@ -79,7 +78,7 @@ void HMUL2(TranslatorVisitor& v, u64 insn, bool sat, bool abs_a, bool neg_a, boo HMUL2(v, insn, hmul2.merge, sat, abs_a, neg_a, hmul2.swizzle_a, abs_b, neg_b, swizzle_b, src_b, hmul2.precision); } -} // namespace +} // Anonymous namespace void TranslatorVisitor::HMUL2_reg(u64 insn) { union { diff --git a/src/shader_recompiler/frontend/maxwell/translate/impl/half_floating_point_set.cpp b/src/shader_recompiler/frontend/maxwell/translate/impl/half_floating_point_set.cpp index 1d28c0531..7f1f4b88c 100644 --- a/src/shader_recompiler/frontend/maxwell/translate/impl/half_floating_point_set.cpp +++ b/src/shader_recompiler/frontend/maxwell/translate/impl/half_floating_point_set.cpp @@ -76,6 +76,7 @@ void TranslatorVisitor::HSET2_reg(u64 insn) { BitField<35, 4, FPCompareOp> compare_op; BitField<28, 2, Swizzle> swizzle_b; } const hset2{insn}; + HSET2(*this, insn, GetReg20(insn), hset2.bf != 0, hset2.ftz != 0, hset2.neg_b != 0, hset2.abs_b != 0, hset2.compare_op, hset2.swizzle_b); } diff --git a/src/shader_recompiler/ir_opt/collect_shader_info_pass.cpp b/src/shader_recompiler/ir_opt/collect_shader_info_pass.cpp index e7fa3fce0..fd6069c65 100644 --- a/src/shader_recompiler/ir_opt/collect_shader_info_pass.cpp +++ b/src/shader_recompiler/ir_opt/collect_shader_info_pass.cpp @@ -74,6 +74,9 @@ void VisitUsages(Info& info, IR::Inst& inst) { 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: @@ -103,6 +106,19 @@ void VisitUsages(Info& info, IR::Inst& inst) { case IR::Opcode::FPRoundEven16: case IR::Opcode::FPSaturate16: 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::FPAbs64: diff --git a/src/shader_recompiler/ir_opt/lower_fp16_to_fp32.cpp b/src/shader_recompiler/ir_opt/lower_fp16_to_fp32.cpp index 7723c9a57..0e8862f45 100644 --- a/src/shader_recompiler/ir_opt/lower_fp16_to_fp32.cpp +++ b/src/shader_recompiler/ir_opt/lower_fp16_to_fp32.cpp @@ -74,6 +74,8 @@ IR::Opcode Replace(IR::Opcode op) { return IR::Opcode::FPOrdGreaterThanEqual32; case IR::Opcode::FPUnordGreaterThanEqual16: return IR::Opcode::FPUnordGreaterThanEqual32; + case IR::Opcode::FPIsNan16: + return IR::Opcode::FPIsNan32; case IR::Opcode::ConvertS16F16: return IR::Opcode::ConvertS16F32; case IR::Opcode::ConvertS32F16: