mirror of
https://git.suyu.dev/suyu/suyu.git
synced 2024-11-25 04:46:27 -05:00
shader: Implement FMNMX
And add a const in FCMP
This commit is contained in:
parent
2d422b2498
commit
8d470c2e63
8 changed files with 101 additions and 25 deletions
|
@ -70,6 +70,7 @@ add_library(shader_recompiler STATIC
|
||||||
frontend/maxwell/translate/impl/floating_point_compare.cpp
|
frontend/maxwell/translate/impl/floating_point_compare.cpp
|
||||||
frontend/maxwell/translate/impl/floating_point_conversion_integer.cpp
|
frontend/maxwell/translate/impl/floating_point_conversion_integer.cpp
|
||||||
frontend/maxwell/translate/impl/floating_point_fused_multiply_add.cpp
|
frontend/maxwell/translate/impl/floating_point_fused_multiply_add.cpp
|
||||||
|
frontend/maxwell/translate/impl/floating_point_min_max.cpp
|
||||||
frontend/maxwell/translate/impl/floating_point_multi_function.cpp
|
frontend/maxwell/translate/impl/floating_point_multi_function.cpp
|
||||||
frontend/maxwell/translate/impl/floating_point_multiply.cpp
|
frontend/maxwell/translate/impl/floating_point_multiply.cpp
|
||||||
frontend/maxwell/translate/impl/floating_point_range_reduction.cpp
|
frontend/maxwell/translate/impl/floating_point_range_reduction.cpp
|
||||||
|
|
|
@ -162,10 +162,10 @@ Id EmitFPAdd64(EmitContext& ctx, IR::Inst* inst, Id a, Id b);
|
||||||
Id EmitFPFma16(EmitContext& ctx, IR::Inst* inst, Id a, Id b, Id c);
|
Id EmitFPFma16(EmitContext& ctx, IR::Inst* inst, Id a, Id b, Id c);
|
||||||
Id EmitFPFma32(EmitContext& ctx, IR::Inst* inst, Id a, Id b, Id c);
|
Id EmitFPFma32(EmitContext& ctx, IR::Inst* inst, Id a, Id b, Id c);
|
||||||
Id EmitFPFma64(EmitContext& ctx, IR::Inst* inst, Id a, Id b, Id c);
|
Id EmitFPFma64(EmitContext& ctx, IR::Inst* inst, Id a, Id b, Id c);
|
||||||
void EmitFPMax32(EmitContext& ctx);
|
Id EmitFPMax32(EmitContext& ctx, Id a, Id b);
|
||||||
void EmitFPMax64(EmitContext& ctx);
|
Id EmitFPMax64(EmitContext& ctx, Id a, Id b);
|
||||||
void EmitFPMin32(EmitContext& ctx);
|
Id EmitFPMin32(EmitContext& ctx, Id a, Id b);
|
||||||
void EmitFPMin64(EmitContext& ctx);
|
Id EmitFPMin64(EmitContext& ctx, Id a, Id b);
|
||||||
Id EmitFPMul16(EmitContext& ctx, IR::Inst* inst, Id a, Id b);
|
Id EmitFPMul16(EmitContext& ctx, IR::Inst* inst, Id a, Id b);
|
||||||
Id EmitFPMul32(EmitContext& ctx, IR::Inst* inst, Id a, Id b);
|
Id EmitFPMul32(EmitContext& ctx, IR::Inst* inst, Id a, Id b);
|
||||||
Id EmitFPMul64(EmitContext& ctx, IR::Inst* inst, Id a, Id b);
|
Id EmitFPMul64(EmitContext& ctx, IR::Inst* inst, Id a, Id b);
|
||||||
|
|
|
@ -60,20 +60,20 @@ Id EmitFPFma64(EmitContext& ctx, IR::Inst* inst, Id a, Id b, Id c) {
|
||||||
return Decorate(ctx, inst, ctx.OpFma(ctx.F64[1], a, b, c));
|
return Decorate(ctx, inst, ctx.OpFma(ctx.F64[1], a, b, c));
|
||||||
}
|
}
|
||||||
|
|
||||||
void EmitFPMax32(EmitContext&) {
|
Id EmitFPMax32(EmitContext& ctx, Id a, Id b) {
|
||||||
throw NotImplementedException("SPIR-V Instruction");
|
return ctx.OpFMax(ctx.F32[1], a, b);
|
||||||
}
|
}
|
||||||
|
|
||||||
void EmitFPMax64(EmitContext&) {
|
Id EmitFPMax64(EmitContext& ctx, Id a, Id b) {
|
||||||
throw NotImplementedException("SPIR-V Instruction");
|
return ctx.OpFMax(ctx.F64[1], a, b);
|
||||||
}
|
}
|
||||||
|
|
||||||
void EmitFPMin32(EmitContext&) {
|
Id EmitFPMin32(EmitContext& ctx, Id a, Id b) {
|
||||||
throw NotImplementedException("SPIR-V Instruction");
|
return ctx.OpFMin(ctx.F32[1], a, b);
|
||||||
}
|
}
|
||||||
|
|
||||||
void EmitFPMin64(EmitContext&) {
|
Id EmitFPMin64(EmitContext& ctx, Id a, Id b) {
|
||||||
throw NotImplementedException("SPIR-V Instruction");
|
return ctx.OpFMin(ctx.F64[1], a, b);
|
||||||
}
|
}
|
||||||
|
|
||||||
Id EmitFPMul16(EmitContext& ctx, IR::Inst* inst, Id a, Id b) {
|
Id EmitFPMul16(EmitContext& ctx, IR::Inst* inst, Id a, Id b) {
|
||||||
|
|
|
@ -831,6 +831,34 @@ U1 IREmitter::FPUnordered(const F32& lhs, const F32& rhs) {
|
||||||
return LogicalOr(FPIsNan(lhs), FPIsNan(rhs));
|
return LogicalOr(FPIsNan(lhs), FPIsNan(rhs));
|
||||||
}
|
}
|
||||||
|
|
||||||
|
F32F64 IREmitter::FPMax(const F32F64& lhs, const F32F64& rhs, FpControl control) {
|
||||||
|
if (lhs.Type() != rhs.Type()) {
|
||||||
|
throw InvalidArgument("Mismatching types {} and {}", lhs.Type(), rhs.Type());
|
||||||
|
}
|
||||||
|
switch (lhs.Type()) {
|
||||||
|
case Type::F32:
|
||||||
|
return Inst<F32>(Opcode::FPMax32, Flags{control}, lhs, rhs);
|
||||||
|
case Type::F64:
|
||||||
|
return Inst<F64>(Opcode::FPMax64, Flags{control}, lhs, rhs);
|
||||||
|
default:
|
||||||
|
ThrowInvalidType(lhs.Type());
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
|
F32F64 IREmitter::FPMin(const F32F64& lhs, const F32F64& rhs, FpControl control) {
|
||||||
|
if (lhs.Type() != rhs.Type()) {
|
||||||
|
throw InvalidArgument("Mismatching types {} and {}", lhs.Type(), rhs.Type());
|
||||||
|
}
|
||||||
|
switch (lhs.Type()) {
|
||||||
|
case Type::F32:
|
||||||
|
return Inst<F32>(Opcode::FPMin32, Flags{control}, lhs, rhs);
|
||||||
|
case Type::F64:
|
||||||
|
return Inst<F64>(Opcode::FPMin64, Flags{control}, lhs, rhs);
|
||||||
|
default:
|
||||||
|
ThrowInvalidType(lhs.Type());
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
U32U64 IREmitter::IAdd(const U32U64& a, const U32U64& b) {
|
U32U64 IREmitter::IAdd(const U32U64& a, const U32U64& b) {
|
||||||
if (a.Type() != b.Type()) {
|
if (a.Type() != b.Type()) {
|
||||||
throw InvalidArgument("Mismatching types {} and {}", a.Type(), b.Type());
|
throw InvalidArgument("Mismatching types {} and {}", a.Type(), b.Type());
|
||||||
|
|
|
@ -155,6 +155,8 @@ public:
|
||||||
[[nodiscard]] U1 FPIsNan(const F32& value);
|
[[nodiscard]] U1 FPIsNan(const F32& value);
|
||||||
[[nodiscard]] U1 FPOrdered(const F32& lhs, const F32& rhs);
|
[[nodiscard]] U1 FPOrdered(const F32& lhs, const F32& rhs);
|
||||||
[[nodiscard]] U1 FPUnordered(const F32& lhs, const F32& rhs);
|
[[nodiscard]] U1 FPUnordered(const F32& lhs, const F32& rhs);
|
||||||
|
[[nodiscard]] F32F64 FPMax(const F32F64& lhs, const F32F64& rhs, FpControl control = {});
|
||||||
|
[[nodiscard]] F32F64 FPMin(const F32F64& lhs, const F32F64& rhs, FpControl control = {});
|
||||||
|
|
||||||
[[nodiscard]] U32U64 IAdd(const U32U64& a, const U32U64& b);
|
[[nodiscard]] U32U64 IAdd(const U32U64& a, const U32U64& b);
|
||||||
[[nodiscard]] U32U64 ISub(const U32U64& a, const U32U64& b);
|
[[nodiscard]] U32U64 ISub(const U32U64& a, const U32U64& b);
|
||||||
|
|
|
@ -88,7 +88,7 @@ void FCMP(TranslatorVisitor& v, u64 insn, const IR::U32& src_a, const IR::F32& o
|
||||||
|
|
||||||
const IR::F32 zero{v.ir.Imm32(0.0f)};
|
const IR::F32 zero{v.ir.Imm32(0.0f)};
|
||||||
const IR::F32 neg_zero{v.ir.Imm32(-0.0f)};
|
const IR::F32 neg_zero{v.ir.Imm32(-0.0f)};
|
||||||
IR::FpControl control{.fmz_mode{fcmp.ftz != 0 ? IR::FmzMode::FTZ : IR::FmzMode::None}};
|
const IR::FpControl control{.fmz_mode{fcmp.ftz != 0 ? IR::FmzMode::FTZ : IR::FmzMode::None}};
|
||||||
const IR::U1 cmp_result{FloatingPointCompare(v.ir, operand, zero, fcmp.compare_op, control)};
|
const IR::U1 cmp_result{FloatingPointCompare(v.ir, operand, zero, fcmp.compare_op, control)};
|
||||||
const IR::U32 src_reg{v.X(fcmp.src_reg)};
|
const IR::U32 src_reg{v.X(fcmp.src_reg)};
|
||||||
const IR::U32 result{v.ir.Select(cmp_result, src_reg, src_a)};
|
const IR::U32 result{v.ir.Select(cmp_result, src_reg, src_a)};
|
||||||
|
|
|
@ -0,0 +1,57 @@
|
||||||
|
// Copyright 2021 yuzu Emulator Project
|
||||||
|
// Licensed under GPLv2 or any later version
|
||||||
|
// Refer to the license.txt file included.
|
||||||
|
|
||||||
|
#include "common/bit_field.h"
|
||||||
|
#include "common/common_types.h"
|
||||||
|
#include "shader_recompiler/frontend/maxwell/translate/impl/impl.h"
|
||||||
|
|
||||||
|
namespace Shader::Maxwell {
|
||||||
|
namespace {
|
||||||
|
void FMNMX(TranslatorVisitor& v, u64 insn, const IR::F32& src_b) {
|
||||||
|
union {
|
||||||
|
u64 insn;
|
||||||
|
BitField<0, 8, IR::Reg> dest_reg;
|
||||||
|
BitField<8, 8, IR::Reg> src_a_reg;
|
||||||
|
BitField<39, 3, IR::Pred> pred;
|
||||||
|
BitField<42, 1, u64> neg_pred;
|
||||||
|
BitField<44, 1, u64> ftz;
|
||||||
|
BitField<45, 1, u64> negate_b;
|
||||||
|
BitField<46, 1, u64> abs_a;
|
||||||
|
BitField<48, 1, u64> negate_a;
|
||||||
|
BitField<49, 1, u64> abs_b;
|
||||||
|
} const fmnmx{insn};
|
||||||
|
|
||||||
|
const IR::U1 pred{v.ir.GetPred(fmnmx.pred)};
|
||||||
|
const IR::F32 op_a{v.ir.FPAbsNeg(v.F(fmnmx.src_a_reg), fmnmx.abs_a != 0, fmnmx.negate_a != 0)};
|
||||||
|
const IR::F32 op_b = v.ir.FPAbsNeg(src_b, fmnmx.abs_b != 0, fmnmx.negate_b != 0);
|
||||||
|
|
||||||
|
const IR::FpControl control{
|
||||||
|
.no_contraction{false},
|
||||||
|
.rounding{IR::FpRounding::DontCare},
|
||||||
|
.fmz_mode{fmnmx.ftz != 0 ? IR::FmzMode::FTZ : IR::FmzMode::None},
|
||||||
|
};
|
||||||
|
IR::F32 max{v.ir.FPMax(op_a, op_b, control)};
|
||||||
|
IR::F32 min{v.ir.FPMin(op_a, op_b, control)};
|
||||||
|
|
||||||
|
if (fmnmx.neg_pred != 0) {
|
||||||
|
std::swap(min, max);
|
||||||
|
}
|
||||||
|
|
||||||
|
v.F(fmnmx.dest_reg, IR::F32{v.ir.Select(pred, min, max)});
|
||||||
|
}
|
||||||
|
} // Anonymous namespace
|
||||||
|
|
||||||
|
void TranslatorVisitor::FMNMX_reg(u64 insn) {
|
||||||
|
FMNMX(*this, insn, GetFloatReg20(insn));
|
||||||
|
}
|
||||||
|
|
||||||
|
void TranslatorVisitor::FMNMX_cbuf(u64 insn) {
|
||||||
|
FMNMX(*this, insn, GetFloatCbuf(insn));
|
||||||
|
}
|
||||||
|
|
||||||
|
void TranslatorVisitor::FMNMX_imm(u64 insn) {
|
||||||
|
FMNMX(*this, insn, GetFloatImm20(insn));
|
||||||
|
}
|
||||||
|
|
||||||
|
} // namespace Shader::Maxwell
|
|
@ -201,18 +201,6 @@ void TranslatorVisitor::FCHK_imm(u64) {
|
||||||
ThrowNotImplemented(Opcode::FCHK_imm);
|
ThrowNotImplemented(Opcode::FCHK_imm);
|
||||||
}
|
}
|
||||||
|
|
||||||
void TranslatorVisitor::FMNMX_reg(u64) {
|
|
||||||
ThrowNotImplemented(Opcode::FMNMX_reg);
|
|
||||||
}
|
|
||||||
|
|
||||||
void TranslatorVisitor::FMNMX_cbuf(u64) {
|
|
||||||
ThrowNotImplemented(Opcode::FMNMX_cbuf);
|
|
||||||
}
|
|
||||||
|
|
||||||
void TranslatorVisitor::FMNMX_imm(u64) {
|
|
||||||
ThrowNotImplemented(Opcode::FMNMX_imm);
|
|
||||||
}
|
|
||||||
|
|
||||||
void TranslatorVisitor::FSET_reg(u64) {
|
void TranslatorVisitor::FSET_reg(u64) {
|
||||||
ThrowNotImplemented(Opcode::FSET_reg);
|
ThrowNotImplemented(Opcode::FSET_reg);
|
||||||
}
|
}
|
||||||
|
|
Loading…
Reference in a new issue