mirror of
https://github.com/yuzu-emu/yuzu.git
synced 2024-07-04 23:31:19 +01:00
shader: Implement FSET and FSETP
Also fix oversight with adding SignedZeroInfNanPreserve execution mode.
This commit is contained in:
parent
17a82b56d7
commit
fa2f6e38f4
9 changed files with 204 additions and 94 deletions
|
@ -66,12 +66,14 @@ add_library(shader_recompiler STATIC
|
|||
frontend/maxwell/translate/impl/find_leading_one.cpp
|
||||
frontend/maxwell/translate/impl/floating_point_add.cpp
|
||||
frontend/maxwell/translate/impl/floating_point_compare.cpp
|
||||
frontend/maxwell/translate/impl/floating_point_compare_and_set.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_min_max.cpp
|
||||
frontend/maxwell/translate/impl/floating_point_multi_function.cpp
|
||||
frontend/maxwell/translate/impl/floating_point_multiply.cpp
|
||||
frontend/maxwell/translate/impl/floating_point_range_reduction.cpp
|
||||
frontend/maxwell/translate/impl/floating_point_set_predicate.cpp
|
||||
frontend/maxwell/translate/impl/half_floating_point_add.cpp
|
||||
frontend/maxwell/translate/impl/impl.cpp
|
||||
frontend/maxwell/translate/impl/impl.h
|
||||
|
|
|
@ -124,10 +124,12 @@ void SetupDenormControl(const Profile& profile, const IR::Program& program, Emit
|
|||
ctx.AddExtension("SPV_KHR_float_controls");
|
||||
|
||||
if (info.uses_fp16 && profile.support_fp16_signed_zero_nan_preserve) {
|
||||
ctx.AddExecutionMode(main_func, spv::ExecutionMode::SignedZeroInfNanPreserve);
|
||||
ctx.AddCapability(spv::Capability::SignedZeroInfNanPreserve);
|
||||
ctx.AddExecutionMode(main_func, spv::ExecutionMode::SignedZeroInfNanPreserve, 16U);
|
||||
}
|
||||
if (profile.support_fp32_signed_zero_nan_preserve) {
|
||||
ctx.AddExecutionMode(main_func, spv::ExecutionMode::SignedZeroInfNanPreserve);
|
||||
ctx.AddCapability(spv::Capability::SignedZeroInfNanPreserve);
|
||||
ctx.AddExecutionMode(main_func, spv::ExecutionMode::SignedZeroInfNanPreserve, 32U);
|
||||
}
|
||||
if (info.uses_fp32_denorms_flush && info.uses_fp32_denorms_preserve) {
|
||||
// LOG_ERROR(HW_GPU, "Fp32 denorm flush and preserve on the same shader");
|
||||
|
|
|
@ -58,4 +58,52 @@ IR::U1 PredicateOperation(IR::IREmitter& ir, const IR::U32& result, PredicateOp
|
|||
}
|
||||
}
|
||||
|
||||
bool IsCompareOpOrdered(FPCompareOp op) {
|
||||
switch (op) {
|
||||
case FPCompareOp::LTU:
|
||||
case FPCompareOp::EQU:
|
||||
case FPCompareOp::LEU:
|
||||
case FPCompareOp::GTU:
|
||||
case FPCompareOp::NEU:
|
||||
case FPCompareOp::GEU:
|
||||
return false;
|
||||
default:
|
||||
return true;
|
||||
}
|
||||
}
|
||||
|
||||
IR::U1 FloatingPointCompare(IR::IREmitter& ir, const IR::F32& operand_1, const IR::F32& operand_2,
|
||||
FPCompareOp compare_op, IR::FpControl control) {
|
||||
const bool ordered{IsCompareOpOrdered(compare_op)};
|
||||
switch (compare_op) {
|
||||
case FPCompareOp::F:
|
||||
return ir.Imm1(false);
|
||||
case FPCompareOp::LT:
|
||||
case FPCompareOp::LTU:
|
||||
return ir.FPLessThan(operand_1, operand_2, control, ordered);
|
||||
case FPCompareOp::EQ:
|
||||
case FPCompareOp::EQU:
|
||||
return ir.FPEqual(operand_1, operand_2, control, ordered);
|
||||
case FPCompareOp::LE:
|
||||
case FPCompareOp::LEU:
|
||||
return ir.FPLessThanEqual(operand_1, operand_2, control, ordered);
|
||||
case FPCompareOp::GT:
|
||||
case FPCompareOp::GTU:
|
||||
return ir.FPGreaterThan(operand_1, operand_2, control, ordered);
|
||||
case FPCompareOp::NE:
|
||||
case FPCompareOp::NEU:
|
||||
return ir.FPNotEqual(operand_1, operand_2, control, ordered);
|
||||
case FPCompareOp::GE:
|
||||
case FPCompareOp::GEU:
|
||||
return ir.FPGreaterThanEqual(operand_1, operand_2, control, ordered);
|
||||
case FPCompareOp::NUM:
|
||||
return ir.FPOrdered(operand_1, operand_2);
|
||||
case FPCompareOp::Nan:
|
||||
return ir.FPUnordered(operand_1, operand_2);
|
||||
case FPCompareOp::T:
|
||||
return ir.Imm1(true);
|
||||
default:
|
||||
throw NotImplementedException("Invalid FP compare op {}", compare_op);
|
||||
}
|
||||
}
|
||||
} // namespace Shader::Maxwell
|
||||
|
|
|
@ -15,4 +15,10 @@ namespace Shader::Maxwell {
|
|||
const IR::U1& predicate_2, BooleanOp bop);
|
||||
|
||||
[[nodiscard]] IR::U1 PredicateOperation(IR::IREmitter& ir, const IR::U32& result, PredicateOp op);
|
||||
|
||||
[[nodiscard]] bool IsCompareOpOrdered(FPCompareOp op);
|
||||
|
||||
[[nodiscard]] IR::U1 FloatingPointCompare(IR::IREmitter& ir, const IR::F32& operand_1,
|
||||
const IR::F32& operand_2, FPCompareOp compare_op,
|
||||
IR::FpControl control = {});
|
||||
} // namespace Shader::Maxwell
|
||||
|
|
|
@ -9,74 +9,6 @@
|
|||
|
||||
namespace Shader::Maxwell {
|
||||
namespace {
|
||||
enum class FPCompareOp : u64 {
|
||||
F,
|
||||
LT,
|
||||
EQ,
|
||||
LE,
|
||||
GT,
|
||||
NE,
|
||||
GE,
|
||||
NUM,
|
||||
Nan,
|
||||
LTU,
|
||||
EQU,
|
||||
LEU,
|
||||
GTU,
|
||||
NEU,
|
||||
GEU,
|
||||
T,
|
||||
};
|
||||
|
||||
bool IsCompareOpOrdered(FPCompareOp op) {
|
||||
switch (op) {
|
||||
case FPCompareOp::LTU:
|
||||
case FPCompareOp::EQU:
|
||||
case FPCompareOp::LEU:
|
||||
case FPCompareOp::GTU:
|
||||
case FPCompareOp::NEU:
|
||||
case FPCompareOp::GEU:
|
||||
return false;
|
||||
default:
|
||||
return true;
|
||||
}
|
||||
}
|
||||
|
||||
IR::U1 FloatingPointCompare(IR::IREmitter& ir, const IR::F32& operand_1, const IR::F32& operand_2,
|
||||
FPCompareOp compare_op, IR::FpControl control) {
|
||||
const bool ordered{IsCompareOpOrdered(compare_op)};
|
||||
switch (compare_op) {
|
||||
case FPCompareOp::F:
|
||||
return ir.Imm1(false);
|
||||
case FPCompareOp::LT:
|
||||
case FPCompareOp::LTU:
|
||||
return ir.FPLessThan(operand_1, operand_2, control, ordered);
|
||||
case FPCompareOp::EQ:
|
||||
case FPCompareOp::EQU:
|
||||
return ir.FPEqual(operand_1, operand_2, control, ordered);
|
||||
case FPCompareOp::LE:
|
||||
case FPCompareOp::LEU:
|
||||
return ir.FPLessThanEqual(operand_1, operand_2, control, ordered);
|
||||
case FPCompareOp::GT:
|
||||
case FPCompareOp::GTU:
|
||||
return ir.FPGreaterThan(operand_1, operand_2, control, ordered);
|
||||
case FPCompareOp::NE:
|
||||
case FPCompareOp::NEU:
|
||||
return ir.FPNotEqual(operand_1, operand_2, control, ordered);
|
||||
case FPCompareOp::GE:
|
||||
case FPCompareOp::GEU:
|
||||
return ir.FPGreaterThanEqual(operand_1, operand_2, control, ordered);
|
||||
case FPCompareOp::NUM:
|
||||
return ir.FPOrdered(operand_1, operand_2);
|
||||
case FPCompareOp::Nan:
|
||||
return ir.FPUnordered(operand_1, operand_2);
|
||||
case FPCompareOp::T:
|
||||
return ir.Imm1(true);
|
||||
default:
|
||||
throw NotImplementedException("Invalid compare op {}", compare_op);
|
||||
}
|
||||
}
|
||||
|
||||
void FCMP(TranslatorVisitor& v, u64 insn, const IR::U32& src_a, const IR::F32& operand) {
|
||||
union {
|
||||
u64 insn;
|
||||
|
|
|
@ -0,0 +1,65 @@
|
|||
// 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/common_funcs.h"
|
||||
#include "shader_recompiler/frontend/maxwell/translate/impl/impl.h"
|
||||
|
||||
namespace Shader::Maxwell {
|
||||
namespace {
|
||||
void FSET(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<43, 1, u64> negate_a;
|
||||
BitField<44, 1, u64> abs_b;
|
||||
BitField<45, 2, BooleanOp> bop;
|
||||
BitField<48, 4, FPCompareOp> compare_op;
|
||||
BitField<52, 1, u64> bf;
|
||||
BitField<53, 1, u64> negate_b;
|
||||
BitField<54, 1, u64> abs_a;
|
||||
BitField<55, 1, u64> ftz;
|
||||
} const fset{insn};
|
||||
|
||||
const IR::F32 op_a{v.ir.FPAbsNeg(v.F(fset.src_a_reg), fset.abs_a != 0, fset.negate_a != 0)};
|
||||
const IR::F32 op_b = v.ir.FPAbsNeg(src_b, fset.abs_b != 0, fset.negate_b != 0);
|
||||
const IR::FpControl control{
|
||||
.no_contraction{false},
|
||||
.rounding{IR::FpRounding::DontCare},
|
||||
.fmz_mode{fset.ftz != 0 ? IR::FmzMode::FTZ : IR::FmzMode::None},
|
||||
};
|
||||
|
||||
IR::U1 pred{v.ir.GetPred(fset.pred)};
|
||||
if (fset.neg_pred != 0) {
|
||||
pred = v.ir.LogicalNot(pred);
|
||||
}
|
||||
const IR::U1 cmp_result{FloatingPointCompare(v.ir, op_a, op_b, fset.compare_op, control)};
|
||||
const IR::U1 bop_result{PredicateCombine(v.ir, cmp_result, pred, fset.bop)};
|
||||
|
||||
const IR::U32 one_mask{v.ir.Imm32(-1)};
|
||||
const IR::U32 fp_one{v.ir.Imm32(0x3f800000)};
|
||||
const IR::U32 fail_result{v.ir.Imm32(0)};
|
||||
const IR::U32 pass_result{fset.bf == 0 ? one_mask : fp_one};
|
||||
|
||||
v.X(fset.dest_reg, IR::U32{v.ir.Select(bop_result, pass_result, fail_result)});
|
||||
}
|
||||
} // Anonymous namespace
|
||||
|
||||
void TranslatorVisitor::FSET_reg(u64 insn) {
|
||||
FSET(*this, insn, GetFloatReg20(insn));
|
||||
}
|
||||
|
||||
void TranslatorVisitor::FSET_cbuf(u64 insn) {
|
||||
FSET(*this, insn, GetFloatCbuf(insn));
|
||||
}
|
||||
|
||||
void TranslatorVisitor::FSET_imm(u64 insn) {
|
||||
FSET(*this, insn, GetFloatImm20(insn));
|
||||
}
|
||||
|
||||
} // namespace Shader::Maxwell
|
|
@ -0,0 +1,60 @@
|
|||
// 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/common_funcs.h"
|
||||
#include "shader_recompiler/frontend/maxwell/translate/impl/impl.h"
|
||||
|
||||
namespace Shader::Maxwell {
|
||||
namespace {
|
||||
void FSETP(TranslatorVisitor& v, u64 insn, const IR::F32& src_b) {
|
||||
union {
|
||||
u64 insn;
|
||||
BitField<0, 3, IR::Pred> dest_pred_b;
|
||||
BitField<3, 3, IR::Pred> dest_pred_a;
|
||||
BitField<6, 1, u64> negate_b;
|
||||
BitField<7, 1, u64> abs_a;
|
||||
BitField<8, 8, IR::Reg> src_a_reg;
|
||||
BitField<39, 3, IR::Pred> bop_pred;
|
||||
BitField<42, 1, u64> neg_bop_pred;
|
||||
BitField<43, 1, u64> negate_a;
|
||||
BitField<44, 1, u64> abs_b;
|
||||
BitField<45, 2, BooleanOp> bop;
|
||||
BitField<47, 1, u64> ftz;
|
||||
BitField<48, 4, FPCompareOp> compare_op;
|
||||
} const fsetp{insn};
|
||||
|
||||
const IR::F32 op_a{v.ir.FPAbsNeg(v.F(fsetp.src_a_reg), fsetp.abs_a != 0, fsetp.negate_a != 0)};
|
||||
const IR::F32 op_b = v.ir.FPAbsNeg(src_b, fsetp.abs_b != 0, fsetp.negate_b != 0);
|
||||
const IR::FpControl control{
|
||||
.no_contraction{false},
|
||||
.rounding{IR::FpRounding::DontCare},
|
||||
.fmz_mode{fsetp.ftz != 0 ? IR::FmzMode::FTZ : IR::FmzMode::None},
|
||||
};
|
||||
|
||||
const BooleanOp bop{fsetp.bop};
|
||||
const FPCompareOp compare_op{fsetp.compare_op};
|
||||
const IR::U1 comparison{FloatingPointCompare(v.ir, op_a, op_b, compare_op, control)};
|
||||
const IR::U1 bop_pred{v.ir.GetPred(fsetp.bop_pred, fsetp.neg_bop_pred != 0)};
|
||||
const IR::U1 result_a{PredicateCombine(v.ir, comparison, bop_pred, bop)};
|
||||
const IR::U1 result_b{PredicateCombine(v.ir, v.ir.LogicalNot(comparison), bop_pred, bop)};
|
||||
v.ir.SetPred(fsetp.dest_pred_a, result_a);
|
||||
v.ir.SetPred(fsetp.dest_pred_b, result_b);
|
||||
}
|
||||
} // Anonymous namespace
|
||||
|
||||
void TranslatorVisitor::FSETP_reg(u64 insn) {
|
||||
FSETP(*this, insn, GetFloatReg20(insn));
|
||||
}
|
||||
|
||||
void TranslatorVisitor::FSETP_cbuf(u64 insn) {
|
||||
FSETP(*this, insn, GetFloatCbuf(insn));
|
||||
}
|
||||
|
||||
void TranslatorVisitor::FSETP_imm(u64 insn) {
|
||||
FSETP(*this, insn, GetFloatImm20(insn));
|
||||
}
|
||||
|
||||
} // namespace Shader::Maxwell
|
|
@ -35,6 +35,25 @@ enum class PredicateOp : u64 {
|
|||
NonZero,
|
||||
};
|
||||
|
||||
enum class FPCompareOp : u64 {
|
||||
F,
|
||||
LT,
|
||||
EQ,
|
||||
LE,
|
||||
GT,
|
||||
NE,
|
||||
GE,
|
||||
NUM,
|
||||
Nan,
|
||||
LTU,
|
||||
EQU,
|
||||
LEU,
|
||||
GTU,
|
||||
NEU,
|
||||
GEU,
|
||||
T,
|
||||
};
|
||||
|
||||
class TranslatorVisitor {
|
||||
public:
|
||||
explicit TranslatorVisitor(Environment& env_, IR::Block& block) : env{env_}, ir(block) {}
|
||||
|
|
|
@ -201,30 +201,6 @@ void TranslatorVisitor::FCHK_imm(u64) {
|
|||
ThrowNotImplemented(Opcode::FCHK_imm);
|
||||
}
|
||||
|
||||
void TranslatorVisitor::FSET_reg(u64) {
|
||||
ThrowNotImplemented(Opcode::FSET_reg);
|
||||
}
|
||||
|
||||
void TranslatorVisitor::FSET_cbuf(u64) {
|
||||
ThrowNotImplemented(Opcode::FSET_cbuf);
|
||||
}
|
||||
|
||||
void TranslatorVisitor::FSET_imm(u64) {
|
||||
ThrowNotImplemented(Opcode::FSET_imm);
|
||||
}
|
||||
|
||||
void TranslatorVisitor::FSETP_reg(u64) {
|
||||
ThrowNotImplemented(Opcode::FSETP_reg);
|
||||
}
|
||||
|
||||
void TranslatorVisitor::FSETP_cbuf(u64) {
|
||||
ThrowNotImplemented(Opcode::FSETP_cbuf);
|
||||
}
|
||||
|
||||
void TranslatorVisitor::FSETP_imm(u64) {
|
||||
ThrowNotImplemented(Opcode::FSETP_imm);
|
||||
}
|
||||
|
||||
void TranslatorVisitor::FSWZADD(u64) {
|
||||
ThrowNotImplemented(Opcode::FSWZADD);
|
||||
}
|
||||
|
|
Loading…
Reference in a new issue