mirror of
https://github.com/dolphin-emu/dolphin.git
synced 2025-07-21 05:09:34 -06:00
Common: Move floating-point utility functions to FloatUtils.h/.cpp
Keeps all of the floating-point utility functions in their own file to keep them all together. This also provides a place for other general-purpose floating-point functions to be added in the future, which will be necessary when improving the flag-setting within the interpreter.
This commit is contained in:
@ -14,6 +14,7 @@ add_library(common
|
||||
File.cpp
|
||||
FileSearch.cpp
|
||||
FileUtil.cpp
|
||||
FloatUtils.cpp
|
||||
GekkoDisassembler.cpp
|
||||
Hash.cpp
|
||||
HttpRequest.cpp
|
||||
|
@ -122,6 +122,7 @@
|
||||
<ClInclude Include="GL\GLInterfaceBase.h" />
|
||||
<ClInclude Include="GL\GLInterface\WGL.h" />
|
||||
<ClInclude Include="GL\GLUtil.h" />
|
||||
<ClInclude Include="FloatUtils.h" />
|
||||
<ClInclude Include="Hash.h" />
|
||||
<ClInclude Include="HttpRequest.h" />
|
||||
<ClInclude Include="IniFile.h" />
|
||||
@ -179,6 +180,7 @@
|
||||
<ClCompile Include="File.cpp" />
|
||||
<ClCompile Include="FileSearch.cpp" />
|
||||
<ClCompile Include="FileUtil.cpp" />
|
||||
<ClCompile Include="FloatUtils.cpp" />
|
||||
<ClCompile Include="GekkoDisassembler.cpp" />
|
||||
<ClCompile Include="GL\GLExtensions\GLExtensions.cpp" />
|
||||
<ClCompile Include="GL\GLInterface\GLInterface.cpp" />
|
||||
|
@ -47,6 +47,7 @@
|
||||
<ClInclude Include="FileUtil.h" />
|
||||
<ClInclude Include="FixedSizeQueue.h" />
|
||||
<ClInclude Include="Flag.h" />
|
||||
<ClInclude Include="FloatUtils.h" />
|
||||
<ClInclude Include="FPURoundMode.h" />
|
||||
<ClInclude Include="Hash.h" />
|
||||
<ClInclude Include="HttpRequest.h" />
|
||||
@ -277,6 +278,7 @@
|
||||
<ClCompile Include="ENetUtil.cpp" />
|
||||
<ClCompile Include="FileSearch.cpp" />
|
||||
<ClCompile Include="FileUtil.cpp" />
|
||||
<ClCompile Include="FloatUtils.cpp" />
|
||||
<ClCompile Include="Hash.cpp" />
|
||||
<ClCompile Include="HttpRequest.cpp" />
|
||||
<ClCompile Include="IniFile.cpp" />
|
||||
|
216
Source/Core/Common/FloatUtils.cpp
Normal file
216
Source/Core/Common/FloatUtils.cpp
Normal file
@ -0,0 +1,216 @@
|
||||
// Copyright 2018 Dolphin Emulator Project
|
||||
// Licensed under GPLv2+
|
||||
// Refer to the license.txt file included.
|
||||
|
||||
#include "Common/FloatUtils.h"
|
||||
|
||||
#include <cmath>
|
||||
|
||||
namespace Common
|
||||
{
|
||||
u32 ClassifyDouble(double dvalue)
|
||||
{
|
||||
// TODO: Optimize the below to be as fast as possible.
|
||||
IntDouble value(dvalue);
|
||||
u64 sign = value.i & DOUBLE_SIGN;
|
||||
u64 exp = value.i & DOUBLE_EXP;
|
||||
if (exp > DOUBLE_ZERO && exp < DOUBLE_EXP)
|
||||
{
|
||||
// Nice normalized number.
|
||||
return sign ? PPC_FPCLASS_NN : PPC_FPCLASS_PN;
|
||||
}
|
||||
else
|
||||
{
|
||||
u64 mantissa = value.i & DOUBLE_FRAC;
|
||||
if (mantissa)
|
||||
{
|
||||
if (exp)
|
||||
{
|
||||
return PPC_FPCLASS_QNAN;
|
||||
}
|
||||
else
|
||||
{
|
||||
// Denormalized number.
|
||||
return sign ? PPC_FPCLASS_ND : PPC_FPCLASS_PD;
|
||||
}
|
||||
}
|
||||
else if (exp)
|
||||
{
|
||||
// Infinite
|
||||
return sign ? PPC_FPCLASS_NINF : PPC_FPCLASS_PINF;
|
||||
}
|
||||
else
|
||||
{
|
||||
// Zero
|
||||
return sign ? PPC_FPCLASS_NZ : PPC_FPCLASS_PZ;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
u32 ClassifyFloat(float fvalue)
|
||||
{
|
||||
// TODO: Optimize the below to be as fast as possible.
|
||||
IntFloat value(fvalue);
|
||||
u32 sign = value.i & FLOAT_SIGN;
|
||||
u32 exp = value.i & FLOAT_EXP;
|
||||
if (exp > FLOAT_ZERO && exp < FLOAT_EXP)
|
||||
{
|
||||
// Nice normalized number.
|
||||
return sign ? PPC_FPCLASS_NN : PPC_FPCLASS_PN;
|
||||
}
|
||||
else
|
||||
{
|
||||
u32 mantissa = value.i & FLOAT_FRAC;
|
||||
if (mantissa)
|
||||
{
|
||||
if (exp)
|
||||
{
|
||||
return PPC_FPCLASS_QNAN; // Quiet NAN
|
||||
}
|
||||
else
|
||||
{
|
||||
// Denormalized number.
|
||||
return sign ? PPC_FPCLASS_ND : PPC_FPCLASS_PD;
|
||||
}
|
||||
}
|
||||
else if (exp)
|
||||
{
|
||||
// Infinite
|
||||
return sign ? PPC_FPCLASS_NINF : PPC_FPCLASS_PINF;
|
||||
}
|
||||
else
|
||||
{
|
||||
// Zero
|
||||
return sign ? PPC_FPCLASS_NZ : PPC_FPCLASS_PZ;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
const std::array<BaseAndDec, 32> frsqrte_expected = {{
|
||||
{0x3ffa000, 0x7a4}, {0x3c29000, 0x700}, {0x38aa000, 0x670}, {0x3572000, 0x5f2},
|
||||
{0x3279000, 0x584}, {0x2fb7000, 0x524}, {0x2d26000, 0x4cc}, {0x2ac0000, 0x47e},
|
||||
{0x2881000, 0x43a}, {0x2665000, 0x3fa}, {0x2468000, 0x3c2}, {0x2287000, 0x38e},
|
||||
{0x20c1000, 0x35e}, {0x1f12000, 0x332}, {0x1d79000, 0x30a}, {0x1bf4000, 0x2e6},
|
||||
{0x1a7e800, 0x568}, {0x17cb800, 0x4f3}, {0x1552800, 0x48d}, {0x130c000, 0x435},
|
||||
{0x10f2000, 0x3e7}, {0x0eff000, 0x3a2}, {0x0d2e000, 0x365}, {0x0b7c000, 0x32e},
|
||||
{0x09e5000, 0x2fc}, {0x0867000, 0x2d0}, {0x06ff000, 0x2a8}, {0x05ab800, 0x283},
|
||||
{0x046a000, 0x261}, {0x0339800, 0x243}, {0x0218800, 0x226}, {0x0105800, 0x20b},
|
||||
}};
|
||||
|
||||
double ApproximateReciprocalSquareRoot(double val)
|
||||
{
|
||||
union
|
||||
{
|
||||
double valf;
|
||||
s64 vali;
|
||||
};
|
||||
valf = val;
|
||||
s64 mantissa = vali & ((1LL << 52) - 1);
|
||||
s64 sign = vali & (1ULL << 63);
|
||||
s64 exponent = vali & (0x7FFLL << 52);
|
||||
|
||||
// Special case 0
|
||||
if (mantissa == 0 && exponent == 0)
|
||||
return sign ? -std::numeric_limits<double>::infinity() :
|
||||
std::numeric_limits<double>::infinity();
|
||||
// Special case NaN-ish numbers
|
||||
if (exponent == (0x7FFLL << 52))
|
||||
{
|
||||
if (mantissa == 0)
|
||||
{
|
||||
if (sign)
|
||||
return std::numeric_limits<double>::quiet_NaN();
|
||||
|
||||
return 0.0;
|
||||
}
|
||||
|
||||
return 0.0 + valf;
|
||||
}
|
||||
|
||||
// Negative numbers return NaN
|
||||
if (sign)
|
||||
return std::numeric_limits<double>::quiet_NaN();
|
||||
|
||||
if (!exponent)
|
||||
{
|
||||
// "Normalize" denormal values
|
||||
do
|
||||
{
|
||||
exponent -= 1LL << 52;
|
||||
mantissa <<= 1;
|
||||
} while (!(mantissa & (1LL << 52)));
|
||||
mantissa &= (1LL << 52) - 1;
|
||||
exponent += 1LL << 52;
|
||||
}
|
||||
|
||||
bool odd_exponent = !(exponent & (1LL << 52));
|
||||
exponent = ((0x3FFLL << 52) - ((exponent - (0x3FELL << 52)) / 2)) & (0x7FFLL << 52);
|
||||
|
||||
int i = (int)(mantissa >> 37);
|
||||
vali = sign | exponent;
|
||||
int index = i / 2048 + (odd_exponent ? 16 : 0);
|
||||
const auto& entry = frsqrte_expected[index];
|
||||
vali |= (s64)(entry.m_base - entry.m_dec * (i % 2048)) << 26;
|
||||
return valf;
|
||||
}
|
||||
|
||||
const std::array<BaseAndDec, 32> fres_expected = {{
|
||||
{0x7ff800, 0x3e1}, {0x783800, 0x3a7}, {0x70ea00, 0x371}, {0x6a0800, 0x340}, {0x638800, 0x313},
|
||||
{0x5d6200, 0x2ea}, {0x579000, 0x2c4}, {0x520800, 0x2a0}, {0x4cc800, 0x27f}, {0x47ca00, 0x261},
|
||||
{0x430800, 0x245}, {0x3e8000, 0x22a}, {0x3a2c00, 0x212}, {0x360800, 0x1fb}, {0x321400, 0x1e5},
|
||||
{0x2e4a00, 0x1d1}, {0x2aa800, 0x1be}, {0x272c00, 0x1ac}, {0x23d600, 0x19b}, {0x209e00, 0x18b},
|
||||
{0x1d8800, 0x17c}, {0x1a9000, 0x16e}, {0x17ae00, 0x15b}, {0x14f800, 0x15b}, {0x124400, 0x143},
|
||||
{0x0fbe00, 0x143}, {0x0d3800, 0x12d}, {0x0ade00, 0x12d}, {0x088400, 0x11a}, {0x065000, 0x11a},
|
||||
{0x041c00, 0x108}, {0x020c00, 0x106},
|
||||
}};
|
||||
|
||||
// Used by fres and ps_res.
|
||||
double ApproximateReciprocal(double val)
|
||||
{
|
||||
// We are using namespace std scoped here because the Android NDK is complete trash as usual
|
||||
// For 32bit targets(mips, ARMv7, x86) it doesn't provide an implementation of std::copysign
|
||||
// but instead provides just global namespace copysign implementations.
|
||||
// The workaround for this is to just use namespace std within this function's scope
|
||||
// That way on real toolchains it will use the std:: variant like normal.
|
||||
using namespace std;
|
||||
union
|
||||
{
|
||||
double valf;
|
||||
s64 vali;
|
||||
};
|
||||
|
||||
valf = val;
|
||||
s64 mantissa = vali & ((1LL << 52) - 1);
|
||||
s64 sign = vali & (1ULL << 63);
|
||||
s64 exponent = vali & (0x7FFLL << 52);
|
||||
|
||||
// Special case 0
|
||||
if (mantissa == 0 && exponent == 0)
|
||||
return copysign(std::numeric_limits<double>::infinity(), valf);
|
||||
|
||||
// Special case NaN-ish numbers
|
||||
if (exponent == (0x7FFLL << 52))
|
||||
{
|
||||
if (mantissa == 0)
|
||||
return copysign(0.0, valf);
|
||||
return 0.0 + valf;
|
||||
}
|
||||
|
||||
// Special case small inputs
|
||||
if (exponent < (895LL << 52))
|
||||
return copysign(std::numeric_limits<float>::max(), valf);
|
||||
|
||||
// Special case large inputs
|
||||
if (exponent >= (1149LL << 52))
|
||||
return copysign(0.0, valf);
|
||||
|
||||
exponent = (0x7FDLL << 52) - exponent;
|
||||
|
||||
int i = (int)(mantissa >> 37);
|
||||
const auto& entry = fres_expected[i / 1024];
|
||||
vali = sign | exponent;
|
||||
vali |= (s64)(entry.m_base - (entry.m_dec * (i % 1024) + 1) / 2) << 29;
|
||||
return valf;
|
||||
}
|
||||
|
||||
} // namespace Common
|
139
Source/Core/Common/FloatUtils.h
Normal file
139
Source/Core/Common/FloatUtils.h
Normal file
@ -0,0 +1,139 @@
|
||||
// Copyright 2018 Dolphin Emulator Project
|
||||
// Licensed under GPLv2+
|
||||
// Refer to the license.txt file included.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <array>
|
||||
#include <limits>
|
||||
|
||||
#include "Common/CommonTypes.h"
|
||||
|
||||
namespace Common
|
||||
{
|
||||
template <typename T>
|
||||
constexpr T SNANConstant()
|
||||
{
|
||||
return std::numeric_limits<T>::signaling_NaN();
|
||||
}
|
||||
|
||||
#ifdef _MSC_VER
|
||||
|
||||
// MSVC needs a workaround, because its std::numeric_limits<double>::signaling_NaN()
|
||||
// will use __builtin_nans, which is improperly handled by the compiler and generates
|
||||
// a bad constant. Here we go back to the version MSVC used before the builtin.
|
||||
// TODO: Remove this and use numeric_limits directly whenever this bug is fixed.
|
||||
|
||||
template <>
|
||||
constexpr double SNANConstant()
|
||||
{
|
||||
return (_CSTD _Snan._Double);
|
||||
}
|
||||
template <>
|
||||
constexpr float SNANConstant()
|
||||
{
|
||||
return (_CSTD _Snan._Float);
|
||||
}
|
||||
|
||||
#endif
|
||||
|
||||
// The most significant bit of the fraction is an is-quiet bit on all architectures we care about.
|
||||
enum : u64
|
||||
{
|
||||
DOUBLE_SIGN = 0x8000000000000000ULL,
|
||||
DOUBLE_EXP = 0x7FF0000000000000ULL,
|
||||
DOUBLE_FRAC = 0x000FFFFFFFFFFFFFULL,
|
||||
DOUBLE_ZERO = 0x0000000000000000ULL,
|
||||
DOUBLE_QBIT = 0x0008000000000000ULL
|
||||
};
|
||||
|
||||
enum : u32
|
||||
{
|
||||
FLOAT_SIGN = 0x80000000,
|
||||
FLOAT_EXP = 0x7F800000,
|
||||
FLOAT_FRAC = 0x007FFFFF,
|
||||
FLOAT_ZERO = 0x00000000
|
||||
};
|
||||
|
||||
union IntDouble
|
||||
{
|
||||
double d;
|
||||
u64 i;
|
||||
|
||||
explicit IntDouble(u64 _i) : i(_i) {}
|
||||
explicit IntDouble(double _d) : d(_d) {}
|
||||
};
|
||||
union IntFloat
|
||||
{
|
||||
float f;
|
||||
u32 i;
|
||||
|
||||
explicit IntFloat(u32 _i) : i(_i) {}
|
||||
explicit IntFloat(float _f) : f(_f) {}
|
||||
};
|
||||
|
||||
inline bool IsQNAN(double d)
|
||||
{
|
||||
IntDouble x(d);
|
||||
return ((x.i & DOUBLE_EXP) == DOUBLE_EXP) && ((x.i & DOUBLE_QBIT) == DOUBLE_QBIT);
|
||||
}
|
||||
|
||||
inline bool IsSNAN(double d)
|
||||
{
|
||||
IntDouble x(d);
|
||||
return ((x.i & DOUBLE_EXP) == DOUBLE_EXP) && ((x.i & DOUBLE_FRAC) != DOUBLE_ZERO) &&
|
||||
((x.i & DOUBLE_QBIT) == DOUBLE_ZERO);
|
||||
}
|
||||
|
||||
inline float FlushToZero(float f)
|
||||
{
|
||||
IntFloat x(f);
|
||||
if ((x.i & FLOAT_EXP) == 0)
|
||||
{
|
||||
x.i &= FLOAT_SIGN; // turn into signed zero
|
||||
}
|
||||
return x.f;
|
||||
}
|
||||
|
||||
inline double FlushToZero(double d)
|
||||
{
|
||||
IntDouble x(d);
|
||||
if ((x.i & DOUBLE_EXP) == 0)
|
||||
{
|
||||
x.i &= DOUBLE_SIGN; // turn into signed zero
|
||||
}
|
||||
return x.d;
|
||||
}
|
||||
|
||||
enum PPCFpClass
|
||||
{
|
||||
PPC_FPCLASS_QNAN = 0x11,
|
||||
PPC_FPCLASS_NINF = 0x9,
|
||||
PPC_FPCLASS_NN = 0x8,
|
||||
PPC_FPCLASS_ND = 0x18,
|
||||
PPC_FPCLASS_NZ = 0x12,
|
||||
PPC_FPCLASS_PZ = 0x2,
|
||||
PPC_FPCLASS_PD = 0x14,
|
||||
PPC_FPCLASS_PN = 0x4,
|
||||
PPC_FPCLASS_PINF = 0x5,
|
||||
};
|
||||
|
||||
// Uses PowerPC conventions for the return value, so it can be easily
|
||||
// used directly in CPU emulation.
|
||||
u32 ClassifyDouble(double dvalue);
|
||||
// More efficient float version.
|
||||
u32 ClassifyFloat(float fvalue);
|
||||
|
||||
struct BaseAndDec
|
||||
{
|
||||
int m_base;
|
||||
int m_dec;
|
||||
};
|
||||
extern const std::array<BaseAndDec, 32> frsqrte_expected;
|
||||
extern const std::array<BaseAndDec, 32> fres_expected;
|
||||
|
||||
// PowerPC approximation algorithms
|
||||
double ApproximateReciprocalSquareRoot(double val);
|
||||
double ApproximateReciprocal(double val);
|
||||
|
||||
} // namespace Common
|
@ -10,215 +10,6 @@
|
||||
#include "Common/CommonTypes.h"
|
||||
#include "Common/MathUtil.h"
|
||||
|
||||
namespace MathUtil
|
||||
{
|
||||
u32 ClassifyDouble(double dvalue)
|
||||
{
|
||||
// TODO: Optimize the below to be as fast as possible.
|
||||
IntDouble value(dvalue);
|
||||
u64 sign = value.i & DOUBLE_SIGN;
|
||||
u64 exp = value.i & DOUBLE_EXP;
|
||||
if (exp > DOUBLE_ZERO && exp < DOUBLE_EXP)
|
||||
{
|
||||
// Nice normalized number.
|
||||
return sign ? PPC_FPCLASS_NN : PPC_FPCLASS_PN;
|
||||
}
|
||||
else
|
||||
{
|
||||
u64 mantissa = value.i & DOUBLE_FRAC;
|
||||
if (mantissa)
|
||||
{
|
||||
if (exp)
|
||||
{
|
||||
return PPC_FPCLASS_QNAN;
|
||||
}
|
||||
else
|
||||
{
|
||||
// Denormalized number.
|
||||
return sign ? PPC_FPCLASS_ND : PPC_FPCLASS_PD;
|
||||
}
|
||||
}
|
||||
else if (exp)
|
||||
{
|
||||
// Infinite
|
||||
return sign ? PPC_FPCLASS_NINF : PPC_FPCLASS_PINF;
|
||||
}
|
||||
else
|
||||
{
|
||||
// Zero
|
||||
return sign ? PPC_FPCLASS_NZ : PPC_FPCLASS_PZ;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
u32 ClassifyFloat(float fvalue)
|
||||
{
|
||||
// TODO: Optimize the below to be as fast as possible.
|
||||
IntFloat value(fvalue);
|
||||
u32 sign = value.i & FLOAT_SIGN;
|
||||
u32 exp = value.i & FLOAT_EXP;
|
||||
if (exp > FLOAT_ZERO && exp < FLOAT_EXP)
|
||||
{
|
||||
// Nice normalized number.
|
||||
return sign ? PPC_FPCLASS_NN : PPC_FPCLASS_PN;
|
||||
}
|
||||
else
|
||||
{
|
||||
u32 mantissa = value.i & FLOAT_FRAC;
|
||||
if (mantissa)
|
||||
{
|
||||
if (exp)
|
||||
{
|
||||
return PPC_FPCLASS_QNAN; // Quiet NAN
|
||||
}
|
||||
else
|
||||
{
|
||||
// Denormalized number.
|
||||
return sign ? PPC_FPCLASS_ND : PPC_FPCLASS_PD;
|
||||
}
|
||||
}
|
||||
else if (exp)
|
||||
{
|
||||
// Infinite
|
||||
return sign ? PPC_FPCLASS_NINF : PPC_FPCLASS_PINF;
|
||||
}
|
||||
else
|
||||
{
|
||||
// Zero
|
||||
return sign ? PPC_FPCLASS_NZ : PPC_FPCLASS_PZ;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
const std::array<BaseAndDec, 32> frsqrte_expected = {{
|
||||
{0x3ffa000, 0x7a4}, {0x3c29000, 0x700}, {0x38aa000, 0x670}, {0x3572000, 0x5f2},
|
||||
{0x3279000, 0x584}, {0x2fb7000, 0x524}, {0x2d26000, 0x4cc}, {0x2ac0000, 0x47e},
|
||||
{0x2881000, 0x43a}, {0x2665000, 0x3fa}, {0x2468000, 0x3c2}, {0x2287000, 0x38e},
|
||||
{0x20c1000, 0x35e}, {0x1f12000, 0x332}, {0x1d79000, 0x30a}, {0x1bf4000, 0x2e6},
|
||||
{0x1a7e800, 0x568}, {0x17cb800, 0x4f3}, {0x1552800, 0x48d}, {0x130c000, 0x435},
|
||||
{0x10f2000, 0x3e7}, {0x0eff000, 0x3a2}, {0x0d2e000, 0x365}, {0x0b7c000, 0x32e},
|
||||
{0x09e5000, 0x2fc}, {0x0867000, 0x2d0}, {0x06ff000, 0x2a8}, {0x05ab800, 0x283},
|
||||
{0x046a000, 0x261}, {0x0339800, 0x243}, {0x0218800, 0x226}, {0x0105800, 0x20b},
|
||||
}};
|
||||
|
||||
double ApproximateReciprocalSquareRoot(double val)
|
||||
{
|
||||
union
|
||||
{
|
||||
double valf;
|
||||
s64 vali;
|
||||
};
|
||||
valf = val;
|
||||
s64 mantissa = vali & ((1LL << 52) - 1);
|
||||
s64 sign = vali & (1ULL << 63);
|
||||
s64 exponent = vali & (0x7FFLL << 52);
|
||||
|
||||
// Special case 0
|
||||
if (mantissa == 0 && exponent == 0)
|
||||
return sign ? -std::numeric_limits<double>::infinity() :
|
||||
std::numeric_limits<double>::infinity();
|
||||
// Special case NaN-ish numbers
|
||||
if (exponent == (0x7FFLL << 52))
|
||||
{
|
||||
if (mantissa == 0)
|
||||
{
|
||||
if (sign)
|
||||
return std::numeric_limits<double>::quiet_NaN();
|
||||
|
||||
return 0.0;
|
||||
}
|
||||
|
||||
return 0.0 + valf;
|
||||
}
|
||||
|
||||
// Negative numbers return NaN
|
||||
if (sign)
|
||||
return std::numeric_limits<double>::quiet_NaN();
|
||||
|
||||
if (!exponent)
|
||||
{
|
||||
// "Normalize" denormal values
|
||||
do
|
||||
{
|
||||
exponent -= 1LL << 52;
|
||||
mantissa <<= 1;
|
||||
} while (!(mantissa & (1LL << 52)));
|
||||
mantissa &= (1LL << 52) - 1;
|
||||
exponent += 1LL << 52;
|
||||
}
|
||||
|
||||
bool odd_exponent = !(exponent & (1LL << 52));
|
||||
exponent = ((0x3FFLL << 52) - ((exponent - (0x3FELL << 52)) / 2)) & (0x7FFLL << 52);
|
||||
|
||||
int i = (int)(mantissa >> 37);
|
||||
vali = sign | exponent;
|
||||
int index = i / 2048 + (odd_exponent ? 16 : 0);
|
||||
const auto& entry = frsqrte_expected[index];
|
||||
vali |= (s64)(entry.m_base - entry.m_dec * (i % 2048)) << 26;
|
||||
return valf;
|
||||
}
|
||||
|
||||
const std::array<BaseAndDec, 32> fres_expected = {{
|
||||
{0x7ff800, 0x3e1}, {0x783800, 0x3a7}, {0x70ea00, 0x371}, {0x6a0800, 0x340}, {0x638800, 0x313},
|
||||
{0x5d6200, 0x2ea}, {0x579000, 0x2c4}, {0x520800, 0x2a0}, {0x4cc800, 0x27f}, {0x47ca00, 0x261},
|
||||
{0x430800, 0x245}, {0x3e8000, 0x22a}, {0x3a2c00, 0x212}, {0x360800, 0x1fb}, {0x321400, 0x1e5},
|
||||
{0x2e4a00, 0x1d1}, {0x2aa800, 0x1be}, {0x272c00, 0x1ac}, {0x23d600, 0x19b}, {0x209e00, 0x18b},
|
||||
{0x1d8800, 0x17c}, {0x1a9000, 0x16e}, {0x17ae00, 0x15b}, {0x14f800, 0x15b}, {0x124400, 0x143},
|
||||
{0x0fbe00, 0x143}, {0x0d3800, 0x12d}, {0x0ade00, 0x12d}, {0x088400, 0x11a}, {0x065000, 0x11a},
|
||||
{0x041c00, 0x108}, {0x020c00, 0x106},
|
||||
}};
|
||||
|
||||
// Used by fres and ps_res.
|
||||
double ApproximateReciprocal(double val)
|
||||
{
|
||||
// We are using namespace std scoped here because the Android NDK is complete trash as usual
|
||||
// For 32bit targets(mips, ARMv7, x86) it doesn't provide an implementation of std::copysign
|
||||
// but instead provides just global namespace copysign implementations.
|
||||
// The workaround for this is to just use namespace std within this function's scope
|
||||
// That way on real toolchains it will use the std:: variant like normal.
|
||||
using namespace std;
|
||||
union
|
||||
{
|
||||
double valf;
|
||||
s64 vali;
|
||||
};
|
||||
|
||||
valf = val;
|
||||
s64 mantissa = vali & ((1LL << 52) - 1);
|
||||
s64 sign = vali & (1ULL << 63);
|
||||
s64 exponent = vali & (0x7FFLL << 52);
|
||||
|
||||
// Special case 0
|
||||
if (mantissa == 0 && exponent == 0)
|
||||
return copysign(std::numeric_limits<double>::infinity(), valf);
|
||||
|
||||
// Special case NaN-ish numbers
|
||||
if (exponent == (0x7FFLL << 52))
|
||||
{
|
||||
if (mantissa == 0)
|
||||
return copysign(0.0, valf);
|
||||
return 0.0 + valf;
|
||||
}
|
||||
|
||||
// Special case small inputs
|
||||
if (exponent < (895LL << 52))
|
||||
return copysign(std::numeric_limits<float>::max(), valf);
|
||||
|
||||
// Special case large inputs
|
||||
if (exponent >= (1149LL << 52))
|
||||
return copysign(0.0, valf);
|
||||
|
||||
exponent = (0x7FDLL << 52) - exponent;
|
||||
|
||||
int i = (int)(mantissa >> 37);
|
||||
const auto& entry = fres_expected[i / 1024];
|
||||
vali = sign | exponent;
|
||||
vali |= (s64)(entry.m_base - (entry.m_dec * (i % 1024) + 1) / 2) << 29;
|
||||
return valf;
|
||||
}
|
||||
|
||||
} // namespace
|
||||
|
||||
inline void MatrixMul(int n, const float* a, const float* b, float* result)
|
||||
{
|
||||
for (int i = 0; i < n; ++i)
|
||||
|
@ -17,32 +17,6 @@
|
||||
|
||||
namespace MathUtil
|
||||
{
|
||||
template <typename T>
|
||||
constexpr T SNANConstant()
|
||||
{
|
||||
return std::numeric_limits<T>::signaling_NaN();
|
||||
}
|
||||
|
||||
#ifdef _MSC_VER
|
||||
|
||||
// MSVC needs a workaround, because its std::numeric_limits<double>::signaling_NaN()
|
||||
// will use __builtin_nans, which is improperly handled by the compiler and generates
|
||||
// a bad constant. Here we go back to the version MSVC used before the builtin.
|
||||
// TODO: Remove this and use numeric_limits directly whenever this bug is fixed.
|
||||
|
||||
template <>
|
||||
constexpr double SNANConstant()
|
||||
{
|
||||
return (_CSTD _Snan._Double);
|
||||
}
|
||||
template <>
|
||||
constexpr float SNANConstant()
|
||||
{
|
||||
return (_CSTD _Snan._Float);
|
||||
}
|
||||
|
||||
#endif
|
||||
|
||||
template <class T>
|
||||
constexpr T Clamp(const T val, const T& min, const T& max)
|
||||
{
|
||||
@ -55,96 +29,6 @@ constexpr bool IsPow2(T imm)
|
||||
return imm > 0 && (imm & (imm - 1)) == 0;
|
||||
}
|
||||
|
||||
// The most significant bit of the fraction is an is-quiet bit on all architectures we care about.
|
||||
|
||||
static const u64 DOUBLE_SIGN = 0x8000000000000000ULL, DOUBLE_EXP = 0x7FF0000000000000ULL,
|
||||
DOUBLE_FRAC = 0x000FFFFFFFFFFFFFULL, DOUBLE_ZERO = 0x0000000000000000ULL,
|
||||
DOUBLE_QBIT = 0x0008000000000000ULL;
|
||||
|
||||
static const u32 FLOAT_SIGN = 0x80000000, FLOAT_EXP = 0x7F800000, FLOAT_FRAC = 0x007FFFFF,
|
||||
FLOAT_ZERO = 0x00000000;
|
||||
|
||||
union IntDouble
|
||||
{
|
||||
double d;
|
||||
u64 i;
|
||||
|
||||
explicit IntDouble(u64 _i) : i(_i) {}
|
||||
explicit IntDouble(double _d) : d(_d) {}
|
||||
};
|
||||
union IntFloat
|
||||
{
|
||||
float f;
|
||||
u32 i;
|
||||
|
||||
explicit IntFloat(u32 _i) : i(_i) {}
|
||||
explicit IntFloat(float _f) : f(_f) {}
|
||||
};
|
||||
|
||||
inline bool IsQNAN(double d)
|
||||
{
|
||||
IntDouble x(d);
|
||||
return ((x.i & DOUBLE_EXP) == DOUBLE_EXP) && ((x.i & DOUBLE_QBIT) == DOUBLE_QBIT);
|
||||
}
|
||||
|
||||
inline bool IsSNAN(double d)
|
||||
{
|
||||
IntDouble x(d);
|
||||
return ((x.i & DOUBLE_EXP) == DOUBLE_EXP) && ((x.i & DOUBLE_FRAC) != DOUBLE_ZERO) &&
|
||||
((x.i & DOUBLE_QBIT) == DOUBLE_ZERO);
|
||||
}
|
||||
|
||||
inline float FlushToZero(float f)
|
||||
{
|
||||
IntFloat x(f);
|
||||
if ((x.i & FLOAT_EXP) == 0)
|
||||
{
|
||||
x.i &= FLOAT_SIGN; // turn into signed zero
|
||||
}
|
||||
return x.f;
|
||||
}
|
||||
|
||||
inline double FlushToZero(double d)
|
||||
{
|
||||
IntDouble x(d);
|
||||
if ((x.i & DOUBLE_EXP) == 0)
|
||||
{
|
||||
x.i &= DOUBLE_SIGN; // turn into signed zero
|
||||
}
|
||||
return x.d;
|
||||
}
|
||||
|
||||
enum PPCFpClass
|
||||
{
|
||||
PPC_FPCLASS_QNAN = 0x11,
|
||||
PPC_FPCLASS_NINF = 0x9,
|
||||
PPC_FPCLASS_NN = 0x8,
|
||||
PPC_FPCLASS_ND = 0x18,
|
||||
PPC_FPCLASS_NZ = 0x12,
|
||||
PPC_FPCLASS_PZ = 0x2,
|
||||
PPC_FPCLASS_PD = 0x14,
|
||||
PPC_FPCLASS_PN = 0x4,
|
||||
PPC_FPCLASS_PINF = 0x5,
|
||||
};
|
||||
|
||||
// Uses PowerPC conventions for the return value, so it can be easily
|
||||
// used directly in CPU emulation.
|
||||
u32 ClassifyDouble(double dvalue);
|
||||
// More efficient float version.
|
||||
u32 ClassifyFloat(float fvalue);
|
||||
|
||||
struct BaseAndDec
|
||||
{
|
||||
int m_base;
|
||||
int m_dec;
|
||||
};
|
||||
extern const std::array<BaseAndDec, 32> frsqrte_expected;
|
||||
extern const std::array<BaseAndDec, 32> fres_expected;
|
||||
|
||||
// PowerPC approximation algorithms
|
||||
double ApproximateReciprocalSquareRoot(double val);
|
||||
double ApproximateReciprocal(double val);
|
||||
|
||||
template <class T>
|
||||
struct Rectangle
|
||||
{
|
||||
|
Reference in New Issue
Block a user