* Copyright (c) 2025 Huawei Technologies Co., Ltd.
* This program is free software, you can redistribute it and/or modify it under the terms and conditions of
* CANN Open Software License Agreement Version 2.0 (the "License").
* Please refer to the License for details. You may not use this file except in compliance with the License.
* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
* See LICENSE in the root of the software repository for the full text of the License.
* The code snippet comes from Huawei's open-source Mindspore project.
* Copyright 2019-2020 Huawei Technologies Co., Ltd.
* Licensed under the Apache License, Version 2.0 (the "License");
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*/
* \file fp16.cpp
* \brief Half precision float
*/
#include "op_common/op_host/util/fp16.h"
namespace Ops {
namespace Base {
constexpr uint16_t FP16_BIT_SHIFT_LEN_2 = 2;
constexpr uint16_t FP16_BIT_SHIFT_LEN_3 = 3;
constexpr uint16_t FP16_BIT_SHIFT_LEN_7 = 7;
constexpr uint16_t FP16_BIT_SHIFT_LEN_10 = 10;
constexpr uint16_t FP16_BIT_SHIFT_LEN_11 = 11;
constexpr uint16_t FP16_BIT_SHIFT_LEN_15 = 15;
constexpr uint16_t FP16_BIT_SHIFT_LEN_31 = 31;
constexpr uint16_t FP16_BIT_SHIFT_LEN_32 = 32;
* @ingroup fp16 basic parameter
* @brief fp16 exponent bias
*/
constexpr uint16_t FP16_EXP_BIAS = 15;
* @ingroup fp16 basic parameter
* @brief the mantissa bit length of fp16 is 10
*/
constexpr uint16_t FP16_MAN_LEN = 10;
* @ingroup fp16 basic parameter
* @brief bit index of sign in fp16
*/
constexpr uint16_t FP16_SIGN_INDEX = 15;
* @ingroup fp16 basic parameter
* @brief sign mask of fp16 (1 00000 00000 00000)
*/
constexpr uint16_t FP16_SIGN_MASK = 0x8000;
* @ingroup fp16 basic parameter
* @brief exponent mask of fp16 ( 11111 00000 00000)
*/
constexpr uint16_t FP16_EXP_MASK = 0x7C00;
* @ingroup fp16 basic parameter
* @brief mantissa mask of fp16 ( 11111 11111)
*/
constexpr uint16_t FP16_MAN_MASK = 0x03FF;
* @ingroup fp16 basic parameter
* @brief conceal bit of mantissa of fp16( 1 00000 00000)
*/
constexpr uint16_t FP16_MAN_HIDE_BIT = 0x0400;
* @ingroup fp16 basic parameter
* @brief maximum value (0111 1011 1111 1111)
*/
constexpr uint16_t FP16_MAX = 0x7BFF;
* @ingroup fp16 basic parameter
* @brief minimum value (1111 1011 1111 1111)
*/
constexpr uint16_t FP16_MIN = 0xFBFF;
* @ingroup fp16 basic parameter
* @brief absolute maximum value (0111 1111 1111 1111)
*/
constexpr uint16_t FP16_ABS_MAX = 0x7FFF;
* @ingroup fp16 basic parameter
* @brief maximum exponent value of fp16 is 15(11111)
*/
constexpr uint16_t FP16_MAX_EXP = 0x001F;
* @ingroup fp16 basic parameter
* @brief maximum mantissa value of fp16(11111 11111)
*/
constexpr uint16_t FP16_MAX_MAN = 0x03FF;
* @ingroup fp16 basic operator
* @brief get sign of fp16
*/
static inline uint16_t Fp16ExtractSign(uint16_t x) { return (x >> FP16_EXP_BIAS) & 1; }
* @ingroup fp16 basic operator
* @brief get exponent of fp16
*/
static inline uint16_t Fp16ExtractExp(uint16_t x) { return (x >> FP16_MAN_LEN) & FP16_MAX_EXP; }
* @ingroup fp16 basic operator
* @brief get mantissa of fp16
*/
static inline uint16_t Fp16ExtractMan(uint16_t x)
{
return static_cast<uint16_t>(((x >> 0) & 0x3FF) | ((((x >> FP16_MAN_LEN) & 0x1F) > 0 ? 1 : 0) * 0x400));
}
* @ingroup fp16 basic operator
* @brief constructor of fp16 from sign exponent and mantissa
*/
static inline uint16_t Fp16Constructor(uint16_t s, uint16_t e, uint16_t m)
{
return static_cast<uint16_t>((s << FP16_SIGN_INDEX) | (e << FP16_MAN_LEN) | (m & FP16_MAX_MAN));
}
* @ingroup fp16 special value judgment
* @brief whether a fp16 is zero
*/
static inline bool Fp16IsZero(uint16_t x) { return (x & FP16_ABS_MAX) == 0; }
* @ingroup fp16 special value judgment
* @brief whether a fp16 is a denormalized value
*/
static inline bool Fp16IsDenorm(uint16_t x) { return (x & FP16_EXP_MASK) == 0; }
* @ingroup fp16 special value judgment
* @brief whether a fp16 is invalid
*/
static inline bool Fp16IsInvalid(uint16_t x) { return (x & FP16_EXP_MASK) == FP16_EXP_MASK; }
* @ingroup fp32 basic parameter
* @brief fp32 exponent bias
*/
constexpr uint32_t FP32_EXP_BIAS = 127;
* @ingroup fp32 basic parameter
* @brief the mantissa bit length of float/fp32 is 23
*/
constexpr uint32_t FP32_MAN_LEN = 23;
* @ingroup fp32 basic parameter
* @brief bit index of sign in float/fp32
*/
constexpr uint32_t FP32_SIGN_INDEX = 31;
* @ingroup fp32 basic parameter
* @brief sign mask of fp32 (1 0000 0000 0000 0000 0000 0000 000)
*/
constexpr uint32_t FP32_SIGN_MASK = 0x80000000u;
* @ingroup fp32 basic parameter
* @brief exponent mask of fp32 ( 1111 1111 0000 0000 0000 0000 000)
*/
constexpr uint32_t FP32_EXP_MASK = 0x7F800000u;
* @ingroup fp32 basic parameter
* @brief mantissa mask of fp32 ( 1111 1111 1111 1111 111)
*/
constexpr uint32_t FP32_MAN_MASK = 0x007FFFFFu;
* @ingroup fp32 basic parameter
* @brief conceal bit of mantissa of fp32 ( 1 0000 0000 0000 0000 000)
*/
constexpr uint32_t FP32_MAN_HIDE_BIT = 0x00800000u;
* @ingroup fp32 basic parameter
* @brief absolute maximum value (0 1111 1111 1111 1111 1111 1111 111)
*/
constexpr uint32_t FP32_ABS_MAX = 0x7FFFFFFFu;
* @ingroup fp32 basic parameter
* @brief maximum mantissa value of fp32 (1111 1111 1111 1111 1111 111)
*/
constexpr uint32_t FP32_MAX_MAN = 0x7FFFFF;
static inline uint32_t Fp32Constructor(uint32_t s, uint32_t e, uint32_t m)
{
return static_cast<uint32_t>(((s) << FP32_SIGN_INDEX) | ((e) << FP32_MAN_LEN) | ((m)&FP32_MAX_MAN));
}
* @ingroup fp64 basic parameter
* @brief fp64 exponent bias
*/
constexpr uint64_t FP64_EXP_BIAS = 1023;
* @ingroup fp64 basic parameter
* @brief the mantissa bit length of double/fp64 is 52
*/
constexpr uint64_t FP64_MAN_LEN = 52;
* @ingroup fp64 basic parameter
* @brief bit index of sign in double/fp64 is 63
*/
constexpr uint64_t FP64_SIGN_INDEX = 63;
* @ingroup fp64 basic parameter
* @brief sign mask of fp64 (1 000 (total 63bits 0))
*/
constexpr uint64_t FP64_SIGN_MASK = 0x8000000000000000LLu;
* @ingroup fp64 basic parameter
* @brief exponent mask of fp64 (0 1 11111 11111 0000?-?-(total 52bits 0))
*/
constexpr uint64_t FP64_EXP_MASK = 0x7FF0000000000000LLu;
* @ingroup fp64 basic parameter
* @brief mantissa mask of fp64 ( 1111?-?-(total 52bits 1))
*/
constexpr uint64_t FP64_MAN_MASK = 0x000FFFFFFFFFFFFFLLu;
* @ingroup fp64 basic parameter
* @brief conceal bit of mantissa of fp64 ( 1 0000?-?-(total 52bits 0))
*/
constexpr uint64_t FP64_MAN_HIDE_BIT = 0x0010000000000000LLu;
* @ingroup integer special value judgment
* @brief maximum positive value of int8_t (0111 1111)
*/
constexpr int8_t INT8_T_MAX = 0x7F;
* @ingroup integer special value judgment
* @brief maximum value of a data with 8 bits length (1111 111)
*/
constexpr uint8_t BIT_LEN8_MAX = 0xFF;
* @ingroup integer special value judgment
* @brief maximum positive value of int16_t (0111 1111 1111 1111)
*/
constexpr int16_t INT16_T_MAX = 0x7FFF;
* @ingroup integer special value judgment
* @brief maximum value of a data with 16 bits length (1111 1111 1111 1111)
*/
constexpr uint16_t BIT_LEN16_MAX = 0xFFFF;
* @ingroup integer special value judgment
* @brief maximum positive value of int32_t (0111 1111 1111 1111 1111 1111 1111 1111)
*/
constexpr int32_t INT32_T_MAX = 0x7FFFFFFFu;
* @ingroup integer special value judgment
* @brief maximum value of a data with 32 bits length (1111 1111 1111 1111 1111 1111 1111 1111)
*/
constexpr uint32_t BIT_LEN32_MAX = 0xFFFFFFFFu;
* @ingroup print switch
* @brief print an error if input fp16 is overflow
*/
* @ingroup fp16_t enum
* @brief round mode of last valid digital
*/
enum class Fp16RoundMode : int {
ROUND_TO_NEAREST = 0,
ROUND_BY_TRUNCATED = 1,
ROUND_MODE_RESERVED = 2,
};
* @ingroup fp16_t global filed
* @brief round mode of last valid digital
*/
constexpr Fp16RoundMode g_RoundMode = Fp16RoundMode::ROUND_TO_NEAREST;
* @ingroup fp16_t public method
* @param [in] negative sign is negative
* @param [in|out] man mantissa to be reverse
* @brief Calculate a mantissa's complement (add ont to it's radix-minus-one complement)
* @return Return complement of man
*/
template <typename T>
static void ReverseMan(bool negative, T* man)
{
if (negative) {
*man = (~(*man)) + 1;
}
}
* @ingroup fp16_t public method
* @param [in] man mantissa to be operate
* @param [in] shift right shift bits
* @brief right shift a mantissa
* @return Return right-shift mantissa
*/
template <typename T>
static auto RightShift(T man, int16_t shift) -> T
{
unsigned int bits = static_cast<unsigned int>(sizeof(T) * 8U);
T mask = (static_cast<T>(1U)) << (static_cast<T>(bits - 1U));
for (int i = 0; i < shift; i++) {
man = ((man & mask) | (man >> 1));
}
return man;
}
* @ingroup fp16_t public method
* @param [in] ea exponent of one temp fp16_t number
* @param [in] ma mantissa of one temp fp16_t number
* @param [in] eb exponent of another temp fp16_t number
* @param [in] mb mantissa of another temp fp16_t number
* @brief Get mantissa sum of two temp fp16_t numbers, T support types: uint16_t/uint32_t/uint64_t
* @return Return mantissa sum
*/
template <typename T>
static auto GetManSum(int16_t ea, const T& ma, int16_t eb, const T& mb) -> T
{
T sum = 0;
if (ea != eb) {
T m_tmp = 0;
int16_t e_tmp = std::abs(ea - eb);
if (ea > eb) {
m_tmp = mb;
m_tmp = RightShift(m_tmp, e_tmp);
sum = ma + m_tmp;
} else {
m_tmp = ma;
m_tmp = RightShift(m_tmp, e_tmp);
sum = m_tmp + mb;
}
} else {
sum = ma + mb;
}
return sum;
}
* @ingroup fp16_t public method
* @param [in] bit0 whether the last preserved bit is 1 before round
* @param [in] bit1 whether the abbreviation's highest bit is 1
* @param [in] bitLeft whether the abbreviation's bits which not contain highest bit grater than 0
* @param [in] man mantissa of a fp16_t or float number, support types: uint16_t/uint32_t/uint64_t
* @param [in] shift abbreviation bits
* @brief Round fp16_t or float mantissa to nearest value
* @return Returns true if round 1,otherwise false;
*/
template <typename T>
static auto ManRoundToNearest(bool bit0, bool bit1, bool bitLeft, T man, uint16_t shift = 0) -> T
{
man = (man >> shift) + ((bit1 && (bitLeft || bit0)) ? 1 : 0);
return man;
}
* @ingroup fp16_t public method
* @param [in] man mantissa of a float number, support types: uint16_t/uint32_t/uint64_t
* @brief Get bit length of a uint32_t number
* @return Return bit length of man
*/
template <typename T>
static int16_t GetManBitLength(T man)
{
int16_t len = 0;
while (man) {
man >>= 1;
len++;
}
return len;
}
* @ingroup fp16_t public method
* @param [in] val signature is negative
* @param [in|out] s sign of fp16_t object
* @param [in|out] e exponent of fp16_t object
* @param [in|out] m mantissa of fp16_t object
* @brief Extract the sign, exponent and mantissa of a fp16_t object, if num is denormal, convert to normal
*/
static void ExtractFP16Norm(const uint16_t& val, uint16_t* s, int16_t* e, uint16_t* m)
{
*s = Fp16ExtractSign(val);
*e = Fp16ExtractExp(val);
*m = Fp16ExtractMan(val);
if (0 == (*e)) {
*e = 1;
}
}
* @ingroup fp16_t public method
* @param [in] val signature is negative
* @param [in|out] s sign of fp16_t object
* @param [in|out] e exponent of fp16_t object
* @param [in|out] m mantissa of fp16_t object
* @brief Extract the sign, exponent and mantissa of a fp16_t object, if num is denormal, convert to normal
*/
static void ExtractFP16(const uint16_t& val, uint8_t* s, uint16_t* e, uint16_t* m)
{
*s = Fp16ExtractSign(val);
*e = Fp16ExtractExp(val);
*m = Fp16ExtractMan(val);
}
* @ingroup fp16_t static method
* @param [in] man truncated mantissa
* @param [in] shiftOut left shift bits based on ten bits
* @brief judge whether to add one to the result while converting fp16_t to other datatype
* @return Return true if add one, otherwise false
*/
static bool IsRoundOne(uint64_t man, uint16_t truncLen)
{
uint16_t shiftOut = truncLen - FP16_BIT_SHIFT_LEN_2;
uint64_t mask0 = 0x4;
mask0 = mask0 << shiftOut;
uint64_t mask1 = 0x2;
mask1 = mask1 << shiftOut;
uint64_t mask2 = mask1 - 1;
bool lastBit = ((man & mask0) > 0);
bool truncHigh = false;
bool truncLeft = false;
if (Fp16RoundMode::ROUND_TO_NEAREST == g_RoundMode) {
truncHigh = ((man & mask1) > 0);
truncLeft = ((man & mask2) > 0);
}
return (truncHigh && (truncLeft || lastBit));
}
* @ingroup fp16_t public method
* @param [in] exp exponent of fp16_t value
* @param [in] man exponent of fp16_t value
* @brief normalize fp16_t value
* @return
*/
static void Fp16Normalize(int16_t& exp, uint16_t& man)
{
if (exp >= FP16_MAX_EXP) {
exp = FP16_MAX_EXP - 1;
man = FP16_MAX_MAN;
} else if (exp == 0 && man == FP16_MAN_HIDE_BIT) {
exp++;
man = 0;
}
}
* @ingroup fp16_t math conversion static method
* @param [in] fpVal uint16_t value of fp16_t object
* @brief Convert fp16_t to float/fp32
* @return Return float/fp32 value of fpVal which is the value of fp16_t object
*/
static float fp16ToFloat(const uint16_t& fpVal)
{
float ret;
uint16_t hfSign, hfMan;
int16_t hfExp;
ExtractFP16Norm(fpVal, &hfSign, &hfExp, &hfMan);
while (hfMan != 0 && (hfMan & FP16_MAN_HIDE_BIT) == 0) {
hfMan <<= 1;
hfExp--;
}
uint32_t sRet, eRet, mRet, fVal;
sRet = hfSign;
if (hfMan == 0) {
eRet = 0;
mRet = 0;
} else {
eRet = static_cast<uint32_t>(hfExp - FP16_EXP_BIAS + FP32_EXP_BIAS);
mRet = hfMan & FP16_MAN_MASK;
mRet = mRet << (FP32_MAN_LEN - FP16_MAN_LEN);
}
fVal = Fp32Constructor(sRet, eRet, mRet);
uint32_t* ptrFVal = &fVal;
float* ptrRet = reinterpret_cast<float*>(ptrFVal);
ret = *ptrRet;
return ret;
}
* @ingroup fp16_t math conversion static method
* @param [in] fpVal uint16_t value of fp16_t object
* @brief Convert fp16_t to double/fp64
* @return Return double/fp64 value of fpVal which is the value of fp16_t object
*/
static double fp16ToDouble(const uint16_t& fpVal)
{
double ret;
uint16_t hfSign, hfMan;
int16_t hfExp;
ExtractFP16Norm(fpVal, &hfSign, &hfExp, &hfMan);
while (hfMan != 0 && (hfMan & FP16_MAN_HIDE_BIT) == 0) {
hfMan <<= 1;
hfExp--;
}
uint64_t sRet, eRet, mRet, fVal;
sRet = hfSign;
if (hfMan == 0) {
eRet = 0;
mRet = 0;
} else {
eRet = static_cast<uint64_t>(hfExp - FP16_EXP_BIAS + FP64_EXP_BIAS);
mRet = hfMan & FP16_MAN_MASK;
mRet = mRet << (FP64_MAN_LEN - FP16_MAN_LEN);
}
fVal = (sRet << FP64_SIGN_INDEX) | (eRet << FP64_MAN_LEN) | (mRet);
uint64_t* ptrFVal = &fVal;
double* ptrRet = reinterpret_cast<double*>(ptrFVal);
ret = *ptrRet;
return ret;
}
static inline void Fp16AlignExp4Int8(uint8_t sRet, uint16_t hfE, uint64_t& longIntM, uint8_t& overflowFlag,
uint16_t& shiftOut)
{
while (hfE != FP16_EXP_BIAS) {
if (hfE > FP16_EXP_BIAS) {
hfE--;
longIntM = longIntM << 1;
if (sRet == 1 && longIntM >= 0x20000u) {
longIntM = 0x20000u;
overflowFlag = 1;
break;
} else if (sRet != 1 && longIntM >= 0x1FFFFu) {
longIntM = 0x1FFFFu;
overflowFlag = 1;
break;
}
} else {
hfE++;
shiftOut++;
}
}
}
* @ingroup fp16_t math conversion static method
* @param [in] fpVal uint16_t value of fp16_t object
* @brief Convert fp16_t to int8_t
* @return Return int8_t value of fpVal which is the value of fp16_t object
*/
static int8_t fp16ToInt8(const uint16_t& fpVal)
{
int8_t ret;
uint8_t retV;
uint8_t sRet;
uint8_t mRet = 0;
uint16_t hfE, hfM;
ExtractFP16(fpVal, &sRet, &hfE, &hfM);
if (Fp16IsDenorm(fpVal)) {
return 0;
}
uint16_t shiftOut = 0;
uint8_t overflowFlag = 0;
uint64_t longIntM = hfM;
if (Fp16IsInvalid(fpVal)) {
overflowFlag = 1;
} else {
Fp16AlignExp4Int8(sRet, hfE, longIntM, overflowFlag, shiftOut);
}
if (overflowFlag != 0) {
retV = INT8_T_MAX + sRet;
} else {
bool needRound = IsRoundOne(longIntM, shiftOut + FP16_MAN_LEN);
mRet = static_cast<uint8_t>((longIntM >> (FP16_MAN_LEN + shiftOut)) & BIT_LEN8_MAX);
needRound = needRound && ((sRet == 0 && mRet < INT8_T_MAX) || (sRet == 1 && mRet <= INT8_T_MAX));
if (needRound) {
mRet++;
}
if (sRet != 0) {
mRet = (~mRet) + 1;
}
if (mRet == 0) {
sRet = 0;
}
retV = (sRet << FP16_BIT_SHIFT_LEN_7) | (mRet);
}
ret = *reinterpret_cast<int8_t*>(&retV);
return ret;
}
* @ingroup fp16_t math conversion static method
* @param [in] fpVal uint16_t value of fp16_t object
* @brief Convert fp16_t to uint8_t
* @return Return uint8_t value of fpVal which is the value of fp16_t object
*/
static uint8_t fp16ToUInt8(const uint16_t& fpVal)
{
uint8_t sRet;
uint8_t mRet = 0;
uint16_t hfE, hfM;
if (Fp16IsDenorm(fpVal)) {
return 0;
}
if (Fp16IsInvalid(fpVal)) {
return ~0;
}
ExtractFP16(fpVal, &sRet, &hfE, &hfM);
uint16_t shiftOut = 0;
uint64_t longIntM = hfM;
uint8_t overflowFlag = 0;
while (hfE != FP16_EXP_BIAS) {
if (hfE > FP16_EXP_BIAS) {
hfE--;
longIntM = longIntM << 1;
if (longIntM >= 0x40000Lu) {
longIntM = 0x3FFFFLu;
overflowFlag = 1;
mRet = ~0;
break;
}
} else {
hfE++;
shiftOut++;
}
}
if (overflowFlag == 0) {
bool needRound = IsRoundOne(longIntM, shiftOut + FP16_MAN_LEN);
mRet = static_cast<uint8_t>((longIntM >> (FP16_MAN_LEN + shiftOut)) & BIT_LEN8_MAX);
if (needRound && mRet != BIT_LEN8_MAX) {
mRet++;
}
}
if (sRet == 1) {
mRet = 0;
}
return mRet;
}
static inline void Fp16AlignExp4Int16(uint8_t sRet, uint16_t hfE, uint64_t& longIntM, uint8_t& overflowFlag,
uint16_t& shiftOut)
{
while (hfE != FP16_EXP_BIAS) {
if (hfE > FP16_EXP_BIAS) {
hfE--;
longIntM = longIntM << 1;
if (sRet == 1 && longIntM > 0x2000000Lu) {
longIntM = 0x2000000Lu;
overflowFlag = 1;
break;
} else if (sRet != 1 && longIntM >= 0x1FFFFFFLu) {
longIntM = 0x1FFFFFFLu;
overflowFlag = 1;
break;
}
} else {
hfE++;
shiftOut++;
}
}
}
* @ingroup fp16_t math conversion static method
* @param [in] fpVal uint16_t value of fp16_t object
* @brief Convert fp16_t to int16_t
* @return Return int16_t value of fpVal which is the value of fp16_t object
*/
static int16_t fp16ToInt16(const uint16_t& fpVal)
{
int16_t ret;
uint16_t retV;
uint8_t sRet;
uint16_t mRet = 0;
uint16_t hfE, hfM;
ExtractFP16(fpVal, &sRet, &hfE, &hfM);
if (Fp16IsDenorm(fpVal)) {
return 0;
}
uint16_t shiftOut = 0;
uint64_t longIntM = hfM;
uint8_t overflowFlag = 0;
if (Fp16IsInvalid(fpVal)) {
overflowFlag = 1;
} else {
Fp16AlignExp4Int16(sRet, hfE, longIntM, overflowFlag, shiftOut);
}
if (overflowFlag == 1) {
retV = INT16_T_MAX + sRet;
} else {
bool needRound = IsRoundOne(longIntM, shiftOut + FP16_MAN_LEN);
mRet = static_cast<uint16_t>((longIntM >> (FP16_MAN_LEN + shiftOut)) & BIT_LEN16_MAX);
if (needRound && mRet < INT16_T_MAX) {
mRet++;
}
if (sRet != 0) {
mRet = (~mRet) + 1;
}
if (mRet == 0) {
sRet = 0;
}
retV = (sRet << FP16_BIT_SHIFT_LEN_15) | (mRet);
}
ret = *reinterpret_cast<int16_t*>(&retV);
return ret;
}
static inline void Fp16AlignExp(uint16_t hfE, uint16_t& shiftOut, uint64_t& longIntM)
{
while (hfE != FP16_EXP_BIAS) {
if (hfE > FP16_EXP_BIAS) {
hfE--;
longIntM = longIntM << 1;
} else {
hfE++;
shiftOut++;
}
}
}
* @ingroup fp16_t math conversion static method
* @param [in] fpVal uint16_t value of fp16_t object
* @brief Convert fp16_t to uint16_t
* @return Return uint16_t value of fpVal which is the value of fp16_t object
*/
static uint16_t fp16ToUInt16(const uint16_t& fpVal)
{
uint8_t sRet = 0;
uint16_t mRet = 0;
uint16_t hfE, hfM;
if (Fp16IsDenorm(fpVal)) {
return 0;
}
if (Fp16IsInvalid(fpVal)) {
mRet = ~0;
return mRet;
}
ExtractFP16(fpVal, &sRet, &hfE, &hfM);
uint16_t shiftOut = 0;
uint64_t longIntM = hfM;
Fp16AlignExp(hfE, shiftOut, longIntM);
bool needRound = IsRoundOne(longIntM, shiftOut + FP16_MAN_LEN);
mRet = static_cast<uint16_t>((longIntM >> (FP16_MAN_LEN + shiftOut)) & BIT_LEN16_MAX);
if (needRound && mRet != BIT_LEN16_MAX) {
mRet++;
}
if (sRet == 1) {
mRet = 0;
}
return mRet;
}
* @ingroup fp16_t math convertion static method
* @param [in] fpVal uint16_t value of fp16_t object
* @brief Convert fp16_t to int32_t
* @return Return int32_t value of fpVal which is the value of fp16_t object
*/
static int32_t fp16ToInt32(const uint16_t& fpVal)
{
int32_t ret;
uint32_t retV;
uint8_t sRet = 0;
uint32_t mRet;
uint16_t hfE, hfM;
ExtractFP16(fpVal, &sRet, &hfE, &hfM);
if (Fp16IsInvalid(fpVal)) {
retV = INT32_T_MAX + sRet;
} else {
uint16_t shiftOut = 0;
uint64_t longIntM = hfM;
Fp16AlignExp(hfE, shiftOut, longIntM);
bool needRound = IsRoundOne(longIntM, shiftOut + FP16_MAN_LEN);
mRet = static_cast<uint32_t>((longIntM >> (FP16_MAN_LEN + shiftOut)) & BIT_LEN32_MAX);
if (needRound && mRet < INT32_T_MAX) {
mRet++;
}
if (sRet == 1) {
mRet = (~mRet) + 1;
}
if (mRet == 0) {
sRet = 0;
}
retV = (sRet << FP16_BIT_SHIFT_LEN_31) | (mRet);
}
ret = *reinterpret_cast<int32_t*>(&retV);
return ret;
}
* @ingroup fp16_t math conversion static method
* @param [in] fpVal uint16_t value of fp16_t object
* @brief Convert fp16_t to uint32_t
* @return Return uint32_t value of fpVal which is the value of fp16_t object
*/
static uint32_t fp16ToUInt32(const uint16_t& fpVal)
{
uint32_t ret;
uint8_t sRet;
uint32_t mRet;
uint16_t hfE, hfM;
ExtractFP16(fpVal, &sRet, &hfE, &hfM);
if (Fp16IsDenorm(fpVal)) {
return 0u;
}
if (Fp16IsInvalid(fpVal)) {
mRet = ~0u;
} else {
uint16_t shiftOut = 0;
uint64_t longIntM = hfM;
Fp16AlignExp(hfE, shiftOut, longIntM);
bool needRound = IsRoundOne(longIntM, shiftOut + FP16_MAN_LEN);
mRet = static_cast<uint32_t>(longIntM >> (FP16_MAN_LEN + shiftOut)) & BIT_LEN32_MAX;
if (needRound && mRet != BIT_LEN32_MAX) {
mRet++;
}
}
if (sRet == 1) {
mRet = 0;
}
ret = mRet;
return ret;
}
static void NormalizeMantissa(uint16_t& mRet, int16_t& eRet, uint32_t& mTrunc, uint16_t shiftOut)
{
uint16_t mMin = static_cast<uint16_t>(FP16_MAN_HIDE_BIT << shiftOut);
uint16_t mMax = mMin << 1;
while (mRet < mMin && eRet > 0) {
mRet = mRet << 1;
mRet += (FP32_SIGN_MASK & mTrunc) >> FP32_SIGN_INDEX;
mTrunc = mTrunc << 1;
eRet = eRet - 1;
}
while (mRet >= mMax) {
mTrunc = mTrunc >> 1;
mTrunc = mTrunc | (FP32_SIGN_MASK * (mRet & 1));
mRet = mRet >> 1;
eRet = eRet + 1;
}
bool bLastBit = ((mRet & 1) > 0);
bool bTruncHigh = 0;
bool bTruncLeft = 0;
bTruncHigh = (Fp16RoundMode::ROUND_TO_NEAREST == g_RoundMode) && ((mTrunc & FP32_SIGN_MASK) > 0);
bTruncLeft = (Fp16RoundMode::ROUND_TO_NEAREST == g_RoundMode) && ((mTrunc & FP32_ABS_MAX) > 0);
mRet = ManRoundToNearest(bLastBit, bTruncHigh, bTruncLeft, mRet, shiftOut);
while (mRet >= mMax) {
mRet = mRet >> 1;
eRet = eRet + 1;
}
if (eRet == 0 && mRet <= mMax) {
mRet = mRet >> 1;
}
Fp16Normalize(eRet, mRet);
}
* @ingroup fp16_t math operator
* @param [in] v1 left operator value of fp16_t object
* @param [in] v2 right operator value of fp16_t object
* @brief Performing fp16_t addition
* @return Return fp16_t result of adding this and fp
*/
static uint16_t fp16Add(uint16_t v1, uint16_t v2)
{
uint16_t ret;
uint16_t sA, sB;
int16_t eA, eB;
uint32_t mA, mB;
uint16_t sRet, mRet;
int16_t eRet;
uint16_t maTmp, mbTmp;
uint16_t shiftOut = 0;
ExtractFP16Norm(v1, &sA, &eA, &maTmp);
ExtractFP16Norm(v2, &sB, &eB, &mbTmp);
mA = maTmp;
mB = mbTmp;
uint16_t sum;
if (sA != sB) {
ReverseMan(sA > 0, &mA);
ReverseMan(sB > 0, &mB);
sum = static_cast<uint16_t>(GetManSum(eA, mA, eB, mB));
sRet = (sum & FP16_SIGN_MASK) >> FP16_SIGN_INDEX;
ReverseMan(sRet > 0, &mA);
ReverseMan(sRet > 0, &mB);
} else {
sum = static_cast<uint16_t>(GetManSum(eA, mA, eB, mB));
sRet = sA;
}
if (sum == 0) {
shiftOut = FP16_BIT_SHIFT_LEN_3;
mA = mA << shiftOut;
mB = mB << shiftOut;
}
int16_t eTmp = 0;
uint32_t mTrunc = 0;
eRet = std::max(eA, eB);
eTmp = std::abs(eA - eB);
if (eA > eB) {
mTrunc = (mB << (FP16_BIT_SHIFT_LEN_32 - static_cast<uint16_t>(eTmp)));
mB = RightShift(mB, eTmp);
} else if (eA < eB) {
mTrunc = (mA << (FP16_BIT_SHIFT_LEN_32 - static_cast<uint16_t>(eTmp)));
mA = RightShift(mA, eTmp);
}
mRet = static_cast<uint16_t>(mA + mB);
NormalizeMantissa(mRet, eRet, mTrunc, shiftOut);
ret = Fp16Constructor(sRet, static_cast<uint16_t>(eRet), mRet);
return ret;
}
* @ingroup fp16_t math operator
* @param [in] v1 left operator value of fp16_t object
* @param [in] v2 right operator value of fp16_t object
* @brief Performing fp16_t subtraction
* @return Return fp16_t result of subtraction fp from this
*/
static uint16_t fp16Sub(uint16_t v1, uint16_t v2)
{
uint16_t ret;
uint16_t tmp;
tmp = ((~(v2)) & FP16_SIGN_MASK) | (v2 & FP16_ABS_MAX);
ret = fp16Add(v1, tmp);
return ret;
}
* @ingroup fp16_t math operator
* @param [in] v1 left operator value of fp16_t object
* @param [in] v2 right operator value of fp16_t object
* @brief Performing fp16_t multiplication
* @return Return fp16_t result of multiplying this and fp
*/
static uint16_t fp16Mul(uint16_t v1, uint16_t v2)
{
uint16_t sA, sB;
int16_t eA, eB;
uint32_t mA, mB;
uint16_t sRet, mRet;
int16_t eRet;
uint32_t mulM;
uint16_t maTmp, mbTmp;
ExtractFP16Norm(v1, &sA, &eA, &maTmp);
ExtractFP16Norm(v2, &sB, &eB, &mbTmp);
mA = maTmp;
mB = mbTmp;
eRet = eA + eB - FP16_EXP_BIAS - FP16_BIT_SHIFT_LEN_10;
mulM = mA * mB;
sRet = sA ^ sB;
uint32_t mMin = FP16_MAN_HIDE_BIT;
uint32_t mMax = mMin << 1;
uint32_t mTrunc = 0;
while (mulM < mMin && eRet > 1) {
mulM = mulM << 1;
eRet = eRet - 1;
}
while (mulM >= mMax || eRet < 1) {
mTrunc = mTrunc >> 1;
mTrunc = mTrunc | (FP32_SIGN_MASK * (mulM & 1));
mulM = mulM >> 1;
eRet = eRet + 1;
}
bool bLastBit = ((mulM & 1) > 0);
bool bTruncHigh = (Fp16RoundMode::ROUND_TO_NEAREST == g_RoundMode) && ((mTrunc & FP32_SIGN_MASK) > 0);
bool bTruncLeft = (Fp16RoundMode::ROUND_TO_NEAREST == g_RoundMode) && ((mTrunc & FP32_ABS_MAX) > 0);
mulM = ManRoundToNearest(bLastBit, bTruncHigh, bTruncLeft, mulM);
while (mulM >= mMax || eRet < 0) {
mulM = mulM >> 1;
eRet = eRet + 1;
}
if (eRet == 1 && mulM < FP16_MAN_HIDE_BIT) {
eRet = 0;
}
mRet = static_cast<uint16_t>(mulM);
Fp16Normalize(eRet, mRet);
uint16_t ret = Fp16Constructor(sRet, static_cast<uint16_t>(eRet), mRet);
return ret;
}
* @ingroup fp16_t math operator divided
* @param [in] v1 left operator value of fp16_t object
* @param [in] v2 right operator value of fp16_t object
* @brief Performing fp16_t division
* @return Return fp16_t result of division this by fp
*/
static uint16_t fp16Div(uint16_t v1, uint16_t v2)
{
uint16_t ret;
if (Fp16IsZero(v2)) {
uint16_t sA, sB;
uint16_t sRet;
sA = Fp16ExtractSign(v1);
sB = Fp16ExtractSign(v2);
sRet = sA ^ sB;
ret = Fp16Constructor(sRet, FP16_MAX_EXP - 1, FP16_MAX_MAN);
} else if (Fp16IsZero(v1)) {
ret = 0u;
} else {
uint16_t sA, sB;
int16_t eA, eB;
uint64_t mA, mB;
float mDiv;
uint16_t maTmp, mbTmp;
ExtractFP16Norm(v1, &sA, &eA, &maTmp);
ExtractFP16Norm(v2, &sB, &eB, &mbTmp);
mA = maTmp;
mB = mbTmp;
uint64_t mTmp;
if (eA > eB) {
mTmp = mA;
uint16_t tmp = static_cast<uint16_t>(eA - eB);
for (int i = 0; i < tmp; i++) {
mTmp = mTmp << 1;
}
mA = mTmp;
} else if (eA < eB) {
mTmp = mB;
uint16_t tmp = static_cast<uint16_t>(eB - eA);
for (int i = 0; i < tmp; i++) {
mTmp = mTmp << 1;
}
mB = mTmp;
}
mDiv = static_cast<float>(mA * 1.0f / mB);
fp16_t fpDiv = mDiv;
ret = fpDiv.val;
if (sA != sB) {
ret |= FP16_SIGN_MASK;
}
}
return ret;
}
fp16_t fp16_t::operator+(const fp16_t fp) const
{
uint16_t retVal = fp16Add(val, fp.val);
fp16_t ret(retVal);
return ret;
}
fp16_t fp16_t::operator-(const fp16_t fp) const
{
uint16_t retVal = fp16Sub(val, fp.val);
fp16_t ret(retVal);
return ret;
}
fp16_t fp16_t::operator*(const fp16_t fp) const
{
uint16_t retVal = fp16Mul(val, fp.val);
fp16_t ret(retVal);
return ret;
}
fp16_t fp16_t::operator/(const fp16_t fp) const
{
uint16_t retVal = fp16Div(val, fp.val);
fp16_t ret(retVal);
return ret;
}
fp16_t fp16_t::operator+=(const fp16_t fp)
{
val = fp16Add(val, fp.val);
return *this;
}
fp16_t fp16_t::operator-=(const fp16_t fp)
{
val = fp16Sub(val, fp.val);
return *this;
}
fp16_t fp16_t::operator*=(const fp16_t fp)
{
val = fp16Mul(val, fp.val);
return *this;
}
fp16_t fp16_t::operator/=(const fp16_t fp)
{
val = fp16Div(val, fp.val);
return *this;
}
bool fp16_t::operator==(const fp16_t& fp) const
{
bool result = true;
if (Fp16IsZero(val) && Fp16IsZero(fp.val)) {
result = true;
} else {
result = ((val & BIT_LEN16_MAX) == (fp.val & BIT_LEN16_MAX));
}
return result;
}
bool fp16_t::operator!=(const fp16_t& fp) const
{
bool result = true;
if (Fp16IsZero(val) && Fp16IsZero(fp.val)) {
result = false;
} else {
result = ((val & BIT_LEN16_MAX) != (fp.val & BIT_LEN16_MAX));
}
return result;
}
bool fp16_t::operator>(const fp16_t& fp) const
{
uint16_t sA, sB;
uint16_t eA, eB;
uint16_t mA, mB;
bool result = true;
sA = Fp16ExtractSign(val);
sB = Fp16ExtractSign(fp.val);
eA = Fp16ExtractExp(val);
eB = Fp16ExtractExp(fp.val);
mA = Fp16ExtractMan(val);
mB = Fp16ExtractMan(fp.val);
if ((sA == 0) && (sB > 0)) {
result = !(Fp16IsZero(val) && Fp16IsZero(fp.val));
} else if ((sA == 0) && (sB == 0)) {
if (eA > eB) {
result = true;
} else if (eA == eB) {
result = mA > mB;
} else {
result = false;
}
} else if ((sA > 0) && (sB > 0)) {
if (eA < eB) {
result = true;
} else if (eA == eB) {
result = mA < mB;
} else {
result = false;
}
} else {
result = false;
}
return result;
}
bool fp16_t::operator>=(const fp16_t& fp) const
{
bool result = true;
if ((*this) > fp) {
result = true;
} else if ((*this) == fp) {
result = true;
} else {
result = false;
}
return result;
}
bool fp16_t::operator<(const fp16_t& fp) const
{
bool result = true;
if ((*this) >= fp) {
result = false;
} else {
result = true;
}
return result;
}
bool fp16_t::operator<=(const fp16_t& fp) const
{
bool result = true;
if ((*this) > fp) {
result = false;
} else {
result = true;
}
return result;
}
fp16_t& fp16_t::operator=(const fp16_t& fp)
{
if (this == &fp) {
return *this;
}
val = fp.val;
return *this;
}
fp16_t& fp16_t::operator=(const float& fVal)
{
uint16_t sRet, manRet;
int16_t eRet;
uint32_t eF, mF;
uint32_t ui32V = *reinterpret_cast<const uint32_t*>(&fVal);
uint32_t mLenDelta;
sRet = static_cast<uint16_t>((ui32V & FP32_SIGN_MASK) >> FP32_SIGN_INDEX);
eF = (ui32V & FP32_EXP_MASK) >> FP32_MAN_LEN;
mF = (ui32V & FP32_MAN_MASK);
mLenDelta = FP32_MAN_LEN - FP16_MAN_LEN;
bool needRound = false;
if (eF > 0x8Fu) {
eRet = FP16_MAX_EXP - 1;
manRet = FP16_MAX_MAN;
} else if (eF <= 0x70u) {
eRet = 0;
if (eF >= 0x67) {
mF = (mF | FP32_MAN_HIDE_BIT);
uint16_t shiftOut = FP32_MAN_LEN;
uint64_t mTmp = static_cast<uint64_t>(mF) << (eF - 0x67);
needRound = IsRoundOne(mTmp, shiftOut);
manRet = static_cast<uint16_t>(mTmp >> shiftOut);
if (needRound) {
manRet++;
}
} else if (eF == 0x66 && mF > 0) {
manRet = 1;
} else {
manRet = 0;
}
} else {
eRet = static_cast<int16_t>(eF - 0x70u);
needRound = IsRoundOne(mF, mLenDelta);
manRet = static_cast<uint16_t>(mF >> mLenDelta);
if (needRound) {
manRet++;
}
if ((manRet & FP16_MAN_HIDE_BIT) != 0) {
eRet++;
}
}
Fp16Normalize(eRet, manRet);
val = Fp16Constructor(sRet, static_cast<uint16_t>(eRet), manRet);
return *this;
}
static inline uint16_t EncodeFp16FromUint8(uint16_t sign, uint16_t mantissa)
{
uint16_t exp = 0;
if (mantissa != 0) {
exp = FP16_MAN_LEN;
while ((mantissa & FP16_MAN_HIDE_BIT) == 0) {
mantissa <<= 1;
exp -= 1;
}
exp += FP16_EXP_BIAS;
}
return Fp16Constructor(sign, exp, mantissa);
}
fp16_t& fp16_t::operator=(const int8_t& iVal)
{
uint16_t sRet = static_cast<uint16_t>((static_cast<uint8_t>(iVal) & 0x80) >> FP16_BIT_SHIFT_LEN_7);
uint16_t mRet = static_cast<uint16_t>((static_cast<uint8_t>(iVal) & INT8_T_MAX));
if (sRet != 0) {
mRet = static_cast<uint16_t>(std::abs(iVal));
}
val = EncodeFp16FromUint8(sRet, mRet);
return *this;
}
fp16_t& fp16_t::operator=(const uint8_t& uiVal)
{
uint16_t sRet = 0;
uint16_t mRet = uiVal;
val = EncodeFp16FromUint8(sRet, mRet);
return *this;
}
static inline uint16_t EncodeFp16FromUint16(uint16_t sign, uint16_t mRet)
{
if (mRet == 0) {
return 0;
}
int16_t eRet;
uint16_t mMin = FP16_MAN_HIDE_BIT;
uint16_t mMax = mMin << 1;
uint16_t len = static_cast<uint16_t>(GetManBitLength(mRet));
if (len > FP16_BIT_SHIFT_LEN_11) {
eRet = FP16_EXP_BIAS + FP16_MAN_LEN;
uint32_t mTrunc;
uint32_t truncMask = 1;
uint16_t eTmp = len - FP16_BIT_SHIFT_LEN_11;
for (int i = 1; i < eTmp; i++) {
truncMask = (truncMask << 1) + 1;
}
mTrunc = (mRet & truncMask) << (FP16_BIT_SHIFT_LEN_32 - eTmp);
for (int i = 0; i < eTmp; i++) {
mRet = (mRet >> 1);
eRet = eRet + 1;
}
bool bLastBit = ((mRet & 1) > 0);
bool bTruncHigh = false;
bool bTruncLeft = false;
if (Fp16RoundMode::ROUND_TO_NEAREST == g_RoundMode) {
bTruncHigh = ((mTrunc & FP32_SIGN_MASK) > 0);
bTruncLeft = ((mTrunc & FP32_ABS_MAX) > 0);
}
mRet = ManRoundToNearest(bLastBit, bTruncHigh, bTruncLeft, mRet);
while (mRet >= mMax || eRet < 0) {
mRet = mRet >> 1;
eRet = eRet + 1;
}
} else {
eRet = FP16_EXP_BIAS;
mRet = mRet << (FP16_BIT_SHIFT_LEN_11 - len);
eRet = eRet + (len - 1);
}
return Fp16Constructor(sign, static_cast<uint16_t>(eRet), mRet);
}
fp16_t& fp16_t::operator=(const int16_t& iVal)
{
if (iVal == 0) {
val = 0;
return *this;
}
uint16_t sRet;
uint16_t uiVal = *reinterpret_cast<const uint16_t*>(&iVal);
sRet = static_cast<uint16_t>(uiVal >> FP16_BIT_SHIFT_LEN_15);
if (sRet != 0) {
int16_t iValM = -iVal;
uiVal = *reinterpret_cast<uint16_t*>(&iValM);
}
uint32_t mTmp = (uiVal & FP32_ABS_MAX);
val = EncodeFp16FromUint16(sRet, mTmp);
return *this;
}
fp16_t& fp16_t::operator=(const uint16_t& uiVal)
{
val = EncodeFp16FromUint16(0U, uiVal);
return *this;
}
static inline uint16_t EncodeFp16FromUint32(uint16_t sign, uint32_t mTmp)
{
if (mTmp == 0) {
return 0;
}
int16_t eRet;
uint32_t mMin = FP16_MAN_HIDE_BIT;
uint32_t mMax = mMin << 1;
uint16_t len = static_cast<uint16_t>(GetManBitLength(mTmp));
if (len > FP16_BIT_SHIFT_LEN_11) {
eRet = FP16_EXP_BIAS + FP16_MAN_LEN;
uint32_t mTrunc = 0;
uint32_t truncMask = 1;
uint16_t eTmp = len - FP16_BIT_SHIFT_LEN_11;
for (int i = 1; i < eTmp; i++) {
truncMask = (truncMask << 1) + 1;
}
mTrunc = (mTmp & truncMask) << (FP16_BIT_SHIFT_LEN_32 - eTmp);
for (int i = 0; i < eTmp; i++) {
mTmp = (mTmp >> 1);
eRet = eRet + 1;
}
bool bLastBit = ((mTmp & 1) > 0);
bool bTruncHigh = false;
bool bTruncLeft = false;
if (Fp16RoundMode::ROUND_TO_NEAREST == g_RoundMode) {
bTruncHigh = ((mTrunc & FP32_SIGN_MASK) > 0);
bTruncLeft = ((mTrunc & FP32_ABS_MAX) > 0);
}
mTmp = ManRoundToNearest(bLastBit, bTruncHigh, bTruncLeft, mTmp);
while (mTmp >= mMax || eRet < 0) {
mTmp = mTmp >> 1;
eRet = eRet + 1;
}
if (eRet >= FP16_MAX_EXP) {
eRet = FP16_MAX_EXP - 1;
mTmp = FP16_MAX_MAN;
}
} else {
eRet = FP16_EXP_BIAS;
mTmp = mTmp << (FP16_BIT_SHIFT_LEN_11 - len);
eRet = eRet + (len - 1);
}
uint16_t mRet = static_cast<uint16_t>(mTmp);
return Fp16Constructor(sign, static_cast<uint16_t>(eRet), mRet);
}
fp16_t& fp16_t::operator=(const int32_t& iVal)
{
if (iVal == 0) {
val = 0;
return *this;
}
uint32_t uiVal = *reinterpret_cast<const uint32_t*>(&iVal);
uint16_t sRet = static_cast<uint16_t>(uiVal >> FP16_BIT_SHIFT_LEN_31);
if (sRet != 0) {
int32_t iValM = -iVal;
uiVal = *reinterpret_cast<uint32_t*>(&iValM);
}
uint32_t mTmp = (uiVal & FP32_ABS_MAX);
val = EncodeFp16FromUint32(sRet, mTmp);
return *this;
}
fp16_t& fp16_t::operator=(const uint32_t& uiVal)
{
val = EncodeFp16FromUint32(0U, uiVal);
return *this;
}
fp16_t& fp16_t::operator=(const double& dVal)
{
uint16_t sRet, mRet;
int16_t eRet;
uint64_t eD, mD;
uint64_t ui64V = *reinterpret_cast<const uint64_t*>(&dVal);
uint32_t mLenDelta;
sRet = static_cast<uint16_t>((ui64V & FP64_SIGN_MASK) >> FP64_SIGN_INDEX);
eD = (ui64V & FP64_EXP_MASK) >> FP64_MAN_LEN;
mD = (ui64V & FP64_MAN_MASK);
mLenDelta = FP64_MAN_LEN - FP16_MAN_LEN;
bool needRound = false;
if (eD >= 0x410u) {
eRet = FP16_MAX_EXP - 1;
mRet = FP16_MAX_MAN;
val = Fp16Constructor(sRet, static_cast<uint16_t>(eRet), mRet);
} else if (eD <= 0x3F0u) {
* Signed zeros, denormalized floats, and floats with small
* exponents all convert to signed zero half precision.
*/
eRet = 0;
if (eD >= 0x3E7u) {
mD = (FP64_MAN_HIDE_BIT | mD);
uint16_t shiftOut = FP64_MAN_LEN;
uint64_t mTmp = (static_cast<uint64_t>(mD)) << (eD - 0x3E7u);
needRound = IsRoundOne(mTmp, shiftOut);
mRet = static_cast<uint16_t>(mTmp >> shiftOut);
if (needRound) {
mRet++;
}
} else if (eD == 0x3E6u && mD > 0) {
mRet = 1;
} else {
mRet = 0;
}
} else {
eRet = static_cast<int16_t>(eD - 0x3F0u);
needRound = IsRoundOne(mD, mLenDelta);
mRet = static_cast<uint16_t>(mD >> mLenDelta);
if (needRound) {
mRet++;
}
if ((mRet & FP16_MAN_HIDE_BIT) != 0) {
eRet++;
}
}
Fp16Normalize(eRet, mRet);
val = Fp16Constructor(sRet, static_cast<uint16_t>(eRet), mRet);
return *this;
}
fp16_t::operator float() const { return fp16ToFloat(val); }
fp16_t::operator double() const { return fp16ToDouble(val); }
fp16_t::operator int8_t() const { return fp16ToInt8(val); }
fp16_t::operator uint8_t() const { return fp16ToUInt8(val); }
fp16_t::operator int16_t() const { return fp16ToInt16(val); }
fp16_t::operator uint16_t() const { return fp16ToUInt16(val); }
fp16_t::operator int32_t() const { return fp16ToInt32(val); }
fp16_t::operator uint32_t() const { return fp16ToUInt32(val); }
int fp16_t::IsInf() const
{
if ((val & FP16_ABS_MAX) == FP16_EXP_MASK) {
if ((val & FP16_SIGN_MASK) != 0) {
return -1;
} else {
return 1;
}
} else {
return 0;
}
}
float fp16_t::toFloat() { return fp16ToFloat(val); }
double fp16_t::toDouble() { return fp16ToDouble(val); }
int8_t fp16_t::toInt8() { return fp16ToInt8(val); }
uint8_t fp16_t::toUInt8() { return fp16ToUInt8(val); }
int16_t fp16_t::toInt16() { return fp16ToInt16(val); }
uint16_t fp16_t::toUInt16() { return fp16ToUInt16(val); }
int32_t fp16_t::toInt32() { return fp16ToInt32(val); }
uint32_t fp16_t::toUInt32() { return fp16ToUInt32(val); }
}
}