wehub-resource-sync
460 lines
14 KiB
C++
460 lines
14 KiB
C++
// Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
|
|
//
|
|
// Licensed under the Apache License, Version 2.0 (the "License");
|
|
// you may not use this file except in compliance with the License.
|
|
// You may obtain a copy of the License at
|
|
//
|
|
// http://www.apache.org/licenses/LICENSE-2.0
|
|
//
|
|
// Unless required by applicable law or agreed to in writing, software
|
|
// distributed under the License is distributed on an "AS IS" BASIS,
|
|
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
// See the License for the specific language governing permissions and
|
|
// limitations under the License.
|
|
|
|
#pragma once
|
|
|
|
#include <stdint.h>
|
|
#include <cmath>
|
|
#include <cstring>
|
|
#include <iostream>
|
|
#include <limits>
|
|
#include "paddle/common/hostdevice.h"
|
|
|
|
#ifdef PADDLE_WITH_CUDA
|
|
#include <cuda.h>
|
|
#endif
|
|
|
|
#if defined(__CUDACC__) && CUDA_VERSION >= 11080
|
|
#define PADDLE_CUDA_FP8
|
|
#include <cuda_fp8.h>
|
|
#endif
|
|
|
|
#ifndef PADDLE_WITH_HIP
|
|
#if !defined(_WIN32)
|
|
#define PADDLE_ALIGN(x) __attribute__((aligned(x)))
|
|
#else
|
|
#define PADDLE_ALIGN(x) __declspec(align(x))
|
|
#endif
|
|
#else
|
|
#define PADDLE_ALIGN(x)
|
|
#endif
|
|
|
|
namespace phi {
|
|
namespace dtype {
|
|
|
|
struct PADDLE_ALIGN(1) float8_e4m3fn {
|
|
public:
|
|
uint8_t x;
|
|
|
|
// Constructors
|
|
float8_e4m3fn() = default;
|
|
float8_e4m3fn(const float8_e4m3fn& o) = default;
|
|
float8_e4m3fn& operator=(const float8_e4m3fn& o) = default;
|
|
float8_e4m3fn(float8_e4m3fn&& o) = default;
|
|
float8_e4m3fn& operator=(float8_e4m3fn&& o) = default;
|
|
~float8_e4m3fn() = default;
|
|
|
|
HOSTDEVICE inline float8_e4m3fn(float val) {
|
|
#if defined(PADDLE_CUDA_FP8)
|
|
__nv_fp8_e4m3 tmp = __nv_fp8_e4m3(val);
|
|
x = *reinterpret_cast<uint8_t*>(&tmp);
|
|
#else
|
|
// CPU implementation.
|
|
Bits fb, denorm_mask;
|
|
fb.f = val;
|
|
constexpr uint32_t fp8_max = UINT32_C(1087) << 20;
|
|
denorm_mask.ui = UINT32_C(141) << 23;
|
|
uint8_t result = 0u;
|
|
const uint32_t sign = fb.ui & UINT32_C(0x80000000);
|
|
fb.ui ^= sign;
|
|
if (fb.ui >= fp8_max) {
|
|
result = 0x7e;
|
|
} else {
|
|
if (fb.ui < (UINT32_C(121) << 23)) {
|
|
fb.f = fb.f + denorm_mask.f;
|
|
fb.ui = fb.ui - denorm_mask.ui;
|
|
result = static_cast<uint8_t>(fb.ui);
|
|
} else {
|
|
uint8_t mant_odd = (fb.ui >> 20) & 1;
|
|
fb.ui += ((uint32_t)(7 - 127) << 23) + 0x7FFFF;
|
|
fb.ui += mant_odd;
|
|
result = static_cast<uint8_t>(fb.ui >> 20);
|
|
}
|
|
}
|
|
|
|
result |= static_cast<uint8_t>(sign >> 24);
|
|
x = result;
|
|
#endif
|
|
}
|
|
|
|
#if defined(PADDLE_CUDA_FP8)
|
|
HOSTDEVICE inline explicit float8_e4m3fn(const __nv_fp8_e4m3& val) {
|
|
x = *reinterpret_cast<const uint8_t*>(&val); // NOLINT
|
|
}
|
|
#endif
|
|
|
|
template <class T>
|
|
HOSTDEVICE inline explicit float8_e4m3fn(const T& val)
|
|
: x(float8_e4m3fn(static_cast<float>(val)).x) {}
|
|
|
|
// Assignment operators
|
|
#if defined(PADDLE_CUDA_FP8)
|
|
HOSTDEVICE inline float8_e4m3fn& operator=(const __nv_fp8_e4m3& val) {
|
|
x = *reinterpret_cast<const uint8_t*>(&val); // NOLINT
|
|
return *this;
|
|
}
|
|
#endif
|
|
|
|
HOSTDEVICE inline float8_e4m3fn& operator=(int8_t val) {
|
|
x = float8_e4m3fn(val).x;
|
|
return *this;
|
|
}
|
|
|
|
HOSTDEVICE inline float8_e4m3fn& operator=(uint8_t val) {
|
|
x = float8_e4m3fn(val).x;
|
|
return *this;
|
|
}
|
|
|
|
HOSTDEVICE inline float8_e4m3fn& operator=(int16_t val) {
|
|
x = float8_e4m3fn(val).x;
|
|
return *this;
|
|
}
|
|
|
|
HOSTDEVICE inline float8_e4m3fn& operator=(uint16_t val) {
|
|
x = float8_e4m3fn(val).x;
|
|
return *this;
|
|
}
|
|
|
|
HOSTDEVICE inline float8_e4m3fn& operator=(int32_t val) {
|
|
x = float8_e4m3fn(val).x;
|
|
return *this;
|
|
}
|
|
|
|
HOSTDEVICE inline float8_e4m3fn& operator=(uint32_t val) {
|
|
x = float8_e4m3fn(val).x;
|
|
return *this;
|
|
}
|
|
|
|
HOSTDEVICE inline float8_e4m3fn& operator=(int64_t val) {
|
|
x = float8_e4m3fn(val).x;
|
|
return *this;
|
|
}
|
|
|
|
HOSTDEVICE inline float8_e4m3fn& operator=(uint64_t val) {
|
|
x = float8_e4m3fn(val).x;
|
|
return *this;
|
|
}
|
|
|
|
HOSTDEVICE inline float8_e4m3fn& operator=(float val) {
|
|
x = float8_e4m3fn(val).x;
|
|
return *this;
|
|
}
|
|
|
|
HOSTDEVICE inline float8_e4m3fn& operator=(double val) {
|
|
x = float8_e4m3fn(val).x;
|
|
return *this;
|
|
}
|
|
|
|
// Conversion operators
|
|
HOSTDEVICE inline operator float() const {
|
|
#ifdef PADDLE_CUDA_FP8
|
|
return static_cast<float>(
|
|
*reinterpret_cast<const __nv_fp8_e4m3*>(&x)); // NOLINT
|
|
#else
|
|
// CPU implementation.
|
|
const uint32_t w = (uint32_t)x << 24;
|
|
const uint32_t sign = w & UINT32_C(0x80000000);
|
|
const uint32_t nonsign = w & UINT32_C(0x7FFFFFFF);
|
|
|
|
// get the leading 0-bits in nonsin.
|
|
uint32_t nonsign_tmp = nonsign;
|
|
uint32_t renorm_shift = 0;
|
|
if (nonsign_tmp == 0) {
|
|
renorm_shift = sizeof(uint32_t) * 8;
|
|
} else {
|
|
if ((nonsign_tmp & 0xFFFF0000) == 0) {
|
|
renorm_shift += 16;
|
|
nonsign_tmp <<= 16;
|
|
}
|
|
if ((nonsign_tmp & 0xFF000000) == 0) {
|
|
renorm_shift += 8;
|
|
nonsign_tmp <<= 8;
|
|
}
|
|
if ((nonsign_tmp & 0xF0000000) == 0) {
|
|
renorm_shift += 4;
|
|
nonsign_tmp <<= 4;
|
|
}
|
|
if ((nonsign_tmp & 0xC0000000) == 0) {
|
|
renorm_shift += 2;
|
|
nonsign_tmp <<= 2;
|
|
}
|
|
if ((nonsign_tmp & 0x80000000) == 0) {
|
|
renorm_shift += 1;
|
|
}
|
|
}
|
|
|
|
renorm_shift = renorm_shift > 4 ? renorm_shift - 4 : 0;
|
|
const int32_t inf_nan_mask =
|
|
((int32_t)(nonsign + 0x01000000) >> 8) & INT32_C(0x7F800000);
|
|
const int32_t zero_mask = (int32_t)(nonsign - 1) >> 31;
|
|
Bits result;
|
|
result.ui =
|
|
sign |
|
|
((((nonsign << renorm_shift >> 4) + ((0x78 - renorm_shift) << 23)) |
|
|
inf_nan_mask) &
|
|
~zero_mask);
|
|
return result.f;
|
|
#endif
|
|
}
|
|
|
|
#ifdef PADDLE_CUDA_FP8
|
|
HOSTDEVICE inline __nv_fp8_e4m3 to_nv_fp8_e4m3() const {
|
|
return *reinterpret_cast<const __nv_fp8_e4m3*>(&x);
|
|
}
|
|
#endif
|
|
|
|
HOSTDEVICE inline explicit operator int8_t() const {
|
|
return static_cast<int8_t>(static_cast<float>(*this));
|
|
}
|
|
|
|
HOSTDEVICE inline explicit operator uint8_t() const {
|
|
return static_cast<uint8_t>(static_cast<float>(*this));
|
|
}
|
|
|
|
HOSTDEVICE inline explicit operator int16_t() const {
|
|
return static_cast<int16_t>(static_cast<float>(*this));
|
|
}
|
|
|
|
HOSTDEVICE inline explicit operator uint16_t() const {
|
|
return static_cast<uint16_t>(static_cast<float>(*this));
|
|
}
|
|
|
|
HOSTDEVICE inline explicit operator int32_t() const {
|
|
return static_cast<int32_t>(static_cast<float>(*this));
|
|
}
|
|
|
|
HOSTDEVICE inline explicit operator uint32_t() const {
|
|
return static_cast<uint32_t>(static_cast<float>(*this));
|
|
}
|
|
|
|
HOSTDEVICE inline explicit operator int64_t() const {
|
|
return static_cast<int64_t>(static_cast<float>(*this));
|
|
}
|
|
|
|
HOSTDEVICE inline explicit operator uint64_t() const {
|
|
return static_cast<uint64_t>(static_cast<float>(*this));
|
|
}
|
|
|
|
private:
|
|
union Bits {
|
|
float f;
|
|
uint32_t ui;
|
|
};
|
|
};
|
|
|
|
HOSTDEVICE inline float8_e4m3fn operator+(const float8_e4m3fn& a,
|
|
const float8_e4m3fn& b) {
|
|
return float8_e4m3fn(static_cast<float>(a) + static_cast<float>(b));
|
|
}
|
|
|
|
HOSTDEVICE inline float8_e4m3fn operator-(const float8_e4m3fn& a,
|
|
const float8_e4m3fn& b) {
|
|
return float8_e4m3fn(static_cast<float>(a) - static_cast<float>(b));
|
|
}
|
|
|
|
HOSTDEVICE inline float8_e4m3fn operator*(const float8_e4m3fn& a,
|
|
const float8_e4m3fn& b) {
|
|
return float8_e4m3fn(static_cast<float>(a) * static_cast<float>(b));
|
|
}
|
|
|
|
HOSTDEVICE inline float8_e4m3fn operator/(const float8_e4m3fn& a,
|
|
const float8_e4m3fn& b) {
|
|
return float8_e4m3fn(static_cast<float>(a) / static_cast<float>(b));
|
|
}
|
|
|
|
HOSTDEVICE inline float8_e4m3fn operator-(const float8_e4m3fn& a) {
|
|
return float8_e4m3fn(-static_cast<float>(a));
|
|
}
|
|
|
|
HOSTDEVICE inline float8_e4m3fn& operator+=(float8_e4m3fn& a, // NOLINT
|
|
const float8_e4m3fn& b) {
|
|
a = float8_e4m3fn(static_cast<float>(a) + static_cast<float>(b));
|
|
return a;
|
|
}
|
|
|
|
HOSTDEVICE inline float8_e4m3fn& operator-=(float8_e4m3fn& a, // NOLINT
|
|
const float8_e4m3fn& b) {
|
|
a = float8_e4m3fn(static_cast<float>(a) - static_cast<float>(b));
|
|
return a;
|
|
}
|
|
|
|
HOSTDEVICE inline float8_e4m3fn& operator*=(float8_e4m3fn& a, // NOLINT
|
|
const float8_e4m3fn& b) {
|
|
a = float8_e4m3fn(static_cast<float>(a) * static_cast<float>(b));
|
|
return a;
|
|
}
|
|
|
|
HOSTDEVICE inline float8_e4m3fn& operator/=(float8_e4m3fn& a, // NOLINT
|
|
const float8_e4m3fn& b) {
|
|
a = float8_e4m3fn(static_cast<float>(a) / static_cast<float>(b));
|
|
return a;
|
|
}
|
|
|
|
HOSTDEVICE inline float8_e4m3fn raw_uint8_to_float8_e4m3fn(uint8_t a) {
|
|
float8_e4m3fn res;
|
|
res.x = a;
|
|
return res;
|
|
}
|
|
|
|
// Comparison operators
|
|
HOSTDEVICE inline bool operator==(const float8_e4m3fn& a,
|
|
const float8_e4m3fn& b) {
|
|
return static_cast<float>(a) == static_cast<float>(b);
|
|
}
|
|
|
|
HOSTDEVICE inline bool operator!=(const float8_e4m3fn& a,
|
|
const float8_e4m3fn& b) {
|
|
return static_cast<float>(a) != static_cast<float>(b);
|
|
}
|
|
|
|
HOSTDEVICE inline bool operator<(const float8_e4m3fn& a,
|
|
const float8_e4m3fn& b) {
|
|
return static_cast<float>(a) < static_cast<float>(b);
|
|
}
|
|
|
|
HOSTDEVICE inline bool operator<=(const float8_e4m3fn& a,
|
|
const float8_e4m3fn& b) {
|
|
return static_cast<float>(a) <= static_cast<float>(b);
|
|
}
|
|
|
|
HOSTDEVICE inline bool operator>(const float8_e4m3fn& a,
|
|
const float8_e4m3fn& b) {
|
|
return static_cast<float>(a) > static_cast<float>(b);
|
|
}
|
|
|
|
HOSTDEVICE inline bool operator>=(const float8_e4m3fn& a,
|
|
const float8_e4m3fn& b) {
|
|
return static_cast<float>(a) >= static_cast<float>(b);
|
|
}
|
|
|
|
HOSTDEVICE inline bool(isnan)(const float8_e4m3fn& a) {
|
|
return (a.x & 0b01111111) == 0b01111111;
|
|
}
|
|
|
|
// TODO(Wanglongzhi2001): According to https://arxiv.org/pdf/2209.05433.pdf,
|
|
// inf in float8_e4m3fn is undefined, to avoid compilation errors,
|
|
// we use 0b01111000 to represent inf for now!
|
|
HOSTDEVICE inline bool(isinf)(const float8_e4m3fn& a) {
|
|
return (a.x & 0b01111111) == 0b01111000;
|
|
}
|
|
|
|
HOSTDEVICE inline bool(isfinite)(const float8_e4m3fn& a) {
|
|
return !((isnan)(a)) && !((isinf)(a));
|
|
}
|
|
|
|
HOSTDEVICE inline float8_e4m3fn(abs)(const float8_e4m3fn& a) {
|
|
return float8_e4m3fn(std::abs(static_cast<float>(a)));
|
|
}
|
|
|
|
inline std::ostream& operator<<(std::ostream& os, const float8_e4m3fn& a) {
|
|
os << static_cast<float>(a);
|
|
return os;
|
|
}
|
|
|
|
} // namespace dtype
|
|
} // namespace phi
|
|
|
|
namespace std {
|
|
|
|
template <>
|
|
struct is_pod<phi::dtype::float8_e4m3fn> {
|
|
static const bool value =
|
|
is_trivial<phi::dtype::float8_e4m3fn>::value &&
|
|
is_standard_layout<phi::dtype::float8_e4m3fn>::value;
|
|
};
|
|
|
|
template <>
|
|
struct is_floating_point<phi::dtype::float8_e4m3fn>
|
|
: std::integral_constant<
|
|
bool,
|
|
std::is_same<phi::dtype::float8_e4m3fn,
|
|
typename std::remove_cv<
|
|
phi::dtype::float8_e4m3fn>::type>::value> {};
|
|
template <>
|
|
struct is_signed<phi::dtype::float8_e4m3fn> {
|
|
static const bool value = true;
|
|
};
|
|
|
|
template <>
|
|
struct is_unsigned<phi::dtype::float8_e4m3fn> {
|
|
static const bool value = false;
|
|
};
|
|
|
|
inline bool isnan(const phi::dtype::float8_e4m3fn& a) {
|
|
return phi::dtype::isnan(a);
|
|
}
|
|
|
|
inline bool isinf(const phi::dtype::float8_e4m3fn& a) {
|
|
return phi::dtype::isinf(a);
|
|
}
|
|
|
|
template <>
|
|
struct numeric_limits<phi::dtype::float8_e4m3fn> {
|
|
static constexpr bool is_specialized = true;
|
|
static constexpr bool is_signed = true;
|
|
static constexpr bool is_integer = false;
|
|
static constexpr bool is_exact = false;
|
|
static constexpr bool has_infinity = false;
|
|
static constexpr bool has_quiet_NaN = true;
|
|
static constexpr bool has_signaling_NaN = false;
|
|
static constexpr auto has_denorm = true;
|
|
static constexpr auto has_denorm_loss = true;
|
|
static constexpr auto round_style = numeric_limits<float>::round_style;
|
|
static constexpr bool is_iec559 = false;
|
|
static constexpr bool is_bounded = true;
|
|
static constexpr bool is_modulo = false;
|
|
static constexpr int digits = 4;
|
|
static constexpr int digits10 = 0;
|
|
static constexpr int max_digits10 = 3;
|
|
static constexpr int radix = 2;
|
|
static constexpr int min_exponent = -5;
|
|
static constexpr int min_exponent10 = -1;
|
|
static constexpr int max_exponent = 8;
|
|
static constexpr int max_exponent10 = 2;
|
|
static constexpr auto traps = numeric_limits<float>::traps;
|
|
static constexpr auto tinyness_before = false;
|
|
|
|
HOSTDEVICE static phi::dtype::float8_e4m3fn(min)() {
|
|
return phi::dtype::raw_uint8_to_float8_e4m3fn(0x08);
|
|
}
|
|
HOSTDEVICE static phi::dtype::float8_e4m3fn lowest() {
|
|
return phi::dtype::raw_uint8_to_float8_e4m3fn(0xFE);
|
|
}
|
|
HOSTDEVICE static phi::dtype::float8_e4m3fn(max)() {
|
|
return phi::dtype::raw_uint8_to_float8_e4m3fn(0x7E);
|
|
}
|
|
HOSTDEVICE static phi::dtype::float8_e4m3fn epsilon() {
|
|
return phi::dtype::raw_uint8_to_float8_e4m3fn(0x20);
|
|
}
|
|
HOSTDEVICE static phi::dtype::float8_e4m3fn round_error() {
|
|
return phi::dtype::raw_uint8_to_float8_e4m3fn(0x30);
|
|
}
|
|
|
|
HOSTDEVICE static phi::dtype::float8_e4m3fn infinity() {
|
|
return phi::dtype::raw_uint8_to_float8_e4m3fn(0x7F);
|
|
}
|
|
HOSTDEVICE static phi::dtype::float8_e4m3fn quiet_NaN() {
|
|
return phi::dtype::raw_uint8_to_float8_e4m3fn(0x7F);
|
|
}
|
|
HOSTDEVICE static phi::dtype::float8_e4m3fn denorm_min() {
|
|
return phi::dtype::raw_uint8_to_float8_e4m3fn(0x01);
|
|
}
|
|
};
|
|
// 新增特化：定义 float 和 float8_e4m3fn 的公共类型为 float
|
|
template <>
|
|
struct common_type<float, phi::dtype::float8_e4m3fn> {
|
|
using type = float;
|
|
};
|
|
} // namespace std
|