Skip to content

N
NiuTrans.Tensor
Commit a89ee126 by xuchen
Options

opimize the float16 CPU implementation

parent 89ad96e6
正在显示 包含 227 行增加 和 55 行删除
source/Main.cpp
......@@ -36,11 +36,65 @@ using namespace nts;
using namespace fnnlm;
using namespace transformer;
int MyTest()
{
float16 x;
printf("%f\n", x.Float());
x = 3.5;
printf("%f\n", x.Float());
x = 0.0F;
printf("%f\n", x.Float());
x.Dump();
x = -3.5;
printf("%f\n", x.Float());
printf("%d\n", sizeof(float16));
FILE* f = fopen("test_fp16", "w");
fwrite(&x, sizeof(float16), 1, f);
fclose(f);
FILE* f2 = fopen("test_fp16", "r");
fread(&x, sizeof(float16), 1, f2);
fclose(f2);
printf("%f\n", x.Float());
return 0;
}
int MyTest2()
{
GDevs.Init();
GDevs.Clear();
XTensor a;
InitTensor2D(&a, 2, 3, X_FLOAT, 0);
a.SetZeroAll();
ScaleAndShift(a, 1);
a.Dump();
printf("dump\n");
getchar();
return 0;
}
int main( int argc, const char ** argv )
{
//_CrtSetDbgFlag(_CrtSetDbgFlag(_CRTDBG_REPORT_FLAG) | _CRTDBG_LEAK_CHECK_DF);
//_CrtSetBreakAlloc(2708);
//MyTest2();
//printf("release\n");
//getchar();
//GDevs.GPUs[0].Reset();
//printf("reset\n");
//getchar();
//printf("bye.\n");
MyTest();
exit(1);
if(argc > 1 && !strcmp(argv[1], "-test"))
Test();
else if(argc > 1 && !strcmp(argv[1], "-fnnlm"))
......
source/tensor/XName.cpp
......@@ -55,6 +55,10 @@ const char * GetOPName(int type)
return "M_ROUND";
else if (type == MATH_RECIPROCAL)
return "M_RECIPROCAL";
else if (type == MATH_EQUAL)
return "M_EQUAL";
else if (type == MATH_NOTEQUAL)
return "M_NOTEQUAL";
else if (type == MATH_CLIP)
return "M_CLIP";
else if (type == MATH_DIV)
......@@ -67,6 +71,10 @@ const char * GetOPName(int type)
return "M_MATRIXMUL";
else if (type == MATH_MATRIXMULBATCHED)
return "M_MATRIXMULBATCHED";
else if (type == MATH_MAX)
return "M_MAX";
else if (type == MATH_MIN)
return "M_MIN";
else if (type == MATH_MULTIPLY)
return "M_MULTIPLY";
else if (type == MATH_MULTIPLYDIM)
......
source/tensor/XName.h
......@@ -46,7 +46,10 @@ namespace nts { // namespace nts(NiuTrans.Tensor)
#define MATH_ROUND MATH_TAN + 1
#define MATH_RECIPROCAL MATH_ROUND + 1
#define MATH_CLIP MATH_RECIPROCAL + 1
#define MATH_EQUAL MATH_RECIPROCAL + 1
#define MATH_NOTEQUAL MATH_EQUAL + 1
#define MATH_CLIP MATH_NOTEQUAL + 1
#define MATH_DIV MATH_CLIP + 1
#define MATH_DIVDIM MATH_DIV + 1
#define MATH_MASK MATH_DIVDIM + 1
......
source/tensor/XTensor.cpp
......@@ -1784,9 +1784,15 @@ void XTensor::BinaryDump(FILE* file)
switch (dataType) {
case X_INT: {
fwrite(tmp.data, sizeof(int), unitNum, file);
break;
}
case X_FLOAT16: {
fwrite(tmp.data, sizeof(float16), unitNum, file);
break;
}
default: {
fwrite(tmp.data, sizeof(float), unitNum, file);
break;
}
}
}
......@@ -1917,12 +1923,21 @@ void XTensor::BinaryRead(FILE* file, size_t offset)
fread(d, sizeof(int), unitNum, file);
SetData(d, unitNum);
delete[] d;
break;
}
case X_FLOAT16: {
int* d = new int[unitNum];
fread(d, sizeof(float16), unitNum, file);
SetData(d, unitNum);
delete[] d;
break;
}
default: {
float* d = new float[unitNum];
fread(d, sizeof(float), unitNum, file);
SetData(d, unitNum);
delete[] d;
break;
}
}
}
......
source/tensor/core/getandset/SetData.cu
......@@ -51,6 +51,7 @@ void KernelSetDataFixed(T * d, T v, int size)
template __global__ void KernelSetDataFixed<int>(int *, int, int);
template __global__ void KernelSetDataFixed<float>(float *, float, int);
template __global__ void KernelSetDataFixed<double>(double *, double, int);
template __global__ void KernelSetDataFixed<__half>(__half*, __half, int);
/*
generate data items with a fixed value
......@@ -79,6 +80,8 @@ void _CudaSetDataFixed(XTensor * tensor, T value)
KernelSetDataFixed << <blocks, threads >> > ((float*)tensor->data, (float)value, tensor->unitNum);
else if (tensor->dataType == X_DOUBLE)
KernelSetDataFixed << <blocks, threads >> > ((double*)tensor->data, (double)value, tensor->unitNum);
else if (tensor->dataType == X_FLOAT16)
KernelSetDataFixed << <blocks, threads >> > ((__half*)tensor->data, (__half)value, tensor->unitNum);
else
ShowNTErrors("TODO! Unsupported datatype!")
......
source/tensor/core/math/Compare.cpp
......@@ -92,6 +92,10 @@ XTensor funcName(const XTensor &a, DTYPE number)
XTensor b(&a); \
b.SetTMPFlag(); \
_funcName(&a, &b, number); \
if (a.enableGrad) { \
XLink::MakeLink(&a, NULL, &b, operationId); \
XLink::AddParamToHead(&b, (DTYPE)number); \
} \
return b; \
}
......@@ -102,6 +106,10 @@ void funcName(const XTensor &a, XTensor &b, DTYPE number)
InitTensorV2(&b, &a); \
} \
_funcName(&a, &b, number); \
if (a.enableGrad) { \
XLink::MakeLink(&a, NULL, &b, operationId); \
XLink::AddParamToHead(&b, (DTYPE)number); \
} \
}
// I think we needn't to make link.
......@@ -186,6 +194,9 @@ XTensor funcName(const XTensor & a, const XTensor & b)
XTensor c(&a); \
c.SetTMPFlag(); \
_funcName(&a, &b, &c); \
if (a.enableGrad && b.enableGrad) { \
XLink::MakeLink(&a, &b, &c, operationId); \
} \
return c; \
}
......@@ -196,16 +207,33 @@ void funcName(const XTensor &a, const XTensor &b, XTensor c)
InitTensor(&c, &a); \
} \
_funcName(&a, &b, &c); \
if (a.enableGrad && b.enableGrad) { \
XLink::MakeLink(&a, &b, &c, operationId); \
} \
}
#ifdef USE_CUDA
_SIMPLE_MAX_MIN_FUNCTION(_Equal, _CudaEqual, myIsEqual)
_SIMPLE_MAX_MIN_FUNCTION(_NotEqual, _CudaNotEqual, myIsNotEqual)
_SIMPLE_MAX_MIN_FUNCTION(_Max, _CudaMax, MAX)
_SIMPLE_MAX_MIN_FUNCTION(_Min, _CudaMin, MIN)
#else
_SIMPLE_MAX_MIN_FUNCTION(_Equal, myIsEqual)
_SIMPLE_MAX_MIN_FUNCTION(_NotEqual, myIsNotEqual)
_SIMPLE_MAX_MIN_FUNCTION(_Max, MAX)
_SIMPLE_MAX_MIN_FUNCTION(_Min, MIN)
#endif
_SIMPLE_MAX_MIN_FUNCTION_ME(_EqualMe, _Equal)
SIMPLE_MAX_MIN_FUNCTION_ME(EqualMe, _Equal)
SIMPLE_MAX_MIN_FUNCTION(Equal, _Equal, MATH_EQUAL)
SIMPLE_MAX_MIN_FUNCTION_VOID(Equal, _Equal, MATH_EQUAL)
_SIMPLE_MAX_MIN_FUNCTION_ME(_NotEqualMe, _NotEqual)
SIMPLE_MAX_MIN_FUNCTION_ME(NotEqualMe, _NotEqual)
SIMPLE_MAX_MIN_FUNCTION(NotEqual, _NotEqual, MATH_NOTEQUAL)
SIMPLE_MAX_MIN_FUNCTION_VOID(NotEqual, _NotEqual, MATH_NOTEQUAL)
_SIMPLE_MAX_MIN_FUNCTION_ME(_MaxMe, _Max)
SIMPLE_MAX_MIN_FUNCTION_ME(MaxMe, _Max)
SIMPLE_MAX_MIN_FUNCTION(Max, _Max, MATH_MAX)
......
source/tensor/core/math/Compare.cu
......@@ -134,6 +134,9 @@ void _Cuda##funcName(const XTensor * a, const XTensor * b, XTensor * c) \
BacktoCudaDev(a->devID, devIDBackup); \
}
SIMPLE_MAX_MIN_FUNCTION_GPU(Equal, cudaIsEqual)
SIMPLE_MAX_MIN_FUNCTION_GPU(NotEqual, cudaIsNotEqual)
SIMPLE_MAX_MIN_FUNCTION_GPU(Max, max)
SIMPLE_MAX_MIN_FUNCTION_GPU(Min, min)
......
source/tensor/core/math/Compare.cuh
......@@ -31,9 +31,15 @@ namespace nts{ // namespace nts(NiuTrans.Tensor)
/* check whether every entry is equal to the given value (cuda version) */
void _CudaEqual(const XTensor * a, XTensor * b, DTYPE value);
/* check whether every entry is equal to the given value (cuda version) */
void _CudaEqual(const XTensor * a, const XTensor * b, XTensor * c);
/* check whether every entry is not equal to the given value (cuda version) */
void _CudaNotEqual(const XTensor * a, XTensor * b, DTYPE value);
/* check whether every entry is not equal to the given value (cuda version) */
void _CudaNotEqual(const XTensor * a, const XTensor * b, XTensor * c);
/* return maximum of two tensor for each items (cuda version) */
void _CudaMax(const XTensor * a, const XTensor * b, XTensor *c);
......
source/tensor/core/math/Compare.h
......@@ -39,7 +39,23 @@ void EqualMe(XTensor & a, DTYPE value);
XTensor Equal(const XTensor & a, DTYPE value);
/* check whether every entry is equal to the given value */
void Equal(const XTensor & a, XTensor & b, DTYPE value);
void Equal(const XTensor & a, XTensor & b, XTensor & c);
/* check whether every entry is equal to the given value */
void _Equal(const XTensor * a, const XTensor * b, XTensor * c);
/* check whether every entry is equal to the given value (do it on site) */
void _EqualMe(XTensor * a, XTensor * b);
/* check whether every entry is equal to the given value (do it on site) */
void EqualMe(XTensor & a, XTensor & b);
/* check whether every entry is equal to the given value (return an XTensor structure) */
XTensor Equal(const XTensor & a, const XTensor & b);
/* check whether every entry is equal to the given value */
void Equal(const XTensor & a, const XTensor & b, XTensor & c);
/* check whether every entry is not equal to the given value */
void _NotEqual(const XTensor * a, XTensor * b, DTYPE value);
......@@ -56,6 +72,22 @@ XTensor NotEqual(const XTensor & a, DTYPE value);
/* check whether every entry is not equal to the given value */
void NotEqual(const XTensor & a, XTensor & b, DTYPE value);
/* check whether every entry is not equal to the given value */
void _NotEqual(const XTensor * a, const XTensor * b, XTensor * c);
/* check whether every entry is not equal to the given value (do it on site) */
void _NotEqualMe(XTensor * a, XTensor * b);
/* check whether every entry is not equal to the given value (do it on site) */
void NotEqualMe(XTensor & a, XTensor * b);
/* check whether every entry is not equal to the given value (return an XTensor structure) */
XTensor NotEqual(const XTensor & a, const XTensor & b);
/* check whether every entry is not equal to the given value */
void NotEqual(const XTensor & a, const XTensor & b, XTensor & c);
/* return maximum of two tensor for each items */
void _Max(const XTensor * a, const XTensor * b, XTensor * c);
......@@ -71,6 +103,7 @@ XTensor Max(const XTensor & a, const XTensor & b);
/* return maximum of two tensor for each items */
void Max(const XTensor & a, const XTensor & b, XTensor & c);
/* return minimum of two tensor for each items */
void _Min(const XTensor * a, const XTensor * b, XTensor * c);
......
source/tensor/core/utilities/Float16.h
//
// float16.h
// 16bit
//
// Created by 管胡昊 on 2020/2/5.
// Copyright © 2020 管胡昊. All rights reserved.
//
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020, Natural Language Processing Lab, Northestern University.
* 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.
*/
/*
* $Creted by: Guan Huhao 2020-02-05
* $Updated by: Xu Chen (email: hello_master1954@163.com) 2020-05-01
*/
#ifndef FLOAT16_H
#define FLOAT16_H
namespace nts { // namespace nts(NiuTrans.Tensor)
struct float16
{
//private member variable
private:
/*
sign is the sign bit 1 means negative, 0 means positive
exp is the exponent with 16 offset
data is the data,similar to ieee-754,the highest is default 1 and ignored
data is the data, similar to ieee-754, the highest is default 1 and ignored
*/
unsigned short data : 10;
unsigned short exp : 5;
unsigned short sign : 1;
// mask for calculate the highest 1
static unsigned int mask[32];
static unsigned int pow2[32];
// private function
int FindHighOne(const int &num, int &l, int &r);
//int FindHighOne(const int &num, int &l, int &r);
int AbsCompare(const float16 & a,const float16 & b);
public:
unsigned short data : 10;
unsigned short exp : 5;
unsigned short sign : 1;
float16 SetOverFlow();
// judge whether overflow
int IsOverlFlow() const;
/* constructor by sign, exp, data
sign:1bit exp:5bit data:10bit similar to ieee 32 floating point */
/* constructor by (sign, exp, data)
similar to ieee 32 floating point
sign: 1bit
exp: 5bit
data: 10bit */
float16(const int& s, const int& e, const int& d);
/* default constructor
This initializes the 16bit floating point to 0. */
float16();
// constructor by a 32-bit floating point
// constructor by a 32-bit float num
float16(const float& data);
template<class T> float16(const T& data);
// constructor by other datatype
//template<class T> float16(const T &data);
template<class T> float16(const T& data);
void Dump();
// change float16 to flaot as you can see the result is a 32-bit floating point
// convert float16 to float and return
float Float();
/* assignment function and tempalte function
float assignment function is the basic function
template assignment function is force change other datetype to float
then call the float assignment function
template assignment function now support int, double */
float16 operator = (const float16& data);
Float assignment function is the basic function.
Template assignment function is force change other datetype to float,
then call the float assignment function.
Template assignment function now support int and double. */
float16 operator = (const float& data);
template<class T> float16 operator = (const T& data);
float16 operator = (const float16& data);
template<class T> float16 operator = (const T& data);
// overload operator (less than) eg. a<b
// overload operator (less than) a < b
int operator < (const float16& data);
template<class T> int operator <(const T& data);
template<class T> int operator < (const T& data);
// overload opertator <= (less or equal than) a<=b
// overload opertator <= (less or equal than) a <= b
int operator <= (const float16& data);
template<class T> int operator <=(const T& data);
template<class T> int operator <= (const T& data);
// overload operator (greater than) eg. a>b
// overload operator (greater than) a > b
int operator > (const float16& data);
template<class T> int operator >(const T& data);
template<class T> int operator > (const T& data);
//overload opertator <= (greater or equal than) a>=b
// overload opertator >= (greater or equal than) a >= b
int operator >= (const float16& data);
template<class T> int operator >=(const T& data);
template<class T> int operator >= (const T& data);
// overload operator + (add) eg. a+b
// overload operator + (add) a + b
float16 operator + (const float16& data);
template<class T> float16 operator +(const T& data);
template<class T> float16 operator + (const T& data);
// overload operator += (add) eg. a+=b
// overload operator += (add) a += b
float16 operator += (const float16& data);
template<class T> float16 operator +=(const T& data);
template<class T> float16 operator += (const T& data);
// overload operator -(negetive) eg. -a
// overload operator -(negetive) -a
float16 operator - ();
// overload operator - (substraction) eg. a-b
// overload operator - (substraction) a - b
float16 operator - (const float16& data);
template<class T> float16 operator -(const T& data);
template<class T> float16 operator - (const T& data);
// overload operator -= (substraction) eg. a-=b
// overload operator -= (substraction) a -= b
float16 operator -= (const float16& data);
template<class T> float16 operator -=(const T& data);
template<class T> float16 operator -= (const T& data);
// overload operator * (multiple) eg. a*b
// overload operator * (multiple) a * b
float16 operator * (const float16& data);
template<class T> float16 operator *(const T& data);
template<class T> float16 operator * (const T& data);
// overload operator *= (multiple) eg. a*=b
// overload operator *= (multiple) a *= b
float16 operator *= (const float16& data);
template<class T> float16 operator *=(const T& data);
template<class T> float16 operator *= (const T& data);
// overload operator / (division) eg. a/b
// overload operator / (division) a / b
float16 GetInverse() const;
float16 operator / (const float16& data);
template<class T> float16 operator /(const T& data);
template<class T> float16 operator / (const T& data);
// overload operator /= (division) eg. a/=b
// overload operator /= (division) a /= b
float16 operator /= (const float16& data);
template<class T> float16 operator /=(const T& data);
template<class T> float16 operator /= (const T& data);
};
} // namespace nts(NiuTrans.Tensor)
#endif /* FLOAT16_H */
Markdown 格式
0% 或
您添加了 0 人 到此讨论。请谨慎行事。
请先完成此评论的编辑！
请 注册 或者 后发表评论