Skip to content

N
NiuTrans.Tensor
Commit 6d05d6c8 by 张裕浩
Options

clean code, remove variable useBLAS and useCUDA permanently.

parent 1f71eb10
正在显示 包含 217 行增加 和 185 行删除
source/tensor/XBLAS.h
......@@ -34,7 +34,6 @@ namespace nts{
/* some of the code below is from OpenBLAS (https://github.com/xianyi/OpenBLAS) */
//#define OPENBLAS
#define OPENBLAS_CONST const
typedef int BLASINT;
......@@ -141,64 +140,6 @@ extern "C" int XBLAS_GET_CORE_NUM();
/* get the parallelization type used by OpenBLAS */
//extern "C" int (*XBLAS_GET_PARALLEL_TYPE)(void);
/* linux systems */
#ifndef _WIN32
/* cblas functions that are imported from the lib. See cblas.h in OpenBlas for more information */
extern "C" void cblas_sgemm(OPENBLAS_CONST enum CBLAS_ORDER Order, OPENBLAS_CONST enum CBLAS_TRANSPOSE TransA, OPENBLAS_CONST enum CBLAS_TRANSPOSE TransB,
OPENBLAS_CONST BLASINT M, OPENBLAS_CONST BLASINT N, OPENBLAS_CONST BLASINT K, OPENBLAS_CONST float alpha,
OPENBLAS_CONST float *A, OPENBLAS_CONST BLASINT lda,
OPENBLAS_CONST float *B, OPENBLAS_CONST BLASINT ldb,
OPENBLAS_CONST float beta, float *C, OPENBLAS_CONST BLASINT ldc);
extern "C" void cblas_dgemm(OPENBLAS_CONST enum CBLAS_ORDER Order, OPENBLAS_CONST enum CBLAS_TRANSPOSE TransA, OPENBLAS_CONST enum CBLAS_TRANSPOSE TransB,
OPENBLAS_CONST BLASINT M, OPENBLAS_CONST BLASINT N, OPENBLAS_CONST BLASINT K, OPENBLAS_CONST double alpha,
OPENBLAS_CONST double *A, OPENBLAS_CONST BLASINT lda,
OPENBLAS_CONST double *B, OPENBLAS_CONST BLASINT ldb,
OPENBLAS_CONST double beta, double *C, OPENBLAS_CONST BLASINT ldc);
extern "C" void cblas_sger (OPENBLAS_CONST enum CBLAS_ORDER order, OPENBLAS_CONST BLASINT M, OPENBLAS_CONST BLASINT N, OPENBLAS_CONST float alpha,
OPENBLAS_CONST float *X, OPENBLAS_CONST BLASINT incX, OPENBLAS_CONST float *Y, OPENBLAS_CONST BLASINT incY,
float *A, OPENBLAS_CONST BLASINT lda);
extern "C" void cblas_dger (OPENBLAS_CONST enum CBLAS_ORDER order, OPENBLAS_CONST BLASINT M, OPENBLAS_CONST BLASINT N, OPENBLAS_CONST double alpha,
OPENBLAS_CONST double *X, OPENBLAS_CONST BLASINT incX, OPENBLAS_CONST double *Y, OPENBLAS_CONST BLASINT incY,
double *A, OPENBLAS_CONST BLASINT lda);
extern "C" void cblas_saxpy (OPENBLAS_CONST BLASINT n, OPENBLAS_CONST float a, OPENBLAS_CONST float *x, OPENBLAS_CONST BLASINT incx, OPENBLAS_CONST float *y, OPENBLAS_CONST BLASINT incy);
extern "C" void cblas_daxpy (OPENBLAS_CONST BLASINT n, OPENBLAS_CONST double a, OPENBLAS_CONST double *x, OPENBLAS_CONST BLASINT incx, OPENBLAS_CONST double *y, OPENBLAS_CONST BLASINT incy);
#endif
#if defined(OPENBLAS)
/* better control of multi-threading */
extern "C" void openblas_set_num_threads(int num_threads);
extern "C" void goto_set_num_threads(int num_threads);
//extern "C" int openblas_get_num_threads(void);
extern "C" int openblas_get_num_procs(void);
extern "C" void cblas_saxpy(OPENBLAS_CONST BLASINT n, OPENBLAS_CONST float a, OPENBLAS_CONST float *x, OPENBLAS_CONST BLASINT incx, OPENBLAS_CONST float *y, OPENBLAS_CONST BLASINT incy);
extern "C" void cblas_daxpy(OPENBLAS_CONST BLASINT n, OPENBLAS_CONST double a, OPENBLAS_CONST double *x, OPENBLAS_CONST BLASINT incx, OPENBLAS_CONST double *y, OPENBLAS_CONST BLASINT incy);
//extern "C" char* openblas_get_config(void);
//extern "C" char* openblas_get_corename(void);
//extern "C" int openblas_get_parallel(void);
#endif
#if defined(MKL)
/* better control of multi-threading */
//_Mkl_Api(void,MKL_Set_Num_Threads,(int nth))
//_Mkl_Api(int,MKL_Get_Max_Threads,(void))
//extern "C" void MKL_Set_Num_Threads(int num_threads);
//extern "C" int MKL_Get_Max_Threads();
//extern "C" void cblas_saxpy(OPENBLAS_CONST BLASINT n, OPENBLAS_CONST float a, OPENBLAS_CONST float *x, OPENBLAS_CONST BLASINT incx, OPENBLAS_CONST float *y, OPENBLAS_CONST BLASINT incy);
//#define mkl_set_num_threads MKL_Set_Num_Threads
//#define mkl_get_max_num_threads MKL_Get_Max_Threads
//extern "C" void mkl_set_num_threads(int num_threads);
//extern "C" void omp_set_num_threads(int num_threads);
//extern "C" int mkl_get_max_num_threads();
#endif
#if defined(CUDA_BLAS)
......
source/tensor/XGlobal.cpp
......@@ -50,8 +50,6 @@ int CONST_MINUSONE = -1;
bool CONST_TRUE = true;
int verboseLevel = 0;
bool useBLAS = false;
bool useCUDA = false;
FILE * tmpLog = NULL;
double myTime = 0;
......
source/tensor/XGlobal.h
......@@ -135,8 +135,6 @@ extern bool CONST_TRUE;
#define NIUTRANSNNDEBUG
extern int verboseLevel;
extern bool useBLAS;
extern bool useCUDA;
#define FFLUSH(FILEH) \
{ \
......@@ -158,8 +156,10 @@ extern bool useCUDA;
/* BLAS interfaces */
#ifdef DOUBELPRICSION
#define GEMM XBLAS_DGEMM
#define AXPY XBLAS_DAXPY
#else
#define GEMM XBLAS_SGEMM
#define AXPY XBLAS_SAXPY
#endif
extern void InitGlobalAll();
......
source/tensor/core/reduce/VectorBuffer.cpp 0 → 100644
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2017, 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.
*/
/*
* $Created by: ZHANG Yuhao (email: zhangyuhao@stu.neu.edu.cn) 2019-07-23
*/
#include "VectorBuffer.h"
namespace nts {
/* data size for each buffer */
int VectorBuffer::size()
{
return 32 / sizeof(DTYPE);
}
/* constructor */
VectorBuffer::VectorBuffer()
{
}
/*
constructor
initial values with val
*/
VectorBuffer::VectorBuffer(DTYPE val)
{
for (int i = 0; i != size(); i++) {
values[i] = val;
}
}
/* load data */
VectorBuffer VectorBuffer::loadu(const DTYPE* ptr, bool isExp , DTYPE power , DTYPE* bias )
{
int count = 32 / sizeof(DTYPE);
VectorBuffer vec;
if (isExp) {
if (bias == NULL) {
if (power == (DTYPE)1.0) {
for (int i = 0; i != count; i++) {
vec.values[i] = (DTYPE)exp(*(ptr + i));
}
}
else if (power == (DTYPE)2.0) {
for (int i = 0; i != count; i++) {
vec.values[i] = (DTYPE)exp((*(ptr + i)) * (*(ptr + i)));
}
}
else if (power == (DTYPE)0.5) {
for (int i = 0; i != count; i++) {
vec.values[i] = (DTYPE)exp(sqrt(*(ptr + i)));
}
}
else {
for (int i = 0; i != count; i++) {
vec.values[i] = (DTYPE)exp(pow(*(ptr + i), power));
}
}
}/*is bias == NULL*/
else {
if (power == (DTYPE)1.0) {
for (int i = 0; i != count; i++) {
vec.values[i] = (DTYPE)exp(*(ptr + i) - bias[i]);
}
}
else if (power == (DTYPE)2.0) {
for (int i = 0; i != count; i++) {
DTYPE value = *(ptr + i) - bias[i];
vec.values[i] = (DTYPE)exp(value * value);
}
}
else if (power == (DTYPE)0.5) {
for (int i = 0; i != count; i++) {
vec.values[i] = (DTYPE)exp(sqrt(*(ptr + i) - bias[i]));
}
}
else {
for (int i = 0; i != count; i++) {
vec.values[i] = (DTYPE)exp(pow(*(ptr + i) - bias[i], power));
}
}
}
}//isExp
else {
if (bias == NULL) {
if (power == (DTYPE)1.0) {
memcpy(vec.values, ptr, count * sizeof(DTYPE));
}
else if (power == (DTYPE)2.0) {
for (int i = 0; i != count; i++) {
vec.values[i] = (*(ptr + i)) * (*(ptr + i));
}
}
else if (power == (DTYPE)0.5) {
for (int i = 0; i != count; i++) {
vec.values[i] = (DTYPE)sqrt(*(ptr + i));
}
}
else {
for (int i = 0; i != count; i++) {
vec.values[i] = (DTYPE)pow(*(ptr + i), power);
}
}
}// if bias == NULL
else {
if (power == (DTYPE)1.0) {
for (int i = 0; i != count; i++) {
vec.values[i] = *(ptr + i) - bias[i];
}
}
else if (power == (DTYPE)2.0) {
for (int i = 0; i != count; i++) {
DTYPE value = *(ptr + i) - bias[i];
vec.values[i] = value * value;
}
}
else if (power == (DTYPE)0.5) {
for (int i = 0; i != count; i++) {
vec.values[i] = (DTYPE)sqrt(*(ptr + i) - bias[i]);
}
}
else {
for (int i = 0; i != count; i++) {
vec.values[i] = (DTYPE)pow(*(ptr + i) - bias[i], power);
}
}
}
}
return vec;
}
/* overloading [] */
const DTYPE& VectorBuffer::operator[](int idx)const
{
return values[idx];
}
/* overloading + */
VectorBuffer VectorBuffer::operator+(const VectorBuffer &a)
{
for (int i = 0; i != a.size(); i++) {
this->values[i] = a[i] + this->values[i];
}
return *this;
}
/* conculte the max of two buffer */
VectorBuffer VectorBuffer::maxData(const VectorBuffer &a) {
for (int i = 0; i != a.size(); i++) {
this->values[i] = MAX(a[i], this->values[i]);
}
return *this;
}
}/* end of the nts (NiuTrans.Tensor) namespace */
\ No newline at end of file
source/tensor/core/reduce/VectorBuffer.h
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2017, 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.
*/
/*
* $Created by: ZHANG Yuhao (email: zhangyuhao@stu.neu.edu.cn) 2019-07-23
*/
#include <cstring>
#include <cmath>
#include "../../XGlobal.h"
namespace nts {
class VectorBuffer {
private:
/* buffer for concluter */
DTYPE values[32 / sizeof(DTYPE)] = { 0 };
public:
static int size() {
return 32 / sizeof(DTYPE);
}
VectorBuffer() {}
VectorBuffer(DTYPE val) {
for (int i = 0; i != size(); i++) {
values[i] = val;
}
}
static VectorBuffer loadu(const DTYPE* ptr, bool isExp = false, DTYPE power = (DTYPE)1.0F, DTYPE* bias = NULL) {
int count = 32 / sizeof(DTYPE);
VectorBuffer vec;
if (isExp) {
if (bias == NULL) {
if (power == (DTYPE)1.0) {
for (int i = 0; i != count; i++) {
vec.values[i] = (DTYPE)exp(*(ptr + i));
}
}
else if (power == (DTYPE)2.0) {
for (int i = 0; i != count; i++) {
vec.values[i] = (DTYPE)exp((*(ptr + i)) * (*(ptr + i)));
}
}
else if (power == (DTYPE)0.5) {
for (int i = 0; i != count; i++) {
vec.values[i] = (DTYPE)exp(sqrt(*(ptr + i)));
}
}
else {
for (int i = 0; i != count; i++) {
vec.values[i] = (DTYPE)exp(pow(*(ptr + i), power));
}
}
}/*is bias == NULL*/
else {
if (power == (DTYPE)1.0) {
for (int i = 0; i != count; i++) {
vec.values[i] = (DTYPE)exp(*(ptr + i) - bias[i]);
}
}
else if (power == (DTYPE)2.0) {
for (int i = 0; i != count; i++) {
DTYPE value = *(ptr + i) - bias[i];
vec.values[i] = (DTYPE)exp(value * value);
}
}
else if (power == (DTYPE)0.5) {
for (int i = 0; i != count; i++) {
vec.values[i] = (DTYPE)exp(sqrt(*(ptr + i) - bias[i]));
}
}
else {
for (int i = 0; i != count; i++) {
vec.values[i] = (DTYPE)exp(pow(*(ptr + i) - bias[i], power));
}
}
}
}//isExp
else {
if (bias == NULL) {
if (power == (DTYPE)1.0) {
memcpy(vec.values, ptr, count * sizeof(DTYPE));
}
else if (power == (DTYPE)2.0) {
for (int i = 0; i != count; i++) {
vec.values[i] = (*(ptr + i)) * (*(ptr + i));
}
}
else if (power == (DTYPE)0.5) {
for (int i = 0; i != count; i++) {
vec.values[i] = (DTYPE)sqrt(*(ptr + i));
}
}
else {
for (int i = 0; i != count; i++) {
vec.values[i] = (DTYPE)pow(*(ptr + i), power);
}
}
}// if bias == NULL
else {
if (power == (DTYPE)1.0) {
for (int i = 0; i != count; i++) {
vec.values[i] = *(ptr + i) - bias[i];
}
}
else if (power == (DTYPE)2.0) {
for (int i = 0; i != count; i++) {
DTYPE value = *(ptr + i) - bias[i];
vec.values[i] = value * value;
}
}
else if (power == (DTYPE)0.5) {
for (int i = 0; i != count; i++) {
vec.values[i] = (DTYPE)sqrt(*(ptr + i) - bias[i]);
}
}
else {
for (int i = 0; i != count; i++) {
vec.values[i] = (DTYPE)pow(*(ptr + i) - bias[i], power);
}
}
}
}
return vec;
}
const DTYPE& operator[](int idx) const {
return values[idx];
}
inline VectorBuffer operator+(const VectorBuffer &a) {
for (int i = 0; i != a.size(); i++) {
this->values[i] = a[i] + this->values[i];
}
return *this;
}
inline VectorBuffer maxData(const VectorBuffer &a) {
for (int i = 0; i != a.size(); i++) {
this->values[i] = MAX(a[i], this->values[i]);
}
return *this;
}
/* data size for each buffer */
static int size();
/* constructor */
VectorBuffer();
/* constructor */
VectorBuffer(DTYPE val);
/* load data */
static VectorBuffer loadu(const DTYPE* ptr, bool isExp = false, DTYPE power = (DTYPE)1.0F, DTYPE* bias = NULL);
/* overloading [] */
const DTYPE& operator[](int idx)const;
/* overloading + */
VectorBuffer operator+(const VectorBuffer &a);
/* conculte the max of two buffer */
VectorBuffer maxData(const VectorBuffer &a);
};
}
\ No newline at end of file
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