【神经网络与深度学习】【C/C++】使用blas做矩阵乘法

 

#define min(x,y) (((x) < (y)) ? (x) : (y))#include <stdio.h>#include <stdlib.h>#include <cublas_v2.h>#include <iostream>#include <vector>//extern "C"//{   #include <cblas.h>//}using namespace std;int main(){    const enum CBLAS_ORDER Order=CblasRowMajor;    const enum CBLAS_TRANSPOSE TransA=CblasNoTrans;    const enum CBLAS_TRANSPOSE TransB=CblasNoTrans;    const int M=4;//A的行数，C的行数    const int N=2;//B的列数，C的列数    const int K=3;//A的列数，B的行数    const float alpha=1;    const float beta=0;    const int lda=K;//A的列    const int ldb=N;//B的列    const int ldc=N;//C的列    const float A[M*K]={1,2,3,4,5,6,7,8,9,8,7,6};    const float B[K*N]={5,4,3,2,1,0};    float C[M*N];       cblas_sgemm(Order, TransA, TransB, M, N, K, alpha, A, lda, B, ldb, beta, C, ldc);         for(int i=0;i<M;i++)    {       for(int j=0;j<N;j++)       {           cout<<C[i*N+j]<<"\n";       }       cout<<endl;    }       return EXIT_SUCCESS; }

g++ testblas.c++ -lopenblas  -o testout

g++ testblas.c++ -lopenblas_piledriverp-r0.2.9 -o testout   本地编译openblas版本

注意library放在引用library的函数的后面

cblas_sgemmMultiplies two matrices (single-precision).void cblas_sgemm (const enum CBLAS_ORDER Order,  // Specifies row-major (C) or column-major (Fortran) data ordering.//typedef enum CBLAS_ORDER     {CblasRowMajor=101, CblasColMajor=102} CBLAS_ORDER;const enum CBLAS_TRANSPOSE TransA,//Specifies whether to transpose matrix A.const enum CBLAS_TRANSPOSE TransB,const int M,   //Number of rows in matrices A and C.const int N,//Number of rows in matrices A and C.const int K,  //Number of columns in matrix A; number of rows in matrix Bconst float alpha, //Scaling factor for the product of matrices A and Bconst float *A, const int lda, //The size of the first dimention of matrix A; if you are passing a matrix A[m][n], the value should be m.  stridelda, ldb and ldc (the strides) are not relevant to my problem after all, but here's an explanation of them : The elements of a matrix (i.e a 2D array) are stored contiguously in memory. However, they may be stored in either column-major or row-major fashion. The stride represents the distance in memory between elements in adjacent rows (if row-major) or in adjacent columns (if column-major). This means that the stride is usually equal to the number of rows/columns in the matrix.Matrix A =[1 2 3][4 5 6]Row-major stores values as {1,2,3,4,5,6}Stride here is 3Col-major stores values as {1, 4, 2, 5, 3, 6}Stride here is 2Matrix B =[1 2 3][4 5 6][7 8 9]Col-major storage is {1, 4, 7, 2, 5, 8, 3, 6, 9}Stride here is 3Read more: http://www.physicsforums.com const float *B,  const int ldb,  //The size of the first dimention of matrix B; if you are passing a matrix B[m][n], the value should be m.const float beta,  //Scaling factor for matrix C.float *C,const int ldc    //The size of the first dimention of matrix C; if you are passing a matrix C[m][n], the value should be m.);Thus, it calculates eitherC←αAB + βCorC←αBA + βCwith optional use of transposed forms of A, B, or both.

 

typedef enum CBLAS_ORDER     {CblasRowMajor=101, CblasColMajor=102} CBLAS_ORDER;typedef enum CBLAS_TRANSPOSE {CblasNoTrans=111, CblasTrans=112, CblasConjTrans=113, CblasConjNoTrans=114} CBLAS_TRANSPOSE;

C=A∗BC=A∗B

CT=(A∗B)T=BT∗ATCT=(A∗B)T=BT∗AT  把A和B的顺序颠倒，可以直接得到转制矩阵乘法的结果，不用作其他变换，（结果C也是转制）。

 

 Y←αAX + βY

cblas_sgemvMultiplies a matrix by a vector (single precision).

void cblas_sgemv (const enum CBLAS_ORDER Order,const enum CBLAS_TRANSPOSE TransA,const int M,const int N,const float alpha,const float *A,const int lda,const float *X,const int incX,const float beta,float *Y,const int incY);

 

 

 

 

 

 

 

 

 

 

 

STL版本

cblas_daxpy
Computes a constant times a vector plus a vector (double-precision).　　

On return, the contents of vector Y are replaced with the result. The value computed is (alpha * X[i]) +
Y[i].

#include <OpenBlas/cblas.h>#include <OpenBlas/common.h>#include <iostream>#include <vector>int main(){    blasint n = 10;    blasint in_x =1;    blasint in_y =1;    std::vector<double> x(n);    std::vector<double> y(n);    double alpha = 10;    std::fill(x.begin(),x.end(),1.0);    std::fill(y.begin(),y.end(),2.0);    cblas_daxpy( n, alpha, &x[0], in_x, &y[0], in_y);    //Print y     for(int j=0;j<n;j++)        std::cout << y[j] << "\t";    std::cout << std::endl;}

 

cublas

cublasStatus_t
cublasCreate(cublasHandle_t *handle)

Return Value MeaningCUBLAS_STATUS_SUCCESS the initialization succeededCUBLAS_STATUS_NOT_INITIALIZED the CUDATM Runtime initialization failedCUBLAS_STATUS_ALLOC_FAILED the resources could not be allocated

cublasStatus_tcublasDestroy(cublasHandle_t handle)

Return Value MeaningCUBLAS_STATUS_SUCCESS the shut down succeededCUBLAS_STATUS_NOT_INITIALIZED the library was not initialized

 

cublasStatus_t cublasSgemm(cublasHandle_t handle,  // 唯一的不同：handle to the cuBLAS library context.

cublasOperation_t transa, cublasOperation_t transbint m, int n, int k,const float *alpha,const float*A, int lda,const float*B, int ldb,const float*beta,float*C, int ldc)

void cblas_sgemm (const enum CBLAS_ORDER Order,  // Specifies row-major (C) or column-major (Fortran) data ordering.//typedef enum CBLAS_ORDER     {CblasRowMajor=101, CblasColMajor=102} CBLAS_ORDER;const enum CBLAS_TRANSPOSE TransA,//Specifies whether to transpose matrix A.const enum CBLAS_TRANSPOSE TransB,const int M,   //Number of rows in matrices A and C.const int N,//Number of rows in matrices A and C.const int K,  //Number of columns in matrix A; number of rows in matrix Bconst float alpha, //Scaling factor for the product of matrices A and Bconst float *A, const int lda, //The size of the first dimention of matrix A; if you are passing a matrix A[m][n], the value should be m.const float *B,  const int ldb,  //The size of the first dimention of matrix B; if you are passing a matrix B[m][n], the value should be m.const float beta,  //Scaling factor for matrix C.float *C,const int ldc    //The size of the first dimention of matrix C; if you are passing a matrix C[m][n], the value should be m.);