matrix_test.h

来自「C++编写的高性能矩阵乘法的Stranssen算法」· C头文件 代码 · 共 109 行

/* Avoid multiple inclusion of this header file */
#ifndef PRISM_MATRIX_TEST_H  
#define PRISM_MATRIX_TEST_H  

#if PRISM_MEM_CHK
#include "tr_include.h"
#endif

#if PRISM_CRAY
#define PRISM_STR_MACHINE "CRAY"
#elif PRISM_T3D
#define PRISM_STR_MACHINE "T3D"
#elif PRISM_SP
#define PRISM_STR_MACHINE "SP"
#elif PRISM_SUN
#define PRISM_STR_MACHINE "SUN"
#else
#define PRISM_STR_MACHINE "UNKNOWN"
#endif

#define STRAS_ERR_MAX 1.0e-5

#if STRAS_PRISM_TIME
/* defines for timer prism_time */
extern double prism_time();
#define CLOCK() prism_time()
#elif PRISM_SP
extern int mclock();
#define CLOCK() (double)(mclock()/100.0)
#elif PRISM_T3D
extern long _rtc();
#define CLK_TICK 150000000.0 
#define CLOCK() (double)((double)_rtc()/CLK_TICK)
#else
/* defines for C library timer clock() */
#if !PRISM_CRAY
extern clock_t clock();
#endif
#define CLK_TICK 1000000.0 
#define CLOCK() (double)((double)clock()/CLK_TICK)
#endif

#define CLOCK_START(A) A=CLOCK()
#define CLOCK_UPDATE(A,B,C,D) A+=CLOCK()-(B)

int 
i_trash,
i_sum_clock
;

double
d_correct1,
d_correct2
;

double t0,t1,gt0,gt1;
#if STRAS_TIME_PARTS
double time_matrix_add, 
time_matrix_sub, 
time_matrix_linear, 
time_matrix_prod, 
time_submatrix,
time_matrix, 
time_free_matrix,
time_fixup_s,
time_fixup_e; 
#endif

/* Used in random matrix generation */
extern double drand48();

extern int i_sub(int i);
extern void print_matrix(double *a,int lda,int rows,int cols);

extern double aij(int i,int j);
extern double bij(int i,int j);
extern double cij(int i,int j);
extern double cij_result(char c_transa,char c_transb,int k,int i,int j,double alpha,double beta);

/* Vendor/Yale Strassen implementations */
#if PRISM_CRAY || PRISM_T3D
#define dgemms_ SGEMMS
#define dgemmb_ SGEMMB
#elif PRISM_SP
#define dgemms_ dgemms
#define dgemmb_ dgemmb
#endif
extern void dgemmb_(char *transa,char *transb,int *m,int *n,int *k,
		   double *alpha,double *a,int *lda,double *b,
		   int *ldb,double *beta,double *c,int *ldc);

#endif

/* Vendor (FORTRAN) Strassen implementations */
#if PRISM_SP && STRAS_ESSL
extern void dgemms_(double *a,int *lda,char *transa,double *b,int *ldb,
		    char *transb,double *c,int *ldc,int *m,int *k,int *n,
		    double *d_aux,int *i_naux);
#elif PRISM_CRAY && STRAS_CRAY
/* Take care of character data between C and FORTRAN */
extern void dgemms_(FCHAR transa,FCHAR transb,int *m,int *n,int *k,double *alpha,
		    double *a,int *lda,double *b,int *ldb,double *beta,double *c,
		    int *ldc,double *d_aux);
#define SGEMMS(a1,a2,a3,a4,a5,a6,a7,a8,a9,a10,a11,a12,a13,a14)\
  cf1=_cptofcd(a1,strlen(a1));cf2=_cptofcd(a2,strlen(a2));SGEMMS(cf1,cf2,a3,a4,a5,a6,a7,a8,a9,a10,a11,a12,a13,a14)
#endif