mt/ds_matmul.c

   1 #include "stdlib.h"
   2
   3 #include "util.h"
   4
   5 #include "dataset.h"
   6 void __attribute__((noinline)) matmul(const int coreid, const int ncores, const int lda,  const data_t A[], const data_t B[], data_t C[] )
   7 {
   8 int i,j,k,a,b,b1,a1,a2,a3,c,c1,c2,c3,b2,b3;
   9         for (j=coreid*4; j<lda; j+=4*ncores){
  10                 a=j*lda;
  11                 a1=(j+1)*lda;
  12                 a2=(j+2)*lda;
  13                 a3=(j+3)*lda;
  14                 for (k=0;k<lda; k+=2)
  15                 {
  16                         b = k*lda;
  17                         b1 = (k+1)*lda;
  18                         for (i=0;i<lda;i++){
  19                                 c = B[b+i];
  20                                 c1 = B[b1+i];
  21                                 C[i+a]+=A[a+k]*c;
  22                                 C[i+a1]+=A[a1+k]*c;
  23                                 C[i+a2]+=A[a2+k]*c;
  24                                 C[i+a3]+=A[a3+k]*c;
  25                                 C[i+a]+=A[a+k+1]*c1;
  26                                 C[i+a1]+=A[a1+k+1]*c1;
  27                                 C[i+a2]+=A[a2+k+1]*c1;
  28                                 C[i+a3]+=A[a3+k+1]*c1;
  29 }
  30 }
  31 }
  32 }