mt/af_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     size_t i, j, k, l;
   9   int row,row2, column, column2, column3, column4, column5, column6, column7, column8;
  10   data_t element, element2, element3, element4, element5, element6, element7, element8;
  11         data_t B1, B2, B3, B4;
  12   data_t temp_mat[32]={0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
  13   data_t temp_mat2[32]={0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
  14         int local_lda = lda;
  15
  16   for (l=coreid*local_lda/ncores; l<local_lda*(1+coreid)/ncores; l+=2){
  17     row=l*32;
  18     row2=(l+1)*32;
  19                 //element = A[row];
  20                 //element5 = A[row2];
  21     for (i=0; i<local_lda; i+=4){
  22       element = A[row+i];
  23       element2 = A[row+i+1];
  24       element3 = A[row+i+2];
  25       element4 = A[row+i+3];
  26
  27       element5 = A[row2+i];
  28       element6 = A[row2+i+1];
  29       element7 = A[row2+i+2];
  30       element8 = A[row2+i+3];
  31
  32       column=i*local_lda;
  33       column2=(i+1)*local_lda;
  34       column3=(i+2)*local_lda;
  35       column4=(i+3)*local_lda;
  36
  37                         B1 = B[column];
  38                         B2 = B[column2];
  39                         B3 = B[column3];
  40                         B4 = B[column4];
  41
  42       for (j=0; j<lda; j+=4){
  43                                 temp_mat[j]+=element*B1+element2*B2+element3*B3+element4*B4;
  44                                 temp_mat[j+1]+=element*B[column+j+1]+element2*B[column2+j+1]+element3*B[column3+j+1]+element4*B[column4+j+1];
  45                                 temp_mat[j+2]+=element*B[column+j+2]+element2*B[column2+j+2]+element3*B[column3+j+2]+element4*B[column4+j+2];
  46                                 temp_mat[j+3]+=element*B[column+j+3]+element2*B[column2+j+3]+element3*B[column3+j+3]+element4*B[column4+j+3];
  47
  48                                 temp_mat2[j]+=element5*B1+element6*B2+element7*B3+element8*B4;
  49                                 temp_mat2[j+1]+=element5*B[column+j+1]+element6*B[column2+j+1]+element7*B[column3+j+1]+element8*B[column4+j+1];
  50                                 temp_mat2[j+2]+=element5*B[column+j+2]+element6*B[column2+j+2]+element7*B[column3+j+2]+element8*B[column4+j+2];
  51                                 temp_mat2[j+3]+=element5*B[column+j+3]+element6*B[column2+j+3]+element7*B[column3+j+3]+element8*B[column4+j+3];
  52
  53                                 B1 = B[column+j+4];
  54                                 B2 = B[column2+j+4];
  55                                 B3 = B[column3+j+4];
  56                                 B4 = B[column4+j+4];
  57
  58                                 }
  59                         //element = A[row+i+4];
  60                         //element5 = A[row2+i+4];
  61       }
  62
  63                         for(k=0; k<local_lda; k++){
  64                                 C[row+k]=temp_mat[k];
  65                                 temp_mat[k]=0;
  66                                 C[row2+k]=temp_mat2[k];
  67                                 temp_mat2[k]=0;
  68
  69                                 }
  70
  71
  72   }
  73    // ***************************** //
  74    // **** ADD YOUR CODE HERE ***** //
  75    // ***************************** //
  76    //
  77    // feel free to make a separate function for MI and MSI versions.
  78
  79 }