mt/du_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   size_t max_dim = 32*32;
  11   data_t element, element2, element3, element4, element5, element6, element7, element8;
  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   //for (i=coreid*max_dim/ncores; i<(max_dim/ncores+coreid*max_dim/ncores); i+=8){
  15   for (l=coreid*32/ncores; l<32*(1+coreid)/ncores; l+=2){
  16     row=l*32;
  17     row2=(l+1)*32;
  18     for (i=0; i<lda; i+=4){
  19       element = A[row+i];
  20       element2 = A[row+i+1];
  21       element3 = A[row+i+2];
  22       element4 = A[row+i+3];
  23       element5 = A[row2+i];
  24       element6 = A[row2+i+1];
  25       element7 = A[row2+i+2];
  26       element8 = A[row2+i+3];
  27       column=i*32;
  28       column2=(i+1)*32;
  29       column3=(i+2)*32;
  30       column4=(i+3)*32;
  31       for (j=0; j<32; j+=4){
  32         temp_mat[j]+=element*B[column+j]+element2*B[column2+j]+element3*B[column3+j]+element4*B[column4+j];
  33         temp_mat[j+1]+=element*B[column+j+1]+element2*B[column2+j+1]+element3*B[column3+j+1]+element4*B[column4+j+1];
  34         temp_mat[j+2]+=element*B[column+j+2]+element2*B[column2+j+2]+element3*B[column3+j+2]+element4*B[column4+j+2];
  35         temp_mat[j+3]+=element*B[column+j+3]+element2*B[column2+j+3]+element3*B[column3+j+3]+element4*B[column4+j+3];
  36         temp_mat2[j]+=element5*B[column+j]+element6*B[column2+j]+element7*B[column3+j]+element8*B[column4+j];
  37         temp_mat2[j+1]+=element5*B[column+j+1]+element6*B[column2+j+1]+element7*B[column3+j+1]+element8*B[column4+j+1];
  38         temp_mat2[j+2]+=element5*B[column+j+2]+element6*B[column2+j+2]+element7*B[column3+j+2]+element8*B[column4+j+2];
  39         temp_mat2[j+3]+=element5*B[column+j+3]+element6*B[column2+j+3]+element7*B[column3+j+3]+element8*B[column4+j+3];
  40       }
  41       /*if (i==28){
  42         for(k=0; k<32; k++){
  43           C[row+k]=temp_mat[k];
  44           C[row2+k]=temp_mat2[k];
  45           temp_mat[k]=0;
  46           temp_mat2[k]=0;
  47         }
  48         }*/
  49     }
  50     for(k=0; k<32; k++){
  51           C[row+k]=temp_mat[k];
  52           C[row2+k]=temp_mat2[k];
  53           temp_mat[k]=0;
  54           temp_mat2[k]=0;
  55     }
  56   }
  57
  58    // ***************************** //
  59    // **** ADD YOUR CODE HERE ***** //
  60    // ***************************** //
  61    //
  62    // feel free to make a separate function for MI and MSI versions.
  63
  64 }