mt/ar_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
   9    // ***************************** //
  10    // **** ADD YOUR CODE HERE ***** //
  11    // ***************************** //
  12    //
  13    // feel free to make a separate function for MI and MSI versions.
  14
  15    int i, j, k, B_t[32*32], x, y;
  16    int ALoc, BLoc, CLoc;
  17 //   int ii = 0, done = 0;
  18    //for(x = coreid*(lda/ncores); x < (coreid+1)*(lda/ncores) && x < lda; x++) {
  19    for (x = 0; x < lda; x++) {
  20                 for(y = 0; y < lda; y++) {
  21                         B_t[y*lda + x] = B[x*lda + y];
  22                 }
  23    }
  24   // for ( ii = lda/4 ; ii < lda ; ii += lda/4)
  25    //{
  26 //         for ( i = coreid*(ii/ncores); i < (coreid+1)*(ii/ncores) && i < ii; i++ )
  27            for ( i = coreid*(lda/ncores); i < (coreid+1)*(lda/ncores) && i < lda; i++ )
  28            {
  29                   ALoc = i*lda;
  30           for ( j = 0; j < lda; j++ )
  31               {
  32                  BLoc = j*lda;
  33                  CLoc = i*lda + j;
  34              for ( k = 0; k < lda; k++ )
  35                  {
  36                 C[CLoc] += A[ALoc + k] * B_t[BLoc + k];
  37                  }
  38           }
  39            }
  40    //}
  41 }