mt/ci_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 //----------MSI--------------
  16 /*
  17    int i,j,k;
  18    barrier(nc);
  19    for(j = coreid*lda/ncores; j < coreid*lda/ncores + lda/ncores; j++) {
  20         for(i = 0; i < lda; i+=4) {
  21                 data_t Cval0 = 0;
  22                 data_t Cval1 = 0;
  23                 data_t Cval2 = 0;
  24                 data_t Cval3 = 0;
  25                 for(k = 0; k < lda; k++) {
  26                         Cval0 += A[j*lda+k]*B[k*lda+i];
  27                         Cval1 += A[j*lda+k]*B[k*lda+i+1];
  28                         Cval2 += A[j*lda+k]*B[k*lda+i+2];
  29                         Cval3 += A[j*lda+k]*B[k*lda+i+3];
  30                 }
  31                 C[j*lda+i] = Cval0;
  32                 C[j*lda+i+1] = Cval1;
  33                 C[j*lda+i+2] = Cval2;
  34                 C[j*lda+i+3] = Cval3;
  35         }
  36    }
  37 */
  38
  39 //------------------MI-------------------
  40
  41    int i,j,k;
  42    barrier(ncores);
  43    for(j = coreid*lda/ncores; j < coreid*lda/ncores + lda/ncores; j++) {
  44         for(i = 0; i < lda; i+=4) {
  45                 data_t Cval0 = 0;
  46                 data_t Cval1 = 0;
  47                 data_t Cval2 = 0;
  48                 data_t Cval3 = 0;
  49                 if(coreid == 0) {
  50                         for(k = 0; k < lda; k++) {
  51                                 Cval0 += A[j*lda+k]*B[k*lda+i];
  52                                 Cval1 += A[j*lda+k]*B[k*lda+i+1];
  53                                 Cval2 += A[j*lda+k]*B[k*lda+i+2];
  54                                 Cval3 += A[j*lda+k]*B[k*lda+i+3];
  55                         }
  56                 } else {
  57                         for(k = lda-1; k >= 0; k--) {
  58                                 Cval0 += A[j*lda+k]*B[k*lda+i];
  59                                 Cval1 += A[j*lda+k]*B[k*lda+i+1];
  60                                 Cval2 += A[j*lda+k]*B[k*lda+i+2];
  61                                 Cval3 += A[j*lda+k]*B[k*lda+i+3];
  62                         }
  63                 }
  64                 C[j*lda+i] = Cval0;
  65                 C[j*lda+i+1] = Cval1;
  66                 C[j*lda+i+2] = Cval2;
  67                 C[j*lda+i+3] = Cval3;
  68         }
  69    }
  70 }