mt/ai_matmul.c

   1 #include "stdlib.h"
   2
   3 #include "util.h"
   4
   5 #include "dataset.h"
   6 #include "util.h"
   7 void __attribute__((noinline)) matmul(const int coreid, const int ncores, const int lda,  const data_t A[], const data_t B[], data_t C[] )
   8 {
   9
  10    // ***************************** //
  11    // **** ADD YOUR CODE HERE ***** //
  12    // ***************************** //
  13    //
  14    // feel free to make a separate function for MI and MSI versions.
  15
  16 //----------MSI--------------
  17 ///*
  18    int i,j,k;
  19    barrier(ncores);
  20    for(j = coreid*lda/ncores; j < coreid*lda/ncores + lda/ncores; j++) {
  21         for(i = 0; i < lda; i+=4) {
  22                 data_t Cval0 = 0;
  23                 data_t Cval1 = 0;
  24                 data_t Cval2 = 0;
  25                 data_t Cval3 = 0;
  26                 for(k = 0; k < lda; k++) {
  27                         Cval0 += A[j*lda+k]*B[k*lda+i];
  28                         Cval1 += A[j*lda+k]*B[k*lda+i+1];
  29                         Cval2 += A[j*lda+k]*B[k*lda+i+2];
  30                         Cval3 += A[j*lda+k]*B[k*lda+i+3];
  31                 }
  32                 C[j*lda+i] = Cval0;
  33                 C[j*lda+i+1] = Cval1;
  34                 C[j*lda+i+2] = Cval2;
  35                 C[j*lda+i+3] = Cval3;
  36         }
  37    }
  38 //*/
  39
  40 //------------------MI-------------------
  41 /*
  42    int i,j,k;
  43    barrier(nc);
  44    for(j = coreid*lda/ncores; j < coreid*lda/ncores + lda/ncores; j++) {
  45         for(i = 0; i < lda; i+=4) {
  46                 data_t Cval0 = 0;
  47                 data_t Cval1 = 0;
  48                 data_t Cval2 = 0;
  49                 data_t Cval3 = 0;
  50                 if(coreid == 0) {
  51                         for(k = 0; k < lda; k++) {
  52                                 Cval0 += A[j*lda+k]*B[k*lda+i];
  53                                 Cval1 += A[j*lda+k]*B[k*lda+i+1];
  54                                 Cval2 += A[j*lda+k]*B[k*lda+i+2];
  55                                 Cval3 += A[j*lda+k]*B[k*lda+i+3];
  56                         }
  57                 } else {
  58                         for(k = lda-1; k >= 0; k--) {
  59                                 Cval0 += A[j*lda+k]*B[k*lda+i];
  60                                 Cval1 += A[j*lda+k]*B[k*lda+i+1];
  61                                 Cval2 += A[j*lda+k]*B[k*lda+i+2];
  62                                 Cval3 += A[j*lda+k]*B[k*lda+i+3];
  63                         }
  64                 }
  65                 C[j*lda+i] = Cval0;
  66                 C[j*lda+i+1] = Cval1;
  67                 C[j*lda+i+2] = Cval2;
  68                 C[j*lda+i+3] = Cval3;
  69         }
  70    }
  71 */
  72 }