mt/bo_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     int i, j, k;
   9     data_t B_trans[32*32];
  10     data_t acc_temp0, acc_temp1;
  11     data_t *A_j, *B_i;
  12     data_t *A_j_k, *B_i_k;
  13     int z;
  14
  15     //for (i = 0; i < 32; i++) {
  16     //    for (j = 0; j < 32; j++) {
  17     //        B_trans[i*lda+j] = B[i+j*lda];
  18     //    }
  19     //}
  20
  21     if (coreid == 0) {
  22         for (i = 0; i < 32; i++) {
  23             B_i = B_trans+i*32;
  24             for (z = 0; z < 32; z++) {
  25                 *(B_i+z) = B[i+z*32];
  26             }
  27             for (j = 0; j < 16; j+=2) {
  28                 A_j = A+j*lda;
  29                 acc_temp0 = 0;
  30                 for (k = 0; k < 32; k+=8) {
  31                     A_j_k = A_j+k;
  32                     B_i_k = B_i+k;
  33                     acc_temp0 += *(A_j_k)     * *(B_i_k);
  34                     acc_temp0 += *(A_j_k + 1) * *(B_i_k + 1);
  35                     acc_temp0 += *(A_j_k + 2) * *(B_i_k + 2);
  36                     acc_temp0 += *(A_j_k + 3) * *(B_i_k + 3);
  37                     acc_temp0 += *(A_j_k + 4) * *(B_i_k + 4);
  38                     acc_temp0 += *(A_j_k + 5) * *(B_i_k + 5);
  39                     acc_temp0 += *(A_j_k + 6) * *(B_i_k + 6);
  40                     acc_temp0 += *(A_j_k + 7) * *(B_i_k + 7);
  41                 }
  42                 A_j += 32;
  43
  44                 acc_temp1 = 0;
  45                 for (k = 0; k < 32; k+=8) {
  46                     acc_temp1 += *(A_j+k) * *(B_i+k);
  47                     acc_temp1 += *(A_j+k + 1) * *(B_i+k + 1);
  48                     acc_temp1 += *(A_j+k + 2) * *(B_i+k + 2);
  49                     acc_temp1 += *(A_j+k + 3) * *(B_i+k + 3);
  50                     acc_temp1 += *(A_j+k + 4) * *(B_i+k + 4);
  51                     acc_temp1 += *(A_j+k + 5) * *(B_i+k + 5);
  52                     acc_temp1 += *(A_j+k + 6) * *(B_i+k + 6);
  53                     acc_temp1 += *(A_j+k + 7) * *(B_i+k + 7);
  54                 }
  55
  56                 C[i + j*lda] = acc_temp0;
  57                 C[i + (j+1)*lda] = acc_temp1;
  58             }
  59         }
  60     }
  61     if (coreid == 1 || ncores == 1) {
  62         for (i = 0; i < 32; i++) {
  63             B_i = B_trans+i*32;
  64             for (z = 0; z < 32; z++) {
  65                 *(B_i+z) = B[i+z*32];
  66             }
  67             for (j = 16; j < 32; j+=2) {
  68                 A_j = A+j*lda;
  69                 acc_temp0 = 0;
  70                 for (k = 0; k < 32; k+=8) {
  71                     acc_temp0 += *(A_j+k) * *(B_i+k);
  72                     acc_temp0 += *(A_j+k + 1) * *(B_i+k + 1);
  73                     acc_temp0 += *(A_j+k + 2) * *(B_i+k + 2);
  74                     acc_temp0 += *(A_j+k + 3) * *(B_i+k + 3);
  75                     acc_temp0 += *(A_j+k + 4) * *(B_i+k + 4);
  76                     acc_temp0 += *(A_j+k + 5) * *(B_i+k + 5);
  77                     acc_temp0 += *(A_j+k + 6) * *(B_i+k + 6);
  78                     acc_temp0 += *(A_j+k + 7) * *(B_i+k + 7);
  79                 }
  80                 A_j += 32;
  81
  82                 acc_temp1 = 0;
  83                 for (k = 0; k < 32; k+=8) {
  84                     acc_temp1 += *(A_j+k) * *(B_i+k);
  85                     acc_temp1 += *(A_j+k + 1) * *(B_i+k + 1);
  86                     acc_temp1 += *(A_j+k + 2) * *(B_i+k + 2);
  87                     acc_temp1 += *(A_j+k + 3) * *(B_i+k + 3);
  88                     acc_temp1 += *(A_j+k + 4) * *(B_i+k + 4);
  89                     acc_temp1 += *(A_j+k + 5) * *(B_i+k + 5);
  90                     acc_temp1 += *(A_j+k + 6) * *(B_i+k + 6);
  91                     acc_temp1 += *(A_j+k + 7) * *(B_i+k + 7);
  92                 }
  93                 C[i + j*lda] = acc_temp0;
  94                 C[i + (j+1)*lda] = acc_temp1;
  95             }
  96         }
  97     }
  98 }