benchmarks/mt-matmul/matmul.c

   1 // See LICENSE for license details.
   2
   3 #include "dataset.h"
   4 #include "util.h"
   5 #include <stddef.h>
   6
   7 #pragma GCC optimize ("unroll-loops")
   8
   9 void matmul(const size_t coreid, const size_t ncores, const size_t lda,  const data_t A[], const data_t B[], data_t C[])
  10 {
  11   size_t i, j, k;
  12   size_t block = lda / ncores;
  13   size_t start = block * coreid;
  14
  15   for (i = 0; i < lda; i++) {
  16     for (j = start; j < (start+block); j++) {
  17       data_t sum = 0;
  18       for (k = 0; k < lda; k++)
  19         sum += A[j*lda + k] * B[k*lda + i];
  20       C[i + j*lda] = sum;
  21     }
  22   }
  23 }