#include "stdlib.h" #include "util.h" #include "dataset.h" void __attribute__((noinline)) matmul(const int coreid, const int ncores, const int lda, const data_t A[], const data_t B[], data_t C[] ) { size_t i, j, k, l; int row,row2, column, column2, column3, column4, column5, column6, column7, column8; data_t element, element2, element3, element4, element5, element6, element7, element8; data_t B1, B2, B3, B4; data_t temp_mat[32]={0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0}; data_t temp_mat2[32]={0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0}; int local_lda = lda; for (l=coreid*local_lda/ncores; l