/*!
* Classical Block Matrix Multiplication example
*
- * Authors: Quintin Jean-Noël
*/
+
#include "Matrix_init.h"
#include "Summa.h"
-#include "timer.h"
#include "xbt/log.h"
+#include <stdio.h>
+
XBT_LOG_NEW_DEFAULT_CATEGORY(MM_Summa,
"Messages specific for this msg example");
-inline double Summa(
- double *a, double *b, double *c,
- size_t lda, size_t ldb, size_t ldc,
- size_t m, size_t k_a, size_t k_b, size_t n,
- size_t Block_size, size_t start, size_t end,
- size_t row, size_t col, size_t size_row, size_t size_col,
- double *a_local, double *b_local,
- MPI_Datatype Block_a, MPI_Datatype Block_a_local,
- MPI_Datatype Block_b,
- MPI_Comm row_comm, MPI_Comm col_comm, int subs)
+double Summa(double *a, double *b, double *c,
+ size_t lda, size_t ldb, size_t ldc,
+ size_t m, size_t k_a, size_t k_b, size_t n,
+ size_t Block_size, size_t start, size_t end,
+ size_t row, size_t col, size_t size_row, size_t size_col,
+ double *a_local, double *b_local,
+ MPI_Datatype Block_a, MPI_Datatype Block_a_local,
+ MPI_Datatype Block_b,
+ MPI_Comm row_comm, MPI_Comm col_comm, int subs)
{
double *B_a , *B_b ; //matrix blocks
size_t err;
- double alpha = 1, beta = 1; //C := alpha * a * b + beta * c
- size_t B_proc_col, B_proc_row; // Number of bloc(row or col) on one processor
+ //double alpha = 1, beta = 1; //C := alpha * a * b + beta * c
+ size_t B_proc_row; // Number of bloc(row or col) on one processor
+#ifndef CYCLIC
+ size_t B_proc_col;
B_proc_col = k_b / Block_size; // Number of block on one processor
+#endif
B_proc_row = k_a / Block_size; // Number of block on one processor
//size_t lda = k_a, ldb = n, ldc = n;
double time, computation_time = 0, communication_time = 0;
- struct timespec start_time, end_time; //time mesure
- struct timespec start_time_intern, end_time_intern; //time mesure
+ double start_time, end_time; //time mesure
+ double start_time_intern, end_time_intern; //time mesure
- get_time(&start_time);
+ start_time = MPI_Wtime();
/*-------------Distributed Matrix Multiplication algorithm-----------------*/
size_t iter;
XBT_DEBUG( "pivot: %zu, iter: %zu, B_proc_col: %zu, "
"size_col:%zu, size_row: %zu\n",
pivot_row, iter, B_proc_row,size_col,size_row);
- MPI_Barrier(row_comm);
- MPI_Barrier(col_comm);
+/* MPI_Barrier(row_comm);*/
+/* MPI_Barrier(col_comm);*/
- get_time(&start_time_intern);
+ start_time_intern = MPI_Wtime();
//Broadcast the row
if(size_row > 1){
MPI_Datatype * Block;
B_b = b + pos_b;
XBT_DEBUG("position of B_b: %zu \n", pos_b);
}
- get_time(&end_time_intern);
- communication_time += get_timediff(&start_time_intern,&end_time_intern);
+ end_time_intern = MPI_Wtime();
+ communication_time += end_time_intern - start_time_intern;
- MPI_Barrier(row_comm);
- MPI_Barrier(col_comm);
- get_time(&start_time_intern);
+/* MPI_Barrier(row_comm);*/
+/* MPI_Barrier(col_comm);*/
+ start_time_intern = MPI_Wtime();
XBT_DEBUG("execute Gemm number: %zu\n", iter);
//We have recieved a line of block and a colomn
// cblas_dgemm( CblasRowMajor, CblasNoTrans, CblasNoTrans,
for(i = 0; i < m; i++)
for(j = 0; j < n; j++)
for(k = 0; k < Block_size; k++)
- c[i*ldc+j] += B_a[j*lda_local+k]*B_b[k*ldb_local+j];
+ c[i*ldc+j] += B_a[i*lda_local+k]*B_b[k*ldb_local+j];
- get_time(&end_time_intern);
- computation_time += get_timediff(&start_time_intern,&end_time_intern);
+ end_time_intern = MPI_Wtime();
+ computation_time += end_time_intern - start_time_intern;
}
MPI_Barrier(row_comm);
MPI_Barrier(col_comm);
- get_time(&end_time);
- time = get_timediff(&start_time,&end_time);
- printf("communication time: %le nanoseconds, "
- "computation time: %le nanoseconds\n",
+ end_time = MPI_Wtime();
+ time = end_time - start_time ;
+ printf("communication time: %e seconds, "
+ "computation time: %e seconds\n",
communication_time, computation_time);
