1 /* Copyright (c) 2012-2014. The SimGrid Team.
2 * All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
8 * Classical Block Matrix Multiplication example
12 #include "Matrix_init.h"
17 XBT_LOG_NEW_DEFAULT_CATEGORY(MM_Summa,
18 "Messages specific for this msg example");
20 double Summa(double *a, double *b, double *c,
21 size_t lda, size_t ldb, size_t ldc,
22 size_t m, size_t k_a, size_t k_b, size_t n,
23 size_t Block_size, size_t start, size_t end,
24 size_t row, size_t col, size_t size_row, size_t size_col,
25 double *a_local, double *b_local,
26 MPI_Datatype Block_a, MPI_Datatype Block_a_local,
28 MPI_Comm row_comm, MPI_Comm col_comm, int subs)
30 double *B_a , *B_b ; //matrix blocks
32 //double alpha = 1, beta = 1; //C := alpha * a * b + beta * c
33 size_t B_proc_row; // Number of bloc(row or col) on one processor
36 B_proc_col = k_b / Block_size; // Number of block on one processor
38 B_proc_row = k_a / Block_size; // Number of block on one processor
40 //size_t lda = k_a, ldb = n, ldc = n;
41 size_t lda_local = lda;
42 size_t ldb_local = ldb;
45 double time, computation_time = 0, communication_time = 0;
46 double start_time, end_time; //time mesure
47 double start_time_intern, end_time_intern; //time mesure
52 start_time = MPI_Wtime();
54 /*-------------Distributed Matrix Multiplication algorithm-----------------*/
56 for( iter = start; iter < end; iter++ ){
57 size_t pivot_row, pivot_col, pos_a, pos_b;
59 // pivot row on processor layer
60 pivot_row = (iter % size_col);
61 pivot_col = (iter % size_row);
62 //position of the block
64 pos_a = (size_t)((iter - start) / size_row) * Block_size;
65 pos_b = (size_t)((iter - start) / size_col) * ldb * Block_size;
67 pos_a = (size_t)(iter / size_row) * Block_size;
68 pos_b = (size_t)(iter / size_col) * ldb * Block_size;
71 // pivot row on processor layer
72 pivot_row = (size_t)(iter / B_proc_col) % size_col;
73 pivot_col = (size_t)(iter / B_proc_row) % size_row;
74 //position of the block
76 pos_a = ((iter - start) % B_proc_row) * Block_size;
77 pos_b = ((iter - start) % B_proc_col) * ldb * Block_size;
79 pos_a = (iter % B_proc_row) * Block_size;
80 pos_b = (iter % B_proc_col) * ldb * Block_size;
83 XBT_DEBUG( "pivot: %zu, iter: %zu, B_proc_col: %zu, "
84 "size_col:%zu, size_row: %zu\n",
85 pivot_row, iter, B_proc_row,size_col,size_row);
86 /* MPI_Barrier(row_comm);*/
87 /* MPI_Barrier(col_comm);*/
89 start_time_intern = MPI_Wtime();
93 if( pivot_col != col ){
95 lda_local = Block_size;
96 XBT_DEBUG("recieve B_a %zu,%zu \n",m , Block_size);
97 Block = &Block_a_local;
101 XBT_DEBUG("sent B_a %zu,%zu \n",m , Block_size);
104 err = MPI_Bcast(B_a, 1, *Block, pivot_col, row_comm);
105 if (err != MPI_SUCCESS) {
106 perror("Error Bcast A\n");
111 XBT_DEBUG("position of B_a: %zu \n", pos_a);
116 if( pivot_row == row ){
118 XBT_DEBUG("sent B_b Block_size: %zu, pos:%zu \n",
122 XBT_DEBUG("recieve B_b %zu,%zu \n", Block_size,n);
124 err = MPI_Bcast(B_b, 1, Block_b, pivot_row, col_comm );
125 if (err != MPI_SUCCESS) {
126 perror("Error Bcast B\n");
132 XBT_DEBUG("position of B_b: %zu \n", pos_b);
134 end_time_intern = MPI_Wtime();
135 communication_time += end_time_intern - start_time_intern;
137 /* MPI_Barrier(row_comm);*/
138 /* MPI_Barrier(col_comm);*/
139 start_time_intern = MPI_Wtime();
140 XBT_DEBUG("execute Gemm number: %zu\n", iter);
141 //We have recieved a line of block and a colomn
142 // cblas_dgemm( CblasRowMajor, CblasNoTrans, CblasNoTrans,
143 // m, n, Block_size, alpha, B_a, lda_local, B_b, ldb_local,
146 for(i = 0; i < m; i++)
147 for(j = 0; j < n; j++)
148 for(k = 0; k < Block_size; k++)
149 c[i*ldc+j] += B_a[i*lda_local+k]*B_b[k*ldb_local+j];
151 end_time_intern = MPI_Wtime();
152 computation_time += end_time_intern - start_time_intern;
155 MPI_Barrier(row_comm);
156 MPI_Barrier(col_comm);
158 end_time = MPI_Wtime();
159 time = end_time - start_time ;
160 printf("communication time: %e seconds, "
161 "computation time: %e seconds\n",
162 communication_time, computation_time);