2 * Classical Block Matrix Multiplication example
4 * Authors: Quintin Jean-Noël
6 #include "Matrix_init.h"
10 XBT_LOG_NEW_DEFAULT_CATEGORY(MM_Summa,
11 "Messages specific for this msg example");
14 double *a, double *b, double *c,
15 size_t lda, size_t ldb, size_t ldc,
16 size_t m, size_t k_a, size_t k_b, size_t n,
17 size_t Block_size, size_t start, size_t end,
18 size_t row, size_t col, size_t size_row, size_t size_col,
19 double *a_local, double *b_local,
20 MPI_Datatype Block_a, MPI_Datatype Block_a_local,
22 MPI_Comm row_comm, MPI_Comm col_comm, int subs)
24 double *B_a , *B_b ; //matrix blocks
26 //double alpha = 1, beta = 1; //C := alpha * a * b + beta * c
27 size_t B_proc_col, B_proc_row; // Number of bloc(row or col) on one processor
28 B_proc_col = k_b / Block_size; // Number of block on one processor
29 B_proc_row = k_a / Block_size; // Number of block on one processor
31 //size_t lda = k_a, ldb = n, ldc = n;
32 size_t lda_local = lda;
33 size_t ldb_local = ldb;
36 double time, computation_time = 0, communication_time = 0;
37 struct timespec start_time, end_time; //time mesure
38 struct timespec start_time_intern, end_time_intern; //time mesure
43 get_time(&start_time);
45 /*-------------Distributed Matrix Multiplication algorithm-----------------*/
47 for( iter = start; iter < end; iter++ ){
48 size_t pivot_row, pivot_col, pos_a, pos_b;
50 // pivot row on processor layer
51 pivot_row = (iter % size_col);
52 pivot_col = (iter % size_row);
53 //position of the block
55 pos_a = (size_t)((iter - start) / size_row) * Block_size;
56 pos_b = (size_t)((iter - start) / size_col) * ldb * Block_size;
58 pos_a = (size_t)(iter / size_row) * Block_size;
59 pos_b = (size_t)(iter / size_col) * ldb * Block_size;
62 // pivot row on processor layer
63 pivot_row = (size_t)(iter / B_proc_col) % size_col;
64 pivot_col = (size_t)(iter / B_proc_row) % size_row;
65 //position of the block
67 pos_a = ((iter - start) % B_proc_row) * Block_size;
68 pos_b = ((iter - start) % B_proc_col) * ldb * Block_size;
70 pos_a = (iter % B_proc_row) * Block_size;
71 pos_b = (iter % B_proc_col) * ldb * Block_size;
74 XBT_DEBUG( "pivot: %zu, iter: %zu, B_proc_col: %zu, "
75 "size_col:%zu, size_row: %zu\n",
76 pivot_row, iter, B_proc_row,size_col,size_row);
77 MPI_Barrier(row_comm);
78 MPI_Barrier(col_comm);
80 get_time(&start_time_intern);
84 if( pivot_col != col ){
86 lda_local = Block_size;
87 XBT_DEBUG("recieve B_a %zu,%zu \n",m , Block_size);
88 Block = &Block_a_local;
92 XBT_DEBUG("sent B_a %zu,%zu \n",m , Block_size);
95 err = MPI_Bcast(B_a, 1, *Block, pivot_col, row_comm);
96 if (err != MPI_SUCCESS) {
97 perror("Error Bcast A\n");
102 XBT_DEBUG("position of B_a: %zu \n", pos_a);
107 if( pivot_row == row ){
109 XBT_DEBUG("sent B_b Block_size: %zu, pos:%zu \n",
113 XBT_DEBUG("recieve B_b %zu,%zu \n", Block_size,n);
115 err = MPI_Bcast(B_b, 1, Block_b, pivot_row, col_comm );
116 if (err != MPI_SUCCESS) {
117 perror("Error Bcast B\n");
123 XBT_DEBUG("position of B_b: %zu \n", pos_b);
125 get_time(&end_time_intern);
126 communication_time += get_timediff(&start_time_intern,&end_time_intern);
128 MPI_Barrier(row_comm);
129 MPI_Barrier(col_comm);
130 get_time(&start_time_intern);
131 XBT_DEBUG("execute Gemm number: %zu\n", iter);
132 //We have recieved a line of block and a colomn
133 // cblas_dgemm( CblasRowMajor, CblasNoTrans, CblasNoTrans,
134 // m, n, Block_size, alpha, B_a, lda_local, B_b, ldb_local,
137 for(i = 0; i < m; i++)
138 for(j = 0; j < n; j++)
139 for(k = 0; k < Block_size; k++)
140 c[i*ldc+j] += B_a[i*lda_local+k]*B_b[k*ldb_local+j];
142 get_time(&end_time_intern);
143 computation_time += get_timediff(&start_time_intern,&end_time_intern);
146 MPI_Barrier(row_comm);
147 MPI_Barrier(col_comm);
150 time = get_timediff(&start_time,&end_time);
151 printf("communication time: %le nanoseconds, "
152 "computation time: %le nanoseconds\n",
153 communication_time, computation_time);