2 * Classical Block Matrix Multiplication example
6 #include "Matrix_init.h"
9 XBT_LOG_NEW_DEFAULT_CATEGORY(MM_Summa,
10 "Messages specific for this msg example");
13 double *a, double *b, double *c,
14 size_t lda, size_t ldb, size_t ldc,
15 size_t m, size_t k_a, size_t k_b, size_t n,
16 size_t Block_size, size_t start, size_t end,
17 size_t row, size_t col, size_t size_row, size_t size_col,
18 double *a_local, double *b_local,
19 MPI_Datatype Block_a, MPI_Datatype Block_a_local,
21 MPI_Comm row_comm, MPI_Comm col_comm, int subs)
23 double *B_a , *B_b ; //matrix blocks
25 //double alpha = 1, beta = 1; //C := alpha * a * b + beta * c
26 size_t B_proc_col, B_proc_row; // Number of bloc(row or col) on one processor
27 B_proc_col = k_b / Block_size; // Number of block on one processor
28 B_proc_row = k_a / Block_size; // Number of block on one processor
30 //size_t lda = k_a, ldb = n, ldc = n;
31 size_t lda_local = lda;
32 size_t ldb_local = ldb;
35 double time, computation_time = 0, communication_time = 0;
36 double start_time, end_time; //time mesure
37 double start_time_intern, end_time_intern; //time mesure
42 start_time = MPI_Wtime();
44 /*-------------Distributed Matrix Multiplication algorithm-----------------*/
46 for( iter = start; iter < end; iter++ ){
47 size_t pivot_row, pivot_col, pos_a, pos_b;
49 // pivot row on processor layer
50 pivot_row = (iter % size_col);
51 pivot_col = (iter % size_row);
52 //position of the block
54 pos_a = (size_t)((iter - start) / size_row) * Block_size;
55 pos_b = (size_t)((iter - start) / size_col) * ldb * Block_size;
57 pos_a = (size_t)(iter / size_row) * Block_size;
58 pos_b = (size_t)(iter / size_col) * ldb * Block_size;
61 // pivot row on processor layer
62 pivot_row = (size_t)(iter / B_proc_col) % size_col;
63 pivot_col = (size_t)(iter / B_proc_row) % size_row;
64 //position of the block
66 pos_a = ((iter - start) % B_proc_row) * Block_size;
67 pos_b = ((iter - start) % B_proc_col) * ldb * Block_size;
69 pos_a = (iter % B_proc_row) * Block_size;
70 pos_b = (iter % B_proc_col) * ldb * Block_size;
73 XBT_DEBUG( "pivot: %zu, iter: %zu, B_proc_col: %zu, "
74 "size_col:%zu, size_row: %zu\n",
75 pivot_row, iter, B_proc_row,size_col,size_row);
76 MPI_Barrier(row_comm);
77 MPI_Barrier(col_comm);
79 start_time_intern = MPI_Wtime();
83 if( pivot_col != col ){
85 lda_local = Block_size;
86 XBT_DEBUG("recieve B_a %zu,%zu \n",m , Block_size);
87 Block = &Block_a_local;
91 XBT_DEBUG("sent B_a %zu,%zu \n",m , Block_size);
94 err = MPI_Bcast(B_a, 1, *Block, pivot_col, row_comm);
95 if (err != MPI_SUCCESS) {
96 perror("Error Bcast A\n");
101 XBT_DEBUG("position of B_a: %zu \n", pos_a);
106 if( pivot_row == row ){
108 XBT_DEBUG("sent B_b Block_size: %zu, pos:%zu \n",
112 XBT_DEBUG("recieve B_b %zu,%zu \n", Block_size,n);
114 err = MPI_Bcast(B_b, 1, Block_b, pivot_row, col_comm );
115 if (err != MPI_SUCCESS) {
116 perror("Error Bcast B\n");
122 XBT_DEBUG("position of B_b: %zu \n", pos_b);
124 end_time_intern = MPI_Wtime();
125 communication_time += start_time_intern - end_time_intern;
127 MPI_Barrier(row_comm);
128 MPI_Barrier(col_comm);
129 start_time_intern = MPI_Wtime();
130 XBT_DEBUG("execute Gemm number: %zu\n", iter);
131 //We have recieved a line of block and a colomn
132 // cblas_dgemm( CblasRowMajor, CblasNoTrans, CblasNoTrans,
133 // m, n, Block_size, alpha, B_a, lda_local, B_b, ldb_local,
136 for(i = 0; i < m; i++)
137 for(j = 0; j < n; j++)
138 for(k = 0; k < Block_size; k++)
139 c[i*ldc+j] += B_a[i*lda_local+k]*B_b[k*ldb_local+j];
141 end_time_intern = MPI_Wtime();
142 computation_time += start_time_intern - end_time_intern;
145 MPI_Barrier(row_comm);
146 MPI_Barrier(col_comm);
148 end_time = MPI_Wtime();
149 time = start_time - end_time;
150 printf("communication time: %le nanoseconds, "
151 "computation time: %le nanoseconds\n",
152 communication_time, computation_time);