int main(int argc, char *argv[]) { int i, j; int ch; extern char *optarg; int edge; int size; /* ARMCI */ void **ptr; double **ptr_loc; MP_INIT(argc,argv); MP_PROCS(&nproc); MP_MYID(&me); while ((ch = getopt(argc, argv, "n:b:p:h")) != -1) { switch(ch) { case 'n': n = atoi(optarg); break; case 'b': block_size = atoi(optarg); break; case 'p': nproc = atoi(optarg); break; case 'h': { printf("Usage: LU, or \n"); printf(" LU -nMATRIXSIZE -bBLOCKSIZE -pNPROC\n"); MP_BARRIER(); MP_FINALIZE(); exit(0); } } } if(me == 0) { printf("\n Blocked Dense LU Factorization\n"); printf(" %d by %d Matrix\n", n, n); printf(" %d Processors\n", nproc); printf(" %d by %d Element Blocks\n", block_size, block_size); printf("\n"); } /* num_rows = (int) sqrt((double) nproc); */ /* for (;;) { */ /* num_cols = nproc/num_rows; */ /* if (num_rows*num_cols == nproc) */ /* break; */ /* num_rows--; */ /* } */ nblocks = n/block_size; if (block_size * nblocks != n) { nblocks++; } nnodes = nproc / 4; if((nnodes * 4) != nproc) { num_cols = nproc - nnodes * 4; nnodes++; num_rows = 1; } else { num_cols = 2; num_rows = 2; } num = (nblocks * nblocks)/nnodes; if((num * nnodes) != (nblocks * nblocks)) num++; #ifdef DEBUG if(me == 0) for (i=0;i<nblocks;i++) { for (j=0;j<nblocks;j++) printf("%d ", block_owner(i, j)); printf("\n"); } MP_BARRIER(); MP_FINALIZE(); exit(0); #endif edge = n%block_size; if (edge == 0) { edge = block_size; } for (i=0;i<nblocks;i++) { for (j=0;j<nblocks;j++) { if(block_owner(i,j) == me) { if ((i == nblocks-1) && (j == nblocks-1)) { size = edge*edge; } else if ((i == nblocks-1) || (j == nblocks-1)) { size = edge*block_size; } else { size = block_size*block_size; } proc_bytes += size*sizeof(double); } } } /* initialize ARMCI */ ARMCI_Init(); ptr = (void **)malloc(nproc * sizeof(void *)); ARMCI_Malloc(ptr, proc_bytes); a = (double **)malloc(nblocks*nblocks*sizeof(double *)); if (a == NULL) { fprintf(stderr, "Could not malloc memory for a\n"); exit(-1); } ptr_loc = (double **)malloc(nproc*sizeof(double *)); for(i=0; i<nproc; i++) ptr_loc[i] = (double *)ptr[i]; for(i=0; i<nblocks;i ++) { for(j=0; j<nblocks; j++) { a[i+j*nblocks] = ptr_loc[block_owner(i, j)]; if ((i == nblocks-1) && (j == nblocks-1)) { size = edge*edge; } else if ((i == nblocks-1) || (j == nblocks-1)) { size = edge*block_size; } else { size = block_size*block_size; } ptr_loc[block_owner(i, j)] += size; } } /* initialize the array */ init_array(); /* barrier to ensure all initialization is done */ MP_BARRIER(); /* to remove cold-start misses, all processors touch their own data */ touch_array(block_size, me); MP_BARRIER(); if(doprint) { if(me == 0) { printf("Matrix before LU decomposition\n"); print_array(me); } MP_BARRIER(); } /* Starting the timer */ if(me == 0) start_timer(); lu(n, block_size, me); MP_BARRIER(); /* Timer Stops here */ if(me == 0) printf("\nRunning time = %lf milliseconds.\n\n", elapsed_time()); if(doprint) { if(me == 0) { printf("after LU\n"); print_array(me); } MP_BARRIER(); } /* done */ ARMCI_Free(ptr[me]); ARMCI_Finalize(); MP_FINALIZE(); return 0; }
main(int argc, char *argv[]) { int i, j; int ch; extern char *optarg; int edge; int size; int nloop=5; double **ptr_loc; MP_INIT(arc,argv); MP_PROCS(&nproc); MP_MYID(&me); while ((ch = getopt(argc, argv, "n:b:p:h")) != -1) { switch(ch) { case 'n': n = atoi(optarg); break; case 'b': block_size = atoi(optarg); break; case 'p': nproc = atoi(optarg); break; case 'h': { printf("Usage: LU, or \n"); printf(" LU -nMATRIXSIZE -bBLOCKSIZE -pNPROC\n"); MP_BARRIER(); MP_FINALIZE(); exit(0); } } } if(me == 0) { printf("\n Blocked Dense LU Factorization\n"); printf(" %d by %d Matrix\n", n, n); printf(" %d Processors\n", nproc); printf(" %d by %d Element Blocks\n", block_size, block_size); printf("\n"); } num_rows = (int) sqrt((double) nproc); for (;;) { num_cols = nproc/num_rows; if (num_rows*num_cols == nproc) break; num_rows--; } nblocks = n/block_size; if (block_size * nblocks != n) { nblocks++; } edge = n%block_size; if (edge == 0) { edge = block_size; } #ifdef DEBUG if(me == 0) for (i=0; i<nblocks; i++) { for (j=0; j<nblocks; j++) printf("%d ", block_owner(i, j)); printf("\n"); } MP_BARRIER(); MP_FINALIZE(); exit(0); #endif for (i=0; i<nblocks; i++) { for (j=0; j<nblocks; j++) { if(block_owner(i,j) == me) { if ((i == nblocks-1) && (j == nblocks-1)) { size = edge*edge; } else if ((i == nblocks-1) || (j == nblocks-1)) { size = edge*block_size; } else { size = block_size*block_size; } proc_bytes += size*sizeof(double); } } } ptr = (void **)malloc(nproc * sizeof(void *)); #ifdef MPI2_ONESIDED MPI_Alloc_mem(proc_bytes, MPI_INFO_NULL, &ptr[me]); MPI_Win_create((void*)ptr[me], proc_bytes, 1, MPI_INFO_NULL, MPI_COMM_WORLD, &win); for(i=0; i<nproc; i++) ptr[i] = (double *)ptr[me]; MPI_Barrier(MPI_COMM_WORLD); #else /* initialize ARMCI */ ARMCI_Init(); ARMCI_Malloc(ptr, proc_bytes); #endif a = (double **)malloc(nblocks*nblocks*sizeof(double *)); if (a == NULL) { fprintf(stderr, "Could not malloc memory for a\n"); exit(-1); } ptr_loc = (double **)malloc(nproc*sizeof(double *)); for(i=0; i<nproc; i++) ptr_loc[i] = (double *)ptr[i]; for(i=0; i<nblocks; i ++) { for(j=0; j<nblocks; j++) { a[i+j*nblocks] = ptr_loc[block_owner(i, j)]; if ((i == nblocks-1) && (j == nblocks-1)) { size = edge*edge; } else if ((i == nblocks-1) || (j == nblocks-1)) { size = edge*block_size; } else { size = block_size*block_size; } ptr_loc[block_owner(i, j)] += size; } } /* initialize the array */ init_array(); /* barrier to ensure all initialization is done */ MP_BARRIER(); /* to remove cold-start misses, all processors touch their own data */ touch_array(block_size, me); MP_BARRIER(); if(doprint) { if(me == 0) { printf("Matrix before LU decomposition\n"); print_array(me); } MP_BARRIER(); } lu(n, block_size, me); /* cold start */ /* Starting the timer */ MP_BARRIER(); if(me == 0) start_timer(); for(i=0; i<nloop; i++) lu(n, block_size, me); MP_BARRIER(); /* Timer Stops here */ if(me == 0) printf("\nRunning time = %lf milliseconds.\n\n", elapsed_time()/nloop); printf("%d: (ngets=%d) Communication (get) time = %e milliseconds\n", me, get_cntr, comm_time*1000/nloop); if(doprint) { if(me == 0) { printf("after LU\n"); print_array(me); } MP_BARRIER(); } /* done */ #ifdef MPI2_ONESIDED MPI_Win_free(&win); MPI_Free_mem(ptr[me]); #else ARMCI_Free(ptr[me]); ARMCI_Finalize(); #endif MP_FINALIZE(); }
main(int argc, char *argv[]) { int i, j; int ch; extern char *optarg; int edge; int size; /* ARMCI */ void **ptr; double **ptr_loc; void **bufr_g, **bufc_g; MP_INIT(arc,argv); MP_PROCS(&nproc); MP_MYID(&me); while ((ch = getopt(argc, argv, "n:b:p:h")) != -1) { switch(ch) { case 'n': n = atoi(optarg); break; case 'b': block_size = atoi(optarg); break; case 'p': nproc = atoi(optarg); break; case 'h': { printf("Usage: LU, or \n"); printf(" LU -nMATRIXSIZE -bBLOCKSIZE -pNPROC\n"); MP_BARRIER(); MP_FINALIZE(); exit(0); } } } if(me == 0) { printf("\nUsing pre-PUTing\n"); printf("\n Blocked Dense LU Factorization\n"); printf(" %d by %d Matrix\n", n, n); printf(" %d Processors\n", nproc); printf(" %d by %d Element Blocks\n", block_size, block_size); printf("\n"); } num_rows = (int) sqrt((double) nproc); for (;;) { num_cols = nproc/num_rows; if (num_rows*num_cols == nproc) break; num_rows--; } nblocks = n/block_size; if (block_size * nblocks != n) { nblocks++; } edge = n%block_size; if (edge == 0) { edge = block_size; } #ifdef DEBUG if(me == 0) for (i=0;i<nblocks;i++) { for (j=0;j<nblocks;j++) printf("%d ", block_owner(i, j)); printf("\n"); } MP_BARRIER(); MP_FINALIZE(); exit(0); #endif for (i=0;i<nblocks;i++) { for (j=0;j<nblocks;j++) { if(block_owner(i,j) == me) { if ((i == nblocks-1) && (j == nblocks-1)) { size = edge*edge; } else if ((i == nblocks-1) || (j == nblocks-1)) { size = edge*block_size; } else { size = block_size*block_size; } proc_bytes += size*sizeof(double); } } } /* initialize ARMCI */ ARMCI_Init(); ptr = (void **)ARMCI_Malloc_local(nproc * sizeof(void *)); ARMCI_Malloc(ptr, proc_bytes); a = (double **)ARMCI_Malloc_local(nblocks*nblocks*sizeof(double *)); if (a == NULL) { fprintf(stderr, "Could not malloc memory for a\n"); exit(-1); } ptr_loc = (double **)ARMCI_Malloc_local(nproc*sizeof(double *)); for(i=0; i<nproc; i++) ptr_loc[i] = (double *)ptr[i]; for(i=0; i<nblocks;i ++) { for(j=0; j<nblocks; j++) { a[i+j*nblocks] = ptr_loc[block_owner(i, j)]; if ((i == nblocks-1) && (j == nblocks-1)) { size = edge*edge; } else if ((i == nblocks-1) || (j == nblocks-1)) { size = edge*block_size; } else { size = block_size*block_size; } ptr_loc[block_owner(i, j)] += size; } } /* initialize the array */ init_array(); bufr = (double **)ARMCI_Malloc_local(nproc*nblocks * sizeof(double *)); bufc = (double **)ARMCI_Malloc_local(nproc*nblocks * sizeof(double *)); if (bufr == NULL || bufc == NULL) printf("Could not ARMCI_Malloc_local() mem\n"); /* bufr points to all k-th row blocks */ /* save all block address in row-major order */ proc_bytes = nblocks*block_size*block_size * sizeof(double); bufr_g = (void **)ARMCI_Malloc_local(nproc * sizeof(void *)); ARMCI_Malloc(bufr_g, proc_bytes); for (i = 0; i < nproc; i++) { bufr[i*nblocks] = (double *) bufr_g[i]; for (j = 1; j < nblocks; j++) { bufr[i*nblocks + j] = bufr[i*nblocks + j-1] + block_size * block_size; } } /* bufc points to all k-th column blocks */ bufc_g = (void **)ARMCI_Malloc_local(nproc * sizeof(void *)); ARMCI_Malloc(bufc_g, proc_bytes); for (i = 0; i < nproc; i++) { bufc[i*nblocks] = (double *) bufc_g[i]; for (j = 1; j < nblocks; j++) { bufc[i*nblocks + j] = bufc[i*nblocks + j-1] + block_size * block_size; } } /* barrier to ensure all initialization is done */ MP_BARRIER(); /* to remove cold-start misses, all processors touch their own data */ touch_array(block_size, me); MP_BARRIER(); if(doprint) { if(me == 0) { printf("Matrix before LU decomposition\n"); print_array(me); } MP_BARRIER(); } /* Starting the timer */ if(me == 0) start_timer(); lu(n, block_size, me); MP_BARRIER(); /* Timer Stops here */ if(me == 0) printf("\nRunning time = %lf milliseconds.\n\n", elapsed_time()); if(doprint) { if(me == 0) { printf("after LU\n"); print_array(me); } MP_BARRIER(); } /* done */ ARMCI_Free(ptr[me]); ARMCI_Free(bufc_g[me]); ARMCI_Free(bufr_g[me]); ARMCI_Finalize(); MP_FINALIZE(); }