int smpi_coll_tuned_allgatherv_mpich_rdb ( void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int *recvcounts, int *displs, MPI_Datatype recvtype, MPI_Comm comm) { int comm_size, rank, j, i; MPI_Status status; MPI_Aint recvtype_extent, recvtype_true_extent, recvtype_true_lb; int curr_cnt, dst, total_count; void *tmp_buf, *tmp_buf_rl; int mask, dst_tree_root, my_tree_root, position, send_offset, recv_offset, last_recv_cnt=0, nprocs_completed, k, offset, tmp_mask, tree_root; comm_size = smpi_comm_size(comm); rank = smpi_comm_rank(comm); total_count = 0; for (i=0; i<comm_size; i++) total_count += recvcounts[i]; if (total_count == 0) return MPI_ERR_COUNT; recvtype_extent=smpi_datatype_get_extent( recvtype); /* need to receive contiguously into tmp_buf because displs could make the recvbuf noncontiguous */ smpi_datatype_extent(recvtype, &recvtype_true_lb, &recvtype_true_extent); tmp_buf_rl= (void*)smpi_get_tmp_sendbuffer(total_count*(MAX(recvtype_true_extent,recvtype_extent))); /* adjust for potential negative lower bound in datatype */ tmp_buf = (void *)((char*)tmp_buf_rl - recvtype_true_lb); /* copy local data into right location in tmp_buf */ position = 0; for (i=0; i<rank; i++) position += recvcounts[i]; if (sendbuf != MPI_IN_PLACE) { smpi_datatype_copy(sendbuf, sendcount, sendtype, ((char *)tmp_buf + position* recvtype_extent), recvcounts[rank], recvtype); } else { /* if in_place specified, local data is found in recvbuf */ smpi_datatype_copy(((char *)recvbuf + displs[rank]*recvtype_extent), recvcounts[rank], recvtype, ((char *)tmp_buf + position* recvtype_extent), recvcounts[rank], recvtype); } curr_cnt = recvcounts[rank]; mask = 0x1; i = 0; while (mask < comm_size) { dst = rank ^ mask; /* find offset into send and recv buffers. zero out the least significant "i" bits of rank and dst to find root of src and dst subtrees. Use ranks of roots as index to send from and recv into buffer */ dst_tree_root = dst >> i; dst_tree_root <<= i; my_tree_root = rank >> i; my_tree_root <<= i; if (dst < comm_size) { send_offset = 0; for (j=0; j<my_tree_root; j++) send_offset += recvcounts[j]; recv_offset = 0; for (j=0; j<dst_tree_root; j++) recv_offset += recvcounts[j]; smpi_mpi_sendrecv(((char *)tmp_buf + send_offset * recvtype_extent), curr_cnt, recvtype, dst, COLL_TAG_ALLGATHERV, ((char *)tmp_buf + recv_offset * recvtype_extent), total_count - recv_offset, recvtype, dst, COLL_TAG_ALLGATHERV, comm, &status); /* for convenience, recv is posted for a bigger amount than will be sent */ last_recv_cnt=smpi_mpi_get_count(&status, recvtype); curr_cnt += last_recv_cnt; } /* if some processes in this process's subtree in this step did not have any destination process to communicate with because of non-power-of-two, we need to send them the data that they would normally have received from those processes. That is, the haves in this subtree must send to the havenots. We use a logarithmic recursive-halfing algorithm for this. */ /* This part of the code will not currently be executed because we are not using recursive doubling for non power of two. Mark it as experimental so that it doesn't show up as red in the coverage tests. */ /* --BEGIN EXPERIMENTAL-- */ if (dst_tree_root + mask > comm_size) { nprocs_completed = comm_size - my_tree_root - mask; /* nprocs_completed is the number of processes in this subtree that have all the data. Send data to others in a tree fashion. First find root of current tree that is being divided into two. k is the number of least-significant bits in this process's rank that must be zeroed out to find the rank of the root */ j = mask; k = 0; while (j) { j >>= 1; k++; } k--; tmp_mask = mask >> 1; while (tmp_mask) { dst = rank ^ tmp_mask; tree_root = rank >> k; tree_root <<= k; /* send only if this proc has data and destination doesn't have data. at any step, multiple processes can send if they have the data */ if ((dst > rank) && (rank < tree_root + nprocs_completed) && (dst >= tree_root + nprocs_completed)) { offset = 0; for (j=0; j<(my_tree_root+mask); j++) offset += recvcounts[j]; offset *= recvtype_extent; smpi_mpi_send(((char *)tmp_buf + offset), last_recv_cnt, recvtype, dst, COLL_TAG_ALLGATHERV, comm); /* last_recv_cnt was set in the previous receive. that's the amount of data to be sent now. */ } /* recv only if this proc. doesn't have data and sender has data */ else if ((dst < rank) && (dst < tree_root + nprocs_completed) && (rank >= tree_root + nprocs_completed)) { offset = 0; for (j=0; j<(my_tree_root+mask); j++) offset += recvcounts[j]; smpi_mpi_recv(((char *)tmp_buf + offset * recvtype_extent), total_count - offset, recvtype, dst, COLL_TAG_ALLGATHERV, comm, &status); /* for convenience, recv is posted for a bigger amount than will be sent */ last_recv_cnt=smpi_mpi_get_count(&status, recvtype); curr_cnt += last_recv_cnt; } tmp_mask >>= 1; k--; } } /* --END EXPERIMENTAL-- */ mask <<= 1; i++; }
int smpi_coll_tuned_bcast_scatter_rdb_allgather(void *buff, int count, MPI_Datatype data_type, int root, MPI_Comm comm) { MPI_Aint extent; MPI_Status status; int i, j, k, src, dst, rank, num_procs, send_offset, recv_offset; int mask, relative_rank, curr_size, recv_size = 0, send_size, nbytes; int scatter_size, tree_root, relative_dst, dst_tree_root; int my_tree_root, offset, tmp_mask, num_procs_completed; int tag = COLL_TAG_BCAST; rank = smpi_comm_rank(comm); num_procs = smpi_comm_size(comm); extent = smpi_datatype_get_extent(data_type); nbytes = extent * count; scatter_size = (nbytes + num_procs - 1) / num_procs; // ceiling division curr_size = (rank == root) ? nbytes : 0; // root starts with all the data relative_rank = (rank >= root) ? rank - root : rank - root + num_procs; mask = 0x1; while (mask < num_procs) { if (relative_rank & mask) { src = rank - mask; if (src < 0) src += num_procs; recv_size = nbytes - relative_rank * scatter_size; // recv_size is larger than what might actually be sent by the // sender. We don't need compute the exact value because MPI // allows you to post a larger recv. if (recv_size <= 0) curr_size = 0; // this process doesn't receive any data // because of uneven division else { smpi_mpi_recv((char *)buff + relative_rank * scatter_size, recv_size, MPI_BYTE, src, tag, comm, &status); curr_size = smpi_mpi_get_count(&status, MPI_BYTE); } break; } mask <<= 1; } // This process is responsible for all processes that have bits // set from the LSB upto (but not including) mask. Because of // the "not including", we start by shifting mask back down // one. mask >>= 1; while (mask > 0) { if (relative_rank + mask < num_procs) { send_size = curr_size - scatter_size * mask; // mask is also the size of this process's subtree if (send_size > 0) { dst = rank + mask; if (dst >= num_procs) dst -= num_procs; smpi_mpi_send((char *)buff + scatter_size * (relative_rank + mask), send_size, MPI_BYTE, dst, tag, comm); curr_size -= send_size; } } mask >>= 1; } // done scatter now do allgather mask = 0x1; i = 0; while (mask < num_procs) { relative_dst = relative_rank ^ mask; dst = (relative_dst + root) % num_procs; /* find offset into send and recv buffers. zero out the least significant "i" bits of relative_rank and relative_dst to find root of src and dst subtrees. Use ranks of roots as index to send from and recv into buffer */ dst_tree_root = relative_dst >> i; dst_tree_root <<= i; my_tree_root = relative_rank >> i; my_tree_root <<= i; send_offset = my_tree_root * scatter_size; recv_offset = dst_tree_root * scatter_size; if (relative_dst < num_procs) { smpi_mpi_sendrecv((char *)buff + send_offset, curr_size, MPI_BYTE, dst, tag, (char *)buff + recv_offset, scatter_size * mask, MPI_BYTE, dst, tag, comm, &status); recv_size = smpi_mpi_get_count(&status, MPI_BYTE); curr_size += recv_size; } /* if some processes in this process's subtree in this step did not have any destination process to communicate with because of non-power-of-two, we need to send them the data that they would normally have received from those processes. That is, the haves in this subtree must send to the havenots. We use a logarithmic recursive-halfing algorithm for this. */ if (dst_tree_root + mask > num_procs) { num_procs_completed = num_procs - my_tree_root - mask; /* num_procs_completed is the number of processes in this subtree that have all the data. Send data to others in a tree fashion. First find root of current tree that is being divided into two. k is the number of least-significant bits in this process's rank that must be zeroed out to find the rank of the root */ j = mask; k = 0; while (j) { j >>= 1; k++; } k--; offset = scatter_size * (my_tree_root + mask); tmp_mask = mask >> 1; while (tmp_mask) { relative_dst = relative_rank ^ tmp_mask; dst = (relative_dst + root) % num_procs; tree_root = relative_rank >> k; tree_root <<= k; /* send only if this proc has data and destination doesn't have data. */ if ((relative_dst > relative_rank) && (relative_rank < tree_root + num_procs_completed) && (relative_dst >= tree_root + num_procs_completed)) { smpi_mpi_send((char *)buff + offset, recv_size, MPI_BYTE, dst, tag, comm); /* recv_size was set in the previous receive. that's the amount of data to be sent now. */ } /* recv only if this proc. doesn't have data and sender has data */ else if ((relative_dst < relative_rank) && (relative_dst < tree_root + num_procs_completed) && (relative_rank >= tree_root + num_procs_completed)) { smpi_mpi_recv((char *)buff + offset, scatter_size * num_procs_completed, MPI_BYTE, dst, tag, comm, &status); /* num_procs_completed is also equal to the no. of processes whose data we don't have */ recv_size = smpi_mpi_get_count(&status, MPI_BYTE); curr_size += recv_size; } tmp_mask >>= 1; k--; } } mask <<= 1; i++; }
int smpi_coll_tuned_allgather_rdb(void *sbuf, int send_count, MPI_Datatype send_type, void *rbuf, int recv_count, MPI_Datatype recv_type, MPI_Comm comm) { // MPI variables MPI_Status status; MPI_Aint send_chunk, recv_chunk; // local int variables unsigned int i, j, k, dst, send_offset, recv_offset, tree_root; int dst_tree_root, rank_tree_root, last_recv_count = 0, num_procs_completed; int offset, tmp_mask; int tag = COLL_TAG_ALLGATHER; int mask = 1; int success = 0; int curr_count = recv_count; // local string variables char *send_ptr = (char *) sbuf; char *recv_ptr = (char *) rbuf; // get size of the communicator, followed by rank unsigned int num_procs = smpi_comm_size(comm); unsigned int rank = smpi_comm_rank(comm); // get size of single element's type for send buffer and recv buffer send_chunk = smpi_datatype_get_extent(send_type); recv_chunk = smpi_datatype_get_extent(recv_type); // multiply size of each element by number of elements to send or recv send_chunk *= send_count; recv_chunk *= recv_count; // perform a local copy smpi_mpi_sendrecv(send_ptr, send_count, send_type, rank, tag, recv_ptr + rank * recv_chunk, recv_count, recv_type, rank, tag, comm, &status); i = 0; while (mask < num_procs) { dst = rank ^ mask; dst_tree_root = dst >> i; dst_tree_root <<= i; rank_tree_root = rank >> i; rank_tree_root <<= i; send_offset = rank_tree_root * send_chunk; recv_offset = dst_tree_root * recv_chunk; if (dst < num_procs) { smpi_mpi_sendrecv(recv_ptr + send_offset, curr_count, send_type, dst, tag, recv_ptr + recv_offset, mask * recv_count, recv_type, dst, tag, comm, &status); last_recv_count = smpi_mpi_get_count(&status, recv_type); curr_count += last_recv_count; } if (dst_tree_root + mask > num_procs) { num_procs_completed = num_procs - rank_tree_root - mask; /* num_procs_completed is the number of processes in this subtree that have all the data. Send data to others in a tree fashion. First find root of current tree that is being divided into two. k is the number of least-significant bits in this process's rank that must be zeroed out to find the rank of the root */ j = mask; k = 0; while (j) { j >>= 1; k++; } k--; offset = recv_chunk * (rank_tree_root + mask); tmp_mask = mask >> 1; while (tmp_mask) { dst = rank ^ tmp_mask; tree_root = rank >> k; tree_root <<= k; /* send only if this proc has data and destination doesn't have data. at any step, multiple processes can send if they have the data */ if ((dst > rank) && (rank < tree_root + num_procs_completed) && (dst >= tree_root + num_procs_completed)) { smpi_mpi_send(recv_ptr + offset, last_recv_count, recv_type, dst, tag, comm); /* last_recv_cnt was set in the previous receive. that's the amount of data to be sent now. */ } /* recv only if this proc. doesn't have data and sender has data */ else if ((dst < rank) && (dst < tree_root + num_procs_completed) && (rank >= tree_root + num_procs_completed)) { smpi_mpi_recv(recv_ptr + offset, recv_count * num_procs_completed, recv_type, dst, tag, comm, &status); // num_procs_completed is also equal to the no. of processes // whose data we don't have last_recv_count = smpi_mpi_get_count(&status, recv_type); curr_count += last_recv_count; } tmp_mask >>= 1; k--; } } mask <<= 1; i++; }