1 #include "colls_private.h"
2 //#include <star-reduction.c>
4 /* change number of core per smp-node
5 we assume that number of core per process will be the same for all implementations */
11 This fucntion performs all-reduce operation as follow.
12 1) binomial_tree reduce inside each SMP node
13 2) reduce-scatter -inter between root of each SMP node
14 3) allgather - inter between root of each SMP node
15 4) binomial_tree bcast inside each SMP node
17 int smpi_coll_tuned_allreduce_smp_rsag_lr(void *send_buf, void *recv_buf,
18 int count, MPI_Datatype dtype,
19 MPI_Op op, MPI_Comm comm)
23 int tag = COLL_TAG_ALLREDUCE;
26 int num_core = NUM_CORE;
28 #ifdef MPICH2_REDUCTION
29 MPI_User_function * uop = MPIR_Op_table[op % 16 - 1];
31 MPI_User_function *uop;
32 struct MPIR_OP *op_ptr;
33 op_ptr = MPIR_ToPointer(op);
37 comm_size = smpi_comm_size(comm);
38 rank = smpi_comm_rank(comm);
40 extent = smpi_datatype_get_extent(dtype);
41 tmp_buf = (void *) xbt_malloc(count * extent);
43 int intra_rank, inter_rank;
44 intra_rank = rank % num_core;
45 inter_rank = rank / num_core;
47 //printf("node %d intra_rank = %d, inter_rank = %d\n", rank, intra_rank, inter_rank);
49 int inter_comm_size = (comm_size + num_core - 1) / num_core;
52 //printf("intra com size = %d\n",num_core);
53 //printf("inter com size = %d\n",inter_comm_size);
57 smpi_mpi_sendrecv(send_buf, count, dtype, rank, tag,
58 recv_buf, count, dtype, rank, tag, comm, &status);
61 // SMP_binomial_reduce
63 while (mask < num_core) {
64 if ((mask & intra_rank) == 0) {
65 src = (inter_rank * num_core) + (intra_rank | mask);
66 // if (src < ((inter_rank + 1) * num_core)) {
67 if (src < comm_size) {
68 smpi_mpi_recv(tmp_buf, count, dtype, src, tag, comm, &status);
69 smpi_op_apply(op, tmp_buf, recv_buf, &count, &dtype);
70 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
74 dst = (inter_rank * num_core) + (intra_rank & (~mask));
75 smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
76 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
84 // INTER: reduce-scatter
85 if (intra_rank == 0) {
86 int send_offset, recv_offset;
87 int send_count, recv_count;
88 int curr_size = count / inter_comm_size;
89 int curr_remainder = count % inter_comm_size;
91 int to = ((inter_rank + 1) % inter_comm_size) * num_core;
93 ((inter_rank + inter_comm_size - 1) % inter_comm_size) * num_core;
96 //printf("node %d to %d from %d\n",rank,to,from);
98 /* last segment may have a larger size since it also include the remainder */
99 int last_segment_ptr =
100 (inter_comm_size - 1) * (count / inter_comm_size) * extent;
102 for (i = 0; i < (inter_comm_size - 1); i++) {
105 ((inter_rank - 1 - i +
106 inter_comm_size) % inter_comm_size) * curr_size * extent;
108 ((inter_rank - 2 - i +
109 inter_comm_size) % inter_comm_size) * curr_size * extent;
112 if (send_offset != last_segment_ptr)
113 send_count = curr_size;
115 send_count = curr_size + curr_remainder;
117 if (recv_offset != last_segment_ptr)
118 recv_count = curr_size;
120 recv_count = curr_size + curr_remainder;
122 smpi_mpi_sendrecv((char *) recv_buf + send_offset, send_count, dtype, to,
123 tag + i, tmp_buf, recv_count, dtype, from, tag + i, comm,
127 smpi_op_apply(op, tmp_buf, (char *) recv_buf + recv_offset, &recv_count,
132 for (i = 0; i < (inter_comm_size - 1); i++) {
136 inter_comm_size) % inter_comm_size) * curr_size * extent;
138 ((inter_rank - 1 - i +
139 inter_comm_size) % inter_comm_size) * curr_size * extent;
142 if (send_offset != last_segment_ptr)
143 send_count = curr_size;
145 send_count = curr_size + curr_remainder;
147 if (recv_offset != last_segment_ptr)
148 recv_count = curr_size;
150 recv_count = curr_size + curr_remainder;
152 smpi_mpi_sendrecv((char *) recv_buf + send_offset, send_count, dtype, to,
153 tag + i, (char *) recv_buf + recv_offset, recv_count, dtype,
154 from, tag + i, comm, &status);
162 // INTER_binomial_reduce
164 // only root node for each SMP
165 if (intra_rank == 0) {
168 while (mask < inter_comm_size) {
169 if ((mask & inter_rank) == 0) {
170 src = (inter_rank | mask) * num_core;
171 if (src < comm_size) {
172 smpi_mpi_recv(tmp_buf, count, dtype, src, tag, comm, &status);
173 (* uop) (tmp_buf, recv_buf, &count, &dtype);
174 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
178 dst = (inter_rank & (~mask)) * num_core;
179 smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
180 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
189 // INTER_binomial_bcast
192 if (intra_rank == 0) {
194 while (mask < inter_comm_size) {
195 if (inter_rank & mask) {
196 src = (inter_rank - mask) * num_core;
197 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
198 smpi_mpi_recv(recv_buf, count, dtype, src, tag, comm, &status);
205 //printf("My rank = %d my mask = %d\n", rank,mask);
208 if (inter_rank < inter_comm_size) {
209 dst = (inter_rank + mask) * num_core;
210 if (dst < comm_size) {
211 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
212 smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
221 // INTRA_binomial_bcast
223 int num_core_in_current_smp = num_core;
224 if (inter_rank == (inter_comm_size - 1)) {
225 num_core_in_current_smp = comm_size - (inter_rank * num_core);
227 // printf("Node %d num_core = %d\n",rank, num_core_in_current_smp);
229 while (mask < num_core_in_current_smp) {
230 if (intra_rank & mask) {
231 src = (inter_rank * num_core) + (intra_rank - mask);
232 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
233 smpi_mpi_recv(recv_buf, count, dtype, src, tag, comm, &status);
240 //printf("My rank = %d my mask = %d\n", rank,mask);
243 dst = (inter_rank * num_core) + (intra_rank + mask);
244 if (dst < comm_size) {
245 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
246 smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);