1 #include "colls_private.h"
3 * implemented by Pitch Patarasuk, 07/01/2007
5 //#include <star-reduction.c>
7 /* change number of core per smp-node
8 we assume that number of core per process will be the same for all implementations */
14 This fucntion performs all-reduce operation as follow.
15 1) binomial_tree reduce inside each SMP node
16 2) reduce-scatter -inter between root of each SMP node
17 3) allgather - inter between root of each SMP node
18 4) binomial_tree bcast inside each SMP node
20 int smpi_coll_tuned_allreduce_smp_rsag_rab(void *sbuf, void *rbuf, int count,
21 MPI_Datatype dtype, MPI_Op op,
26 int tag = COLL_TAG_ALLREDUCE;
29 int num_core = simcall_host_get_core(SIMIX_host_self());
30 // do we use the default one or the number of cores in the platform ?
31 // if the number of cores is one, the platform may be simulated with 1 node = 1 core
32 if (num_core == 1) num_core = NUM_CORE;
34 comm_size = smpi_comm_size(comm);
36 if((comm_size&(comm_size-1)))
37 THROWF(arg_error,0, "allreduce smp rsag rab algorithm can't be used with non power of two number of processes ! ");
39 rank = smpi_comm_rank(comm);
41 extent = smpi_datatype_get_extent(dtype);
42 tmp_buf = (void *) xbt_malloc(count * extent);
44 int intra_rank, inter_rank;
45 intra_rank = rank % num_core;
46 inter_rank = rank / num_core;
48 //printf("node %d intra_rank = %d, inter_rank = %d\n", rank, intra_rank, inter_rank);
50 int inter_comm_size = (comm_size + num_core - 1) / num_core;
52 smpi_mpi_sendrecv(sbuf, count, dtype, rank, tag,
53 rbuf, count, dtype, rank, tag, comm, &status);
55 // SMP_binomial_reduce
57 while (mask < num_core) {
58 if ((mask & intra_rank) == 0) {
59 src = (inter_rank * num_core) + (intra_rank | mask);
60 // if (src < ((inter_rank + 1) * num_core)) {
61 if (src < comm_size) {
62 smpi_mpi_recv(tmp_buf, count, dtype, src, tag, comm, &status);
63 smpi_op_apply(op, tmp_buf, rbuf, &count, &dtype);
64 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
68 dst = (inter_rank * num_core) + (intra_rank & (~mask));
69 smpi_mpi_send(rbuf, count, dtype, dst, tag, comm);
70 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
77 // INTER: reduce-scatter
78 if (intra_rank == 0) {
80 int dst, base_offset, send_base_offset, recv_base_offset, recv_chunk;
81 int curr_count, i, recv_offset, send_offset;
85 recv_chunk = extent * count / (comm_size / num_core);
89 curr_count = count / 2;
95 while (mask < (comm_size / num_core)) {
96 dst = inter_rank ^ mask;
99 send_base_offset = base_offset;
102 if (inter_rank & mask) {
103 recv_base_offset = base_offset + curr_count;
104 send_base_offset = base_offset;
105 base_offset = recv_base_offset;
109 recv_base_offset = base_offset;
110 send_base_offset = base_offset + curr_count;
112 send_offset = send_base_offset * extent;
113 recv_offset = recv_base_offset * extent;
116 // printf("node %d send to %d in phase %d s_offset = %d r_offset = %d count = %d\n",rank,dst,phase, send_offset, recv_offset, curr_count);
118 smpi_mpi_sendrecv((char *)rbuf + send_offset, curr_count, dtype, (dst * num_core), tag,
119 tmp_buf, curr_count, dtype, (dst * num_core), tag,
122 smpi_op_apply(op, tmp_buf, (char *)rbuf + recv_offset, &curr_count, &dtype);
132 int size = (comm_size / num_core) / 2;
135 while (mask < (comm_size / num_core)) {
136 if (inter_rank & mask) {
148 // destination pair for both send and recv
149 dst = inter_rank ^ mask;
152 send_base_offset = base_offset;
153 if (inter_rank & mask) {
154 recv_base_offset = base_offset - i;
157 recv_base_offset = base_offset + i;
159 send_offset = send_base_offset * recv_chunk;
160 recv_offset = recv_base_offset * recv_chunk;
163 //printf("node %d send to %d in phase %d s_offset = %d r_offset = %d count = %d\n",rank,dst,phase, send_offset, recv_offset, curr_count);
165 smpi_mpi_sendrecv((char *)rbuf + send_offset, curr_count, dtype, (dst * num_core), tag,
166 (char *)rbuf + recv_offset, curr_count, dtype, (dst * num_core), tag,
179 // intra SMP binomial bcast
181 int num_core_in_current_smp = num_core;
182 if (inter_rank == (inter_comm_size - 1)) {
183 num_core_in_current_smp = comm_size - (inter_rank * num_core);
185 // printf("Node %d num_core = %d\n",rank, num_core_in_current_smp);
187 while (mask < num_core_in_current_smp) {
188 if (intra_rank & mask) {
189 src = (inter_rank * num_core) + (intra_rank - mask);
190 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
191 smpi_mpi_recv(rbuf, count, dtype, src, tag, comm, &status);
198 //printf("My rank = %d my mask = %d\n", rank,mask);
201 dst = (inter_rank * num_core) + (intra_rank + mask);
202 if (dst < comm_size) {
203 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
204 smpi_mpi_send(rbuf, count, dtype, dst, tag, comm);