X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/46da24e9999dbd79a6deafaa4e4ed0ac0db7f1c4..8ec703af1cf974ed8097c2e96cc051f1f0d38ac7:/src/smpi/colls/smpi_mvapich2_selector.c diff --git a/src/smpi/colls/smpi_mvapich2_selector.c b/src/smpi/colls/smpi_mvapich2_selector.c index 49906afab8..32408edf52 100644 --- a/src/smpi/colls/smpi_mvapich2_selector.c +++ b/src/smpi/colls/smpi_mvapich2_selector.c @@ -3,7 +3,7 @@ /* Copyright (c) 2009-2010, 2013-2014. The SimGrid Team. * All rights reserved. */ -/* This program is free software; you can redistribute it and/or modify it +/* This program is xbt_free software; you can redistribute it and/or modify it * under the terms of the license (GNU LGPL) which comes with this package. */ #include "colls_private.h" @@ -11,249 +11,6 @@ #include "smpi_mvapich2_selector_stampede.h" -static void init_mv2_alltoall_tables_stampede(){ -int i; - int agg_table_sum = 0; -mv2_alltoall_tuning_table **table_ptrs = NULL; - mv2_alltoall_num_ppn_conf = 3; - mv2_alltoall_thresholds_table - = malloc(sizeof(mv2_alltoall_tuning_table *) - * mv2_alltoall_num_ppn_conf); - table_ptrs = malloc(sizeof(mv2_alltoall_tuning_table *) - * mv2_alltoall_num_ppn_conf); - mv2_size_alltoall_tuning_table = malloc(sizeof(int) * - mv2_alltoall_num_ppn_conf); - mv2_alltoall_table_ppn_conf =malloc(mv2_alltoall_num_ppn_conf * sizeof(int)); - mv2_alltoall_table_ppn_conf[0] = 1; - mv2_size_alltoall_tuning_table[0] = 6; - mv2_alltoall_tuning_table mv2_tmp_alltoall_thresholds_table_1ppn[] = { - {2, - 1, - {{0, -1, &MPIR_Alltoall_pairwise_MV2}, - }, - - {{0, -1, &MPIR_Alltoall_inplace_MV2}, - }, - }, - - {4, - 2, - {{0, 262144, &MPIR_Alltoall_Scatter_dest_MV2}, - {262144, -1, &MPIR_Alltoall_pairwise_MV2}, - }, - - {{0, -1, &MPIR_Alltoall_inplace_MV2}, - }, - }, - - {8, - 2, - {{0, 8, &MPIR_Alltoall_RD_MV2}, - {8, -1, &MPIR_Alltoall_Scatter_dest_MV2}, - }, - - {{0, -1, &MPIR_Alltoall_inplace_MV2}, - }, - }, - - {16, - 3, - {{0, 64, &MPIR_Alltoall_RD_MV2}, - {64, 512, &MPIR_Alltoall_bruck_MV2}, - {512, -1, &MPIR_Alltoall_Scatter_dest_MV2}, - }, - - {{0,-1, &MPIR_Alltoall_inplace_MV2}, - }, - }, - - {32, - 3, - {{0, 32, &MPIR_Alltoall_RD_MV2}, - {32, 2048, &MPIR_Alltoall_bruck_MV2}, - {2048, -1, &MPIR_Alltoall_Scatter_dest_MV2}, - }, - - {{0, -1, &MPIR_Alltoall_inplace_MV2}, - }, - }, - - {64, - 3, - {{0, 8, &MPIR_Alltoall_RD_MV2}, - {8, 1024, &MPIR_Alltoall_bruck_MV2}, - {1024, -1, &MPIR_Alltoall_Scatter_dest_MV2}, - }, - - {{0, -1, &MPIR_Alltoall_inplace_MV2}, - }, - }, - }; - table_ptrs[0] = mv2_tmp_alltoall_thresholds_table_1ppn; - mv2_alltoall_table_ppn_conf[1] = 2; - mv2_size_alltoall_tuning_table[1] = 6; - mv2_alltoall_tuning_table mv2_tmp_alltoall_thresholds_table_2ppn[] = { - {4, - 2, - {{0, 32, &MPIR_Alltoall_RD_MV2}, - {32, -1, &MPIR_Alltoall_Scatter_dest_MV2}, - }, - - {{0, -1, &MPIR_Alltoall_inplace_MV2}, - }, - }, - - {8, - 2, - {{0, 64, &MPIR_Alltoall_RD_MV2}, - {64, -1, &MPIR_Alltoall_Scatter_dest_MV2}, - }, - - {{0, -1, &MPIR_Alltoall_inplace_MV2}, - }, - }, - - {16, - 3, - {{0, 64, &MPIR_Alltoall_RD_MV2}, - {64, 2048, &MPIR_Alltoall_bruck_MV2}, - {2048, -1, &MPIR_Alltoall_Scatter_dest_MV2}, - }, - - {{0,-1, &MPIR_Alltoall_inplace_MV2}, - }, - }, - - {32, - 3, - {{0, 16, &MPIR_Alltoall_RD_MV2}, - {16, 2048, &MPIR_Alltoall_bruck_MV2}, - {2048, -1, &MPIR_Alltoall_Scatter_dest_MV2}, - }, - - {{0, -1, &MPIR_Alltoall_inplace_MV2}, - }, - }, - - {64, - 3, - {{0, 8, &MPIR_Alltoall_RD_MV2}, - {8, 1024, &MPIR_Alltoall_bruck_MV2}, - {1024, -1, &MPIR_Alltoall_Scatter_dest_MV2}, - }, - - {{0, -1, &MPIR_Alltoall_inplace_MV2}, - }, - }, - - {128, - 3, - {{0, 4, &MPIR_Alltoall_RD_MV2}, - {4, 2048, &MPIR_Alltoall_bruck_MV2}, - {2048, -1, &MPIR_Alltoall_Scatter_dest_MV2}, - }, - - {{0, -1, &MPIR_Alltoall_inplace_MV2}, - }, - }, - }; - table_ptrs[1] = mv2_tmp_alltoall_thresholds_table_2ppn; - mv2_alltoall_table_ppn_conf[2] = 16; - mv2_size_alltoall_tuning_table[2] = 7; - mv2_alltoall_tuning_table mv2_tmp_alltoall_thresholds_table_16ppn[] = { - {16, - 2, - {{0, 2048, &MPIR_Alltoall_bruck_MV2}, - {2048, -1, &MPIR_Alltoall_Scatter_dest_MV2}, - }, - - {{32768, -1, &MPIR_Alltoall_inplace_MV2}, - }, - }, - - {32, - 2, - {{0, 2048, &MPIR_Alltoall_bruck_MV2}, - {2048, -1, &MPIR_Alltoall_Scatter_dest_MV2}, - }, - - {{16384, -1, &MPIR_Alltoall_inplace_MV2}, - }, - }, - - {64, - 3, - {{0, 2048, &MPIR_Alltoall_bruck_MV2}, - {2048, 16384, &MPIR_Alltoall_Scatter_dest_MV2}, - {16384, -1, &MPIR_Alltoall_pairwise_MV2}, - }, - - {{32768, 131072, &MPIR_Alltoall_inplace_MV2}, - }, - }, - - {128, - 2, - {{0, 2048, &MPIR_Alltoall_bruck_MV2}, - {2048, -1, &MPIR_Alltoall_pairwise_MV2}, - }, - - {{16384,65536, &MPIR_Alltoall_inplace_MV2}, - }, - }, - - {256, - 2, - {{0, 1024, &MPIR_Alltoall_bruck_MV2}, - {1024, -1, &MPIR_Alltoall_pairwise_MV2}, - }, - - {{16384, 65536, &MPIR_Alltoall_inplace_MV2}, - }, - }, - - {512, - 2, - {{0, 1024, &MPIR_Alltoall_bruck_MV2}, - {1024, -1, &MPIR_Alltoall_pairwise_MV2}, - }, - - {{16384, 65536, &MPIR_Alltoall_inplace_MV2}, - }, - }, - {1024, - 2, - {{0, 1024, &MPIR_Alltoall_bruck_MV2}, - {1024, -1, &MPIR_Alltoall_pairwise_MV2}, - }, - - {{16384, 65536, &MPIR_Alltoall_inplace_MV2}, - }, - }, - - }; - table_ptrs[2] = mv2_tmp_alltoall_thresholds_table_16ppn; - agg_table_sum = 0; - for (i = 0; i < mv2_alltoall_num_ppn_conf; i++) { - agg_table_sum += mv2_size_alltoall_tuning_table[i]; - } - mv2_alltoall_thresholds_table[0] = - malloc(agg_table_sum * sizeof (mv2_alltoall_tuning_table)); - memcpy(mv2_alltoall_thresholds_table[0], table_ptrs[0], - (sizeof(mv2_alltoall_tuning_table) - * mv2_size_alltoall_tuning_table[0])); - for (i = 1; i < mv2_alltoall_num_ppn_conf; i++) { - mv2_alltoall_thresholds_table[i] = - mv2_alltoall_thresholds_table[i - 1] - + mv2_size_alltoall_tuning_table[i - 1]; - memcpy(mv2_alltoall_thresholds_table[i], table_ptrs[i], - (sizeof(mv2_alltoall_tuning_table) - * mv2_size_alltoall_tuning_table[i])); - } - free(table_ptrs); - - -} int smpi_coll_tuned_alltoall_mvapich2( void *sendbuf, int sendcount, MPI_Datatype sendtype, @@ -304,7 +61,7 @@ int smpi_coll_tuned_alltoall_mvapich2( void *sendbuf, int sendcount, mv2_alltoall_thresholds_table[conf_index][range].in_place_algo_table[range_threshold].min ||nbytes > mv2_alltoall_thresholds_table[conf_index][range].in_place_algo_table[range_threshold].max ) { - tmp_buf = (char *)malloc( comm_size * recvcount * recvtype_size ); + tmp_buf = (char *)xbt_malloc( comm_size * recvcount * recvtype_size ); mpi_errno = smpi_datatype_copy((char *)recvbuf, comm_size*recvcount, recvtype, (char *)tmp_buf, @@ -313,7 +70,7 @@ int smpi_coll_tuned_alltoall_mvapich2( void *sendbuf, int sendcount, mpi_errno = MV2_Alltoall_function(tmp_buf, recvcount, recvtype, recvbuf, recvcount, recvtype, comm ); - free(tmp_buf); + xbt_free(tmp_buf); } else { mpi_errno = MPIR_Alltoall_inplace_MV2(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, @@ -326,221 +83,6 @@ int smpi_coll_tuned_alltoall_mvapich2( void *sendbuf, int sendcount, } -static void init_mv2_allgather_tables_stampede(){ -int i; - int agg_table_sum = 0; -mv2_allgather_tuning_table **table_ptrs = NULL; - mv2_allgather_num_ppn_conf = 3; - mv2_allgather_thresholds_table - = malloc(sizeof(mv2_allgather_tuning_table *) - * mv2_allgather_num_ppn_conf); - table_ptrs = malloc(sizeof(mv2_allgather_tuning_table *) - * mv2_allgather_num_ppn_conf); - mv2_size_allgather_tuning_table = malloc(sizeof(int) * - mv2_allgather_num_ppn_conf); - mv2_allgather_table_ppn_conf - = malloc(mv2_allgather_num_ppn_conf * sizeof(int)); - mv2_allgather_table_ppn_conf[0] = 1; - mv2_size_allgather_tuning_table[0] = 6; - mv2_allgather_tuning_table mv2_tmp_allgather_thresholds_table_1ppn[] = { - { - 2, - {0}, - 1, - { - {0, -1, &MPIR_Allgather_Ring_MV2}, - }, - }, - { - 4, - {0,0}, - 2, - { - {0, 262144, &MPIR_Allgather_RD_MV2}, - {262144, -1, &MPIR_Allgather_Ring_MV2}, - }, - }, - { - 8, - {0,0}, - 2, - { - {0, 131072, &MPIR_Allgather_RD_MV2}, - {131072, -1, &MPIR_Allgather_Ring_MV2}, - }, - }, - { - 16, - {0,0}, - 2, - { - {0, 131072, &MPIR_Allgather_RD_MV2}, - {131072, -1, &MPIR_Allgather_Ring_MV2}, - }, - }, - { - 32, - {0,0}, - 2, - { - {0, 65536, &MPIR_Allgather_RD_MV2}, - {65536, -1, &MPIR_Allgather_Ring_MV2}, - }, - }, - { - 64, - {0,0}, - 2, - { - {0, 32768, &MPIR_Allgather_RD_MV2}, - {32768, -1, &MPIR_Allgather_Ring_MV2}, - }, - }, - }; - table_ptrs[0] = mv2_tmp_allgather_thresholds_table_1ppn; - mv2_allgather_table_ppn_conf[1] = 2; - mv2_size_allgather_tuning_table[1] = 6; - mv2_allgather_tuning_table mv2_tmp_allgather_thresholds_table_2ppn[] = { - { - 4, - {0,0}, - 2, - { - {0, 524288, &MPIR_Allgather_RD_MV2}, - {524288, -1, &MPIR_Allgather_Ring_MV2}, - }, - }, - { - 8, - {0,1,0}, - 2, - { - {0, 32768, &MPIR_Allgather_RD_MV2}, - {32768, 524288, &MPIR_Allgather_Ring_MV2}, - {524288, -1, &MPIR_Allgather_Ring_MV2}, - }, - }, - { - 16, - {0,1,0}, - 2, - { - {0, 16384, &MPIR_Allgather_RD_MV2}, - {16384, 524288, &MPIR_Allgather_Ring_MV2}, - {524288, -1, &MPIR_Allgather_Ring_MV2}, - }, - }, - { - 32, - {1,1,0}, - 2, - { - {0, 65536, &MPIR_Allgather_RD_MV2}, - {65536, 524288, &MPIR_Allgather_Ring_MV2}, - {524288, -1, &MPIR_Allgather_Ring_MV2}, - }, - }, - { - 64, - {1,1,0}, - 2, - { - {0, 32768, &MPIR_Allgather_RD_MV2}, - {32768, 524288, &MPIR_Allgather_Ring_MV2}, - {524288, -1, &MPIR_Allgather_Ring_MV2}, - }, - }, - { - 128, - {1,1,0}, - 2, - { - {0, 65536, &MPIR_Allgather_RD_MV2}, - {65536, 524288, &MPIR_Allgather_Ring_MV2}, - {524288, -1, &MPIR_Allgather_Ring_MV2}, - }, - }, - }; - table_ptrs[1] = mv2_tmp_allgather_thresholds_table_2ppn; - mv2_allgather_table_ppn_conf[2] = 16; - mv2_size_allgather_tuning_table[2] = 6; - mv2_allgather_tuning_table mv2_tmp_allgather_thresholds_table_16ppn[] = { - { - 16, - {0,0}, - 2, - { - {0, 1024, &MPIR_Allgather_RD_MV2}, - {1024, -1, &MPIR_Allgather_Ring_MV2}, - }, - }, - { - 32, - {0,0}, - 2, - { - {0, 1024, &MPIR_Allgather_RD_Allgather_Comm_MV2}, - {1024, -1, &MPIR_Allgather_Ring_MV2}, - }, - }, - { - 64, - {0,0}, - 2, - { - {0, 1024, &MPIR_Allgather_RD_Allgather_Comm_MV2}, - {1024, -1, &MPIR_Allgather_Ring_MV2}, - }, - }, - { - 128, - {0,0}, - 2, - { - {0, 1024, &MPIR_Allgather_RD_Allgather_Comm_MV2}, - {1024, -1, &MPIR_Allgather_Ring_MV2}, - }, - }, - { - 256, - {0,0}, - 2, - { - {0, 1024, &MPIR_Allgather_RD_Allgather_Comm_MV2}, - {1024, -1, &MPIR_Allgather_Ring_MV2}, - }, - }, - { - 512, - {0,0}, - 2, - { - {0, 1024, &MPIR_Allgather_RD_Allgather_Comm_MV2}, - {1024, -1, &MPIR_Allgather_Ring_MV2}, - }, - }, - - }; - table_ptrs[2] = mv2_tmp_allgather_thresholds_table_16ppn; - agg_table_sum = 0; - for (i = 0; i < mv2_allgather_num_ppn_conf; i++) { - agg_table_sum += mv2_size_allgather_tuning_table[i]; - } - mv2_allgather_thresholds_table[0] = - malloc(agg_table_sum * sizeof (mv2_allgather_tuning_table)); - memcpy(mv2_allgather_thresholds_table[0], table_ptrs[0], - (sizeof(mv2_allgather_tuning_table) - * mv2_size_allgather_tuning_table[0])); - for (i = 1; i < mv2_allgather_num_ppn_conf; i++) { - mv2_allgather_thresholds_table[i] = - mv2_allgather_thresholds_table[i - 1] - + mv2_size_allgather_tuning_table[i - 1]; - memcpy(mv2_allgather_thresholds_table[i], table_ptrs[i], - (sizeof(mv2_allgather_tuning_table) - * mv2_size_allgather_tuning_table[i])); - } - free(table_ptrs); -} int smpi_coll_tuned_allgather_mvapich2(void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, @@ -647,53 +189,6 @@ int smpi_coll_tuned_allgather_mvapich2(void *sendbuf, int sendcount, MPI_Datatyp return mpi_errno; } -static void init_mv2_gather_tables_stampede(){ - - mv2_size_gather_tuning_table=7; - mv2_gather_thresholds_table = malloc(mv2_size_gather_tuning_table* - sizeof (mv2_gather_tuning_table)); - mv2_gather_tuning_table mv2_tmp_gather_thresholds_table[]={ - {16, - 2,{{0, 524288, &MPIR_Gather_MV2_Direct}, - {524288, -1, &MPIR_Gather_intra}}, - 1,{{0, -1, &MPIR_Gather_MV2_Direct}}}, - {32, - 3,{{0, 16384, &MPIR_Gather_MV2_Direct}, - {16384, 131072, &MPIR_Gather_intra}, - {131072, -1, &MPIR_Gather_MV2_two_level_Direct}}, - 1,{{0, -1, &MPIR_Gather_intra}}}, - {64, - 3,{{0, 256, &MPIR_Gather_MV2_two_level_Direct}, - {256, 16384, &MPIR_Gather_MV2_Direct}, - {256, -1, &MPIR_Gather_MV2_two_level_Direct}}, - 1,{{0, -1, &MPIR_Gather_intra}}}, - {128, - 3,{{0, 512, &MPIR_Gather_MV2_two_level_Direct}, - {512, 16384, &MPIR_Gather_MV2_Direct}, - {16384, -1, &MPIR_Gather_MV2_two_level_Direct}}, - 1,{{0, -1, &MPIR_Gather_intra}}}, - {256, - 3,{{0, 512, &MPIR_Gather_MV2_two_level_Direct}, - {512, 16384, &MPIR_Gather_MV2_Direct}, - {16384, -1, &MPIR_Gather_MV2_two_level_Direct}}, - 1,{{0, -1, &MPIR_Gather_intra}}}, - {512, - 3,{{0, 512, &MPIR_Gather_MV2_two_level_Direct}, - {512, 16384, &MPIR_Gather_MV2_Direct}, - {8196, -1, &MPIR_Gather_MV2_two_level_Direct}}, - 1,{{0, -1, &MPIR_Gather_intra}}}, - {1024, - 3,{{0, 512, &MPIR_Gather_MV2_two_level_Direct}, - {512, 16384, &MPIR_Gather_MV2_Direct}, - {8196, -1, &MPIR_Gather_MV2_two_level_Direct}}, - 1,{{0, -1, &MPIR_Gather_intra}}}, - }; - - memcpy(mv2_gather_thresholds_table, mv2_tmp_gather_thresholds_table, - mv2_size_gather_tuning_table * sizeof (mv2_gather_tuning_table)); - -} - int smpi_coll_tuned_gather_mvapich2(void *sendbuf, int sendcnt, @@ -774,72 +269,6 @@ int smpi_coll_tuned_gather_mvapich2(void *sendbuf, } - -static void init_mv2_allgatherv_tables_stampede(){ - mv2_size_allgatherv_tuning_table = 6; - mv2_allgatherv_thresholds_table = malloc(mv2_size_allgatherv_tuning_table * - sizeof (mv2_allgatherv_tuning_table)); - mv2_allgatherv_tuning_table mv2_tmp_allgatherv_thresholds_table[] = { - { - 16, - 2, - { - {0, 512, &MPIR_Allgatherv_Rec_Doubling_MV2}, - {512, -1, &MPIR_Allgatherv_Ring_MV2}, - }, - }, - { - 32, - 2, - { - {0, 512, &MPIR_Allgatherv_Rec_Doubling_MV2}, - {512, -1, &MPIR_Allgatherv_Ring_MV2}, - }, - }, - { - 64, - 2, - { - {0, 256, &MPIR_Allgatherv_Rec_Doubling_MV2}, - {256, -1, &MPIR_Allgatherv_Ring_MV2}, - }, - }, - { - 128, - 2, - { - {0, 256, &MPIR_Allgatherv_Rec_Doubling_MV2}, - {256, -1, &MPIR_Allgatherv_Ring_MV2}, - }, - }, - { - 256, - 2, - { - {0, 256, &MPIR_Allgatherv_Rec_Doubling_MV2}, - {256, -1, &MPIR_Allgatherv_Ring_MV2}, - }, - }, - { - 512, - 2, - { - {0, 256, &MPIR_Allgatherv_Rec_Doubling_MV2}, - {256, -1, &MPIR_Allgatherv_Ring_MV2}, - }, - }, - - }; - memcpy(mv2_allgatherv_thresholds_table, mv2_tmp_allgatherv_thresholds_table, - mv2_size_allgatherv_tuning_table * sizeof (mv2_allgatherv_tuning_table)); -} - - - - - - - int smpi_coll_tuned_allgatherv_mvapich2(void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int *recvcounts, int *displs, MPI_Datatype recvtype, MPI_Comm comm ) @@ -905,147 +334,6 @@ int smpi_coll_tuned_allgatherv_mvapich2(void *sendbuf, int sendcount, MPI_Dataty } -static void init_mv2_allreduce_tables_stampede(){ -mv2_size_allreduce_tuning_table = 8; - mv2_allreduce_thresholds_table = malloc(mv2_size_allreduce_tuning_table * - sizeof (mv2_allreduce_tuning_table)); - mv2_allreduce_tuning_table mv2_tmp_allreduce_thresholds_table[] = { - { - 16, - 0, - {1, 0}, - 2, - { - {0, 1024, &MPIR_Allreduce_pt2pt_rd_MV2}, - {1024, -1, &MPIR_Allreduce_pt2pt_rs_MV2}, - }, - 2, - { - {0, 1024, &MPIR_Allreduce_reduce_shmem_MV2}, - {1024, -1, &MPIR_Allreduce_reduce_p2p_MV2}, - }, - }, - { - 32, - 0, - {1, 1, 0}, - 3, - { - {0, 1024, &MPIR_Allreduce_pt2pt_rd_MV2}, - {1024, 16384, &MPIR_Allreduce_pt2pt_rd_MV2}, - {16384, -1, &MPIR_Allreduce_pt2pt_rs_MV2}, - }, - 2, - { - {0, 1024, &MPIR_Allreduce_reduce_shmem_MV2}, - {1024, 16384, &MPIR_Allreduce_reduce_p2p_MV2}, - }, - }, - { - 64, - 0, - {1, 1, 0}, - 3, - { - {0, 512, &MPIR_Allreduce_pt2pt_rd_MV2}, - {512, 16384, &MPIR_Allreduce_pt2pt_rd_MV2}, - {16384, -1, &MPIR_Allreduce_pt2pt_rs_MV2}, - }, - 2, - { - {0, 512, &MPIR_Allreduce_reduce_shmem_MV2}, - {512, 16384, &MPIR_Allreduce_reduce_p2p_MV2}, - }, - }, - { - 128, - 0, - {1, 1, 0}, - 3, - { - {0, 512, &MPIR_Allreduce_pt2pt_rd_MV2}, - {512, 16384, &MPIR_Allreduce_pt2pt_rd_MV2}, - {16384, -1, &MPIR_Allreduce_pt2pt_rs_MV2}, - }, - 2, - { - {0, 512, &MPIR_Allreduce_reduce_shmem_MV2}, - {512, 16384, &MPIR_Allreduce_reduce_p2p_MV2}, - }, - }, - { - 256, - 0, - {1, 1, 0}, - 3, - { - {0, 512, &MPIR_Allreduce_pt2pt_rd_MV2}, - {512, 16384, &MPIR_Allreduce_pt2pt_rd_MV2}, - {16384, -1, &MPIR_Allreduce_pt2pt_rs_MV2}, - }, - 2, - { - {0, 512, &MPIR_Allreduce_reduce_shmem_MV2}, - {512, -1, &MPIR_Allreduce_reduce_p2p_MV2}, - }, - }, - { - 512, - 0, - {1, 1, 0}, - 3, - { - {0, 512, &MPIR_Allreduce_pt2pt_rd_MV2}, - {512, 16384, &MPIR_Allreduce_pt2pt_rd_MV2}, - {16384, -1, &MPIR_Allreduce_pt2pt_rs_MV2}, - }, - 2, - { - {0, 512, &MPIR_Allreduce_reduce_shmem_MV2}, - {512, 16384, &MPIR_Allreduce_reduce_p2p_MV2}, - }, - }, - { - 1024, - 0, - {1, 1, 1, 0}, - 4, - { - {0, 512, &MPIR_Allreduce_pt2pt_rd_MV2}, - {512, 8192, &MPIR_Allreduce_pt2pt_rd_MV2}, - {8192, 65536, &MPIR_Allreduce_pt2pt_rs_MV2}, - {65536, -1, &MPIR_Allreduce_pt2pt_rs_MV2}, - }, - 2, - { - {0, 512, &MPIR_Allreduce_reduce_shmem_MV2}, - {512, -1, &MPIR_Allreduce_reduce_p2p_MV2}, - }, - }, - { - 2048, - 0, - {1, 1, 1, 0}, - 4, - { - {0, 64, &MPIR_Allreduce_pt2pt_rd_MV2}, - {64, 512, &MPIR_Allreduce_reduce_p2p_MV2}, - {512, 4096, &MPIR_Allreduce_mcst_reduce_two_level_helper_MV2}, - {4096, 16384, &MPIR_Allreduce_pt2pt_rs_MV2}, - {16384, -1, &MPIR_Allreduce_pt2pt_rs_MV2}, - }, - 2, - { - {0, 512, &MPIR_Allreduce_reduce_shmem_MV2}, - {512, -1, &MPIR_Allreduce_reduce_p2p_MV2}, - }, - }, - - }; - memcpy(mv2_allreduce_thresholds_table, mv2_tmp_allreduce_thresholds_table, - mv2_size_allreduce_tuning_table * sizeof (mv2_allreduce_tuning_table)); -} - int smpi_coll_tuned_allreduce_mvapich2(void *sendbuf, void *recvbuf, @@ -1179,7 +467,7 @@ if (sbuf == MPI_IN_PLACE) { rbuf, rcounts, rdisps,rdtype, comm); } else /* For starters, just keep the original algorithm. */ - return smpi_coll_tuned_alltoallv_pair(sbuf, scounts, sdisps, sdtype, + return smpi_coll_tuned_alltoallv_ring(sbuf, scounts, sdisps, sdtype, rbuf, rcounts, rdisps,rdtype, comm); } @@ -1191,215 +479,6 @@ int smpi_coll_tuned_barrier_mvapich2(MPI_Comm comm) } -/* -static void init_mv2_bcast_tables_stampede(){ - //Stampede, - mv2_size_bcast_tuning_table=8; - mv2_bcast_thresholds_table = malloc(mv2_size_bcast_tuning_table * - sizeof (mv2_bcast_tuning_table)); - - mv2_bcast_tuning_table mv2_tmp_bcast_thresholds_table[]={ - { - 16, - 8192, 4, 4, - {1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1}, - 11, - { - {0, 8, &MPIR_Pipelined_Bcast_Zcpy_MV2, 2}, - {8, 16, &MPIR_Pipelined_Bcast_Zcpy_MV2, 4}, - {16, 1024, &MPIR_Pipelined_Bcast_Zcpy_MV2, 2}, - {1024, 8192, &MPIR_Pipelined_Bcast_Zcpy_MV2, 4}, - {8192, 16384, &MPIR_Bcast_binomial_MV2, -1}, - {16384, 32768, &MPIR_Pipelined_Bcast_Zcpy_MV2, 4}, - {32768, 65536, &MPIR_Pipelined_Bcast_Zcpy_MV2, 2}, - {65536, 131072, &MPIR_Bcast_scatter_ring_allgather_shm_MV2, -1}, - {131072, 262144, &MPIR_Bcast_scatter_ring_allgather_MV2, -1}, - {262144, 524288, &MPIR_Bcast_scatter_doubling_allgather_MV2, -1}, - {524288, -1, &MPIR_Bcast_scatter_ring_allgather_MV2, -1} - }, - 11, - { - {0, 8, &MPIR_Shmem_Bcast_MV2, 2}, - {8, 16, &MPIR_Shmem_Bcast_MV2, 4}, - {16, 1024, &MPIR_Shmem_Bcast_MV2, 2}, - {1024, 8192, &MPIR_Shmem_Bcast_MV2, 4}, - {8192, 16384, &MPIR_Shmem_Bcast_MV2, -1}, - {16384, 32768, &MPIR_Shmem_Bcast_MV2, 4}, - {32768, 65536, &MPIR_Shmem_Bcast_MV2, 2}, - {65536, 131072, &MPIR_Shmem_Bcast_MV2, -1}, - {131072, 262144, &MPIR_Shmem_Bcast_MV2, -1}, - {262144, 524288, &MPIR_Shmem_Bcast_MV2, -1}, - {524288, -1, &MPIR_Shmem_Bcast_MV2, -1} - } - }, - { - 32, - 8192, 4, 4, - {1, 1, 1, 1, 1, 1, 1, 1}, - 8, - { - {0, 128, &MPIR_Pipelined_Bcast_Zcpy_MV2, 2}, - {128, 256, &MPIR_Pipelined_Bcast_Zcpy_MV2, 4}, - {256, 32768, &MPIR_Pipelined_Bcast_Zcpy_MV2, 2}, - {32768, 65536, &MPIR_Pipelined_Bcast_Zcpy_MV2, 4}, - {65536, 131072, &MPIR_Pipelined_Bcast_Zcpy_MV2, 2}, - {131072, 262144, &MPIR_Pipelined_Bcast_Zcpy_MV2, 8}, - {262144, 524288, &MPIR_Pipelined_Bcast_Zcpy_MV2, 2}, - {524288, -1, &MPIR_Pipelined_Bcast_Zcpy_MV2, 8} - }, - 8, - { - {0, 128, &MPIR_Shmem_Bcast_MV2, 2}, - {128, 256, &MPIR_Shmem_Bcast_MV2, 4}, - {256, 32768, &MPIR_Shmem_Bcast_MV2, 2}, - {32768, 65536, &MPIR_Shmem_Bcast_MV2, 4}, - {65536, 131072, &MPIR_Shmem_Bcast_MV2, 2}, - {131072, 262144, &MPIR_Shmem_Bcast_MV2, 8}, - {262144, 524288, &MPIR_Shmem_Bcast_MV2, 2}, - {524288, -1, &MPIR_Shmem_Bcast_MV2, 8} - } - }, - { - 64, - 8192, 4, 4, - {1, 1, 1, 1, 1, 1, 1, 1, 1}, - 9, - { - {0, 2, &MPIR_Pipelined_Bcast_Zcpy_MV2, 4}, - {2, 4, &MPIR_Pipelined_Bcast_Zcpy_MV2, 8}, - {4, 16, &MPIR_Pipelined_Bcast_Zcpy_MV2, 4}, - {16, 32, &MPIR_Pipelined_Bcast_Zcpy_MV2, 8}, - {32, 128, &MPIR_Pipelined_Bcast_Zcpy_MV2, 4}, - {128, 256, &MPIR_Pipelined_Bcast_Zcpy_MV2, 8}, - {256, 4096, &MPIR_Pipelined_Bcast_Zcpy_MV2, 4}, - {4096, 32768, &MPIR_Pipelined_Bcast_Zcpy_MV2, 8}, - {32768, -1, &MPIR_Pipelined_Bcast_Zcpy_MV2, 2} - }, - 9, - { - {0, 2, &MPIR_Shmem_Bcast_MV2, 4}, - {2, 4, &MPIR_Shmem_Bcast_MV2, 8}, - {4, 16, &MPIR_Shmem_Bcast_MV2, 4}, - {16, 32, &MPIR_Shmem_Bcast_MV2, 8}, - {32, 128, &MPIR_Shmem_Bcast_MV2, 4}, - {128, 256, &MPIR_Shmem_Bcast_MV2, 8}, - {256, 4096, &MPIR_Shmem_Bcast_MV2, 4}, - {4096, 32768, &MPIR_Shmem_Bcast_MV2, 8}, - {32768, -1, &MPIR_Shmem_Bcast_MV2, 2} - } - }, - { - 128, - 8192, 4, 4, - {1, 1, 1, 0}, - 4, - { - {0, 8192, &MPIR_Pipelined_Bcast_Zcpy_MV2, 8}, - {8192, 16384, &MPIR_Pipelined_Bcast_Zcpy_MV2, 4}, - {16384, 524288, &MPIR_Pipelined_Bcast_Zcpy_MV2, 2}, - {524288, -1, &MPIR_Bcast_scatter_ring_allgather_MV2, -1} - }, - 4, - { - {0, 8192, &MPIR_Shmem_Bcast_MV2, 8}, - {8192, 16384, &MPIR_Shmem_Bcast_MV2, 4}, - {16384, 524288, &MPIR_Shmem_Bcast_MV2, 2}, - {524288, -1, NULL, -1} - } - }, - { - 256, - 8192, 4, 4, - {1, 1, 1, 1, 1}, - 5, - { - {0, 16384, &MPIR_Pipelined_Bcast_Zcpy_MV2, 4}, - {16384, 131072, &MPIR_Pipelined_Bcast_Zcpy_MV2, 2}, - {131072, 262144, &MPIR_Bcast_scatter_ring_allgather_shm_MV2, -1}, - {262144, 524288, &MPIR_Pipelined_Bcast_Zcpy_MV2, 2}, - {524288, -1, &MPIR_Bcast_scatter_ring_allgather_shm_MV2, -1} - }, - 5, - { - {0, 16384, &MPIR_Shmem_Bcast_MV2, 4}, - {16384, 131072, &MPIR_Shmem_Bcast_MV2, 2}, - {131072, 262144, &MPIR_Shmem_Bcast_MV2, -1}, - {262144, 524288, &MPIR_Shmem_Bcast_MV2, 2}, - {524288, -1, &MPIR_Shmem_Bcast_MV2, -1} - } - }, - { - 512, - 8192, 4, 4, - {1, 1, 1, 1, 1}, - 5, - { - {0, 4096, &MPIR_Pipelined_Bcast_Zcpy_MV2, 8}, - {4096, 16384, &MPIR_Pipelined_Bcast_Zcpy_MV2, 4}, - {16384, 131072, &MPIR_Pipelined_Bcast_Zcpy_MV2, 2}, - {131072, 262144, &MPIR_Pipelined_Bcast_MV2, -1}, - {262144, -1, &MPIR_Bcast_scatter_ring_allgather_shm_MV2, -1} - }, - 5, - { - {0, 4096, &MPIR_Shmem_Bcast_MV2, 8}, - {4096, 16384, &MPIR_Shmem_Bcast_MV2, 4}, - {16384, 131072, &MPIR_Shmem_Bcast_MV2, 2}, - {131072, 262144, &MPIR_Shmem_Bcast_MV2, -1}, - {262144, -1, &MPIR_Shmem_Bcast_MV2, -1} - } - }, - { - 1024, - 8192, 4, 4, - {1, 1, 1, 1, 1}, - 5, - { - {0, 8192, &MPIR_Pipelined_Bcast_Zcpy_MV2, 8}, - {8192, 16384, &MPIR_Pipelined_Bcast_Zcpy_MV2, 4}, - {16384, 65536, &MPIR_Pipelined_Bcast_Zcpy_MV2, 2}, - {65536, 524288, &MPIR_Pipelined_Bcast_MV2, -1}, - {524288, -1, &MPIR_Bcast_scatter_ring_allgather_shm_MV2, -1} - }, - 5, - { - {0, 8192, &MPIR_Shmem_Bcast_MV2, 8}, - {8192, 16384, &MPIR_Shmem_Bcast_MV2, 4}, - {16384, 65536, &MPIR_Shmem_Bcast_MV2, 2}, - {65536, 524288, &MPIR_Shmem_Bcast_MV2, -1}, - {524288, -1, &MPIR_Shmem_Bcast_MV2, -1} - } - }, - { - 2048, - 8192, 4, 4, - {1, 1, 1, 1, 1, 1, 1}, - 7, - { - {0, 16, &MPIR_Pipelined_Bcast_Zcpy_MV2, 8}, - {16, 32, &MPIR_Pipelined_Bcast_Zcpy_MV2, 4}, - {32, 4096, &MPIR_Pipelined_Bcast_Zcpy_MV2, 8}, - {4096, 16384, &MPIR_Pipelined_Bcast_Zcpy_MV2, 4}, - {16384, 32768, &MPIR_Pipelined_Bcast_Zcpy_MV2, 2}, - {32768, 524288, &MPIR_Pipelined_Bcast_MV2, -1}, - {524288, -1, &MPIR_Bcast_scatter_ring_allgather_shm_MV2, -1} - }, - 7, - { - {0, 16, &MPIR_Shmem_Bcast_MV2, 8}, - {16, 32, &MPIR_Shmem_Bcast_MV2, 4}, - {32, 4096, &MPIR_Shmem_Bcast_MV2, 8}, - {4096, 16384, &MPIR_Shmem_Bcast_MV2, 4}, - {16384, 32768, &MPIR_Shmem_Bcast_MV2, 2}, - {32768, 524288, &MPIR_Shmem_Bcast_MV2, -1}, - {524288, -1, &MPIR_Shmem_Bcast_MV2, -1} - } - } - }; - - memcpy(mv2_bcast_thresholds_table, mv2_tmp_bcast_thresholds_table, - mv2_size_bcast_tuning_table * sizeof (mv2_bcast_tuning_table)); -}*/ int smpi_coll_tuned_bcast_mvapich2(void *buffer, @@ -1413,199 +492,6 @@ int smpi_coll_tuned_bcast_mvapich2(void *buffer, } -static void init_mv2_reduce_tables_stampede(){ - /*Stampede*/ - mv2_size_reduce_tuning_table = 8; - mv2_reduce_thresholds_table = malloc(mv2_size_reduce_tuning_table * - sizeof (mv2_reduce_tuning_table)); - mv2_reduce_tuning_table mv2_tmp_reduce_thresholds_table[] = { - { - 16, - 4, - 4, - {1, 0, 0}, - 3, - { - {0, 262144, &MPIR_Reduce_inter_knomial_wrapper_MV2}, - {262144, 1048576, &MPIR_Reduce_binomial_MV2}, - {1048576, -1, &MPIR_Reduce_redscat_gather_MV2}, - }, - 2, - { - {0, 65536, &MPIR_Reduce_shmem_MV2}, - {65536,-1, &MPIR_Reduce_binomial_MV2}, - }, - }, - { - 32, - 4, - 4, - {1, 1, 1, 1, 0, 0, 0}, - 7, - { - {0, 8192, &MPIR_Reduce_inter_knomial_wrapper_MV2}, - {8192, 16384, &MPIR_Reduce_inter_knomial_wrapper_MV2}, - {16384, 32768, &MPIR_Reduce_inter_knomial_wrapper_MV2}, - {32768, 65536, &MPIR_Reduce_binomial_MV2}, - {65536, 262144, &MPIR_Reduce_inter_knomial_wrapper_MV2}, - {262144, 1048576, &MPIR_Reduce_binomial_MV2}, - {1048576, -1, &MPIR_Reduce_redscat_gather_MV2}, - }, - 6, - { - {0, 8192, &MPIR_Reduce_shmem_MV2}, - {8192, 16384, &MPIR_Reduce_intra_knomial_wrapper_MV2}, - {16384, 32768, &MPIR_Reduce_shmem_MV2}, - {32768, 65536, &MPIR_Reduce_shmem_MV2}, - {65536, 262144, &MPIR_Reduce_shmem_MV2}, - {262144,-1, &MPIR_Reduce_binomial_MV2}, - }, - }, - { - 64, - 4, - 4, - {1, 1, 1, 1, 0}, - 5, - { - {0, 8192, &MPIR_Reduce_inter_knomial_wrapper_MV2}, - {8192, 16384, &MPIR_Reduce_inter_knomial_wrapper_MV2}, - {16384, 65536, &MPIR_Reduce_binomial_MV2}, - {65536, 262144, &MPIR_Reduce_inter_knomial_wrapper_MV2}, - {262144, -1, &MPIR_Reduce_redscat_gather_MV2}, - }, - 5, - { - {0, 8192, &MPIR_Reduce_shmem_MV2}, - {8192, 16384, &MPIR_Reduce_intra_knomial_wrapper_MV2}, - {16384, 65536, &MPIR_Reduce_shmem_MV2}, - {65536, 262144, &MPIR_Reduce_intra_knomial_wrapper_MV2}, - {262144, -1, &MPIR_Reduce_binomial_MV2}, - }, - }, - { - 128, - 4, - 4, - {1, 0, 1, 0, 1, 0}, - 6, - { - {0, 8192, &MPIR_Reduce_inter_knomial_wrapper_MV2}, - {8192, 16384, &MPIR_Reduce_inter_knomial_wrapper_MV2}, - {16384, 65536, &MPIR_Reduce_binomial_MV2}, - {65536, 262144, &MPIR_Reduce_inter_knomial_wrapper_MV2}, - {262144, 1048576, &MPIR_Reduce_binomial_MV2}, - {1048576, -1, &MPIR_Reduce_redscat_gather_MV2}, - }, - 5, - { - {0, 8192, &MPIR_Reduce_shmem_MV2}, - {8192, 16384, &MPIR_Reduce_intra_knomial_wrapper_MV2}, - {16384, 65536, &MPIR_Reduce_shmem_MV2}, - {65536, 262144, &MPIR_Reduce_intra_knomial_wrapper_MV2}, - {262144, -1, &MPIR_Reduce_binomial_MV2}, - }, - }, - { - 256, - 4, - 4, - {1, 1, 1, 0, 1, 1, 0}, - 7, - { - {0, 8192, &MPIR_Reduce_inter_knomial_wrapper_MV2}, - {8192, 16384, &MPIR_Reduce_inter_knomial_wrapper_MV2}, - {16384, 32768, &MPIR_Reduce_binomial_MV2}, - {32768, 65536, &MPIR_Reduce_binomial_MV2}, - {65536, 262144, &MPIR_Reduce_binomial_MV2}, - {262144, 1048576, &MPIR_Reduce_binomial_MV2}, - {1048576, -1, &MPIR_Reduce_redscat_gather_MV2}, - }, - 6, - { - {0, 8192, &MPIR_Reduce_shmem_MV2}, - {8192, 16384, &MPIR_Reduce_intra_knomial_wrapper_MV2}, - {16384, 32768, &MPIR_Reduce_shmem_MV2}, - {32768, 65536, &MPIR_Reduce_shmem_MV2}, - {65536, 262144, &MPIR_Reduce_intra_knomial_wrapper_MV2}, - {262144, -1, &MPIR_Reduce_binomial_MV2}, - }, - }, - { - 512, - 4, - 4, - {1, 0, 1, 1, 1, 0}, - 6, - { - {0, 8192, &MPIR_Reduce_inter_knomial_wrapper_MV2}, - {8192, 16384, &MPIR_Reduce_inter_knomial_wrapper_MV2}, - {16384, 65536, &MPIR_Reduce_binomial_MV2}, - {65536, 262144, &MPIR_Reduce_binomial_MV2}, - {262144, 1048576, &MPIR_Reduce_binomial_MV2}, - {1048576, -1, &MPIR_Reduce_redscat_gather_MV2}, - }, - 5, - { - {0, 8192, &MPIR_Reduce_shmem_MV2}, - {8192, 16384, &MPIR_Reduce_intra_knomial_wrapper_MV2}, - {16384, 65536, &MPIR_Reduce_shmem_MV2}, - {65536, 262144, &MPIR_Reduce_intra_knomial_wrapper_MV2}, - {262144, -1, &MPIR_Reduce_binomial_MV2}, - }, - }, - { - 1024, - 4, - 4, - {1, 0, 1, 1, 1}, - 5, - { - {0, 8192, &MPIR_Reduce_inter_knomial_wrapper_MV2}, - {8192, 16384, &MPIR_Reduce_inter_knomial_wrapper_MV2}, - {16384, 65536, &MPIR_Reduce_binomial_MV2}, - {65536, 262144, &MPIR_Reduce_binomial_MV2}, - {262144, -1, &MPIR_Reduce_binomial_MV2}, - }, - 5, - { - {0, 8192, &MPIR_Reduce_shmem_MV2}, - {8192, 16384, &MPIR_Reduce_intra_knomial_wrapper_MV2}, - {16384, 65536, &MPIR_Reduce_shmem_MV2}, - {65536, 262144, &MPIR_Reduce_intra_knomial_wrapper_MV2}, - {262144, -1, &MPIR_Reduce_binomial_MV2}, - }, - }, - { - 2048, - 4, - 4, - {1, 0, 1, 1, 1,1}, - 6, - { - {0, 2048, &MPIR_Reduce_inter_knomial_wrapper_MV2}, - {2048, 4096, &MPIR_Reduce_inter_knomial_wrapper_MV2}, - {4096, 16384, &MPIR_Reduce_binomial_MV2}, - {16384, 65536, &MPIR_Reduce_binomial_MV2}, - {65536, 131072, &MPIR_Reduce_binomial_MV2}, - {131072, -1, &MPIR_Reduce_binomial_MV2}, - }, - 6, - { - {0, 2048, &MPIR_Reduce_shmem_MV2}, - {2048, 4096, &MPIR_Reduce_shmem_MV2}, - {4096, 16384, &MPIR_Reduce_shmem_MV2}, - {16384, 65536, &MPIR_Reduce_intra_knomial_wrapper_MV2}, - {65536, 131072, &MPIR_Reduce_binomial_MV2}, - {131072, -1, &MPIR_Reduce_shmem_MV2}, - }, - }, - - }; - memcpy(mv2_reduce_thresholds_table, mv2_tmp_reduce_thresholds_table, - mv2_size_reduce_tuning_table * sizeof (mv2_reduce_tuning_table)); -} - int smpi_coll_tuned_reduce_mvapich2( void *sendbuf, @@ -1724,69 +610,6 @@ int smpi_coll_tuned_reduce_mvapich2( void *sendbuf, } - -static void init_mv2_reduce_scatter_tables_stampede(){ - mv2_size_red_scat_tuning_table = 6; - mv2_red_scat_thresholds_table = malloc(mv2_size_red_scat_tuning_table * - sizeof (mv2_red_scat_tuning_table)); - mv2_red_scat_tuning_table mv2_tmp_red_scat_thresholds_table[] = { - { - 16, - 3, - { - {0, 64, &MPIR_Reduce_Scatter_Basic_MV2}, - {64, 65536, &MPIR_Reduce_scatter_Rec_Halving_MV2}, - {65536, -1, &MPIR_Reduce_scatter_Pair_Wise_MV2}, - }, - }, - { - 32, - 3, - { - {0, 64, &MPIR_Reduce_Scatter_Basic_MV2}, - {64, 131072, &MPIR_Reduce_scatter_Rec_Halving_MV2}, - {131072, -1, &MPIR_Reduce_scatter_Pair_Wise_MV2}, - }, - }, - { - 64, - 3, - { - {0, 1024, &MPIR_Reduce_Scatter_Basic_MV2}, - {1024, 262144, &MPIR_Reduce_scatter_Rec_Halving_MV2}, - {262144, -1, &MPIR_Reduce_scatter_Pair_Wise_MV2}, - }, - }, - { - 128, - 2, - { - {0, 128, &MPIR_Reduce_Scatter_Basic_MV2}, - {128, -1, &MPIR_Reduce_scatter_Rec_Halving_MV2}, - }, - }, - { - 256, - 2, - { - {0, 128, &MPIR_Reduce_Scatter_Basic_MV2}, - {128, -1, &MPIR_Reduce_scatter_Rec_Halving_MV2}, - }, - }, - { - 512, - 2, - { - {0, 256, &MPIR_Reduce_Scatter_Basic_MV2}, - {256, -1, &MPIR_Reduce_scatter_Rec_Halving_MV2}, - }, - }, - - }; - memcpy(mv2_red_scat_thresholds_table, mv2_tmp_red_scat_thresholds_table, - mv2_size_red_scat_tuning_table * sizeof (mv2_red_scat_tuning_table)); -} - int smpi_coll_tuned_reduce_scatter_mvapich2(void *sendbuf, void *recvbuf, int *recvcnts, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) @@ -1797,7 +620,7 @@ int smpi_coll_tuned_reduce_scatter_mvapich2(void *sendbuf, void *recvbuf, int *r int range = 0; int range_threshold = 0; int is_commutative = 0; - int *disps = malloc(comm_size * sizeof (int)); + int *disps = xbt_malloc(comm_size * sizeof (int)); if(mv2_red_scat_thresholds_table==NULL) init_mv2_reduce_scatter_tables_stampede(); @@ -1836,9 +659,24 @@ int smpi_coll_tuned_reduce_scatter_mvapich2(void *sendbuf, void *recvbuf, int *r recvcnts, datatype, op, comm); } else { - mpi_errno = MPIR_Reduce_scatter_non_comm_MV2(sendbuf, recvbuf, - recvcnts, datatype, - op, comm); + int is_block_regular = 1; + for (i = 0; i < (comm_size - 1); ++i) { + if (recvcnts[i] != recvcnts[i+1]) { + is_block_regular = 0; + break; + } + } + int pof2 = 1; + while (pof2 < comm_size) pof2 <<= 1; + if (pof2 == comm_size && is_block_regular) { + /* noncommutative, pof2 size, and block regular */ + mpi_errno = MPIR_Reduce_scatter_non_comm_MV2(sendbuf, recvbuf, + recvcnts, datatype, + op, comm); + } + mpi_errno = smpi_coll_tuned_reduce_scatter_mpich_rdb(sendbuf, recvbuf, + recvcnts, datatype, + op, comm); } return mpi_errno; @@ -1847,317 +685,6 @@ int smpi_coll_tuned_reduce_scatter_mvapich2(void *sendbuf, void *recvbuf, int *r -static void init_mv2_scatter_tables_stampede(){ -{ - int agg_table_sum = 0; - int i; - mv2_scatter_tuning_table **table_ptrs = NULL; - mv2_scatter_num_ppn_conf = 3; - mv2_scatter_thresholds_table - = malloc(sizeof(mv2_scatter_tuning_table *) - * mv2_scatter_num_ppn_conf); - table_ptrs = malloc(sizeof(mv2_scatter_tuning_table *) - * mv2_scatter_num_ppn_conf); - mv2_size_scatter_tuning_table = malloc(sizeof(int) * - mv2_scatter_num_ppn_conf); - mv2_scatter_table_ppn_conf - = malloc(mv2_scatter_num_ppn_conf * sizeof(int)); - mv2_scatter_table_ppn_conf[0] = 1; - mv2_size_scatter_tuning_table[0] = 6; - mv2_scatter_tuning_table mv2_tmp_scatter_thresholds_table_1ppn[] = { - {2, - 1, - { - {0, -1, &MPIR_Scatter_MV2_Binomial}, - }, - 1, - { - {0, -1, &MPIR_Scatter_MV2_Binomial}, - }, - }, - - {4, - 1, - { - {0, -1, &MPIR_Scatter_MV2_Direct}, - }, - 1, - { - {0, -1, &MPIR_Scatter_MV2_Direct}, - }, - }, - - {8, - 1, - { - {0, -1, &MPIR_Scatter_MV2_Direct}, - }, - 1, - { - {0, -1, &MPIR_Scatter_MV2_Direct}, - }, - }, - - {16, - 1, - { - {0, -1, &MPIR_Scatter_MV2_Direct}, - }, - 1, - { - {0, -1, &MPIR_Scatter_MV2_Direct}, - }, - }, - - {32, - 1, - { - {0, -1, &MPIR_Scatter_MV2_Direct}, - }, - 1, - { - {0, -1, &MPIR_Scatter_MV2_Direct}, - }, - }, - - {64, - 2, - { - {0, 32, &MPIR_Scatter_MV2_Binomial}, - {32, -1, &MPIR_Scatter_MV2_Direct}, - }, - 1, - { - {0, -1, &MPIR_Scatter_MV2_Binomial}, - }, - }, - }; - table_ptrs[0] = mv2_tmp_scatter_thresholds_table_1ppn; - mv2_scatter_table_ppn_conf[1] = 2; - mv2_size_scatter_tuning_table[1] = 6; - mv2_scatter_tuning_table mv2_tmp_scatter_thresholds_table_2ppn[] = { - {4, - 2, - { - {0, 4096, &MPIR_Scatter_MV2_Binomial}, - {4096, -1, &MPIR_Scatter_MV2_Direct}, - }, - 1, - { - {0, -1, &MPIR_Scatter_MV2_Direct}, - }, - }, - - {8, - 2, - { - {0, 512, &MPIR_Scatter_MV2_two_level_Direct}, - {512, -1, &MPIR_Scatter_MV2_Direct}, - }, - 1, - { - {0, -1, &MPIR_Scatter_MV2_Binomial}, - }, - }, - - {16, - 2, - { - {0, 2048, &MPIR_Scatter_MV2_two_level_Direct}, - {2048, -1, &MPIR_Scatter_MV2_Direct}, - }, - 1, - { - {0, -1, &MPIR_Scatter_MV2_Binomial}, - }, - }, - - {32, - 2, - { - {0, 2048, &MPIR_Scatter_MV2_two_level_Direct}, - {2048, -1, &MPIR_Scatter_MV2_Direct}, - }, - 1, - { - {0, -1, &MPIR_Scatter_MV2_Binomial}, - }, - }, - - {64, - 2, - { - {0, 8192, &MPIR_Scatter_MV2_two_level_Direct}, - {8192, -1, &MPIR_Scatter_MV2_Direct}, - }, - 1, - { - {0, -1, &MPIR_Scatter_MV2_Binomial}, - }, - }, - - {128, - 4, - { - {0, 16, &MPIR_Scatter_MV2_Binomial}, - {16, 128, &MPIR_Scatter_MV2_two_level_Binomial}, - {128, 16384, &MPIR_Scatter_MV2_two_level_Direct}, - {16384, -1, &MPIR_Scatter_MV2_Direct}, - }, - 1, - { - {0, 128, &MPIR_Scatter_MV2_Direct}, - {128, -1, &MPIR_Scatter_MV2_Binomial}, - }, - }, - }; - table_ptrs[1] = mv2_tmp_scatter_thresholds_table_2ppn; - mv2_scatter_table_ppn_conf[2] = 16; - mv2_size_scatter_tuning_table[2] = 8; - mv2_scatter_tuning_table mv2_tmp_scatter_thresholds_table_16ppn[] = { - { - 16, - 2, - { - {0, 256, &MPIR_Scatter_MV2_Binomial}, - {256, -1, &MPIR_Scatter_MV2_Direct}, - }, - 1, - { - { 0, -1, &MPIR_Scatter_MV2_Direct}, - }, - }, - - { - 32, - 2, - { - {0, 512, &MPIR_Scatter_MV2_Binomial}, - {512, -1, &MPIR_Scatter_MV2_Direct}, - }, - 1, - { - { 0, -1, &MPIR_Scatter_MV2_Direct}, - }, - }, - - { - 64, - 2, - { - {0, 1024, &MPIR_Scatter_MV2_two_level_Direct}, - {1024, -1, &MPIR_Scatter_MV2_Direct}, - }, - 1, - { - { 0, -1, &MPIR_Scatter_MV2_Direct}, - }, - }, - - { - 128, - 4, - { - {0, 16, &MPIR_Scatter_mcst_wrap_MV2}, - {0, 16, &MPIR_Scatter_MV2_two_level_Direct}, - {16, 2048, &MPIR_Scatter_MV2_two_level_Direct}, - {2048, -1, &MPIR_Scatter_MV2_Direct}, - }, - 1, - { - { 0, -1, &MPIR_Scatter_MV2_Direct}, - }, - }, - - { - 256, - 4, - { - {0, 16, &MPIR_Scatter_mcst_wrap_MV2}, - {0, 16, &MPIR_Scatter_MV2_two_level_Direct}, - {16, 2048, &MPIR_Scatter_MV2_two_level_Direct}, - {2048, -1, &MPIR_Scatter_MV2_Direct}, - }, - 1, - { - { 0, -1, &MPIR_Scatter_MV2_Direct}, - }, - }, - - { - 512, - 4, - { - {0, 16, &MPIR_Scatter_mcst_wrap_MV2}, - {16, 16, &MPIR_Scatter_MV2_two_level_Direct}, - {16, 4096, &MPIR_Scatter_MV2_two_level_Direct}, - {4096, -1, &MPIR_Scatter_MV2_Direct}, - }, - 1, - { - { 0, -1, &MPIR_Scatter_MV2_Binomial}, - }, - }, - { - 1024, - 5, - { - {0, 16, &MPIR_Scatter_mcst_wrap_MV2}, - {0, 16, &MPIR_Scatter_MV2_Binomial}, - {16, 32, &MPIR_Scatter_MV2_Binomial}, - {32, 4096, &MPIR_Scatter_MV2_two_level_Direct}, - {4096, -1, &MPIR_Scatter_MV2_Direct}, - }, - 1, - { - { 0, -1, &MPIR_Scatter_MV2_Binomial}, - }, - }, - { - 2048, - 7, - { - {0, 16, &MPIR_Scatter_mcst_wrap_MV2}, - {0, 16, &MPIR_Scatter_MV2_two_level_Binomial}, - {16, 128, &MPIR_Scatter_MV2_two_level_Binomial}, - {128, 1024, &MPIR_Scatter_MV2_two_level_Direct}, - {1024, 16384, &MPIR_Scatter_MV2_two_level_Direct}, - {16384, 65536, &MPIR_Scatter_MV2_Direct}, - {65536, -1, &MPIR_Scatter_MV2_two_level_Direct}, - }, - 6, - { - {0, 16, &MPIR_Scatter_MV2_Binomial}, - {16, 128, &MPIR_Scatter_MV2_Binomial}, - {128, 1024, &MPIR_Scatter_MV2_Binomial}, - {1024, 16384, &MPIR_Scatter_MV2_Direct}, - {16384, 65536, &MPIR_Scatter_MV2_Direct}, - {65536, -1, &MPIR_Scatter_MV2_Direct}, - }, - }, - }; - table_ptrs[2] = mv2_tmp_scatter_thresholds_table_16ppn; - agg_table_sum = 0; - for (i = 0; i < mv2_scatter_num_ppn_conf; i++) { - agg_table_sum += mv2_size_scatter_tuning_table[i]; - } - mv2_scatter_thresholds_table[0] = - malloc(agg_table_sum * sizeof (mv2_scatter_tuning_table)); - memcpy(mv2_scatter_thresholds_table[0], table_ptrs[0], - (sizeof(mv2_scatter_tuning_table) - * mv2_size_scatter_tuning_table[0])); - for (i = 1; i < mv2_scatter_num_ppn_conf; i++) { - mv2_scatter_thresholds_table[i] = - mv2_scatter_thresholds_table[i - 1] - + mv2_size_scatter_tuning_table[i - 1]; - memcpy(mv2_scatter_thresholds_table[i], table_ptrs[i], - (sizeof(mv2_scatter_tuning_table) - * mv2_size_scatter_tuning_table[i])); - } - free(table_ptrs); - } -} - int smpi_coll_tuned_scatter_mvapich2(void *sendbuf, int sendcnt, MPI_Datatype sendtype,