X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/9cca09ad88fcaeb49f2637de97df519c3920aaa4..2d37e348a09783cda723c7019640ee69de168324:/src/smpi/colls/smpi_mvapich2_selector_stampede.hpp diff --git a/src/smpi/colls/smpi_mvapich2_selector_stampede.hpp b/src/smpi/colls/smpi_mvapich2_selector_stampede.hpp index a922738a50..417bb1b58b 100644 --- a/src/smpi/colls/smpi_mvapich2_selector_stampede.hpp +++ b/src/smpi/colls/smpi_mvapich2_selector_stampede.hpp @@ -2,16 +2,21 @@ /* This is the tuning used by MVAPICH for Stampede platform based on (MV2_ARCH_INTEL_XEON_E5_2680_16, * MV2_HCA_MLX_CX_FDR) */ -/* Copyright (c) 2009-2017. The SimGrid Team. All rights reserved. */ +/* Copyright (c) 2009-2019. The SimGrid Team. All rights reserved. */ /* This program is free software; you can redistribute it and/or modify it * under the terms of the license (GNU LGPL) which comes with this package. */ /************ Alltoall variables and initializers */ +#ifndef SMPI_MVAPICH2_SELECTOR_STAMPEDE_HPP +#define SMPI_MVAPICH2_SELECTOR_STAMPEDE_HPP + +#include + #define MV2_MAX_NB_THRESHOLDS 32 -XBT_PUBLIC(void) smpi_coll_cleanup_mvapich2(void); +XBT_PUBLIC void smpi_coll_cleanup_mvapich2(void); struct mv2_alltoall_tuning_element { int min; @@ -45,18 +50,15 @@ mv2_alltoall_tuning_table** mv2_alltoall_thresholds_table = NULL; 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; if (simgrid::smpi::Colls::smpi_coll_cleanup_callback == NULL) simgrid::smpi::Colls::smpi_coll_cleanup_callback = &smpi_coll_cleanup_mvapich2; - mv2_alltoall_thresholds_table = static_cast( - xbt_malloc(sizeof(mv2_alltoall_tuning_table*) * mv2_alltoall_num_ppn_conf)); - table_ptrs = static_cast( - xbt_malloc(sizeof(mv2_alltoall_tuning_table*) * mv2_alltoall_num_ppn_conf)); - mv2_size_alltoall_tuning_table = static_cast(xbt_malloc(sizeof(int) * mv2_alltoall_num_ppn_conf)); - mv2_alltoall_table_ppn_conf = static_cast(xbt_malloc(mv2_alltoall_num_ppn_conf * sizeof(int))); + mv2_alltoall_thresholds_table = new mv2_alltoall_tuning_table*[mv2_alltoall_num_ppn_conf]; + table_ptrs = new mv2_alltoall_tuning_table*[mv2_alltoall_num_ppn_conf]; + mv2_size_alltoall_tuning_table = new int[mv2_alltoall_num_ppn_conf]; + mv2_alltoall_table_ppn_conf = new int[mv2_alltoall_num_ppn_conf]; mv2_alltoall_table_ppn_conf[0] = 1; mv2_size_alltoall_tuning_table[0] = 6; mv2_alltoall_tuning_table mv2_tmp_alltoall_thresholds_table_1ppn[] = { @@ -314,19 +316,16 @@ static void init_mv2_alltoall_tables_stampede() }; table_ptrs[2] = mv2_tmp_alltoall_thresholds_table_16ppn; agg_table_sum = 0; - for (i = 0; i < mv2_alltoall_num_ppn_conf; i++) { + for (int i = 0; i < mv2_alltoall_num_ppn_conf; i++) { agg_table_sum += mv2_size_alltoall_tuning_table[i]; } - mv2_alltoall_thresholds_table[0] = - static_cast(xbt_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[0] = new mv2_alltoall_tuning_table[agg_table_sum]; + std::copy_n(table_ptrs[0], mv2_size_alltoall_tuning_table[0], mv2_alltoall_thresholds_table[0]); + for (int 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])); + std::copy_n(table_ptrs[i], mv2_size_alltoall_tuning_table[i], mv2_alltoall_thresholds_table[i]); } - xbt_free(table_ptrs); + delete[] table_ptrs; } /************ Allgather variables and initializers */ @@ -366,19 +365,16 @@ static int MPIR_Allgather_RD_Allgather_Comm_MV2(void* sendbuf, int sendcount, MP static void init_mv2_allgather_tables_stampede() { - int i; int agg_table_sum = 0; if (simgrid::smpi::Colls::smpi_coll_cleanup_callback == NULL) simgrid::smpi::Colls::smpi_coll_cleanup_callback = &smpi_coll_cleanup_mvapich2; mv2_allgather_tuning_table** table_ptrs = NULL; mv2_allgather_num_ppn_conf = 3; - mv2_allgather_thresholds_table = static_cast( - xbt_malloc(sizeof(mv2_allgather_tuning_table*) * mv2_allgather_num_ppn_conf)); - table_ptrs = static_cast( - xbt_malloc(sizeof(mv2_allgather_tuning_table*) * mv2_allgather_num_ppn_conf)); - mv2_size_allgather_tuning_table = static_cast(xbt_malloc(sizeof(int) * mv2_allgather_num_ppn_conf)); - mv2_allgather_table_ppn_conf = static_cast(xbt_malloc(mv2_allgather_num_ppn_conf * sizeof(int))); + mv2_allgather_thresholds_table = new mv2_allgather_tuning_table*[mv2_allgather_num_ppn_conf]; + table_ptrs = new mv2_allgather_tuning_table*[mv2_allgather_num_ppn_conf]; + mv2_size_allgather_tuning_table = new int[mv2_allgather_num_ppn_conf]; + mv2_allgather_table_ppn_conf = new int[mv2_allgather_num_ppn_conf]; mv2_allgather_table_ppn_conf[0] = 1; mv2_size_allgather_tuning_table[0] = 6; mv2_allgather_tuning_table mv2_tmp_allgather_thresholds_table_1ppn[] = { @@ -550,19 +546,16 @@ static void init_mv2_allgather_tables_stampede() }; table_ptrs[2] = mv2_tmp_allgather_thresholds_table_16ppn; agg_table_sum = 0; - for (i = 0; i < mv2_allgather_num_ppn_conf; i++) { + for (int i = 0; i < mv2_allgather_num_ppn_conf; i++) { agg_table_sum += mv2_size_allgather_tuning_table[i]; } - mv2_allgather_thresholds_table[0] = - static_cast(xbt_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[0] = new mv2_allgather_tuning_table[agg_table_sum]; + std::copy_n(table_ptrs[0], mv2_size_allgather_tuning_table[0], mv2_allgather_thresholds_table[0]); + for (int 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])); + std::copy_n(table_ptrs[i], mv2_size_allgather_tuning_table[i], mv2_allgather_thresholds_table[i]); } - xbt_free(table_ptrs); + delete[] table_ptrs; } /************ Gather variables and initializers */ @@ -601,8 +594,7 @@ static void init_mv2_gather_tables_stampede() if (simgrid::smpi::Colls::smpi_coll_cleanup_callback == NULL) simgrid::smpi::Colls::smpi_coll_cleanup_callback = &smpi_coll_cleanup_mvapich2; mv2_size_gather_tuning_table = 7; - mv2_gather_thresholds_table = - static_cast(xbt_malloc(mv2_size_gather_tuning_table * sizeof(mv2_gather_tuning_table))); + mv2_gather_thresholds_table = new mv2_gather_tuning_table[mv2_size_gather_tuning_table]; mv2_gather_tuning_table mv2_tmp_gather_thresholds_table[] = { {16, 2, @@ -653,8 +645,7 @@ static void init_mv2_gather_tables_stampede() {{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)); + std::copy_n(mv2_tmp_gather_thresholds_table, mv2_size_gather_tuning_table, mv2_gather_thresholds_table); } /************ Allgatherv variables and initializers */ @@ -687,8 +678,7 @@ static void init_mv2_allgatherv_tables_stampede() if (simgrid::smpi::Colls::smpi_coll_cleanup_callback == NULL) simgrid::smpi::Colls::smpi_coll_cleanup_callback = &smpi_coll_cleanup_mvapich2; mv2_size_allgatherv_tuning_table = 6; - mv2_allgatherv_thresholds_table = static_cast( - xbt_malloc(mv2_size_allgatherv_tuning_table * sizeof(mv2_allgatherv_tuning_table))); + mv2_allgatherv_thresholds_table = new mv2_allgatherv_tuning_table[mv2_size_allgatherv_tuning_table]; mv2_allgatherv_tuning_table mv2_tmp_allgatherv_thresholds_table[] = { { 16, @@ -734,8 +724,7 @@ static void init_mv2_allgatherv_tables_stampede() }, }; - memcpy(mv2_allgatherv_thresholds_table, mv2_tmp_allgatherv_thresholds_table, - mv2_size_allgatherv_tuning_table * sizeof(mv2_allgatherv_tuning_table)); + std::copy_n(mv2_tmp_allgatherv_thresholds_table, mv2_size_allgatherv_tuning_table, mv2_allgatherv_thresholds_table); } /************ Allreduce variables and initializers */ @@ -801,8 +790,7 @@ static void init_mv2_allreduce_tables_stampede() if (simgrid::smpi::Colls::smpi_coll_cleanup_callback == NULL) simgrid::smpi::Colls::smpi_coll_cleanup_callback = &smpi_coll_cleanup_mvapich2; mv2_size_allreduce_tuning_table = 8; - mv2_allreduce_thresholds_table = static_cast( - xbt_malloc(mv2_size_allreduce_tuning_table * sizeof(mv2_allreduce_tuning_table))); + mv2_allreduce_thresholds_table = new mv2_allreduce_tuning_table[mv2_size_allreduce_tuning_table]; mv2_allreduce_tuning_table mv2_tmp_allreduce_thresholds_table[] = { { 16, @@ -927,8 +915,7 @@ static void init_mv2_allreduce_tables_stampede() }, }; - memcpy(mv2_allreduce_thresholds_table, mv2_tmp_allreduce_thresholds_table, - mv2_size_allreduce_tuning_table * sizeof(mv2_allreduce_tuning_table)); + std::copy_n(mv2_tmp_allreduce_thresholds_table, mv2_size_allreduce_tuning_table, mv2_allreduce_thresholds_table); } struct mv2_bcast_tuning_element { @@ -987,8 +974,7 @@ static void init_mv2_bcast_tables_stampede() if (simgrid::smpi::Colls::smpi_coll_cleanup_callback == NULL) simgrid::smpi::Colls::smpi_coll_cleanup_callback = &smpi_coll_cleanup_mvapich2; mv2_size_bcast_tuning_table = 8; - mv2_bcast_thresholds_table = - static_cast(xbt_malloc(mv2_size_bcast_tuning_table * sizeof(mv2_bcast_tuning_table))); + mv2_bcast_thresholds_table = new mv2_bcast_tuning_table[mv2_size_bcast_tuning_table]; mv2_bcast_tuning_table mv2_tmp_bcast_thresholds_table[] = { {16, @@ -1156,8 +1142,7 @@ static void init_mv2_bcast_tables_stampede() {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)); + std::copy_n(mv2_tmp_bcast_thresholds_table, mv2_size_bcast_tuning_table, mv2_bcast_thresholds_table); } /************ Reduce variables and initializers */ @@ -1205,8 +1190,7 @@ static void init_mv2_reduce_tables_stampede() simgrid::smpi::Colls::smpi_coll_cleanup_callback = &smpi_coll_cleanup_mvapich2; /*Stampede*/ mv2_size_reduce_tuning_table = 8; - mv2_reduce_thresholds_table = - static_cast(xbt_malloc(mv2_size_reduce_tuning_table * sizeof(mv2_reduce_tuning_table))); + mv2_reduce_thresholds_table = new mv2_reduce_tuning_table[mv2_size_reduce_tuning_table]; mv2_reduce_tuning_table mv2_tmp_reduce_thresholds_table[] = { { 16, @@ -1390,8 +1374,7 @@ static void init_mv2_reduce_tables_stampede() }, }; - memcpy(mv2_reduce_thresholds_table, mv2_tmp_reduce_thresholds_table, - mv2_size_reduce_tuning_table * sizeof(mv2_reduce_tuning_table)); + std::copy_n(mv2_tmp_reduce_thresholds_table, mv2_size_reduce_tuning_table, mv2_reduce_thresholds_table); } /************ Reduce scatter variables and initializers */ @@ -1431,8 +1414,7 @@ static void init_mv2_reduce_scatter_tables_stampede() if (simgrid::smpi::Colls::smpi_coll_cleanup_callback == NULL) simgrid::smpi::Colls::smpi_coll_cleanup_callback = &smpi_coll_cleanup_mvapich2; mv2_size_red_scat_tuning_table = 6; - mv2_red_scat_thresholds_table = static_cast( - xbt_malloc(mv2_size_red_scat_tuning_table * sizeof(mv2_red_scat_tuning_table))); + mv2_red_scat_thresholds_table = new mv2_red_scat_tuning_table[mv2_size_red_scat_tuning_table]; mv2_red_scat_tuning_table mv2_tmp_red_scat_thresholds_table[] = { { 16, @@ -1484,8 +1466,7 @@ static void init_mv2_reduce_scatter_tables_stampede() }, }; - memcpy(mv2_red_scat_thresholds_table, mv2_tmp_red_scat_thresholds_table, - mv2_size_red_scat_tuning_table * sizeof(mv2_red_scat_tuning_table)); + std::copy_n(mv2_tmp_red_scat_thresholds_table, mv2_size_red_scat_tuning_table, mv2_red_scat_thresholds_table); } /************ Scatter variables and initializers */ @@ -1535,15 +1516,12 @@ static void init_mv2_scatter_tables_stampede() simgrid::smpi::Colls::smpi_coll_cleanup_callback = &smpi_coll_cleanup_mvapich2; int agg_table_sum = 0; - int i; mv2_scatter_tuning_table** table_ptrs = NULL; mv2_scatter_num_ppn_conf = 3; - mv2_scatter_thresholds_table = - static_cast(xbt_malloc(sizeof(mv2_scatter_tuning_table*) * mv2_scatter_num_ppn_conf)); - table_ptrs = - static_cast(xbt_malloc(sizeof(mv2_scatter_tuning_table*) * mv2_scatter_num_ppn_conf)); - mv2_size_scatter_tuning_table = static_cast(xbt_malloc(sizeof(int) * mv2_scatter_num_ppn_conf)); - mv2_scatter_table_ppn_conf = static_cast(xbt_malloc(mv2_scatter_num_ppn_conf * sizeof(int))); + mv2_scatter_thresholds_table = new mv2_scatter_tuning_table*[mv2_scatter_num_ppn_conf]; + table_ptrs = new mv2_scatter_tuning_table*[mv2_scatter_num_ppn_conf]; + mv2_size_scatter_tuning_table = new int[mv2_scatter_num_ppn_conf]; + mv2_scatter_table_ppn_conf = new int[mv2_scatter_num_ppn_conf]; mv2_scatter_table_ppn_conf[0] = 1; mv2_size_scatter_tuning_table[0] = 6; mv2_scatter_tuning_table mv2_tmp_scatter_thresholds_table_1ppn[] = { @@ -1822,17 +1800,16 @@ static void init_mv2_scatter_tables_stampede() }; table_ptrs[2] = mv2_tmp_scatter_thresholds_table_16ppn; agg_table_sum = 0; - for (i = 0; i < mv2_scatter_num_ppn_conf; i++) { + for (int i = 0; i < mv2_scatter_num_ppn_conf; i++) { agg_table_sum += mv2_size_scatter_tuning_table[i]; } - mv2_scatter_thresholds_table[0] = - static_cast(xbt_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[0] = new mv2_scatter_tuning_table[agg_table_sum]; + std::copy_n(table_ptrs[0], mv2_size_scatter_tuning_table[0], mv2_scatter_thresholds_table[0]); + for (int 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])); + std::copy_n(table_ptrs[i], mv2_size_scatter_tuning_table[i], mv2_scatter_thresholds_table[i]); } - xbt_free(table_ptrs); + delete[] table_ptrs; } + +#endif