oops

[simgrid.git] / src / smpi / colls / smpi_mvapich2_selector_stampede.h

diff --git a/src/smpi/colls/smpi_mvapich2_selector_stampede.h b/src/smpi/colls/smpi_mvapich2_selector_stampede.h
index db05c60..6c5601e 100644 (file)
--- a/src/smpi/colls/smpi_mvapich2_selector_stampede.h
+++ b/src/smpi/colls/smpi_mvapich2_selector_stampede.h
@@ -10,6 +10,8 @@
 /************ Alltoall variables and initializers                        */
 
 #define MV2_MAX_NB_THRESHOLDS  32
+
+
 typedef struct {
   int min;
   int max;
@@ -37,7 +39,7 @@ mv2_alltoall_tuning_table **mv2_alltoall_thresholds_table = NULL;
 
 #define MPIR_Alltoall_bruck_MV2 smpi_coll_tuned_alltoall_bruck
 #define MPIR_Alltoall_RD_MV2 smpi_coll_tuned_alltoall_rdb
-#define MPIR_Alltoall_Scatter_dest_MV2 smpi_coll_tuned_alltoall_ring
+#define MPIR_Alltoall_Scatter_dest_MV2 smpi_coll_tuned_alltoall_mvapich2_scatter_dest
 #define MPIR_Alltoall_pairwise_MV2 smpi_coll_tuned_alltoall_pair
 #define MPIR_Alltoall_inplace_MV2 smpi_coll_tuned_alltoall_ring 
 
@@ -47,6 +49,8 @@ static void init_mv2_alltoall_tables_stampede(){
   int agg_table_sum = 0;
   mv2_alltoall_tuning_table **table_ptrs = NULL;
   mv2_alltoall_num_ppn_conf = 3;
+  if(smpi_coll_cleanup_callback==NULL)
+    smpi_coll_cleanup_callback=&smpi_coll_cleanup_mvapich2;
   mv2_alltoall_thresholds_table = xbt_malloc(sizeof(mv2_alltoall_tuning_table *)
       * mv2_alltoall_num_ppn_conf);
   table_ptrs = xbt_malloc(sizeof(mv2_alltoall_tuning_table *)
@@ -318,15 +322,27 @@ int mv2_allgather_num_ppn_conf = 1;
 int *mv2_size_allgather_tuning_table = NULL;
 mv2_allgather_tuning_table **mv2_allgather_thresholds_table = NULL;
 
+static int MPIR_Allgather_RD_Allgather_Comm_MV2( void *sendbuf,
+                                 int sendcount,
+                                 MPI_Datatype sendtype,
+                                 void *recvbuf,
+                                 int recvcount,
+                                 MPI_Datatype recvtype, MPI_Comm comm_ptr)
+{
+    return 0;
+}
+
 #define MPIR_Allgather_Bruck_MV2 smpi_coll_tuned_allgather_bruck
 #define MPIR_Allgather_RD_MV2 smpi_coll_tuned_allgather_rdb
-#define MPIR_Allgather_RD_Allgather_Comm_MV2 smpi_coll_tuned_allgather_rdb
 #define MPIR_Allgather_Ring_MV2 smpi_coll_tuned_allgather_ring
-
+#define MPIR_2lvl_Allgather_MV2 smpi_coll_tuned_allgather_mvapich2_smp
 
 static void init_mv2_allgather_tables_stampede(){
   int i;
   int agg_table_sum = 0;
+
+  if(smpi_coll_cleanup_callback==NULL)
+    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
@@ -576,12 +592,14 @@ MV2_Gather_function_ptr MV2_Gather_intra_node_function = NULL;
 
 
 #define MPIR_Gather_MV2_Direct smpi_coll_tuned_gather_ompi_basic_linear
-#define MPIR_Gather_MV2_two_level_Direct smpi_coll_tuned_gather_ompi_basic_linear
+#define MPIR_Gather_MV2_two_level_Direct smpi_coll_tuned_gather_mvapich2_two_level
 #define MPIR_Gather_intra smpi_coll_tuned_gather_mpich
 
 
 static void init_mv2_gather_tables_stampede(){
 
+  if(smpi_coll_cleanup_callback==NULL)
+    smpi_coll_cleanup_callback=&smpi_coll_cleanup_mvapich2;
   mv2_size_gather_tuning_table=7;
   mv2_gather_thresholds_table = xbt_malloc(mv2_size_gather_tuning_table*
       sizeof (mv2_gather_tuning_table));
@@ -667,6 +685,8 @@ mv2_allgatherv_tuning_table *mv2_allgatherv_thresholds_table = NULL;
 
 
 static void init_mv2_allgatherv_tables_stampede(){
+  if(smpi_coll_cleanup_callback==NULL)
+    smpi_coll_cleanup_callback=&smpi_coll_cleanup_mvapich2;
   mv2_size_allgatherv_tuning_table = 6;
   mv2_allgatherv_thresholds_table = xbt_malloc(mv2_size_allgatherv_tuning_table *
       sizeof (mv2_allgatherv_tuning_table));
@@ -809,10 +829,12 @@ static  int MPIR_Allreduce_reduce_shmem_MV2( void *sendbuf,
 
 #define MPIR_Allreduce_pt2pt_rd_MV2 smpi_coll_tuned_allreduce_rdb
 #define MPIR_Allreduce_pt2pt_rs_MV2 smpi_coll_tuned_allreduce_mvapich2_rs
-
+#define MPIR_Allreduce_two_level_MV2 smpi_coll_tuned_allreduce_mvapich2_two_level
 
 
 static void init_mv2_allreduce_tables_stampede(){
+  if(smpi_coll_cleanup_callback==NULL)
+    smpi_coll_cleanup_callback=&smpi_coll_cleanup_mvapich2;
   mv2_size_allreduce_tuning_table = 8;
   mv2_allreduce_thresholds_table = xbt_malloc(mv2_size_allreduce_tuning_table *
       sizeof (mv2_allreduce_tuning_table));
@@ -954,8 +976,8 @@ static void init_mv2_allreduce_tables_stampede(){
 }
 
 
-/*
-Bcast deactivated for now, defaults to mpich one
+
+
 typedef struct {
     int min;
     int max;
@@ -986,16 +1008,36 @@ int (*MV2_Bcast_function) (void *buffer, int count, MPI_Datatype datatype,
 int (*MV2_Bcast_intra_node_function) (void *buffer, int count, MPI_Datatype datatype,
                                       int root, MPI_Comm comm_ptr) = NULL;
 
+int zcpy_knomial_factor = 2;
+int mv2_pipelined_zcpy_knomial_factor = -1;
+int bcast_segment_size = 8192;
+int mv2_inter_node_knomial_factor = 4;
+int mv2_intra_node_knomial_factor = 4;
+#define mv2_bcast_two_level_system_size  64
+#define mv2_bcast_short_msg             16384
+#define mv2_bcast_large_msg            512*1024
+
+#define INTRA_NODE_ROOT 0
+
+#define MPIR_Pipelined_Bcast_Zcpy_MV2 smpi_coll_tuned_bcast_mpich
+#define MPIR_Pipelined_Bcast_MV2 smpi_coll_tuned_bcast_mpich
+#define MPIR_Bcast_binomial_MV2 smpi_coll_tuned_bcast_binomial_tree
+#define MPIR_Bcast_scatter_ring_allgather_shm_MV2 smpi_coll_tuned_bcast_scatter_LR_allgather
+#define MPIR_Bcast_scatter_doubling_allgather_MV2 smpi_coll_tuned_bcast_scatter_rdb_allgather
+#define MPIR_Bcast_scatter_ring_allgather_MV2 smpi_coll_tuned_bcast_scatter_LR_allgather
+#define MPIR_Shmem_Bcast_MV2 smpi_coll_tuned_bcast_mpich
+#define MPIR_Bcast_tune_inter_node_helper_MV2 smpi_coll_tuned_bcast_mvapich2_inter_node
+#define MPIR_Bcast_inter_node_helper_MV2 smpi_coll_tuned_bcast_mvapich2_inter_node
+#define MPIR_Knomial_Bcast_intra_node_MV2 smpi_coll_tuned_bcast_mvapich2_knomial_intra_node
+#define MPIR_Bcast_intra_MV2 smpi_coll_tuned_bcast_mvapich2_intra_node
 
- */
-
-
-/*
 static void init_mv2_bcast_tables_stampede(){
  //Stampede,
-        mv2_size_bcast_tuning_table=8;
-        mv2_bcast_thresholds_table = xbt_malloc(mv2_size_bcast_tuning_table *
-                                                 sizeof (mv2_bcast_tuning_table));
+  if(smpi_coll_cleanup_callback==NULL)
+    smpi_coll_cleanup_callback=&smpi_coll_cleanup_mvapich2;
+  mv2_size_bcast_tuning_table=8;
+  mv2_bcast_thresholds_table = xbt_malloc(mv2_size_bcast_tuning_table *
+  sizeof (mv2_bcast_tuning_table));
 
   mv2_bcast_tuning_table mv2_tmp_bcast_thresholds_table[]={
     {
@@ -1198,7 +1240,7 @@ static void init_mv2_bcast_tables_stampede(){
 
         memcpy(mv2_bcast_thresholds_table, mv2_tmp_bcast_thresholds_table,
                     mv2_size_bcast_tuning_table * sizeof (mv2_bcast_tuning_table));
-}*/
+}
 
 
 /************ Reduce variables and initializers                        */
@@ -1250,15 +1292,17 @@ int (*MV2_Reduce_intra_function)( void *sendbuf,
     MPI_Comm  comm_ptr)=NULL;
 
 
-#define MPIR_Reduce_inter_knomial_wrapper_MV2 smpi_coll_tuned_reduce_binomial
-#define MPIR_Reduce_intra_knomial_wrapper_MV2 smpi_coll_tuned_reduce_binomial
+#define MPIR_Reduce_inter_knomial_wrapper_MV2 smpi_coll_tuned_reduce_mvapich2_knomial
+#define MPIR_Reduce_intra_knomial_wrapper_MV2 smpi_coll_tuned_reduce_mvapich2_knomial
 #define MPIR_Reduce_binomial_MV2 smpi_coll_tuned_reduce_binomial
 #define MPIR_Reduce_redscat_gather_MV2 smpi_coll_tuned_reduce_scatter_gather
 #define MPIR_Reduce_shmem_MV2 smpi_coll_tuned_reduce_ompi_basic_linear
-
+#define MPIR_Reduce_two_level_helper_MV2 smpi_coll_tuned_reduce_mvapich2_two_level
 
 
 static void init_mv2_reduce_tables_stampede(){
+  if(smpi_coll_cleanup_callback==NULL)
+    smpi_coll_cleanup_callback=&smpi_coll_cleanup_mvapich2;
   /*Stampede*/
   mv2_size_reduce_tuning_table = 8;
   mv2_reduce_thresholds_table = xbt_malloc(mv2_size_reduce_tuning_table *
@@ -1501,6 +1545,8 @@ static  int MPIR_Reduce_Scatter_Basic_MV2(void *sendbuf,
 
 
 static void init_mv2_reduce_scatter_tables_stampede(){
+  if(smpi_coll_cleanup_callback==NULL)
+    smpi_coll_cleanup_callback=&smpi_coll_cleanup_mvapich2;
   mv2_size_red_scat_tuning_table = 6;
   mv2_red_scat_thresholds_table = xbt_malloc(mv2_size_red_scat_tuning_table *
       sizeof (mv2_red_scat_tuning_table));
@@ -1618,14 +1664,16 @@ int MPIR_Scatter_mcst_wrap_MV2(void *sendbuf,
 
 #define MPIR_Scatter_MV2_Binomial smpi_coll_tuned_scatter_ompi_binomial
 #define MPIR_Scatter_MV2_Direct smpi_coll_tuned_scatter_ompi_basic_linear
-#define MPIR_Scatter_MV2_two_level_Binomial smpi_coll_tuned_scatter_ompi_binomial
-#define MPIR_Scatter_MV2_two_level_Direct smpi_coll_tuned_scatter_ompi_basic_linear
+#define MPIR_Scatter_MV2_two_level_Binomial smpi_coll_tuned_scatter_mvapich2_two_level_binomial
+#define MPIR_Scatter_MV2_two_level_Direct smpi_coll_tuned_scatter_mvapich2_two_level_direct
 
 
 
 
 static void init_mv2_scatter_tables_stampede(){
-  {
+    if(smpi_coll_cleanup_callback==NULL)
+      smpi_coll_cleanup_callback=&smpi_coll_cleanup_mvapich2;
+
     int agg_table_sum = 0;
     int i;
     mv2_scatter_tuning_table **table_ptrs = NULL;
@@ -1932,6 +1980,6 @@ static void init_mv2_scatter_tables_stampede(){
                 * mv2_size_scatter_tuning_table[i]));
     }
     xbt_free(table_ptrs);
-  }
+  
 }