have smp-aware algorithms use number of cores on the node as basis for their computat...

[simgrid.git] / src / smpi / colls / allreduce-smp-rsag.c

diff --git a/src/smpi/colls/allreduce-smp-rsag.c b/src/smpi/colls/allreduce-smp-rsag.c
index a40f189..9de2bcf 100644 (file)
--- a/src/smpi/colls/allreduce-smp-rsag.c
+++ b/src/smpi/colls/allreduce-smp-rsag.c
@@ -1,4 +1,4 @@
-#include "colls.h"
+#include "colls_private.h"
 
 /* change number of core per smp-node
    we assume that number of core per process will be the same for all implementations */
@@ -19,10 +19,13 @@ int smpi_coll_tuned_allreduce_smp_rsag(void *send_buf, void *recv_buf,
 {
   int comm_size, rank;
   void *tmp_buf;
-  int tag = 50;
+  int tag = COLL_TAG_ALLREDUCE;
   int mask, src, dst;
   MPI_Status status;
-  int num_core = NUM_CORE;
+  int num_core = simcall_host_get_core(SIMIX_host_self());
+  // do we use the default one or the number of cores in the platform ?
+  // if the number of cores is one, the platform may be simulated with 1 node = 1 core
+  if (num_core == 1) num_core = NUM_CORE;
   /*
      #ifdef MPICH2_REDUCTION
      MPI_User_function * uop = MPIR_Op_table[op % 16 - 1];
@@ -33,11 +36,11 @@ int smpi_coll_tuned_allreduce_smp_rsag(void *send_buf, void *recv_buf,
      uop  = op_ptr->op;
      #endif
    */
-  MPI_Comm_size(comm, &comm_size);
-  MPI_Comm_rank(comm, &rank);
+  comm_size = smpi_comm_size(comm);
+  rank = smpi_comm_rank(comm);
   MPI_Aint extent;
-  MPI_Type_extent(dtype, &extent);
-  tmp_buf = (void *) malloc(count * extent);
+  extent = smpi_datatype_get_extent(dtype);
+  tmp_buf = (void *) xbt_malloc(count * extent);
 
   int intra_rank, inter_rank;
   intra_rank = rank % num_core;
@@ -53,7 +56,7 @@ int smpi_coll_tuned_allreduce_smp_rsag(void *send_buf, void *recv_buf,
   }
 
 
-  MPI_Sendrecv(send_buf, count, dtype, rank, tag,
+  smpi_mpi_sendrecv(send_buf, count, dtype, rank, tag,
                recv_buf, count, dtype, rank, tag, comm, &status);
 
 
@@ -64,14 +67,14 @@ int smpi_coll_tuned_allreduce_smp_rsag(void *send_buf, void *recv_buf,
       src = (inter_rank * num_core) + (intra_rank | mask);
       //      if (src < ((inter_rank + 1) * num_core)) {
       if (src < comm_size) {
-        MPI_Recv(tmp_buf, count, dtype, src, tag, comm, &status);
-        star_reduction(op, tmp_buf, recv_buf, &count, &dtype);
+        smpi_mpi_recv(tmp_buf, count, dtype, src, tag, comm, &status);
+        smpi_op_apply(op, tmp_buf, recv_buf, &count, &dtype);
         //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
       }
     } else {
 
       dst = (inter_rank * num_core) + (intra_rank & (~mask));
-      MPI_Send(recv_buf, count, dtype, dst, tag, comm);
+      smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
       //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
       break;
     }
@@ -100,12 +103,12 @@ int smpi_coll_tuned_allreduce_smp_rsag(void *send_buf, void *recv_buf,
           ((inter_rank - 2 - i +
             inter_comm_size) % inter_comm_size) * seg_count * extent;
 
-      MPI_Sendrecv((char *) recv_buf + send_offset, seg_count, dtype, to,
+      smpi_mpi_sendrecv((char *) recv_buf + send_offset, seg_count, dtype, to,
                    tag + i, tmp_buf, seg_count, dtype, from, tag + i, comm,
                    &status);
 
       // result is in rbuf
-      star_reduction(op, tmp_buf, (char *) recv_buf + recv_offset, &seg_count,
+      smpi_op_apply(op, tmp_buf, (char *) recv_buf + recv_offset, &seg_count,
                      &dtype);
     }
 
@@ -119,7 +122,7 @@ int smpi_coll_tuned_allreduce_smp_rsag(void *send_buf, void *recv_buf,
           ((inter_rank - 1 - i +
             inter_comm_size) % inter_comm_size) * seg_count * extent;
 
-      MPI_Sendrecv((char *) recv_buf + send_offset, seg_count, dtype, to,
+      smpi_mpi_sendrecv((char *) recv_buf + send_offset, seg_count, dtype, to,
                    tag + i, (char *) recv_buf + recv_offset, seg_count, dtype,
                    from, tag + i, comm, &status);
 
@@ -139,14 +142,14 @@ int smpi_coll_tuned_allreduce_smp_rsag(void *send_buf, void *recv_buf,
      if ((mask & inter_rank) == 0) {
      src = (inter_rank | mask) * num_core;
      if (src < comm_size) {
-     MPI_Recv(tmp_buf, count, dtype, src, tag, comm, &status);
+     smpi_mpi_recv(tmp_buf, count, dtype, src, tag, comm, &status);
      (* uop) (tmp_buf, recv_buf, &count, &dtype);
      //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
      }
      }
      else {
      dst = (inter_rank & (~mask)) * num_core;
-     MPI_Send(recv_buf, count, dtype, dst, tag, comm);
+     smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
      //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
      break;
      }
@@ -165,7 +168,7 @@ int smpi_coll_tuned_allreduce_smp_rsag(void *send_buf, void *recv_buf,
      if (inter_rank & mask) {
      src = (inter_rank - mask) * num_core;
      //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
-     MPI_Recv(recv_buf, count, dtype, src, tag, comm, &status);
+     smpi_mpi_recv(recv_buf, count, dtype, src, tag, comm, &status);
      break;
      }
      mask <<= 1;
@@ -179,7 +182,7 @@ int smpi_coll_tuned_allreduce_smp_rsag(void *send_buf, void *recv_buf,
      dst = (inter_rank + mask) * num_core;
      if (dst < comm_size) {
      //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
-     MPI_Send(recv_buf, count, dtype, dst, tag, comm);
+     smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
      }
      }
      mask >>= 1;
@@ -200,7 +203,7 @@ int smpi_coll_tuned_allreduce_smp_rsag(void *send_buf, void *recv_buf,
     if (intra_rank & mask) {
       src = (inter_rank * num_core) + (intra_rank - mask);
       //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
-      MPI_Recv(recv_buf, count, dtype, src, tag, comm, &status);
+      smpi_mpi_recv(recv_buf, count, dtype, src, tag, comm, &status);
       break;
     }
     mask <<= 1;
@@ -213,7 +216,7 @@ int smpi_coll_tuned_allreduce_smp_rsag(void *send_buf, void *recv_buf,
     dst = (inter_rank * num_core) + (intra_rank + mask);
     if (dst < comm_size) {
       //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
-      MPI_Send(recv_buf, count, dtype, dst, tag, comm);
+      smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
     }
     mask >>= 1;
   }