ompi/coll/accelerator/cuda: implement reduce_local

Signed-off-by: Akshay Venkatesh <[email protected]>

bot:notacherrypick

Author: Akshay-Venkatesh

ompi/mca/coll/accelerator/coll_accelerator.h

@@ -1,4 +1,5 @@
 /*
+ * Copyright (c) 2024      NVIDIA Corporation.  All rights reserved.
  * Copyright (c) 2014      The University of Tennessee and The University
  *                         of Tennessee Research Foundation.  All rights
  *                         reserved.
@@ -45,6 +46,11 @@ mca_coll_accelerator_allreduce(const void *sbuf, void *rbuf, size_t count,
                         struct ompi_communicator_t *comm,
                         mca_coll_base_module_t *module);
 
+int mca_coll_accelerator_reduce_local(const void *sbuf, void *rbuf, size_t count,
+                         struct ompi_datatype_t *dtype,
+                         struct ompi_op_t *op,
+                         mca_coll_base_module_t *module);
+
 int mca_coll_accelerator_reduce(const void *sbuf, void *rbuf, size_t count,
                          struct ompi_datatype_t *dtype,
                          struct ompi_op_t *op,

ompi/mca/coll/accelerator/coll_accelerator_module.c

@@ -1,4 +1,5 @@
 /*
+ * Copyright (c) 2024      NVIDIA Corporation.  All rights reserved.
  * Copyright (c) 2014-2017 The University of Tennessee and The University
  *                         of Tennessee Research Foundation.  All rights
  *                         reserved.
@@ -94,6 +95,7 @@ mca_coll_accelerator_comm_query(struct ompi_communicator_t *comm,
 
     accelerator_module->super.coll_allreduce  = mca_coll_accelerator_allreduce;
     accelerator_module->super.coll_reduce     = mca_coll_accelerator_reduce;
+    accelerator_module->super.coll_reduce_local         = mca_coll_accelerator_reduce_local;
     accelerator_module->super.coll_reduce_scatter_block = mca_coll_accelerator_reduce_scatter_block;
     if (!OMPI_COMM_IS_INTER(comm)) {
         accelerator_module->super.coll_scan       = mca_coll_accelerator_scan;
@@ -141,6 +143,7 @@ mca_coll_accelerator_module_enable(mca_coll_base_module_t *module,
 
     ACCELERATOR_INSTALL_COLL_API(comm, s, allreduce);
     ACCELERATOR_INSTALL_COLL_API(comm, s, reduce);
+    ACCELERATOR_INSTALL_COLL_API(comm, s, reduce_local);
     ACCELERATOR_INSTALL_COLL_API(comm, s, reduce_scatter_block);
     if (!OMPI_COMM_IS_INTER(comm)) {
         /* MPI does not define scan/exscan on intercommunicators */
@@ -159,6 +162,7 @@ mca_coll_accelerator_module_disable(mca_coll_base_module_t *module,
 
     ACCELERATOR_UNINSTALL_COLL_API(comm, s, allreduce);
     ACCELERATOR_UNINSTALL_COLL_API(comm, s, reduce);
+    ACCELERATOR_UNINSTALL_COLL_API(comm, s, reduce_local);
     ACCELERATOR_UNINSTALL_COLL_API(comm, s, reduce_scatter_block);
     if (!OMPI_COMM_IS_INTER(comm))
     {

ompi/mca/coll/accelerator/coll_accelerator_reduce.c

@@ -1,4 +1,5 @@
 /*
+ * Copyright (c) 2024      NVIDIA Corporation.  All rights reserved.
  * Copyright (c) 2004-2023 The University of Tennessee and The University
  *                         of Tennessee Research Foundation.  All rights
  *                         reserved.
@@ -35,7 +36,7 @@ mca_coll_accelerator_reduce(const void *sbuf, void *rbuf, size_t count,
                      mca_coll_base_module_t *module)
 {
     mca_coll_accelerator_module_t *s = (mca_coll_accelerator_module_t*) module;
-    int rank = ompi_comm_rank(comm);
+    int rank = (comm == NULL) ? -1 : ompi_comm_rank(comm);
     ptrdiff_t gap;
     char *rbuf1 = NULL, *sbuf1 = NULL, *rbuf2 = NULL;
     size_t bufsize;
@@ -70,11 +71,18 @@ mca_coll_accelerator_reduce(const void *sbuf, void *rbuf, size_t count,
         rbuf2 = rbuf; /* save away original buffer */
         rbuf = rbuf1 - gap;
     }
-    rc = s->c_coll.coll_reduce((void *) sbuf, rbuf, count,
-                               dtype, op, root, comm,
-                               s->c_coll.coll_reduce_module);
+
+    if ((comm == NULL) && (root == -1)) {
+        ompi_op_reduce(op, (void *)sbuf, rbuf, count, dtype);
+        rc = OMPI_SUCCESS;
+    } else {
+        rc = s->c_coll.coll_reduce((void *) sbuf, rbuf, count,
+                                   dtype, op, root, comm,
+                                   s->c_coll.coll_reduce_module);
+    }
 
     if (NULL != sbuf1) {
+        sbuf = sbuf2;
         free(sbuf1);
     }
     if (NULL != rbuf1) {
@@ -84,3 +92,13 @@ mca_coll_accelerator_reduce(const void *sbuf, void *rbuf, size_t count,
     }
     return rc;
 }
+
+int
+mca_coll_accelerator_reduce_local(const void *sbuf, void *rbuf, size_t count,
+                                  struct ompi_datatype_t *dtype,
+                                  struct ompi_op_t *op,
+                                  mca_coll_base_module_t *module)
+{
+    return mca_coll_accelerator_reduce(sbuf, rbuf, count, dtype, op, -1, NULL,
+                                       module);
+}