Merge pull request #12838 from Akshay-Venkatesh/topic/5.0.x/reduce-local-impl

5.0.x/ompi/coll/cuda: Implement reduce local

Author: janjust

ompi/mca/coll/cuda/coll_cuda.h

@@ -54,6 +54,11 @@ int mca_coll_cuda_reduce(const void *sbuf, void *rbuf, int count,
                          struct ompi_communicator_t *comm,
                          mca_coll_base_module_t *module);
 
+int mca_coll_cuda_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_cuda_exscan(const void *sbuf, void *rbuf, int count,
                          struct ompi_datatype_t *dtype,
                          struct ompi_op_t *op,

ompi/mca/coll/cuda/coll_cuda_module.c

@@ -40,6 +40,7 @@ static void mca_coll_cuda_module_destruct(mca_coll_cuda_module_t *module)
 {
     OBJ_RELEASE(module->c_coll.coll_allreduce_module);
     OBJ_RELEASE(module->c_coll.coll_reduce_module);
+    OBJ_RELEASE(module->c_coll.coll_reduce_local_module);
     OBJ_RELEASE(module->c_coll.coll_reduce_scatter_block_module);
     OBJ_RELEASE(module->c_coll.coll_scatter_module);
     /* If the exscan module is not NULL, then this was an
@@ -103,6 +104,7 @@ mca_coll_cuda_comm_query(struct ompi_communicator_t *comm,
     cuda_module->super.coll_gather     = NULL;
     cuda_module->super.coll_gatherv    = NULL;
     cuda_module->super.coll_reduce     = mca_coll_cuda_reduce;
+    cuda_module->super.coll_reduce_local = mca_coll_cuda_reduce_local;
     cuda_module->super.coll_reduce_scatter = NULL;
     cuda_module->super.coll_reduce_scatter_block = mca_coll_cuda_reduce_scatter_block;
     cuda_module->super.coll_scan       = mca_coll_cuda_scan;
@@ -135,6 +137,7 @@ int mca_coll_cuda_module_enable(mca_coll_base_module_t *module,
 
     CHECK_AND_RETAIN(comm, s, allreduce);
     CHECK_AND_RETAIN(comm, s, reduce);
+    CHECK_AND_RETAIN(comm, s, reduce_local);
     CHECK_AND_RETAIN(comm, s, reduce_scatter_block);
     CHECK_AND_RETAIN(comm, s, scatter);
     if (!OMPI_COMM_IS_INTER(comm)) {

ompi/mca/coll/cuda/coll_cuda_reduce.c

@@ -83,3 +83,60 @@ mca_coll_cuda_reduce(const void *sbuf, void *rbuf, int count,
     }
     return rc;
 }
+
+int
+mca_coll_cuda_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)
+{
+    ptrdiff_t gap;
+    char *rbuf1 = NULL, *sbuf1 = NULL, *rbuf2 = NULL;
+    size_t bufsize;
+    int rc;
+
+    bufsize = opal_datatype_span(&dtype->super, count, &gap);
+
+    rc = mca_coll_cuda_check_buf((void *)sbuf);
+    if (rc < 0) {
+        return rc;
+    }
+
+    if ((MPI_IN_PLACE != sbuf) && (rc > 0)) {
+        sbuf1 = (char*)malloc(bufsize);
+        if (NULL == sbuf1) {
+            return OMPI_ERR_OUT_OF_RESOURCE;
+        }
+        mca_coll_cuda_memcpy(sbuf1, sbuf, bufsize);
+        sbuf = sbuf1 - gap;
+    }
+
+    rc = mca_coll_cuda_check_buf(rbuf);
+    if (rc < 0) {
+        return rc;
+    }
+
+    if (rc > 0) {
+        rbuf1 = (char*)malloc(bufsize);
+        if (NULL == rbuf1) {
+            if (NULL != sbuf1) free(sbuf1);
+            return OMPI_ERR_OUT_OF_RESOURCE;
+        }
+        mca_coll_cuda_memcpy(rbuf1, rbuf, bufsize);
+        rbuf2 = rbuf; /* save away original buffer */
+        rbuf = rbuf1 - gap;
+    }
+
+    ompi_op_reduce(op, (void *)sbuf, rbuf, count, dtype);
+    rc = OMPI_SUCCESS;
+
+    if (NULL != sbuf1) {
+        free(sbuf1);
+    }
+    if (NULL != rbuf1) {
+        rbuf = rbuf2;
+        mca_coll_cuda_memcpy(rbuf, rbuf1, bufsize);
+        free(rbuf1);
+    }
+    return rc;
+}