coll/ch4: activate CH4 POSIX release_gather bcast

Enable CVARs and JSONs to select ch4-posix layer release_gather
algorithms.

Select MPIDI_POSIX_mpi_bcast_release_gather if it passes
MPIDI_CH4_release_gather condition check, which only passes if comm is
an posix intranode comm.

Author: hzhou

src/mpi/coll/coll_algorithms.txt

@@ -70,6 +70,9 @@ conditions:
     avg_msg_size(thresh):   MPIR_Csel_avg_msg_size
     total_msg_size(thresh): MPIR_Csel_total_msg_size
 
+    # conditional conditions - only call the condition function under macro_guard
+    MPIDI_CH4_release_gather: MPIDI_POSIX_check_release_gather #if defined(MPIDI_CH4_SHM_POSIX)
+
 # ----
 general:
     MPIR_Coll_auto
@@ -110,6 +113,11 @@ bcast-intra:
     pipelined_tree
         extra_params: tree_type, k, is_non_blocking, chunk_size, recv_pre_posted
         cvar_params: TREE_TYPE, TREE_KVAL, IS_NON_BLOCKING, TREE_PIPELINE_CHUNK_SIZE, RECV_PRE_POST 
+    release_gather
+        func_name: MPIDI_POSIX_mpi_bcast_release_gather
+        inline: 1
+        macro_guard: defined(MPIDI_CH4_SHM_POSIX)
+        restrictions: MPIDI_CH4_release_gather
 ibcast-intra:
     sched_binomial
     sched_smp

src/mpi/coll/coll_selection.json

@@ -1,6 +1,10 @@
 {
     "collective=bcast-intra":
     {
+            "MPIDI_CH4_release_gather":
+            {
+                "algorithm=MPIDI_POSIX_mpi_bcast_release_gather":{}
+            },
             "comm_size(8)":
             {
                 "algorithm=MPIR_Bcast_intra_binomial":{}

src/mpi/coll/cvars.txt

@@ -202,6 +202,7 @@ cvars:
         scatter_ring_allgather                  - Force Scatter Ring
         pipelined_tree                          - Force tree-based pipelined algorithm
         tree                                    - Force tree-based algorithm
+        release_gather                          - Force CH4 POSIX release_gather algorithm
 
     - name        : MPIR_CVAR_BCAST_TREE_KVAL
       category    : COLLECTIVE

src/mpid/ch4/shm/posix/posix_coll_release_gather.h

@@ -29,6 +29,47 @@
 #include "algo_common.h"
 #include "release_gather.h"
 
+MPL_STATIC_INLINE_PREFIX bool MPIDI_POSIX_check_release_gather(MPIR_Csel_coll_sig_s * coll_sig)
+{
+    if (MPIR_IS_THREADED) {
+        return false;
+    }
+
+    /* Check whether comm is an intranode comm */
+    MPIR_Comm *comm_ptr = coll_sig->comm_ptr;
+    MPIR_Assert(comm_ptr->attr & MPIR_COMM_ATTR__HIERARCHY);
+    if (comm_ptr->num_external > 1) {
+        return false;
+    }
+
+    /* check coll_type */
+    MPIDI_POSIX_release_gather_opcode_t opcode;
+    switch (coll_sig->coll_type) {
+        case MPIR_CSEL_COLL_TYPE__INTRA_BCAST:
+            opcode = MPIDI_POSIX_RELEASE_GATHER_OPCODE_BCAST;
+            break;
+        default:
+            return false;
+    }
+
+    /* Check repeats if the algorithm CVAR is not set */
+    if (!(coll_sig->flags & MPIR_COLL_SIG_FLAG__CVAR)) {
+        MPIDI_POSIX_COMM(comm_ptr, release_gather).num_collective_calls++;
+        if (MPIDI_POSIX_COMM(comm_ptr, release_gather).num_collective_calls <
+            MPIR_CVAR_POSIX_NUM_COLLS_THRESHOLD) {
+            return false;
+        }
+    }
+
+    /* Lazy initialization of release_gather specific struct */
+    int mpi_errno = MPIDI_POSIX_mpi_release_gather_comm_init(comm_ptr, opcode);
+    if (mpi_errno != MPI_SUCCESS) {
+        return false;
+    }
+
+    return true;
+}
+
 /* Intra-node bcast is implemented as a release step followed by gather step in release_gather
  * framework. The actual data movement happens in release step. Gather step makes sure that
  * the shared bcast buffer can be reused for next bcast call. Release gather framework has
@@ -59,20 +100,6 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_POSIX_mpi_bcast_release_gather(void *buffer,
         goto fn_exit;
     }
 
-    MPIDI_POSIX_COMM(comm_ptr, release_gather).num_collective_calls++;
-    if (MPIDI_POSIX_COMM(comm_ptr, release_gather).num_collective_calls <
-        MPIR_CVAR_POSIX_NUM_COLLS_THRESHOLD) {
-        /* Fallback to pt2pt algorithms if the total number of release_gather collective calls is
-         * less than the specified threshold */
-        goto fallback;
-    }
-
-    /* Lazy initialization of release_gather specific struct */
-    mpi_errno =
-        MPIDI_POSIX_mpi_release_gather_comm_init(comm_ptr, MPIDI_POSIX_RELEASE_GATHER_OPCODE_BCAST);
-    MPII_COLLECTIVE_FALLBACK_CHECK(MPIR_Comm_rank(comm_ptr), !mpi_errno, mpi_errno,
-                                   "release_gather bcast cannot create more shared memory. Falling back to pt2pt algorithms.\n");
-
     my_rank = MPIR_Comm_rank(comm_ptr);
     MPIR_Type_get_extent_impl(datatype, &lb, &extent);
     MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
@@ -150,10 +177,6 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_POSIX_mpi_bcast_release_gather(void *buffer,
     return mpi_errno;
   fn_fail:
     goto fn_exit;
-  fallback:
-    /* FIXME: proper error */
-    mpi_errno = MPI_ERR_OTHER;
-    goto fn_exit;
 }
 
 /* Intra-node reduce is implemented as a release step followed by gather step in release_gather

src/mpid/ch4/shm/posix/posix_pre.h

@@ -9,6 +9,10 @@
 #include <mpi.h>
 #include "release_gather_types.h"
 
+/* define a macro to sigify that we have CH4 POSIX. An example usage is
+ * to macro-guard posix collective algorithms */
+#define MPIDI_CH4_SHM_POSIX 1
+
 #define MPIDI_POSIX_MAX_AM_HDR_SIZE     800     /* constrained by MPIDI_POSIX_AM_HDR_POOL_CELL_SIZE */
 #define MPIDI_POSIX_AM_MSG_HEADER_SIZE  (sizeof(MPIDI_POSIX_am_header_t))
 #define MPIDI_POSIX_MAX_IOV_NUM         (3)     /* am_hdr, [padding], payload */

src/mpid/ch4/shm/posix/release_gather/release_gather.c

@@ -481,6 +481,7 @@ int MPIDI_POSIX_mpi_release_gather_comm_init(MPIR_Comm * comm_ptr,
     if (mpi_errno != MPI_SUCCESS) {
         MPIDI_POSIX_mpi_release_gather_comm_free(comm_ptr);
         RELEASE_GATHER_FIELD(comm_ptr, is_initialized) = 0;
+        /* TODO: set a flag so next time we skip the try */
     }
     return mpi_errno;
   fn_fail: