coll-portals4: add gather and igather implementations that use Portals4 triggered operations

This commit adds implementations of gather and igather using
Portals4 triggered operations.  The default algorithm is linear,
but binomial can be selected using an MCA parameter -
coll_portals4_use_binomial_gather_algorithm.

Author: tkordenbrock

ompi/mca/coll/portals4/Makefile.am

@@ -15,6 +15,7 @@ local_sources = \
 	coll_portals4_barrier.c \
 	coll_portals4_bcast.c \
 	coll_portals4_reduce.c \
+	coll_portals4_gather.c \
 	coll_portals4_request.h \
 	coll_portals4_request.c
 

ompi/mca/coll/portals4/coll_portals4.h

@@ -24,13 +24,18 @@
 #include "ompi/datatype/ompi_datatype_internal.h"
 #include "ompi/op/op.h"
 #include "ompi/mca/mca.h"
+#include "opal/datatype/opal_convertor.h"
 #include "ompi/mca/coll/coll.h"
 #include "ompi/request/request.h"
 #include "ompi/communicator/communicator.h"
 #include "ompi/mca/coll/base/base.h"
+#include "ompi/datatype/ompi_datatype.h"
+#include "ompi/mca/mtl/portals4/mtl_portals4_endpoint.h"
 
 #include "ompi/mca/mtl/portals4/mtl_portals4.h"
 
+#define MAXTREEFANOUT 32
+
 BEGIN_C_DECLS
 
 #define COLL_PORTALS4_NO_OP ((ptl_op_t)-1)
@@ -61,10 +66,27 @@ struct mca_coll_portals4_component_t {
 
     ptl_ni_limits_t ni_limits;
 
+    int use_binomial_gather_algorithm;
+
 };
 typedef struct mca_coll_portals4_component_t mca_coll_portals4_component_t;
 OMPI_MODULE_DECLSPEC extern mca_coll_portals4_component_t mca_coll_portals4_component;
 
+
+/*
+ * Borrowed with thanks from the coll-tuned component, then modified for Portals4.
+ */
+typedef struct ompi_coll_portals4_tree_t {
+    int32_t tree_root;
+    int32_t tree_fanout;
+    int32_t tree_bmtree;
+    int32_t tree_prev;
+    int32_t tree_next[MAXTREEFANOUT];
+    int32_t tree_nextsize;
+    int32_t tree_numdescendants;
+} ompi_coll_portals4_tree_t;
+
+
 struct mca_coll_portals4_module_t {
     mca_coll_base_module_t super;
     size_t coll_count;
@@ -79,6 +101,13 @@ struct mca_coll_portals4_module_t {
     mca_coll_base_module_t *previous_allreduce_module;
     mca_coll_base_module_iallreduce_fn_t previous_iallreduce;
     mca_coll_base_module_t *previous_iallreduce_module;
+
+    /* binomial tree */
+    ompi_coll_portals4_tree_t *cached_in_order_bmtree;
+    int                        cached_in_order_bmtree_root;
+
+    size_t barrier_count;
+    size_t gather_count;
 };
 typedef struct mca_coll_portals4_module_t mca_coll_portals4_module_t;
 OBJ_CLASS_DECLARATION(mca_coll_portals4_module_t);
@@ -135,6 +164,22 @@ int
 opal_stderr(const char *msg, const char *file,
         const int line, const int ret);
 
+/*
+ * Borrowed with thanks from the coll-tuned component.
+ */
+#define COLL_PORTALS4_UPDATE_IN_ORDER_BMTREE( OMPI_COMM, PORTALS4_MODULE, ROOT ) \
+do {                                                                                         \
+    if( !( ((PORTALS4_MODULE)->cached_in_order_bmtree)                                               \
+           && ((PORTALS4_MODULE)->cached_in_order_bmtree_root == (ROOT)) ) ) {                       \
+        if( (PORTALS4_MODULE)->cached_in_order_bmtree ) { /* destroy previous binomial if defined */ \
+            ompi_coll_portals4_destroy_tree( &((PORTALS4_MODULE)->cached_in_order_bmtree) );       \
+        }                                                                                    \
+        (PORTALS4_MODULE)->cached_in_order_bmtree = ompi_coll_portals4_build_in_order_bmtree( (OMPI_COMM), (ROOT) ); \
+        (PORTALS4_MODULE)->cached_in_order_bmtree_root = (ROOT);                                     \
+    }                                                                                        \
+} while (0)
+
+
 int ompi_coll_portals4_barrier_intra(struct ompi_communicator_t *comm,
         mca_coll_base_module_t *module);
 int ompi_coll_portals4_ibarrier_intra(struct ompi_communicator_t *comm,
@@ -177,6 +222,20 @@ int ompi_coll_portals4_iallreduce_intra(const void* sendbuf, void* recvbuf, int
 int
 ompi_coll_portals4_iallreduce_intra_fini(struct ompi_coll_portals4_request_t *request);
 
+int ompi_coll_portals4_gather_intra(const void *sbuf, int scount, struct ompi_datatype_t *sdtype,
+                                    void *rbuf, int rcount, struct ompi_datatype_t *rdtype,
+                                    int root,
+                                    struct ompi_communicator_t *comm,
+                                    mca_coll_base_module_t *module);
+int ompi_coll_portals4_igather_intra(const void *sbuf, int scount, struct ompi_datatype_t *sdtype,
+                                     void *rbuf, int rcount, struct ompi_datatype_t *rdtype,
+                                     int root,
+                                     struct ompi_communicator_t *comm,
+                                     ompi_request_t **request,
+                                     mca_coll_base_module_t *module);
+int ompi_coll_portals4_igather_intra_fini(struct ompi_coll_portals4_request_t *request);
+
+
 static inline ptl_process_t
 ompi_coll_portals4_get_peer(struct ompi_communicator_t *comm, int rank)
 {
@@ -357,6 +416,43 @@ void get_k_ary_tree(const unsigned int k_ary,
     return;
 }
 
+
+static inline void
+ompi_coll_portals4_create_recv_converter (opal_convertor_t *converter,
+                                          void *target,
+                                          ompi_proc_t *proc,
+                                          int count,
+                                          ompi_datatype_t *datatype)
+{
+    /* create converter */
+    OBJ_CONSTRUCT(converter, opal_convertor_t);
+
+    /* initialize converter */
+    opal_convertor_copy_and_prepare_for_recv(proc->super.proc_convertor,
+                                             &datatype->super,
+                                             count,
+                                             target,
+                                             0,
+                                             converter);
+}
+
+static inline void
+ompi_coll_portals4_create_send_converter (opal_convertor_t *converter,
+                                          const void *source,
+                                          ompi_proc_t *proc,
+                                          int count,
+                                          ompi_datatype_t *datatype)
+{
+    OBJ_CONSTRUCT(converter, opal_convertor_t);
+
+    opal_convertor_copy_and_prepare_for_send(proc->super.proc_convertor,
+                                             &datatype->super,
+                                             count,
+                                             source,
+                                             0,
+                                             converter);
+}
+
 END_C_DECLS
 
 #endif /* MCA_COLL_PORTALS4_EXPORT_H */

ompi/mca/coll/portals4/coll_portals4_component.c

@@ -203,6 +203,14 @@ portals4_register(void)
             MCA_BASE_VAR_SCOPE_READONLY,
             &mca_coll_portals4_priority);
 
+    mca_coll_portals4_component.use_binomial_gather_algorithm = 0;
+    (void) mca_base_component_var_register(&mca_coll_portals4_component.super.collm_version, "use_binomial_gather_algorithm",
+            "if 1 use a binomial tree algorithm for gather, otherwise use linear",
+            MCA_BASE_VAR_TYPE_INT, NULL, 0, 0,
+            OPAL_INFO_LVL_9,
+            MCA_BASE_VAR_SCOPE_READONLY,
+            &mca_coll_portals4_component.use_binomial_gather_algorithm);
+
     return OMPI_SUCCESS;
 }
 
@@ -463,7 +471,7 @@ portals4_init_query(bool enable_progress_threads,
                 __FILE__, __LINE__, ret);
         return OMPI_ERROR;
     }
-    OPAL_OUTPUT_VERBOSE((90, ompi_coll_base_framework.framework_output, "PtlMDBind start=%p length=%x\n", md.start, md.length));
+    OPAL_OUTPUT_VERBOSE((90, ompi_coll_base_framework.framework_output, "PtlMDBind start=%p length=%lx\n", md.start, md.length));
 
     /* setup finish ack ME */
     me.start = NULL;
@@ -472,7 +480,7 @@ portals4_init_query(bool enable_progress_threads,
     me.min_free = 0;
     me.uid = mca_coll_portals4_component.uid;
     me.options = PTL_ME_OP_PUT |
-            PTL_ME_EVENT_LINK_DISABLE | PTL_ME_EVENT_UNLINK_DISABLE;
+        PTL_ME_EVENT_LINK_DISABLE | PTL_ME_EVENT_UNLINK_DISABLE;
     me.match_id.phys.nid = PTL_NID_ANY;
     me.match_id.phys.pid = PTL_PID_ANY;
     me.match_bits = 0;
@@ -584,6 +592,12 @@ portals4_comm_query(struct ompi_communicator_t *comm,
     portals4_module->super.coll_barrier = ompi_coll_portals4_barrier_intra;
     portals4_module->super.coll_ibarrier = ompi_coll_portals4_ibarrier_intra;
 
+    portals4_module->super.coll_gather   = ompi_coll_portals4_gather_intra;
+    portals4_module->super.coll_igather  = ompi_coll_portals4_igather_intra;
+
+    portals4_module->cached_in_order_bmtree=NULL;
+    portals4_module->cached_in_order_bmtree_root=-1;
+
     portals4_module->super.coll_bcast = ompi_coll_portals4_bcast_intra;
     portals4_module->super.coll_ibcast = ompi_coll_portals4_ibcast_intra;
 
@@ -593,6 +607,9 @@ portals4_comm_query(struct ompi_communicator_t *comm,
     portals4_module->super.coll_reduce = ompi_coll_portals4_reduce_intra;
     portals4_module->super.coll_ireduce = ompi_coll_portals4_ireduce_intra;
 
+    portals4_module->barrier_count = 0;
+    portals4_module->gather_count = 0;
+
     return &(portals4_module->super);
 }
 
@@ -689,9 +706,11 @@ portals4_progress(void)
                         ompi_coll_portals4_iallreduce_intra_fini(ptl_request);
                         break;
                     case OMPI_COLL_PORTALS4_TYPE_SCATTER:
-                    case OMPI_COLL_PORTALS4_TYPE_GATHER:
                         opal_output(ompi_coll_base_framework.framework_output,
-                                "allreduce is not supported yet\n");
+                                "scatter is not supported yet\n");
+                        break;
+                    case OMPI_COLL_PORTALS4_TYPE_GATHER:
+                        ompi_coll_portals4_igather_intra_fini(ptl_request);
                         break;
                     }
                 }