Fix ADAPT for few corner cases

Add support for fallback to previous coll module on non-commutative operations (open-mpi#30)
Replace mutexes by atomic operations.
Use the correct nbc request type (open-mpi#31)
* coll/base: document type casts in ompi_coll_base_retain_*
Other minor fixes.

Signed-off-by: George Bosilca <[email protected]>
Signed-off-by: Joseph Schuchart <[email protected]>

ompi/mca/coll/adapt/coll_adapt.h

@@ -73,11 +73,44 @@ typedef struct mca_coll_adapt_component_t {
 
 } mca_coll_adapt_component_t;
 
+/*
+ * Structure used to store what is necessary for the collective operations
+ * routines in case of fallback.
+ */
+typedef struct mca_coll_adapt_collective_fallback_s {
+    union {
+        mca_coll_base_module_reduce_fn_t   reduce;
+        mca_coll_base_module_ireduce_fn_t ireduce;
+    } previous_routine;
+    mca_coll_base_module_t *previous_module;
+} mca_coll_adapt_collective_fallback_t;
+
+
+typedef enum mca_coll_adapt_colltype {
+    ADAPT_REDUCE  = 0,
+    ADAPT_IREDUCE = 1,
+    ADAPT_COLLCOUNT
+} mca_coll_adapt_colltype_t;
+
+/*
+ * Some defines to stick to the naming used in the other components in terms of
+ * fallback routines
+ */
+#define previous_reduce     previous_routines[ADAPT_REDUCE].previous_routine.reduce
+#define previous_ireduce    previous_routines[ADAPT_IREDUCE].previous_routine.ireduce
+
+#define previous_reduce_module     previous_routines[ADAPT_REDUCE].previous_module
+#define previous_ireduce_module    previous_routines[ADAPT_IREDUCE].previous_module
+
+
 /* Coll adapt module per communicator*/
 struct mca_coll_adapt_module_t {
     /* Base module */
     mca_coll_base_module_t super;
 
+    /* To be able to fallback when the cases are not supported */
+    struct mca_coll_adapt_collective_fallback_s previous_routines[ADAPT_COLLCOUNT];
+
     /* Whether this module has been lazily initialized or not yet */
     bool adapt_enabled;
 };

ompi/mca/coll/adapt/coll_adapt_context.h

@@ -77,17 +77,13 @@ struct ompi_coll_adapt_constant_reduce_context_s {
     ompi_request_t *request;
     int rank;
     /* Length of the fragment array, which is the number of recevied segments */
-    int32_t num_recv_segs;
+    opal_atomic_int32_t num_recv_segs;
     /* Number of sent segments */
-    int32_t num_sent_segs;
+    opal_atomic_int32_t num_sent_segs;
     /* Next seg need to be received for every children */
     opal_atomic_int32_t *next_recv_segs;
     /* Mutex to protect recv_list */
     opal_mutex_t *mutex_recv_list;
-    /* Mutex to protect num_recv_segs */
-    opal_mutex_t *mutex_num_recv_segs;
-    /* Mutex to protect num_sent */
-    opal_mutex_t *mutex_num_sent;
     /* Mutex to protect each segment when do the reduce op */
     opal_mutex_t *mutex_op_list;
     /* Reduce operation */