zgetrf: Add all_reduce algorithm choice

compute/zgetrf.c

@@ -98,6 +98,21 @@ CHAMELEON_zgetrf_WS_Alloc( const CHAM_desc_t *A )
         chameleon_cleanenv( algostr );
     }
 
+    {
+        char *allreduce = chameleon_getenv( "CHAMELEON_GETRF_ALL_REDUCE" );
+
+        if ( allreduce != NULL ) {
+            if ( strcasecmp( allreduce, "cham_spu_tasks" ) == 0 ) {
+                ws->alg_allreduce = ChamStarPUTasks;
+            }
+            else {
+                chameleon_error( "CHAMELEON_zgetrf_WS_Alloc", "CHAMELEON_GETRF_ALL_REDUCE is not one of chameleon_starpu_tasks, chameleon_starpu, chameleon_starpu_mpi, chameleon_mpi => Switch back to chameleon_starpu_tasks\n" );
+                ws->alg_allreduce = ChamStarPUTasks;
+            }
+        }
+        chameleon_cleanenv( allreduce );
+    }
+
     ws->batch_size = chameleon_getenv_get_value_int( "CHAMELEON_GETRF_BATCH_SIZE", 0 );
     if ( ws->batch_size > CHAMELEON_BATCH_SIZE ) {
         chameleon_warning( "CHAMELEON_BATCH_SIZE", "CHAMELEON_GETRF_BATCH_SIZE must be smaller than CHAMELEON_BATCH_SIZE, please recompile with the right CHAMELEON_BATCH_SIZE, or reduce the CHAMELEON_GETRF_BATCH_SIZE value\n" );

control/compute_z.h

@@ -43,17 +43,18 @@ struct chameleon_pzgemm_s {
  * @brief Data structure to handle the GETRF workspaces with partial pivoting
  */
 struct chameleon_pzgetrf_s {
-    cham_getrf_t   alg;
-    int            ib;         /**< Internal blocking parameter                   */
-    int            batch_size; /**< Batch size for the panel                      */
-    int            ringswitch; /**< Define when to switch to ring bcast           */
-    CHAM_desc_t    U;
-    CHAM_desc_t    Up;         /**< Workspace used for the panel factorization    */
-    CHAM_desc_t    Wu;         /**< Workspace used for the permutation and update */
-    CHAM_desc_t    Wl;         /**< Workspace used the update                     */
-    int           *proc_involved;
-    unsigned int   involved;
-    int            np_involved;
+    cham_getrf_t            alg;
+    cham_getrf_allreduce_t  alg_allreduce;
+    int                     ib;         /**< Internal blocking parameter */
+    int                     batch_size; /**< Batch size for the panel    */
+    int                     ringswitch; /**< Define when to switch to ring bcast           */
+    CHAM_desc_t             U;
+    CHAM_desc_t             Up; /**< Workspace used for the panel factorization    */
+    CHAM_desc_t             Wu; /**< Workspace used for the permutation and update */
+    CHAM_desc_t             Wl; /**< Workspace used the update                     */
+    int                    *proc_involved;
+    unsigned int            involved;
+    int                     np_involved;
 };
 
 /**

include/chameleon/constants.h

@@ -290,6 +290,13 @@ typedef enum chameleon_getrf_e {
     ChamGetrfPPivPerColumn  = 3,
 } cham_getrf_t;
 
+/**
+ * @brief Chameleon GETRF all reduce algorithm variants
+ */
+typedef enum chameleon_getrf_allreduce_e {
+    ChamStarPUTasks,
+} cham_getrf_allreduce_t;
+
 #define ChameleonTrd            1001
 #define ChameleonBrd            1002