getrf: Add a ringswitch parameter to define when to switch from classic bcast to ring bcast

compute/zgetrf.c

@@ -27,6 +27,7 @@
  *
  */
 #include "control/common.h"
+#include <limits.h>
 
 /**
  ********************************************************************************
@@ -103,6 +104,8 @@ CHAMELEON_zgetrf_WS_Alloc( const CHAM_desc_t *A )
         ws->batch_size = CHAMELEON_BATCH_SIZE;
     }
 
+    ws->ringswitch = chameleon_getenv_get_value_int( "CHAMELEON_GETRF_RINGSWITCH", INT_MAX );
+
     /* Allocation of U for permutation of the panels */
     if ( ws->alg == ChamGetrfNoPivPerColumn ) {
         chameleon_desc_init( &(ws->U), CHAMELEON_MAT_ALLOC_TILE,

control/compute_z.h

@@ -44,12 +44,13 @@ struct chameleon_pzgemm_s {
  */
 struct chameleon_pzgetrf_s {
     cham_getrf_t   alg;
-    int            ib;         /**< Internal blocking parameter */
-    int            batch_size; /**< Batch size for the panel    */
+    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                     */
+    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;