starpu/codelet: New version of the potrf starpu codelet

runtime/starpu/codelets/codelet_zpotrf.c

@@ -29,11 +29,11 @@
 #include "runtime_codelet_z.h"
 
 struct cl_zpotrf_args_s {
-    cham_uplo_t uplo;
-    int n;
-    int iinfo;
+    cham_uplo_t         uplo;
+    int                 n;
+    int                 iinfo;
     RUNTIME_sequence_t *sequence;
-    RUNTIME_request_t *request;
+    RUNTIME_request_t  *request;
 };
 
 #if !defined(CHAMELEON_SIMULATION)
@@ -65,14 +65,15 @@ CODELETS( zpotrf, cl_zpotrf_cpu_func, cl_zpotrf_cuda_func, STARPU_CUDA_ASYNC )
 CODELETS_CPU( zpotrf, cl_zpotrf_cpu_func )
 #endif
 
+#if defined(CHAMELEON_STARPU_USE_INSERT)
 void INSERT_TASK_zpotrf( const RUNTIME_option_t *options,
                          cham_uplo_t uplo, int n, int nb,
                          const CHAM_desc_t *A, int Am, int An,
                          int iinfo )
 {
-    struct cl_zpotrf_args_s *clargs = NULL;
     void (*callback)(void*);
-    int                      exec = 0;
+    struct cl_zpotrf_args_s *clargs  = NULL;
+    int                      exec    = 0;
     const char              *cl_name = "zpotrf";
 
     /* Handle cache */
@@ -114,3 +115,69 @@ void INSERT_TASK_zpotrf( const RUNTIME_option_t *options,
 
     (void)nb;
 }
+
+#else
+
+void INSERT_TASK_zpotrf( const RUNTIME_option_t *options,
+                         cham_uplo_t uplo, int n, int nb,
+                         const CHAM_desc_t *A, int Am, int An,
+                         int iinfo )
+{
+    INSERT_TASK_COMMON_PARAMETERS( zpotrf, 1 );
+
+    /*
+     * Set the data handles and initialize exchanges if needed
+     */
+    starpu_cham_exchange_init_params( options, &params, A->get_rankof( A, Am, An ) );
+    starpu_cham_exchange_data_before_execution( options, params, &nbdata, descrs, A, Am, An, STARPU_RW );
+
+    /*
+     * Not involved, let's return
+     */
+    if ( nbdata == 0 ) {
+        return;
+    }
+
+    if ( params.do_execute )
+    {
+        int ret;
+        struct starpu_task *task = starpu_task_create();
+        task->cl = cl;
+
+        /* Set codelet parameters */
+        clargs = malloc( sizeof( struct cl_zpotrf_args_s ) );
+        clargs->uplo     = uplo;
+        clargs->n        = n;
+        clargs->iinfo    = iinfo;
+        clargs->sequence = options->sequence;
+        clargs->request  = options->request;
+
+        task->cl_arg      = clargs;
+        task->cl_arg_size = sizeof( struct cl_zpotrf_args_s );
+        task->cl_arg_free = 1;
+
+        /* Set common parameters */
+        starpu_cham_task_set_options( options, task, nbdata, descrs, cl_zpotrf_callback );
+
+        /* Flops */
+        task->flops = flops_zpotrf( n );
+
+        /* Refine name */
+        task->name = chameleon_codelet_name( cl_name, 1,
+                                             A->get_blktile( A, Am, An ) );
+
+        ret = starpu_task_submit( task );
+        if ( ret == -ENODEV ) {
+            task->destroy = 0;
+            starpu_task_destroy( task );
+            chameleon_error( "INSERT_TASK_zpotrf", "Failed to submit the task to StarPU" );
+            return;
+        }
+    }
+
+    starpu_cham_task_exchange_data_after_execution( options, params, nbdata, descrs );
+
+    (void)nb;
+}
+
+#endif