starpu/codelet: Add new task submit to codelet_ztrsm.c

runtime/starpu/codelets/codelet_ztrsm.c

@@ -30,12 +30,12 @@
 #include "runtime_codelet_z.h"
 
 struct cl_ztrsm_args_s {
-    cham_side_t side;
-    cham_uplo_t uplo;
-    cham_trans_t transA;
-    cham_diag_t diag;
-    int m;
-    int n;
+    cham_side_t           side;
+    cham_uplo_t           uplo;
+    cham_trans_t          transA;
+    cham_diag_t           diag;
+    int                   m;
+    int                   n;
     CHAMELEON_Complex64_t alpha;
 };
 
@@ -113,15 +113,16 @@ CODELETS_GPU( ztrsm, cl_ztrsm_cpu_func, cl_ztrsm_hip_func, STARPU_HIP_ASYNC )
 CODELETS( ztrsm, cl_ztrsm_cpu_func, cl_ztrsm_cuda_func, STARPU_CUDA_ASYNC )
 #endif
 
+#if defined(CHAMELEON_STARPU_USE_INSERT)
 void INSERT_TASK_ztrsm( const RUNTIME_option_t *options,
                         cham_side_t side, cham_uplo_t uplo, cham_trans_t transA, cham_diag_t diag,
                         int m, int n, int nb,
                         CHAMELEON_Complex64_t alpha, const CHAM_desc_t *A, int Am, int An,
                         const CHAM_desc_t *B, int Bm, int Bn )
 {
-    struct cl_ztrsm_args_s  *clargs = NULL;
     void (*callback)(void*);
-    int                      exec = 0;
+    struct cl_ztrsm_args_s  *clargs  = NULL;
+    int                      exec    = 0;
     const char              *cl_name = "ztrsm";
 
     /* Handle cache */
@@ -168,3 +169,74 @@ void INSERT_TASK_ztrsm( const RUNTIME_option_t *options,
 
     (void)nb;
 }
+
+#else
+
+void INSERT_TASK_ztrsm( const RUNTIME_option_t *options,
+                        cham_side_t side, cham_uplo_t uplo, cham_trans_t transA, cham_diag_t diag,
+                        int m, int n, int nb,
+                        CHAMELEON_Complex64_t alpha, const CHAM_desc_t *A, int Am, int An,
+                        const CHAM_desc_t *B, int Bm, int Bn )
+{
+    INSERT_TASK_COMMON_PARAMETERS( ztrsm, 2 );
+
+    /*
+     * Set the data handles and initialize exchanges if needed
+     */
+    starpu_cham_exchange_init_params( options, &params, B->get_rankof( B, Bm, Bn ) );
+    starpu_cham_exchange_data_before_execution( options, params, &nbdata, descrs, A, Am, An, STARPU_R  );
+    starpu_cham_exchange_data_before_execution( options, params, &nbdata, descrs, B, Bm, Bn, 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_ztrsm_args_s ) );
+        clargs->side   = side;
+        clargs->uplo   = uplo;
+        clargs->transA = transA;
+        clargs->diag   = diag;
+        clargs->m      = m;
+        clargs->n      = n;
+        clargs->alpha  = alpha;
+
+        task->cl_arg      = clargs;
+        task->cl_arg_size = sizeof( struct cl_ztrsm_args_s );
+        task->cl_arg_free = 1;
+
+        /* Set common parameters */
+        starpu_cham_task_set_options( options, task, nbdata, descrs, cl_ztrsm_callback );
+
+        /* Flops */
+        task->flops = flops_ztrsm( side, m, n );
+
+        /* Refine name */
+        task->name = chameleon_codelet_name( cl_name, 2,
+                                             A->get_blktile( A, Am, An ),
+                                             B->get_blktile( B, Bm, Bn ) );
+
+        ret = starpu_task_submit( task );
+        if ( ret == -ENODEV ) {
+            task->destroy = 0;
+            starpu_task_destroy( task );
+            chameleon_error( "INSERT_TASK_ztrsm", "Failed to submit the task to StarPU" );
+            return;
+        }
+    }
+
+    starpu_cham_task_exchange_data_after_execution( options, params, nbdata, descrs );
+
+    (void)nb;
+}
+
+#endif