check the size of data to decide where to execute the task for those involved in POTRS and GEQRS

compute/zgetrs_nopiv.c

@@ -218,7 +218,7 @@ int MORSE_zgetrs_nopiv_Tile(MORSE_desc_t *A, MORSE_desc_t *B)
     RUNTIME_barrier(morse);
     RUNTIME_desc_getoncpu(A);
     RUNTIME_desc_getoncpu(B);
-    
+
     status = sequence->status;
     morse_sequence_destroy(morse, sequence);
     return status;

runtime/starpu/codelets/codelet_zgemm.c

@@ -47,9 +47,9 @@ void MORSE_TASK_zgemm(MORSE_option_t *options,
     (void)nb;
     struct starpu_codelet *codelet = &cl_zgemm;
     void (*callback)(void*) = options->profiling ? cl_zgemm_callback : NULL;
-    int sizeA = m*k;
-    int sizeB = k*n;
-    int sizeC = m*n;
+    int sizeA = lda*k;
+    int sizeB = ldb*n;
+    int sizeC = ldc*n;
     int execution_rank = C->get_rankof( C, Cm, Cn );
     int rank_changed=0;
 

runtime/starpu/codelets/codelet_ztrsm.c

@@ -46,8 +46,8 @@ void MORSE_TASK_ztrsm(MORSE_option_t *options,
     (void)nb;
     struct starpu_codelet *codelet = &cl_ztrsm;
     void (*callback)(void*) = options->profiling ? cl_ztrsm_callback : NULL;
-    int sizeA = m*m;
-    int sizeB = m*n;
+    int sizeA = lda*m;
+    int sizeB = ldb*n;
     int execution_rank = B->get_rankof( B, Bm, Bn );
     int rank_changed=0;
 

runtime/starpu/codelets/codelet_ztsmqr.c

@@ -144,11 +144,36 @@ void MORSE_TASK_ztsmqr(MORSE_option_t *options,
     struct starpu_codelet *codelet = &cl_ztsmqr;
     void (*callback)(void*) = options->profiling ? cl_ztsmqr_callback : NULL;
     int ldwork = side == MorseLeft ? ib : nb;
+    int sizeA1 = lda1*n1;
+    int sizeA2 = lda2*n2;
+    int sizeV = ldv*k;
+    int sizeT = ldt*n1;
+    int execution_rank = A2->get_rankof( A2, A2m, A2n );
+    int rank_changed=0;
+
+    // force execution on the rank owning the largest data (tile)
+    int threshold;
+    char* env = getenv("MORSE_COMM_FACTOR_THRESHOLD");
+    if (env != NULL)
+        threshold = (unsigned)atoi(env);
+    else
+        threshold = 10;
+    if ( sizeA1 > threshold*sizeA2 ){
+        execution_rank = A1->get_rankof( A1, A1m, A1n );
+        rank_changed = 1;
+    }else if( sizeV > threshold*sizeA2 ){
+        execution_rank = V->get_rankof( V, Vm, Vn );
+        rank_changed = 1;
+    }else if( sizeT > threshold*sizeA2 ){
+        execution_rank = T->get_rankof( T, Tm, Tn );
+        rank_changed = 1;
+    }
 
     if ( morse_desc_islocal( A1, A1m, A1n ) ||
          morse_desc_islocal( A2, A2m, A2n ) ||
          morse_desc_islocal( V,  Vm,  Vn  ) ||
-         morse_desc_islocal( T,  Tm,  Tn  ) )
+         morse_desc_islocal( T,  Tm,  Tn  ) ||
+         rank_changed )
     {
         starpu_insert_task(
             codelet,
@@ -173,7 +198,9 @@ void MORSE_TASK_ztsmqr(MORSE_option_t *options,
             STARPU_VALUE,    &ldwork,            sizeof(int),
             STARPU_PRIORITY,  options->priority,
             STARPU_CALLBACK,  callback,
-            STARPU_EXECUTE_ON_NODE, A2->get_rankof(A2, A2m, A2n),
+#if defined(CHAMELEON_USE_MPI)
+            STARPU_EXECUTE_ON_NODE, execution_rank,
+#endif
             0);
     }
 }

runtime/starpu/codelets/codelet_zttmqr.c

@@ -137,11 +137,36 @@ void MORSE_TASK_zttmqr(MORSE_option_t *options,
     struct starpu_codelet *codelet = &cl_zttmqr;
     void (*callback)(void*) = options->profiling ? cl_zttmqr_callback : NULL;
     int ldwork = side == MorseLeft ? ib : nb;
+    int sizeA1 = lda1*n1;
+    int sizeA2 = lda2*n2;
+    int sizeV = ldv*k;
+    int sizeT = ldt*n1;
+    int execution_rank = A2->get_rankof( A2, A2m, A2n );
+    int rank_changed=0;
+
+    // force execution on the rank owning the largest data (tile)
+    int threshold;
+    char* env = getenv("MORSE_COMM_FACTOR_THRESHOLD");
+    if (env != NULL)
+        threshold = (unsigned)atoi(env);
+    else
+        threshold = 10;
+    if ( sizeA1 > threshold*sizeA2 ){
+        execution_rank = A1->get_rankof( A1, A1m, A1n );
+        rank_changed = 1;
+    }else if( sizeV > threshold*sizeA2 ){
+        execution_rank = V->get_rankof( V, Vm, Vn );
+        rank_changed = 1;
+    }else if( sizeT > threshold*sizeA2 ){
+        execution_rank = T->get_rankof( T, Tm, Tn );
+        rank_changed = 1;
+    }
 
     if ( morse_desc_islocal( A1, A1m, A1n ) ||
          morse_desc_islocal( A2, A2m, A2n ) ||
          morse_desc_islocal( V,  Vm,  Vn  ) ||
-         morse_desc_islocal( T,  Tm,  Tn  ) )
+         morse_desc_islocal( T,  Tm,  Tn  ) ||
+         rank_changed )
     {
         starpu_insert_task(
             codelet,
@@ -166,6 +191,9 @@ void MORSE_TASK_zttmqr(MORSE_option_t *options,
             STARPU_VALUE,    &ldwork,            sizeof(int),
             STARPU_PRIORITY,  options->priority,
             STARPU_CALLBACK,  callback,
+#if defined(CHAMELEON_USE_MPI)
+            STARPU_EXECUTE_ON_NODE, execution_rank,
+#endif
             0);
     }
 }

runtime/starpu/codelets/codelet_zunmqr.c

@@ -126,10 +126,31 @@ void MORSE_TASK_zunmqr(MORSE_option_t *options,
 {
     struct starpu_codelet *codelet = &cl_zunmqr;
     void (*callback)(void*) = options->profiling ? cl_zunmqr_callback : NULL;
+    int sizeA = lda*k;
+    int sizeT = ldt*n;
+    int sizeC = ldc*n;
+    int execution_rank = C->get_rankof( C, Cm, Cn );
+    int rank_changed=0;
+
+    // force execution on the rank owning the largest data (tile)
+    int threshold;
+    char* env = getenv("MORSE_COMM_FACTOR_THRESHOLD");
+    if (env != NULL)
+        threshold = (unsigned)atoi(env);
+    else
+        threshold = 10;
+    if ( sizeA > threshold*sizeC ){
+        execution_rank = A->get_rankof( A, Am, An );
+        rank_changed = 1;
+    }else if( sizeT > threshold*sizeC ){
+        execution_rank = T->get_rankof( T, Tm, Tn );
+        rank_changed = 1;
+    }
 
     if ( morse_desc_islocal( A, Am, An ) ||
          morse_desc_islocal( T, Tm, Tn ) ||
-         morse_desc_islocal( C, Cm, Cn ) )
+         morse_desc_islocal( C, Cm, Cn ) ||
+         rank_changed )
     {
         starpu_insert_task(
             codelet,
@@ -150,6 +171,9 @@ void MORSE_TASK_zunmqr(MORSE_option_t *options,
             STARPU_VALUE,    &nb,                sizeof(int),
             STARPU_PRIORITY,  options->priority,
             STARPU_CALLBACK,  callback,
+#if defined(CHAMELEON_USE_MPI)
+            STARPU_EXECUTE_ON_NODE, execution_rank,
+#endif
             0);
     }
 }