Merge branch 'hotfix/qdwh' into 'master'

Hotfix/qdwh

See merge request solverstack/chameleon!139

compute/pzlange.c

@@ -517,7 +517,11 @@ void chameleon_pzlange_generic( cham_normtype_t norm, cham_uplo_t uplo, cham_dia
         }
     }
 
+    if ( wcol_init ) {
+        CHAMELEON_Desc_Flush( &Wcol, sequence );
+    }
     CHAMELEON_Desc_Flush( &Welt, sequence );
+    CHAMELEON_Desc_Flush( A, sequence );
     RUNTIME_sequence_wait(chamctxt, sequence);
 
     *result = *((double *)Welt.get_blkaddr( &Welt, A->myrank / A->q, A->myrank % A->q ));

compute/pzlansy.c

@@ -423,7 +423,11 @@ void chameleon_pzlansy_generic( cham_normtype_t norm, cham_uplo_t uplo, cham_tra
         }
     }
 
+    if ( wcol_init ) {
+        CHAMELEON_Desc_Flush( &Wcol, sequence );
+    }
     CHAMELEON_Desc_Flush( &Welt, sequence );
+    CHAMELEON_Desc_Flush( A, sequence );
     RUNTIME_sequence_wait(chamctxt, sequence);
 
     *result = *(double *)Welt.get_blkaddr( &Welt, A->myrank / A->q, A->myrank % A->q );

compute/zlange.c

@@ -188,7 +188,7 @@ double CHAMELEON_zlange(cham_normtype_t norm, int M, int N,
  * @sa CHAMELEON_slange_Tile
  *
  */
-double CHAMELEON_zlange_Tile(cham_normtype_t norm, CHAM_desc_t *A )
+double CHAMELEON_zlange_Tile( cham_normtype_t norm, CHAM_desc_t *A )
 {
     CHAM_context_t *chamctxt;
     RUNTIME_sequence_t *sequence = NULL;

include/chameleon/tasks_z.h

@@ -34,6 +34,10 @@ void INSERT_TASK_dzasum( const RUNTIME_option_t *options,
                          cham_store_t storev, cham_uplo_t uplo, int M, int N,
                          const CHAM_desc_t *A, int Am, int An, int lda,
                          const CHAM_desc_t *B, int Bm, int Bn );
+void INSERT_TASK_zaxpy( const RUNTIME_option_t *options,
+                        int M, CHAMELEON_Complex64_t alpha,
+                        const CHAM_desc_t *A, int Am, int An, int incA,
+                        const CHAM_desc_t *B, int Bm, int Bn, int incB );
 void INSERT_TASK_zgeadd( const RUNTIME_option_t *options,
                          cham_trans_t trans, int m, int n, int nb,
                          CHAMELEON_Complex64_t alpha, const CHAM_desc_t *A, int Am, int An, int lda,

runtime/parsec/codelets/codelet_zaxpy.c

@@ -49,11 +49,11 @@ void INSERT_TASK_zaxpy(const RUNTIME_option_t *options,
 
     parsec_dtd_taskpool_insert_task(
         PARSEC_dtd_taskpool, CORE_zaxpy_parsec, options->priority,  "axpy",
-        sizeof(int),               &M,     VALUE,
+        sizeof(int),                   &M,     VALUE,
         sizeof(CHAMELEON_Complex64_t), &alpha, VALUE,
-        PASSED_BY_REF,  RTBLKADDR( A, CHAMELEON_Complex64_t, Am, An ), chameleon_parsec_get_arena_index( A ) | INPUT,
-        sizeof(int),               &incA, VALUE,
-        PASSED_BY_REF,  RTBLKADDR( B, CHAMELEON_Complex64_t, Bm, Bn ), chameleon_parsec_get_arena_index( B ) | INOUT | AFFINITY,
-        sizeof(int),               &incB, VALUE,
+        PASSED_BY_REF, RTBLKADDR( A, CHAMELEON_Complex64_t, Am, An ), chameleon_parsec_get_arena_index( A ) | INPUT,
+        sizeof(int),                   &incA,  VALUE,
+        PASSED_BY_REF, RTBLKADDR( B, CHAMELEON_Complex64_t, Bm, Bn ), chameleon_parsec_get_arena_index( B ) | INOUT | AFFINITY,
+        sizeof(int),                   &incB,  VALUE,
         PARSEC_DTD_ARG_END );
 }

runtime/quark/codelets/codelet_zaxpy.c

@@ -37,7 +37,7 @@ void CORE_zaxpy_quark(Quark *quark)
 }
 
 void INSERT_TASK_zaxpy(const RUNTIME_option_t *options,
-                      int M, CHAMELEON_Complex64_t *alpha,
+                      int M, CHAMELEON_Complex64_t alpha,
                       const CHAM_desc_t *A, int Am, int An, int incA,
                       const CHAM_desc_t *B, int Bm, int Bn, int incB)
 {
@@ -45,10 +45,10 @@ void INSERT_TASK_zaxpy(const RUNTIME_option_t *options,
     DAG_CORE_AXPY;
     QUARK_Insert_Task(opt->quark, CORE_zaxpy_quark, (Quark_Task_Flags*)opt,
         sizeof(int),                        &M,         VALUE,
-        sizeof(CHAMELEON_Complex64_t),          alpha,      VALUE,
-        sizeof(CHAMELEON_Complex64_t)*M,        RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An), INPUT,
+        sizeof(CHAMELEON_Complex64_t),      &alpha,     VALUE,
+        sizeof(CHAMELEON_Complex64_t)*M,     RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An), INPUT,
         sizeof(int),                        &incA,      VALUE,
-        sizeof(CHAMELEON_Complex64_t)*M,        RTBLKADDR(B, CHAMELEON_Complex64_t, Bm, Bn), INOUT,
+        sizeof(CHAMELEON_Complex64_t)*M,     RTBLKADDR(B, CHAMELEON_Complex64_t, Bm, Bn), INOUT,
         sizeof(int),                        &incB,      VALUE,
         0);
 }

runtime/starpu/codelets/codelet_zaxpy.c

@@ -36,7 +36,7 @@ void INSERT_TASK_zaxpy(const RUNTIME_option_t *options,
     starpu_insert_task(
             starpu_mpi_codelet(codelet),
             STARPU_VALUE,    &M,                           sizeof(int),
-            STARPU_VALUE,    alpha,                       sizeof(CHAMELEON_Complex64_t),
+            STARPU_VALUE,    &alpha,                       sizeof(CHAMELEON_Complex64_t),
             STARPU_R,        RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
             STARPU_VALUE,    &incA,                        sizeof(int),
             STARPU_RW,       RTBLKADDR(B, CHAMELEON_Complex64_t, Bm, Bn),