diff --git a/compute/zgetrf_nopiv.c b/compute/zgetrf_nopiv.c
index 1f45349147c331212cc6b72e4b802683fc69424e..3409e97f7a80cfb0844b4eefa66d6646be6239f1 100644
--- a/compute/zgetrf_nopiv.c
+++ b/compute/zgetrf_nopiv.c
@@ -167,6 +167,7 @@ int CHAMELEON_zgetrf_nopiv( int M, int N,
CHAM_context_t *chamctxt;
RUNTIME_sequence_t *sequence = NULL;
RUNTIME_request_t request = RUNTIME_REQUEST_INITIALIZER;
+ void *ws = NULL;
chamctxt = chameleon_context_self();
if (chamctxt == NULL) {
@@ -207,7 +208,8 @@ int CHAMELEON_zgetrf_nopiv( int M, int N,
A, NB, NB, LDA, N, M, N, sequence, &request );
/* Call the tile interface */
- CHAMELEON_zgetrf_nopiv_Tile_Async( &descAt, sequence, &request );
+ ws = CHAMELEON_zgetrf_nopiv_WS_Alloc( &descAt );
+ CHAMELEON_zgetrf_nopiv_Tile_Async( &descAt, ws, sequence, &request );
/* Submit the matrix conversion back */
chameleon_ztile2lap( chamctxt, &descAl, &descAt,
@@ -216,6 +218,7 @@ int CHAMELEON_zgetrf_nopiv( int M, int N,
chameleon_sequence_wait( chamctxt, sequence );
/* Cleanup the temporary data */
+ CHAMELEON_zgetrf_nopiv_WS_Free( ws );
chameleon_ztile2lap_cleanup( chamctxt, &descAl, &descAt );
status = sequence->status;
@@ -259,10 +262,11 @@ int CHAMELEON_zgetrf_nopiv( int M, int N,
*/
int CHAMELEON_zgetrf_nopiv_Tile( CHAM_desc_t *A )
{
- CHAM_context_t *chamctxt;
+ CHAM_context_t *chamctxt;
RUNTIME_sequence_t *sequence = NULL;
- RUNTIME_request_t request = RUNTIME_REQUEST_INITIALIZER;
- int status;
+ RUNTIME_request_t request = RUNTIME_REQUEST_INITIALIZER;
+ int status;
+ void *ws;
chamctxt = chameleon_context_self();
if (chamctxt == NULL) {
@@ -271,11 +275,14 @@ int CHAMELEON_zgetrf_nopiv_Tile( CHAM_desc_t *A )
}
chameleon_sequence_create( chamctxt, &sequence );
- CHAMELEON_zgetrf_nopiv_Tile_Async( A, sequence, &request );
+ ws = CHAMELEON_zgetrf_nopiv_WS_Alloc( A );
+ CHAMELEON_zgetrf_nopiv_Tile_Async( A, ws, sequence, &request );
CHAMELEON_Desc_Flush( A, sequence );
chameleon_sequence_wait( chamctxt, sequence );
+ CHAMELEON_zgetrf_nopiv_WS_Free( ws );
+
status = sequence->status;
chameleon_sequence_destroy( chamctxt, sequence );
return status;
@@ -314,11 +321,13 @@ int CHAMELEON_zgetrf_nopiv_Tile( CHAM_desc_t *A )
* @sa CHAMELEON_zgetrs_Tile_Async
*
*/
-int CHAMELEON_zgetrf_nopiv_Tile_Async( CHAM_desc_t *A,
- RUNTIME_sequence_t *sequence,
- RUNTIME_request_t *request )
+int CHAMELEON_zgetrf_nopiv_Tile_Async( CHAM_desc_t *A,
+ void *user_ws,
+ RUNTIME_sequence_t *sequence,
+ RUNTIME_request_t *request )
{
- CHAM_context_t *chamctxt;
+ CHAM_context_t *chamctxt;
+ struct chameleon_pzgetrf_nopiv_s *ws;
chamctxt = chameleon_context_self();
if (chamctxt == NULL) {
@@ -353,7 +362,19 @@ int CHAMELEON_zgetrf_nopiv_Tile_Async( CHAM_desc_t *A,
return chameleon_request_fail(sequence, request, CHAMELEON_ERR_ILLEGAL_VALUE);
}
- chameleon_pzgetrf_nopiv( NULL, A, sequence, request );
+ if ( user_ws == NULL ) {
+ ws = CHAMELEON_zgetrf_nopiv_WS_Alloc( A );
+ }
+ else {
+ ws = user_ws;
+ }
+ chameleon_pzgetrf_nopiv( ws, A, sequence, request );
+
+ if ( user_ws == NULL ) {
+ CHAMELEON_Desc_Flush( A, sequence );
+ chameleon_sequence_wait( chamctxt, sequence );
+ CHAMELEON_zgetrf_nopiv_WS_Free( ws );
+ }
return CHAMELEON_SUCCESS;
}
diff --git a/control/chameleon_zf77.c b/control/chameleon_zf77.c
index 70541b792b2b3ee527c1737af4f314a86a7f57d3..64fad284a4f118b79bef98e0c622f415e3a2ac0a 100644
--- a/control/chameleon_zf77.c
+++ b/control/chameleon_zf77.c
@@ -11,7 +11,7 @@
*
* @brief Chameleon Fortran77 computational routines
*
- * @version 1.2.0
+ * @version 1.3.0
* @comment This file has been automatically generated
* from Plasma 2.5.0 for CHAMELEON 0.9.2
* @comment This file is automatically generated by tools/genf77interface.pl
@@ -21,7 +21,8 @@
* @author Cedric Castagnede
* @author Florent Pruvost
* @author Alycia Lisito
- * @date 2022-02-22
+ * @author Matthieu Kuhn
+ * @date 2024-10-17
* @precisions normal z -> c d s
*
*/
@@ -727,7 +728,7 @@ void CHAMELEON_ZGETRF_INCPIV_TILE_ASYNC(CHAM_desc_t *A, CHAM_desc_t *L, int *IPI
{ *info = CHAMELEON_zgetrf_incpiv_Tile_Async(A, L, IPIV, sequence, request); }
void CHAMELEON_ZGETRF_NOPIV_TILE_ASYNC(CHAM_desc_t *A, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request, int *info)
-{ *info = CHAMELEON_zgetrf_nopiv_Tile_Async(A, sequence, request); }
+{ *info = CHAMELEON_zgetrf_nopiv_Tile_Async(A, NULL, sequence, request); }
//void CHAMELEON_ZGETRI_TILE_ASYNC(CHAM_desc_t *A, int *IPIV, CHAM_desc_t *W, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request, int *info)
//{ *info = CHAMELEON_zgetri_Tile_Async(A, IPIV, W, sequence, request); }
diff --git a/include/chameleon/chameleon_z.h b/include/chameleon/chameleon_z.h
index 9bd22083b926e2e51051848557500a5c83571e47..c139111befda8146f20194193b908b87b5548fa1 100644
--- a/include/chameleon/chameleon_z.h
+++ b/include/chameleon/chameleon_z.h
@@ -24,7 +24,7 @@
* @author Alycia Lisito
* @author Matthieu Kuhn
* @author Ana Hourcau
- * @date 2024-07-17
+ * @date 2024-10-17
* @precisions normal z -> c d s
*
*/
@@ -215,7 +215,7 @@ int CHAMELEON_zgesv_nopiv_Tile_Async(CHAM_desc_t *A, CHAM_desc_t *B, RUNTIME_seq
int CHAMELEON_zgesvd_Tile_Async(cham_job_t jobu, cham_job_t jobvt, CHAM_desc_t *A, double *S, CHAM_desc_t *T, CHAMELEON_Complex64_t *U, int LDU, CHAMELEON_Complex64_t *VT, int LDVT, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
//int CHAMELEON_zgetrf_Tile_Async(CHAM_desc_t *A, int *IPIV, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
int CHAMELEON_zgetrf_incpiv_Tile_Async(CHAM_desc_t *A, CHAM_desc_t *L, int *IPIV, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
-int CHAMELEON_zgetrf_nopiv_Tile_Async(CHAM_desc_t *A, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
+int CHAMELEON_zgetrf_nopiv_Tile_Async(CHAM_desc_t *A, void * ws, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
int CHAMELEON_zgetrf_Tile_Async( CHAM_desc_t *A, CHAM_ipiv_t *IPIV, void *ws, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request );
//int CHAMELEON_zgetri_Tile_Async(CHAM_desc_t *A, int *IPIV, CHAM_desc_t *W, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
//int CHAMELEON_zgetrs_Tile_Async(cham_trans_t trans, CHAM_desc_t *A, int *IPIV, CHAM_desc_t *B, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
diff --git a/testing/testing_zgetrf_nopiv.c b/testing/testing_zgetrf_nopiv.c
index afa44ede3d749a4caa695a8f780ee9f30615e793..657513eaf26371399a7402e1e0b305afdc8fb854 100644
--- a/testing/testing_zgetrf_nopiv.c
+++ b/testing/testing_zgetrf_nopiv.c
@@ -13,7 +13,9 @@
* @author Lucas Barros de Assis
* @author Mathieu Faverge
* @author Alycia Lisito
- * @date 2023-07-05
+ * @author Lucas Barros De Assis
+ * @author Matthieu Kuhn
+ * @date 2024-10-17
* @precisions normal z -> c d s
*
*/
@@ -53,7 +55,7 @@ testing_zgetrf_nopiv_desc( run_arg_list_t *args, int check )
/* Calculates the solution */
testing_start( &test_data );
if ( async ) {
- hres = CHAMELEON_zgetrf_nopiv_Tile_Async( descA, test_data.sequence, &test_data.request );
+ hres = CHAMELEON_zgetrf_nopiv_Tile_Async( descA, NULL, test_data.sequence, &test_data.request );
CHAMELEON_Desc_Flush( descA, test_data.sequence );
}
else {