diff --git a/include/chameleon/tasks_z.h b/include/chameleon/tasks_z.h
index f444409b8279007d43d810f6b603595c414f9819..90b4578d47eadeab248d9d47cb45a4a93b74a1b2 100644
--- a/include/chameleon/tasks_z.h
+++ b/include/chameleon/tasks_z.h
@@ -25,7 +25,7 @@
* @author Romain Peressoni
* @author Matthieu Kuhn
* @author Ana Hourcau
- * @date 2024-11-12
+ * @date 2025-03-24
* @precisions normal z -> c d s
*
*/
@@ -188,12 +188,12 @@ void INSERT_TASK_zlaset( const RUNTIME_option_t *options,
void INSERT_TASK_zlaset2( const RUNTIME_option_t *options,
cham_uplo_t uplo, int n1, int n2, CHAMELEON_Complex64_t alpha,
const CHAM_desc_t *tileA, int tileAm, int tileAn );
-void INSERT_TASK_zlaswp_get( const RUNTIME_option_t *options,
+void INSERT_TASK_zlaswp_get( const RUNTIME_option_t *options, cham_dir_t dir,
int m0, int k,
const CHAM_ipiv_t *tIPIV, int tIPIVk,
const CHAM_desc_t *tileA, int tileAm, int tileAn,
const CHAM_desc_t *tileB, int tileBm, int tileBn );
-void INSERT_TASK_zlaswp_set( const RUNTIME_option_t *options,
+void INSERT_TASK_zlaswp_set( const RUNTIME_option_t *options, cham_dir_t dir,
int m0, int k,
const CHAM_ipiv_t *tIPIV, int tIPIVk,
const CHAM_desc_t *tileA, int tileAm, int tileAn,
@@ -588,15 +588,20 @@ void INSERT_TASK_zipiv_allreduce( const RUNTIME_option_t *options,
*
* @ingroup CHAMELEON_Complex64_t
*
- * INSERT_TASK_zperm_allreduce - Perfoms an allreduce operation on the tile
- * U(Um, Un) according to the permutation ipiv. This task is used in the LU
- * factorization with partial pivoting.
+ * @brief Perfoms an allreduce operation on the tile
+ * U(Um, Un) according to the permutation ipiv. This task is used in the LU
+ * factorization with partial pivoting.
*
*******************************************************************************
*
* @param[in] options
* The runtime options data structure to pass through all insert_task calls.
*
+ * @param[in] dir
+ * Specifies the order of the permutation.
+ * = ChamDirForward: Natural order
+ * = ChamDirBackward: Reverse order
+ *
* @param[in] A
* The descriptor of the matrix A.
*
@@ -630,6 +635,7 @@ void INSERT_TASK_zipiv_allreduce( const RUNTIME_option_t *options,
*******************************************************************************
*/
void INSERT_TASK_zperm_allreduce( const RUNTIME_option_t *options,
+ cham_dir_t dir,
const CHAM_desc_t *A,
CHAM_desc_t *U,
int Um,
@@ -645,9 +651,9 @@ void INSERT_TASK_zperm_allreduce( const RUNTIME_option_t *options,
*
* @ingroup CHAMELEON_Complex64_t
*
- * INSERT_TASK_zperm_allreduce_send_A - Sends the tile A(Am, An) to the processus
- * involved in the permutation. This task is used in the LU factorization with
- * partial pivoting.
+ * @brief Sends the tile A(Am, An) to the processus
+ * involved in the permutation. This task is used in the LU factorization with
+ * partial pivoting.
*
*******************************************************************************
*
@@ -687,15 +693,20 @@ void INSERT_TASK_zperm_allreduce_send_A( const RUNTIME_option_t *options,
*
* @ingroup CHAMELEON_Complex64_t
*
- * INSERT_TASK_zperm_allreduce_send_perm - Sends the permutation ipivk to the
- * processus involved in the permutation. This task is used in the LU
- * factorization with partial pivoting.
+ * @brief - Sends the permutation ipivk to the
+ * processus involved in the permutation. This task is used in the LU
+ * factorization with partial pivoting.
*
*******************************************************************************
*
* @param[in] options
* The runtime options data structure to pass through all insert_task calls.
*
+ * @param[in] dir
+ * Specifies the order of the permutation.
+ * = ChamDirForward: Natural order
+ * = ChamDirBackward: Reverse order
+ *
* @param[in] ipiv
* The pivot structure that contains the informations for the LU
* factorization with partial pivoting.
@@ -715,6 +726,7 @@ void INSERT_TASK_zperm_allreduce_send_A( const RUNTIME_option_t *options,
*******************************************************************************
*/
void INSERT_TASK_zperm_allreduce_send_perm( const RUNTIME_option_t *options,
+ cham_dir_t dir,
CHAM_ipiv_t *ipiv,
int ipivk,
int myrank,
@@ -726,15 +738,20 @@ void INSERT_TASK_zperm_allreduce_send_perm( const RUNTIME_option_t *options,
*
* @ingroup CHAMELEON_Complex64_t
*
- * INSERT_TASK_zperm_allreduce_send_invp - Sends the inverse permutation ipivk
- * to the processus involved in the permutation. This task is used in the LU
- * factorization with partial pivoting.
+ * @brief Sends the inverse permutation ipivk
+ * to the processus involved in the permutation. This task is used in the LU
+ * factorization with partial pivoting.
*
*******************************************************************************
*
* @param[in] options
* The runtime options data structure to pass through all insert_task calls.
*
+ * @param[in] dir
+ * Specifies the order of the permutation.
+ * = ChamDirForward: Natural order
+ * = ChamDirBackward: Reverse order
+ *
* @param[in] ipiv
* The pivot structure that contains the informations for the LU
* factorization with partial pivoting.
@@ -754,6 +771,7 @@ void INSERT_TASK_zperm_allreduce_send_perm( const RUNTIME_option_t *options,
*******************************************************************************
*/
void INSERT_TASK_zperm_allreduce_send_invp( const RUNTIME_option_t *options,
+ cham_dir_t dir,
CHAM_ipiv_t *ipiv,
int ipivk,
const CHAM_desc_t *A,
@@ -761,3 +779,4 @@ void INSERT_TASK_zperm_allreduce_send_invp( const RUNTIME_option_t *options,
int n );
#endif /* _chameleon_tasks_z_h_ */
+
diff --git a/runtime/openmp/codelets/codelet_zlaswp.c b/runtime/openmp/codelets/codelet_zlaswp.c
index bce58c771ef3052ce4d20d16232082cd9a746f66..93bf20aef11964fa548adb7739b000af575b04ba 100644
--- a/runtime/openmp/codelets/codelet_zlaswp.c
+++ b/runtime/openmp/codelets/codelet_zlaswp.c
@@ -11,7 +11,7 @@
*
* @version 1.3.0
* @author Mathieu Faverge
- * @date 2024-02-18
+ * @date 2025-03-24
* @precisions normal z -> c d s
*
*/
@@ -20,7 +20,7 @@
#include "coreblas/coreblas_ztile.h"
void INSERT_TASK_zlaswp_get( const RUNTIME_option_t *options,
- int m0, int k,
+ cham_dir_t dir, int m0, int k,
const CHAM_ipiv_t *ipiv, int ipivk,
const CHAM_desc_t *A, int Am, int An,
const CHAM_desc_t *U, int Um, int Un )
@@ -38,10 +38,11 @@ void INSERT_TASK_zlaswp_get( const RUNTIME_option_t *options,
}
(void)options;
+ (void)dir;
}
void INSERT_TASK_zlaswp_set( const RUNTIME_option_t *options,
- int m0, int k,
+ cham_dir_t dir, int m0, int k,
const CHAM_ipiv_t *ipiv, int ipivk,
const CHAM_desc_t *A, int Am, int An,
const CHAM_desc_t *B, int Bm, int Bn )
@@ -59,4 +60,5 @@ void INSERT_TASK_zlaswp_set( const RUNTIME_option_t *options,
}
(void)options;
+ (void)dir;
}
diff --git a/runtime/openmp/codelets/codelet_zperm_allreduce.c b/runtime/openmp/codelets/codelet_zperm_allreduce.c
index eac34fdfd1f8a0814c277f7acb8a9b85cb594ec7..8b20a60fd43332dac7373edcb2de40ee552d050a 100644
--- a/runtime/openmp/codelets/codelet_zperm_allreduce.c
+++ b/runtime/openmp/codelets/codelet_zperm_allreduce.c
@@ -11,7 +11,7 @@
*
* @version 1.3.0
* @author Alycia Lisito
- * @date 2024-11-12
+ * @date 2025-03-24
* @precisions normal z -> c d s
*
*/
@@ -38,6 +38,7 @@ INSERT_TASK_zperm_allreduce_send_A( const RUNTIME_option_t *options,
void
INSERT_TASK_zperm_allreduce_send_perm( const RUNTIME_option_t *options,
+ cham_dir_t dir,
CHAM_ipiv_t *ipiv,
int ipivk,
int myrank,
@@ -45,6 +46,7 @@ INSERT_TASK_zperm_allreduce_send_perm( const RUNTIME_option_t *options,
int *proc_involved )
{
(void)options;
+ (void)dir;
(void)ipiv;
(void)ipivk;
(void)myrank;
@@ -54,6 +56,7 @@ INSERT_TASK_zperm_allreduce_send_perm( const RUNTIME_option_t *options,
void
INSERT_TASK_zperm_allreduce_send_invp( const RUNTIME_option_t *options,
+ cham_dir_t dir,
CHAM_ipiv_t *ipiv,
int ipivk,
const CHAM_desc_t *A,
@@ -61,6 +64,7 @@ INSERT_TASK_zperm_allreduce_send_invp( const RUNTIME_option_t *options,
int n )
{
(void)options;
+ (void)dir;
(void)ipiv;
(void)ipivk;
(void)A;
@@ -70,6 +74,7 @@ INSERT_TASK_zperm_allreduce_send_invp( const RUNTIME_option_t *options,
void
INSERT_TASK_zperm_allreduce( const RUNTIME_option_t *options,
+ cham_dir_t dir,
const CHAM_desc_t *A,
CHAM_desc_t *U,
int Um,
@@ -81,6 +86,7 @@ INSERT_TASK_zperm_allreduce( const RUNTIME_option_t *options,
void *ws )
{
(void)options;
+ (void)dir;
(void)A;
(void)U;
(void)Um;
diff --git a/runtime/parsec/codelets/codelet_zlaswp.c b/runtime/parsec/codelets/codelet_zlaswp.c
index 12aaf7089ff41f4e4090e0fb6f18e518c9813fd3..65849c96d1aae96cc1000dd93e5efbebe481c7d9 100644
--- a/runtime/parsec/codelets/codelet_zlaswp.c
+++ b/runtime/parsec/codelets/codelet_zlaswp.c
@@ -11,7 +11,7 @@
*
* @version 1.3.0
* @author Mathieu Faverge
- * @date 2024-02-18
+ * @date 2025-03-24
* @precisions normal z -> c d s
*
*/
@@ -33,7 +33,7 @@ CORE_zlaswp_get_parsec( parsec_execution_stream_t *context,
}
void INSERT_TASK_zlaswp_get( const RUNTIME_option_t *options,
- int m0, int k,
+ cham_dir_t dir, int m0, int k,
const CHAM_ipiv_t *ipiv, int ipivk,
const CHAM_desc_t *A, int Am, int An,
const CHAM_desc_t *U, int Um, int Un )
@@ -54,6 +54,8 @@ void INSERT_TASK_zlaswp_get( const RUNTIME_option_t *options,
sizeof(int), &(tileU->ld), VALUE,
PASSED_BY_REF, RUNTIME_perm_getaddr( ipiv, ipivk ), chameleon_parsec_get_arena_index_perm( ipiv ) | INPUT,
PARSEC_DTD_ARG_END );
+
+ (void)dir;
}
static inline int
@@ -70,7 +72,7 @@ CORE_zlaswp_set_parsec( parsec_execution_stream_t *context,
}
void INSERT_TASK_zlaswp_set( const RUNTIME_option_t *options,
- int m0, int k,
+ cham_dir_t dir, int m0, int k,
const CHAM_ipiv_t *ipiv, int ipivk,
const CHAM_desc_t *A, int Am, int An,
const CHAM_desc_t *B, int Bm, int Bn )
@@ -91,4 +93,6 @@ void INSERT_TASK_zlaswp_set( const RUNTIME_option_t *options,
sizeof(int), &(tileB->ld), VALUE,
PASSED_BY_REF, RUNTIME_invp_getaddr( ipiv, ipivk ), chameleon_parsec_get_arena_index_invp( ipiv ) | INPUT,
PARSEC_DTD_ARG_END );
+
+ (void)dir;
}
diff --git a/runtime/parsec/codelets/codelet_zperm_allreduce.c b/runtime/parsec/codelets/codelet_zperm_allreduce.c
index 9ceb440c8a4e677630a68355daa7defda7f904fa..f68148e24b5e4c4e7d42d7248c8bf4a9948477c6 100644
--- a/runtime/parsec/codelets/codelet_zperm_allreduce.c
+++ b/runtime/parsec/codelets/codelet_zperm_allreduce.c
@@ -11,7 +11,7 @@
*
* @version 1.3.0
* @author Alycia Lisito
- * @date 2024-11-12
+ * @date 2025-03-24
* @precisions normal z -> c d s
*
*/
@@ -38,6 +38,7 @@ INSERT_TASK_zperm_allreduce_send_A( const RUNTIME_option_t *options,
void
INSERT_TASK_zperm_allreduce_send_perm( const RUNTIME_option_t *options,
+ cham_dir_t dir,
CHAM_ipiv_t *ipiv,
int ipivk,
int myrank,
@@ -45,6 +46,7 @@ INSERT_TASK_zperm_allreduce_send_perm( const RUNTIME_option_t *options,
int *proc_involved )
{
(void)options;
+ (void)dir;
(void)ipiv;
(void)ipivk;
(void)myrank;
@@ -54,6 +56,7 @@ INSERT_TASK_zperm_allreduce_send_perm( const RUNTIME_option_t *options,
void
INSERT_TASK_zperm_allreduce_send_invp( const RUNTIME_option_t *options,
+ cham_dir_t dir,
CHAM_ipiv_t *ipiv,
int ipivk,
const CHAM_desc_t *A,
@@ -61,6 +64,7 @@ INSERT_TASK_zperm_allreduce_send_invp( const RUNTIME_option_t *options,
int n )
{
(void)options;
+ (void)dir;
(void)ipiv;
(void)ipivk;
(void)A;
@@ -70,6 +74,7 @@ INSERT_TASK_zperm_allreduce_send_invp( const RUNTIME_option_t *options,
void
INSERT_TASK_zperm_allreduce( const RUNTIME_option_t *options,
+ cham_dir_t dir,
const CHAM_desc_t *A,
CHAM_desc_t *U,
int Um,
@@ -81,6 +86,7 @@ INSERT_TASK_zperm_allreduce( const RUNTIME_option_t *options,
void *ws )
{
(void)options;
+ (void)dir;
(void)A;
(void)U;
(void)Um;
diff --git a/runtime/quark/codelets/codelet_zlaswp.c b/runtime/quark/codelets/codelet_zlaswp.c
index 176dd16916eb51e1b698ad0d17dbd0d37c1a1d61..8f5a1b57fd52bd2e401273171584ebcca1478e50 100644
--- a/runtime/quark/codelets/codelet_zlaswp.c
+++ b/runtime/quark/codelets/codelet_zlaswp.c
@@ -11,7 +11,7 @@
*
* @version 1.3.0
* @author Mathieu Faverge
- * @date 2024-02-18
+ * @date 2025-03-24
* @precisions normal z -> c d s
*
*/
@@ -30,7 +30,7 @@ static void CORE_zlaswp_get_quark( Quark *quark )
}
void INSERT_TASK_zlaswp_get( const RUNTIME_option_t *options,
- int m0, int k,
+ cham_dir_t dir, int m0, int k,
const CHAM_ipiv_t *ipiv, int ipivk,
const CHAM_desc_t *A, int Am, int An,
const CHAM_desc_t *U, int Um, int Un )
@@ -46,6 +46,8 @@ void INSERT_TASK_zlaswp_get( const RUNTIME_option_t *options,
sizeof(CHAM_tile_t*), RTBLKADDR(A, ChamComplexDouble, Am, An), INPUT,
sizeof(CHAM_tile_t*), RTBLKADDR(U, ChamComplexDouble, Um, Un), INOUT,
0 );
+
+ (void)dir;
}
static void CORE_zlaswp_set_quark( Quark *quark )
@@ -59,7 +61,7 @@ static void CORE_zlaswp_set_quark( Quark *quark )
}
void INSERT_TASK_zlaswp_set( const RUNTIME_option_t *options,
- int m0, int k,
+ cham_dir_t dir, int m0, int k,
const CHAM_ipiv_t *ipiv, int ipivk,
const CHAM_desc_t *A, int Am, int An,
const CHAM_desc_t *B, int Bm, int Bn )
@@ -75,4 +77,6 @@ void INSERT_TASK_zlaswp_set( const RUNTIME_option_t *options,
sizeof(CHAM_tile_t*), RTBLKADDR(A, ChamComplexDouble, Am, An), INPUT,
sizeof(CHAM_tile_t*), RTBLKADDR(B, ChamComplexDouble, Bm, Bn), INOUT,
0 );
+
+ (void)dir;
}
diff --git a/runtime/quark/codelets/codelet_zperm_allreduce.c b/runtime/quark/codelets/codelet_zperm_allreduce.c
index f297d343b33455ba6340f0b81c45e8d01d29600f..1a2a7089c8addc5715d074a6c04bc5e8732aed1b 100644
--- a/runtime/quark/codelets/codelet_zperm_allreduce.c
+++ b/runtime/quark/codelets/codelet_zperm_allreduce.c
@@ -11,7 +11,7 @@
*
* @version 1.3.0
* @author Alycia Lisito
- * @date 2024-11-12
+ * @date 2025-03-24
* @precisions normal z -> c d s
*
*/
@@ -38,6 +38,7 @@ INSERT_TASK_zperm_allreduce_send_A( const RUNTIME_option_t *options,
void
INSERT_TASK_zperm_allreduce_send_perm( const RUNTIME_option_t *options,
+ cham_dir_t dir,
CHAM_ipiv_t *ipiv,
int ipivk,
int myrank,
@@ -45,6 +46,7 @@ INSERT_TASK_zperm_allreduce_send_perm( const RUNTIME_option_t *options,
int *proc_involved )
{
(void)options;
+ (void)dir;
(void)ipiv;
(void)ipivk;
(void)myrank;
@@ -54,6 +56,7 @@ INSERT_TASK_zperm_allreduce_send_perm( const RUNTIME_option_t *options,
void
INSERT_TASK_zperm_allreduce_send_invp( const RUNTIME_option_t *options,
+ cham_dir_t dir,
CHAM_ipiv_t *ipiv,
int ipivk,
const CHAM_desc_t *A,
@@ -61,6 +64,7 @@ INSERT_TASK_zperm_allreduce_send_invp( const RUNTIME_option_t *options,
int n )
{
(void)options;
+ (void)dir;
(void)ipiv;
(void)ipivk;
(void)A;
@@ -70,6 +74,7 @@ INSERT_TASK_zperm_allreduce_send_invp( const RUNTIME_option_t *options,
void
INSERT_TASK_zperm_allreduce( const RUNTIME_option_t *options,
+ cham_dir_t dir,
const CHAM_desc_t *A,
CHAM_desc_t *U,
int Um,
@@ -81,6 +86,7 @@ INSERT_TASK_zperm_allreduce( const RUNTIME_option_t *options,
void *ws )
{
(void)options;
+ (void)dir;
(void)A;
(void)U;
(void)Um;
diff --git a/runtime/starpu/codelets/codelet_zlaswp.c b/runtime/starpu/codelets/codelet_zlaswp.c
index 81c28d92f05d6c23e85e743b8402b79db31815b1..3829763abd896ca9db917a9d0573ac4d9b9b5255 100644
--- a/runtime/starpu/codelets/codelet_zlaswp.c
+++ b/runtime/starpu/codelets/codelet_zlaswp.c
@@ -13,7 +13,7 @@
* @author Mathieu Faverge
* @author Matthieu Kuhn
* @author Alycia Lisito
- * @date 2024-11-12
+ * @date 2025-03-24
* @precisions normal z -> c d s
*
*/
@@ -48,11 +48,12 @@ CODELETS_CPU( zlaswp_get, cl_zlaswp_get_cpu_func )
#if defined(CHAMELEON_STARPU_USE_INSERT)
void INSERT_TASK_zlaswp_get( const RUNTIME_option_t *options,
- int m0, int k,
+ cham_dir_t dir, int m0, int k,
const CHAM_ipiv_t *ipiv, int ipivk,
const CHAM_desc_t *A, int Am, int An,
const CHAM_desc_t *U, int Um, int Un )
{
+ void *ipiv_handle;
struct starpu_codelet *codelet = &cl_zlaswp_get;
if ( A->get_rankof( A, Am, An) != A->myrank ) {
return;
@@ -63,12 +64,18 @@ void INSERT_TASK_zlaswp_get( const RUNTIME_option_t *options,
clargs->m0 = m0;
clargs->k = k;
+ if ( dir == ChamDirForward ) {
+ ipiv_handle = RUNTIME_perm_getaddr( ipiv, ipivk );
+ }
+ else {
+ ipiv_handle = RUNTIME_invp_getaddr( ipiv, ipivk );
+ }
//void (*callback)(void*) = options->profiling ? cl_zlaswp_get_callback : NULL;
rt_starpu_insert_task(
codelet,
STARPU_CL_ARGS, clargs, sizeof(struct cl_zlaswp_args_s),
- STARPU_R, RUNTIME_perm_getaddr( ipiv, ipivk ),
+ STARPU_R, ipiv_handle,
STARPU_R, RTBLKADDR(A, ChamComplexDouble, Am, An),
STARPU_RW | STARPU_COMMUTE, RTBLKADDR(U, ChamComplexDouble, Um, Un),
STARPU_PRIORITY, options->priority,
@@ -80,18 +87,26 @@ void INSERT_TASK_zlaswp_get( const RUNTIME_option_t *options,
#else /* defined(CHAMELEON_STARPU_USE_INSERT) */
void INSERT_TASK_zlaswp_get( const RUNTIME_option_t *options,
- int m0, int k,
+ cham_dir_t dir, int m0, int k,
const CHAM_ipiv_t *ipiv, int ipivk,
const CHAM_desc_t *A, int Am, int An,
const CHAM_desc_t *U, int Um, int Un )
{
- int ret;
+ int ret;
struct starpu_task *task;
+ void *ipiv_handle;
if ( A->get_rankof( A, Am, An) != A->myrank ) {
return;
}
+ if ( dir == ChamDirForward ) {
+ ipiv_handle = RUNTIME_perm_getaddr( ipiv, ipivk );
+ }
+ else {
+ ipiv_handle = RUNTIME_invp_getaddr( ipiv, ipivk );
+ }
+
INSERT_TASK_COMMON_PARAMETERS_EXTENDED( zlaswp_get, zlaswp_get, zlaswp, 3);
/*
@@ -99,8 +114,7 @@ void INSERT_TASK_zlaswp_get( const RUNTIME_option_t *options,
*/
starpu_cham_exchange_init_params( options, ¶ms, U->get_rankof( U, Um, Un ) );
starpu_cham_exchange_handle_before_execution( options, ¶ms, &nbdata, descrs,
- RUNTIME_perm_getaddr( ipiv, ipivk ),
- STARPU_R );
+ ipiv_handle, STARPU_R );
starpu_cham_register_descr( &nbdata, descrs, RTBLKADDR( A, ChamComplexDouble, Am, An ), STARPU_R );
starpu_cham_register_descr( &nbdata, descrs, RTBLKADDR( U, ChamComplexDouble, Um, Un ),
STARPU_RW | STARPU_COMMUTE );
@@ -157,12 +171,14 @@ static void cl_zlaswp_set_cpu_func( void *descr[], void *cl_arg )
CODELETS_CPU( zlaswp_set, cl_zlaswp_set_cpu_func )
#if defined(CHAMELEON_STARPU_USE_INSERT)
+
void INSERT_TASK_zlaswp_set( const RUNTIME_option_t *options,
- int m0, int k,
+ cham_dir_t dir, int m0, int k,
const CHAM_ipiv_t *ipiv, int ipivk,
const CHAM_desc_t *A, int Am, int An,
const CHAM_desc_t *B, int Bm, int Bn )
{
+ void *ipiv_handle;
struct starpu_codelet *codelet = &cl_zlaswp_set;
if ( B->get_rankof( B, Bm, Bn) != A->myrank ) {
return;
@@ -173,12 +189,19 @@ void INSERT_TASK_zlaswp_set( const RUNTIME_option_t *options,
clargs->m0 = m0;
clargs->k = k;
+ if ( dir == ChamDirForward ) {
+ ipiv_handle = RUNTIME_invp_getaddr( ipiv, ipivk );
+ }
+ else {
+ ipiv_handle = RUNTIME_perm_getaddr( ipiv, ipivk );
+ }
+
//void (*callback)(void*) = options->profiling ? cl_zlaswp_set_callback : NULL;
rt_starpu_insert_task(
codelet,
STARPU_CL_ARGS, clargs, sizeof(struct cl_zlaswp_args_s),
- STARPU_R, RUNTIME_invp_getaddr( ipiv, ipivk ),
+ STARPU_R, ipiv_handle,
STARPU_R, RTBLKADDR(A, ChamComplexDouble, Am, An),
STARPU_RW, RTBLKADDR(B, ChamComplexDouble, Bm, Bn),
STARPU_PRIORITY, options->priority,
@@ -186,20 +209,30 @@ void INSERT_TASK_zlaswp_set( const RUNTIME_option_t *options,
STARPU_EXECUTE_ON_WORKER, options->workerid,
0 );
}
-#else
+
+#else /* defined(CHAMELEON_STARPU_USE_INSERT) */
+
void INSERT_TASK_zlaswp_set( const RUNTIME_option_t *options,
- int m0, int k,
+ cham_dir_t dir, int m0, int k,
const CHAM_ipiv_t *ipiv, int ipivk,
const CHAM_desc_t *A, int Am, int An,
const CHAM_desc_t *B, int Bm, int Bn )
{
- int ret;
+ int ret;
struct starpu_task *task;
+ void *ipiv_handle;
if ( B->get_rankof( B, Bm, Bn) != A->myrank ) {
return;
}
+ if( dir == ChamDirForward ) {
+ ipiv_handle = RUNTIME_invp_getaddr( ipiv, ipivk );
+ }
+ else {
+ ipiv_handle = RUNTIME_perm_getaddr( ipiv, ipivk );
+ }
+
INSERT_TASK_COMMON_PARAMETERS_EXTENDED( zlaswp_set, zlaswp_set, zlaswp, 3);
/*
@@ -207,8 +240,7 @@ void INSERT_TASK_zlaswp_set( const RUNTIME_option_t *options,
*/
starpu_cham_exchange_init_params( options, ¶ms, B->get_rankof( B, Bm, Bn ) );
starpu_cham_exchange_handle_before_execution( options, ¶ms, &nbdata, descrs,
- RUNTIME_invp_getaddr( ipiv, ipivk ),
- STARPU_R );
+ ipiv_handle, STARPU_R );
starpu_cham_register_descr( &nbdata, descrs, RTBLKADDR( A, ChamComplexDouble, Am, An ), STARPU_R );
starpu_cham_register_descr( &nbdata, descrs, RTBLKADDR( B, ChamComplexDouble, Bm, Bn ), STARPU_RW );
@@ -242,4 +274,5 @@ void INSERT_TASK_zlaswp_set( const RUNTIME_option_t *options,
}
starpu_cham_task_exchange_data_after_execution( options, params, nbdata, descrs );
}
-#endif
+#endif /* defined(CHAMELEON_STARPU_USE_INSERT) */
+
diff --git a/runtime/starpu/codelets/codelet_zperm_allreduce.c b/runtime/starpu/codelets/codelet_zperm_allreduce.c
index e32b7ad9c46a2303eb1c4c6a18d442935fca6d3a..a479056c5f9321b75cd89a99349fd1ef1c3f3976 100644
--- a/runtime/starpu/codelets/codelet_zperm_allreduce.c
+++ b/runtime/starpu/codelets/codelet_zperm_allreduce.c
@@ -12,7 +12,7 @@
* @version 1.3.0
* @author Alycia Lisito
* @author Pierre Esterie
- * @date 2024-11-14
+ * @date 2025-03-24
* @precisions normal z -> c d s
*
*/
@@ -21,6 +21,7 @@
#include <coreblas/cblas_wrapper.h>
#if defined(CHAMELEON_USE_MPI)
+
struct cl_redux_args_s {
int tempmm;
int mb;
@@ -91,6 +92,7 @@ INSERT_TASK_zperm_allreduce_send( const RUNTIME_option_t *options,
static void
INSERT_TASK_zperm_allreduce_recv( const RUNTIME_option_t *options,
+ cham_dir_t dir,
CHAM_desc_t *U,
CHAM_ipiv_t *ipiv,
int ipivk,
@@ -105,6 +107,15 @@ INSERT_TASK_zperm_allreduce_recv( const RUNTIME_option_t *options,
int p_first )
{
struct cl_redux_args_s *clargs;
+ void *ipiv_handle;
+
+ if ( dir == ChamDirForward ) {
+ ipiv_handle = RUNTIME_perm_getaddr( ipiv, ipivk );
+ }
+ else {
+ ipiv_handle = RUNTIME_invp_getaddr( ipiv, ipivk );
+ }
+
clargs = malloc( sizeof( struct cl_redux_args_s ) );
clargs->tempmm = tempmm;
clargs->mb = U->mb;
@@ -121,7 +132,7 @@ INSERT_TASK_zperm_allreduce_recv( const RUNTIME_option_t *options,
STARPU_CL_ARGS, clargs, sizeof(struct cl_redux_args_s),
STARPU_RW, RTBLKADDR(U, CHAMELEON_Complex64_t, me, n),
STARPU_R, RTBLKADDR(U, CHAMELEON_Complex64_t, src, n),
- STARPU_R, RUNTIME_perm_getaddr( ipiv, ipivk ),
+ STARPU_R, ipiv_handle,
STARPU_EXECUTE_ON_NODE, me,
STARPU_EXECUTE_ON_WORKER, options->workerid,
STARPU_PRIORITY, options->priority,
@@ -151,6 +162,7 @@ INSERT_TASK_zperm_allreduce_send( const RUNTIME_option_t *options,
static void
INSERT_TASK_zperm_allreduce_recv( const RUNTIME_option_t *options,
+ cham_dir_t dir,
CHAM_desc_t *U,
CHAM_ipiv_t *ipiv,
int ipivk,
@@ -164,8 +176,16 @@ INSERT_TASK_zperm_allreduce_recv( const RUNTIME_option_t *options,
int np,
int p_first )
{
- int ret;
+ int ret;
struct starpu_task *task;
+ void *ipiv_handle
+
+ if ( dir == ChamDirForward ) {
+ ipiv_handle = RUNTIME_perm_getaddr( ipiv, ipivk );
+ }
+ else {
+ ipiv_handle = RUNTIME_invp_getaddr( ipiv, ipivk );
+ }
INSERT_TASK_COMMON_PARAMETERS_EXTENDED( zperm_allreduce_send, zperm_allreduce, redux, 3 );
@@ -176,7 +196,7 @@ INSERT_TASK_zperm_allreduce_recv( const RUNTIME_option_t *options,
starpu_cham_exchange_handle_before_execution( options, ¶ms, &nbdata, descrs,
RTBLKADDR( U, ChamComplexDouble, src, n ),
STARPU_R );
- starpu_cham_register_descr( &nbdata, descrs, RUNTIME_perm_getaddr( ipiv, ipivk ), STARPU_R );
+ starpu_cham_register_descr( &nbdata, descrs, ipiv_handle, STARPU_R );
task = starpu_task_create();
task->cl = cl;
@@ -221,6 +241,7 @@ INSERT_TASK_zperm_allreduce_recv( const RUNTIME_option_t *options,
static void
zperm_allreduce_chameleon_starpu_task( const RUNTIME_option_t *options,
+ cham_dir_t dir,
const CHAM_desc_t *A,
CHAM_desc_t *U,
int Um,
@@ -229,10 +250,10 @@ zperm_allreduce_chameleon_starpu_task( const RUNTIME_option_t *options,
int ipivk,
int k,
int n,
- struct chameleon_pzgetrf_s *ws)
+ struct chameleon_pzgetrf_s *ws )
{
int *proc_involved = ws->proc_involved;
- int np_involved = chameleon_min( chameleon_desc_datadist_get_iparam(A, 0), A->mt - k);
+ int np_involved = chameleon_min( chameleon_desc_datadist_get_iparam(A, 0), A->mt - k );
int np_iter = np_involved;
int p_recv, p_send, me, p_first;
int shift = 1;
@@ -253,9 +274,11 @@ zperm_allreduce_chameleon_starpu_task( const RUNTIME_option_t *options,
p_recv = proc_involved[ ( me - shift + np_involved ) % np_involved ];
INSERT_TASK_zperm_allreduce_send( options, U, A->myrank, p_send, n );
- INSERT_TASK_zperm_allreduce_recv( options, U, ipiv, ipivk, A->myrank, p_recv,
+ INSERT_TASK_zperm_allreduce_recv( options, dir, U, ipiv, ipivk, A->myrank, p_recv,
n, k == (A->mt-1) ? A->m - k * A->mb : A->mb,
- chameleon_desc_datadist_get_iparam(A, 0), chameleon_desc_datadist_get_iparam(A, 1), shift, np_involved, p_first );
+ chameleon_desc_datadist_get_iparam(A, 0),
+ chameleon_desc_datadist_get_iparam(A, 1),
+ shift, np_involved, p_first );
shift = shift << 1;
np_iter = chameleon_ceil( np_iter, 2 );
@@ -265,6 +288,7 @@ zperm_allreduce_chameleon_starpu_task( const RUNTIME_option_t *options,
void
INSERT_TASK_zperm_allreduce( const RUNTIME_option_t *options,
+ cham_dir_t dir,
const CHAM_desc_t *A,
CHAM_desc_t *U,
int Um,
@@ -280,7 +304,7 @@ INSERT_TASK_zperm_allreduce( const RUNTIME_option_t *options,
switch( alg ) {
case ChamStarPUTasks:
default:
- zperm_allreduce_chameleon_starpu_task( options, A, U, Um, Un, ipiv, ipivk, k, n, tmp );
+ zperm_allreduce_chameleon_starpu_task( options, dir, A, U, Um, Un, ipiv, ipivk, k, n, tmp );
}
}
@@ -307,33 +331,51 @@ INSERT_TASK_zperm_allreduce_send_A( const RUNTIME_option_t *options,
void
INSERT_TASK_zperm_allreduce_send_perm( const RUNTIME_option_t *options,
+ cham_dir_t dir,
CHAM_ipiv_t *ipiv,
int ipivk,
int myrank,
int np,
int *proc_involved )
{
- int p;
+ int p;
+ void *ipiv_handle;
+
+ if ( dir == ChamDirForward ) {
+ ipiv_handle = RUNTIME_perm_getaddr( ipiv, ipivk );
+ }
+ else {
+ ipiv_handle = RUNTIME_invp_getaddr( ipiv, ipivk );
+ }
for ( p = 0; p < np; p++ ) {
if ( proc_involved[ p ] == myrank ) {
continue;
}
starpu_mpi_get_data_on_node_detached( options->sequence->comm,
- RUNTIME_perm_getaddr( ipiv, ipivk ),
+ ipiv_handle,
proc_involved[ p ], NULL, NULL );
}
}
void
INSERT_TASK_zperm_allreduce_send_invp( const RUNTIME_option_t *options,
+ cham_dir_t dir,
CHAM_ipiv_t *ipiv,
int ipivk,
const CHAM_desc_t *A,
int k,
int n )
{
- int b, rank;
+ int b, rank;
+ void *ipiv_handle;
+
+ if ( dir == ChamDirForward ) {
+ ipiv_handle = RUNTIME_invp_getaddr( ipiv, ipivk );
+ }
+ else {
+ ipiv_handle = RUNTIME_perm_getaddr( ipiv, ipivk );
+ }
for ( b = k+1; (b < A->mt) && ((b-(k+1)) < chameleon_desc_datadist_get_iparam(A, 0)); b ++ ) {
rank = A->get_rankof( A, b, n );
@@ -341,10 +383,11 @@ INSERT_TASK_zperm_allreduce_send_invp( const RUNTIME_option_t *options,
continue;
}
starpu_mpi_get_data_on_node_detached( options->sequence->comm,
- RUNTIME_invp_getaddr( ipiv, ipivk ),
+ ipiv_handle,
rank, NULL, NULL );
}
}
+
#else
void
INSERT_TASK_zperm_allreduce_send_A( const RUNTIME_option_t *options,
@@ -353,7 +396,7 @@ INSERT_TASK_zperm_allreduce_send_A( const RUNTIME_option_t *options,
int An,
int myrank,
int np,
- int *proc_involved )
+ int *proc_involved )
{
(void)options;
(void)A;
@@ -366,11 +409,12 @@ INSERT_TASK_zperm_allreduce_send_A( const RUNTIME_option_t *options,
void
INSERT_TASK_zperm_allreduce_send_perm( const RUNTIME_option_t *options,
+ cham_dir_t dir,
CHAM_ipiv_t *ipiv,
int ipivk,
int myrank,
int np,
- int *proc_involved )
+ int *proc_involved )
{
(void)options;
(void)ipiv;
@@ -382,6 +426,7 @@ INSERT_TASK_zperm_allreduce_send_perm( const RUNTIME_option_t *options,
void
INSERT_TASK_zperm_allreduce_send_invp( const RUNTIME_option_t *options,
+ cham_dir_t dir,
CHAM_ipiv_t *ipiv,
int ipivk,
const CHAM_desc_t *A,
@@ -398,6 +443,7 @@ INSERT_TASK_zperm_allreduce_send_invp( const RUNTIME_option_t *options,
void
INSERT_TASK_zperm_allreduce( const RUNTIME_option_t *options,
+ cham_dir_t dir,
const CHAM_desc_t *A,
CHAM_desc_t *U,
int Um,
@@ -419,4 +465,5 @@ INSERT_TASK_zperm_allreduce( const RUNTIME_option_t *options,
(void)n;
(void)ws;
}
+
#endif