diff --git a/cmake_modules/local_subs.py b/cmake_modules/local_subs.py
index afd17c16f2a60d1b5cb35616151072872fdb3de2..892e1405401236a94252ff1d3281d65a571e0880 100644
--- a/cmake_modules/local_subs.py
+++ b/cmake_modules/local_subs.py
@@ -52,6 +52,7 @@ _extra_blas = [
('', 'sgered', 'dgered', 'cgered', 'zgered' ),
('', 'sgerst', 'dgerst', 'cgerst', 'zgerst' ),
('', 'sipiv_allreduce', 'dipiv_allreduce', 'cipiv_allreduce', 'zipiv_allreduce' ),
+ ('', 'sperm_allreduce', 'dperm_allreduce', 'cperm_allreduce', 'zperm_allreduce' ),
]
_extra_BLAS = [ [ x.upper() for x in row ] for row in _extra_blas ]
diff --git a/include/chameleon/tasks_z.h b/include/chameleon/tasks_z.h
index 236482682266032654bcc6a8e6050b617134fa98..9b843c60a057a2c5fe2e4e3321b94c02e968fe62 100644
--- a/include/chameleon/tasks_z.h
+++ b/include/chameleon/tasks_z.h
@@ -583,4 +583,181 @@ void INSERT_TASK_zipiv_allreduce( CHAM_desc_t *A,
int h,
int n );
+/**
+ ********************************************************************************
+ *
+ * @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.
+ *
+ *******************************************************************************
+ *
+ * @param[in] options
+ * The runtime options data structure to pass through all insert_task calls.
+ *
+ * @param[in] A
+ * The descriptor of the matrix A.
+ *
+ * @param[in] ipiv
+ * The pivot structure that contains the informations for the LU
+ * factorization with partial pivoting.
+ *
+ * @param[in] ipivk
+ * The index of the permutation.
+ *
+ * @param[in] k
+ * The number of rows in the tile U(Um, Un).
+ *
+ * @param[in] n
+ * The number of columns in the tile U(Um, Un).
+ *
+ * @param[inout] U
+ * The descriptor of the worskpace used for the permutation in the LU
+ * factorization with partial pivoting.
+ *
+ * @param[in] Um
+ * The row index of the tile used in U.
+ *
+ * @param[in] Un
+ * The column index of the tile used in U.
+ *
+ * @param[in] ws
+ * The workspace to handle the data in the LU factorization with
+ * partial pivoting.
+ *
+ *******************************************************************************
+ */
+void INSERT_TASK_zperm_allreduce( const RUNTIME_option_t *options,
+ const CHAM_desc_t *A,
+ CHAM_ipiv_t *ipiv,
+ int ipivk,
+ int k,
+ int n,
+ CHAM_desc_t *U,
+ int Um,
+ int Un,
+ void *ws );
+
+/**
+ ********************************************************************************
+ *
+ * @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.
+ *
+ *******************************************************************************
+ *
+ * @param[in] options
+ * The runtime options data structure to pass through all insert_task calls.
+ *
+ * @param[in] A
+ * The descriptor of the matrix A.
+ *
+ * @param[in] Am
+ * The row index of the tile used in A.
+ *
+ * @param[in] An
+ * The column index of the tile used in A.
+ *
+ * @param[in] myrank
+ * The rank of the current process.
+ *
+ * @param[in] np
+ * The number of processus involved in the permutation.
+ *
+ * @param[in] proc_involved
+ * The list of the processus involved in the permutation.
+ *
+ *******************************************************************************
+ */
+void INSERT_TASK_zperm_allreduce_send_A( const RUNTIME_option_t *options,
+ CHAM_desc_t *A,
+ int Am,
+ int An,
+ int myrank,
+ int np,
+ int *proc_involved );
+
+/**
+ ********************************************************************************
+ *
+ * @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.
+ *
+ *******************************************************************************
+ *
+ * @param[in] options
+ * The runtime options data structure to pass through all insert_task calls.
+ *
+ * @param[in] ipiv
+ * The pivot structure that contains the informations for the LU
+ * factorization with partial pivoting.
+ *
+ * @param[in] ipivk
+ * The index of the permutation.
+ *
+ * @param[in] myrank
+ * The rank of the current process.
+ *
+ * @param[in] np
+ * The number of processus involved in the permutation.
+ *
+ * @param[in] proc_involved
+ * The list of the processus involved in the permutation.
+ *
+ *******************************************************************************
+ */
+void INSERT_TASK_zperm_allreduce_send_perm( const RUNTIME_option_t *options,
+ CHAM_ipiv_t *ipiv,
+ int ipivk,
+ int myrank,
+ int np,
+ int *proc_involved );
+
+/**
+ ********************************************************************************
+ *
+ * @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.
+ *
+ *******************************************************************************
+ *
+ * @param[in] options
+ * The runtime options data structure to pass through all insert_task calls.
+ *
+ * @param[in] ipiv
+ * The pivot structure that contains the informations for the LU
+ * factorization with partial pivoting.
+ *
+ * @param[in] ipivk
+ * The index of the permutation.
+ *
+ * @param[in] A
+ * The descriptor of the matrix A.
+ *
+ * @param[in] k
+ * The index of the panel factorized.
+ *
+ * @param[in] n
+ * The index of the panel to permute.
+ *
+ *******************************************************************************
+ */
+void INSERT_TASK_zperm_allreduce_send_invp( const RUNTIME_option_t *options,
+ CHAM_ipiv_t *ipiv,
+ int ipivk,
+ const CHAM_desc_t *A,
+ int k,
+ int n );
+
#endif /* _chameleon_tasks_z_h_ */
diff --git a/runtime/CMakeLists.txt b/runtime/CMakeLists.txt
index 6b24081b2bd7f58f330e28b142f8c714ba208009..e46fd45b105edfcf96d1bccb2a6780f481a1a9a7 100644
--- a/runtime/CMakeLists.txt
+++ b/runtime/CMakeLists.txt
@@ -86,6 +86,7 @@ set(CODELETS_ZSRC
codelets/codelet_zlaswp_batched.c
codelets/codelet_zlatro.c
codelets/codelet_zlauum.c
+ codelets/codelet_zperm_allreduce.c
codelets/codelet_zplghe.c
codelets/codelet_zplgsy.c
codelets/codelet_zplrnt.c
diff --git a/runtime/starpu/codelets/codelet_zperm_allreduce.c b/runtime/starpu/codelets/codelet_zperm_allreduce.c
new file mode 100644
index 0000000000000000000000000000000000000000..c21490d02f42447b07e8516842b2b9b840850006
--- /dev/null
+++ b/runtime/starpu/codelets/codelet_zperm_allreduce.c
@@ -0,0 +1,173 @@
+/**
+ *
+ * @file starpu/codelet_zperm_allreduce.c
+ *
+ * @copyright 2012-2024 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria,
+ * Univ. Bordeaux. All rights reserved.
+ *
+ ***
+ *
+ * @brief Chameleon StarPU codelets to do the reduction
+ *
+ * @version 1.3.0
+ * @author Alycia Lisito
+ * @date 2024-06-11
+ * @precisions normal z -> c d s
+ *
+ */
+#include "chameleon_starpu_internal.h"
+#include "runtime_codelet_z.h"
+#include <coreblas/cblas_wrapper.h>
+
+#if defined(CHAMELEON_USE_MPI)
+void
+INSERT_TASK_zperm_allreduce( const RUNTIME_option_t *options,
+ const CHAM_desc_t *A,
+ CHAM_ipiv_t *ipiv,
+ int ipivk,
+ int k,
+ int n,
+ CHAM_desc_t *U,
+ int Um,
+ int Un,
+ void *ws )
+{
+ struct chameleon_pzgetrf_s *tmp = (struct chameleon_pzgetrf_s *)ws;
+}
+
+void
+INSERT_TASK_zperm_allreduce_send_A( const RUNTIME_option_t *options,
+ CHAM_desc_t *A,
+ int Am,
+ int An,
+ int myrank,
+ int np,
+ int *proc_involved )
+{
+ int p, rank;
+
+ for ( p = 0; p < np; p ++ ) {
+ if ( proc_involved[ p ] == myrank ) {
+ continue;
+ }
+ starpu_mpi_get_data_on_node_detached( options->sequence->comm,
+ RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
+ proc_involved[ p ], NULL, NULL );
+ }
+}
+
+void
+INSERT_TASK_zperm_allreduce_send_perm( const RUNTIME_option_t *options,
+ CHAM_ipiv_t *ipiv,
+ int ipivk,
+ int myrank,
+ int np,
+ int *proc_involved )
+{
+ int p;
+
+ 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 ),
+ proc_involved[ p ], NULL, NULL );
+ }
+}
+
+void
+INSERT_TASK_zperm_allreduce_send_invp( const RUNTIME_option_t *options,
+ CHAM_ipiv_t *ipiv,
+ int ipivk,
+ const CHAM_desc_t *A,
+ int k,
+ int n )
+{
+ int b, rank;
+
+ for ( b = k+1; (b < A->mt) && ((b-(k+1)) < A->p); b ++ ) {
+ rank = A->get_rankof( A, b, n );
+ if ( rank == A->myrank ) {
+ continue;
+ }
+ starpu_mpi_get_data_on_node_detached( options->sequence->comm,
+ RUNTIME_invp_getaddr( ipiv, ipivk ),
+ rank, NULL, NULL );
+ }
+}
+#else
+void
+INSERT_TASK_zperm_allreduce_send_A( const RUNTIME_option_t *options,
+ CHAM_desc_t *A,
+ int Am,
+ int An,
+ int myrank,
+ int np,
+ int *proc_involved )
+{
+ (void)options;
+ (void)A;
+ (void)Am;
+ (void)An;
+ (void)myrank;
+ (void)np;
+ (void)proc_involved;
+}
+
+void
+INSERT_TASK_zperm_allreduce_send_perm( const RUNTIME_option_t *options,
+ CHAM_ipiv_t *ipiv,
+ int ipivk,
+ int myrank,
+ int np,
+ int *proc_involved )
+{
+ (void)options;
+ (void)ipiv;
+ (void)ipivk;
+ (void)myrank;
+ (void)np;
+ (void)proc_involved;
+}
+
+void
+INSERT_TASK_zperm_allreduce_send_invp( const RUNTIME_option_t *options,
+ CHAM_ipiv_t *ipiv,
+ int ipivk,
+ const CHAM_desc_t *A,
+ int k,
+ int n )
+{
+ (void)options;
+ (void)ipiv;
+ (void)ipivk;
+ (void)A;
+ (void)k;
+ (void)n;
+}
+
+void
+INSERT_TASK_zperm_allreduce( const RUNTIME_option_t *options,
+ const CHAM_desc_t *A,
+ CHAM_ipiv_t *ipiv,
+ int ipivk,
+ int k,
+ int n,
+ CHAM_desc_t *U,
+ int Um,
+ int Un,
+ void *ws )
+{
+ (void)options;
+ (void)A;
+ (void)ipiv;
+ (void)ipivk;
+ (void)k;
+ (void)n;
+ (void)U;
+ (void)Um;
+ (void)Un;
+ (void)ws;
+}
+#endif