From a3d1533fddcb630afde4a3b9530ec02f5c18e581 Mon Sep 17 00:00:00 2001
From: Matthieu KUHN <bkuhnm@l0.spartan.bench.local>
Date: Fri, 1 Apr 2022 13:33:50 +0200
Subject: [PATCH] getrf_nopiv: Call the option structure alloc/free functions
in the getrf nopiv and gesv nopiv functions to switch from initial algorithm
to the one with explicit communication management
---
compute/pzgetrf_nopiv.c | 18 +++++++++++++-----
compute/zgesv_nopiv.c | 11 ++++++-----
compute/zgetrf_nopiv.c | 4 ++--
control/compute_z.h | 2 +-
4 files changed, 22 insertions(+), 13 deletions(-)
diff --git a/compute/pzgetrf_nopiv.c b/compute/pzgetrf_nopiv.c
index daac4bad9..fe5bdaeec 100644
--- a/compute/pzgetrf_nopiv.c
+++ b/compute/pzgetrf_nopiv.c
@@ -11,7 +11,7 @@
*
* @brief Chameleon zgetrf_nopiv parallel algorithm
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Omar Zenati
* @author Mathieu Faverge
* @author Emmanuel Agullo
@@ -20,7 +20,7 @@
* @author Samuel Thibault
* @author Terry Cojean
* @author Matthieu Kuhn
- * @date 2022-02-22
+ * @date 2024-10-17
* @precisions normal z -> s d c
*
*/
@@ -31,9 +31,9 @@
/**
* Parallel tile LU factorization with no pivoting - dynamic scheduling
*/
-void chameleon_pzgetrf_nopiv( CHAM_desc_t *A,
- RUNTIME_sequence_t *sequence,
- RUNTIME_request_t *request )
+void chameleon_pzgetrf_nopiv_generic( CHAM_desc_t *A,
+ RUNTIME_sequence_t *sequence,
+ RUNTIME_request_t *request )
{
CHAM_context_t *chamctxt;
RUNTIME_option_t options;
@@ -121,3 +121,11 @@ void chameleon_pzgetrf_nopiv( CHAM_desc_t *A,
RUNTIME_options_finalize(&options, chamctxt);
}
+
+void chameleon_pzgetrf_nopiv( struct chameleon_pzgetrf_nopiv_s *ws,
+ CHAM_desc_t *A,
+ RUNTIME_sequence_t *sequence,
+ RUNTIME_request_t *request )
+{
+ chameleon_pzgetrf_nopiv_generic( A, sequence, request );
+}
diff --git a/compute/zgesv_nopiv.c b/compute/zgesv_nopiv.c
index c7c703a45..3ad6b0c25 100644
--- a/compute/zgesv_nopiv.c
+++ b/compute/zgesv_nopiv.c
@@ -11,7 +11,7 @@
*
* @brief Chameleon zgesv_nopiv wrappers
*
- * @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
* @author Jakub Kurzak
@@ -19,7 +19,8 @@
* @author Emmanuel Agullo
* @author Cedric Castagnede
* @author Florent Pruvost
- * @date 2022-02-22
+ * @author Matthieu Kuhn
+ * @date 2024-10-17
* @precisions normal z -> s d c
*
*/
@@ -82,8 +83,8 @@
*
*/
int CHAMELEON_zgesv_nopiv( int N, int NRHS,
- CHAMELEON_Complex64_t *A, int LDA,
- CHAMELEON_Complex64_t *B, int LDB )
+ CHAMELEON_Complex64_t *A, int LDA,
+ CHAMELEON_Complex64_t *B, int LDB )
{
int NB;
int status;
@@ -294,7 +295,7 @@ int CHAMELEON_zgesv_nopiv_Tile_Async( CHAM_desc_t *A, CHAM_desc_t *B,
return CHAMELEON_SUCCESS;
*/
- chameleon_pzgetrf_nopiv( A, sequence, request );
+ chameleon_pzgetrf_nopiv( NULL, A, sequence, request );
chameleon_pztrsm( ChamLeft, ChamLower, ChamNoTrans, ChamUnit, (CHAMELEON_Complex64_t)1.0, A, B, sequence, request );
diff --git a/compute/zgetrf_nopiv.c b/compute/zgetrf_nopiv.c
index 07ebfa2e9..1f4534914 100644
--- a/compute/zgetrf_nopiv.c
+++ b/compute/zgetrf_nopiv.c
@@ -159,7 +159,7 @@ void CHAMELEON_zgetrf_nopiv_WS_Free( void *user_ws )
*
*/
int CHAMELEON_zgetrf_nopiv( int M, int N,
- CHAMELEON_Complex64_t *A, int LDA )
+ CHAMELEON_Complex64_t *A, int LDA )
{
int NB;
int status;
@@ -353,7 +353,7 @@ int CHAMELEON_zgetrf_nopiv_Tile_Async( CHAM_desc_t *A,
return chameleon_request_fail(sequence, request, CHAMELEON_ERR_ILLEGAL_VALUE);
}
- chameleon_pzgetrf_nopiv( A, sequence, request );
+ chameleon_pzgetrf_nopiv( NULL, A, sequence, request );
return CHAMELEON_SUCCESS;
}
diff --git a/control/compute_z.h b/control/compute_z.h
index 12ec9e909..c0bac9755 100644
--- a/control/compute_z.h
+++ b/control/compute_z.h
@@ -146,7 +146,7 @@ void chameleon_pzgeqrf( int genD, CHAM_desc_t *A, CHAM_desc_t *T, CHAM_desc_t *D
void chameleon_pzgeqrfrh( int genD, int BS, CHAM_desc_t *A, CHAM_desc_t *T, CHAM_desc_t *D, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
void chameleon_pzgetrf( struct chameleon_pzgetrf_s *ws, CHAM_desc_t *A, CHAM_ipiv_t *IPIV, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request );
void chameleon_pzgetrf_incpiv(CHAM_desc_t *A, CHAM_desc_t *L, CHAM_desc_t *D, int *IPIV, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
-void chameleon_pzgetrf_nopiv(CHAM_desc_t *A, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
+void chameleon_pzgetrf_nopiv(struct chameleon_pzgetrf_nopiv_s *ws, CHAM_desc_t *A, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
void chameleon_pzgetrf_reclap(CHAM_desc_t *A, int *IPIV, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
void chameleon_pzgetrf_rectil(CHAM_desc_t *A, int *IPIV, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
void chameleon_pzhegst(int itype, cham_uplo_t uplo, CHAM_desc_t *A, CHAM_desc_t *B, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
--
GitLab