From 7ec28e8a4e0cad7515a6daa11dd20a2380c70b31 Mon Sep 17 00:00:00 2001
From: Lionel Eyraud-Dubois <lionel.eyraud-dubois@inria.fr>
Date: Thu, 12 May 2022 14:32:52 +0200
Subject: [PATCH] descriptors: Add an arg parameter to the rank_of functions to
enable the use of more complex solutions.
---
compute/pzgenm2.c | 15 +++++----
compute/pzgepdf_qdwh.c | 15 +++++----
compute/pzhetrd_he2hb.c | 7 ++--
compute/pzlange.c | 17 +++++-----
compute/pzlansy.c | 13 ++++----
compute/pzlatms.c | 13 ++++----
compute/pzplrnk.c | 9 ++---
compute/zcesca.c | 15 +++++----
compute/zgemm.c | 5 +--
compute/zgetrf.c | 7 ++--
compute/zgram.c | 9 ++---
compute/zhemm.c | 9 ++---
compute/zsymm.c | 9 ++---
compute/ztile.c | 9 ++---
control/chameleon_f77.c | 11 +++---
control/compute_z.h | 17 +++++-----
control/descriptor.c | 52 +++++++++++++++++++++--------
control/descriptor.h | 11 +++---
control/descriptor_rec.c | 17 ++++++----
example/lapack_to_chameleon/step3.c | 15 +++++----
include/chameleon.h | 13 +++++---
include/chameleon/struct.h | 6 ++--
testing/testing_zgetrf.c | 9 ++---
testing/testing_zprint.c | 11 +++---
24 files changed, 187 insertions(+), 127 deletions(-)
diff --git a/compute/pzgenm2.c b/compute/pzgenm2.c
index 36b0ae6ee..77c07aa17 100644
--- a/compute/pzgenm2.c
+++ b/compute/pzgenm2.c
@@ -12,11 +12,12 @@
*
* @brief Chameleon parallel algorithm to compute 2-norm estimator.
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Mathieu Faverge
* @author Dalal Sukkari
* @author Alycia Lisito
- * @date 2022-02-22
+ * @author Lionel Eyraud-Dubois
+ * @date 2023-07-05
* @precisions normal z -> s d c
*
*/
@@ -67,7 +68,7 @@ chameleon_pzgenm2( double tol, const CHAM_desc_t *A, double *result,
*/
chameleon_desc_init( &DROW, CHAMELEON_MAT_ALLOC_GLOBAL, ChamRealDouble, 1, A->nb, A->nb,
A->p, A->n, 0, 0, A->p, A->n, A->p, A->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
/**
* NRMX must be allocated with GLOBAL to be able to access the norm value
* after flushing the descriptor.
@@ -75,7 +76,7 @@ chameleon_pzgenm2( double tol, const CHAM_desc_t *A, double *result,
*/
chameleon_desc_init( &NRMX, CHAMELEON_MAT_ALLOC_GLOBAL, ChamRealDouble, 2, 1, 2,
A->p * 2, A->q, 0, 0, A->p * 2, A->q, A->p, A->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
/**
* Start by computing the initial vector of the iterative process, and that
@@ -171,13 +172,13 @@ chameleon_pzgenm2( double tol, const CHAM_desc_t *A, double *result,
chameleon_desc_init( &NRMSX, CHAMELEON_MAT_ALLOC_GLOBAL, ChamRealDouble, 2, 1, 2,
A->p * 2, A->q, 0, 0, A->p * 2, A->q, A->p, A->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
chameleon_desc_init( &X, CHAMELEON_MAT_ALLOC_GLOBAL, ChamComplexDouble, 1, A->nb, A->nb,
A->p, A->n, 0, 0, A->p, A->n, A->p, A->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
chameleon_desc_init( &SX, CHAMELEON_MAT_ALLOC_GLOBAL, ChamComplexDouble, A->mb, 1, A->mb,
A->m, A->q, 0, 0, A->m, A->q, A->p, A->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
cnt = 0;
e0 = 0.;
diff --git a/compute/pzgepdf_qdwh.c b/compute/pzgepdf_qdwh.c
index 1f369ff3a..3cace25b8 100644
--- a/compute/pzgepdf_qdwh.c
+++ b/compute/pzgepdf_qdwh.c
@@ -11,10 +11,11 @@
*
* @brief Chameleon QDWH Polar Decomposition parallel algorithm
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Mathieu Faverge
* @author Hatem Ltaief
- * @date 2022-02-22
+ * @author Lionel Eyraud-Dubois
+ * @date 2023-07-05
* @precisions normal z -> s d c
*
*/
@@ -109,23 +110,23 @@ chameleon_pzgepdf_qdwh_init( const CHAM_desc_t *U, const CHAM_desc_t *H,
ChamComplexDouble, ib, nb, ib * nb,
ib * U->mt, nb * U->nt, 0, 0,
ib * U->mt, nb * U->nt, U->p, U->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
chameleon_desc_init( TT1, CHAMELEON_MAT_ALLOC_TILE,
ChamComplexDouble, ib, nb, ib * nb,
ib * U->mt, nb * U->nt, 0, 0,
ib * U->mt, nb * U->nt, H->p, H->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
chameleon_desc_init( TS2, CHAMELEON_MAT_ALLOC_TILE,
ChamComplexDouble, ib, nb, ib * nb,
ib * H->mt, nb * H->nt, 0, 0,
ib * H->mt, nb * H->nt, U->p, U->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
chameleon_desc_init( TT2, CHAMELEON_MAT_ALLOC_TILE,
ChamComplexDouble, ib, nb, ib * nb,
ib * H->mt, nb * H->nt, 0, 0,
ib * H->mt, nb * H->nt, H->p, H->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
/*
* Create the adapted trees to perform the QR factorizations
@@ -170,7 +171,7 @@ chameleon_pzgepdf_qdwh_init( const CHAM_desc_t *U, const CHAM_desc_t *H,
ChamComplexDouble, U->mb, U->nb, U->mb * U->nb,
U->n, U->m, 0, 0,
U->n, U->m, U->p, U->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
/*
* Allocate the data descriptors for the lookahead if needed
diff --git a/compute/pzhetrd_he2hb.c b/compute/pzhetrd_he2hb.c
index 7d416d185..288cc9382 100644
--- a/compute/pzhetrd_he2hb.c
+++ b/compute/pzhetrd_he2hb.c
@@ -11,13 +11,14 @@
*
* @brief Chameleon zhetrd_he2hb parallel algorithm
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Hatem Ltaief
* @author Azzam Haidar
* @author Mathieu Faverge
* @author Samuel Thibault
* @author Alycia Lisito
- * @date 2022-02-22
+ * @author Lionel Eyraud-Dubois
+ * @date 2023-07-05
* @precisions normal z -> s d c
*
*/
@@ -94,7 +95,7 @@ void chameleon_pzhetrd_he2hb(cham_uplo_t uplo,
chameleon_desc_init( &AT, CHAMELEON_MAT_ALLOC_GLOBAL, ChamComplexDouble, A->mb, A->nb, (A->mb*A->nb),
chameleon_min(A->mt, A->nt) * A->mb, A->nb, 0, 0,
chameleon_min(A->mt, A->nt) * A->mb, A->nb, 1, 1,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
/* Let's extract the diagonal in a temporary copy that contains A and A' */
for (k = 1; k < A->nt; k++){
diff --git a/compute/pzlange.c b/compute/pzlange.c
index fd29e505e..4944d8fb4 100644
--- a/compute/pzlange.c
+++ b/compute/pzlange.c
@@ -11,14 +11,15 @@
*
* @brief Chameleon zlange parallel algorithm
*
- * @version 1.2.0
+ * @version 1.3.0
* @comment This file has been automatically generated
* from Plasma 2.6.0 for CHAMELEON 0.9.2
* @author Emmanuel Agullo
* @author Mathieu Faverge
* @author Florent Pruvost
* @author Alycia Lisito
- * @date 2022-02-22
+ * @author Lionel Eyraud-Dubois
+ * @date 2023-07-05
* @precisions normal z -> s d c
*
*/
@@ -403,7 +404,7 @@ void chameleon_pzlange_generic( cham_normtype_t norm, cham_uplo_t uplo, cham_dia
chameleon_desc_init( &Wcol, CHAMELEON_MAT_ALLOC_TILE, ChamRealDouble, 1, A->nb, A->nb,
workmt, worknt * A->nb, 0, 0, workmt, worknt * A->nb, A->p, A->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
wcol_init = 1;
/*
@@ -411,7 +412,7 @@ void chameleon_pzlange_generic( cham_normtype_t norm, cham_uplo_t uplo, cham_dia
*/
chameleon_desc_init( &Welt, CHAMELEON_MAT_ALLOC_GLOBAL, ChamRealDouble, 1, 1, 1,
A->p, worknt, 0, 0, A->p, worknt, A->p, A->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
break;
@@ -423,12 +424,12 @@ void chameleon_pzlange_generic( cham_normtype_t norm, cham_uplo_t uplo, cham_dia
chameleon_desc_init( &Wcol, CHAMELEON_MAT_ALLOC_TILE, ChamRealDouble, A->mb, 1, A->mb,
workmt * A->mb, worknt, 0, 0, workmt * A->mb, worknt, A->p, A->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
wcol_init = 1;
chameleon_desc_init( &Welt, CHAMELEON_MAT_ALLOC_GLOBAL, ChamRealDouble, 1, 1, 1,
workmt, A->q, 0, 0, workmt, A->q, A->p, A->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
break;
/*
@@ -440,7 +441,7 @@ void chameleon_pzlange_generic( cham_normtype_t norm, cham_uplo_t uplo, cham_dia
alpha = 1.;
chameleon_desc_init( &Welt, CHAMELEON_MAT_ALLOC_GLOBAL, ChamRealDouble, 2, 1, 2,
workmt*2, worknt, 0, 0, workmt*2, worknt, A->p, A->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
break;
/*
@@ -452,7 +453,7 @@ void chameleon_pzlange_generic( cham_normtype_t norm, cham_uplo_t uplo, cham_dia
chameleon_desc_init( &Welt, CHAMELEON_MAT_ALLOC_GLOBAL, ChamRealDouble, 1, 1, 1,
workmt, worknt, 0, 0, workmt, worknt, A->p, A->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
}
/* Initialize workspaces */
diff --git a/compute/pzlansy.c b/compute/pzlansy.c
index c19aada11..9c427cee9 100644
--- a/compute/pzlansy.c
+++ b/compute/pzlansy.c
@@ -11,14 +11,15 @@
*
* @brief Chameleon zlansy parallel algorithm
*
- * @version 1.2.0
+ * @version 1.3.0
* @comment This file has been automatically generated
* from Plasma 2.6.0 for CHAMELEON 0.9.2
* @author Emmanuel Agullo
* @author Mathieu Faverge
* @author Florent Pruvost
* @author Alycia Lisito
- * @date 2022-02-22
+ * @author Lionel Eyraud-Dubois
+ * @date 2023-07-05
* @precisions normal z -> c d s
*
*/
@@ -330,7 +331,7 @@ void chameleon_pzlansy_generic( cham_normtype_t norm, cham_uplo_t uplo, cham_tra
chameleon_desc_init( &Wcol, CHAMELEON_MAT_ALLOC_TILE, ChamRealDouble, A->mb, 1, A->mb,
workmt * A->mb, worknt, 0, 0, workmt * A->mb, worknt, A->p, A->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
wcol_init = 1;
/*
@@ -338,7 +339,7 @@ void chameleon_pzlansy_generic( cham_normtype_t norm, cham_uplo_t uplo, cham_tra
*/
chameleon_desc_init( &Welt, CHAMELEON_MAT_ALLOC_GLOBAL, ChamRealDouble, 1, 1, 1,
workmt, A->q, 0, 0, workmt, A->q, A->p, A->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
break;
/*
@@ -350,7 +351,7 @@ void chameleon_pzlansy_generic( cham_normtype_t norm, cham_uplo_t uplo, cham_tra
alpha = 1.;
chameleon_desc_init( &Welt, CHAMELEON_MAT_ALLOC_GLOBAL, ChamRealDouble, 2, 1, 2,
workmt*2, worknt, 0, 0, workmt*2, worknt, A->p, A->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
break;
/*
@@ -362,7 +363,7 @@ void chameleon_pzlansy_generic( cham_normtype_t norm, cham_uplo_t uplo, cham_tra
chameleon_desc_init( &Welt, CHAMELEON_MAT_ALLOC_GLOBAL, ChamRealDouble, 1, 1, 1,
workmt, worknt, 0, 0, workmt, worknt, A->p, A->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
}
/* Initialize workspaces */
diff --git a/compute/pzlatms.c b/compute/pzlatms.c
index 3881a9f9d..c8e1fa5a9 100644
--- a/compute/pzlatms.c
+++ b/compute/pzlatms.c
@@ -10,9 +10,10 @@
*
* @brief Chameleon zlatms parallel algorithm
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Mathieu Faverge
- * @date 2022-02-22
+ * @author Lionel Eyraud-Dubois
+ * @date 2023-07-05
* @precisions normal z -> s d c
*
*/
@@ -212,12 +213,12 @@ void chameleon_pzlatms( cham_dist_t idist, unsigned long long int seed, cham_sym
ChamComplexDouble, ib, descU.nb, ib * descU.nb,
ib * descU.mt, descU.nb * descU.nt, 0, 0,
ib * descU.mt, descU.nb * descU.nt, descU.p, descU.q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
chameleon_desc_init( &descTT, CHAMELEON_MAT_ALLOC_TILE,
ChamComplexDouble, ib, descU.nb, ib * descU.nb,
ib * descU.mt, descU.nb * descU.nt, 0, 0,
ib * descU.mt, descU.nb * descU.nt, descU.p, descU.q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
/* U <= qr(U) */
chameleon_pzgeqrf_param( 1, kt, &qrtree, &descU,
@@ -283,12 +284,12 @@ void chameleon_pzlatms( cham_dist_t idist, unsigned long long int seed, cham_sym
ChamComplexDouble, ib, descV.nb, ib * descV.nb,
ib * descV.mt, descV.nb * descV.nt, 0, 0,
ib * descV.mt, descV.nb * descV.nt, descV.p, descV.q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
chameleon_desc_init( &descTT, CHAMELEON_MAT_ALLOC_TILE,
ChamComplexDouble, ib, descV.nb, ib * descV.nb,
ib * descV.mt, descV.nb * descV.nt, 0, 0,
ib * descV.mt, descV.nb * descV.nt, descV.p, descV.q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
/* V <= qr(V) */
chameleon_pzgelqf_param( 1, kt, &qrtree, &descV,
diff --git a/compute/pzplrnk.c b/compute/pzplrnk.c
index c4367c4cc..368ddec75 100644
--- a/compute/pzplrnk.c
+++ b/compute/pzplrnk.c
@@ -9,10 +9,11 @@
*
* @brief Chameleon zplrnk parallel algorithm
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Mathieu Faverge
* @author Florent Pruvost
- * @date 2022-02-22
+ * @author Lionel Eyraud-Dubois
+ * @date 2023-07-05
* @precisions normal z -> s d c
*
*/
@@ -48,12 +49,12 @@ void chameleon_pzplrnk( int K, CHAM_desc_t *C,
ChamComplexDouble, C->mb, C->nb, (C->mb * C->nb),
C->mt * C->mb, C->nb * C->q, 0, 0,
C->mt * C->mb, C->nb * C->q, C->p, C->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
chameleon_desc_init( &WB, CHAMELEON_MAT_ALLOC_TILE,
ChamComplexDouble, C->mb, C->nb, (C->mb * C->nb),
C->mb * C->p, C->nt * C->nb, 0, 0,
C->mb * C->p, C->nt * C->nb, C->p, C->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
KT = (K + C->mb - 1) / C->mb;
myp = C->myrank / C->q;
diff --git a/compute/zcesca.c b/compute/zcesca.c
index 36c735842..76489d42f 100644
--- a/compute/zcesca.c
+++ b/compute/zcesca.c
@@ -9,9 +9,10 @@
*
* @brief Chameleon zcesca wrappers
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Florent Pruvost
- * @date 2022-02-22
+ * @author Lionel Eyraud-Dubois
+ * @date 2023-07-05
* @precisions normal z -> s d c z
*
*/
@@ -61,31 +62,31 @@ void *CHAMELEON_zcesca_WS_Alloc( const CHAM_desc_t *A )
ChamComplexDouble, 1, A->nb, A->nb,
workmt, A->n, 0, 0,
workmt, A->n, A->p, A->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
chameleon_desc_init( &(options->Wgrow), CHAMELEON_MAT_ALLOC_TILE,
ChamComplexDouble, A->mb, 1, A->mb,
A->m, worknt, 0, 0,
A->m, worknt, A->p, A->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
chameleon_desc_init( &(options->Wgelt), CHAMELEON_MAT_ALLOC_TILE,
ChamComplexDouble, 1, 1, 1,
1, worknt, 0, 0,
1, worknt, A->p, A->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
chameleon_desc_init( &(options->Wdcol), CHAMELEON_MAT_ALLOC_TILE,
ChamRealDouble, 2, A->nb, 2*A->nb,
2*workmt, A->n, 0, 0,
2*workmt, A->n, A->p, A->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
chameleon_desc_init( &(options->Wdrow), CHAMELEON_MAT_ALLOC_TILE,
ChamRealDouble, A->mb, 2, 2*A->mb,
A->m, 2*worknt, 0, 0,
A->m, 2*worknt, A->p, A->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
return (void*)options;
}
diff --git a/compute/zgemm.c b/compute/zgemm.c
index def663153..d8892fe79 100644
--- a/compute/zgemm.c
+++ b/compute/zgemm.c
@@ -18,6 +18,7 @@
* @author Emmanuel Agullo
* @author Cedric Castagnede
* @author Florent Pruvost
+ * @author Lionel Eyraud-Dubois
* @date 2023-07-05
* @precisions normal z -> s d c
*
@@ -182,12 +183,12 @@ void *CHAMELEON_zgemm_WS_Alloc( cham_trans_t transA __attribute__((unused)
ChamComplexDouble, C->mb, C->nb, (C->mb * C->nb),
C->mt * C->mb, C->nb * C->q * lookahead, 0, 0,
C->mt * C->mb, C->nb * C->q * lookahead, C->p, C->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
chameleon_desc_init( &(options->WB), CHAMELEON_MAT_ALLOC_TILE,
ChamComplexDouble, C->mb, C->nb, (C->mb * C->nb),
C->mb * C->p * lookahead, C->nt * C->nb, 0, 0,
C->mb * C->p * lookahead, C->nt * C->nb, C->p, C->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
}
return (void*)options;
diff --git a/compute/zgetrf.c b/compute/zgetrf.c
index 72c595373..4a39cb5bb 100644
--- a/compute/zgetrf.c
+++ b/compute/zgetrf.c
@@ -11,14 +11,15 @@
*
* @brief Chameleon zgetrf wrappers
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Omar Zenati
* @author Mathieu Faverge
* @author Emmanuel Agullo
* @author Cedric Castagnede
* @author Florent Pruvost
* @author Matthieu Kuhn
- * @date 2023-02-21
+ * @author Lionel Eyraud-Dubois
+ * @date 2023-07-05
*
* @precisions normal z -> s d c
*
@@ -86,7 +87,7 @@ CHAMELEON_zgetrf_WS_Alloc( const CHAM_desc_t *A )
ChamComplexDouble, 1, A->nb, A->nb,
A->mt, A->nt * A->nb, 0, 0,
A->mt, A->nt * A->nb, A->p, A->q,
- NULL, NULL, A->get_rankof_init );
+ NULL, NULL, A->get_rankof_init, A->get_rankof_init_arg );
}
return ws;
diff --git a/compute/zgram.c b/compute/zgram.c
index fac51926e..141f0ffcd 100644
--- a/compute/zgram.c
+++ b/compute/zgram.c
@@ -9,11 +9,12 @@
*
* @brief Chameleon zgram wrappers
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Mathieu Faverge
* @author Florent Pruvost
* @author Philippe Swartvagher
- * @date 2022-02-22
+ * @author Lionel Eyraud-Dubois
+ * @date 2023-07-05
* @precisions normal z -> s d c z
*
*/
@@ -64,13 +65,13 @@ void *CHAMELEON_zgram_WS_Alloc( const CHAM_desc_t *A )
ChamRealDouble, 2, A->nb, 2*A->nb,
2*workmt, A->n, 0, 0,
2*workmt, A->n, A->p, A->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
chameleon_desc_init( &(options->Welt), CHAMELEON_MAT_ALLOC_TILE,
ChamRealDouble, 2, 1, 2,
2, worknt, 0, 0,
2, worknt, A->p, A->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
return (void*)options;
}
diff --git a/compute/zhemm.c b/compute/zhemm.c
index 92ea46502..8b01715c4 100644
--- a/compute/zhemm.c
+++ b/compute/zhemm.c
@@ -11,14 +11,15 @@
*
* @brief Chameleon zhemm 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 Mathieu Faverge
* @author Emmanuel Agullo
* @author Cedric Castagnede
* @author Florent Pruvost
- * @date 2022-02-22
+ * @author Lionel Eyraud-Dubois
+ * @date 2023-07-05
* @precisions normal z -> c
*
*/
@@ -158,12 +159,12 @@ void *CHAMELEON_zhemm_WS_Alloc( cham_side_t side __attribute__((unused)),
ChamComplexDouble, C->mb, C->nb, (C->mb * C->nb),
C->mt * C->mb, C->nb * C->q * lookahead, 0, 0,
C->mt * C->mb, C->nb * C->q * lookahead, C->p, C->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
chameleon_desc_init( &(options->WB), CHAMELEON_MAT_ALLOC_TILE,
ChamComplexDouble, C->mb, C->nb, (C->mb * C->nb),
C->mb * C->p * lookahead, C->nt * C->nb, 0, 0,
C->mb * C->p * lookahead, C->nt * C->nb, C->p, C->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
}
return (void*)options;
diff --git a/compute/zsymm.c b/compute/zsymm.c
index fe680aa28..f493497b8 100644
--- a/compute/zsymm.c
+++ b/compute/zsymm.c
@@ -11,14 +11,15 @@
*
* @brief Chameleon zsymm 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 Mathieu Faverge
* @author Emmanuel Agullo
* @author Cedric Castagnede
* @author Florent Pruvost
- * @date 2022-02-22
+ * @author Lionel Eyraud-Dubois
+ * @date 2023-07-05
* @precisions normal z -> s d c
*
*/
@@ -158,12 +159,12 @@ void *CHAMELEON_zsymm_WS_Alloc( cham_side_t side __attribute__((unused)),
ChamComplexDouble, C->mb, C->nb, (C->mb * C->nb),
C->mt * C->mb, C->nb * C->q * lookahead, 0, 0,
C->mt * C->mb, C->nb * C->q * lookahead, C->p, C->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
chameleon_desc_init( &(options->WB), CHAMELEON_MAT_ALLOC_TILE,
ChamComplexDouble, C->mb, C->nb, (C->mb * C->nb),
C->mb * C->p * lookahead, C->nt * C->nb, 0, 0,
C->mb * C->p * lookahead, C->nt * C->nb, C->p, C->q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
}
return (void*)options;
diff --git a/compute/ztile.c b/compute/ztile.c
index bdcfd29e7..2e2707908 100644
--- a/compute/ztile.c
+++ b/compute/ztile.c
@@ -11,12 +11,13 @@
*
* @brief Chameleon auxiliary routines
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Jakub Kurzak
* @author Mathieu Faverge
* @author Cedric Castagnede
* @author Florent Pruvost
- * @date 2022-02-22
+ * @author Lionel Eyraud-Dubois
+ * @date 2023-07-05
* @precisions normal z -> s d c
*
*/
@@ -139,7 +140,7 @@ int CHAMELEON_zLap2Desc( cham_uplo_t uplo, CHAMELEON_Complex64_t *Af77, int LDA,
/* Create the B descriptor to handle the Lapack format matrix */
status = CHAMELEON_Desc_Create_User( &B, Af77, ChamComplexDouble, A->mb, A->nb, A->bsiz,
LDA, A->n, 0, 0, A->m, A->n, 1, 1,
- chameleon_getaddr_cm, chameleon_getblkldd_cm, NULL );
+ chameleon_getaddr_cm, chameleon_getblkldd_cm, NULL, NULL );
if ( status != CHAMELEON_SUCCESS ) {
chameleon_error("CHAMELEON_zTile_to_Lapack", "Failed to create the descriptor");
return status;
@@ -223,7 +224,7 @@ int CHAMELEON_zDesc2Lap( cham_uplo_t uplo, CHAM_desc_t *A, CHAMELEON_Complex64_t
/* Create the B descriptor to handle the Lapack format matrix */
status = CHAMELEON_Desc_Create_User( &B, Af77, ChamComplexDouble, A->mb, A->nb, A->bsiz,
LDA, A->n, 0, 0, A->m, A->n, 1, 1,
- chameleon_getaddr_cm, chameleon_getblkldd_cm, NULL );
+ chameleon_getaddr_cm, chameleon_getblkldd_cm, NULL, NULL );
if ( status != CHAMELEON_SUCCESS ) {
chameleon_error("CHAMELEON_zTile_to_Lapack", "Failed to create the descriptor");
return status;
diff --git a/control/chameleon_f77.c b/control/chameleon_f77.c
index 21a6316ee..648c80793 100644
--- a/control/chameleon_f77.c
+++ b/control/chameleon_f77.c
@@ -11,13 +11,14 @@
*
* @brief Chameleon Fortran77 interface
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Bilel Hadri
* @author Cedric Castagnede
* @author Florent Pruvost
* @author Mathieu Faverge
* @author Philippe Virouleau
- * @date 2022-02-22
+ * @author Lionel Eyraud-Dubois
+ * @date 2023-07-05
*
*/
#include "control/common.h"
@@ -80,16 +81,18 @@ extern "C" {
void* (*get_blkaddr)( const CHAM_desc_t*, int, int ),
int (*get_blkldd) ( const CHAM_desc_t*, int ),
int (*get_rankof) ( const CHAM_desc_t*, int, int ),
+ void* get_rankof_arg,
int *INFO)
{ *INFO = CHAMELEON_Desc_Create_User(descptr, mat, *dtyp, *mb, *nb, *bsiz, *lm, *ln, *i, *j, *m, *n, *p, *q,
- get_blkaddr, get_blkldd, get_rankof); }
+ get_blkaddr, get_blkldd, get_rankof, get_rankof_arg); }
void CHAMELEON_DESC_CREATE_OOC_USER(CHAM_desc_t **descptr, cham_flttype_t *dtyp,
int *mb, int *nb, int *bsiz, int *lm, int *ln,
int *i, int *j, int *m, int *n, int *p, int *q,
int (*get_rankof) ( const CHAM_desc_t*, int, int ),
+ void* get_rankof_arg,
int *INFO)
{ *INFO = CHAMELEON_Desc_Create_OOC_User(descptr, *dtyp, *mb, *nb, *bsiz, *lm, *ln, *i, *j, *m, *n, *p, *q,
- get_rankof); }
+ get_rankof, get_rankof_arg); }
void CHAMELEON_DESC_DESTROY(CHAM_desc_t **desc, int *INFO)
{ *INFO = CHAMELEON_Desc_Destroy(desc); }
diff --git a/control/compute_z.h b/control/compute_z.h
index 634bd2d5c..8bec9da5f 100644
--- a/control/compute_z.h
+++ b/control/compute_z.h
@@ -11,7 +11,7 @@
*
* @brief Chameleon computational functions header
*
- * @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
@@ -21,7 +21,8 @@
* @author Florent Pruvost
* @author Alycia Lisito
* @author Matthieu Kuhn
- * @date 2023-02-21
+ * @author Lionel Eyraud-Dubois
+ * @date 2023-07-05
* @precisions normal z -> c d s
*
*/
@@ -219,7 +220,7 @@ chameleon_zdesc_alloc_diag( CHAM_desc_t *descA, int nb, int m, int n, int p, int
diag_m, nb, 0, 0, diag_m, nb, p, q,
chameleon_getaddr_diag,
chameleon_getblkldd_ccrb,
- chameleon_getrankof_2d_diag );
+ chameleon_getrankof_2d_diag, NULL );
}
#define chameleon_zdesc_alloc( descA, mb, nb, lm, ln, i, j, m, n, free) \
@@ -228,7 +229,7 @@ chameleon_zdesc_alloc_diag( CHAM_desc_t *descA, int nb, int m, int n, int p, int
rc = chameleon_desc_init( &(descA), CHAMELEON_MAT_ALLOC_GLOBAL, \
ChamComplexDouble, (mb), (nb), ((mb)*(nb)), \
(m), (n), (i), (j), (m), (n), 1, 1, \
- NULL, NULL, NULL ); \
+ NULL, NULL, NULL, NULL ); \
if ( rc != CHAMELEON_SUCCESS ) { \
{free;} \
return rc; \
@@ -254,7 +255,7 @@ chameleon_zdesc_copy_and_restrict( const CHAM_desc_t *descIn,
m, n, 0, 0, m, n, descIn->p, descIn->q,
descIn->get_blkaddr,
descIn->get_blkldd,
- descIn->get_rankof_init );
+ descIn->get_rankof_init, descIn->get_rankof_init_arg );
return rc;
}
@@ -273,13 +274,13 @@ chameleon_zlap2tile( CHAM_context_t *chamctxt,
/* Initialize the Lapack descriptor */
chameleon_desc_init( descAl, A, ChamComplexDouble, mb, nb, (mb)*(nb),
lm, ln, 0, 0, m, n, 1, 1,
- chameleon_getaddr_cm, chameleon_getblkldd_cm, NULL );
+ chameleon_getaddr_cm, chameleon_getblkldd_cm, NULL, NULL );
descAl->styp = ChamCM;
/* Initialize the tile descriptor */
chameleon_desc_init( descAt, CHAMELEON_MAT_ALLOC_TILE, ChamComplexDouble, mb, nb, (mb)*(nb),
lm, ln, 0, 0, m, n, 1, 1,
- chameleon_getaddr_ccrb, chameleon_getblkldd_ccrb, NULL );
+ chameleon_getaddr_ccrb, chameleon_getblkldd_ccrb, NULL, NULL );
if ( mode & ChamDescInput ) {
chameleon_pzlacpy( uplo, descAl, descAt, seq, req );
@@ -289,7 +290,7 @@ chameleon_zlap2tile( CHAM_context_t *chamctxt,
/* Initialize the tile descriptor */
chameleon_desc_init( descAt, A, ChamComplexDouble, mb, nb, (mb)*(nb),
lm, ln, 0, 0, m, n, 1, 1,
- chameleon_getaddr_cm, chameleon_getblkldd_cm, NULL );
+ chameleon_getaddr_cm, chameleon_getblkldd_cm, NULL, NULL );
}
return CHAMELEON_SUCCESS;
}
diff --git a/control/descriptor.c b/control/descriptor.c
index c5e4ca2d2..1436ad25d 100644
--- a/control/descriptor.c
+++ b/control/descriptor.c
@@ -18,7 +18,8 @@
* @author Guillaume Sylvand
* @author Raphael Boucherie
* @author Samuel Thibault
- * @date 2023-07-04
+ * @author Lionel Eyraud-Dubois
+ * @date 2023-07-05
*
***
*
@@ -169,6 +170,9 @@ void chameleon_desc_init_tiles( CHAM_desc_t *desc, blkrankof_fct_t rankof )
* @param[in] get_rankof
* A function that return the MPI rank of the tile A(m,n).
*
+ * @param[in] get_rankof_arg
+ * A pointer to custom data that can be used by the get_rankof function
+ *
******************************************************************************
*
* @return The descriptor with the matrix description parameters set.
@@ -179,7 +183,8 @@ int chameleon_desc_init_internal( CHAM_desc_t *desc, const char *name, void *mat
int lm, int ln, int m, int n, int p, int q,
blkaddr_fct_t get_blkaddr,
blkldd_fct_t get_blkldd,
- blkrankof_fct_t get_rankof )
+ blkrankof_fct_t get_rankof,
+ void *get_rankof_arg )
{
CHAM_context_t *chamctxt;
int rc = CHAMELEON_SUCCESS;
@@ -200,6 +205,7 @@ int chameleon_desc_init_internal( CHAM_desc_t *desc, const char *name, void *mat
desc->get_blkldd = get_blkldd ? get_blkldd : chameleon_getblkldd_ccrb;
desc->get_rankof = chameleon_getrankof_tile;
desc->get_rankof_init = get_rankof ? get_rankof : chameleon_getrankof_2d;
+ desc->get_rankof_init_arg = get_rankof_arg;
/* Matrix properties */
desc->dtyp = dtyp;
@@ -492,9 +498,17 @@ CHAMELEON_Desc_SubMatrix( CHAM_desc_t *descA, int i, int j, int m, int n )
int CHAMELEON_Desc_Create( CHAM_desc_t **descptr, void *mat, cham_flttype_t dtyp, int mb, int nb, int bsiz,
int lm, int ln, int i, int j, int m, int n, int p, int q )
{
+ blkrankof_fct_t get_rankof = NULL;
+
+ /* if (getenv("CHAMELEON_1Dx1D_DISTRIBUTION")){ */
+ /* printf("[CHAMELEON] : Using 1Dx1D distribubtion\n"); */
+ /* get_rankof = chameleon_getrankof_custom; */
+ /* load_dist(&custom_dist_a, m, n, mb, nb); */
+ /* } */
+
return CHAMELEON_Desc_Create_User( descptr, mat, dtyp, mb, nb, bsiz,
lm, ln, i, j, m, n, p, q,
- NULL, NULL, NULL );
+ NULL, NULL, get_rankof, NULL );
}
/**
@@ -564,6 +578,9 @@ int CHAMELEON_Desc_Create( CHAM_desc_t **descptr, void *mat, cham_flttype_t dtyp
* @param[in] get_rankof
* A function that return the MPI rank of the tile A(m,n).
*
+ * @param[in] get_rankof_arg
+ * A pointer to custom data that can be used by the get_rankof function
+ *
******************************************************************************
*
* @retval CHAMELEON_SUCCESS successful exit
@@ -573,7 +590,8 @@ int CHAMELEON_Desc_Create_User( CHAM_desc_t **descptr, void *mat, cham_flttype_t
int lm, int ln, int i, int j, int m, int n, int p, int q,
blkaddr_fct_t get_blkaddr,
blkldd_fct_t get_blkldd,
- blkrankof_fct_t get_rankof )
+ blkrankof_fct_t get_rankof,
+ void* get_rankof_arg )
{
CHAM_context_t *chamctxt;
CHAM_desc_t *desc;
@@ -594,7 +612,7 @@ int CHAMELEON_Desc_Create_User( CHAM_desc_t **descptr, void *mat, cham_flttype_t
chameleon_desc_init( desc, mat, dtyp, mb, nb, bsiz,
lm, ln, i, j, m, n, p, q,
- get_blkaddr, get_blkldd, get_rankof );
+ get_blkaddr, get_blkldd, get_rankof, get_rankof_arg );
status = chameleon_desc_check( desc );
if (status != CHAMELEON_SUCCESS) {
@@ -663,6 +681,9 @@ int CHAMELEON_Desc_Create_User( CHAM_desc_t **descptr, void *mat, cham_flttype_t
* @param[in] get_rankof
* A function that return the MPI rank of the tile A(m,n).
*
+ * @param[in] get_rankof_arg
+ * A pointer to custom data that can be used by the get_rankof function
+ *
******************************************************************************
*
* @retval CHAMELEON_SUCCESS successful exit
@@ -670,7 +691,7 @@ int CHAMELEON_Desc_Create_User( CHAM_desc_t **descptr, void *mat, cham_flttype_t
*/
int CHAMELEON_Desc_Create_OOC_User(CHAM_desc_t **descptr, cham_flttype_t dtyp, int mb, int nb, int bsiz,
int lm, int ln, int i, int j, int m, int n, int p, int q,
- blkrankof_fct_t get_rankof )
+ blkrankof_fct_t get_rankof, void* get_rankof_arg )
{
#if !defined (CHAMELEON_SCHED_STARPU)
(void)descptr; (void)dtyp; (void)mb; (void)nb; (void)bsiz;
@@ -683,7 +704,7 @@ int CHAMELEON_Desc_Create_OOC_User(CHAM_desc_t **descptr, cham_flttype_t dtyp, i
int rc;
rc = CHAMELEON_Desc_Create_User( descptr, CHAMELEON_MAT_OOC, dtyp, mb, nb, bsiz,
lm, ln, i, j, m, n, p, q,
- chameleon_getaddr_null, NULL, get_rankof );
+ chameleon_getaddr_null, NULL, get_rankof, get_rankof_arg );
return rc;
#endif
}
@@ -751,7 +772,7 @@ int CHAMELEON_Desc_Create_OOC(CHAM_desc_t **descptr, cham_flttype_t dtyp, int mb
{
return CHAMELEON_Desc_Create_User( descptr, CHAMELEON_MAT_OOC, dtyp, mb, nb, bsiz,
lm, ln, i, j, m, n, p, q,
- chameleon_getaddr_null, NULL, NULL );
+ chameleon_getaddr_null, NULL, NULL, NULL );
}
/**
@@ -785,7 +806,7 @@ CHAM_desc_t *CHAMELEON_Desc_Copy( const CHAM_desc_t *descin, void *mat )
CHAMELEON_Desc_Create_User( &descout, mat,
descin->dtyp, descin->mb, descin->nb, descin->bsiz,
descin->lm, descin->ln, descin->i, descin->j, descin->m, descin->n, descin->p, descin->q,
- NULL, NULL, descin->get_rankof_init );
+ NULL, NULL, descin->get_rankof_init, descin->get_rankof_init_arg );
return descout;
}
@@ -820,7 +841,7 @@ CHAM_desc_t *CHAMELEON_Desc_CopyOnZero( const CHAM_desc_t *descin, void *mat )
CHAMELEON_Desc_Create_User( &descout, mat,
descin->dtyp, descin->mb, descin->nb, descin->bsiz,
descin->lm, descin->ln, descin->i, descin->j, descin->m, descin->n, 1, 1,
- NULL, NULL, descin->get_rankof_init );
+ NULL, NULL, descin->get_rankof_init, descin->get_rankof_init_arg );
return descout;
}
@@ -1043,7 +1064,8 @@ CHAMELEON_Desc_Print( const CHAM_desc_t *desc )
*/
int CHAMELEON_Desc_Change_Distribution( cham_uplo_t uplo,
CHAM_desc_t *desc,
- blkrankof_fct_t new_get_rankof )
+ blkrankof_fct_t new_get_rankof,
+ void* new_get_rankof_arg )
{
int status;
CHAM_context_t *chamctxt;
@@ -1057,7 +1079,7 @@ int CHAMELEON_Desc_Change_Distribution( cham_uplo_t uplo,
chameleon_sequence_create( chamctxt, &sequence );
- CHAMELEON_Desc_Change_Distribution_Async( uplo, desc, new_get_rankof, sequence );
+ CHAMELEON_Desc_Change_Distribution_Async( uplo, desc, new_get_rankof, new_get_rankof_arg, sequence );
RUNTIME_desc_flush( desc, sequence );
chameleon_sequence_wait( chamctxt, sequence );
@@ -1098,6 +1120,7 @@ int CHAMELEON_Desc_Change_Distribution( cham_uplo_t uplo,
int CHAMELEON_Desc_Change_Distribution_Async( cham_uplo_t uplo,
CHAM_desc_t *desc,
blkrankof_fct_t new_get_rankof,
+ void* new_get_rankof_arg,
RUNTIME_sequence_t *sequence )
{
CHAM_context_t *chamctxt;
@@ -1111,7 +1134,8 @@ int CHAMELEON_Desc_Change_Distribution_Async( cham_uplo_t uplo,
}
/* Nothing to do if the new mapping is the same as the original one */
- if ( ( new_get_rankof == desc->get_rankof_init ) ||
+ if ( ( ( new_get_rankof == desc->get_rankof_init ) &&
+ ( new_get_rankof_arg == desc->get_rankof_init_arg) ) ||
( RUNTIME_comm_size( chamctxt ) == 1 ) )
{
return CHAMELEON_SUCCESS;
@@ -1132,6 +1156,8 @@ int CHAMELEON_Desc_Change_Distribution_Async( cham_uplo_t uplo,
}
#endif
+ desc->get_rankof_init_arg = new_get_rankof_arg;
+
for ( n = 0; n < desc->nt; n++ ) {
mmin = ( uplo == ChamLower ) ? chameleon_min( n, desc->mt ) : 0;
mmax = ( uplo == ChamUpper ) ? chameleon_min( n+1, desc->mt ) : desc->mt;
diff --git a/control/descriptor.h b/control/descriptor.h
index 96bf0bf0b..38153b6b2 100644
--- a/control/descriptor.h
+++ b/control/descriptor.h
@@ -19,7 +19,8 @@
* @author Guillaume Sylvand
* @author Raphael Boucherie
* @author Samuel Thibault
- * @date 2023-07-04
+ * @author Lionel Eyraud-Dubois
+ * @date 2023-07-05
*
*/
#ifndef _chameleon_descriptor_h_
@@ -48,7 +49,8 @@ int chameleon_desc_init_internal( CHAM_desc_t *desc, const char *name, void *mat
int lm, int ln, int m, int n, int p, int q,
void* (*get_blkaddr)( const CHAM_desc_t*, int, int ),
int (*get_blkldd) ( const CHAM_desc_t*, int ),
- int (*get_rankof) ( const CHAM_desc_t*, int, int ) );
+ int (*get_rankof) ( const CHAM_desc_t*, int, int ),
+ void* get_rankof_arg );
static inline int chameleon_desc_init( CHAM_desc_t *desc, void *mat,
@@ -57,7 +59,8 @@ static inline int chameleon_desc_init( CHAM_desc_t *desc, void *mat,
int m, int n, int p, int q,
void* (*get_blkaddr)( const CHAM_desc_t*, int, int ),
int (*get_blkldd) ( const CHAM_desc_t*, int ),
- int (*get_rankof) ( const CHAM_desc_t*, int, int ) )
+ int (*get_rankof) ( const CHAM_desc_t*, int, int ),
+ void* get_rankof_arg)
{
assert( i == 0 );
assert( j == 0 );
@@ -66,7 +69,7 @@ static inline int chameleon_desc_init( CHAM_desc_t *desc, void *mat,
(void)i;
(void)j;
return chameleon_desc_init_internal( desc, NULL, mat, dtyp, mb, nb, lm, ln, m, n, p, q,
- get_blkaddr, get_blkldd, get_rankof );
+ get_blkaddr, get_blkldd, get_rankof, get_rankof_arg );
}
CHAM_desc_t* chameleon_desc_submatrix( CHAM_desc_t *descA, int i, int j, int m, int n );
diff --git a/control/descriptor_rec.c b/control/descriptor_rec.c
index a69171411..058758b66 100644
--- a/control/descriptor_rec.c
+++ b/control/descriptor_rec.c
@@ -11,10 +11,11 @@
*
* @brief Chameleon descriptors routines
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Mathieu Faverge
* @author Gwenole Lucas
- * @date 2022-02-22
+ * @author Lionel Eyraud-Dubois
+ * @date 2023-07-05
*
*/
#include "control/common.h"
@@ -24,7 +25,8 @@ static int
chameleon_recdesc_create( const char *name, CHAM_desc_t **descptr, void *mat, cham_flttype_t dtyp,
int *mb, int *nb,
int lm, int ln, int m, int n, int p, int q,
- blkaddr_fct_t get_blkaddr, blkldd_fct_t get_blkldd, blkrankof_fct_t get_rankof )
+ blkaddr_fct_t get_blkaddr, blkldd_fct_t get_blkldd,
+ blkrankof_fct_t get_rankof, void* get_rankof_arg )
{
CHAM_context_t *chamctxt;
CHAM_desc_t *desc;
@@ -49,7 +51,7 @@ chameleon_recdesc_create( const char *name, CHAM_desc_t **descptr, void *mat, ch
desc = (CHAM_desc_t*)malloc(sizeof(CHAM_desc_t));
rc = chameleon_desc_init_internal( desc, name, mat, dtyp, mb[0], nb[0],
lm, ln, m, n, p, q,
- get_blkaddr, get_blkldd, get_rankof );
+ get_blkaddr, get_blkldd, get_rankof, get_rankof_arg );
*descptr = desc;
if ( rc != CHAMELEON_SUCCESS ) {
@@ -81,7 +83,7 @@ chameleon_recdesc_create( const char *name, CHAM_desc_t **descptr, void *mat, ch
tile->ld, tempnn, /* Abuse as ln is not used */
tempmm, tempnn,
1, 1, /* can recurse only on local data */
- chameleon_getaddr_cm, chameleon_getblkldd_cm, NULL);
+ chameleon_getaddr_cm, chameleon_getblkldd_cm, NULL, NULL);
tile->format = CHAMELEON_TILE_DESC;
tile->mat = tiledesc;
@@ -98,7 +100,8 @@ chameleon_recdesc_create( const char *name, CHAM_desc_t **descptr, void *mat, ch
int
CHAMELEON_Recursive_Desc_Create( CHAM_desc_t **descptr, void *mat, cham_flttype_t dtyp,
int *mb, int *nb, int lm, int ln, int m, int n, int p, int q,
- blkaddr_fct_t get_blkaddr, blkldd_fct_t get_blkldd, blkrankof_fct_t get_rankof )
+ blkaddr_fct_t get_blkaddr, blkldd_fct_t get_blkldd,
+ blkrankof_fct_t get_rankof, void* get_rankof_arg )
{
/*
* The first layer must be allocated, otherwise we will give unitialized
@@ -109,5 +112,5 @@ CHAMELEON_Recursive_Desc_Create( CHAM_desc_t **descptr, void *mat, cham_flttype_
return chameleon_recdesc_create( "A", descptr, mat, dtyp,
mb, nb, lm, ln, m, n, p, q,
- get_blkaddr, get_blkldd, get_rankof );
+ get_blkaddr, get_blkldd, get_rankof, get_rankof_arg );
}
diff --git a/example/lapack_to_chameleon/step3.c b/example/lapack_to_chameleon/step3.c
index ac32b0d3f..aae96bab7 100644
--- a/example/lapack_to_chameleon/step3.c
+++ b/example/lapack_to_chameleon/step3.c
@@ -11,11 +11,12 @@
*
* @brief Chameleon step3 example
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Florent Pruvost
* @author Mathieu Faverge
* @author Philippe Virouleau
- * @date 2022-02-22
+ * @author Lionel Eyraud-Dubois
+ * @date 2023-07-05
*
*/
#include "step3.h"
@@ -101,11 +102,11 @@ int main(int argc, char *argv[]) {
* returns the MPI rank of the tile m, n (0 here because we do not
* intend to use this program with MPI)
*/
- CHAMELEON_Desc_Create_User(&descA, matA, ChamRealDouble,
- NB, NB, NB*NB, N, N, 0, 0, N, N, 1, 1,
- user_getaddr_arrayofpointers,
- user_getblkldd_arrayofpointers,
- user_getrankof_zero);
+ CHAMELEON_Desc_Create_User( &descA, matA, ChamRealDouble,
+ NB, NB, NB*NB, N, N, 0, 0, N, N, 1, 1,
+ user_getaddr_arrayofpointers,
+ user_getblkldd_arrayofpointers,
+ user_getrankof_zero, NULL );
/*
* We use the classical CHAMELEON way for accessing tiles for descripotrs
diff --git a/include/chameleon.h b/include/chameleon.h
index e8246474d..cce6539fc 100644
--- a/include/chameleon.h
+++ b/include/chameleon.h
@@ -17,6 +17,7 @@
* @author Cedric Castagnede
* @author Florent Pruvost
* @author Philippe Virouleau
+ * @author Lionel Eyraud-Dubois
* @date 2023-07-05
*
*/
@@ -124,7 +125,8 @@ int CHAMELEON_Element_Size( cham_flttype_t type );
int CHAMELEON_Desc_Create_User( CHAM_desc_t **desc, void *mat, cham_flttype_t dtyp, int mb, int nb, int bsiz,
int lm, int ln, int i, int j, int m, int n, int p, int q,
- blkaddr_fct_t get_blkaddr, blkldd_fct_t get_blkldd, blkrankof_fct_t get_rankof );
+ blkaddr_fct_t get_blkaddr, blkldd_fct_t get_blkldd,
+ blkrankof_fct_t get_rankof, void* get_rankof_arg );
int CHAMELEON_Desc_Create( CHAM_desc_t **desc, void *mat, cham_flttype_t dtyp,
int mb, int nb, int bsiz, int lm, int ln,
@@ -133,7 +135,7 @@ int CHAMELEON_Desc_Create( CHAM_desc_t **desc, void *mat, cham_flttype_t dtyp,
int CHAMELEON_Desc_Create_OOC_User( CHAM_desc_t **desc, cham_flttype_t dtyp,
int mb, int nb, int bsiz, int lm, int ln,
int i, int j, int m, int n, int p, int q,
- blkrankof_fct_t get_rankof );
+ blkrankof_fct_t get_rankof, void* get_rankof_arg );
int CHAMELEON_Desc_Create_OOC( CHAM_desc_t **desc, cham_flttype_t dtyp,
int mb, int nb, int bsiz, int lm, int ln,
int i, int j, int m, int n, int p, int q );
@@ -150,10 +152,12 @@ int CHAMELEON_Desc_Flush ( const CHAM_desc_t *desc,
int CHAMELEON_Desc_Change_Distribution( cham_uplo_t uplo,
CHAM_desc_t *desc,
- blkrankof_fct_t new_get_rankof );
+ blkrankof_fct_t new_get_rankof,
+ void* new_get_rankof_arg );
int CHAMELEON_Desc_Change_Distribution_Async( cham_uplo_t uplo,
CHAM_desc_t *desc,
blkrankof_fct_t new_get_rankof,
+ void* new_get_rankof_arg,
RUNTIME_sequence_t *sequence );
void CHAMELEON_Desc_Print( const CHAM_desc_t *desc );
@@ -188,7 +192,8 @@ int CHAMELEON_Request_Set (RUNTIME_request_t *request, int param, int value)
int CHAMELEON_Recursive_Desc_Create( CHAM_desc_t **descptr, void *mat, cham_flttype_t dtyp,
int *mb, int *nb, int lm, int ln, int m, int n, int p, int q,
- blkaddr_fct_t get_blkaddr, blkldd_fct_t get_blkldd, blkrankof_fct_t get_rankof );
+ blkaddr_fct_t get_blkaddr, blkldd_fct_t get_blkldd,
+ blkrankof_fct_t get_rankof, void* get_rankof_arg );
/**
*
diff --git a/include/chameleon/struct.h b/include/chameleon/struct.h
index 2b324ef54..efa64a1c5 100644
--- a/include/chameleon/struct.h
+++ b/include/chameleon/struct.h
@@ -11,14 +11,15 @@
*
* @brief Chameleon structures
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Cedric Augonnet
* @author Mathieu Faverge
* @author Cedric Castagnede
* @author Florent Pruvost
* @author Samuel Thibault
* @author Matthieu Kuhn
- * @date 2022-12-13
+ * @author Lionel Eyraud-Dubois
+ * @date 2023-07-05
*
*/
#ifndef _chameleon_struct_h_
@@ -93,6 +94,7 @@ struct chameleon_desc_s {
blkrankof_fct_t get_rankof;
// function to get chameleon tiles MPI rank
blkrankof_fct_t get_rankof_init;
+ void* get_rankof_init_arg;
CHAM_tile_t *tiles; // pointer to the array of tiles descriptors
void *mat; // pointer to the beginning of the matrix
size_t A21; // pointer to the beginning of the matrix A21
diff --git a/testing/testing_zgetrf.c b/testing/testing_zgetrf.c
index 92a82eb6b..4fb585532 100644
--- a/testing/testing_zgetrf.c
+++ b/testing/testing_zgetrf.c
@@ -9,12 +9,13 @@
*
* @brief Chameleon zgetrf testing
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Lucas Barros de Assis
* @author Mathieu Faverge
* @author Alycia Lisito
* @author Matthieu Kuhn
- * @date 2023-02-21
+ * @author Lionel Eyraud-Dubois
+ * @date 2023-07-05
* @precisions normal z -> c d s
*
*/
@@ -93,11 +94,11 @@ testing_zgetrf_desc( run_arg_list_t *args, int check )
CHAMELEON_Desc_Create_User(
&descA0c, (void*)CHAMELEON_MAT_ALLOC_GLOBAL, ChamComplexDouble,
nb, nb, nb*nb, M, N, 0, 0, M, N, 1, 1,
- chameleon_getaddr_cm, chameleon_getblkldd_cm, NULL );
+ chameleon_getaddr_cm, chameleon_getblkldd_cm, NULL, NULL );
CHAMELEON_Desc_Create_User(
&descIPIVc, (void*)CHAMELEON_MAT_ALLOC_GLOBAL, ChamInteger,
nb, 1, nb, M, 1, 0, 0, M, 1, 1, 1,
- chameleon_getaddr_cm, chameleon_getblkldd_cm, NULL );
+ chameleon_getaddr_cm, chameleon_getblkldd_cm, NULL, NULL );
if ( diag == ChamUnit ) {
CHAMELEON_zplgtr_Tile( 0, ChamUpper, descA0c, seedA );
diff --git a/testing/testing_zprint.c b/testing/testing_zprint.c
index 3aab544aa..bff288a02 100644
--- a/testing/testing_zprint.c
+++ b/testing/testing_zprint.c
@@ -9,11 +9,12 @@
*
* @brief Chameleon zprint testing
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Lucas Barros de Assis
* @author Mathieu Faverge
* @author Alycia Lisito
- * @date 2022-02-22
+ * @author Lionel Eyraud-Dubois
+ * @date 2023-07-05
* @precisions normal z -> c d s
*
*/
@@ -62,7 +63,7 @@ testing_zprint_desc( run_arg_list_t *args, int check )
fprintf( stdout, "--- Lapack layout ---\n" );
rc = CHAMELEON_Desc_Create_User(
&descA, (void*)(-mtxfmt), ChamComplexDouble, nb, nb, nb * nb, LDA, N, 0, 0, M, N, P, Q,
- chameleon_getaddr_cm, chameleon_getblkldd_cm, NULL );
+ chameleon_getaddr_cm, chameleon_getblkldd_cm, NULL, NULL );
if ( rc == CHAMELEON_SUCCESS ) {
CHAMELEON_Desc_Print( descA );
@@ -79,7 +80,7 @@ testing_zprint_desc( run_arg_list_t *args, int check )
rc = CHAMELEON_Recursive_Desc_Create(
&descA, CHAMELEON_MAT_ALLOC_GLOBAL, ChamComplexDouble,
list_nb, list_nb, LDA, N, M, N, P, Q,
- NULL, NULL, NULL );
+ NULL, NULL, NULL, NULL );
if ( rc == CHAMELEON_SUCCESS ) {
CHAMELEON_Desc_Print( descA );
@@ -90,7 +91,7 @@ testing_zprint_desc( run_arg_list_t *args, int check )
rc = CHAMELEON_Recursive_Desc_Create(
&descA, CHAMELEON_MAT_ALLOC_GLOBAL, ChamComplexDouble,
list_nb, list_nb, LDA, N, M, N, P, Q,
- chameleon_getaddr_cm, chameleon_getblkldd_cm, NULL );
+ chameleon_getaddr_cm, chameleon_getblkldd_cm, NULL, NULL );
if ( rc == CHAMELEON_SUCCESS ) {
CHAMELEON_Desc_Print( descA );
--
GitLab