From 0d58113d1dc30324b806f489c98be586aa5882c5 Mon Sep 17 00:00:00 2001
From: Matthieu KUHN <bkuhnm@l0.spartan.bench.local>
Date: Thu, 31 Mar 2022 14:46:46 +0200
Subject: [PATCH] getrf: Add a version with no pivoting per column
---
compute/pzgetrf.c | 88 ++++++++++++++++++++++++++++++++-------------
compute/zgetrf.c | 31 ++++++++++------
control/compute_z.h | 12 +++++--
3 files changed, 94 insertions(+), 37 deletions(-)
diff --git a/compute/pzgetrf.c b/compute/pzgetrf.c
index 7e6237c7e..967e83f2f 100644
--- a/compute/pzgetrf.c
+++ b/compute/pzgetrf.c
@@ -4,7 +4,7 @@
*
* @copyright 2009-2014 The University of Tennessee and The University of
* Tennessee Research Foundation. All rights reserved.
- * @copyright 2012-2022 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria,
+ * @copyright 2012-2023 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria,
* Univ. Bordeaux. All rights reserved.
*
***
@@ -16,13 +16,14 @@
* @author Mathieu Faverge
* @author Emmanuel Agullo
* @author Matthieu Kuhn
- * @date 2022-02-22
+ * @date 2023-02-21
* @precisions normal z -> s d c
*
*/
#include "control/common.h"
#define A(m,n) A, m, n
+#define U(m,n) &(ws->U), m, n
/*
* All the functions below are panel factorization variant.
@@ -44,10 +45,10 @@
* The runtime options data structure to pass through all insert_task calls.
*/
static inline void
-chameleon_pzgetrf_panel_facto_nopiv( void *ws,
- CHAM_desc_t *A,
- int k,
- RUNTIME_option_t *options )
+chameleon_pzgetrf_panel_facto_nopiv( struct chameleon_pzgetrf_s *ws,
+ CHAM_desc_t *A,
+ int k,
+ RUNTIME_option_t *options )
{
const CHAMELEON_Complex64_t zone = (CHAMELEON_Complex64_t) 1.0;
int m, tempkm, tempkn, tempmm;
@@ -60,7 +61,7 @@ chameleon_pzgetrf_panel_facto_nopiv( void *ws,
*/
INSERT_TASK_zgetrf_nopiv(
options,
- tempkm, tempkn, 32, A->mb,
+ tempkm, tempkn, ws->ib, A->mb,
A(k, k), 0);
for (m = k+1; m < A->mt; m++) {
@@ -73,24 +74,61 @@ chameleon_pzgetrf_panel_facto_nopiv( void *ws,
A(m, k) );
}
}
+
+static inline void
+chameleon_pzgetrf_panel_facto_nopiv_percol( struct chameleon_pzgetrf_s *ws,
+ CHAM_desc_t *A,
+ int k,
+ RUNTIME_option_t *options )
+{
+ int m, h;
+ int tempkm, tempkn, tempmm, minmn;
+
+ tempkm = k == A->mt-1 ? A->m-k*A->mb : A->mb;
+ tempkn = k == A->nt-1 ? A->n-k*A->nb : A->nb;
+ minmn = chameleon_min( tempkm, tempkn );
+
+ /*
+ * Algorithm per column without pivoting
+ */
+ for(h=0; h<minmn; h++){
+ INSERT_TASK_zgetrf_panel_nopiv_percol_diag(
+ options, tempkm, tempkn, h,
+ A( k, k ), U( k, k ), A->mb * k );
+
+ for (m = k+1; m < A->mt; m++) {
+ tempmm = (m == (A->mt - 1)) ? A->m - m * A->mb : A->mb;
+ INSERT_TASK_zgetrf_panel_nopiv_percol_trsm(
+ options, tempmm, tempkn, h,
+ A( m, k ), U( k, k ) );
+ }
+ }
+
+ RUNTIME_data_flush( options->sequence, U(k, k) );
+}
+
static inline void
-chameleon_pzgetrf_panel_facto( void *ws,
- CHAM_desc_t *A,
- int k,
- RUNTIME_option_t *options )
+chameleon_pzgetrf_panel_facto( struct chameleon_pzgetrf_s *ws,
+ CHAM_desc_t *A,
+ int k,
+ RUNTIME_option_t *options )
{
+#if defined(GETRF_NOPIV_PER_COLUMN)
+ chameleon_pzgetrf_panel_facto_nopiv_percol( ws, A, k, options );
+#else
chameleon_pzgetrf_panel_facto_nopiv( ws, A, k, options );
+#endif
}
/**
* Permutation of the panel n at step k
*/
static inline void
-chameleon_pzgetrf_panel_permute( void *ws,
- CHAM_desc_t *A,
- int k,
- int n,
- RUNTIME_option_t *options )
+chameleon_pzgetrf_panel_permute( struct chameleon_pzgetrf_s *ws,
+ CHAM_desc_t *A,
+ int k,
+ int n,
+ RUNTIME_option_t *options )
{
(void)ws;
(void)A;
@@ -100,11 +138,11 @@ chameleon_pzgetrf_panel_permute( void *ws,
}
static inline void
-chameleon_pzgetrf_panel_update( void *ws,
- CHAM_desc_t *A,
- int k,
- int n,
- RUNTIME_option_t *options )
+chameleon_pzgetrf_panel_update( struct chameleon_pzgetrf_s *ws,
+ CHAM_desc_t *A,
+ int k,
+ int n,
+ RUNTIME_option_t *options )
{
const CHAMELEON_Complex64_t zone = (CHAMELEON_Complex64_t) 1.0;
const CHAMELEON_Complex64_t mzone = (CHAMELEON_Complex64_t)-1.0;
@@ -141,10 +179,10 @@ chameleon_pzgetrf_panel_update( void *ws,
/**
* Parallel tile LU factorization with no pivoting - dynamic scheduling
*/
-void chameleon_pzgetrf( void *ws,
- CHAM_desc_t *A,
- RUNTIME_sequence_t *sequence,
- RUNTIME_request_t *request )
+void chameleon_pzgetrf( struct chameleon_pzgetrf_s *ws,
+ CHAM_desc_t *A,
+ RUNTIME_sequence_t *sequence,
+ RUNTIME_request_t *request )
{
CHAM_context_t *chamctxt;
RUNTIME_option_t options;
diff --git a/compute/zgetrf.c b/compute/zgetrf.c
index 52c33a740..c99ae3f80 100644
--- a/compute/zgetrf.c
+++ b/compute/zgetrf.c
@@ -4,7 +4,7 @@
*
* @copyright 2009-2014 The University of Tennessee and The University of
* Tennessee Research Foundation. All rights reserved.
- * @copyright 2012-2022 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria,
+ * @copyright 2012-2023 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria,
* Univ. Bordeaux. All rights reserved.
*
***
@@ -18,7 +18,7 @@
* @author Cedric Castagnede
* @author Florent Pruvost
* @author Matthieu Kuhn
- * @date 2022-09-19
+ * @date 2023-02-21
*
* @precisions normal z -> s d c
*
@@ -52,14 +52,26 @@
void *
CHAMELEON_zgetrf_WS_Alloc( const CHAM_desc_t *A )
{
- CHAM_context_t *chamctxt;
+ CHAM_context_t *chamctxt;
+ struct chameleon_pzgetrf_s *options;
chamctxt = chameleon_context_self();
if ( chamctxt == NULL ) {
return NULL;
}
- return NULL;
+ options = calloc( 1, sizeof( struct chameleon_pzgetrf_s ) );
+ options->ib = CHAMELEON_IB;
+
+#if defined(GETRF_NOPIV_PER_COLUMN)
+ chameleon_desc_init( &(options->U), CHAMELEON_MAT_ALLOC_TILE,
+ 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 );
+#endif
+
+ return options;
}
/**
@@ -84,14 +96,13 @@ CHAMELEON_zgetrf_WS_Alloc( const CHAM_desc_t *A )
void
CHAMELEON_zgetrf_WS_Free( const CHAM_desc_t *A, void *user_ws )
{
- CHAM_context_t *chamctxt;
+ struct chameleon_pzgetrf_s *ws = (struct chameleon_pzgetrf_s *)user_ws;
- chamctxt = chameleon_context_self();
- if ( chamctxt == NULL ) {
- return;
- }
+#if defined(GETRF_NOPIV_PER_COLUMN)
+ chameleon_desc_destroy( &(ws->U) );
+#endif
- (void)user_ws;
+ free( ws );
}
#if defined(NOT_AVAILABLE_YET)
diff --git a/control/compute_z.h b/control/compute_z.h
index b938b178f..bf9fdc8dc 100644
--- a/control/compute_z.h
+++ b/control/compute_z.h
@@ -21,7 +21,7 @@
* @author Florent Pruvost
* @author Alycia Lisito
* @author Matthieu Kuhn
- * @date 2022-09-19
+ * @date 2023-02-21
* @precisions normal z -> c d s
*
*/
@@ -37,6 +37,14 @@ struct chameleon_pzgemm_s {
CHAM_desc_t WB;
};
+/**
+ * @brief Data structure to handle the GETRF workspaces with partial pivoting
+ */
+struct chameleon_pzgetrf_s {
+ int ib; /* Internal blocking parameter */
+ CHAM_desc_t U;
+};
+
/**
* @brief Data structure to handle the Centering-Scaled workspaces
*/
@@ -78,7 +86,7 @@ void chameleon_pzgepdf_qdwh( cham_mtxtype_t trans, CHAM_desc_t *descU, CHAM_desc
void chameleon_pzgepdf_qr( int genD, int doqr, int optid, const libhqr_tree_t *qrtreeT, const libhqr_tree_t *qrtreeB, CHAM_desc_t *A1, CHAM_desc_t *TS1, CHAM_desc_t *TT1, CHAM_desc_t *D1, CHAM_desc_t *Q1, CHAM_desc_t *A2, CHAM_desc_t *TS2, CHAM_desc_t *TT2, CHAM_desc_t *D2, CHAM_desc_t *Q2, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request );
void chameleon_pzgeqrf( int genD, CHAM_desc_t *A, CHAM_desc_t *T, CHAM_desc_t *D, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
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( void *ws, CHAM_desc_t *A, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request );
+void chameleon_pzgetrf( struct chameleon_pzgetrf_s *ws, CHAM_desc_t *A, 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_reclap(CHAM_desc_t *A, int *IPIV, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
--
GitLab