From 1f31e2b6731350362f47c2cd5811836c39356484 Mon Sep 17 00:00:00 2001
From: Alycia Lisito <alycia.lisito@inria.fr>
Date: Wed, 14 Feb 2024 16:11:37 +0100
Subject: [PATCH] zgetrf batched: Add an a field in the data structure and
environment variable to define the batch size
---
compute/zgetrf.c | 12 ++++++++----
control/compute_z.h | 3 ++-
2 files changed, 10 insertions(+), 5 deletions(-)
diff --git a/compute/zgetrf.c b/compute/zgetrf.c
index a1887f832..b15e25b11 100644
--- a/compute/zgetrf.c
+++ b/compute/zgetrf.c
@@ -88,6 +88,8 @@ CHAMELEON_zgetrf_WS_Alloc( const CHAM_desc_t *A )
chameleon_cleanenv( algostr );
}
+ ws->batch_size = chameleon_getenv_get_value_int( "CHAMELEON_GETRF_BATCH_SIZE", 1 );
+
/* Allocation of U for permutation of the panels */
if ( ws->alg == ChamGetrfNoPivPerColumn ) {
chameleon_desc_init( &(ws->U), CHAMELEON_MAT_ALLOC_TILE,
@@ -96,7 +98,7 @@ CHAMELEON_zgetrf_WS_Alloc( const CHAM_desc_t *A )
A->mt, A->nt * A->nb, A->p, A->q,
NULL, NULL, A->get_rankof_init, A->get_rankof_init_arg );
}
- else if ( ( ws->alg == ChamGetrfPPiv ) ||
+ else if ( ( ws->alg == ChamGetrfPPiv ) ||
( ws->alg == ChamGetrfPPivPerColumn ) )
{
chameleon_desc_init( &(ws->U), CHAMELEON_MAT_ALLOC_TILE,
@@ -108,13 +110,14 @@ CHAMELEON_zgetrf_WS_Alloc( const CHAM_desc_t *A )
/* Set ib to 1 if per column algorithm */
if ( ( ws->alg == ChamGetrfNoPivPerColumn ) ||
- ( ws->alg == ChamGetrfPPivPerColumn ) )
+ ( ws->alg == ChamGetrfPPivPerColumn ) )
{
ws->ib = 1;
}
/* Allocation of Up for the permutation of the diagonal panel per block */
- if ( ws->alg == ChamGetrfPPiv ) {
+ if ( ws->alg == ChamGetrfPPiv )
+ {
/* TODO: Should be restricted to diagonal tiles */
/* Possibly to a single handle with a permutation of the ownership */
chameleon_desc_init( &(ws->Up), CHAMELEON_MAT_ALLOC_TILE,
@@ -157,7 +160,8 @@ CHAMELEON_zgetrf_WS_Free( void *user_ws )
{
chameleon_desc_destroy( &(ws->U) );
}
- if ( ws->alg == ChamGetrfPPiv ) {
+ if ( ws->alg == ChamGetrfPPiv )
+ {
chameleon_desc_destroy( &(ws->Up) );
}
free( ws );
diff --git a/control/compute_z.h b/control/compute_z.h
index 2d0923155..645018f83 100644
--- a/control/compute_z.h
+++ b/control/compute_z.h
@@ -43,7 +43,8 @@ struct chameleon_pzgemm_s {
*/
struct chameleon_pzgetrf_s {
cham_getrf_t alg;
- int ib; /* Internal blocking parameter */
+ int ib; /**< Internal blocking parameter */
+ int batch_size; /**< Batch size for the panel */
CHAM_desc_t U;
CHAM_desc_t Up;
};
--
GitLab