From 7ff0faacbeb264b59e189270e3cb8f8e26a6c4dd Mon Sep 17 00:00:00 2001
From: Mathieu Faverge <mathieu.faverge@inria.fr>
Date: Thu, 4 May 2023 15:25:52 +0200
Subject: [PATCH] getrf: add an alg parameter to dynamically switch from one
algorithm to another through environment variable
---
compute/pzgetrf.c | 18 +++++++++--------
compute/zgetrf.c | 38 ++++++++++++++++++++++++++---------
control/compute_z.h | 5 +++--
include/chameleon/constants.h | 8 ++++++++
4 files changed, 49 insertions(+), 20 deletions(-)
diff --git a/compute/pzgetrf.c b/compute/pzgetrf.c
index 967e83f2f..165801efb 100644
--- a/compute/pzgetrf.c
+++ b/compute/pzgetrf.c
@@ -113,11 +113,13 @@ chameleon_pzgetrf_panel_facto( struct chameleon_pzgetrf_s *ws,
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
+ /* TODO: Should be replaced by a function pointer */
+ if ( ws->alg == ChamGetrfNoPivPerColumn ) {
+ chameleon_pzgetrf_panel_facto_nopiv_percol( ws, A, k, options );
+ }
+ else {
+ chameleon_pzgetrf_panel_facto_nopiv( ws, A, k, options );
+ }
}
/**
@@ -180,9 +182,9 @@ chameleon_pzgetrf_panel_update( struct chameleon_pzgetrf_s *ws,
* Parallel tile LU factorization with no pivoting - dynamic scheduling
*/
void chameleon_pzgetrf( struct chameleon_pzgetrf_s *ws,
- CHAM_desc_t *A,
- RUNTIME_sequence_t *sequence,
- RUNTIME_request_t *request )
+ 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 97f010bb9..bcb8ee0c8 100644
--- a/compute/zgetrf.c
+++ b/compute/zgetrf.c
@@ -61,15 +61,33 @@ CHAMELEON_zgetrf_WS_Alloc( const CHAM_desc_t *A )
}
ws = calloc( 1, sizeof( struct chameleon_pzgetrf_s ) );
+ ws->alg = ChamGetrfNoPiv;
ws->ib = CHAMELEON_IB;
-#if defined(GETRF_NOPIV_PER_COLUMN)
- chameleon_desc_init( &(ws->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
+ {
+ char *algostr = chameleon_getenv( "CHAMELEON_GETRF_ALGO" );
+
+ if ( algostr ) {
+ if ( strcasecmp( algostr, "nopiv" ) ) {
+ ws->alg = ChamGetrfNoPiv;
+ }
+ else if ( strcasecmp( algostr, "nopivpercolumn" ) == 0 ) {
+ ws->alg = ChamGetrfNoPivPerColumn;
+ }
+ else {
+ fprintf( stderr, "ERROR: CHAMELEON_GETRF_ALGO is not one of NoPiv, NoPivPerColumn => Switch back to NoPiv\n" );
+ }
+ }
+ chameleon_cleanenv( algostr );
+ }
+
+ if ( ws->alg == ChamGetrfNoPivPerColumn ) {
+ chameleon_desc_init( &(ws->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 );
+ }
return ws;
}
@@ -98,9 +116,9 @@ CHAMELEON_zgetrf_WS_Free( void *user_ws )
{
struct chameleon_pzgetrf_s *ws = (struct chameleon_pzgetrf_s *)user_ws;
-#if defined(GETRF_NOPIV_PER_COLUMN)
- chameleon_desc_destroy( &(ws->U) );
-#endif
+ if ( ws->alg == ChamGetrfNoPivPerColumn ) {
+ chameleon_desc_destroy( &(ws->U) );
+ }
free( ws );
}
diff --git a/control/compute_z.h b/control/compute_z.h
index bf9fdc8dc..1dd9b1305 100644
--- a/control/compute_z.h
+++ b/control/compute_z.h
@@ -41,8 +41,9 @@ struct chameleon_pzgemm_s {
* @brief Data structure to handle the GETRF workspaces with partial pivoting
*/
struct chameleon_pzgetrf_s {
- int ib; /* Internal blocking parameter */
- CHAM_desc_t U;
+ cham_getrf_t alg;
+ int ib; /* Internal blocking parameter */
+ CHAM_desc_t U;
};
/**
diff --git a/include/chameleon/constants.h b/include/chameleon/constants.h
index ba252005b..d3d2b4317 100644
--- a/include/chameleon/constants.h
+++ b/include/chameleon/constants.h
@@ -194,6 +194,14 @@ typedef enum chameleon_gemm_e {
ChamGemmAlgSummaC
} cham_gemm_t;
+/**
+ * @brief Chameleon GETRF algorithm variants
+ */
+typedef enum chameleon_getrf_e {
+ ChamGetrfNoPiv,
+ ChamGetrfNoPivPerColumn,
+} cham_getrf_t;
+
#define ChameleonTrd 1001
#define ChameleonBrd 1002
--
GitLab