diff --git a/compute/pzgetrf.c b/compute/pzgetrf.c
index 967e83f2fbb24a949412fb127a2f5f5cac9137f9..165801efb0e1bf4cfc5f5f36edf33def54063797 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 c99ae3f80b2d2c3afaee44832c10a5519bbccdf6..bcb8ee0c8c560863cf27248f180ce6af0d11e628 100644
--- a/compute/zgetrf.c
+++ b/compute/zgetrf.c
@@ -52,26 +52,44 @@
void *
CHAMELEON_zgetrf_WS_Alloc( const CHAM_desc_t *A )
{
- CHAM_context_t *chamctxt;
- struct chameleon_pzgetrf_s *options;
+ CHAM_context_t *chamctxt;
+ struct chameleon_pzgetrf_s *ws;
chamctxt = chameleon_context_self();
if ( chamctxt == NULL ) {
return NULL;
}
- options = calloc( 1, sizeof( struct chameleon_pzgetrf_s ) );
- options->ib = CHAMELEON_IB;
+ ws = calloc( 1, sizeof( struct chameleon_pzgetrf_s ) );
+ ws->alg = ChamGetrfNoPiv;
+ ws->ib = CHAMELEON_IB;
+
+ {
+ 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 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
+ 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 options;
+ return ws;
}
/**
@@ -94,14 +112,13 @@ CHAMELEON_zgetrf_WS_Alloc( const CHAM_desc_t *A )
*
*/
void
-CHAMELEON_zgetrf_WS_Free( const CHAM_desc_t *A, void *user_ws )
+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 );
}
@@ -210,7 +227,7 @@ CHAMELEON_zgetrf( int M, int N, CHAMELEON_Complex64_t *A, int *IPIV, int LDA )
chameleon_sequence_wait( chamctxt, sequence );
/* Cleanup the temporary data */
- CHAMELEON_zgetrf_WS_Free( &descAt, ws );
+ CHAMELEON_zgetrf_WS_Free( ws );
chameleon_ztile2lap_cleanup( chamctxt, &descAl, &descAt );
status = sequence->status;
@@ -275,7 +292,7 @@ CHAMELEON_zgetrf_Tile( CHAM_desc_t *A )
CHAMELEON_Desc_Flush( A, sequence );
chameleon_sequence_wait( chamctxt, sequence );
- CHAMELEON_zgetrf_WS_Free( A, ws );
+ CHAMELEON_zgetrf_WS_Free( ws );
status = sequence->status;
chameleon_sequence_destroy( chamctxt, sequence );
@@ -321,7 +338,8 @@ CHAMELEON_zgetrf_Tile_Async( CHAM_desc_t *A,
RUNTIME_sequence_t *sequence,
RUNTIME_request_t *request )
{
- CHAM_context_t *chamctxt;
+ CHAM_context_t *chamctxt;
+ struct chameleon_pzgetrf_s *ws;
chamctxt = chameleon_context_self();
if ( chamctxt == NULL ) {
@@ -364,7 +382,19 @@ CHAMELEON_zgetrf_Tile_Async( CHAM_desc_t *A,
return chameleon_request_fail( sequence, request, CHAMELEON_ERR_ILLEGAL_VALUE );
}
+ if ( user_ws == NULL ) {
+ ws = CHAMELEON_zgetrf_WS_Alloc( A );
+ }
+ else {
+ ws = user_ws;
+ }
+
chameleon_pzgetrf( user_ws, A, sequence, request );
+ if ( user_ws == NULL ) {
+ CHAMELEON_Desc_Flush( A, sequence );
+ chameleon_sequence_wait( chamctxt, sequence );
+ CHAMELEON_zgetrf_WS_Free( ws );
+ }
return CHAMELEON_SUCCESS;
}
diff --git a/control/compute_z.h b/control/compute_z.h
index bf9fdc8dca77911c439b8071b2264180bbf0ce19..1dd9b13055672181c935ebf67fe8ff1db6b0199e 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/chameleon_z.h b/include/chameleon/chameleon_z.h
index e587d015c4239a90fb10c228d9bb41faaca7dda1..1f04c61aad249f13d4653e94886eba891236da35 100644
--- a/include/chameleon/chameleon_z.h
+++ b/include/chameleon/chameleon_z.h
@@ -324,7 +324,7 @@ void CHAMELEON_zcesca_WS_Free( void *ws );
void *CHAMELEON_zgram_WS_Alloc( const CHAM_desc_t *A );
void CHAMELEON_zgram_WS_Free( void *ws );
void *CHAMELEON_zgetrf_WS_Alloc( const CHAM_desc_t *A );
-void CHAMELEON_zgetrf_WS_Free( const CHAM_desc_t *A, void *ws );
+void CHAMELEON_zgetrf_WS_Free( void *ws );
int CHAMELEON_Alloc_Workspace_zgesv_incpiv( int N, CHAM_desc_t **descL, int **IPIV, int p, int q);
int CHAMELEON_Alloc_Workspace_zgetrf_incpiv(int M, int N, CHAM_desc_t **descL, int **IPIV, int p, int q);
diff --git a/include/chameleon/constants.h b/include/chameleon/constants.h
index ba252005b72f862af1896ca1cac14ab0742f62a3..d3d2b43178ad3f0dcca7532a5c3f9b233875eb54 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
diff --git a/testing/testing_zgetrf.c b/testing/testing_zgetrf.c
index e075741c3f99764cae4b5c2eea957d5cc79d0dab..b41ee056dfbcdfb89590841041c11014b9f296d6 100644
--- a/testing/testing_zgetrf.c
+++ b/testing/testing_zgetrf.c
@@ -122,7 +122,7 @@ testing_zgetrf_desc( run_arg_list_t *args, int check )
#endif /* !defined(CHAMELEON_SIMULATION) */
if ( ws != NULL ) {
- CHAMELEON_zgetrf_WS_Free( descA, ws );
+ CHAMELEON_zgetrf_WS_Free( ws );
}
CHAMELEON_Desc_Destroy( &descA );