diff --git a/ChangeLog b/ChangeLog
index 2517d5d027d2a8c248466b97428388ce6cdab625..1e299d40f81f0d0b6ca075464c76cffbc7a3224d 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,6 +1,6 @@
chameleon-1.3.0
------------------------------------------------------------------------
- - mixed-precision: introduce descripto with precision adapted to local norms
+ - mixed-precision: introduce descriptor with precision adapted to local norms
- Add CHAMELEON_[dz]gered... functions to reduce the precision of the tiles based on a requested accuracy
- Add CHAMELEON_[dz]gerst... functions to restore the original numerical precision of the tiles in a descriptor
- types: add support for half precision arithmetic into the data descriptors
@@ -11,6 +11,8 @@ chameleon-1.3.0
CHAMELEON_Desc_Create_User that requires aan additional `, NULL`
parameters in the general case.
- compute/poinv: Add the possibility to use an intermediate distribution for the TRTRI operation
+ - compute/getrf_nopiv: Add lookahead through temporary buffers to better regulate the communication allocations
+ - runtime/starpu: Whenever possible replace the lacpy codelet by a direct memory copy from the input handler to the output one
chameleon-1.2.0
------------------------------------------------------------------------
diff --git a/compute/pzgepdf_qdwh.c b/compute/pzgepdf_qdwh.c
index 420ad6f502ceccea3a253363e5b00333c5f49aea..0d1b66cb267f374b5ce493c01a2ecb619603d645 100644
--- a/compute/pzgepdf_qdwh.c
+++ b/compute/pzgepdf_qdwh.c
@@ -15,7 +15,7 @@
* @author Mathieu Faverge
* @author Hatem Ltaief
* @author Lionel Eyraud-Dubois
- * @date 2023-07-05
+ * @date 2024-10-17
* @precisions normal z -> s d c
*
*/
@@ -35,8 +35,8 @@ static int _zgepdf_qdwh_opt_genD = 1;
static int _zgepdf_qdwh_opt_genD = 0;
#endif
-static int _zgepdf_qdwh_opt_qr = 1;
-static int _zgepdf_qdwh_opt_id = 1; // There is a numerical issue when combining this optimization and the StarPU lacpy
+static int _zgepdf_qdwh_opt_qr = 1;
+static int _zgepdf_qdwh_opt_id = 1;
static int _zgepdf_qdwh_verbose = 0;
/**
@@ -603,6 +603,7 @@ chameleon_pzgepdf_qdwh( cham_mtxtype_t mtxtype, CHAM_desc_t *descU, CHAM_desc_t
double conv = 100.;
double normest, Unorm;
int it, itconv, facto = -1;
+ cham_bool_t optlacpy_backup;
double eps = CHAMELEON_dlamch();
double tol1 = 5. * eps;
@@ -615,6 +616,10 @@ chameleon_pzgepdf_qdwh( cham_mtxtype_t mtxtype, CHAM_desc_t *descU, CHAM_desc_t
}
assert( chamctxt->scheduler != RUNTIME_SCHED_PARSEC );
+ /* Force unoptimized lacpy */
+ optlacpy_backup = chamctxt->optlacpy_enabled;
+ chamctxt->optlacpy_enabled = CHAMELEON_FALSE;
+
if ( info ) {
info->itQR = 0;
info->itPO = 0;
@@ -848,5 +853,8 @@ chameleon_pzgepdf_qdwh( cham_mtxtype_t mtxtype, CHAM_desc_t *descU, CHAM_desc_t
&descB2, &descTS2, &descTT2, &descQ2, &descD2 );
CHAMELEON_zgemm_WS_Free( gemm_ws );
+ /* Restore optimized lacpy value */
+ chamctxt->optlacpy_enabled = optlacpy_backup;
+
return;
}
diff --git a/compute/pzgetrf_nopiv.c b/compute/pzgetrf_nopiv.c
index daac4bad9652cd1512b023431c5ad1b325cd8c5c..18fb33ed81656b462cdcff0f52fe82e6080b90de 100644
--- a/compute/pzgetrf_nopiv.c
+++ b/compute/pzgetrf_nopiv.c
@@ -11,7 +11,7 @@
*
* @brief Chameleon zgetrf_nopiv parallel algorithm
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Omar Zenati
* @author Mathieu Faverge
* @author Emmanuel Agullo
@@ -20,20 +20,23 @@
* @author Samuel Thibault
* @author Terry Cojean
* @author Matthieu Kuhn
- * @date 2022-02-22
+ * @date 2024-10-17
* @precisions normal z -> s d c
*
*/
#include "control/common.h"
-#define A(m,n) A, m, n
+#define A(m, n) A, m, n
+#define WD(m) WL, m, m
+#define WL(m, n) WL, m, n
+#define WU(m, n) WU, m, n
/**
* Parallel tile LU factorization with no pivoting - dynamic scheduling
*/
-void chameleon_pzgetrf_nopiv( CHAM_desc_t *A,
- RUNTIME_sequence_t *sequence,
- RUNTIME_request_t *request )
+void chameleon_pzgetrf_nopiv_generic( CHAM_desc_t *A,
+ RUNTIME_sequence_t *sequence,
+ RUNTIME_request_t *request )
{
CHAM_context_t *chamctxt;
RUNTIME_option_t options;
@@ -121,3 +124,196 @@ void chameleon_pzgetrf_nopiv( CHAM_desc_t *A,
RUNTIME_options_finalize(&options, chamctxt);
}
+
+void chameleon_pzgetrf_nopiv_ws( CHAM_desc_t *A,
+ CHAM_desc_t *WL,
+ CHAM_desc_t *WU,
+ RUNTIME_sequence_t *sequence,
+ RUNTIME_request_t *request )
+{
+ CHAM_context_t *chamctxt;
+ RUNTIME_option_t options;
+
+ int k, m, n, ib, p, q, lp, lq;
+ int tempkm, tempkn, tempmm, tempnn;
+ int lookahead, myp, myq;
+
+ CHAMELEON_Complex64_t zone = (CHAMELEON_Complex64_t) 1.0;
+ CHAMELEON_Complex64_t mzone = (CHAMELEON_Complex64_t)-1.0;
+
+ chamctxt = chameleon_context_self();
+ if (sequence->status != CHAMELEON_SUCCESS) {
+ return;
+ }
+ RUNTIME_options_init(&options, chamctxt, sequence, request);
+
+ ib = CHAMELEON_IB;
+ lookahead = chamctxt->lookahead;
+ myp = A->myrank / A->q;
+ myq = A->myrank % A->q;
+
+ for (k = 0; k < chameleon_min(A->mt, A->nt); k++) {
+ RUNTIME_iteration_push(chamctxt, k);
+ lp = (k % lookahead) * A->p;
+ lq = (k % lookahead) * A->q;
+
+ tempkm = k == A->mt-1 ? A->m-k*A->mb : A->mb;
+ tempkn = k == A->nt-1 ? A->n-k*A->nb : A->nb;
+
+ options.priority = 2*A->nt - 2*k;
+ INSERT_TASK_zgetrf_nopiv(
+ &options,
+ tempkm, tempkn, ib, A->mb,
+ A(k, k), A->mb*k);
+
+ /**
+ * Broadcast of A(k,k) along rings in both directions
+ */
+ {
+ INSERT_TASK_zlacpy(
+ &options,
+ ChamUpperLower, tempkm, tempkn,
+ A( k, k ),
+ WL( k, (k % A->q) + lq ) );
+ INSERT_TASK_zlacpy(
+ &options,
+ ChamUpperLower, tempkm, tempkn,
+ A( k, k ),
+ WU( (k % A->p) + lp, k ) );
+
+ for ( q=1; q < A->q; q++ ) {
+ INSERT_TASK_zlacpy(
+ &options,
+ ChamUpperLower, tempkm, tempkn,
+ WL( k, ((k+q-1) % A->q) + lq ),
+ WL( k, ((k+q) % A->q) + lq ) );
+ }
+
+ for ( p=1; p < A->p; p++ ) {
+ INSERT_TASK_zlacpy(
+ &options,
+ ChamUpperLower, tempkm, tempkn,
+ WU( ((k+p-1) % A->p) + lp, k ),
+ WU( ((k+p) % A->p) + lp, k ) );
+ }
+ }
+ RUNTIME_data_flush( sequence, A( k, k ) );
+
+ for (m = k+1; m < A->mt; m++) {
+
+ /* Skip the row if you are not involved with */
+ if ( m%A->p != myp ) {
+ continue;
+ }
+
+ options.priority = 2*A->nt - 2*k - m;
+ tempmm = m == A->mt-1 ? A->m-m*A->mb : A->mb;
+
+ assert( A->get_rankof( A, m, k ) == WU->get_rankof( WU, myp + lp, k) );
+ INSERT_TASK_ztrsm(
+ &options,
+ ChamRight, ChamUpper, ChamNoTrans, ChamNonUnit,
+ tempmm, tempkn, A->mb,
+ zone, WU( myp + lp, k ),
+ A( m, k ) );
+
+ /* Broadcast A(m,k) into temp buffers through a ring */
+ {
+ assert( A->get_rankof( A, m, k ) == WL->get_rankof( WL, m, (k % A->q) + lq) );
+ INSERT_TASK_zlacpy(
+ &options,
+ ChamUpperLower, tempmm, tempkn,
+ A( m, k ),
+ WL( m, (k % A->q) + lq) );
+
+ for ( q=1; q < A->q; q++ ) {
+ INSERT_TASK_zlacpy(
+ &options,
+ ChamUpperLower, tempmm, tempkn,
+ WL( m, ((k+q-1) % A->q) + lq ),
+ WL( m, ((k+q) % A->q) + lq ) );
+ }
+ }
+ RUNTIME_data_flush( sequence, A( m, k ) );
+ }
+
+ for (n = k+1; n < A->nt; n++) {
+
+ /* Skip the column if you are not involved with */
+ if ( n%A->q != myq ) {
+ continue;
+ }
+
+ tempnn = n == A->nt-1 ? A->n-n*A->nb : A->nb;
+ options.priority = 2*A->nt - 2*k - n;
+
+ assert( A->get_rankof( A, k, n ) == WL->get_rankof( WL, k, myq+lq) );
+ INSERT_TASK_ztrsm(
+ &options,
+ ChamLeft, ChamLower, ChamNoTrans, ChamUnit,
+ tempkm, tempnn, A->mb,
+ zone, WL( k, myq + lq ),
+ A( k, n ));
+
+ /* Broadcast A(k,n) into temp buffers through a ring */
+ {
+ assert( A->get_rankof( A, k, n ) == WU->get_rankof( WU, (k%A->p) + lp, n) );
+ INSERT_TASK_zlacpy(
+ &options,
+ ChamUpperLower, tempkm, tempnn,
+ A( k, n ),
+ WU( (k % A->p) + lp, n ) );
+
+ for ( p=1; p < A->p; p++ ) {
+ INSERT_TASK_zlacpy(
+ &options,
+ ChamUpperLower, tempkm, tempnn,
+ WU( ((k+p-1) % A->p) + lp, n ),
+ WU( ((k+p) % A->p) + lp, n ) );
+ }
+ }
+ RUNTIME_data_flush( sequence, A( k, n ) );
+
+ for (m = k+1; m < A->mt; m++) {
+
+ /* Skip the row if you are not involved with */
+ if ( m%A->p != myp ) {
+ continue;
+ }
+
+ tempmm = m == A->mt-1 ? A->m-m*A->mb : A->mb;
+ options.priority = 2*A->nt - 2*k - n - m;
+
+ assert( A->get_rankof( A, m, n ) == WL->get_rankof( WL, m, myq + lq) );
+ assert( A->get_rankof( A, m, n ) == WU->get_rankof( WU, myp + lp, n) );
+
+ INSERT_TASK_zgemm(
+ &options,
+ ChamNoTrans, ChamNoTrans,
+ tempmm, tempnn, A->mb, A->mb,
+ mzone, WL( m, myq + lq ),
+ WU( myp + lp, n ),
+ zone, A( m, n ));
+ }
+ }
+ RUNTIME_iteration_pop( chamctxt );
+ }
+
+ CHAMELEON_Desc_Flush( WL, sequence );
+ CHAMELEON_Desc_Flush( WU, sequence );
+
+ RUNTIME_options_finalize( &options, chamctxt );
+}
+
+void chameleon_pzgetrf_nopiv( struct chameleon_pzgetrf_nopiv_s *ws,
+ CHAM_desc_t *A,
+ RUNTIME_sequence_t *sequence,
+ RUNTIME_request_t *request )
+{
+ if ( ws && ws->use_workspace ) {
+ chameleon_pzgetrf_nopiv_ws( A, &(ws->WL), &(ws->WU), sequence, request );
+ }
+ else {
+ chameleon_pzgetrf_nopiv_generic( A, sequence, request );
+ }
+}
diff --git a/compute/zgesv_nopiv.c b/compute/zgesv_nopiv.c
index c7c703a45dcf31b7129d31b5baefd549fb78d5f9..cf5c756eb00259ded3da6bc1c8dcfb5965768089 100644
--- a/compute/zgesv_nopiv.c
+++ b/compute/zgesv_nopiv.c
@@ -11,7 +11,7 @@
*
* @brief Chameleon zgesv_nopiv 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 Jakub Kurzak
@@ -19,7 +19,8 @@
* @author Emmanuel Agullo
* @author Cedric Castagnede
* @author Florent Pruvost
- * @date 2022-02-22
+ * @author Matthieu Kuhn
+ * @date 2024-10-17
* @precisions normal z -> s d c
*
*/
@@ -82,8 +83,8 @@
*
*/
int CHAMELEON_zgesv_nopiv( int N, int NRHS,
- CHAMELEON_Complex64_t *A, int LDA,
- CHAMELEON_Complex64_t *B, int LDB )
+ CHAMELEON_Complex64_t *A, int LDA,
+ CHAMELEON_Complex64_t *B, int LDB )
{
int NB;
int status;
@@ -92,6 +93,7 @@ int CHAMELEON_zgesv_nopiv( int N, int NRHS,
RUNTIME_request_t request = RUNTIME_REQUEST_INITIALIZER;
CHAM_desc_t descAl, descAt;
CHAM_desc_t descBl, descBt;
+ void *ws = NULL;
chamctxt = chameleon_context_self();
if (chamctxt == NULL) {
@@ -138,7 +140,8 @@ int CHAMELEON_zgesv_nopiv( int N, int NRHS,
B, NB, NB, LDB, NRHS, N, NRHS, sequence, &request );
/* Call the tile interface */
- CHAMELEON_zgesv_nopiv_Tile_Async( &descAt, &descBt, sequence, &request );
+ ws = CHAMELEON_zgetrf_nopiv_WS_Alloc( &descAt );
+ CHAMELEON_zgesv_nopiv_Tile_Async( &descAt, &descBt, ws, sequence, &request );
/* Submit the matrix conversion back */
chameleon_ztile2lap( chamctxt, &descAl, &descAt,
@@ -149,6 +152,7 @@ int CHAMELEON_zgesv_nopiv( int N, int NRHS,
chameleon_sequence_wait( chamctxt, sequence );
/* Cleanup the temporary data */
+ CHAMELEON_zgetrf_nopiv_WS_Free( ws );
chameleon_ztile2lap_cleanup( chamctxt, &descAl, &descAt );
chameleon_ztile2lap_cleanup( chamctxt, &descBl, &descBt );
@@ -195,10 +199,11 @@ int CHAMELEON_zgesv_nopiv( int N, int NRHS,
*/
int CHAMELEON_zgesv_nopiv_Tile( CHAM_desc_t *A, CHAM_desc_t *B )
{
- CHAM_context_t *chamctxt;
+ CHAM_context_t *chamctxt;
RUNTIME_sequence_t *sequence = NULL;
- RUNTIME_request_t request = RUNTIME_REQUEST_INITIALIZER;
- int status;
+ RUNTIME_request_t request = RUNTIME_REQUEST_INITIALIZER;
+ int status;
+ void *ws;
chamctxt = chameleon_context_self();
if (chamctxt == NULL) {
@@ -207,12 +212,15 @@ int CHAMELEON_zgesv_nopiv_Tile( CHAM_desc_t *A, CHAM_desc_t *B )
}
chameleon_sequence_create( chamctxt, &sequence );
- CHAMELEON_zgesv_nopiv_Tile_Async( A, B, sequence, &request );
+ ws = CHAMELEON_zgetrf_nopiv_WS_Alloc( A );
+ CHAMELEON_zgesv_nopiv_Tile_Async( A, B, ws, sequence, &request );
CHAMELEON_Desc_Flush( A, sequence );
CHAMELEON_Desc_Flush( B, sequence );
chameleon_sequence_wait( chamctxt, sequence );
+ CHAMELEON_zgetrf_nopiv_WS_Free( ws );
+
status = sequence->status;
chameleon_sequence_destroy( chamctxt, sequence );
return status;
@@ -248,10 +256,14 @@ int CHAMELEON_zgesv_nopiv_Tile( CHAM_desc_t *A, CHAM_desc_t *B )
* @sa CHAMELEON_zcgesv_Tile_Async
*
*/
-int CHAMELEON_zgesv_nopiv_Tile_Async( CHAM_desc_t *A, CHAM_desc_t *B,
- RUNTIME_sequence_t *sequence, RUNTIME_request_t *request )
+int CHAMELEON_zgesv_nopiv_Tile_Async( CHAM_desc_t *A,
+ CHAM_desc_t *B,
+ void *user_ws,
+ RUNTIME_sequence_t *sequence,
+ RUNTIME_request_t *request )
{
- CHAM_context_t *chamctxt;
+ CHAM_context_t *chamctxt;
+ struct chameleon_pzgetrf_nopiv_s *ws;
chamctxt = chameleon_context_self();
if (chamctxt == NULL) {
@@ -294,11 +306,23 @@ int CHAMELEON_zgesv_nopiv_Tile_Async( CHAM_desc_t *A, CHAM_desc_t *B,
return CHAMELEON_SUCCESS;
*/
- chameleon_pzgetrf_nopiv( A, sequence, request );
+ if ( user_ws == NULL ) {
+ ws = CHAMELEON_zgetrf_nopiv_WS_Alloc( A );
+ }
+ else {
+ ws = user_ws;
+ }
+ chameleon_pzgetrf_nopiv( ws, A, sequence, request );
chameleon_pztrsm( ChamLeft, ChamLower, ChamNoTrans, ChamUnit, (CHAMELEON_Complex64_t)1.0, A, B, sequence, request );
chameleon_pztrsm( ChamLeft, ChamUpper, ChamNoTrans, ChamNonUnit, (CHAMELEON_Complex64_t)1.0, A, B, sequence, request );
+ if ( user_ws == NULL ) {
+ CHAMELEON_Desc_Flush( A, sequence );
+ CHAMELEON_Desc_Flush( B, sequence );
+ chameleon_sequence_wait( chamctxt, sequence );
+ CHAMELEON_zgetrf_nopiv_WS_Free( ws );
+ }
return CHAMELEON_SUCCESS;
}
diff --git a/compute/zgetrf_nopiv.c b/compute/zgetrf_nopiv.c
index d6e1c27ec6423bb5c0db35a0f96615003aa4fbe8..3409e97f7a80cfb0844b4eefa66d6646be6239f1 100644
--- a/compute/zgetrf_nopiv.c
+++ b/compute/zgetrf_nopiv.c
@@ -11,19 +11,109 @@
*
* @brief Chameleon zgetrf_nopiv 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 Alycia Lisito
- * @date 2022-02-22
+ * @author Matthieu Kuhn
+ * @date 2024-10-17
*
* @precisions normal z -> s d c
*
*/
#include "control/common.h"
+/**
+ ********************************************************************************
+ *
+ * @ingroup CHAMELEON_Complex64_t
+ *
+ * @brief Allocate the required workspaces for asynchronous getrf
+ *
+ *******************************************************************************
+ *
+ * @param[in] A
+ * The descriptor of the matrix A.
+ *
+ *******************************************************************************
+ *
+ * @retval An allocated opaque pointer to use in CHAMELEON_zgetrf_nopiv_Tile_Async()
+ * and to free with CHAMELEON_zgetrf_nopiv_WS_Free().
+ *
+ *******************************************************************************
+ *
+ * @sa CHAMELEON_zgetrf_nopiv_Tile_Async
+ * @sa CHAMELEON_zgetrf_nopiv_WS_Free
+ *
+ */
+void *CHAMELEON_zgetrf_nopiv_WS_Alloc( const CHAM_desc_t *A )
+{
+ CHAM_context_t *chamctxt;
+ struct chameleon_pzgetrf_nopiv_s *options;
+
+ chamctxt = chameleon_context_self();
+ if ( chamctxt == NULL ) {
+ return NULL;
+ }
+
+ options = calloc( 1, sizeof(struct chameleon_pzgetrf_nopiv_s) );
+ options->use_workspace = 0;
+
+ if ( ( ( A->p > 1 ) || ( A->q > 1 ) ) &&
+ ( A->get_rankof_init == chameleon_getrankof_2d ) &&
+ ( chamctxt->generic_enabled != CHAMELEON_TRUE ) )
+ {
+ int lookahead = chamctxt->lookahead;
+ options->use_workspace = 1;
+
+ chameleon_desc_init( &(options->WL), CHAMELEON_MAT_ALLOC_TILE,
+ ChamComplexDouble, A->mb, A->nb, (A->mb * A->nb),
+ A->mt * A->mb, A->nb * A->q * lookahead, 0, 0,
+ A->mt * A->mb, A->nb * A->q * lookahead, A->p, A->q,
+ NULL, NULL, A->get_rankof_init, A->get_rankof_init_arg );
+
+ chameleon_desc_init( &(options->WU), CHAMELEON_MAT_ALLOC_TILE,
+ ChamComplexDouble,
+ A->mb, A->nb, (A->mb * A->nb),
+ A->mb * A->p * lookahead, A->nt * A->nb, 0, 0,
+ A->mb * A->p * lookahead, A->nt * A->nb, A->p, A->q,
+ NULL, NULL, A->get_rankof_init, A->get_rankof_init_arg );
+ }
+
+ return (void*)options;
+}
+
+/**
+ ********************************************************************************
+ *
+ * @ingroup CHAMELEON_Complex64_t
+ *
+ * @brief Free the allocated workspaces for asynchronous getrf
+ *
+ *******************************************************************************
+ *
+ * @param[in,out] user_ws
+ * On entry, the opaque pointer allocated by CHAMELEON_zgetrf_nopiv_WS_Alloc()
+ * On exit, all data are freed.
+ *
+ *******************************************************************************
+ *
+ * @sa CHAMELEON_zgetrf_nopiv_Tile_Async
+ * @sa CHAMELEON_zgetrf_nopiv_WS_Alloc
+ *
+ */
+void CHAMELEON_zgetrf_nopiv_WS_Free( void *user_ws )
+{
+ struct chameleon_pzgetrf_nopiv_s *ws = (struct chameleon_pzgetrf_nopiv_s*)user_ws;
+
+ if ( ws->use_workspace ) {
+ chameleon_desc_destroy( &(ws->WL) );
+ chameleon_desc_destroy( &(ws->WU) );
+ }
+ free( ws );
+}
/**
********************************************************************************
@@ -69,7 +159,7 @@
*
*/
int CHAMELEON_zgetrf_nopiv( int M, int N,
- CHAMELEON_Complex64_t *A, int LDA )
+ CHAMELEON_Complex64_t *A, int LDA )
{
int NB;
int status;
@@ -77,6 +167,7 @@ int CHAMELEON_zgetrf_nopiv( int M, int N,
CHAM_context_t *chamctxt;
RUNTIME_sequence_t *sequence = NULL;
RUNTIME_request_t request = RUNTIME_REQUEST_INITIALIZER;
+ void *ws = NULL;
chamctxt = chameleon_context_self();
if (chamctxt == NULL) {
@@ -117,7 +208,8 @@ int CHAMELEON_zgetrf_nopiv( int M, int N,
A, NB, NB, LDA, N, M, N, sequence, &request );
/* Call the tile interface */
- CHAMELEON_zgetrf_nopiv_Tile_Async( &descAt, sequence, &request );
+ ws = CHAMELEON_zgetrf_nopiv_WS_Alloc( &descAt );
+ CHAMELEON_zgetrf_nopiv_Tile_Async( &descAt, ws, sequence, &request );
/* Submit the matrix conversion back */
chameleon_ztile2lap( chamctxt, &descAl, &descAt,
@@ -126,6 +218,7 @@ int CHAMELEON_zgetrf_nopiv( int M, int N,
chameleon_sequence_wait( chamctxt, sequence );
/* Cleanup the temporary data */
+ CHAMELEON_zgetrf_nopiv_WS_Free( ws );
chameleon_ztile2lap_cleanup( chamctxt, &descAl, &descAt );
status = sequence->status;
@@ -169,10 +262,11 @@ int CHAMELEON_zgetrf_nopiv( int M, int N,
*/
int CHAMELEON_zgetrf_nopiv_Tile( CHAM_desc_t *A )
{
- CHAM_context_t *chamctxt;
+ CHAM_context_t *chamctxt;
RUNTIME_sequence_t *sequence = NULL;
- RUNTIME_request_t request = RUNTIME_REQUEST_INITIALIZER;
- int status;
+ RUNTIME_request_t request = RUNTIME_REQUEST_INITIALIZER;
+ int status;
+ void *ws;
chamctxt = chameleon_context_self();
if (chamctxt == NULL) {
@@ -181,11 +275,14 @@ int CHAMELEON_zgetrf_nopiv_Tile( CHAM_desc_t *A )
}
chameleon_sequence_create( chamctxt, &sequence );
- CHAMELEON_zgetrf_nopiv_Tile_Async( A, sequence, &request );
+ ws = CHAMELEON_zgetrf_nopiv_WS_Alloc( A );
+ CHAMELEON_zgetrf_nopiv_Tile_Async( A, ws, sequence, &request );
CHAMELEON_Desc_Flush( A, sequence );
chameleon_sequence_wait( chamctxt, sequence );
+ CHAMELEON_zgetrf_nopiv_WS_Free( ws );
+
status = sequence->status;
chameleon_sequence_destroy( chamctxt, sequence );
return status;
@@ -224,11 +321,13 @@ int CHAMELEON_zgetrf_nopiv_Tile( CHAM_desc_t *A )
* @sa CHAMELEON_zgetrs_Tile_Async
*
*/
-int CHAMELEON_zgetrf_nopiv_Tile_Async( CHAM_desc_t *A,
- RUNTIME_sequence_t *sequence,
- RUNTIME_request_t *request )
+int CHAMELEON_zgetrf_nopiv_Tile_Async( CHAM_desc_t *A,
+ void *user_ws,
+ RUNTIME_sequence_t *sequence,
+ RUNTIME_request_t *request )
{
- CHAM_context_t *chamctxt;
+ CHAM_context_t *chamctxt;
+ struct chameleon_pzgetrf_nopiv_s *ws;
chamctxt = chameleon_context_self();
if (chamctxt == NULL) {
@@ -263,7 +362,19 @@ int CHAMELEON_zgetrf_nopiv_Tile_Async( CHAM_desc_t *A,
return chameleon_request_fail(sequence, request, CHAMELEON_ERR_ILLEGAL_VALUE);
}
- chameleon_pzgetrf_nopiv( A, sequence, request );
+ if ( user_ws == NULL ) {
+ ws = CHAMELEON_zgetrf_nopiv_WS_Alloc( A );
+ }
+ else {
+ ws = user_ws;
+ }
+
+ chameleon_pzgetrf_nopiv( ws, A, sequence, request );
+ if ( user_ws == NULL ) {
+ CHAMELEON_Desc_Flush( A, sequence );
+ chameleon_sequence_wait( chamctxt, sequence );
+ CHAMELEON_zgetrf_nopiv_WS_Free( ws );
+ }
return CHAMELEON_SUCCESS;
}
diff --git a/control/chameleon_zf77.c b/control/chameleon_zf77.c
index 70541b792b2b3ee527c1737af4f314a86a7f57d3..64fad284a4f118b79bef98e0c622f415e3a2ac0a 100644
--- a/control/chameleon_zf77.c
+++ b/control/chameleon_zf77.c
@@ -11,7 +11,7 @@
*
* @brief Chameleon Fortran77 computational routines
*
- * @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
* @comment This file is automatically generated by tools/genf77interface.pl
@@ -21,7 +21,8 @@
* @author Cedric Castagnede
* @author Florent Pruvost
* @author Alycia Lisito
- * @date 2022-02-22
+ * @author Matthieu Kuhn
+ * @date 2024-10-17
* @precisions normal z -> c d s
*
*/
@@ -727,7 +728,7 @@ void CHAMELEON_ZGETRF_INCPIV_TILE_ASYNC(CHAM_desc_t *A, CHAM_desc_t *L, int *IPI
{ *info = CHAMELEON_zgetrf_incpiv_Tile_Async(A, L, IPIV, sequence, request); }
void CHAMELEON_ZGETRF_NOPIV_TILE_ASYNC(CHAM_desc_t *A, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request, int *info)
-{ *info = CHAMELEON_zgetrf_nopiv_Tile_Async(A, sequence, request); }
+{ *info = CHAMELEON_zgetrf_nopiv_Tile_Async(A, NULL, sequence, request); }
//void CHAMELEON_ZGETRI_TILE_ASYNC(CHAM_desc_t *A, int *IPIV, CHAM_desc_t *W, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request, int *info)
//{ *info = CHAMELEON_zgetri_Tile_Async(A, IPIV, W, sequence, request); }
diff --git a/control/compute_z.h b/control/compute_z.h
index 06c8854c1b05a7c1bffbca0fa8e615218363672d..c0bac97556b86fbe2aaa5b61abcb9a645abf246b 100644
--- a/control/compute_z.h
+++ b/control/compute_z.h
@@ -23,7 +23,7 @@
* @author Matthieu Kuhn
* @author Lionel Eyraud-Dubois
* @author Ana Hourcau
- * @date 2024-07-17
+ * @date 2024-10-17
* @precisions normal z -> c d s
*
*/
@@ -52,6 +52,49 @@ struct chameleon_pzgetrf_s {
int involved:1;
};
+/**
+ * @brief Data structure to handle the GETRF temporary workspaces
+ * for MPI transfers.
+ *
+ * @comment The idea is to manage explicitely temporary
+ * blocks arising from MPI transfers automatically
+ * inferred by StarPU, hence limiting the total number
+ * of temporary data allocated for these blocks.
+ *
+ * The blocks to be sent/received on the network are
+ * copied into those buffers. These copies are
+ * then used by the algorithm in place of the regular
+ * blocks of the problem matrix.
+ *
+ * For WL (resp. WU), the number of allocated blocks
+ * corresponds to the number of blocks on the column
+ * (resp. on the line) multiplied by lookahead number
+ * from the current chameleon context.
+ *
+ * Then, depending on the block panel index, we access
+ * one of the temporary column blocks of WL and row blocks
+ * of WU int a circular way.
+ *
+ * For instance, for the block panel index k, the block
+ * A(m,k) produced by the TRSM(A(k,k),A(m,k)) is stored
+ * into temporary buffer WL(m,k%chamctxt->lookahead).
+ * Similarly, the block A(k,n) is stored into the temporary
+ * block WU(k%chamctxt->lookahead, n).
+ *
+ * Notice that, by doing so, the notion of look ahead is
+ * reintroduced : artificial dependencies are implied by
+ * the circular usage of WL and WU temporary workspaces.
+ *
+ */
+struct chameleon_pzgetrf_nopiv_s {
+ int use_workspace;
+
+ CHAM_desc_t WL; /* Workspace to store temporary blocks of the */
+ /* diagonal and the lower part of the problem matrix */
+ CHAM_desc_t WU; /* Workspace to store temporary blocks of the */
+ /* upper part of the problem matrix */
+};
+
/**
* @brief Data structure to handle the Centering-Scaled workspaces
*/
@@ -103,7 +146,7 @@ void chameleon_pzgeqrf( int genD, CHAM_desc_t *A, CHAM_desc_t *T, CHAM_desc_t *D
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( struct chameleon_pzgetrf_s *ws, CHAM_desc_t *A, CHAM_ipiv_t *IPIV, 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_nopiv(struct chameleon_pzgetrf_nopiv_s *ws, 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);
void chameleon_pzgetrf_rectil(CHAM_desc_t *A, int *IPIV, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
void chameleon_pzhegst(int itype, cham_uplo_t uplo, CHAM_desc_t *A, CHAM_desc_t *B, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
diff --git a/control/context.c b/control/context.c
index cfbe13a714ec97d1f817aaff6e4762e7e85b7ef4..5bcd6d598f9c2a6d90590934b51787a7e0b50162 100644
--- a/control/context.c
+++ b/control/context.c
@@ -21,7 +21,7 @@
* @author Matthieu Kuhn
* @author Loris Lucido
* @author Terry Cojean
- * @date 2023-09-11
+ * @date 2024-10-17
*
***
*
@@ -140,6 +140,7 @@ CHAM_context_t *chameleon_context_create()
chamctxt->progress_enabled = chameleon_env_on_off( "CHAMELEON_PROGRESS", CHAMELEON_FALSE );
chamctxt->generic_enabled = chameleon_env_on_off( "CHAMELEON_GENERIC", CHAMELEON_FALSE );
chamctxt->autominmax_enabled = chameleon_env_on_off( "CHAMELEON_AUTOMINMAX", CHAMELEON_TRUE );
+ chamctxt->optlacpy_enabled = chameleon_env_on_off( "CHAMELEON_OPTIMIZED_LACPY", CHAMELEON_TRUE );
chamctxt->runtime_paused = CHAMELEON_FALSE;
diff --git a/include/chameleon/chameleon_z.h b/include/chameleon/chameleon_z.h
index 3f33260f4436ec195323874d3c13b9b44d2c62e4..9b085201650e266247ade2308e6522b58d4070a9 100644
--- a/include/chameleon/chameleon_z.h
+++ b/include/chameleon/chameleon_z.h
@@ -24,7 +24,7 @@
* @author Alycia Lisito
* @author Matthieu Kuhn
* @author Ana Hourcau
- * @date 2024-07-17
+ * @date 2024-10-17
* @precisions normal z -> c d s
*
*/
@@ -211,11 +211,11 @@ int CHAMELEON_zgeqrf_Tile_Async(CHAM_desc_t *A, CHAM_desc_t *T, RUNTIME_sequence
int CHAMELEON_zgeqrs_Tile_Async(CHAM_desc_t *A, CHAM_desc_t *T, CHAM_desc_t *B, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
//int CHAMELEON_zgesv_Tile_Async(CHAM_desc_t *A, int *IPIV, CHAM_desc_t *B, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
int CHAMELEON_zgesv_incpiv_Tile_Async(CHAM_desc_t *A, CHAM_desc_t *L, int *IPIV, CHAM_desc_t *B, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
-int CHAMELEON_zgesv_nopiv_Tile_Async(CHAM_desc_t *A, CHAM_desc_t *B, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
+int CHAMELEON_zgesv_nopiv_Tile_Async(CHAM_desc_t *A, CHAM_desc_t *B, void * ws, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
int CHAMELEON_zgesvd_Tile_Async(cham_job_t jobu, cham_job_t jobvt, CHAM_desc_t *A, double *S, CHAM_desc_t *T, CHAMELEON_Complex64_t *U, int LDU, CHAMELEON_Complex64_t *VT, int LDVT, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
//int CHAMELEON_zgetrf_Tile_Async(CHAM_desc_t *A, int *IPIV, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
int CHAMELEON_zgetrf_incpiv_Tile_Async(CHAM_desc_t *A, CHAM_desc_t *L, int *IPIV, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
-int CHAMELEON_zgetrf_nopiv_Tile_Async(CHAM_desc_t *A, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
+int CHAMELEON_zgetrf_nopiv_Tile_Async(CHAM_desc_t *A, void * ws, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
int CHAMELEON_zgetrf_Tile_Async( CHAM_desc_t *A, CHAM_ipiv_t *IPIV, void *ws, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request );
//int CHAMELEON_zgetri_Tile_Async(CHAM_desc_t *A, int *IPIV, CHAM_desc_t *W, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
//int CHAMELEON_zgetrs_Tile_Async(cham_trans_t trans, CHAM_desc_t *A, int *IPIV, CHAM_desc_t *B, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
@@ -333,6 +333,8 @@ 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( void *ws );
+void *CHAMELEON_zgetrf_nopiv_WS_Alloc( const CHAM_desc_t *A );
+void CHAMELEON_zgetrf_nopiv_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/runtime_struct.h b/include/chameleon/runtime_struct.h
index 72e353accaba3542c4ac36dd129fdc4552f72989..4c1b9c5210f9153221a2b6668c1868e6fbd5af42 100644
--- a/include/chameleon/runtime_struct.h
+++ b/include/chameleon/runtime_struct.h
@@ -17,7 +17,7 @@
* @author Cedric Castagnede
* @author Florent Pruvost
* @author Philippe Virouleau
- * @date 2024-03-16
+ * @date 2024-10-17
*
*/
#ifndef _chameleon_runtime_struct_h_
@@ -103,6 +103,7 @@ typedef struct runtime_option_s {
int priority; /**< Define the submitted task priority */
int workerid; /**< Define the prefered worker id to perform the tasks */
int forcesub; /**< Force task submission if true */
+ int withlacpy; /**< Flag to force the use of lacpy copies */
int withcuda; /**< Flag to know if cuda is enabled/disabled */
size_t ws_wsize; /**< Define the worker workspace size */
size_t ws_hsize; /**< Define the host workspace size for hybrid CPU/GPU kernel */
diff --git a/include/chameleon/struct_context.h b/include/chameleon/struct_context.h
index 471dcfc11acfbdf27fc58593950ae739ed4b32f9..b66621cca2b625d35a7b92f9aaf6dbd70f5f2113 100644
--- a/include/chameleon/struct_context.h
+++ b/include/chameleon/struct_context.h
@@ -14,7 +14,7 @@
* @version 1.3.0
* @author Mathieu Faverge
* @author Florent Pruvost
- * @date 2024-03-16
+ * @date 2024-10-17
*
*/
#ifndef _struct_context_h_
@@ -43,18 +43,19 @@ typedef struct chameleon_context_s {
cham_bool_t progress_enabled;
cham_bool_t generic_enabled;
cham_bool_t autominmax_enabled;
+ cham_bool_t optlacpy_enabled; /**< Enable runtime cpy instead of lacpy kernel */
cham_bool_t runtime_paused;
- cham_householder_t householder; /**> "domino" (flat) or tree-based (reduction) Householder */
- cham_translation_t translation; /**> In place or Out of place layout conversion */
+ cham_householder_t householder; /**< "domino" (flat) or tree-based (reduction) Householder */
+ cham_translation_t translation; /**< In place or Out of place layout conversion */
int nb;
int ib;
- int rhblock; /**> block size for tree-based (reduction) Householder */
- int lookahead; /**> depth of the look ahead in algorithms */
- void *schedopt; /**> structure for runtimes */
- int mpi_outer_init; /**> MPI has been initialized outside our functions */
- MPI_Comm comm; /**> MPI communicator */
+ int rhblock; /**< block size for tree-based (reduction) Householder */
+ int lookahead; /**< depth of the look ahead in algorithms */
+ void *schedopt; /**< structure for runtimes */
+ int mpi_outer_init; /**< MPI has been initialized outside our functions */
+ MPI_Comm comm; /**< MPI communicator */
} CHAM_context_t;
END_C_DECLS
diff --git a/runtime/openmp/control/runtime_options.c b/runtime/openmp/control/runtime_options.c
index 77f4a51320216dc19eb043937dd2758bce71869a..11ecfb4dee30a518cb127114483cec03e0942d5d 100644
--- a/runtime/openmp/control/runtime_options.c
+++ b/runtime/openmp/control/runtime_options.c
@@ -11,12 +11,12 @@
*
* @brief Chameleon OpenMP options routines
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Cedric Augonnet
* @author Mathieu Faverge
* @author Cedric Castagnede
* @author Philippe Virouleau
- * @date 2022-02-22
+ * @date 2024-10-17
*
*/
#include "chameleon_openmp.h"
@@ -31,6 +31,7 @@ void RUNTIME_options_init( RUNTIME_option_t *options, CHAM_context_t *chamctxt,
options->priority = RUNTIME_PRIORITY_MIN;
options->workerid = -1;
options->forcesub = 0;
+ options->withlacpy = 0;
options->ws_wsize = 0;
options->ws_hsize = 0;
options->ws_worker = NULL;
diff --git a/runtime/parsec/control/runtime_options.c b/runtime/parsec/control/runtime_options.c
index edb2017e599215a46c683013a311c9c07883e23d..a4f8c0ffffec83029f3d2d1bdbaac3da9bf53a72 100644
--- a/runtime/parsec/control/runtime_options.c
+++ b/runtime/parsec/control/runtime_options.c
@@ -11,10 +11,10 @@
*
* @brief Chameleon PaRSEC options routines
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Reazul Hoque
* @author Mathieu Faverge
- * @date 2022-02-22
+ * @date 2024-10-17
*
*/
#include "chameleon_parsec.h"
@@ -22,17 +22,18 @@
void RUNTIME_options_init( RUNTIME_option_t *options, CHAM_context_t *chamctxt,
RUNTIME_sequence_t *sequence, RUNTIME_request_t *request )
{
- options->sequence = sequence;
- options->request = request;
- options->profiling = CHAMELEON_STATISTICS == CHAMELEON_TRUE;
- options->parallel = CHAMELEON_PARALLEL == CHAMELEON_TRUE;
- options->priority = RUNTIME_PRIORITY_MIN;
- options->workerid = -1;
- options->forcesub = 0;
- options->ws_wsize = 0;
- options->ws_hsize = 0;
- options->ws_worker = NULL;
- options->ws_host = NULL;
+ options->sequence = sequence;
+ options->request = request;
+ options->profiling = CHAMELEON_STATISTICS == CHAMELEON_TRUE;
+ options->parallel = CHAMELEON_PARALLEL == CHAMELEON_TRUE;
+ options->priority = RUNTIME_PRIORITY_MIN;
+ options->workerid = -1;
+ options->forcesub = 0;
+ options->withlacpy = 0;
+ options->ws_wsize = 0;
+ options->ws_hsize = 0;
+ options->ws_worker = NULL;
+ options->ws_host = NULL;
return;
}
diff --git a/runtime/quark/control/runtime_options.c b/runtime/quark/control/runtime_options.c
index 8c090263308a8a861fdaab945c3d3759228da37e..5cb4ac31371ac8fcce34002ae14ef4f78f2d522a 100644
--- a/runtime/quark/control/runtime_options.c
+++ b/runtime/quark/control/runtime_options.c
@@ -11,12 +11,12 @@
*
* @brief Chameleon Quark options routines
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Vijay Joshi
* @author Cedric Castagnede
* @author Florent Pruvost
* @author Mathieu Faverge
- * @date 2022-02-22
+ * @date 2024-10-17
*
*/
#include "chameleon_quark.h"
@@ -34,18 +34,19 @@ void RUNTIME_options_init( RUNTIME_option_t *options, CHAM_context_t *chamctxt,
qopt->flags.thread_set_to_manual_scheduling = -1;
/* Initialize options */
- options->sequence = sequence;
- options->request = request;
- options->profiling = CHAMELEON_STATISTICS == CHAMELEON_TRUE;
- options->parallel = CHAMELEON_PARALLEL == CHAMELEON_TRUE;
- options->priority = RUNTIME_PRIORITY_MIN;
- options->workerid = -1;
- options->forcesub = 0;
+ options->sequence = sequence;
+ options->request = request;
+ options->profiling = CHAMELEON_STATISTICS == CHAMELEON_TRUE;
+ options->parallel = CHAMELEON_PARALLEL == CHAMELEON_TRUE;
+ options->priority = RUNTIME_PRIORITY_MIN;
+ options->workerid = -1;
+ options->forcesub = 0;
+ options->withlacpy = 0;
- options->ws_wsize = 0;
- options->ws_hsize = 0;
- options->ws_worker = NULL;
- options->ws_host = NULL;
+ options->ws_wsize = 0;
+ options->ws_hsize = 0;
+ options->ws_worker = NULL;
+ options->ws_host = NULL;
/* quark in options */
qopt->quark = (Quark*)(chamctxt->schedopt);
diff --git a/runtime/starpu/codelets/codelet_zlacpy.c b/runtime/starpu/codelets/codelet_zlacpy.c
index aa8d73ed3774d60b34f1b9d6641c984db59fe820..93c951a246545a970c272e50a544c62b9fd986d8 100644
--- a/runtime/starpu/codelets/codelet_zlacpy.c
+++ b/runtime/starpu/codelets/codelet_zlacpy.c
@@ -21,7 +21,7 @@
* @author Florent Pruvost
* @author Samuel Thibault
* @author Alycia Lisito
- * @date 2024-03-16
+ * @date 2024-10-17
* @precisions normal z -> c d s
*
*/
@@ -148,7 +148,8 @@ void INSERT_TASK_zlacpyx( const RUNTIME_option_t *options,
#if !defined(CHAMELEON_USE_MPI) || defined(HAVE_STARPU_MPI_DATA_CPY_PRIORITY)
/* Insert the task */
- if ( (uplo == ChamUpperLower) &&
+ if ( (!options->withlacpy) &&
+ (uplo == ChamUpperLower) &&
(tileA->m == m) && (tileA->n == n) &&
(tileB->m == m) && (tileB->n == n) &&
(displA == 0) && (displB == 0) )
@@ -225,7 +226,8 @@ void INSERT_TASK_zlacpy( const RUNTIME_option_t *options,
#if !defined(CHAMELEON_USE_MPI) || defined(HAVE_STARPU_MPI_DATA_CPY_PRIORITY)
/* Insert the task */
- if ( (uplo == ChamUpperLower) &&
+ if ( (!options->withlacpy) &&
+ (uplo == ChamUpperLower) &&
(tileA->m == m) && (tileA->n == n) &&
(tileB->m == m) && (tileB->n == n) )
{
diff --git a/runtime/starpu/control/runtime_options.c b/runtime/starpu/control/runtime_options.c
index 8ec2551f59d59e94c0eb763196576911db49d0cc..8423d9d7d5530bb1fe15bee327be2c0b2a3d7f11 100644
--- a/runtime/starpu/control/runtime_options.c
+++ b/runtime/starpu/control/runtime_options.c
@@ -16,7 +16,7 @@
* @author Mathieu Faverge
* @author Cedric Castagnede
* @author Florent Pruvost
- * @date 2023-07-04
+ * @date 2024-10-17
*
*/
#include "chameleon_starpu_internal.h"
@@ -32,6 +32,7 @@ void RUNTIME_options_init( RUNTIME_option_t *options, CHAM_context_t *chamctxt,
options->priority = RUNTIME_PRIORITY_MIN;
options->workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
options->forcesub = 0;
+ options->withlacpy = !(chamctxt->optlacpy_enabled);
options->withcuda = (chamctxt->ncudas > 0);
options->ws_wsize = 0;
options->ws_hsize = 0;
diff --git a/testing/testing_zgesv_nopiv.c b/testing/testing_zgesv_nopiv.c
index 8855334349498a9a28f1ca2cc3e70cfe1f95ea9e..d1fd9b1df9c7b745c178f42b69892b781c87efed 100644
--- a/testing/testing_zgesv_nopiv.c
+++ b/testing/testing_zgesv_nopiv.c
@@ -13,7 +13,7 @@
* @author Lucas Barros de Assis
* @author Mathieu Faverge
* @author Alycia Lisito
- * @date 2023-07-05
+ * @date 2024-10-17
* @precisions normal z -> c d s
*
*/
@@ -64,7 +64,7 @@ testing_zgesv_nopiv_desc( run_arg_list_t *args, int check )
/* Calculates the solution */
testing_start( &test_data );
if ( async ) {
- hres = CHAMELEON_zgesv_nopiv_Tile_Async( descA, descX,
+ hres = CHAMELEON_zgesv_nopiv_Tile_Async( descA, descX, NULL,
test_data.sequence, &test_data.request );
CHAMELEON_Desc_Flush( descA, test_data.sequence );
CHAMELEON_Desc_Flush( descX, test_data.sequence );
diff --git a/testing/testing_zgetrf_nopiv.c b/testing/testing_zgetrf_nopiv.c
index afa44ede3d749a4caa695a8f780ee9f30615e793..657513eaf26371399a7402e1e0b305afdc8fb854 100644
--- a/testing/testing_zgetrf_nopiv.c
+++ b/testing/testing_zgetrf_nopiv.c
@@ -13,7 +13,9 @@
* @author Lucas Barros de Assis
* @author Mathieu Faverge
* @author Alycia Lisito
- * @date 2023-07-05
+ * @author Lucas Barros De Assis
+ * @author Matthieu Kuhn
+ * @date 2024-10-17
* @precisions normal z -> c d s
*
*/
@@ -53,7 +55,7 @@ testing_zgetrf_nopiv_desc( run_arg_list_t *args, int check )
/* Calculates the solution */
testing_start( &test_data );
if ( async ) {
- hres = CHAMELEON_zgetrf_nopiv_Tile_Async( descA, test_data.sequence, &test_data.request );
+ hres = CHAMELEON_zgetrf_nopiv_Tile_Async( descA, NULL, test_data.sequence, &test_data.request );
CHAMELEON_Desc_Flush( descA, test_data.sequence );
}
else {