diff --git a/include/chameleon/runtime.h b/include/chameleon/runtime.h
index a8aaaef56a42b2dbaa25664d86022c51c2f4cd09..2655ef34669642404a79a566db7f0020323f9c65 100644
--- a/include/chameleon/runtime.h
+++ b/include/chameleon/runtime.h
@@ -18,7 +18,7 @@
* @author Samuel Thibault
* @author Philippe Swartvagher
* @author Matthieu Kuhn
- * @date 2023-08-22
+ * @date 2023-08-31
*
*/
#ifndef _chameleon_runtime_h_
@@ -710,9 +710,11 @@ void RUNTIME_ipiv_destroy( CHAM_ipiv_t *ipiv );
void RUNTIME_ipiv_init ( CHAM_ipiv_t *ipiv );
void RUNTIME_ipiv_gather ( CHAM_ipiv_t *desc, int *ipiv, int node );
-void *RUNTIME_ipiv_getaddr ( CHAM_ipiv_t *ipiv, int m );
-void *RUNTIME_nextpiv_getaddr( CHAM_ipiv_t *ipiv, int m, int h );
-void *RUNTIME_prevpiv_getaddr( CHAM_ipiv_t *ipiv, int m, int h );
+void *RUNTIME_ipiv_getaddr ( const CHAM_ipiv_t *ipiv, int m );
+void *RUNTIME_nextpiv_getaddr( const CHAM_ipiv_t *ipiv, int m, int h );
+void *RUNTIME_prevpiv_getaddr( const CHAM_ipiv_t *ipiv, int m, int h );
+void *RUNTIME_perm_getaddr ( const CHAM_ipiv_t *ipiv, int m );
+void *RUNTIME_invp_getaddr ( const CHAM_ipiv_t *ipiv, int m );
static inline void *
RUNTIME_pivot_getaddr( CHAM_ipiv_t *ipiv, int m, int h ) {
@@ -730,6 +732,8 @@ void RUNTIME_ipiv_flush ( const CHAM_ipiv_t *ipiv,
const RUNTIME_sequence_t *sequence );
void RUNTIME_ipiv_reducek( const RUNTIME_option_t *options,
CHAM_ipiv_t *ws, int k, int h );
+void RUNTIME_perm_flushk( const RUNTIME_sequence_t *sequence,
+ const CHAM_ipiv_t *ipiv, int m );
/**
* @}
diff --git a/include/chameleon/struct.h b/include/chameleon/struct.h
index d7dd07f48dc6fad1cff9359bf3410defdd657357..e995dcbe18321d0b9bfef2c375318cee8c336b1b 100644
--- a/include/chameleon/struct.h
+++ b/include/chameleon/struct.h
@@ -19,7 +19,7 @@
* @author Samuel Thibault
* @author Matthieu Kuhn
* @author Lionel Eyraud-Dubois
- * @date 2023-08-22
+ * @date 2023-08-31
*
*/
#ifndef _chameleon_struct_h_
@@ -143,13 +143,17 @@ struct chameleon_desc_s {
typedef struct chameleon_piv_s {
const CHAM_desc_t *desc; /**> Reference descriptor to compute data mapping based on diagonal tiles,
and get floating reference type */
- int *data; /**> Pointer to the data */
- void *ipiv; /**> Opaque array of pointers for the runtimes to handle the ipiv array */
- void *nextpiv; /**> Opaque array of pointers for the runtimes to handle the pivot computation structure */
- void *prevpiv; /**> Opaque array of pointers for the runtimes to handle the pivot computation structure */
+ int *data; /**> Pointer to the data */
+ void *ipiv; /**> Opaque array of pointers for the runtimes to handle the ipiv array */
+ void *nextpiv; /**> Opaque array of pointers for the runtimes to handle the pivot computation structure */
+ void *prevpiv; /**> Opaque array of pointers for the runtimes to handle the pivot computation structure */
+ void *perm; /**> Opaque array of pointers for the runtimes to handle the temporary permutation array */
+ void *invp; /**> Opaque array of pointers for the runtimes to handle the temporary inverse permutation array */
int64_t mpitag_ipiv; /**> Initial mpi tag values for the ipiv handles */
int64_t mpitag_nextpiv; /**> Initial mpi tag values for the nextpiv handles */
int64_t mpitag_prevpiv; /**> Initial mpi tag values for the prevpiv handles */
+ int64_t mpitag_perm; /**> Initial mpi tag values for the nextpiv handles */
+ int64_t mpitag_invp; /**> Initial mpi tag values for the prevpiv handles */
int i; /**> row index to the beginning of the submatrix */
int m; /**> The number of row in the vector ipiv */
int mb; /**> The number of row per block */
diff --git a/runtime/openmp/control/runtime_descriptor_ipiv.c b/runtime/openmp/control/runtime_descriptor_ipiv.c
index 03886ca650340279207c8163bc30eac81f4a1054..f10c4156d83f3e50d4b523f3942b0757475b913f 100644
--- a/runtime/openmp/control/runtime_descriptor_ipiv.c
+++ b/runtime/openmp/control/runtime_descriptor_ipiv.c
@@ -12,7 +12,7 @@
* @version 1.3.0
* @author Mathieu Faverge
* @author Matthieu Kuhn
- * @date 2023-08-22
+ * @date 2023-08-31
*
*/
#include "chameleon_openmp.h"
@@ -29,7 +29,7 @@ void RUNTIME_ipiv_destroy( CHAM_ipiv_t *ipiv )
(void)ipiv;
}
-void *RUNTIME_ipiv_getaddr( CHAM_ipiv_t *ipiv, int m )
+void *RUNTIME_ipiv_getaddr( const CHAM_ipiv_t *ipiv, int m )
{
assert( 0 );
(void)ipiv;
@@ -37,7 +37,7 @@ void *RUNTIME_ipiv_getaddr( CHAM_ipiv_t *ipiv, int m )
return NULL;
}
-void *RUNTIME_nextpiv_getaddr( CHAM_ipiv_t *ipiv, int m, int h )
+void *RUNTIME_nextpiv_getaddr( const CHAM_ipiv_t *ipiv, int m, int h )
{
assert( 0 );
(void)ipiv;
@@ -46,7 +46,7 @@ void *RUNTIME_nextpiv_getaddr( CHAM_ipiv_t *ipiv, int m, int h )
return NULL;
}
-void *RUNTIME_prevpiv_getaddr( CHAM_ipiv_t *ipiv, int m, int h )
+void *RUNTIME_prevpiv_getaddr( const CHAM_ipiv_t *ipiv, int m, int h )
{
assert( 0 );
(void)ipiv;
@@ -55,6 +55,22 @@ void *RUNTIME_prevpiv_getaddr( CHAM_ipiv_t *ipiv, int m, int h )
return NULL;
}
+void *RUNTIME_perm_getaddr( const CHAM_ipiv_t *ipiv, int k )
+{
+ assert( 0 );
+ (void)ipiv;
+ (void)k;
+ return NULL;
+}
+
+void *RUNTIME_invp_getaddr( const CHAM_ipiv_t *ipiv, int k )
+{
+ assert( 0 );
+ (void)ipiv;
+ (void)k;
+ return NULL;
+}
+
void RUNTIME_ipiv_flushk( const RUNTIME_sequence_t *sequence,
const CHAM_ipiv_t *ipiv, int m )
{
@@ -72,6 +88,15 @@ void RUNTIME_ipiv_flush( const CHAM_ipiv_t *ipiv,
(void)sequence;
}
+void RUNTIME_perm_flushk( const RUNTIME_sequence_t *sequence,
+ const CHAM_ipiv_t *ipiv, int m )
+{
+ assert( 0 );
+ (void)sequence;
+ (void)ipiv;
+ (void)m;
+}
+
void RUNTIME_ipiv_reducek( const RUNTIME_option_t *options,
CHAM_ipiv_t *ipiv, int k, int h )
{
diff --git a/runtime/parsec/control/runtime_descriptor_ipiv.c b/runtime/parsec/control/runtime_descriptor_ipiv.c
index 04a0b791139d5c6a247b25630e126d4a3eb467bf..fefb42abf9aaa65f98e2959bf09ca24779c95a7d 100644
--- a/runtime/parsec/control/runtime_descriptor_ipiv.c
+++ b/runtime/parsec/control/runtime_descriptor_ipiv.c
@@ -12,7 +12,7 @@
* @version 1.3.0
* @author Mathieu Faverge
* @author Matthieu Kuhn
- * @date 2023-08-22
+ * @date 2023-08-31
*
*/
#include "chameleon_parsec.h"
@@ -29,7 +29,7 @@ void RUNTIME_ipiv_destroy( CHAM_ipiv_t *ipiv )
(void)ipiv;
}
-void *RUNTIME_ipiv_getaddr( CHAM_ipiv_t *ipiv, int m )
+void *RUNTIME_ipiv_getaddr( const CHAM_ipiv_t *ipiv, int m )
{
assert( 0 );
(void)ipiv;
@@ -37,7 +37,7 @@ void *RUNTIME_ipiv_getaddr( CHAM_ipiv_t *ipiv, int m )
return NULL;
}
-void *RUNTIME_nextpiv_getaddr( CHAM_ipiv_t *ipiv, int m, int h )
+void *RUNTIME_nextpiv_getaddr( const CHAM_ipiv_t *ipiv, int m, int h )
{
assert( 0 );
(void)ipiv;
@@ -46,7 +46,7 @@ void *RUNTIME_nextpiv_getaddr( CHAM_ipiv_t *ipiv, int m, int h )
return NULL;
}
-void *RUNTIME_prevpiv_getaddr( CHAM_ipiv_t *ipiv, int m, int h )
+void *RUNTIME_prevpiv_getaddr( const CHAM_ipiv_t *ipiv, int m, int h )
{
assert( 0 );
(void)ipiv;
@@ -55,6 +55,22 @@ void *RUNTIME_prevpiv_getaddr( CHAM_ipiv_t *ipiv, int m, int h )
return NULL;
}
+void *RUNTIME_perm_getaddr( const CHAM_ipiv_t *ipiv, int k )
+{
+ assert( 0 );
+ (void)ipiv;
+ (void)k;
+ return NULL;
+}
+
+void *RUNTIME_invp_getaddr( const CHAM_ipiv_t *ipiv, int k )
+{
+ assert( 0 );
+ (void)ipiv;
+ (void)k;
+ return NULL;
+}
+
void RUNTIME_ipiv_flushk( const RUNTIME_sequence_t *sequence,
const CHAM_ipiv_t *ipiv, int m )
{
@@ -72,6 +88,15 @@ void RUNTIME_ipiv_flush( const CHAM_ipiv_t *ipiv,
(void)sequence;
}
+void RUNTIME_perm_flushk( const RUNTIME_sequence_t *sequence,
+ const CHAM_ipiv_t *ipiv, int m )
+{
+ assert( 0 );
+ (void)sequence;
+ (void)ipiv;
+ (void)m;
+}
+
void RUNTIME_ipiv_reducek( const RUNTIME_option_t *options,
CHAM_ipiv_t *ipiv, int k, int h )
{
diff --git a/runtime/quark/control/runtime_descriptor_ipiv.c b/runtime/quark/control/runtime_descriptor_ipiv.c
index 34706a55518f95f0e4b229a772534e3f062d05d2..88e8f886e8578f99e066868e6dfb2880fc4035d0 100644
--- a/runtime/quark/control/runtime_descriptor_ipiv.c
+++ b/runtime/quark/control/runtime_descriptor_ipiv.c
@@ -12,7 +12,7 @@
* @version 1.3.0
* @author Mathieu Faverge
* @author Matthieu Kuhn
- * @date 2023-08-22
+ * @date 2023-08-31
*
*/
#include "chameleon_quark.h"
@@ -29,7 +29,7 @@ void RUNTIME_ipiv_destroy( CHAM_ipiv_t *ipiv )
(void)ipiv;
}
-void *RUNTIME_ipiv_getaddr( CHAM_ipiv_t *ipiv, int m )
+void *RUNTIME_ipiv_getaddr( const CHAM_ipiv_t *ipiv, int m )
{
assert( 0 );
(void)ipiv;
@@ -37,7 +37,7 @@ void *RUNTIME_ipiv_getaddr( CHAM_ipiv_t *ipiv, int m )
return NULL;
}
-void *RUNTIME_nextpiv_getaddr( CHAM_ipiv_t *ipiv, int m, int h )
+void *RUNTIME_nextpiv_getaddr( const CHAM_ipiv_t *ipiv, int m, int h )
{
assert( 0 );
(void)ipiv;
@@ -46,7 +46,7 @@ void *RUNTIME_nextpiv_getaddr( CHAM_ipiv_t *ipiv, int m, int h )
return NULL;
}
-void *RUNTIME_prevpiv_getaddr( CHAM_ipiv_t *ipiv, int m, int h )
+void *RUNTIME_prevpiv_getaddr( const CHAM_ipiv_t *ipiv, int m, int h )
{
assert( 0 );
(void)ipiv;
@@ -55,6 +55,22 @@ void *RUNTIME_prevpiv_getaddr( CHAM_ipiv_t *ipiv, int m, int h )
return NULL;
}
+void *RUNTIME_perm_getaddr( const CHAM_ipiv_t *ipiv, int k )
+{
+ assert( 0 );
+ (void)ipiv;
+ (void)k;
+ return NULL;
+}
+
+void *RUNTIME_invp_getaddr( const CHAM_ipiv_t *ipiv, int k )
+{
+ assert( 0 );
+ (void)ipiv;
+ (void)k;
+ return NULL;
+}
+
void RUNTIME_ipiv_flushk( const RUNTIME_sequence_t *sequence,
const CHAM_ipiv_t *ipiv, int m )
{
@@ -72,6 +88,15 @@ void RUNTIME_ipiv_flush( const CHAM_ipiv_t *ipiv,
(void)sequence;
}
+void RUNTIME_perm_flushk( const RUNTIME_sequence_t *sequence,
+ const CHAM_ipiv_t *ipiv, int m )
+{
+ assert( 0 );
+ (void)sequence;
+ (void)ipiv;
+ (void)m;
+}
+
void RUNTIME_ipiv_reducek( const RUNTIME_option_t *options,
CHAM_ipiv_t *ipiv, int k, int h )
{
diff --git a/runtime/starpu/control/runtime_descriptor_ipiv.c b/runtime/starpu/control/runtime_descriptor_ipiv.c
index ffe3d1e47a24dbfa3e9986a477df07a95c2cdf32..69f09cb0ae291365ee9aafae8e579662c9ffd387 100644
--- a/runtime/starpu/control/runtime_descriptor_ipiv.c
+++ b/runtime/starpu/control/runtime_descriptor_ipiv.c
@@ -12,7 +12,7 @@
* @version 1.3.0
* @author Mathieu Faverge
* @author Matthieu Kuhn
- * @date 2023-08-22
+ * @date 2023-08-31
*
*/
#include "chameleon_starpu.h"
@@ -23,10 +23,16 @@
void RUNTIME_ipiv_create( CHAM_ipiv_t *ipiv )
{
assert( ipiv );
-
- ipiv->ipiv = (void*)calloc( ipiv->mt, sizeof(starpu_data_handle_t) );
- ipiv->nextpiv = (void*)calloc( ipiv->mt, sizeof(starpu_data_handle_t) );
- ipiv->prevpiv = (void*)calloc( ipiv->mt, sizeof(starpu_data_handle_t) );
+ starpu_data_handle_t *handles = calloc( 5 * ipiv->mt, sizeof(starpu_data_handle_t) );
+ ipiv->ipiv = handles;
+ handles += ipiv->mt;
+ ipiv->nextpiv = handles;
+ handles += ipiv->mt;
+ ipiv->prevpiv = handles;
+ handles += ipiv->mt;
+ ipiv->perm = handles;
+ handles += ipiv->mt;
+ ipiv->invp = handles;
#if defined(CHAMELEON_USE_MPI)
/*
* Book the number of tags required to describe pivot structure
@@ -34,13 +40,15 @@ void RUNTIME_ipiv_create( CHAM_ipiv_t *ipiv )
*/
{
chameleon_starpu_tag_init();
- ipiv->mpitag_ipiv = chameleon_starpu_tag_book( (int64_t)(ipiv->mt) * 3 );
+ ipiv->mpitag_ipiv = chameleon_starpu_tag_book( (int64_t)(ipiv->mt) * 5 );
if ( ipiv->mpitag_ipiv == -1 ) {
chameleon_fatal_error("RUNTIME_ipiv_create", "Can't pursue computation since no more tags are available for ipiv structure");
return;
}
ipiv->mpitag_nextpiv = ipiv->mpitag_ipiv + ipiv->mt;
ipiv->mpitag_prevpiv = ipiv->mpitag_nextpiv + ipiv->mt;
+ ipiv->mpitag_perm = ipiv->mpitag_prevpiv + ipiv->mt;
+ ipiv->mpitag_invp = ipiv->mpitag_perm + ipiv->mt;
}
#endif
}
@@ -51,37 +59,26 @@ void RUNTIME_ipiv_create( CHAM_ipiv_t *ipiv )
void RUNTIME_ipiv_destroy( CHAM_ipiv_t *ipiv )
{
int i;
- starpu_data_handle_t *ipiv_handle = (starpu_data_handle_t*)(ipiv->ipiv);
- starpu_data_handle_t *nextpiv_handle = (starpu_data_handle_t*)(ipiv->nextpiv);
- starpu_data_handle_t *prevpiv_handle = (starpu_data_handle_t*)(ipiv->prevpiv);
-
- for(i=0; i<ipiv->mt; i++) {
- if ( *ipiv_handle != NULL ) {
- starpu_data_unregister( *ipiv_handle );
- *ipiv_handle = NULL;
- }
- ipiv_handle++;
-
- if ( *nextpiv_handle != NULL ) {
- starpu_data_unregister( *nextpiv_handle );
- *nextpiv_handle = NULL;
- }
- nextpiv_handle++;
+ starpu_data_handle_t *handle = (starpu_data_handle_t*)(ipiv->ipiv);
- if ( *prevpiv_handle != NULL ) {
- starpu_data_unregister( *prevpiv_handle );
- *prevpiv_handle = NULL;
+ for(i=0; i<(5 * ipiv->mt); i++) {
+ if ( *handle != NULL ) {
+ starpu_data_unregister( *handle );
+ *handle = NULL;
}
- prevpiv_handle++;
+ handle++;
}
free( ipiv->ipiv );
- free( ipiv->nextpiv );
- free( ipiv->prevpiv );
+ ipiv->ipiv = NULL;
+ ipiv->nextpiv = NULL;
+ ipiv->prevpiv = NULL;
+ ipiv->perm = NULL;
+ ipiv->invp = NULL;
chameleon_starpu_tag_release( ipiv->mpitag_ipiv );
}
-void *RUNTIME_ipiv_getaddr( CHAM_ipiv_t *ipiv, int m )
+void *RUNTIME_ipiv_getaddr( const CHAM_ipiv_t *ipiv, int m )
{
starpu_data_handle_t *handle = (starpu_data_handle_t*)(ipiv->ipiv);
int64_t mm = m + (ipiv->i / ipiv->mb);
@@ -110,7 +107,7 @@ void *RUNTIME_ipiv_getaddr( CHAM_ipiv_t *ipiv, int m )
return *handle;
}
-void *RUNTIME_nextpiv_getaddr( CHAM_ipiv_t *ipiv, int m, int h )
+void *RUNTIME_nextpiv_getaddr( const CHAM_ipiv_t *ipiv, int m, int h )
{
starpu_data_handle_t *nextpiv = (starpu_data_handle_t*)(ipiv->nextpiv);
int64_t mm = m + (ipiv->i / ipiv->mb);
@@ -133,7 +130,7 @@ void *RUNTIME_nextpiv_getaddr( CHAM_ipiv_t *ipiv, int m, int h )
return *nextpiv;
}
-void *RUNTIME_prevpiv_getaddr( CHAM_ipiv_t *ipiv, int m, int h )
+void *RUNTIME_prevpiv_getaddr( const CHAM_ipiv_t *ipiv, int m, int h )
{
starpu_data_handle_t *prevpiv = (starpu_data_handle_t*)(ipiv->prevpiv);
int64_t mm = m + (ipiv->i / ipiv->mb);
@@ -156,6 +153,64 @@ void *RUNTIME_prevpiv_getaddr( CHAM_ipiv_t *ipiv, int m, int h )
return *prevpiv;
}
+void *RUNTIME_perm_getaddr( const CHAM_ipiv_t *ipiv, int m )
+{
+ starpu_data_handle_t *handle = (starpu_data_handle_t*)(ipiv->perm);
+ int64_t mm = m + (ipiv->i / ipiv->mb);
+
+ handle += mm;
+ assert( handle );
+
+ if ( *handle != NULL ) {
+ return *handle;
+ }
+
+ const CHAM_desc_t *A = ipiv->desc;
+ int owner = A->get_rankof( A, m, m );
+ int ncols = ipiv->mb;
+
+ starpu_vector_data_register( handle, -1, (uintptr_t)NULL, ncols, sizeof(int) );
+
+#if defined(CHAMELEON_USE_MPI)
+ {
+ int64_t tag = ipiv->mpitag_perm + mm;
+ starpu_mpi_data_register( *handle, tag, owner );
+ }
+#endif /* defined(CHAMELEON_USE_MPI) */
+
+ assert( *handle );
+ return *handle;
+}
+
+void *RUNTIME_invp_getaddr( const CHAM_ipiv_t *ipiv, int m )
+{
+ starpu_data_handle_t *handle = (starpu_data_handle_t*)(ipiv->invp);
+ int64_t mm = m + (ipiv->i / ipiv->mb);
+
+ handle += mm;
+ assert( handle );
+
+ if ( *handle != NULL ) {
+ return *handle;
+ }
+
+ const CHAM_desc_t *A = ipiv->desc;
+ int owner = A->get_rankof( A, m, m );
+ int ncols = ipiv->mb;
+
+ starpu_vector_data_register( handle, -1, (uintptr_t)NULL, ncols, sizeof(int) );
+
+#if defined(CHAMELEON_USE_MPI)
+ {
+ int64_t tag = ipiv->mpitag_invp + mm;
+ starpu_mpi_data_register( *handle, tag, owner );
+ }
+#endif /* defined(CHAMELEON_USE_MPI) */
+
+ assert( *handle );
+ return *handle;
+}
+
void RUNTIME_ipiv_flushk( const RUNTIME_sequence_t *sequence,
const CHAM_ipiv_t *ipiv, int m )
{
@@ -205,6 +260,44 @@ void RUNTIME_ipiv_flush( const CHAM_ipiv_t *ipiv,
}
}
+void RUNTIME_perm_flushk( const RUNTIME_sequence_t *sequence,
+ const CHAM_ipiv_t *ipiv, int m )
+{
+ starpu_data_handle_t *handle;
+ const CHAM_desc_t *A = ipiv->desc;
+ int64_t mm = m + ( ipiv->i / ipiv->mb );
+
+ handle = (starpu_data_handle_t*)(ipiv->perm);
+ handle += mm;
+
+ if ( *handle != NULL ) {
+#if defined(CHAMELEON_USE_MPI)
+ starpu_mpi_cache_flush( MPI_COMM_WORLD, *handle );
+ if ( starpu_mpi_data_get_rank( *handle ) == A->myrank )
+#endif
+ {
+ chameleon_starpu_data_wont_use( *handle );
+ }
+ }
+
+ handle = (starpu_data_handle_t*)(ipiv->invp);
+ handle += mm;
+
+ if ( *handle != NULL ) {
+#if defined(CHAMELEON_USE_MPI)
+ starpu_mpi_cache_flush( MPI_COMM_WORLD, *handle );
+ if ( starpu_mpi_data_get_rank( *handle ) == A->myrank )
+#endif
+ {
+ chameleon_starpu_data_wont_use( *handle );
+ }
+ }
+
+ (void)sequence;
+ (void)ipiv;
+ (void)m;
+}
+
void RUNTIME_ipiv_reducek( const RUNTIME_option_t *options,
CHAM_ipiv_t *ipiv, int k, int h )
{