From 3e6a06b1133b24262ab403a70fb31361789e5851 Mon Sep 17 00:00:00 2001
From: Matthieu Kuhn <matthieu.kuhn@atos.net>
Date: Thu, 20 Apr 2023 18:06:04 +0200
Subject: [PATCH] runtime: Add functions to manipulate CHAM_ipiv_t data
structure at the runtime level
---
include/chameleon/runtime.h | 30 +-
.../openmp/control/runtime_descriptor_ipiv.c | 97 ++++++
.../parsec/control/runtime_descriptor_ipiv.c | 97 ++++++
.../quark/control/runtime_descriptor_ipiv.c | 97 ++++++
.../starpu/control/runtime_descriptor_ipiv.c | 306 ++++++++++++++++++
5 files changed, 625 insertions(+), 2 deletions(-)
create mode 100644 runtime/openmp/control/runtime_descriptor_ipiv.c
create mode 100644 runtime/parsec/control/runtime_descriptor_ipiv.c
create mode 100644 runtime/quark/control/runtime_descriptor_ipiv.c
create mode 100644 runtime/starpu/control/runtime_descriptor_ipiv.c
diff --git a/include/chameleon/runtime.h b/include/chameleon/runtime.h
index 82818ba75..a8aaaef56 100644
--- a/include/chameleon/runtime.h
+++ b/include/chameleon/runtime.h
@@ -10,7 +10,7 @@
***
*
* @brief The common runtimes API
- * @version 1.2.0
+ * @version 1.3.0
* @author Mathieu Faverge
* @author Cedric Augonnet
* @author Cedric Castagnede
@@ -18,7 +18,7 @@
* @author Samuel Thibault
* @author Philippe Swartvagher
* @author Matthieu Kuhn
- * @date 2022-02-22
+ * @date 2023-08-22
*
*/
#ifndef _chameleon_runtime_h_
@@ -705,6 +705,32 @@ void RUNTIME_ddisplay_oneprofile (cham_tasktype_t task);
void RUNTIME_sdisplay_allprofile ();
void RUNTIME_sdisplay_oneprofile (cham_tasktype_t task);
+void RUNTIME_ipiv_create ( CHAM_ipiv_t *ipiv );
+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 );
+
+static inline void *
+RUNTIME_pivot_getaddr( CHAM_ipiv_t *ipiv, int m, int h ) {
+ if ( h%2 == 0 ) {
+ return RUNTIME_nextpiv_getaddr( ipiv, m, -1 );
+ }
+ else {
+ return RUNTIME_prevpiv_getaddr( ipiv, m, -1 );
+ }
+}
+
+void RUNTIME_ipiv_flushk ( const RUNTIME_sequence_t *sequence,
+ const CHAM_ipiv_t *ipiv, int m );
+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 );
+
/**
* @}
*/
diff --git a/runtime/openmp/control/runtime_descriptor_ipiv.c b/runtime/openmp/control/runtime_descriptor_ipiv.c
new file mode 100644
index 000000000..03886ca65
--- /dev/null
+++ b/runtime/openmp/control/runtime_descriptor_ipiv.c
@@ -0,0 +1,97 @@
+/**
+ *
+ * @file openmp/runtime_descriptor_ipiv.c
+ *
+ * @copyright 2022-2023 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria,
+ * Univ. Bordeaux. All rights reserved.
+ *
+ ***
+ *
+ * @brief Chameleon OpenMP descriptor routines
+ *
+ * @version 1.3.0
+ * @author Mathieu Faverge
+ * @author Matthieu Kuhn
+ * @date 2023-08-22
+ *
+ */
+#include "chameleon_openmp.h"
+
+void RUNTIME_ipiv_create( CHAM_ipiv_t *ipiv )
+{
+ assert( 0 );
+ (void)ipiv;
+}
+
+void RUNTIME_ipiv_destroy( CHAM_ipiv_t *ipiv )
+{
+ assert( 0 );
+ (void)ipiv;
+}
+
+void *RUNTIME_ipiv_getaddr( CHAM_ipiv_t *ipiv, int m )
+{
+ assert( 0 );
+ (void)ipiv;
+ (void)m;
+ return NULL;
+}
+
+void *RUNTIME_nextpiv_getaddr( CHAM_ipiv_t *ipiv, int m, int h )
+{
+ assert( 0 );
+ (void)ipiv;
+ (void)m;
+ (void)h;
+ return NULL;
+}
+
+void *RUNTIME_prevpiv_getaddr( CHAM_ipiv_t *ipiv, int m, int h )
+{
+ assert( 0 );
+ (void)ipiv;
+ (void)m;
+ (void)h;
+ return NULL;
+}
+
+void RUNTIME_ipiv_flushk( const RUNTIME_sequence_t *sequence,
+ const CHAM_ipiv_t *ipiv, int m )
+{
+ assert( 0 );
+ (void)sequence;
+ (void)ipiv;
+ (void)m;
+}
+
+void RUNTIME_ipiv_flush( const CHAM_ipiv_t *ipiv,
+ const RUNTIME_sequence_t *sequence )
+{
+ assert( 0 );
+ (void)ipiv;
+ (void)sequence;
+}
+
+void RUNTIME_ipiv_reducek( const RUNTIME_option_t *options,
+ CHAM_ipiv_t *ipiv, int k, int h )
+{
+ assert( 0 );
+ (void)options;
+ (void)ipiv;
+ (void)k;
+ (void)h;
+}
+
+void RUNTIME_ipiv_init( CHAM_ipiv_t *ipiv )
+{
+ assert( 0 );
+ (void)ipiv;
+}
+
+void RUNTIME_ipiv_gather( CHAM_ipiv_t *desc, int *ipiv, int node )
+{
+ assert( 0 );
+ (void)desc;
+ (void)ipiv;
+ (void)node;
+}
diff --git a/runtime/parsec/control/runtime_descriptor_ipiv.c b/runtime/parsec/control/runtime_descriptor_ipiv.c
new file mode 100644
index 000000000..04a0b7911
--- /dev/null
+++ b/runtime/parsec/control/runtime_descriptor_ipiv.c
@@ -0,0 +1,97 @@
+/**
+ *
+ * @file parsec/runtime_descriptor_ipiv.c
+ *
+ * @copyright 2022-2023 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria,
+ * Univ. Bordeaux. All rights reserved.
+ *
+ ***
+ *
+ * @brief Chameleon PaRSEC descriptor routines
+ *
+ * @version 1.3.0
+ * @author Mathieu Faverge
+ * @author Matthieu Kuhn
+ * @date 2023-08-22
+ *
+ */
+#include "chameleon_parsec.h"
+
+void RUNTIME_ipiv_create( CHAM_ipiv_t *ipiv )
+{
+ assert( 0 );
+ (void)ipiv;
+}
+
+void RUNTIME_ipiv_destroy( CHAM_ipiv_t *ipiv )
+{
+ assert( 0 );
+ (void)ipiv;
+}
+
+void *RUNTIME_ipiv_getaddr( CHAM_ipiv_t *ipiv, int m )
+{
+ assert( 0 );
+ (void)ipiv;
+ (void)m;
+ return NULL;
+}
+
+void *RUNTIME_nextpiv_getaddr( CHAM_ipiv_t *ipiv, int m, int h )
+{
+ assert( 0 );
+ (void)ipiv;
+ (void)m;
+ (void)h;
+ return NULL;
+}
+
+void *RUNTIME_prevpiv_getaddr( CHAM_ipiv_t *ipiv, int m, int h )
+{
+ assert( 0 );
+ (void)ipiv;
+ (void)m;
+ (void)h;
+ return NULL;
+}
+
+void RUNTIME_ipiv_flushk( const RUNTIME_sequence_t *sequence,
+ const CHAM_ipiv_t *ipiv, int m )
+{
+ assert( 0 );
+ (void)sequence;
+ (void)ipiv;
+ (void)m;
+}
+
+void RUNTIME_ipiv_flush( const CHAM_ipiv_t *ipiv,
+ const RUNTIME_sequence_t *sequence )
+{
+ assert( 0 );
+ (void)ipiv;
+ (void)sequence;
+}
+
+void RUNTIME_ipiv_reducek( const RUNTIME_option_t *options,
+ CHAM_ipiv_t *ipiv, int k, int h )
+{
+ assert( 0 );
+ (void)options;
+ (void)ipiv;
+ (void)k;
+ (void)h;
+}
+
+void RUNTIME_ipiv_init( CHAM_ipiv_t *ipiv )
+{
+ assert( 0 );
+ (void)ipiv;
+}
+
+void RUNTIME_ipiv_gather( CHAM_ipiv_t *desc, int *ipiv, int node )
+{
+ assert( 0 );
+ (void)desc;
+ (void)ipiv;
+ (void)node;
+}
diff --git a/runtime/quark/control/runtime_descriptor_ipiv.c b/runtime/quark/control/runtime_descriptor_ipiv.c
new file mode 100644
index 000000000..34706a555
--- /dev/null
+++ b/runtime/quark/control/runtime_descriptor_ipiv.c
@@ -0,0 +1,97 @@
+/**
+ *
+ * @file quark/runtime_descriptor_ipiv.c
+ *
+ * @copyright 2022-2023 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria,
+ * Univ. Bordeaux. All rights reserved.
+ *
+ ***
+ *
+ * @brief Chameleon Quark descriptor routines
+ *
+ * @version 1.3.0
+ * @author Mathieu Faverge
+ * @author Matthieu Kuhn
+ * @date 2023-08-22
+ *
+ */
+#include "chameleon_quark.h"
+
+void RUNTIME_ipiv_create( CHAM_ipiv_t *ipiv )
+{
+ assert( 0 );
+ (void)ipiv;
+}
+
+void RUNTIME_ipiv_destroy( CHAM_ipiv_t *ipiv )
+{
+ assert( 0 );
+ (void)ipiv;
+}
+
+void *RUNTIME_ipiv_getaddr( CHAM_ipiv_t *ipiv, int m )
+{
+ assert( 0 );
+ (void)ipiv;
+ (void)m;
+ return NULL;
+}
+
+void *RUNTIME_nextpiv_getaddr( CHAM_ipiv_t *ipiv, int m, int h )
+{
+ assert( 0 );
+ (void)ipiv;
+ (void)m;
+ (void)h;
+ return NULL;
+}
+
+void *RUNTIME_prevpiv_getaddr( CHAM_ipiv_t *ipiv, int m, int h )
+{
+ assert( 0 );
+ (void)ipiv;
+ (void)m;
+ (void)h;
+ return NULL;
+}
+
+void RUNTIME_ipiv_flushk( const RUNTIME_sequence_t *sequence,
+ const CHAM_ipiv_t *ipiv, int m )
+{
+ assert( 0 );
+ (void)sequence;
+ (void)ipiv;
+ (void)m;
+}
+
+void RUNTIME_ipiv_flush( const CHAM_ipiv_t *ipiv,
+ const RUNTIME_sequence_t *sequence )
+{
+ assert( 0 );
+ (void)ipiv;
+ (void)sequence;
+}
+
+void RUNTIME_ipiv_reducek( const RUNTIME_option_t *options,
+ CHAM_ipiv_t *ipiv, int k, int h )
+{
+ assert( 0 );
+ (void)options;
+ (void)ipiv;
+ (void)k;
+ (void)h;
+}
+
+void RUNTIME_ipiv_init( CHAM_ipiv_t *ipiv )
+{
+ assert( 0 );
+ (void)ipiv;
+}
+
+void RUNTIME_ipiv_gather( CHAM_ipiv_t *desc, int *ipiv, int node )
+{
+ assert( 0 );
+ (void)desc;
+ (void)ipiv;
+ (void)node;
+}
diff --git a/runtime/starpu/control/runtime_descriptor_ipiv.c b/runtime/starpu/control/runtime_descriptor_ipiv.c
new file mode 100644
index 000000000..4131f7d6c
--- /dev/null
+++ b/runtime/starpu/control/runtime_descriptor_ipiv.c
@@ -0,0 +1,306 @@
+/**
+ *
+ * @file starpu/runtime_descriptor_ipiv.c
+ *
+ * @copyright 2022-2023 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria,
+ * Univ. Bordeaux. All rights reserved.
+ *
+ ***
+ *
+ * @brief Chameleon StarPU descriptor routines
+ *
+ * @version 1.3.0
+ * @author Mathieu Faverge
+ * @author Matthieu Kuhn
+ * @date 2023-08-22
+ *
+ */
+#include "chameleon_starpu.h"
+
+/**
+ * Create ws_pivot runtime structures
+ */
+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) );
+#if defined(CHAMELEON_USE_MPI)
+ /*
+ * Book the number of tags required to describe pivot structure
+ * One per handle type
+ */
+ {
+ chameleon_starpu_tag_init();
+ ipiv->mpitag_ipiv = chameleon_starpu_tag_book( (int64_t)(ipiv->mt) * 3 );
+ 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;
+ }
+#endif
+}
+
+/**
+ * Destroy ws_pivot runtime structures
+ */
+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++;
+
+ if ( *prevpiv_handle != NULL ) {
+ starpu_data_unregister( *prevpiv_handle );
+ *prevpiv_handle = NULL;
+ }
+ prevpiv_handle++;
+ }
+
+ free( ipiv->ipiv );
+ free( ipiv->nextpiv );
+ free( ipiv->prevpiv );
+ chameleon_starpu_tag_release( ipiv->mpitag_ipiv );
+}
+
+void *RUNTIME_ipiv_getaddr( 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);
+
+ 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 = (mm == (ipiv->mt-1)) ? ipiv->m - mm * ipiv->mb : ipiv->mb;
+
+ starpu_vector_data_register( handle, -1, (uintptr_t)NULL, ncols, sizeof(int) );
+
+#if defined(CHAMELEON_USE_MPI)
+ {
+ int64_t tag = ipiv->mpitag_ipiv + mm;
+ starpu_mpi_data_register( *handle, tag, owner );
+ }
+#endif /* defined(CHAMELEON_USE_MPI) */
+
+ assert( *handle );
+ return *handle;
+}
+
+void *RUNTIME_nextpiv_getaddr( 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);
+
+ nextpiv += mm;
+ assert( nextpiv );
+
+ if ( *nextpiv != NULL ) {
+ return *nextpiv;
+ }
+
+ const CHAM_desc_t *A = ipiv->desc;
+ int owner = A->get_rankof( A, m, m );
+ int ncols = (mm == (ipiv->mt-1)) ? ipiv->m - mm * ipiv->mb : ipiv->mb;
+ int64_t tag = ipiv->mpitag_nextpiv + mm;
+
+ cppi_register( nextpiv, A->dtyp, ncols, tag, owner );
+
+ assert( *nextpiv );
+ return *nextpiv;
+}
+
+void *RUNTIME_prevpiv_getaddr( 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);
+
+ prevpiv += mm;
+ assert( prevpiv );
+
+ if ( *prevpiv != NULL ) {
+ return *prevpiv;
+ }
+
+ const CHAM_desc_t *A = ipiv->desc;
+ int owner = A->get_rankof( A, m, m );
+ int ncols = (mm == (ipiv->mt-1)) ? ipiv->m - mm * ipiv->mb : ipiv->mb;
+ int64_t tag = ipiv->mpitag_prevpiv + mm;
+
+ cppi_register( prevpiv, A->dtyp, ncols, tag, owner );
+
+ assert( *prevpiv );
+ return *prevpiv;
+}
+
+void RUNTIME_ipiv_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->nextpiv);
+ 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->prevpiv);
+ 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_flush( const CHAM_ipiv_t *ipiv,
+ const RUNTIME_sequence_t *sequence )
+{
+ int m;
+
+ for (m = 0; m < ipiv->mt; m++)
+ {
+ RUNTIME_ipiv_flushk( sequence, ipiv, m );
+ }
+}
+
+void RUNTIME_ipiv_reducek( const RUNTIME_option_t *options,
+ CHAM_ipiv_t *ipiv, int k, int h )
+{
+ starpu_data_handle_t nextpiv = RUNTIME_pivot_getaddr( ipiv, k, h );
+ starpu_data_handle_t prevpiv = RUNTIME_pivot_getaddr( ipiv, k, h-1 );
+
+ if ( h < ipiv->n ) {
+#if defined(HAVE_STARPU_MPI_REDUX) && defined(CHAMELEON_USE_MPI)
+#if !defined(HAVE_STARPU_MPI_REDUX_WRAPUP)
+ starpu_mpi_redux_data_prio_tree( MPI_COMM_WORLD, nextpiv,
+ options->priority, 2 /* Binary tree */ );
+#endif
+#endif
+ }
+
+ /* Invalidate the previous pivot structure for correct initialization in later reuse */
+ if ( h > 0 ) {
+ starpu_data_invalidate_submit( prevpiv );
+ }
+
+ (void)options;
+}
+
+static void cl_ipiv_init_cpu_func(void *descr[], void *cl_arg)
+{
+ int *ipiv = (int *)STARPU_VECTOR_GET_PTR(descr[0]);
+
+#if !defined(CHAMELEON_SIMULATION)
+ {
+ int i, m0, n;
+ starpu_codelet_unpack_args( cl_arg, &m0, &n );
+
+ for( i=0; i<n; i++ ) {
+ ipiv[i] = m0 + i + 1;
+ }
+ }
+#endif
+}
+
+struct starpu_codelet cl_ipiv_init = {
+ .where = STARPU_CPU,
+ .cpu_func = cl_ipiv_init_cpu_func,
+ .nbuffers = 1,
+};
+
+void RUNTIME_ipiv_init( CHAM_ipiv_t *ipiv )
+{
+ int64_t mt = ipiv->mt;
+ int64_t mb = ipiv->mb;
+ int m;
+
+ for (m = 0; m < mt; m++) {
+ starpu_data_handle_t ipiv_src = RUNTIME_ipiv_getaddr( ipiv, m );
+ int m0 = m * mb;
+ int n = (m == (mt-1)) ? ipiv->m - m0 : mb;
+
+ rt_starpu_insert_task(
+ &cl_ipiv_init,
+ STARPU_VALUE, &m0, sizeof(int),
+ STARPU_VALUE, &n, sizeof(int),
+ STARPU_W, ipiv_src,
+ 0);
+ }
+}
+
+void RUNTIME_ipiv_gather( CHAM_ipiv_t *desc, int *ipiv, int node )
+{
+ int64_t mt = desc->mt;
+ int64_t mb = desc->mb;
+ int64_t tag = chameleon_starpu_tag_book( (int64_t)(desc->mt) );
+ int rank = CHAMELEON_Comm_rank();
+ int m;
+
+ for (m = 0; m < mt; m++, ipiv += mb) {
+ starpu_data_handle_t ipiv_src = RUNTIME_ipiv_getaddr( desc, m );
+
+#if defined(CHAMELEON_USE_MPI)
+ if ( (rank == node) ||
+ (rank == starpu_mpi_data_get_rank(ipiv_src)) )
+#endif
+ {
+ starpu_data_handle_t ipiv_dst;
+ int ncols = (m == (mt-1)) ? desc->m - m * mb : mb;
+ uintptr_t ipivptr = (rank == node) ? (uintptr_t)ipiv : 0;
+ int home_node = (rank == node) ? STARPU_MAIN_RAM : -1;
+
+ starpu_vector_data_register( &ipiv_dst, home_node, ipivptr, ncols, sizeof(int) );
+
+#if defined(CHAMELEON_USE_MPI)
+ starpu_mpi_data_register( ipiv_dst, tag + m, 0 );
+#endif /* defined(CHAMELEON_USE_MPI) */
+
+ assert( ipiv_dst );
+
+ starpu_data_cpy( ipiv_dst, ipiv_src, 0, NULL, NULL );
+ starpu_data_unregister( ipiv_dst );
+ }
+ }
+}
--
GitLab