diff --git a/control/control.c b/control/control.c
index c3cacb10a68b4328b6b2bc9ce24b0b7915a37f23..fbc777a11f487ae9a56a4dac4fc1ee12c708ea95 100644
--- a/control/control.c
+++ b/control/control.c
@@ -20,7 +20,7 @@
* @author Samuel Thibault
* @author Philippe Swartvagher
* @author Loris Lucido
- * @date 2023-07-04
+ * @date 2024-03-16
*
***
*
@@ -35,7 +35,8 @@
*
* @ingroup Control
*
- * @brief Initialize CHAMELEON.
+ * @brief Initialize CHAMELEON with number of cpus and gpus (using
+ * MPI_COMM_WORLD).
*
******************************************************************************
*
@@ -59,7 +60,8 @@ int __chameleon_init(int cores, int gpus)
*
* @ingroup Control
*
- * @brief Initialize CHAMELEON.
+ * @brief Initialize CHAMELEON with number of cpus and gpus and threads per
+ * worker (using MPI_COMM_WORLD).
*
******************************************************************************
*
@@ -78,6 +80,37 @@ int __chameleon_init(int cores, int gpus)
*
*/
int __chameleon_initpar(int ncpus, int ngpus, int nthreads_per_worker)
+{
+ return __chameleon_initparcomm( ncpus, ngpus, nthreads_per_worker, MPI_COMM_WORLD );
+}
+
+/**
+ *
+ * @ingroup Control
+ *
+ * @brief Initialize CHAMELEON with number of cpus and gpus and threads per
+ * worker and using a given MPI communicator.
+ *
+ ******************************************************************************
+ *
+ * @param[in] ncpus
+ * Number of cores to use.
+ *
+ * @param[in] ngpus
+ * Number of cuda devices to use.
+ *
+ * @param[in] nthreads_per_worker
+ * Number of threads per worker (cpu, cuda device).
+ *
+ * @param[in] comm
+ * The MPI communicator.
+ *
+ ******************************************************************************
+ *
+ * @retval CHAMELEON_SUCCESS successful exit
+ *
+ */
+int __chameleon_initparcomm(int ncpus, int ngpus, int nthreads_per_worker, MPI_Comm comm)
{
CHAM_context_t *chamctxt;
@@ -124,6 +157,7 @@ int __chameleon_initpar(int ncpus, int ngpus, int nthreads_per_worker)
#endif
chamctxt->ncudas = ngpus;
+ chamctxt->comm = comm;
return RUNTIME_init( chamctxt, ncpus, ngpus, nthreads_per_worker );
}
@@ -145,15 +179,23 @@ int __chameleon_finalize(void)
chameleon_error("CHAMELEON_Finalize", "CHAMELEON not initialized");
return CHAMELEON_ERR_NOT_INITIALIZED;
}
- RUNTIME_flush();
+
+ /* Make sure all data are flushed */
+ RUNTIME_flush( chamctxt );
+
+ /* Wait for anything running */
# if !defined(CHAMELEON_SIMULATION)
RUNTIME_barrier(chamctxt);
# endif
+
+ /* Stop the runtime system */
RUNTIME_finalize( chamctxt );
#if defined(CHAMELEON_USE_MPI)
- if (!chamctxt->mpi_outer_init)
+ /* Finalize MPI if initialized by Chameleon */
+ if ( !chamctxt->mpi_outer_init ) {
MPI_Finalize();
+ }
#endif
chameleon_context_destroy();
diff --git a/coreblas/compute/CMakeLists.txt b/coreblas/compute/CMakeLists.txt
index 137adfbba9179c7f5dc49ca0d0ad8f4bc1ac1dbe..c314957fa12d6e2e09c7ac48cec96637c3904c78 100644
--- a/coreblas/compute/CMakeLists.txt
+++ b/coreblas/compute/CMakeLists.txt
@@ -24,7 +24,7 @@
# @author Florent Pruvost
# @author Guillaume Sylvand
# @author Matthieu Kuhn
-# @date 2023-08-31
+# @date 2024-03-16
#
###
@@ -164,6 +164,9 @@ endif()
target_link_libraries(coreblas PRIVATE MORSE::LAPACKE)
target_link_libraries(coreblas PRIVATE MORSE::CBLAS)
target_link_libraries(coreblas PUBLIC MORSE::M)
+if (CHAMELEON_USE_MPI)
+ target_link_libraries(coreblas PUBLIC MPI::MPI_C)
+endif()
# export target coreblas
install(EXPORT coreblasTargets
diff --git a/example/lapack_to_chameleon/step6.h b/example/lapack_to_chameleon/step6.h
index b834863c1f91b6606f3fe985030df8f8c9ba06cb..d60683987471c510aafa20ce55a6a87f0e358bcb 100644
--- a/example/lapack_to_chameleon/step6.h
+++ b/example/lapack_to_chameleon/step6.h
@@ -11,10 +11,10 @@
*
* @brief Chameleon step6 example header
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Florent Pruvost
* @author Mathieu Faverge
- * @date 2022-02-22
+ * @date 2024-03-16
*
*/
#ifndef _step6_h_
@@ -34,7 +34,7 @@
enum iparam_step6 {
IPARAM_THRDNBR, /* Number of cores */
IPARAM_NCUDAS, /* Number of cuda devices */
- IPARAM_NMPI, /* Number of cuda devices */
+ IPARAM_NMPI, /* Number of MPI PROCS */
IPARAM_N, /* Number of columns of the matrix */
IPARAM_NB, /* Number of columns in a tile */
IPARAM_IB, /* Inner-blocking size */
diff --git a/include/chameleon.h b/include/chameleon.h
index 2abadc0dfd0fa4156c3fbd8647f3a126981223c2..7ac09793f4931b420310728d04760ea74172ea20 100644
--- a/include/chameleon.h
+++ b/include/chameleon.h
@@ -18,7 +18,7 @@
* @author Florent Pruvost
* @author Philippe Virouleau
* @author Lionel Eyraud-Dubois
- * @date 2024-03-11
+ * @date 2024-03-16
*
*/
#ifndef _chameleon_h_
@@ -117,6 +117,7 @@ int CHAMELEON_Initialized (void);
int CHAMELEON_My_Mpi_Rank (void) __attribute__((deprecated));
int __chameleon_init (int nworkers, int ncudas);
int __chameleon_initpar (int nworkers, int ncudas, int nthreads_per_worker);
+int __chameleon_initparcomm (int nworkers, int ncudas, int nthreads_per_worker, MPI_Comm comm);
int __chameleon_finalize (void);
int CHAMELEON_Pause (void);
int CHAMELEON_Resume (void);
@@ -237,16 +238,23 @@ void CHAMELEON_Ipiv_Print ( const CHAM_ipiv_t *ipiv );
*
*/
#if defined(CHAMELEON_SCHED_OPENMP)
-#define CHAMELEON_Init( _nworkers_, _ncudas_ ) \
+
+#define CHAMELEON_Init( _nworkers_, _ncudas_ ) \
__chameleon_init( (_nworkers_), (_ncudas_) ); \
- _Pragma("omp parallel") \
- _Pragma("omp master") \
+ _Pragma("omp parallel") \
+ _Pragma("omp master") \
{
-#define CHAMELEON_InitPar( _nworkers_, _ncudas_, _nthreads_per_worker_ ) \
+#define CHAMELEON_InitPar( _nworkers_, _ncudas_, _nthreads_per_worker_ ) \
__chameleon_initpar( (_nworkers_), (_ncudas_), (_nthreads_per_worker_) ); \
- _Pragma("omp parallel")\
- _Pragma("omp master")\
+ _Pragma("omp parallel") \
+ _Pragma("omp master") \
+ {
+
+#define CHAMELEON_InitParComm( _nworkers_, _ncudas_, _nthreads_per_worker_, _comm_ ) \
+ __chameleon_initparcomm( (_nworkers_), (_ncudas_), (_nthreads_per_worker_), (_comm_) ); \
+ _Pragma("omp parallel") \
+ _Pragma("omp master") \
{
#define CHAMELEON_Finalize() \
@@ -255,11 +263,14 @@ void CHAMELEON_Ipiv_Print ( const CHAM_ipiv_t *ipiv );
#else
-#define CHAMELEON_Init( _nworkers_, _ncudas_ ) \
+#define CHAMELEON_Init( _nworkers_, _ncudas_ ) \
__chameleon_init( (_nworkers_), (_ncudas_) );
#define CHAMELEON_InitPar( _nworkers_, _ncudas_, _nthreads_per_worker_ ) \
- __chameleon_initpar( (_nworkers_), (_ncudas_), (_nthreads_per_worker_) );
+ __chameleon_initpar( (_nworkers_), (_ncudas_), (_nthreads_per_worker_), MPI_COMM_WORLD );
+
+#define CHAMELEON_InitParComm( _nworkers_, _ncudas_, _nthreads_per_worker_, _comm_ ) \
+ __chameleon_initparcomm( (_nworkers_), (_ncudas_), (_nthreads_per_worker_), (_comm_) );
#define CHAMELEON_Finalize() \
__chameleon_finalize();
diff --git a/include/chameleon/runtime.h b/include/chameleon/runtime.h
index 9b2239c1bc53591a5d0fec3609b2ef7add1953b7..010b8b1b5eb68c9f841021a7072f8192579bb545 100644
--- a/include/chameleon/runtime.h
+++ b/include/chameleon/runtime.h
@@ -480,7 +480,7 @@ RUNTIME_desc_flush( const CHAM_desc_t *desc,
* This function flushes all data from the distributed cache of the runtime system.
*/
void
-RUNTIME_flush( );
+RUNTIME_flush( CHAM_context_t *chamctxt );
/**
* @brief Flush a single piece of data.
diff --git a/include/chameleon/runtime_struct.h b/include/chameleon/runtime_struct.h
index 3028d328bac41300310dd4f1eadeb9656466798f..c83da6fd9d1af72851219f06c9aba880331ccd8f 100644
--- a/include/chameleon/runtime_struct.h
+++ b/include/chameleon/runtime_struct.h
@@ -17,12 +17,23 @@
* @author Cedric Castagnede
* @author Florent Pruvost
* @author Philippe Virouleau
- * @date 2023-07-04
+ * @date 2024-03-16
*
*/
#ifndef _chameleon_runtime_struct_h_
#define _chameleon_runtime_struct_h_
+#if defined(CHAMELEON_USE_MPI)
+#include <mpi.h>
+#else
+#ifndef MPI_Comm
+typedef uintptr_t MPI_Comm;
+#endif
+#ifndef MPI_COMM_WORLD
+#define MPI_COMM_WORLD 0
+#endif
+#endif
+
BEGIN_C_DECLS
/**
@@ -70,6 +81,7 @@ typedef struct runtime_sequence_s {
int status; /**< Return status registered by the tasks for the request */
RUNTIME_request_t *request; /**< Pointer to the request that failed if any, NULL otherwise */
void *schedopt; /**< Specific runtime data pointer to handle the sequence */
+ MPI_Comm comm; /**< MPI communicator */
} RUNTIME_sequence_t;
/**
diff --git a/include/chameleon/struct.h b/include/chameleon/struct.h
index 00a79664c17b4d051ea0858ae910b8f87460a8d9..3664d5c2bdca39950d494c4863f1cb4b2b150a42 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 2023-08-31
+ * @date 2024-03-16
*
*/
#ifndef _chameleon_struct_h_
@@ -30,6 +30,17 @@
#include "chameleon/constants.h"
#include "chameleon/runtime_struct.h"
+#if defined(CHAMELEON_USE_MPI)
+#include <mpi.h>
+#else
+#ifndef MPI_Comm
+typedef uintptr_t MPI_Comm;
+#endif
+#ifndef MPI_COMM_WORLD
+#define MPI_COMM_WORLD 0
+#endif
+#endif
+
BEGIN_C_DECLS
#define CHAMELEON_TILE_FULLRANK (1 << 0)
@@ -191,6 +202,7 @@ typedef struct chameleon_context_s {
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;
static inline void *
diff --git a/runtime/openmp/control/runtime_descriptor.c b/runtime/openmp/control/runtime_descriptor.c
index 38ea0a3b1dbb198bee1d1bd5db1b55d49dd3c572..075724b9b9e5dbcfd250fd34eaf946391ea3ad26 100644
--- a/runtime/openmp/control/runtime_descriptor.c
+++ b/runtime/openmp/control/runtime_descriptor.c
@@ -11,12 +11,12 @@
*
* @brief Chameleon OpenMP descriptor routines
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Vijay Joshi
* @author Cedric Castagnede
* @author Philippe Virouleau
* @author Mathieu Faverge
- * @date 2022-02-22
+ * @date 2024-03-16
*
*/
#include "chameleon_openmp.h"
@@ -59,7 +59,7 @@ int RUNTIME_desc_release( const CHAM_desc_t *desc )
}
void
-RUNTIME_desc_flush( const CHAM_desc_t *desc,
+RUNTIME_desc_flush( const CHAM_desc_t *desc,
const RUNTIME_sequence_t *sequence )
{
(void)desc;
@@ -69,8 +69,9 @@ RUNTIME_desc_flush( const CHAM_desc_t *desc,
void
-RUNTIME_flush( )
+RUNTIME_flush( CHAM_context_t *chamctxt )
{
+ (void)chamctxt;
return;
}
diff --git a/runtime/parsec/control/runtime_descriptor.c b/runtime/parsec/control/runtime_descriptor.c
index 360d673261c5493d196e7f2638b427fa9df8b0cb..b1266e284513e234c72859355a79b5f3dc342b3f 100644
--- a/runtime/parsec/control/runtime_descriptor.c
+++ b/runtime/parsec/control/runtime_descriptor.c
@@ -11,12 +11,12 @@
*
* @brief Chameleon PaRSEC descriptor routines
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Reazul Hoque
* @author Mathieu Faverge
* @author Guillaume Sylvand
* @author Samuel Thibault
- * @date 2022-02-22
+ * @date 2024-03-16
*
*/
#include "chameleon_parsec.h"
@@ -345,8 +345,10 @@ int RUNTIME_desc_release( const CHAM_desc_t *desc )
/**
* Flush cached data
*/
-void RUNTIME_flush()
+void RUNTIME_flush( CHAM_context_t *chamctxt )
{
+ (void)chamctxt;
+ return;
}
void RUNTIME_desc_flush( const CHAM_desc_t *desc,
diff --git a/runtime/quark/control/runtime_descriptor.c b/runtime/quark/control/runtime_descriptor.c
index 6301b9c1408feb4f8883b73a338bcc31dedcf441..4435e06629694e45492cb5884d9797d3d3501195 100644
--- a/runtime/quark/control/runtime_descriptor.c
+++ b/runtime/quark/control/runtime_descriptor.c
@@ -11,13 +11,13 @@
*
* @brief Chameleon Quark descriptor routines
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Vijay Joshi
* @author Cedric Castagnede
* @author Florent Pruvost
* @author Mathieu Faverge
* @author Samuel Thibault
- * @date 2022-02-22
+ * @date 2024-03-16
*
*/
#include "chameleon_quark.h"
@@ -60,7 +60,7 @@ int RUNTIME_desc_release( const CHAM_desc_t *desc )
}
void
-RUNTIME_desc_flush( const CHAM_desc_t *desc,
+RUNTIME_desc_flush( const CHAM_desc_t *desc,
const RUNTIME_sequence_t *sequence )
{
(void)desc;
@@ -70,8 +70,9 @@ RUNTIME_desc_flush( const CHAM_desc_t *desc,
void
-RUNTIME_flush( )
+RUNTIME_flush( CHAM_context_t *chamctxt )
{
+ (void)chamctxt;
return;
}
diff --git a/runtime/starpu/codelets/codelet_zgersum.c b/runtime/starpu/codelets/codelet_zgersum.c
index 8f8b2eaebc2d5701e734bbf59111c4c0474d13f2..dd44fb9f61150ebf4219af469b79bfd779cf8085 100644
--- a/runtime/starpu/codelets/codelet_zgersum.c
+++ b/runtime/starpu/codelets/codelet_zgersum.c
@@ -15,7 +15,7 @@
* @author Romain Peressoni
* @author Mathieu Faverge
* @author Antoine Jego
- * @date 2023-07-06
+ * @date 2024-03-16
* @precisions normal z -> c d s
*
*/
@@ -128,7 +128,7 @@ RUNTIME_zgersum_submit_tree( const RUNTIME_option_t *options,
const CHAM_desc_t *A, int Am, int An )
{
#if defined(HAVE_STARPU_MPI_REDUX) && defined(CHAMELEON_USE_MPI)
- starpu_mpi_redux_data_prio_tree( MPI_COMM_WORLD,
+ starpu_mpi_redux_data_prio_tree( options->sequence->comm,
RTBLKADDR(A, ChamComplexDouble, Am, An),
options->priority + 1,
2 /* Binary tree */ );
diff --git a/runtime/starpu/codelets/codelet_zlacpy.c b/runtime/starpu/codelets/codelet_zlacpy.c
index af940bf2879a6b5a8985f163cb3d349b34d784b3..53647940eac27d0b6ac51d7f2f64042100841b5e 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 2023-07-06
+ * @date 2024-03-16
* @precisions normal z -> c d s
*
*/
@@ -120,9 +120,9 @@ insert_task_zlacpy_on_remote_node( const RUNTIME_option_t *options,
{
void (*callback)(void*) = options->profiling ? cl_zlacpy_callback : NULL;
#if defined(CHAMELEON_RUNTIME_SYNC)
- starpu_mpi_data_cpy_priority( handleB, handleA, MPI_COMM_WORLD, 0, callback, NULL, options->priority );
+ starpu_mpi_data_cpy_priority( handleB, handleA, options->sequence->comm, 0, callback, NULL, options->priority );
#else
- starpu_mpi_data_cpy_priority( handleB, handleA, MPI_COMM_WORLD, 1, callback, NULL, options->priority );
+ starpu_mpi_data_cpy_priority( handleB, handleA, options->sequence->comm, 1, callback, NULL, options->priority );
#endif
}
#endif
diff --git a/runtime/starpu/control/runtime_async.c b/runtime/starpu/control/runtime_async.c
index a439e5d8fe17986708508dceeab859ae922356dc..ea19203fcf95effe511255ff690ccdb7810f28eb 100644
--- a/runtime/starpu/control/runtime_async.c
+++ b/runtime/starpu/control/runtime_async.c
@@ -11,12 +11,12 @@
*
* @brief Chameleon StarPU asynchronous routines
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Mathieu Faverge
* @author Cedric Castagnede
* @author Florent Pruvost
* @author Samuel Thibault
- * @date 2022-02-22
+ * @date 2024-03-16
*
*/
#include "chameleon_starpu.h"
@@ -28,7 +28,7 @@ int RUNTIME_sequence_create( CHAM_context_t *chamctxt,
RUNTIME_sequence_t *sequence )
{
(void)chamctxt;
- (void)sequence;
+ sequence->comm = chamctxt->comm;
return CHAMELEON_SUCCESS;
}
@@ -58,10 +58,10 @@ int RUNTIME_sequence_wait( CHAM_context_t *chamctxt,
#if defined(CHAMELEON_USE_MPI)
# if defined(HAVE_STARPU_MPI_WAIT_FOR_ALL)
- starpu_mpi_wait_for_all(MPI_COMM_WORLD);
+ starpu_mpi_wait_for_all(sequence->comm);
# else
starpu_task_wait_for_all();
- starpu_mpi_barrier(MPI_COMM_WORLD);
+ starpu_mpi_barrier(sequence->comm);
# endif
#else
starpu_task_wait_for_all();
diff --git a/runtime/starpu/control/runtime_control.c b/runtime/starpu/control/runtime_control.c
index b8fd4003fa742629c5ffb518609067d67e97609d..96e1c3ff7cbad2e0c5336a2804ab5d98e08ee56d 100644
--- a/runtime/starpu/control/runtime_control.c
+++ b/runtime/starpu/control/runtime_control.c
@@ -21,7 +21,7 @@
* @author Matthieu Kuhn
* @author Loris Lucido
* @author Terry Cojean
- * @date 2023-08-22
+ * @date 2024-03-16
*
*/
#include "chameleon_starpu.h"
@@ -100,7 +100,7 @@ void chameleon_starpu_parallel_worker_fini( starpu_sched_opt_t *sched_opt )
/**
*
*/
-static int chameleon_starpu_init( struct starpu_conf *conf )
+static int chameleon_starpu_init( MPI_Comm comm, struct starpu_conf *conf )
{
int hres = CHAMELEON_SUCCESS;
int rc;
@@ -118,7 +118,7 @@ static int chameleon_starpu_init( struct starpu_conf *conf )
# endif
# if defined(HAVE_STARPU_MPI_INIT_CONF)
- rc = starpu_mpi_init_conf(NULL, NULL, !flag, MPI_COMM_WORLD, conf);
+ rc = starpu_mpi_init_conf(NULL, NULL, !flag, comm, conf);
# else
rc = starpu_init(conf);
if (rc < 0) {
@@ -186,7 +186,7 @@ int RUNTIME_init( CHAM_context_t *chamctxt,
if ((ncpus == -1)||(nthreads_per_worker == -1))
{
- hres = chameleon_starpu_init( conf );
+ hres = chameleon_starpu_init( chamctxt->comm, conf );
chamctxt->nworkers = ncpus;
chamctxt->nthreads_per_worker = nthreads_per_worker;
@@ -202,7 +202,7 @@ int RUNTIME_init( CHAM_context_t *chamctxt,
conf->use_explicit_workers_bindid = 1;
- hres = chameleon_starpu_init( conf );
+ hres = chameleon_starpu_init( chamctxt->comm, conf );
chamctxt->nworkers = ncpus;
chamctxt->nthreads_per_worker = nthreads_per_worker;
@@ -300,11 +300,11 @@ void RUNTIME_barrier( CHAM_context_t *chamctxt )
#if defined(CHAMELEON_USE_MPI)
# if defined(HAVE_STARPU_MPI_WAIT_FOR_ALL)
- starpu_mpi_wait_for_all(MPI_COMM_WORLD);
- starpu_mpi_barrier(MPI_COMM_WORLD);
+ starpu_mpi_wait_for_all( chamctxt->comm );
+ starpu_mpi_barrier( chamctxt->comm );
# else
starpu_task_wait_for_all();
- starpu_mpi_barrier(MPI_COMM_WORLD);
+ starpu_mpi_barrier( chamctxt->comm );
# endif
#else
starpu_task_wait_for_all();
@@ -380,9 +380,9 @@ int RUNTIME_comm_rank( CHAM_context_t *chamctxt )
#if defined(CHAMELEON_USE_MPI)
# if defined(HAVE_STARPU_MPI_COMM_RANK)
- starpu_mpi_comm_rank( MPI_COMM_WORLD, &rank );
+ starpu_mpi_comm_rank( chamctxt->comm, &rank );
# else
- MPI_Comm_rank( MPI_COMM_WORLD, &rank );
+ MPI_Comm_rank( chamctxt->comm, &rank );
# endif
#endif
@@ -398,9 +398,9 @@ int RUNTIME_comm_size( CHAM_context_t *chamctxt )
int size;
#if defined(CHAMELEON_USE_MPI)
# if defined(HAVE_STARPU_MPI_COMM_RANK)
- starpu_mpi_comm_size( MPI_COMM_WORLD, &size );
+ starpu_mpi_comm_size( chamctxt->comm, &size );
# else
- MPI_Comm_size( MPI_COMM_WORLD, &size );
+ MPI_Comm_size( chamctxt->comm, &size );
# endif
#else
size = 1;
diff --git a/runtime/starpu/control/runtime_descriptor.c b/runtime/starpu/control/runtime_descriptor.c
index 1e98660118946a75eec93f00bed0091b3301b7c0..2e72133c6acf979cba775db9488a0da93ffd2ab1 100644
--- a/runtime/starpu/control/runtime_descriptor.c
+++ b/runtime/starpu/control/runtime_descriptor.c
@@ -20,7 +20,7 @@
* @author Raphael Boucherie
* @author Samuel Thibault
* @author Loris Lucido
- * @date 2023-08-22
+ * @date 2024-03-16
*
*/
#include "chameleon_starpu.h"
@@ -149,7 +149,7 @@ void RUNTIME_desc_create( CHAM_desc_t *desc )
* Book the number of tags required to describe this matrix
*/
{
- chameleon_starpu_tag_init();
+ chameleon_starpu_tag_init( );
desc->mpitag = chameleon_starpu_tag_book( nbtiles );
if ( desc->mpitag == -1 ) {
@@ -267,10 +267,10 @@ int RUNTIME_desc_release( const CHAM_desc_t *desc )
/**
* Flush cached data
*/
-void RUNTIME_flush()
+void RUNTIME_flush( CHAM_context_t *chamctxt )
{
#if defined(CHAMELEON_USE_MPI)
- starpu_mpi_cache_flush_all_data(MPI_COMM_WORLD);
+ starpu_mpi_cache_flush_all_data( chamctxt->comm );
#endif
}
@@ -317,7 +317,7 @@ void RUNTIME_data_flush( const RUNTIME_sequence_t *sequence,
}
#if defined(CHAMELEON_USE_MPI)
- starpu_mpi_cache_flush( MPI_COMM_WORLD, *handlebis );
+ starpu_mpi_cache_flush( sequence->comm, *handlebis );
#endif
if ( local ) {
@@ -345,7 +345,7 @@ void RUNTIME_data_migrate( const RUNTIME_sequence_t *sequence,
old_rank = starpu_mpi_data_get_rank( lhandle );
if ( old_rank != new_rank ) {
- starpu_mpi_data_migrate( MPI_COMM_WORLD, lhandle, new_rank );
+ starpu_mpi_data_migrate( sequence->comm, lhandle, new_rank );
}
(void)sequence;
diff --git a/runtime/starpu/control/runtime_descriptor_ipiv.c b/runtime/starpu/control/runtime_descriptor_ipiv.c
index 640ffa83e0a3dac177b72d0e95b049ba8a4427cf..634378e8dc41bb41b9a0610fa7642711c9371f10 100644
--- a/runtime/starpu/control/runtime_descriptor_ipiv.c
+++ b/runtime/starpu/control/runtime_descriptor_ipiv.c
@@ -223,7 +223,7 @@ void RUNTIME_ipiv_flushk( const RUNTIME_sequence_t *sequence,
if ( *handle != NULL ) {
#if defined(CHAMELEON_USE_MPI)
- starpu_mpi_cache_flush( MPI_COMM_WORLD, *handle );
+ starpu_mpi_cache_flush( sequence->comm, *handle );
if ( starpu_mpi_data_get_rank( *handle ) == A->myrank )
#endif
{
@@ -236,7 +236,7 @@ void RUNTIME_ipiv_flushk( const RUNTIME_sequence_t *sequence,
if ( *handle != NULL ) {
#if defined(CHAMELEON_USE_MPI)
- starpu_mpi_cache_flush( MPI_COMM_WORLD, *handle );
+ starpu_mpi_cache_flush( sequence->comm, *handle );
if ( starpu_mpi_data_get_rank( *handle ) == A->myrank )
#endif
{
@@ -272,7 +272,7 @@ void RUNTIME_perm_flushk( const RUNTIME_sequence_t *sequence,
if ( *handle != NULL ) {
#if defined(CHAMELEON_USE_MPI)
- starpu_mpi_cache_flush( MPI_COMM_WORLD, *handle );
+ starpu_mpi_cache_flush( sequence->comm, *handle );
if ( starpu_mpi_data_get_rank( *handle ) == A->myrank )
#endif
{
@@ -285,7 +285,7 @@ void RUNTIME_perm_flushk( const RUNTIME_sequence_t *sequence,
if ( *handle != NULL ) {
#if defined(CHAMELEON_USE_MPI)
- starpu_mpi_cache_flush( MPI_COMM_WORLD, *handle );
+ starpu_mpi_cache_flush( sequence->comm, *handle );
if ( starpu_mpi_data_get_rank( *handle ) == A->myrank )
#endif
{
@@ -323,7 +323,7 @@ void RUNTIME_ipiv_gather( const RUNTIME_sequence_t *sequence,
if (already_received == 0)
{
MPI_Status status;
- starpu_mpi_recv( ipiv_src, owner, tag, MPI_COMM_WORLD, &status );
+ starpu_mpi_recv( ipiv_src, owner, tag, sequence->comm, &status );
}
}
else if ( rank == owner )
@@ -332,7 +332,7 @@ void RUNTIME_ipiv_gather( const RUNTIME_sequence_t *sequence,
int already_sent = starpu_mpi_cached_send_set( ipiv_src, node );
if (already_sent == 0)
{
- starpu_mpi_send( ipiv_src, node, tag, MPI_COMM_WORLD );
+ starpu_mpi_send( ipiv_src, node, tag, sequence->comm );
}
}
}
diff --git a/runtime/starpu/control/runtime_tags.c b/runtime/starpu/control/runtime_tags.c
index 57c9f859e4d3343eedc0660b98ace968e99e0335..031a556641fd4a59b57f1dd78a548d84e0787628 100644
--- a/runtime/starpu/control/runtime_tags.c
+++ b/runtime/starpu/control/runtime_tags.c
@@ -5,11 +5,11 @@
* @copyright 2017-2024 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria,
* Univ. Bordeaux. All rights reserved.
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Pierre Ramet
* @author Mathieu Faverge
* @author Florent Pruvost
- * @date 2021-10-04
+ * @date 2024-03-16
*
* Functions to manage the MPI data tags with StarPU (originated from PaStiX).
*
@@ -58,7 +58,8 @@ chameleon_starpu_tag_init( void )
int ok = 0;
void *tag_ub_p = NULL;
- starpu_mpi_comm_get_attr( MPI_COMM_WORLD, STARPU_MPI_TAG_UB, &tag_ub_p, &ok );
+ CHAM_context_t *chamctxt = chameleon_context_self();
+ starpu_mpi_comm_get_attr( chamctxt->comm, STARPU_MPI_TAG_UB, &tag_ub_p, &ok );
starpu_tag_ub = (uint64_t)((intptr_t)tag_ub_p);
if ( !ok ) {
diff --git a/runtime/starpu/include/chameleon_starpu.h.in b/runtime/starpu/include/chameleon_starpu.h.in
index fd6d0e4688bd4ce9e8ba80b1084a00c999e5cb56..b795b4c79454e65ad9e22ca9a37b124bfe6c734a 100644
--- a/runtime/starpu/include/chameleon_starpu.h.in
+++ b/runtime/starpu/include/chameleon_starpu.h.in
@@ -20,7 +20,7 @@
* @author Loris Lucido
* @author Terry Cojean
* @author Matthieu Kuhn
- * @date 2023-08-22
+ * @date 2024-03-16
*
*/
#ifndef _chameleon_starpu_h_
@@ -131,10 +131,10 @@ void *RUNTIME_data_getaddr_withconversion( const RUNTIME_option_t *options,
#if defined(CHAMELEON_RUNTIME_SYNC)
#define rt_starpu_insert_task( _codelet_, ... ) \
- starpu_mpi_insert_task( MPI_COMM_WORLD, (_codelet_), STARPU_TASK_SYNCHRONOUS, 1, ##__VA_ARGS__ )
+ starpu_mpi_insert_task( options->sequence->comm, (_codelet_), STARPU_TASK_SYNCHRONOUS, 1, ##__VA_ARGS__ )
#else
#define rt_starpu_insert_task( _codelet_, ... ) \
- starpu_mpi_insert_task( MPI_COMM_WORLD, (_codelet_), ##__VA_ARGS__ )
+ starpu_mpi_insert_task( options->sequence->comm, (_codelet_), ##__VA_ARGS__ )
#endif
#else