From 410358afefe23d4341569522592bbe5640777dd5 Mon Sep 17 00:00:00 2001
From: Alycia Lisito <alycia.lisito@inria.fr>
Date: Thu, 8 Aug 2024 16:59:08 +0200
Subject: [PATCH 1/6] zgetrf: Add all_reduce algorithm choice
---
compute/zgetrf.c | 15 +++++++++++++++
control/compute_z.h | 23 ++++++++++++-----------
include/chameleon/constants.h | 7 +++++++
3 files changed, 34 insertions(+), 11 deletions(-)
diff --git a/compute/zgetrf.c b/compute/zgetrf.c
index 514e89d3e..976ba2ad5 100644
--- a/compute/zgetrf.c
+++ b/compute/zgetrf.c
@@ -98,6 +98,21 @@ CHAMELEON_zgetrf_WS_Alloc( const CHAM_desc_t *A )
chameleon_cleanenv( algostr );
}
+ {
+ char *allreduce = chameleon_getenv( "CHAMELEON_GETRF_ALL_REDUCE" );
+
+ if ( allreduce != NULL ) {
+ if ( strcasecmp( allreduce, "cham_spu_tasks" ) == 0 ) {
+ ws->alg_allreduce = ChamStarPUTasks;
+ }
+ else {
+ chameleon_error( "CHAMELEON_zgetrf_WS_Alloc", "CHAMELEON_GETRF_ALL_REDUCE is not one of chameleon_starpu_tasks, chameleon_starpu, chameleon_starpu_mpi, chameleon_mpi => Switch back to chameleon_starpu_tasks\n" );
+ ws->alg_allreduce = ChamStarPUTasks;
+ }
+ }
+ chameleon_cleanenv( allreduce );
+ }
+
ws->batch_size = chameleon_getenv_get_value_int( "CHAMELEON_GETRF_BATCH_SIZE", 0 );
if ( ws->batch_size > CHAMELEON_BATCH_SIZE ) {
chameleon_warning( "CHAMELEON_BATCH_SIZE", "CHAMELEON_GETRF_BATCH_SIZE must be smaller than CHAMELEON_BATCH_SIZE, please recompile with the right CHAMELEON_BATCH_SIZE, or reduce the CHAMELEON_GETRF_BATCH_SIZE value\n" );
diff --git a/control/compute_z.h b/control/compute_z.h
index b75c303a1..855f820b1 100644
--- a/control/compute_z.h
+++ b/control/compute_z.h
@@ -43,17 +43,18 @@ struct chameleon_pzgemm_s {
* @brief Data structure to handle the GETRF workspaces with partial pivoting
*/
struct chameleon_pzgetrf_s {
- cham_getrf_t alg;
- int ib; /**< Internal blocking parameter */
- int batch_size; /**< Batch size for the panel */
- int ringswitch; /**< Define when to switch to ring bcast */
- CHAM_desc_t U;
- CHAM_desc_t Up; /**< Workspace used for the panel factorization */
- CHAM_desc_t Wu; /**< Workspace used for the permutation and update */
- CHAM_desc_t Wl; /**< Workspace used the update */
- int *proc_involved;
- unsigned int involved;
- int np_involved;
+ cham_getrf_t alg;
+ cham_getrf_allreduce_t alg_allreduce;
+ int ib; /**< Internal blocking parameter */
+ int batch_size; /**< Batch size for the panel */
+ int ringswitch; /**< Define when to switch to ring bcast */
+ CHAM_desc_t U;
+ CHAM_desc_t Up; /**< Workspace used for the panel factorization */
+ CHAM_desc_t Wu; /**< Workspace used for the permutation and update */
+ CHAM_desc_t Wl; /**< Workspace used the update */
+ int *proc_involved;
+ unsigned int involved;
+ int np_involved;
};
/**
diff --git a/include/chameleon/constants.h b/include/chameleon/constants.h
index 09bfc942f..88c1f653b 100644
--- a/include/chameleon/constants.h
+++ b/include/chameleon/constants.h
@@ -290,6 +290,13 @@ typedef enum chameleon_getrf_e {
ChamGetrfPPivPerColumn = 3,
} cham_getrf_t;
+/**
+ * @brief Chameleon GETRF all reduce algorithm variants
+ */
+typedef enum chameleon_getrf_allreduce_e {
+ ChamStarPUTasks,
+} cham_getrf_allreduce_t;
+
#define ChameleonTrd 1001
#define ChameleonBrd 1002
--
GitLab
From 6ee85f43a4a7a0711f9873d8c02e71e040863d53 Mon Sep 17 00:00:00 2001
From: Alycia Lisito <alycia.lisito@inria.fr>
Date: Fri, 25 Oct 2024 11:32:49 +0200
Subject: [PATCH 2/6] zgetrf: zipiv allreduce MPI in task prepare codelet
---
compute/pzgetrf.c | 8 +-
include/chameleon/tasks_z.h | 8 +-
.../openmp/codelets/codelet_zipiv_allreduce.c | 12 +-
.../parsec/codelets/codelet_zipiv_allreduce.c | 12 +-
.../quark/codelets/codelet_zipiv_allreduce.c | 12 +-
.../starpu/codelets/codelet_zipiv_allreduce.c | 126 +++++++++++-------
6 files changed, 103 insertions(+), 75 deletions(-)
diff --git a/compute/pzgetrf.c b/compute/pzgetrf.c
index 6db9a8a40..4347f5710 100644
--- a/compute/pzgetrf.c
+++ b/compute/pzgetrf.c
@@ -150,7 +150,7 @@ chameleon_pzgetrf_panel_facto_percol( struct chameleon_pzgetrf_s *ws,
}
/* Reduce globally (between MPI processes) */
- INSERT_TASK_zipiv_allreduce( A, options, ipiv, ws->proc_involved, k, h, tempkn );
+ INSERT_TASK_zipiv_allreduce( options, A, ipiv, k, h, tempkn, ws );
}
/* Flush temporary data used for the pivoting */
@@ -196,7 +196,7 @@ chameleon_pzgetrf_panel_facto_percol_batched( struct chameleon_pzgetrf_s *ws,
}
INSERT_TASK_zgetrf_panel_offdiag_batched_flush( options, A, k, clargs, ipiv );
- INSERT_TASK_zipiv_allreduce( A, options, ipiv, ws->proc_involved, k, h, tempkn );
+ INSERT_TASK_zipiv_allreduce( options, A, ipiv, k, h, tempkn, ws );
}
free( clargs );
@@ -250,7 +250,7 @@ chameleon_pzgetrf_panel_facto_blocked( struct chameleon_pzgetrf_s *ws,
assert( j <= minmn );
/* Reduce globally (between MPI processes) */
- INSERT_TASK_zipiv_allreduce( A, options, ipiv, ws->proc_involved, k, j, tempkn );
+ INSERT_TASK_zipiv_allreduce( options, A, ipiv, k, j, tempkn, ws );
if ( ( b < (nbblock-1) ) && ( h == hmax-1 ) ) {
INSERT_TASK_zgetrf_blocked_trsm(
@@ -312,7 +312,7 @@ chameleon_pzgetrf_panel_facto_blocked_batched( struct chameleon_pzgetrf_s *ws,
assert( j <= minmn );
/* Reduce globally (between MPI processes) */
- INSERT_TASK_zipiv_allreduce( A, options, ipiv, ws->proc_involved, k, j, tempkn );
+ INSERT_TASK_zipiv_allreduce( options, A, ipiv, k, j, tempkn, ws );
if ( (b < (nbblock-1)) && (h == hmax-1) ) {
INSERT_TASK_zgetrf_blocked_trsm(
diff --git a/include/chameleon/tasks_z.h b/include/chameleon/tasks_z.h
index 5f1bbcd32..402c92a3f 100644
--- a/include/chameleon/tasks_z.h
+++ b/include/chameleon/tasks_z.h
@@ -575,13 +575,13 @@ void INSERT_TASK_zgetrf_blocked_trsm( const RUNTIME_option_t *options,
CHAM_desc_t *U, int Um, int Un,
CHAM_ipiv_t *ws );
-void INSERT_TASK_zipiv_allreduce( CHAM_desc_t *A,
- const RUNTIME_option_t *options,
+void INSERT_TASK_zipiv_allreduce( const RUNTIME_option_t *options,
+ CHAM_desc_t *A,
CHAM_ipiv_t *ipiv,
- int *proc_involved,
int k,
int h,
- int n );
+ int n,
+ void *ws );
/**
********************************************************************************
diff --git a/runtime/openmp/codelets/codelet_zipiv_allreduce.c b/runtime/openmp/codelets/codelet_zipiv_allreduce.c
index b08828325..197842ea3 100644
--- a/runtime/openmp/codelets/codelet_zipiv_allreduce.c
+++ b/runtime/openmp/codelets/codelet_zipiv_allreduce.c
@@ -17,19 +17,19 @@
*/
#include "chameleon_openmp.h"
-void INSERT_TASK_zipiv_allreduce( CHAM_desc_t *A,
- const RUNTIME_option_t *options,
+void INSERT_TASK_zipiv_allreduce( const RUNTIME_option_t *options,
+ CHAM_desc_t *A,
CHAM_ipiv_t *ipiv,
- int *proc_involved,
int k,
int h,
- int n )
+ int n,
+ void *ws )
{
- (void)A;
(void)options;
+ (void)A;
(void)ipiv;
- (void)proc_involved;
(void)k;
(void)h;
(void)n;
+ (void)ws;
}
diff --git a/runtime/parsec/codelets/codelet_zipiv_allreduce.c b/runtime/parsec/codelets/codelet_zipiv_allreduce.c
index 75e061164..d6bd3f4c0 100644
--- a/runtime/parsec/codelets/codelet_zipiv_allreduce.c
+++ b/runtime/parsec/codelets/codelet_zipiv_allreduce.c
@@ -17,19 +17,19 @@
*/
#include "chameleon_parsec.h"
-void INSERT_TASK_zipiv_allreduce( CHAM_desc_t *A,
- const RUNTIME_option_t *options,
+void INSERT_TASK_zipiv_allreduce( const RUNTIME_option_t *options,
+ CHAM_desc_t *A,
CHAM_ipiv_t *ipiv,
- int *proc_involved,
int k,
int h,
- int n )
+ int n,
+ void *ws )
{
- (void)A;
(void)options;
+ (void)A;
(void)ipiv;
- (void)proc_involved;
(void)k;
(void)h;
(void)n;
+ (void)ws;
}
diff --git a/runtime/quark/codelets/codelet_zipiv_allreduce.c b/runtime/quark/codelets/codelet_zipiv_allreduce.c
index e88269e93..0186fd142 100644
--- a/runtime/quark/codelets/codelet_zipiv_allreduce.c
+++ b/runtime/quark/codelets/codelet_zipiv_allreduce.c
@@ -17,19 +17,19 @@
*/
#include "chameleon_quark.h"
-void INSERT_TASK_zipiv_allreduce( CHAM_desc_t *A,
- const RUNTIME_option_t *options,
+void INSERT_TASK_zipiv_allreduce( const RUNTIME_option_t *options,
+ CHAM_desc_t *A,
CHAM_ipiv_t *ipiv,
- int *proc_involved,
int k,
int h,
- int n )
+ int n,
+ void *ws )
{
- (void)A;
(void)options;
+ (void)A;
(void)ipiv;
- (void)proc_involved;
(void)k;
(void)h;
(void)n;
+ (void)ws;
}
diff --git a/runtime/starpu/codelets/codelet_zipiv_allreduce.c b/runtime/starpu/codelets/codelet_zipiv_allreduce.c
index a81f0d08e..48ecdd0c3 100644
--- a/runtime/starpu/codelets/codelet_zipiv_allreduce.c
+++ b/runtime/starpu/codelets/codelet_zipiv_allreduce.c
@@ -22,18 +22,18 @@
struct cl_redux_args_t {
int h;
int n;
- int k;
};
-static void cl_zipiv_allreduce_cpu_func( void *descr[], void *cl_arg )
+static void
+zipiv_allreduce_cpu_func( cppi_interface_t *cppi_me,
+ cppi_interface_t *cppi_src,
+ int h,
+ int n )
{
- struct cl_redux_args_t *clargs = (struct cl_redux_args_t *) cl_arg;
- cppi_interface_t *cppi_me = ((cppi_interface_t *) descr[0]);
- cppi_interface_t *cppi_src = ((cppi_interface_t *) descr[1]);
- CHAM_pivot_t *nextpiv_me = &(cppi_me->pivot);
- CHAM_pivot_t *nextpiv_src = &(cppi_src->pivot);
- CHAMELEON_Complex64_t *pivrow_me = (CHAMELEON_Complex64_t *)(nextpiv_me->pivrow);
- CHAMELEON_Complex64_t *pivrow_src = (CHAMELEON_Complex64_t *)(nextpiv_src->pivrow);
+ CHAM_pivot_t *nextpiv_me = &(cppi_me->pivot);
+ CHAM_pivot_t *nextpiv_src = &(cppi_src->pivot);
+ CHAMELEON_Complex64_t *pivrow_me = (CHAMELEON_Complex64_t *)(nextpiv_me->pivrow);
+ CHAMELEON_Complex64_t *pivrow_src = (CHAMELEON_Complex64_t *)(nextpiv_src->pivrow);
cppi_display_dbg( cppi_me, stderr, "Global redux Inout: ");
cppi_display_dbg( cppi_src, stderr, "Global redux Input: ");
@@ -43,33 +43,42 @@ static void cl_zipiv_allreduce_cpu_func( void *descr[], void *cl_arg )
assert( cppi_me->flttype == cppi_src->flttype );
assert( cppi_me->arraysize == cppi_src->arraysize );
- if ( cabs( pivrow_src[ clargs->h ] ) > cabs( pivrow_me[ clargs->h ] ) ) {
+ if ( cabs( pivrow_src[ h ] ) > cabs( pivrow_me[ h ] ) ) {
nextpiv_me->blkm0 = nextpiv_src->blkm0;
nextpiv_me->blkidx = nextpiv_src->blkidx;
- cblas_zcopy( clargs->n, pivrow_src, 1, pivrow_me, 1 );
+ cblas_zcopy( n, pivrow_src, 1, pivrow_me, 1 );
}
/* Let's copy the diagonal row if needed */
if ( ( cppi_src->has_diag == 1 ) &&
( cppi_me->has_diag == -1 ) )
{
- cblas_zcopy( clargs->n, nextpiv_src->diagrow, 1, nextpiv_me->diagrow, 1 );
- assert( cppi_src->arraysize == sizeof(CHAMELEON_Complex64_t) * clargs->n );
+ cblas_zcopy( n, nextpiv_src->diagrow, 1, nextpiv_me->diagrow, 1 );
+ assert( cppi_src->arraysize == sizeof(CHAMELEON_Complex64_t) * n );
cppi_me->has_diag = 1;
}
cppi_display_dbg( cppi_me, stderr, "Global redux Inout(After): ");
}
+static void
+cl_zipiv_allreduce_cpu_func( void *descr[], void *cl_arg )
+{
+ struct cl_redux_args_t *clargs = (struct cl_redux_args_t *) cl_arg;
+ cppi_interface_t *cppi_me = ((cppi_interface_t *) descr[0]);
+ cppi_interface_t *cppi_src = ((cppi_interface_t *) descr[1]);
+ zipiv_allreduce_cpu_func( cppi_me, cppi_src, clargs->h, clargs->n );
+}
+
CODELETS_CPU( zipiv_allreduce, cl_zipiv_allreduce_cpu_func )
-void
-INSERT_TASK_zipiv_allreduce_send( CHAM_ipiv_t *ipiv,
- int me,
- int dst,
- int k,
- int h,
- const RUNTIME_option_t *options )
+static void
+INSERT_TASK_zipiv_allreduce_send( const RUNTIME_option_t *options,
+ CHAM_ipiv_t *ipiv,
+ int me,
+ int dst,
+ int k,
+ int h )
{
rt_starpu_insert_task(
NULL,
@@ -79,20 +88,19 @@ INSERT_TASK_zipiv_allreduce_send( CHAM_ipiv_t *ipiv,
0 );
}
-void
-INSERT_TASK_zipiv_allreduce_recv( CHAM_ipiv_t *ipiv,
- int me,
- int src,
- int k,
- int h,
- int n,
- const RUNTIME_option_t *options )
+static void
+INSERT_TASK_zipiv_allreduce_recv( const RUNTIME_option_t *options,
+ CHAM_ipiv_t *ipiv,
+ int me,
+ int src,
+ int k,
+ int h,
+ int n )
{
struct cl_redux_args_t *clargs;
- clargs = malloc( sizeof( struct cl_redux_args_t ) );
+ clargs = malloc( sizeof( struct cl_redux_args_t ) );
clargs->h = h;
clargs->n = n;
- clargs->k = k;
rt_starpu_insert_task(
&cl_zipiv_allreduce,
@@ -106,16 +114,17 @@ INSERT_TASK_zipiv_allreduce_recv( CHAM_ipiv_t *ipiv,
starpu_mpi_cache_flush( options->sequence->comm, RUNTIME_pivot_getaddr( ipiv, src, k, h ) );
}
-void INSERT_TASK_zipiv_allreduce( CHAM_desc_t *A,
- const RUNTIME_option_t *options,
- CHAM_ipiv_t *ipiv,
- int *proc_involved,
- int k,
- int h,
- int n )
+static void
+zipiv_allreduce_chameleon_starpu_task( const RUNTIME_option_t *options,
+ CHAM_desc_t *A,
+ CHAM_ipiv_t *ipiv,
+ int *proc_involved,
+ int k,
+ int h,
+ int n )
{
- int np_involved = chameleon_min( chameleon_desc_datadist_get_iparam(A, 0), A->mt - k);
- int np_iter = np_involved;
+ int np_involved = chameleon_min( chameleon_desc_datadist_get_iparam(A, 0), A->mt - k);
+ int np_iter = np_involved;
int p_recv, p_send, me;
int shift = 1;
@@ -140,29 +149,48 @@ void INSERT_TASK_zipiv_allreduce( CHAM_desc_t *A,
p_send = proc_involved[ ( me + shift ) % np_involved ];
p_recv = proc_involved[ ( me - shift + np_involved ) % np_involved ];
- INSERT_TASK_zipiv_allreduce_send( ipiv, A->myrank, p_send, k, h, options );
- INSERT_TASK_zipiv_allreduce_recv( ipiv, A->myrank, p_recv, k, h, n, options );
+ INSERT_TASK_zipiv_allreduce_send( options, ipiv, A->myrank, p_send, k, h );
+ INSERT_TASK_zipiv_allreduce_recv( options, ipiv, A->myrank, p_recv, k, h, n );
shift = shift << 1;
np_iter = chameleon_ceil( np_iter, 2 );
}
}
}
+
+void
+INSERT_TASK_zipiv_allreduce( const RUNTIME_option_t *options,
+ CHAM_desc_t *A,
+ CHAM_ipiv_t *ipiv,
+ int k,
+ int h,
+ int n,
+ void *ws )
+{
+ struct chameleon_pzgetrf_s *tmp = (struct chameleon_pzgetrf_s *)ws;
+ cham_getrf_allreduce_t alg = tmp->alg_allreduce;
+ switch( alg ) {
+ case ChamStarPUTasks:
+ default:
+ zipiv_allreduce_chameleon_starpu_task( options, A, ipiv, tmp->proc_involved, k, h, n );
+ }
+}
#else
-void INSERT_TASK_zipiv_allreduce( CHAM_desc_t *A,
- const RUNTIME_option_t *options,
- CHAM_ipiv_t *ipiv,
- int *proc_involved,
- int k,
- int h,
- int n )
+void
+INSERT_TASK_zipiv_allreduce( const RUNTIME_option_t *options,
+ CHAM_desc_t *A,
+ CHAM_ipiv_t *ipiv,
+ int k,
+ int h,
+ int n,
+ void *ws )
{
if ( h > 0 ) {
starpu_data_invalidate_submit( RUNTIME_pivot_getaddr( ipiv, A->myrank, k, h-1 ) );
}
(void)options;
- (void)proc_involved;
+ (void)ws;
(void)n;
}
#endif
--
GitLab
From 488866d4968e6d2b0bf44eefe5a8004684f27638 Mon Sep 17 00:00:00 2001
From: Alycia Lisito <alycia.lisito@inria.fr>
Date: Mon, 21 Oct 2024 14:03:18 +0200
Subject: [PATCH 3/6] zgetrf: zperm allreduce MPI in task prepare codelet
---
compute/pzgetrf.c | 5 +-
include/chameleon/tasks_z.h | 26 +++----
.../openmp/codelets/codelet_zperm_allreduce.c | 12 +--
.../parsec/codelets/codelet_zperm_allreduce.c | 12 +--
.../quark/codelets/codelet_zperm_allreduce.c | 12 +--
.../starpu/codelets/codelet_zperm_allreduce.c | 74 ++++++++++++-------
6 files changed, 80 insertions(+), 61 deletions(-)
diff --git a/compute/pzgetrf.c b/compute/pzgetrf.c
index 4347f5710..ca7f4d120 100644
--- a/compute/pzgetrf.c
+++ b/compute/pzgetrf.c
@@ -433,8 +433,7 @@ chameleon_pzgetrf_panel_permute( struct chameleon_pzgetrf_s *ws,
ipiv, k, A(k, n), A(m, n) );
}
- INSERT_TASK_zperm_allreduce( options, A, ipiv, k, k, n,
- Wu(A->myrank, n), ws );
+ INSERT_TASK_zperm_allreduce( options, A, Wu(A->myrank, n), ipiv, k, k, n, ws );
}
break;
default:
@@ -499,7 +498,7 @@ chameleon_pzgetrf_panel_permute_batched( struct chameleon_pzgetrf_s *ws,
}
INSERT_TASK_zlaswp_batched_flush( options, ipiv, k, A(k, n), Wu(A->myrank, n), clargs );
- INSERT_TASK_zperm_allreduce( options, A, ipiv, k, k, n, Wu(A->myrank, n), ws );
+ INSERT_TASK_zperm_allreduce( options, A, Wu(A->myrank, n), ipiv, k, k, n, ws );
free( clargs );
}
diff --git a/include/chameleon/tasks_z.h b/include/chameleon/tasks_z.h
index 402c92a3f..bf3831af5 100644
--- a/include/chameleon/tasks_z.h
+++ b/include/chameleon/tasks_z.h
@@ -600,6 +600,16 @@ void INSERT_TASK_zipiv_allreduce( const RUNTIME_option_t *options,
* @param[in] A
* The descriptor of the matrix A.
*
+ * @param[inout] U
+ * The descriptor of the worskpace used for the permutation in the LU
+ * factorization with partial pivoting.
+ *
+ * @param[in] Um
+ * The row index of the tile used in U.
+ *
+ * @param[in] Un
+ * The column index of the tile used in U.
+ *
* @param[in] ipiv
* The pivot structure that contains the informations for the LU
* factorization with partial pivoting.
@@ -613,16 +623,6 @@ void INSERT_TASK_zipiv_allreduce( const RUNTIME_option_t *options,
* @param[in] n
* The number of columns in the tile U(Um, Un).
*
- * @param[inout] U
- * The descriptor of the worskpace used for the permutation in the LU
- * factorization with partial pivoting.
- *
- * @param[in] Um
- * The row index of the tile used in U.
- *
- * @param[in] Un
- * The column index of the tile used in U.
- *
* @param[in] ws
* The workspace to handle the data in the LU factorization with
* partial pivoting.
@@ -631,13 +631,13 @@ void INSERT_TASK_zipiv_allreduce( const RUNTIME_option_t *options,
*/
void INSERT_TASK_zperm_allreduce( const RUNTIME_option_t *options,
const CHAM_desc_t *A,
+ CHAM_desc_t *U,
+ int Um,
+ int Un,
CHAM_ipiv_t *ipiv,
int ipivk,
int k,
int n,
- CHAM_desc_t *U,
- int Um,
- int Un,
void *ws );
/**
diff --git a/runtime/openmp/codelets/codelet_zperm_allreduce.c b/runtime/openmp/codelets/codelet_zperm_allreduce.c
index cb77c806b..7aeb24fae 100644
--- a/runtime/openmp/codelets/codelet_zperm_allreduce.c
+++ b/runtime/openmp/codelets/codelet_zperm_allreduce.c
@@ -71,23 +71,23 @@ INSERT_TASK_zperm_allreduce_send_invp( const RUNTIME_option_t *options,
void
INSERT_TASK_zperm_allreduce( const RUNTIME_option_t *options,
const CHAM_desc_t *A,
+ CHAM_desc_t *U,
+ int Um,
+ int Un,
CHAM_ipiv_t *ipiv,
int ipivk,
int k,
int n,
- CHAM_desc_t *U,
- int Um,
- int Un,
void *ws )
{
(void)options;
(void)A;
+ (void)U;
+ (void)Um;
+ (void)Un;
(void)ipiv;
(void)ipivk;
(void)k;
(void)n;
- (void)U;
- (void)Um;
- (void)Un;
(void)ws;
}
diff --git a/runtime/parsec/codelets/codelet_zperm_allreduce.c b/runtime/parsec/codelets/codelet_zperm_allreduce.c
index 30890f811..5acfa4a2b 100644
--- a/runtime/parsec/codelets/codelet_zperm_allreduce.c
+++ b/runtime/parsec/codelets/codelet_zperm_allreduce.c
@@ -71,23 +71,23 @@ INSERT_TASK_zperm_allreduce_send_invp( const RUNTIME_option_t *options,
void
INSERT_TASK_zperm_allreduce( const RUNTIME_option_t *options,
const CHAM_desc_t *A,
+ CHAM_desc_t *U,
+ int Um,
+ int Un,
CHAM_ipiv_t *ipiv,
int ipivk,
int k,
int n,
- CHAM_desc_t *U,
- int Um,
- int Un,
void *ws )
{
(void)options;
(void)A;
+ (void)U;
+ (void)Um;
+ (void)Un;
(void)ipiv;
(void)ipivk;
(void)k;
(void)n;
- (void)U;
- (void)Um;
- (void)Un;
(void)ws;
}
diff --git a/runtime/quark/codelets/codelet_zperm_allreduce.c b/runtime/quark/codelets/codelet_zperm_allreduce.c
index 52281451d..f6c5f98e6 100644
--- a/runtime/quark/codelets/codelet_zperm_allreduce.c
+++ b/runtime/quark/codelets/codelet_zperm_allreduce.c
@@ -71,23 +71,23 @@ INSERT_TASK_zperm_allreduce_send_invp( const RUNTIME_option_t *options,
void
INSERT_TASK_zperm_allreduce( const RUNTIME_option_t *options,
const CHAM_desc_t *A,
+ CHAM_desc_t *U,
+ int Um,
+ int Un,
CHAM_ipiv_t *ipiv,
int ipivk,
int k,
int n,
- CHAM_desc_t *U,
- int Um,
- int Un,
void *ws )
{
(void)options;
(void)A;
+ (void)U;
+ (void)Um;
+ (void)Un;
(void)ipiv;
(void)ipivk;
(void)k;
(void)n;
- (void)U;
- (void)Um;
- (void)Un;
(void)ws;
}
diff --git a/runtime/starpu/codelets/codelet_zperm_allreduce.c b/runtime/starpu/codelets/codelet_zperm_allreduce.c
index 4c33a2e50..1c8d44164 100644
--- a/runtime/starpu/codelets/codelet_zperm_allreduce.c
+++ b/runtime/starpu/codelets/codelet_zperm_allreduce.c
@@ -102,14 +102,14 @@ INSERT_TASK_zperm_allreduce_recv( const RUNTIME_option_t *options,
{
struct cl_redux_args_t *clargs;
clargs = malloc( sizeof( struct cl_redux_args_t ) );
- clargs->tempmm = tempmm;
- clargs->n = n;
- clargs->p = p;
- clargs->q = q;
- clargs->p_first = p_first;
- clargs->me = me;
- clargs->shift = shift;
- clargs->np_inv = np;
+ clargs->tempmm = tempmm;
+ clargs->n = n;
+ clargs->p = p;
+ clargs->q = q;
+ clargs->p_first = p_first;
+ clargs->me = me;
+ clargs->shift = shift;
+ clargs->np_inv = np;
rt_starpu_insert_task(
&cl_zperm_allreduce,
@@ -124,20 +124,19 @@ INSERT_TASK_zperm_allreduce_recv( const RUNTIME_option_t *options,
starpu_mpi_cache_flush( options->sequence->comm, RTBLKADDR(U, CHAMELEON_Complex64_t, src, n) );
}
-void
-INSERT_TASK_zperm_allreduce( const RUNTIME_option_t *options,
- const CHAM_desc_t *A,
- CHAM_ipiv_t *ipiv,
- int ipivk,
- int k,
- int n,
- CHAM_desc_t *U,
- int Um,
- int Un,
- void *ws )
+static void
+zperm_allreduce_chameleon_starpu_task( const RUNTIME_option_t *options,
+ const CHAM_desc_t *A,
+ CHAM_desc_t *U,
+ int Um,
+ int Un,
+ CHAM_ipiv_t *ipiv,
+ int ipivk,
+ int k,
+ int n,
+ struct chameleon_pzgetrf_s *ws)
{
- struct chameleon_pzgetrf_s *tmp = (struct chameleon_pzgetrf_s *)ws;
- int *proc_involved = tmp->proc_involved;
+ int *proc_involved = ws->proc_involved;
int np_involved = chameleon_min( chameleon_desc_datadist_get_iparam(A, 0), A->mt - k);
int np_iter = np_involved;
int p_recv, p_send, me, p_first;
@@ -169,6 +168,27 @@ INSERT_TASK_zperm_allreduce( const RUNTIME_option_t *options,
}
}
+void
+INSERT_TASK_zperm_allreduce( const RUNTIME_option_t *options,
+ const CHAM_desc_t *A,
+ CHAM_desc_t *U,
+ int Um,
+ int Un,
+ CHAM_ipiv_t *ipiv,
+ int ipivk,
+ int k,
+ int n,
+ void *ws )
+{
+ struct chameleon_pzgetrf_s *tmp = (struct chameleon_pzgetrf_s *)ws;
+ cham_getrf_allreduce_t alg = tmp->alg_allreduce;
+ switch( alg ) {
+ case ChamStarPUTasks:
+ default:
+ zperm_allreduce_chameleon_starpu_task( options, A, U, Um, Un, ipiv, ipivk, k, n, tmp );
+ }
+}
+
void
INSERT_TASK_zperm_allreduce_send_A( const RUNTIME_option_t *options,
CHAM_desc_t *A,
@@ -284,24 +304,24 @@ INSERT_TASK_zperm_allreduce_send_invp( const RUNTIME_option_t *options,
void
INSERT_TASK_zperm_allreduce( const RUNTIME_option_t *options,
const CHAM_desc_t *A,
+ CHAM_desc_t *U,
+ int Um,
+ int Un,
CHAM_ipiv_t *ipiv,
int ipivk,
int k,
int n,
- CHAM_desc_t *U,
- int Um,
- int Un,
void *ws )
{
(void)options;
(void)A;
+ (void)U;
+ (void)Um;
+ (void)Un;
(void)ipiv;
(void)ipivk;
(void)k;
(void)n;
- (void)U;
- (void)Um;
- (void)Un;
(void)ws;
}
#endif
--
GitLab
From 77a4aaebbb4f39d9af9c20c5448f8b1827eef242 Mon Sep 17 00:00:00 2001
From: Alycia Lisito <alycia.lisito@inria.fr>
Date: Wed, 11 Dec 2024 13:13:37 +0100
Subject: [PATCH 4/6] zgetrf: different values for batch size (blas 2
operations, blas 3 operations and swap)
---
compute/pzgetrf.c | 6 +++---
compute/zgetrf.c | 18 ++++++++++++++----
control/compute_z.h | 4 +++-
.../starpu/codelets/codelet_zgetrf_batched.c | 7 ++++---
.../starpu/codelets/codelet_zlaswp_batched.c | 2 +-
5 files changed, 25 insertions(+), 12 deletions(-)
diff --git a/compute/pzgetrf.c b/compute/pzgetrf.c
index ca7f4d120..98f9d0470 100644
--- a/compute/pzgetrf.c
+++ b/compute/pzgetrf.c
@@ -350,7 +350,7 @@ chameleon_pzgetrf_panel_facto( struct chameleon_pzgetrf_s *ws,
break;
case ChamGetrfPPivPerColumn:
- if ( ws->batch_size > 0 ) {
+ if ( ws->batch_size_blas2 > 0 ) {
chameleon_pzgetrf_panel_facto_percol_batched( ws, A, ipiv, k, options );
}
else {
@@ -359,7 +359,7 @@ chameleon_pzgetrf_panel_facto( struct chameleon_pzgetrf_s *ws,
break;
case ChamGetrfPPiv:
- if ( ws->batch_size > 0 ) {
+ if ( ws->batch_size_blas2 > 0 ) {
chameleon_pzgetrf_panel_facto_blocked_batched( ws, A, ipiv, k, options );
}
else {
@@ -583,7 +583,7 @@ chameleon_pzgetrf_panel_update( struct chameleon_pzgetrf_s *ws,
tempkm = A->get_blkdim( A, k, DIM_m, A->m );
tempnn = A->get_blkdim( A, n, DIM_n, A->n );
- if ( ws->batch_size > 0 ) {
+ if ( ws->batch_size_swap > 0 ) {
chameleon_pzgetrf_panel_permute_batched( ws, A, ipiv, k, n, options );
}
else {
diff --git a/compute/zgetrf.c b/compute/zgetrf.c
index 976ba2ad5..254020a55 100644
--- a/compute/zgetrf.c
+++ b/compute/zgetrf.c
@@ -113,10 +113,20 @@ CHAMELEON_zgetrf_WS_Alloc( const CHAM_desc_t *A )
chameleon_cleanenv( allreduce );
}
- ws->batch_size = chameleon_getenv_get_value_int( "CHAMELEON_GETRF_BATCH_SIZE", 0 );
- if ( ws->batch_size > CHAMELEON_BATCH_SIZE ) {
- chameleon_warning( "CHAMELEON_BATCH_SIZE", "CHAMELEON_GETRF_BATCH_SIZE must be smaller than CHAMELEON_BATCH_SIZE, please recompile with the right CHAMELEON_BATCH_SIZE, or reduce the CHAMELEON_GETRF_BATCH_SIZE value\n" );
- ws->batch_size = CHAMELEON_BATCH_SIZE;
+ ws->batch_size_blas2 = chameleon_getenv_get_value_int( "CHAMELEON_GETRF_BATCH_SIZE_BLAS2", 0 );
+ if ( ws->batch_size_blas2 > CHAMELEON_BATCH_SIZE ) {
+ chameleon_warning( "CHAMELEON_BATCH_SIZE", "CHAMELEON_GETRF_BATCH_SIZE_BLAS2 must be smaller than CHAMELEON_BATCH_SIZE, please recompile with the right CHAMELEON_BATCH_SIZE, or reduce the CHAMELEON_GETRF_BATCH_SIZE_BLAS2 value\n" );
+ ws->batch_size_blas2 = CHAMELEON_BATCH_SIZE;
+ }
+ ws->batch_size_blas3 = chameleon_getenv_get_value_int( "CHAMELEON_GETRF_BATCH_SIZE_BLAS3", 0 );
+ if ( ws->batch_size_blas3 > CHAMELEON_BATCH_SIZE ) {
+ chameleon_warning( "CHAMELEON_BATCH_SIZE", "CHAMELEON_GETRF_BATCH_SIZE_BLAS3 must be smaller than CHAMELEON_BATCH_SIZE, please recompile with the right CHAMELEON_BATCH_SIZE, or reduce the CHAMELEON_GETRF_BATCH_SIZE_BLAS3 value\n" );
+ ws->batch_size_blas3 = CHAMELEON_BATCH_SIZE;
+ }
+ ws->batch_size_swap = chameleon_getenv_get_value_int( "CHAMELEON_GETRF_BATCH_SIZE_SWAP", 0 );
+ if ( ws->batch_size_swap > CHAMELEON_BATCH_SIZE ) {
+ chameleon_warning( "CHAMELEON_BATCH_SIZE", "CHAMELEON_GETRF_BATCH_SIZE_SWAP must be smaller than CHAMELEON_BATCH_SIZE, please recompile with the right CHAMELEON_BATCH_SIZE, or reduce the CHAMELEON_GETRF_BATCH_SIZE_SWAP value\n" );
+ ws->batch_size_swap = CHAMELEON_BATCH_SIZE;
}
ws->ringswitch = chameleon_getenv_get_value_int( "CHAMELEON_GETRF_RINGSWITCH", INT_MAX );
diff --git a/control/compute_z.h b/control/compute_z.h
index 855f820b1..1229a1797 100644
--- a/control/compute_z.h
+++ b/control/compute_z.h
@@ -46,7 +46,9 @@ struct chameleon_pzgetrf_s {
cham_getrf_t alg;
cham_getrf_allreduce_t alg_allreduce;
int ib; /**< Internal blocking parameter */
- int batch_size; /**< Batch size for the panel */
+ int batch_size_blas2; /**< Batch size for the blas 2 operations of the panel factorization */
+ int batch_size_blas3; /**< Batch size for the blas 3 operations of the panel factorization */
+ int batch_size_swap; /**< Batch size for the permutation */
int ringswitch; /**< Define when to switch to ring bcast */
CHAM_desc_t U;
CHAM_desc_t Up; /**< Workspace used for the panel factorization */
diff --git a/runtime/starpu/codelets/codelet_zgetrf_batched.c b/runtime/starpu/codelets/codelet_zgetrf_batched.c
index 2e04493df..011785aa2 100644
--- a/runtime/starpu/codelets/codelet_zgetrf_batched.c
+++ b/runtime/starpu/codelets/codelet_zgetrf_batched.c
@@ -74,7 +74,7 @@ INSERT_TASK_zgetrf_panel_offdiag_batched( const RUNTIME_option_t *options,
CHAM_ipiv_t *ipiv )
{
int task_num = 0;
- int batch_size = ((struct chameleon_pzgetrf_s *)ws)->batch_size;
+ int batch_size = ((struct chameleon_pzgetrf_s *)ws)->batch_size_blas2;
void (*callback)(void*) = NULL;
struct cl_getrf_batched_args_t *clargs = *clargs_ptr;
int rankA = A->get_rankof( A, Am, An );
@@ -241,8 +241,9 @@ INSERT_TASK_zgetrf_panel_blocked_batched( const RUNTIME_option_t *options,
void **clargs_ptr,
CHAM_ipiv_t *ipiv )
{
- int batch_size = ((struct chameleon_pzgetrf_s *)ws)->batch_size;
- int ib = ((struct chameleon_pzgetrf_s *)ws)->ib;
+ struct chameleon_pzgetrf_s *tmp = (struct chameleon_pzgetrf_s *) ws;
+ int ib = tmp->ib;
+ int batch_size = ( (h % ib) != 0 ) ? tmp->batch_size_blas2 : tmp->batch_size_blas3;
int task_num = 0;
void (*callback)(void*) = NULL;
int accessU, access_npiv, access_ipiv, access_ppiv;
diff --git a/runtime/starpu/codelets/codelet_zlaswp_batched.c b/runtime/starpu/codelets/codelet_zlaswp_batched.c
index b17f26a48..303e6a674 100644
--- a/runtime/starpu/codelets/codelet_zlaswp_batched.c
+++ b/runtime/starpu/codelets/codelet_zlaswp_batched.c
@@ -72,7 +72,7 @@ void INSERT_TASK_zlaswp_batched( const RUNTIME_option_t *options,
void **clargs_ptr )
{
int task_num = 0;
- int batch_size = ((struct chameleon_pzgetrf_s *)ws)->batch_size;
+ int batch_size = ((struct chameleon_pzgetrf_s *)ws)->batch_size_swap;
int nhandles;
struct cl_laswp_batched_args_t *clargs = *clargs_ptr;
if ( Am->get_rankof( Am, Amm, Amn) != Am->myrank ) {
--
GitLab
From 470e5c5fd3f1e6a68c1feb5079dbc4a0c3572113 Mon Sep 17 00:00:00 2001
From: Alycia Lisito <alycia.lisito@inria.fr>
Date: Tue, 7 Jan 2025 12:28:27 +0100
Subject: [PATCH 5/6] zgetrf: clean permutation
---
compute/pzgetrf.c | 160 ++++++++++++++++++++++++++--------------------
1 file changed, 89 insertions(+), 71 deletions(-)
diff --git a/compute/pzgetrf.c b/compute/pzgetrf.c
index 98f9d0470..7081a1f7f 100644
--- a/compute/pzgetrf.c
+++ b/compute/pzgetrf.c
@@ -338,10 +338,12 @@ chameleon_pzgetrf_panel_facto( struct chameleon_pzgetrf_s *ws,
int k,
RUNTIME_option_t *options )
{
+#if defined(CHAMELEON_USE_MPI)
chameleon_get_proc_involved_in_panelk_2dbc( A, k, k, ws );
if ( !ws->involved ) {
return;
}
+#endif
/* TODO: Should be replaced by a function pointer */
switch( ws->alg ) {
@@ -392,19 +394,6 @@ chameleon_pzgetrf_panel_permute( struct chameleon_pzgetrf_s *ws,
int tempkm, tempkn, tempnn, minmn;
int withlacpy;
- chameleon_get_proc_involved_in_panelk_2dbc( A, k, n, ws );
- if ( A->myrank == chameleon_getrankof_2d( A, k, k ) ) {
- INSERT_TASK_zperm_allreduce_send_perm( options, ipiv, k, A->myrank, ws->np_involved, ws->proc_involved );
- INSERT_TASK_zperm_allreduce_send_invp( options, ipiv, k, A, k, n );
- }
- if ( A->myrank == chameleon_getrankof_2d( A, k, n ) ) {
- INSERT_TASK_zperm_allreduce_send_A( options, A, k, n, A->myrank, ws->np_involved, ws->proc_involved );
- }
-
- if ( !ws->involved ) {
- return;
- }
-
tempkm = A->get_blkdim( A, k, DIM_m, A->m );
tempkn = A->get_blkdim( A, k, DIM_n, A->n );
tempnn = A->get_blkdim( A, n, DIM_n, A->n );
@@ -457,19 +446,6 @@ chameleon_pzgetrf_panel_permute_batched( struct chameleon_pzgetrf_s *ws,
int tempkm, tempkn, tempnn, minmn;
int withlacpy;
- chameleon_get_proc_involved_in_panelk_2dbc( A, k, n, ws );
- if ( A->myrank == chameleon_getrankof_2d( A, k, k ) ) {
- INSERT_TASK_zperm_allreduce_send_perm( options, ipiv, k, A->myrank, ws->np_involved, ws->proc_involved );
- INSERT_TASK_zperm_allreduce_send_invp( options, ipiv, k, A, k, n );
- }
- if ( A->myrank == chameleon_getrankof_2d( A, k, n ) ) {
- INSERT_TASK_zperm_allreduce_send_A( options, A, k, n, A->myrank, ws->np_involved, ws->proc_involved );
- }
-
- if ( !ws->involved ) {
- return;
- }
-
void **clargs = malloc( sizeof(char *) );
*clargs = NULL;
@@ -508,6 +484,80 @@ chameleon_pzgetrf_panel_permute_batched( struct chameleon_pzgetrf_s *ws,
}
}
+static inline void
+chameleon_pzgetrf_panel_permute_forward( struct chameleon_pzgetrf_s *ws,
+ CHAM_desc_t *A,
+ CHAM_ipiv_t *ipiv,
+ int k,
+ int n,
+ RUNTIME_option_t *options )
+{
+#if defined(CHAMELEON_USE_MPI)
+ chameleon_get_proc_involved_in_panelk_2dbc( A, k, n, ws );
+ if ( A->myrank == chameleon_getrankof_2d( A, k, k ) ) {
+ INSERT_TASK_zperm_allreduce_send_perm( options, ipiv, k, A->myrank, ws->np_involved, ws->proc_involved );
+ INSERT_TASK_zperm_allreduce_send_invp( options, ipiv, k, A, k, n );
+ }
+ if ( A->myrank == chameleon_getrankof_2d( A, k, n ) ) {
+ INSERT_TASK_zperm_allreduce_send_A( options, A, k, n, A->myrank, ws->np_involved, ws->proc_involved );
+ }
+
+ if ( !ws->involved ) {
+ return;
+ }
+#endif
+
+ if ( ws->batch_size_swap > 0 ) {
+ chameleon_pzgetrf_panel_permute_batched( ws, A, ipiv, k, n, options );
+ }
+ else {
+ chameleon_pzgetrf_panel_permute( ws, A, ipiv, k, n, options );
+ }
+}
+
+static inline void
+chameleon_pzgetrf_panel_permute_backward( struct chameleon_pzgetrf_s *ws,
+ CHAM_desc_t *A,
+ CHAM_ipiv_t *ipiv,
+ int k,
+ int n,
+ RUNTIME_option_t *options,
+ RUNTIME_sequence_t *sequence )
+{
+ int tempkm, tempnn;
+
+#if defined(CHAMELEON_USE_MPI)
+ chameleon_get_proc_involved_in_panelk_2dbc( A, k, n, ws );
+ if ( A->myrank == chameleon_getrankof_2d( A, k, k ) ) {
+ INSERT_TASK_zperm_allreduce_send_perm( options, ipiv, k, A->myrank, ws->np_involved, ws->proc_involved );
+ INSERT_TASK_zperm_allreduce_send_invp( options, ipiv, k, A, k, n );
+ }
+ if ( A->myrank == chameleon_getrankof_2d( A, k, n ) ) {
+ INSERT_TASK_zperm_allreduce_send_A( options, A, k, n, A->myrank, ws->np_involved, ws->proc_involved );
+ }
+
+ if ( !ws->involved ) {
+ return;
+ }
+#endif
+
+ if ( ws->batch_size_swap > 0 ) {
+ chameleon_pzgetrf_panel_permute_batched( ws, A, ipiv, k, n, options );
+ }
+ else {
+ chameleon_pzgetrf_panel_permute( ws, A, ipiv, k, n, options );
+ }
+
+ if ( A->myrank == chameleon_getrankof_2d( A, k, n ) ) {
+
+ tempkm = A->get_blkdim( A, k, DIM_m, A->m );
+ tempnn = A->get_blkdim( A, n, DIM_n, A->n );
+ INSERT_TASK_zlacpy( options, ChamUpperLower, tempkm, tempnn,
+ Wu(A->myrank, n), A(k, n) );
+ RUNTIME_data_flush( sequence, A(k, n) );
+ }
+}
+
static inline void
chameleon_pzgetrf_panel_update_ws( struct chameleon_pzgetrf_s *ws,
CHAM_desc_t *A,
@@ -515,7 +565,7 @@ chameleon_pzgetrf_panel_update_ws( struct chameleon_pzgetrf_s *ws,
RUNTIME_option_t *options )
{
CHAM_context_t *chamctxt = chameleon_context_self();
- int m, tempmm, tempkn, q;
+ int m, n, tempmm, tempkn, q;
int lookahead = chamctxt->lookahead;
int P = chameleon_desc_datadist_get_iparam(A, 0);
int Q = chameleon_desc_datadist_get_iparam(A, 1);
@@ -583,12 +633,7 @@ chameleon_pzgetrf_panel_update( struct chameleon_pzgetrf_s *ws,
tempkm = A->get_blkdim( A, k, DIM_m, A->m );
tempnn = A->get_blkdim( A, n, DIM_n, A->n );
- if ( ws->batch_size_swap > 0 ) {
- chameleon_pzgetrf_panel_permute_batched( ws, A, ipiv, k, n, options );
- }
- else {
- chameleon_pzgetrf_panel_permute( ws, A, ipiv, k, n, options );
- }
+ chameleon_pzgetrf_panel_permute_forward( ws, A, ipiv, k, n, options );
if ( A->myrank == chameleon_getrankof_2d( A, k, k ) ) {
for ( p = 0; p < ws->np_involved; p++ ) {
@@ -607,6 +652,7 @@ chameleon_pzgetrf_panel_update( struct chameleon_pzgetrf_s *ws,
ChamLeft, ChamLower, ChamNoTrans, ChamUnit,
tempkm, tempnn, A->mb,
zone, A(k, k),
+ zone, Wu(A->myrank, k),
Wu(A->myrank, n) );
}
@@ -682,54 +728,26 @@ void chameleon_pzgetrf( struct chameleon_pzgetrf_s *ws,
}
/* Flush panel k */
- for (m = k; m < A->mt; m++) {
+ for (m = k+1; m < A->mt; m++) {
RUNTIME_data_flush( sequence, A(m, k) );
}
+ RUNTIME_data_flush( sequence, Wu(A->myrank, k) );
RUNTIME_iteration_pop( chamctxt );
}
CHAMELEON_Desc_Flush( &(ws->Wl), sequence );
/* Backward pivoting */
- if ( ws->batch_size > 0 ) {
- for (k = 1; k < min_mnt; k++) {
- for (n = 0; n < k; n++) {
- if ( chameleon_involved_in_panelk_2dbc( A, k ) ||
- chameleon_involved_in_panelk_2dbc( A, n ) )
- {
- chameleon_pzgetrf_panel_permute_batched( ws, A, IPIV, k, n, &options );
- if ( A->myrank == chameleon_getrankof_2d( A, k, n ) ) {
- tempkm = A->get_blkdim( A, k, DIM_m, A->m );
- tempnn = A->get_blkdim( A, n, DIM_n, A->n );
- INSERT_TASK_zlacpy( &options, ChamUpperLower, tempkm, tempnn,
- Wu(A->myrank, n), A(k, n) );
- RUNTIME_data_flush( sequence, A(k, n) );
- }
- }
- RUNTIME_data_flush( sequence, Wu(A->myrank, n) );
- }
- RUNTIME_perm_flushk( sequence, IPIV, k );
- }
- }
- else {
- for (k = 1; k < min_mnt; k++) {
- for (n = 0; n < k; n++) {
- if ( chameleon_involved_in_panelk_2dbc( A, k ) ||
- chameleon_involved_in_panelk_2dbc( A, n ) )
- {
- chameleon_pzgetrf_panel_permute( ws, A, IPIV, k, n, &options );
- if ( A->myrank == chameleon_getrankof_2d( A, k, n ) ) {
- tempkm = A->get_blkdim( A, k, DIM_m, A->m );
- tempnn = A->get_blkdim( A, n, DIM_n, A->n );
- INSERT_TASK_zlacpy( &options, ChamUpperLower, tempkm, tempnn,
- Wu(A->myrank, n), A(k, n) );
- RUNTIME_data_flush( sequence, A(k, n) );
- }
- }
- RUNTIME_data_flush( sequence, Wu(A->myrank, n) );
+ for (k = 1; k < min_mnt; k++) {
+ for (n = 0; n < k; n++) {
+ if ( chameleon_involved_in_panelk_2dbc( A, k ) ||
+ chameleon_involved_in_panelk_2dbc( A, n ) )
+ {
+ chameleon_pzgetrf_panel_permute_backward( ws, A, IPIV, k, n, &options, sequence );
}
- RUNTIME_perm_flushk( sequence, IPIV, k );
+ RUNTIME_data_flush( sequence, Wu(A->myrank, n) );
}
+ RUNTIME_perm_flushk( sequence, IPIV, k );
}
CHAMELEON_Desc_Flush( &(ws->Wu), sequence );
--
GitLab
From 0e64bc8b2455a84da029f732e170c12d66a05e8a Mon Sep 17 00:00:00 2001
From: Alycia Lisito <alycia.lisito@inria.fr>
Date: Thu, 13 Feb 2025 17:43:55 +0100
Subject: [PATCH 6/6] zgetrf: send Akk before allreduce
---
compute/pzgetrf.c | 58 +++++++++++++++++++-------
control/descriptor_helpers.c | 20 +++++++++
include/chameleon/descriptor_helpers.h | 1 +
3 files changed, 64 insertions(+), 15 deletions(-)
diff --git a/compute/pzgetrf.c b/compute/pzgetrf.c
index 7081a1f7f..635bbbb84 100644
--- a/compute/pzgetrf.c
+++ b/compute/pzgetrf.c
@@ -565,7 +565,7 @@ chameleon_pzgetrf_panel_update_ws( struct chameleon_pzgetrf_s *ws,
RUNTIME_option_t *options )
{
CHAM_context_t *chamctxt = chameleon_context_self();
- int m, n, tempmm, tempkn, q;
+ int m, n, tempmm, tempkn, tempkm, p, q, involved, np;
int lookahead = chamctxt->lookahead;
int P = chameleon_desc_datadist_get_iparam(A, 0);
int Q = chameleon_desc_datadist_get_iparam(A, 1);
@@ -610,6 +610,44 @@ chameleon_pzgetrf_panel_update_ws( struct chameleon_pzgetrf_s *ws,
RUNTIME_data_flush( options->sequence, A(m, k) );
}
}
+
+ tempkm = A->get_blkdim( A, k, DIM_m, A->m );
+ np = chameleon_desc_datadist_get_iparam(A, 1) * chameleon_desc_datadist_get_iparam(A, 0);
+#if defined(CHAMELEON_USE_MPI)
+ /* Send Akk for replicated trsm */
+ if ( A->myrank == chameleon_getrankof_2d( A, k, k ) ) {
+ for ( p = 0; p < np; p++ ) {
+ involved = 0;
+ for ( n = k+1; n < A->nt; n++ ) {
+ if ( chameleon_p_involved_in_panelk_2dbc( A, n, p ) ) {
+ involved = 1;
+ break;
+ }
+ }
+ if ( involved ) {
+ INSERT_TASK_zlacpy( options, ChamUpperLower, tempkm, tempkn,
+ A(k, k), Wu(p, k) );
+ }
+ }
+ }
+ else {
+ involved = 0;
+ for ( n = k+1; n < A->nt; n++ ) {
+ if ( chameleon_involved_in_panelk_2dbc( A, n ) ) {
+ involved = 1;
+ break;
+ }
+ }
+ if ( involved ) {
+ INSERT_TASK_zlacpy( options, ChamUpperLower, tempkm, tempkn,
+ A(k, k), Wu(A->myrank, k) );
+ }
+ }
+#else
+ INSERT_TASK_zlacpy( options, ChamUpperLower, tempkm, tempkn,
+ A(k, k), Wu(A->myrank, k) );
+#endif
+ RUNTIME_data_flush( options->sequence, A(k, k) );
}
static inline void
@@ -635,23 +673,14 @@ chameleon_pzgetrf_panel_update( struct chameleon_pzgetrf_s *ws,
chameleon_pzgetrf_panel_permute_forward( ws, A, ipiv, k, n, options );
- if ( A->myrank == chameleon_getrankof_2d( A, k, k ) ) {
- for ( p = 0; p < ws->np_involved; p++ ) {
- INSERT_TASK_ztrsm(
- options,
- ChamLeft, ChamLower, ChamNoTrans, ChamUnit,
- tempkm, tempnn, A->mb,
- zone, A(k, k),
- Wu(ws->proc_involved[p], n) );
- RUNTIME_data_flush( options->sequence, Wu(ws->proc_involved[p], n) );
- }
- }
- else if ( ws->involved ) {
+#if defined(CHAMELEON_USE_MPI)
+ if ( ws->involved )
+#endif
+ {
INSERT_TASK_ztrsm(
options,
ChamLeft, ChamLower, ChamNoTrans, ChamUnit,
tempkm, tempnn, A->mb,
- zone, A(k, k),
zone, Wu(A->myrank, k),
Wu(A->myrank, n) );
}
@@ -677,7 +706,6 @@ chameleon_pzgetrf_panel_update( struct chameleon_pzgetrf_s *ws,
}
RUNTIME_data_flush( options->sequence, Wu(A->myrank, n) );
- RUNTIME_data_flush( options->sequence, A(k, k) );
RUNTIME_data_flush( options->sequence, A(k, n) );
}
diff --git a/control/descriptor_helpers.c b/control/descriptor_helpers.c
index d5e143063..6a0492111 100644
--- a/control/descriptor_helpers.c
+++ b/control/descriptor_helpers.c
@@ -100,6 +100,26 @@ int chameleon_involved_in_panelk_2dbc( const CHAM_desc_t *A, int k ) {
return ( myrank % chameleon_desc_datadist_get_iparam(A,1) == k % chameleon_desc_datadist_get_iparam(A,1) );
}
+/**
+ * @brief Test if the MPI process p is involved in the panel k for 2DBC distributions.
+ *
+ * @param[in] A
+ * The matrix descriptor.
+ *
+ * @param[in] k
+ * The index of the panel to test.
+ *
+ * @param[in] p
+ * The rank of the MPI process.
+ *
+ * @return 1 if the current MPI process contributes to the panel k.
+ * 0 if the current MPI process doesn't contribute to the panel k.
+ *
+ */
+int chameleon_p_involved_in_panelk_2dbc( const CHAM_desc_t *A, int k, int p ) {
+ return ( p % chameleon_desc_datadist_get_iparam(A,1) == k % chameleon_desc_datadist_get_iparam(A,1) );
+}
+
/**
* @brief Test if the current MPI process is involved in the panel k for 2DBC distributions.
*
diff --git a/include/chameleon/descriptor_helpers.h b/include/chameleon/descriptor_helpers.h
index 9e60ef27d..f8caf5080 100644
--- a/include/chameleon/descriptor_helpers.h
+++ b/include/chameleon/descriptor_helpers.h
@@ -64,6 +64,7 @@ int chameleon_getrankof_custom ( const CHAM_desc_t *A, int m, int n );
*/
int chameleon_involved_in_panelk_2dbc( const CHAM_desc_t *A, int An );
+int chameleon_p_involved_in_panelk_2dbc( const CHAM_desc_t *A, int k, int p );
void chameleon_get_proc_involved_in_panelk_2dbc( const CHAM_desc_t *A,
int k,
int n,
--
GitLab