diff --git a/compute/pzgetrf.c b/compute/pzgetrf.c
index 6db9a8a40148a67fcaeda74f9a718c949b12e59d..4347f5710f23d5ff450f9eb77bf7f7f5c9f41205 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 5f1bbcd322293e3104fca313974096bcf711de71..402c92a3f6d9dcb4ff8e1039b64b62ef051a4a7c 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 b088283254cd64e1bada1628939436327b8a2789..197842ea3e96fdba1a9e1d67152a8a5b3e6196ea 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 75e0611647a464cad9c37e59a5619ebefaae19ed..d6bd3f4c06baf9b1c44e4db6971c88c09acd432f 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 e88269e931f3f210282a1382d44a6ff9516c7453..0186fd142b67d08dcfca01e9b8184b471362ce1c 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 a81f0d08eef1fb94b6846606b5e63aae64ab075c..48ecdd0c33fa07f9cfd326f775b0f31fb48a67b8 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