From 0cc694735ddcd0645e4c4a2c23dd2ee3a5d2934d Mon Sep 17 00:00:00 2001
From: Mathieu Faverge <mathieu.faverge@inria.fr>
Date: Tue, 12 Mar 2024 14:40:46 +0100
Subject: [PATCH] parsec/map: Fix the map codelet to always use a tile
interface and do not enforce the user to check for the runtime used in the
map functions.
---
compute/pzlatms.c | 28 +++-------
compute/zprint.c | 26 ++-------
runtime/parsec/codelets/codelet_map.c | 78 ++++++++++++++++++++++-----
3 files changed, 75 insertions(+), 57 deletions(-)
diff --git a/compute/pzlatms.c b/compute/pzlatms.c
index 8e6ba0f5f..e8daf897e 100644
--- a/compute/pzlatms.c
+++ b/compute/pzlatms.c
@@ -13,7 +13,7 @@
* @version 1.3.0
* @author Mathieu Faverge
* @author Lionel Eyraud-Dubois
- * @date 2024-03-11
+ * @date 2024-03-14
* @precisions normal z -> s d c
*
*/
@@ -26,39 +26,26 @@
#define A(m, n) A, m, n
-/*
- * Static variable to know how to handle the data within the kernel
- * This assumes that only one runtime is enabled at a time.
- */
-static RUNTIME_id_t zlatms_runtime_id = RUNTIME_SCHED_STARPU;
-
static inline int
zlaset_diag_cpu( void *op_args,
cham_uplo_t uplo, int m, int n, int ndata,
const CHAM_desc_t *descA, CHAM_tile_t *tileA, ... )
{
- CHAMELEON_Complex64_t *A;
- const double *D = (const double *)op_args;
+ const double *D = (const double *)op_args;
+ CHAMELEON_Complex64_t *A = CHAM_tile_get_ptr( tileA );
int tempmm = m == descA->mt-1 ? descA->m-m*descA->mb : descA->mb;
int tempnn = n == descA->nt-1 ? descA->n-n*descA->nb : descA->nb;
- int minmn = chameleon_min( tempmm, tempnn );
- int lda, i;
+ int minmn = chameleon_min( tempmm, tempnn );
+ int lda = tileA->ld;
+ int i;
if ( ndata > 1 ) {
fprintf( stderr, "zlaset_diag_cpu: supports only one piece of data and %d have been given\n", ndata );
}
- if ( zlatms_runtime_id == RUNTIME_SCHED_PARSEC ) {
- A = (CHAMELEON_Complex64_t*)tileA;
- lda = descA->get_blkldd( descA, m );
- }
- else {
- A = tileA->mat;
- lda = tileA->ld;
- }
-
assert( m == n );
+ assert( tileA->format & CHAMELEON_TILE_FULLRANK );
/* Shift to the values corresponding to the tile */
D += m * descA->mb;
@@ -104,7 +91,6 @@ void chameleon_pzlatms( cham_dist_t idist, unsigned long long int seed, cham_sym
return;
}
ib = CHAMELEON_IB;
- zlatms_runtime_id = chamctxt->scheduler;
RUNTIME_options_init(&options, chamctxt, sequence, request);
diff --git a/compute/zprint.c b/compute/zprint.c
index cc2be4076..e17a406a1 100644
--- a/compute/zprint.c
+++ b/compute/zprint.c
@@ -12,7 +12,7 @@
* @version 1.3.0
* @author Mathieu Faverge
* @author Matthieu Kuhn
- * @date 2024-03-11
+ * @date 2024-03-14
* @precisions normal z -> s d c
*
*/
@@ -21,12 +21,6 @@
#include <coreblas/coreblas_z.h>
#endif
-/*
- * Static variable to know how to handle the data within the kernel
- * This assumes that only one runtime is enabled at a time.
- */
-static RUNTIME_id_t zprint_runtime_id = RUNTIME_SCHED_STARPU;
-
struct zprint_args_s {
FILE *file;
const char *header;
@@ -37,25 +31,17 @@ zprint_cpu( void *op_args,
cham_uplo_t uplo, int m, int n, int ndata,
const CHAM_desc_t *descA, CHAM_tile_t *tileA, ... )
{
- CHAMELEON_Complex64_t *A;
struct zprint_args_s *options = (struct zprint_args_s *)op_args;
+ CHAMELEON_Complex64_t *A = CHAM_tile_get_ptr( tileA );
int tempmm = m == descA->mt-1 ? descA->m-m*descA->mb : descA->mb;
int tempnn = n == descA->nt-1 ? descA->n-n*descA->nb : descA->nb;
- int lda;
+ int lda = tileA->ld;
if ( ndata > 1 ) {
fprintf( stderr, "zprint_cpu: supports only one piece of data and %d have been given\n", ndata );
}
-
- if ( zprint_runtime_id == RUNTIME_SCHED_PARSEC ) {
- A = (CHAMELEON_Complex64_t*)tileA;
- lda = descA->get_blkldd( descA, m );
- }
- else {
- A = CHAM_tile_get_ptr( tileA );
- lda = tileA->ld;
- }
+ assert( tileA->format & CHAMELEON_TILE_FULLRANK );
#if !defined(CHAMELEON_SIMULATION)
CORE_zprint( options->file, options->header, uplo,
@@ -162,8 +148,6 @@ int CHAMELEON_zprint( FILE *file, const char *header,
A, NB, NB, LDA, N, M, N, sequence, &request );
/* Call the tile interface */
- zprint_runtime_id = chamctxt->scheduler;
-
data.access = ChamR;
data.desc = &descAt;
chameleon_pmap( uplo, 1, &data, &zprint_map, &options, sequence, &request );
@@ -230,8 +214,6 @@ int CHAMELEON_zprint_Tile( FILE *file, const char *header,
}
chameleon_sequence_create( chamctxt, &sequence );
- zprint_runtime_id = chamctxt->scheduler;
-
data.access = ChamR;
data.desc = A;
diff --git a/runtime/parsec/codelets/codelet_map.c b/runtime/parsec/codelets/codelet_map.c
index 12a701bf3..d14ee9996 100644
--- a/runtime/parsec/codelets/codelet_map.c
+++ b/runtime/parsec/codelets/codelet_map.c
@@ -11,7 +11,7 @@
*
* @version 1.3.0
* @author Mathieu Faverge
- * @date 2024-03-11
+ * @date 2024-03-14
*
*/
#include "chameleon_parsec.h"
@@ -30,11 +30,21 @@ CORE_map_one_parsec( parsec_execution_stream_t *context,
parsec_task_t *this_task )
{
struct parsec_map_args_s *pargs = NULL;
- CHAM_tile_t *tileA;
+ const CHAM_desc_t *descA;
+ CHAM_tile_t tileA;
+
+ parsec_dtd_unpack_args( this_task, &pargs, &(tileA.mat) );
+
+ descA = pargs->desc[0];
+ tileA.rank = 0;
+ tileA.m = (pargs->m == (descA->mt-1)) ? (descA->m - pargs->m * descA->mb) : descA->mb;
+ tileA.n = (pargs->n == (descA->nt-1)) ? (descA->n - pargs->n * descA->nb) : descA->nb;
+ tileA.ld = descA->get_blkldd( descA, pargs->m );
+ tileA.format = CHAMELEON_TILE_FULLRANK;
+ tileA.flttype = descA->dtyp;
- parsec_dtd_unpack_args( this_task, &pargs, &tileA );
pargs->op_fcts->cpufunc( pargs->op_args, pargs->uplo, pargs->m, pargs->n, 1,
- pargs->desc[0], tileA );
+ descA, &tileA );
free( pargs );
}
@@ -44,12 +54,29 @@ CORE_map_two_parsec( parsec_execution_stream_t *context,
parsec_task_t *this_task )
{
struct parsec_map_args_s *pargs = NULL;
- CHAM_tile_t *tileA;
- CHAM_tile_t *tileB;
+ const CHAM_desc_t *descA, *descB;
+ CHAM_tile_t tileA, tileB;
+
+ parsec_dtd_unpack_args( this_task, &pargs, &(tileA.mat), &(tileB.mat) );
+
+ descA = pargs->desc[0];
+ tileA.rank = 0;
+ tileA.m = (pargs->m == (descA->mt-1)) ? (descA->m - pargs->m * descA->mb) : descA->mb;
+ tileA.n = (pargs->n == (descA->nt-1)) ? (descA->n - pargs->n * descA->nb) : descA->nb;
+ tileA.ld = descA->get_blkldd( descA, pargs->m );
+ tileA.format = CHAMELEON_TILE_FULLRANK;
+ tileA.flttype = descA->dtyp;
+
+ descB = pargs->desc[1];
+ tileB.rank = 0;
+ tileB.m = (pargs->m == (descB->mt-1)) ? (descB->m - pargs->m * descB->mb) : descB->mb;
+ tileB.n = (pargs->n == (descB->nt-1)) ? (descB->n - pargs->n * descB->nb) : descB->nb;
+ tileB.ld = descB->get_blkldd( descB, pargs->m );
+ tileB.format = CHAMELEON_TILE_FULLRANK;
+ tileB.flttype = descB->dtyp;
- parsec_dtd_unpack_args( this_task, &pargs, &tileA, &tileB );
pargs->op_fcts->cpufunc( pargs->op_args, pargs->uplo, pargs->m, pargs->n, 2,
- pargs->desc[0], tileA, pargs->desc[1], tileB );
+ descA, &tileA, descB, &tileB );
free( pargs );
}
@@ -59,14 +86,37 @@ CORE_map_three_parsec( parsec_execution_stream_t *context,
parsec_task_t *this_task )
{
struct parsec_map_args_s *pargs = NULL;
- CHAM_tile_t *tileA;
- CHAM_tile_t *tileB;
- CHAM_tile_t *tileC;
+ const CHAM_desc_t *descA, *descB, *descC;
+ CHAM_tile_t tileA, tileB, tileC;
+
+ parsec_dtd_unpack_args( this_task, &pargs, &(tileA.mat), &(tileB.mat), &(tileC.mat) );
+
+ descA = pargs->desc[0];
+ tileA.rank = 0;
+ tileA.m = (pargs->m == (descA->mt-1)) ? (descA->m - pargs->m * descA->mb) : descA->mb;
+ tileA.n = (pargs->n == (descA->nt-1)) ? (descA->n - pargs->n * descA->nb) : descA->nb;
+ tileA.ld = descA->get_blkldd( descA, pargs->m );
+ tileA.format = CHAMELEON_TILE_FULLRANK;
+ tileA.flttype = descA->dtyp;
+
+ descB = pargs->desc[1];
+ tileB.rank = 0;
+ tileB.m = (pargs->m == (descB->mt-1)) ? (descB->m - pargs->m * descB->mb) : descB->mb;
+ tileB.n = (pargs->n == (descB->nt-1)) ? (descB->n - pargs->n * descB->nb) : descB->nb;
+ tileB.ld = descB->get_blkldd( descB, pargs->m );
+ tileB.format = CHAMELEON_TILE_FULLRANK;
+ tileB.flttype = descB->dtyp;
+
+ descC = pargs->desc[2];
+ tileC.rank = 0;
+ tileC.m = (pargs->m == (descC->mt-1)) ? (descC->m - pargs->m * descC->mb) : descC->mb;
+ tileC.n = (pargs->n == (descC->nt-1)) ? (descC->n - pargs->n * descC->nb) : descC->nb;
+ tileC.ld = descC->get_blkldd( descC, pargs->m );
+ tileC.format = CHAMELEON_TILE_FULLRANK;
+ tileC.flttype = descC->dtyp;
- parsec_dtd_unpack_args( this_task, &pargs, &tileA, &tileB, &tileC );
pargs->op_fcts->cpufunc( pargs->op_args, pargs->uplo, pargs->m, pargs->n, 3,
- pargs->desc[0], tileA, pargs->desc[1], tileB,
- pargs->desc[2], tileC );
+ descA, &tileA, descB, &tileB, descC, &tileC );
free( pargs );
}
--
GitLab