diff --git a/compute/mapv.c b/compute/mapv.c
index a27cd2d49842a671ec83d7ecf458d231294e4b19..c492aa621921e2a35c983bd0aa80082adbd07287 100644
--- a/compute/mapv.c
+++ b/compute/mapv.c
@@ -11,7 +11,7 @@
*
* @version 1.3.0
* @author Mathieu Faverge
- * @date 2024-03-11
+ * @date 2024-03-14
*
*/
#include "control/common.h"
@@ -21,35 +21,34 @@
*
* @ingroup CHAMELEON_Tile
*
- * Apply a given operator on each tile of the given matrix. Operates on
- * matrices stored by tiles. All matrices are passed through descriptors. All
- * dimensions are taken from the descriptors.
+ * Apply a given operator f on each tile of the given matrices.
+ * For each (m,n) \in [1,MT]x[1xNT], apply f( data[0](m, n), data[1](m, n), ... )
*
*******************************************************************************
*
- * @param[in] access
- * - ChamR: A is accessed in read-only mode.
- * - ChamW: A is accessed in write-only mode.
- * WARNING: if the descriptor is set for allocation on the fly, the
- * flush call included in this synchronous API will free all allocated
- * data, prefer asynchronous call if you want to initialiaze data
- * before submitting another algorithm.
- * - ChamRW: A is accessed in read-write mode.
- *
* @param[in] uplo
- * - ChamUpper: Only the upper triangular part of the matrix is touched
- * - ChamLower: Only the lower triangular part of the matrix is touched
- * - ChamUpperLower: The entire the matrix is touched
+ * - ChamUpper: only the upper part of the matrices are referenced.
+ * - ChamLower: only the lower part of the matrices are referenced.
+ * - ChamUpperLower: the full matrices are references.
+ *
+ * @param[in] ndata
+ * - the numbe of matrices given to the map function.
+ * - ChamLower: only the lower part of the matrices are referenced.
+ * - ChamUpperLower: the full matrices are references.
*
- * @param[in,out] A
- * On exit, the operator has been applied on each tile of the matrix A.
+ * @param[in] data
+ * Array of size ndata that contains the ndata couples { access,
+ * descriptor } used in the operator.
*
* @param[in] op_fct
- * The operator function to apply on each tile of the matrix.
+ * The operator function to apply on each tile of the matrices. Must
+ * support the number of data ndata given as parameter.
*
* @param[in,out] op_args
* The arguments structure passed to the operator function when applied
- * on each tile. May be updated by the operator function.
+ * on each tile. May be updated by the operator function. If concurrent
+ * accesses must be protected, it is let to the user to do it in the
+ * op_fct.
*
*******************************************************************************
*
@@ -93,20 +92,38 @@ int CHAMELEON_mapv_Tile( cham_uplo_t uplo,
/**
********************************************************************************
*
- * @ingroup CHAMELEON_Tile_Async
+ * @ingroup CHAMELEON_Tile
*
- * Apply a given operator on each tile of the given matrix. Non-blocking equivalent of
- * CHAMELEON_mapv_Tile(). May return before the computation is finished.
- * Allows for pipelining of operations at runtime.
+ * Apply a given operator f on each tile of the given matrices.
+ * For each (m,n) \in [1,MT]x[1xNT], apply f( data[0](m, n), data[1](m, n), ... )
+ * Non-blocking equivalent of CHAMELEON_map_Tile(). May return before the
+ * computation is finished. Allows for pipelining of operations at runtime.
*
*******************************************************************************
*
- * @param[in] access
- * - ChamR: A is accessed in read-only mode.
- * - ChamW: A is accessed in write-only mode.
- * INFO: tile of A can be unallocated before the call if the
- * descriptor is set for allocation on the fly.
- * - ChamRW: A is accessed in read-write mode.
+ * @param[in] uplo
+ * - ChamUpper: only the upper part of the matrices are referenced.
+ * - ChamLower: only the lower part of the matrices are referenced.
+ * - ChamUpperLower: the full matrices are references.
+ *
+ * @param[in] ndata
+ * - the numbe of matrices given to the map function.
+ * - ChamLower: only the lower part of the matrices are referenced.
+ * - ChamUpperLower: the full matrices are references.
+ *
+ * @param[in] data
+ * Array of size ndata that contains the ndata couples { access,
+ * descriptor } used in the operator.
+ *
+ * @param[in] op_fct
+ * The operator function to apply on each tile of the matrices. Must
+ * support the number of data ndata given as parameter.
+ *
+ * @param[in,out] op_args
+ * The arguments structure passed to the operator function when applied
+ * on each tile. May be updated by the operator function. If concurrent
+ * accesses must be protected, it is let to the user to do it in the
+ * op_fct.
*
* @param[in] sequence
* Identifies the sequence of function calls that this call belongs to
diff --git a/compute/pmap.c b/compute/pmap.c
index bcc1a868e82fb6c7defec63d90ffcbdbed311e58..ef30f09de23cf92e71d20b6ae86ac69cc83da819 100644
--- a/compute/pmap.c
+++ b/compute/pmap.c
@@ -11,7 +11,7 @@
*
* @version 1.3.0
* @author Mathieu Faverge
- * @date 2024-03-11
+ * @date 2024-03-14
*
*/
#include "control/common.h"
@@ -28,31 +28,34 @@ void chameleon_pmap( cham_uplo_t uplo, int ndata, cham_map_data_t *data,
CHAM_context_t *chamctxt;
RUNTIME_option_t options;
const CHAM_desc_t *A = data[0].desc;
- int m, n;
+ int m, n, minmn;
chamctxt = chameleon_context_self();
if (sequence->status != CHAMELEON_SUCCESS)
return;
RUNTIME_options_init( &options, chamctxt, sequence, request );
+ minmn = chameleon_min( A->mt, A->nt );
+
switch( uplo ) {
case ChamUpper:
- for (n = 0; n < A->nt; n++) {
- for (m = 0; m < n; m++) {
+ for (m = 0; m < minmn; m++) {
+ INSERT_TASK_map(
+ &options, uplo, m, m,
+ ndata, data,
+ op_fct, op_args );
+
+ for (n = m+1; n < A->nt; n++) {
INSERT_TASK_map(
&options, ChamUpperLower, m, n,
ndata, data,
op_fct, op_args );
}
- INSERT_TASK_map(
- &options, uplo, n, n,
- ndata, data,
- op_fct, op_args );
}
break;
case ChamLower:
- for (n = 0; n < A->nt; n++) {
+ for (n = 0; n < minmn; n++){
INSERT_TASK_map(
&options, uplo, n, n,
ndata, data,
diff --git a/compute/pzlatms.c b/compute/pzlatms.c
index 8e6ba0f5fc1b85a612ff835b351332d3fe831c44..e8daf897e08937862fe4a08563ab050209984794 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/zbuild.c b/compute/zbuild.c
index 6a56eaadbe738e78a4498d7344b5cf1bc67f502e..f36980ecc536d4cd24614b168d7b04c91651a053 100644
--- a/compute/zbuild.c
+++ b/compute/zbuild.c
@@ -128,7 +128,7 @@ int CHAMELEON_zbuild( cham_uplo_t uplo, int M, int N,
A, NB, NB, LDA, N, M, N, sequence, &request );
/* Call the tile interface */
- CHAMELEON_zbuild_Tile_Async( uplo, &descAt, user_data, user_build_callback, sequence, &request );
+ chameleon_pzbuild( uplo, &descAt, user_data, user_build_callback, sequence, &request );
/* Submit the matrix conversion back */
chameleon_ztile2lap( chamctxt, &descAl, &descAt,
@@ -201,7 +201,7 @@ int CHAMELEON_zbuild_Tile( cham_uplo_t uplo, CHAM_desc_t *A,
}
chameleon_sequence_create( chamctxt, &sequence );
- CHAMELEON_zbuild_Tile_Async( uplo, A, user_data, user_build_callback, sequence, &request );
+ chameleon_pzbuild( uplo, A, user_data, user_build_callback, sequence, &request );
CHAMELEON_Desc_Flush( A, sequence );
diff --git a/compute/zprint.c b/compute/zprint.c
index cc2be40769abf806dd9012e92b705fb297ab9ca7..e17a406a1b5adcdf80ff90e2d49a87ec62c104eb 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/coreblas/include/coreblas.h b/coreblas/include/coreblas.h
index ee6cda933cc791aa70bf7308c21c0262c2cfbb38..cc889a9ee28a46915d4deb115407d568c2616cba 100644
--- a/coreblas/include/coreblas.h
+++ b/coreblas/include/coreblas.h
@@ -18,7 +18,7 @@
* @author Guillaume Sylvand
* @author Mathieu Faverge
* @author Raphael Boucherie
- * @date 2023-08-31
+ * @date 2024-03-14
*
*/
#ifndef _coreblas_h_
@@ -51,6 +51,13 @@ BEGIN_C_DECLS
#include "coreblas/coreblas_zc.h"
#include "coreblas/coreblas_ds.h"
+#include "coreblas/coreblas_ztile.h"
+#include "coreblas/coreblas_dtile.h"
+#include "coreblas/coreblas_ctile.h"
+#include "coreblas/coreblas_stile.h"
+#include "coreblas/coreblas_zctile.h"
+#include "coreblas/coreblas_dstile.h"
+
END_C_DECLS
/**
diff --git a/example/lapack_to_chameleon/step7.c b/example/lapack_to_chameleon/step7.c
index 7390979f401662f70f45d0c9eb7dcfae5bd459a7..8802ee7e74dfcdd147e67e47070b0f253fa8981e 100644
--- a/example/lapack_to_chameleon/step7.c
+++ b/example/lapack_to_chameleon/step7.c
@@ -11,12 +11,12 @@
*
* @brief Chameleon step7 example
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Florent Pruvost
* @author Guillaume Sylvand
* @author Mathieu Faverge
* @author Philippe Virouleau
- * @date 2022-02-22
+ * @date 2024-03-14
*
*/
#include "step7.h"
@@ -135,12 +135,32 @@ int main(int argc, char *argv[]) {
/* generate A matrix with random values such that it is spd.
We use the callback function Cham_build_callback_plgsy() defined in step7.h
In this example, it is just a wrapper toward CORE_dplgsy() */
- struct data_pl data_A={(double)N, 51, N};
- CHAMELEON_dbuild_Tile(ChamUpperLower, descA, (void*)&data_A, Cham_build_callback_plgsy);
+ struct data_pl plgsy_args = { (double)N, 51 };
+ struct cham_map_operator_s plgsy_op = {
+ .name = "plgsy",
+ .cpufunc = Cham_build_plgsy_cpu,
+ .cudafunc = NULL,
+ .hipfunc = NULL,
+ };
+ struct cham_map_data_s plgsy_data = {
+ .access = ChamW,
+ .desc = descA,
+ };
+ CHAMELEON_mapv_Tile( ChamUpperLower, 1, &plgsy_data, &plgsy_op, &plgsy_args );
/* generate RHS with the callback Cham_build_callback_plrnt() */
- struct data_pl data_B={0., 5673, N};
- CHAMELEON_dbuild_Tile(ChamUpperLower, descB, (void*)&data_B, Cham_build_callback_plrnt);
+ struct data_pl plrnt_args = { 0., 5673 };
+ struct cham_map_operator_s plrnt_op = {
+ .name = "plrnt",
+ .cpufunc = Cham_build_plrnt_cpu,
+ .cudafunc = NULL,
+ .hipfunc = NULL,
+ };
+ struct cham_map_data_s plrnt_data = {
+ .access = ChamW,
+ .desc = descA,
+ };
+ CHAMELEON_mapv_Tile( ChamUpperLower, 1, &plrnt_data, &plrnt_op, &plrnt_args );
/* copy A before facto. in order to check the result */
CHAMELEON_dlacpy_Tile(ChamUpperLower, descA, descAC);
diff --git a/example/lapack_to_chameleon/step7.h b/example/lapack_to_chameleon/step7.h
index 5026f8750456e486f12846eb83a92589987ea47f..33d74144209dd9b3cdbfc06fd14ffc1d65c456fc 100644
--- a/example/lapack_to_chameleon/step7.h
+++ b/example/lapack_to_chameleon/step7.h
@@ -11,11 +11,11 @@
*
* @brief Chameleon step7 example header
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Florent Pruvost
* @author Guillaume Sylvand
* @author Mathieu Faverge
- * @date 2022-02-22
+ * @date 2024-03-14
*
*/
#ifndef _step7_h_
@@ -74,19 +74,42 @@ static void init_iparam(int iparam[IPARAM_SIZEOF]){
* and store it at the adresse 'buffer' with leading dimension 'ld'
*/
struct data_pl {
- double bump;
+ double bump;
unsigned long long int seed;
- int bigM;
};
-static void Cham_build_callback_plgsy(int row_min, int row_max, int col_min, int col_max, void *buffer, int ld, void *user_data) {
- struct data_pl *data=(struct data_pl *)user_data;
- CORE_dplgsy(data->bump, row_max-row_min+1, col_max-col_min+1, buffer, ld, data->bigM, row_min, col_min, data->seed);
+static int Cham_build_plgsy_cpu( void *op_args, cham_uplo_t uplo, int m, int n, int ndata,
+ const CHAM_desc_t *descA, CHAM_tile_t *tileA, ... )
+{
+ struct data_pl *data = (struct data_pl *)op_args;
+ int tempmm, tempnn;
+
+ /* Get the dimension of the tile */
+ tempmm = (m == (descA->mt-1)) ? (descA->m - m * descA->mb) : descA->mb;
+ tempnn = (n == (descA->nt-1)) ? (descA->n - n * descA->nb) : descA->nb;
+
+ TCORE_dplgsy( data->bump, tempmm, tempnn, tileA,
+ descA->m, m * descA->mb, n * descA->nb, data->seed );
+
+ (void)uplo;
+ return 0;
}
-static void Cham_build_callback_plrnt(int row_min, int row_max, int col_min, int col_max, void *buffer, int ld, void *user_data) {
- struct data_pl *data=(struct data_pl *)user_data;
- CORE_dplrnt(row_max-row_min+1, col_max-col_min+1, buffer, ld, data->bigM, row_min, col_min, data->seed);
+static int Cham_build_plrnt_cpu( void *op_args, cham_uplo_t uplo, int m, int n, int ndata,
+ const CHAM_desc_t *descA, CHAM_tile_t *tileA, ... )
+{
+ struct data_pl *data = (struct data_pl *)op_args;
+ int tempmm, tempnn;
+
+ /* Get the dimension of the tile */
+ tempmm = (m == (descA->mt-1)) ? (descA->m - m * descA->mb) : descA->mb;
+ tempnn = (n == (descA->nt-1)) ? (descA->n - n * descA->nb) : descA->nb;
+
+ TCORE_dplrnt( tempmm, tempnn, tileA,
+ descA->m, m * descA->mb, n * descA->nb, data->seed );
+
+ (void)uplo;
+ return 0;
}
/**
diff --git a/include/chameleon/chameleon_z.h b/include/chameleon/chameleon_z.h
index 6fbf8a1584e6ec9a0ba071d0c852f213cf244335..166feeef619018307a7f0873d0dc3ccf1b4983b4 100644
--- a/include/chameleon/chameleon_z.h
+++ b/include/chameleon/chameleon_z.h
@@ -23,7 +23,7 @@
* @author Florent Pruvost
* @author Alycia Lisito
* @author Matthieu Kuhn
- * @date 2023-08-22
+ * @date 2024-03-14
* @precisions normal z -> c d s
*
*/
@@ -374,9 +374,9 @@ int CHAMELEON_zDesc2Lap( cham_uplo_t uplo, CHAM_desc_t *A, CHAMELEON_Complex64_t
/**
* User Builder function prototypes
*/
-int CHAMELEON_zbuild(cham_uplo_t uplo, int M, int N, CHAMELEON_Complex64_t *A, int LDA, void *user_data, void* user_build_callback);
-int CHAMELEON_zbuild_Tile(cham_uplo_t uplo, CHAM_desc_t *A, void *user_data, void* user_build_callback );
-int CHAMELEON_zbuild_Tile_Async(cham_uplo_t uplo, CHAM_desc_t *A, void *user_data, void* user_build_callback, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
+int CHAMELEON_zbuild(cham_uplo_t uplo, int M, int N, CHAMELEON_Complex64_t *A, int LDA, void *user_data, void* user_build_callback) __attribute__((deprecated("Please refer to CHAMELEON_mapv_Tile() instead")));
+int CHAMELEON_zbuild_Tile(cham_uplo_t uplo, CHAM_desc_t *A, void *user_data, void* user_build_callback ) __attribute__((deprecated("Please refer to CHAMELEON_mapv_Tile() instead")));
+int CHAMELEON_zbuild_Tile_Async(cham_uplo_t uplo, CHAM_desc_t *A, void *user_data, void* user_build_callback, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request) __attribute__((deprecated("Please refer to CHAMELEON_mapv_Tile_Async() instead")));
/**
* Centered-Scaled function prototypes
diff --git a/runtime/parsec/codelets/codelet_map.c b/runtime/parsec/codelets/codelet_map.c
index 12a701bf379b41e5e9e42a61b4df3153c4b3cf60..d14ee9996087df3d6ee4357e649f9e93c832a0a4 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 );
}