diff --git a/compute/pzhetrd_he2hb.c b/compute/pzhetrd_he2hb.c
index 623a25517808356acaa7b61373e831a9b043452b..602f7f4e6cb923454d2381961ac7c89ff3f8af11 100644
--- a/compute/pzhetrd_he2hb.c
+++ b/compute/pzhetrd_he2hb.c
@@ -23,9 +23,9 @@
#define A(m, n) A, m, n
#define T(m, n) T, m, n
-#define D(k) D, (k)-1, 0
+#define D(k) &D, (k)-1, 0
-#define AT(k) AT, k, 0
+#define AT(k) &AT, k, 0
#if defined(CHAMELEON_COPY_DIAG)
#define E(m, n) E, m, 0
@@ -42,8 +42,8 @@ void chameleon_pzhetrd_he2hb(cham_uplo_t uplo,
{
CHAM_context_t *chamctxt;
RUNTIME_option_t options;
- CHAM_desc_t *D = NULL;
- CHAM_desc_t *AT = NULL;
+ CHAM_desc_t D;
+ CHAM_desc_t AT;
size_t ws_worker = 0;
size_t ws_host = 0;
@@ -87,15 +87,12 @@ void chameleon_pzhetrd_he2hb(cham_uplo_t uplo,
RUNTIME_options_ws_alloc( &options, ws_worker, ws_host );
/* Copy of the diagonal tiles to keep the general version of the tile all along the computation */
- D = (CHAM_desc_t*)malloc(sizeof(CHAM_desc_t));
- chameleon_zdesc_alloc_diag(*D, A->mb, A->nb, chameleon_min(A->m, A->n) - A->mb, A->nb, 0, 0, chameleon_min(A->m, A->n) - A->mb, A->nb, A->p, A->q);
+ chameleon_zdesc_alloc_diag( &D, A->mb, A->m, A->n, A->p, A->q );
- AT = (CHAM_desc_t*)malloc(sizeof(CHAM_desc_t));
- *AT = chameleon_desc_init(
- ChamComplexDouble, A->mb, A->nb, (A->mb*A->nb),
- chameleon_min(A->mt, A->nt) * A->mb, A->nb, 0, 0, chameleon_min(A->mt, A->nt) * A->mb, A->nb, 1, 1);
- chameleon_desc_mat_alloc( AT );
- RUNTIME_desc_create( AT );
+ chameleon_desc_init( &AT, CHAMELEON_MAT_ALLOC_GLOBAL, ChamComplexDouble, A->mb, A->nb, (A->mb*A->nb),
+ chameleon_min(A->mt, A->nt) * A->mb, A->nb, 0, 0,
+ chameleon_min(A->mt, A->nt) * A->mb, A->nb, 1, 1,
+ NULL, NULL, NULL );
/* Let's extract the diagonal in a temporary copy that contains A and A' */
for (k = 1; k < A->nt; k++){
@@ -437,8 +434,8 @@ void chameleon_pzhetrd_he2hb(cham_uplo_t uplo,
RUNTIME_options_finalize(&options, chamctxt);
CHAMELEON_Sequence_Wait(sequence);
- CHAMELEON_Desc_Destroy( &D );
- CHAMELEON_Desc_Destroy( &AT );
+ chameleon_desc_destroy( &D );
+ chameleon_desc_destroy( &AT );
(void)E;
}
diff --git a/compute/pzlange.c b/compute/pzlange.c
index cefbf531476995539dbb808d028c11dd5cd572f3..3f549b17c355764ce3c18eba3d201c0f53f394bf 100644
--- a/compute/pzlange.c
+++ b/compute/pzlange.c
@@ -26,9 +26,8 @@
//WS_ADD : A->mb + A->nb
#include "control/common.h"
-#define A(m, n) A, (m), (n)
-#define Wcol(m, n) Wcol, (m), (n)
-#define Welt(m, n) Welt, (m), (n)
+#define A( m, n ) A, (m), (n)
+#define W( desc, m, n ) (desc), (m), (n)
static inline void
chameleon_pzlange_one( cham_uplo_t uplo, cham_diag_t diag, CHAM_desc_t *A,
@@ -63,21 +62,21 @@ chameleon_pzlange_one( cham_uplo_t uplo, cham_diag_t diag, CHAM_desc_t *A,
INSERT_TASK_ztrasm(
options,
ChamColumnwise, uplo, diag, tempmm, tempnn,
- A(m, n), ldam, Wcol(m, n) );
+ A(m, n), ldam, W( Wcol, m, n ) );
}
else {
INSERT_TASK_dzasum(
options,
ChamColumnwise, ChamUpperLower, tempmm, tempnn,
- A(m, n), ldam, Wcol(m, n) );
+ A(m, n), ldam, W( Wcol, m, n ) );
}
if ( m >= P ) {
INSERT_TASK_dgeadd(
options,
- ChamNoTrans, tempnn, 1, A->nb,
- 1.0, Wcol(m, n), tempnn,
- 1.0, Wcol(m%P, n), tempnn );
+ ChamNoTrans, 1, tempnn, A->nb,
+ 1.0, W( Wcol, m, n ), 1,
+ 1.0, W( Wcol, m%P, n ), 1 );
}
}
@@ -88,15 +87,15 @@ chameleon_pzlange_one( cham_uplo_t uplo, cham_diag_t diag, CHAM_desc_t *A,
for(m = 1; m < P; n++) {
INSERT_TASK_dgeadd(
options,
- ChamNoTrans, tempnn, 1, A->nb,
- 1.0, Wcol(m, n), tempnn,
- 1.0, Wcol(0, n), tempnn );
+ ChamNoTrans, 1, tempnn, A->nb,
+ 1.0, W( Wcol, m, n ), 1,
+ 1.0, W( Wcol, 0, n ), 1 );
}
INSERT_TASK_dlange(
options,
- ChamMaxNorm, tempnn, 1, A->nb,
- Wcol(0, n), tempnn, Welt(0, n));
+ ChamMaxNorm, 1, tempnn, A->nb,
+ W( Wcol, 0, n), 1, W( Welt, 0, n));
}
/**
@@ -106,7 +105,7 @@ chameleon_pzlange_one( cham_uplo_t uplo, cham_diag_t diag, CHAM_desc_t *A,
for(n = Q; n < NT; n++) {
INSERT_TASK_dlange_max(
options,
- Welt(0, n), Welt(0, n%Q) );
+ W( Welt, 0, n), W( Welt, 0, n%Q) );
}
/**
@@ -116,7 +115,7 @@ chameleon_pzlange_one( cham_uplo_t uplo, cham_diag_t diag, CHAM_desc_t *A,
for(n = 1; n < Q; n++) {
INSERT_TASK_dlange_max(
options,
- Welt(0, n), Welt(0, 0) );
+ W( Welt, 0, n), W( Welt, 0, 0) );
}
}
@@ -153,21 +152,21 @@ chameleon_pzlange_inf( cham_uplo_t uplo, cham_diag_t diag, CHAM_desc_t *A,
INSERT_TASK_ztrasm(
options,
ChamRowwise, uplo, diag, tempmm, tempnn,
- A(m, n), ldam, Wcol(m, n) );
+ A(m, n), ldam, W( Wcol, m, n) );
}
else {
INSERT_TASK_dzasum(
options,
ChamRowwise, ChamUpperLower, tempmm, tempnn,
- A(m, n), ldam, Wcol(m, n) );
+ A(m, n), ldam, W( Wcol, m, n) );
}
if ( n >= Q ) {
INSERT_TASK_dgeadd(
options,
ChamNoTrans, tempmm, 1, A->mb,
- 1.0, Wcol(m, n ), tempmm,
- 1.0, Wcol(m, n%Q), tempmm );
+ 1.0, W( Wcol, m, n ), tempmm,
+ 1.0, W( Wcol, m, n%Q), tempmm );
}
}
@@ -179,14 +178,14 @@ chameleon_pzlange_inf( cham_uplo_t uplo, cham_diag_t diag, CHAM_desc_t *A,
INSERT_TASK_dgeadd(
options,
ChamNoTrans, tempmm, 1, A->mb,
- 1.0, Wcol(m, n), tempmm,
- 1.0, Wcol(m, 0), tempmm );
+ 1.0, W( Wcol, m, n), tempmm,
+ 1.0, W( Wcol, m, 0), tempmm );
}
INSERT_TASK_dlange(
options,
ChamMaxNorm, tempmm, 1, A->nb,
- Wcol(m, 0), 1, Welt(m, 0));
+ W( Wcol, m, 0), 1, W( Welt, m, 0));
}
/**
@@ -196,7 +195,7 @@ chameleon_pzlange_inf( cham_uplo_t uplo, cham_diag_t diag, CHAM_desc_t *A,
for(m = P; m < MT; m++) {
INSERT_TASK_dlange_max(
options,
- Welt(m, 0), Welt(m%P, 0) );
+ W( Welt, m, 0), W( Welt, m%P, 0) );
}
/**
@@ -206,7 +205,7 @@ chameleon_pzlange_inf( cham_uplo_t uplo, cham_diag_t diag, CHAM_desc_t *A,
for(m = 1; m < P; m++) {
INSERT_TASK_dlange_max(
options,
- Welt(m, 0), Welt(0, 0) );
+ W( Welt, m, 0), W( Welt, 0, 0) );
}
}
@@ -242,19 +241,19 @@ chameleon_pzlange_max( cham_uplo_t uplo, cham_diag_t diag, CHAM_desc_t *A, CHAM_
INSERT_TASK_zlantr(
options,
ChamMaxNorm, uplo, diag, tempmm, tempnn, A->nb,
- A(m, n), ldam, Welt(m, n));
+ A(m, n), ldam, W( Welt, m, n));
}
else {
INSERT_TASK_zlange(
options,
ChamMaxNorm, tempmm, tempnn, A->nb,
- A(m, n), ldam, Welt(m, n));
+ A(m, n), ldam, W( Welt, m, n));
}
if ( n >= Q ) {
INSERT_TASK_dlange_max(
options,
- Welt(m, n), Welt(m, n%Q) );
+ W( Welt, m, n), W( Welt, m, n%Q) );
}
}
@@ -265,7 +264,7 @@ chameleon_pzlange_max( cham_uplo_t uplo, cham_diag_t diag, CHAM_desc_t *A, CHAM_
for(n = 1; n < Q; n++) {
INSERT_TASK_dlange_max(
options,
- Welt(m, n), Welt(m, 0) );
+ W( Welt, m, n), W( Welt, m, 0) );
}
}
@@ -276,7 +275,7 @@ chameleon_pzlange_max( cham_uplo_t uplo, cham_diag_t diag, CHAM_desc_t *A, CHAM_
for(m = P; m < MT; m++) {
INSERT_TASK_dlange_max(
options,
- Welt(m, 0), Welt(m%P, 0) );
+ W( Welt, m, 0), W( Welt, m%P, 0) );
}
/**
@@ -286,7 +285,7 @@ chameleon_pzlange_max( cham_uplo_t uplo, cham_diag_t diag, CHAM_desc_t *A, CHAM_
for(m = 1; m < P; m++) {
INSERT_TASK_dlange_max(
options,
- Welt(m, 0), Welt(0, 0) );
+ W( Welt, m, 0), W( Welt, 0, 0) );
}
}
@@ -322,18 +321,18 @@ chameleon_pzlange_frb( cham_uplo_t uplo, cham_diag_t diag, CHAM_desc_t *A, CHAM_
INSERT_TASK_ztrssq(
options,
uplo, diag, tempmm, tempnn,
- A(m, n), ldam, Welt(m, n) );
+ A(m, n), ldam, W( Welt, m, n) );
}
else {
INSERT_TASK_zgessq(
options,
tempmm, tempnn,
- A(m, n), ldam, Welt(m, n) );
+ A(m, n), ldam, W( Welt, m, n) );
}
if ( n >= Q ) {
INSERT_TASK_dplssq(
- options, Welt(m, n), Welt(m, n%Q) );
+ options, W( Welt, m, n), W( Welt, m, n%Q) );
}
}
@@ -343,7 +342,7 @@ chameleon_pzlange_frb( cham_uplo_t uplo, cham_diag_t diag, CHAM_desc_t *A, CHAM_
*/
for(n = 1; n < Q; n++) {
INSERT_TASK_dplssq(
- options, Welt(m, n), Welt(m, 0) );
+ options, W( Welt, m, n), W( Welt, m, 0) );
}
}
@@ -353,7 +352,7 @@ chameleon_pzlange_frb( cham_uplo_t uplo, cham_diag_t diag, CHAM_desc_t *A, CHAM_
*/
for(m = P; m < MT; m++) {
INSERT_TASK_dplssq(
- options, Welt(m, 0), Welt(m%P, 0) );
+ options, W( Welt, m, 0), W( Welt, m%P, 0) );
}
/**
@@ -362,11 +361,11 @@ chameleon_pzlange_frb( cham_uplo_t uplo, cham_diag_t diag, CHAM_desc_t *A, CHAM_
*/
for(m = 1; m < P; m++) {
INSERT_TASK_dplssq(
- options, Welt(m, 0), Welt(0, 0) );
+ options, W( Welt, m, 0), W( Welt, 0, 0) );
}
INSERT_TASK_dplssq2(
- options, Welt(0, 0) );
+ options, W( Welt, 0, 0) );
}
/**
@@ -378,13 +377,13 @@ void chameleon_pzlange_generic( cham_normtype_t norm, cham_uplo_t uplo, cham_dia
{
CHAM_context_t *chamctxt;
RUNTIME_option_t options;
- CHAM_desc_t *Wcol = NULL;
- CHAM_desc_t *Welt = NULL;
+ CHAM_desc_t Wcol;
+ CHAM_desc_t Welt;
double alpha = 0.0;
double beta = 0.0;
int workn, workmt, worknt;
- int m, n;
+ int m, n, wcol_init = 0;
chamctxt = chameleon_context_self();
if ( sequence->status != CHAMELEON_SUCCESS ) {
@@ -402,11 +401,14 @@ void chameleon_pzlange_generic( cham_normtype_t norm, cham_uplo_t uplo, cham_dia
case ChamOneNorm:
RUNTIME_options_ws_alloc( &options, 1, 0 );
- CHAMELEON_Desc_Create( &Wcol, NULL, ChamRealDouble, 1, A->nb, A->nb,
- workmt, worknt * A->nb, 0, 0, workmt, worknt * A->nb, A->p, A->q );
+ chameleon_desc_init( &Wcol, CHAMELEON_MAT_ALLOC_GLOBAL, ChamRealDouble, 1, A->nb, A->nb,
+ workmt, worknt * A->nb, 0, 0, workmt, worknt * A->nb, A->p, A->q,
+ NULL, NULL, NULL );
+ wcol_init = 1;
- CHAMELEON_Desc_Create( &Welt, NULL, ChamRealDouble, 1, 1, 1,
- A->p, worknt, 0, 0, A->p, worknt, A->p, A->q );
+ chameleon_desc_init( &Welt, CHAMELEON_MAT_ALLOC_GLOBAL, ChamRealDouble, 1, 1, 1,
+ A->p, worknt, 0, 0, A->p, worknt, A->p, A->q,
+ NULL, NULL, NULL );
break;
@@ -416,11 +418,14 @@ void chameleon_pzlange_generic( cham_normtype_t norm, cham_uplo_t uplo, cham_dia
case ChamInfNorm:
RUNTIME_options_ws_alloc( &options, A->mb, 0 );
- CHAMELEON_Desc_Create( &Wcol, NULL, ChamRealDouble, A->mb, 1, A->mb,
- workmt * A->mb, worknt, 0, 0, workmt * A->mb, worknt, A->p, A->q );
+ chameleon_desc_init( &Wcol, CHAMELEON_MAT_ALLOC_GLOBAL, ChamRealDouble, A->mb, 1, A->mb,
+ workmt * A->mb, worknt, 0, 0, workmt * A->mb, worknt, A->p, A->q,
+ NULL, NULL, NULL );
+ wcol_init = 1;
- CHAMELEON_Desc_Create( &Welt, NULL, ChamRealDouble, 1, 1, 1,
- workmt, A->q, 0, 0, workmt, A->q, A->p, A->q );
+ chameleon_desc_init( &Welt, CHAMELEON_MAT_ALLOC_GLOBAL, ChamRealDouble, 1, 1, 1,
+ workmt, A->q, 0, 0, workmt, A->q, A->p, A->q,
+ NULL, NULL, NULL );
break;
/*
@@ -430,8 +435,9 @@ void chameleon_pzlange_generic( cham_normtype_t norm, cham_uplo_t uplo, cham_dia
RUNTIME_options_ws_alloc( &options, 1, 0 );
alpha = 1.;
- CHAMELEON_Desc_Create( &Welt, NULL, ChamRealDouble, 2, 1, 2,
- workmt*2, workn, 0, 0, workmt*2, workn, A->p, A->q );
+ chameleon_desc_init( &Welt, CHAMELEON_MAT_ALLOC_GLOBAL, ChamRealDouble, 2, 1, 2,
+ workmt*2, workn, 0, 0, workmt*2, workn, A->p, A->q,
+ NULL, NULL, NULL );
break;
/*
@@ -441,8 +447,9 @@ void chameleon_pzlange_generic( cham_normtype_t norm, cham_uplo_t uplo, cham_dia
default:
RUNTIME_options_ws_alloc( &options, 1, 0 );
- CHAMELEON_Desc_Create( &Welt, NULL, ChamRealDouble, 1, 1, 1,
- workmt, workn, 0, 0, workmt, workn, A->p, A->q );
+ chameleon_desc_init( &Welt, CHAMELEON_MAT_ALLOC_GLOBAL, ChamRealDouble, 1, 1, 1,
+ workmt, workn, 0, 0, workmt, workn, A->p, A->q,
+ NULL, NULL, NULL );
}
/* Initialize workspaces */
@@ -450,44 +457,44 @@ void chameleon_pzlange_generic( cham_normtype_t norm, cham_uplo_t uplo, cham_dia
(norm == ChamOneNorm) )
{
/* Initialize Wcol tile */
- for(m = 0; m < Wcol->mt; m++) {
- for(n = 0; n < Wcol->nt; n++) {
+ for(m = 0; m < Wcol.mt; m++) {
+ for(n = 0; n < Wcol.nt; n++) {
INSERT_TASK_dlaset(
&options,
- ChamUpperLower, Wcol->mb, Wcol->nb,
+ ChamUpperLower, Wcol.mb, Wcol.nb,
alpha, beta,
- Wcol(m,n), Wcol->mb );
+ W( &Wcol, m, n ), Wcol.mb );
}
}
}
- for(m = 0; m < Welt->mt; m++) {
- for(n = 0; n < Welt->nt; n++) {
+ for(m = 0; m < Welt.mt; m++) {
+ for(n = 0; n < Welt.nt; n++) {
INSERT_TASK_dlaset(
&options,
- ChamUpperLower, Welt->mb, Welt->nb,
+ ChamUpperLower, Welt.mb, Welt.nb,
alpha, beta,
- Welt(m,n), Welt->mb );
+ W( &Welt, m, n ), Welt.mb );
}
}
switch ( norm ) {
case ChamOneNorm:
- chameleon_pzlange_one( uplo, diag, A, Wcol, Welt, &options );
- CHAMELEON_Desc_Flush( Wcol, sequence );
+ chameleon_pzlange_one( uplo, diag, A, &Wcol, &Welt, &options );
+ CHAMELEON_Desc_Flush( &Wcol, sequence );
break;
case ChamInfNorm:
- chameleon_pzlange_inf( uplo, diag, A, Wcol, Welt, &options );
- CHAMELEON_Desc_Flush( Wcol, sequence );
+ chameleon_pzlange_inf( uplo, diag, A, &Wcol, &Welt, &options );
+ CHAMELEON_Desc_Flush( &Wcol, sequence );
break;
case ChamFrobeniusNorm:
- chameleon_pzlange_frb( uplo, diag, A, Welt, &options );
+ chameleon_pzlange_frb( uplo, diag, A, &Welt, &options );
break;
case ChamMaxNorm:
default:
- chameleon_pzlange_max( uplo, diag, A, Welt, &options );
+ chameleon_pzlange_max( uplo, diag, A, &Welt, &options );
}
/**
@@ -499,20 +506,20 @@ void chameleon_pzlange_generic( cham_normtype_t norm, cham_uplo_t uplo, cham_dia
INSERT_TASK_dlacpy(
&options,
ChamUpperLower, 1, 1, 1,
- Welt(0,0), 1, Welt(m, n), 1);
+ W( &Welt, 0, 0 ), 1, W( &Welt, m, n ), 1);
}
}
}
- CHAMELEON_Desc_Flush( Welt, sequence );
+ CHAMELEON_Desc_Flush( &Welt, sequence );
RUNTIME_sequence_wait(chamctxt, sequence);
- *result = *(double *)Welt->get_blkaddr(Welt, A->myrank / A->q, A->myrank % A->q );
+ *result = *(double *)Welt.get_blkaddr( &Welt, A->myrank / A->q, A->myrank % A->q );
- if ( Wcol != NULL ) {
- CHAMELEON_Desc_Destroy( &Wcol );
+ if ( wcol_init ) {
+ chameleon_desc_destroy( &Wcol );
}
- CHAMELEON_Desc_Destroy( &Welt );
+ chameleon_desc_destroy( &Welt );
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, chamctxt);
diff --git a/compute/pzlansy.c b/compute/pzlansy.c
index 1d327383865353c2a1f3c835ea0a69c2749340e4..b81041526222324726d7cb19febdf33300694b5f 100644
--- a/compute/pzlansy.c
+++ b/compute/pzlansy.c
@@ -27,9 +27,8 @@
#include <math.h>
#include "control/common.h"
-#define A(m, n) A, (m), (n)
-#define Wcol(m, n) Wcol, (m), (n)
-#define Welt(m, n) Welt, (m), (n)
+#define A( m, n ) A, (m), (n)
+#define W( desc, m, n ) (desc), (m), (n)
static inline void
chameleon_pzlansy_inf( cham_uplo_t uplo, CHAM_desc_t *A,
@@ -62,18 +61,18 @@ chameleon_pzlansy_inf( cham_uplo_t uplo, CHAM_desc_t *A,
INSERT_TASK_dzasum(
options,
ChamRowwise, uplo, tempmm, tempnn,
- A(m, n), ldam, Wcol(m, n) );
+ A(m, n), ldam, W( Wcol, m, n) );
}
else {
INSERT_TASK_dzasum(
options,
ChamRowwise, ChamUpperLower, tempmm, tempnn,
- A(m, n), ldam, Wcol(m, n) );
+ A(m, n), ldam, W( Wcol, m, n) );
INSERT_TASK_dzasum(
options,
ChamColumnwise, ChamUpperLower, tempmm, tempnn,
- A(m, n), ldam, Wcol(n, m) );
+ A(m, n), ldam, W( Wcol, n, m) );
}
}
}
@@ -85,8 +84,8 @@ chameleon_pzlansy_inf( cham_uplo_t uplo, CHAM_desc_t *A,
INSERT_TASK_dgeadd(
options,
ChamNoTrans, tempmm, 1, A->nb,
- 1.0, Wcol(m, n ), tempmm,
- 1.0, Wcol(m, n%Q), tempmm );
+ 1.0, W( Wcol, m, n ), tempmm,
+ 1.0, W( Wcol, m, n%Q), tempmm );
}
/**
@@ -97,34 +96,34 @@ chameleon_pzlansy_inf( cham_uplo_t uplo, CHAM_desc_t *A,
INSERT_TASK_dgeadd(
options,
ChamNoTrans, tempmm, 1, A->mb,
- 1.0, Wcol(m, n), tempmm,
- 1.0, Wcol(m, 0), tempmm );
+ 1.0, W( Wcol, m, n), tempmm,
+ 1.0, W( Wcol, m, 0), tempmm );
}
INSERT_TASK_dlange(
options,
ChamMaxNorm, tempmm, 1, A->nb,
- Wcol(m, 0), 1, Welt(m, 0));
+ W( Wcol, m, 0), 1, W( Welt, m, 0));
}
/**
* Step 3:
- * For m in 0..P-1, Welt(m, n) = max( Wcol(m..mt[P], n ) )
+ * For m in 0..P-1, W( Welt, m, n) = max( Wcol(m..mt[P], n ) )
*/
for(m = P; m < MT; m++) {
INSERT_TASK_dlange_max(
options,
- Welt(m, 0), Welt(m%P, 0) );
+ W( Welt, m, 0), W( Welt, m%P, 0) );
}
/**
* Step 4:
- * For each i, Welt(i, n) = max( Welt(0..P-1, n) )
+ * For each i, W( Welt, i, n) = max( W( Welt, 0..P-1, n) )
*/
for(m = 1; m < P; m++) {
INSERT_TASK_dlange_max(
options,
- Welt(m, 0), Welt(0, 0) );
+ W( Welt, m, 0), W( Welt, 0, 0) );
}
}
@@ -159,26 +158,26 @@ chameleon_pzlansy_max( cham_trans_t trans, cham_uplo_t uplo, CHAM_desc_t *A,
INSERT_TASK_zlanhe(
options,
ChamMaxNorm, uplo, tempmm, A->nb,
- A(m, n), ldam, Welt(m, n));
+ A(m, n), ldam, W( Welt, m, n));
}
else {
INSERT_TASK_zlansy(
options,
ChamMaxNorm, uplo, tempmm, A->nb,
- A(m, n), ldam, Welt(m, n));
+ A(m, n), ldam, W( Welt, m, n));
}
}
else {
INSERT_TASK_zlange(
options,
ChamMaxNorm, tempmm, tempnn, A->nb,
- A(m, n), ldam, Welt(m, n));
+ A(m, n), ldam, W( Welt, m, n));
}
if ( n >= Q ) {
INSERT_TASK_dlange_max(
options,
- Welt(m, n), Welt(m, n%Q) );
+ W( Welt, m, n), W( Welt, m, n%Q) );
}
}
@@ -189,7 +188,7 @@ chameleon_pzlansy_max( cham_trans_t trans, cham_uplo_t uplo, CHAM_desc_t *A,
for(n = 1; n < Q; n++) {
INSERT_TASK_dlange_max(
options,
- Welt(m, n), Welt(m, 0) );
+ W( Welt, m, n), W( Welt, m, 0) );
}
}
@@ -200,7 +199,7 @@ chameleon_pzlansy_max( cham_trans_t trans, cham_uplo_t uplo, CHAM_desc_t *A,
for(m = P; m < MT; m++) {
INSERT_TASK_dlange_max(
options,
- Welt(m, 0), Welt(m%P, 0) );
+ W( Welt, m, 0), W( Welt, m%P, 0) );
}
/**
@@ -210,7 +209,7 @@ chameleon_pzlansy_max( cham_trans_t trans, cham_uplo_t uplo, CHAM_desc_t *A,
for(m = 1; m < P; m++) {
INSERT_TASK_dlange_max(
options,
- Welt(m, 0), Welt(0, 0) );
+ W( Welt, m, 0), W( Welt, 0, 0) );
}
}
@@ -245,21 +244,21 @@ chameleon_pzlansy_frb( cham_trans_t trans, cham_uplo_t uplo,
if ( trans == ChamConjTrans) {
INSERT_TASK_zhessq(
options, uplo, tempmm,
- A(m, n), ldam, Welt(m, n) );
+ A(m, n), ldam, W( Welt, m, n) );
}
else {
INSERT_TASK_zsyssq(
options, uplo, tempmm,
- A(m, n), ldam, Welt(m, n) );
+ A(m, n), ldam, W( Welt, m, n) );
}
}
else {
INSERT_TASK_zgessq(
options, tempmm, tempnn,
- A(m, n), ldam, Welt(m, n) );
+ A(m, n), ldam, W( Welt, m, n) );
INSERT_TASK_zgessq(
options, tempmm, tempnn,
- A(m, n), ldam, Welt(n, m) );
+ A(m, n), ldam, W( Welt, n, m) );
}
}
}
@@ -267,16 +266,16 @@ chameleon_pzlansy_frb( cham_trans_t trans, cham_uplo_t uplo,
for(m = 0; m < MT; m++) {
for(n = Q; n < NT; n++) {
INSERT_TASK_dplssq(
- options, Welt(m, n), Welt(m, n%Q) );
+ options, W( Welt, m, n), W( Welt, m, n%Q) );
}
/**
* Step 2:
- * For each j, W(m, j) = reduce( Welt(m, 0..Q-1) )
+ * For each j, W(m, j) = reduce( W( Welt, m, 0..Q-1) )
*/
for(n = 1; n < Q; n++) {
INSERT_TASK_dplssq(
- options, Welt(m, n), Welt(m, 0) );
+ options, W( Welt, m, n), W( Welt, m, 0) );
}
}
@@ -286,7 +285,7 @@ chameleon_pzlansy_frb( cham_trans_t trans, cham_uplo_t uplo,
*/
for(m = P; m < MT; m++) {
INSERT_TASK_dplssq(
- options, Welt(m, 0), Welt(m%P, 0) );
+ options, W( Welt, m, 0), W( Welt, m%P, 0) );
}
/**
@@ -295,11 +294,11 @@ chameleon_pzlansy_frb( cham_trans_t trans, cham_uplo_t uplo,
*/
for(m = 1; m < P; m++) {
INSERT_TASK_dplssq(
- options, Welt(m, 0), Welt(0, 0) );
+ options, W( Welt, m, 0), W( Welt, 0, 0) );
}
INSERT_TASK_dplssq2(
- options, Welt(0, 0) );
+ options, W( Welt, 0, 0) );
}
/**
@@ -311,13 +310,13 @@ void chameleon_pzlansy_generic( cham_normtype_t norm, cham_uplo_t uplo, cham_tra
{
CHAM_context_t *chamctxt;
RUNTIME_option_t options;
- CHAM_desc_t *Wcol = NULL;
- CHAM_desc_t *Welt = NULL;
+ CHAM_desc_t Wcol;
+ CHAM_desc_t Welt;
double alpha = 0.0;
double beta = 0.0;
int workn, workmt, worknt;
- int m, n;
+ int m, n, wcol_init = 0;
chamctxt = chameleon_context_self();
if ( sequence->status != CHAMELEON_SUCCESS ) {
@@ -336,11 +335,14 @@ void chameleon_pzlansy_generic( cham_normtype_t norm, cham_uplo_t uplo, cham_tra
case ChamInfNorm:
RUNTIME_options_ws_alloc( &options, 1, 0 );
- CHAMELEON_Desc_Create( &Wcol, NULL, ChamRealDouble, A->mb, 1, A->mb,
- workmt * A->mb, worknt, 0, 0, workmt * A->mb, worknt, A->p, A->q );
+ chameleon_desc_init( &Wcol, CHAMELEON_MAT_ALLOC_GLOBAL, ChamRealDouble, A->mb, 1, A->mb,
+ workmt * A->mb, worknt, 0, 0, workmt * A->mb, worknt, A->p, A->q,
+ NULL, NULL, NULL );
+ wcol_init = 1;
- CHAMELEON_Desc_Create( &Welt, NULL, ChamRealDouble, 1, 1, 1,
- workmt, A->q, 0, 0, workmt, A->q, A->p, A->q );
+ chameleon_desc_init( &Welt, CHAMELEON_MAT_ALLOC_GLOBAL, ChamRealDouble, 1, 1, 1,
+ workmt, A->q, 0, 0, workmt, A->q, A->p, A->q,
+ NULL, NULL, NULL );
break;
/*
@@ -350,8 +352,9 @@ void chameleon_pzlansy_generic( cham_normtype_t norm, cham_uplo_t uplo, cham_tra
RUNTIME_options_ws_alloc( &options, 1, 0 );
alpha = 1.;
- CHAMELEON_Desc_Create( &Welt, NULL, ChamRealDouble, 2, 1, 2,
- workmt*2, workn, 0, 0, workmt*2, workn, A->p, A->q );
+ chameleon_desc_init( &Welt, CHAMELEON_MAT_ALLOC_GLOBAL, ChamRealDouble, 2, 1, 2,
+ workmt*2, workn, 0, 0, workmt*2, workn, A->p, A->q,
+ NULL, NULL, NULL );
break;
/*
@@ -361,8 +364,9 @@ void chameleon_pzlansy_generic( cham_normtype_t norm, cham_uplo_t uplo, cham_tra
default:
RUNTIME_options_ws_alloc( &options, 1, 0 );
- CHAMELEON_Desc_Create( &Welt, NULL, ChamRealDouble, 1, 1, 1,
- workmt, workn, 0, 0, workmt, workn, A->p, A->q );
+ chameleon_desc_init( &Welt, CHAMELEON_MAT_ALLOC_GLOBAL, ChamRealDouble, 1, 1, 1,
+ workmt, workn, 0, 0, workmt, workn, A->p, A->q,
+ NULL, NULL, NULL );
}
/* Initialize workspaces */
@@ -370,40 +374,40 @@ void chameleon_pzlansy_generic( cham_normtype_t norm, cham_uplo_t uplo, cham_tra
(norm == ChamOneNorm) )
{
/* Initialize Wcol tile */
- for(m = 0; m < Wcol->mt; m++) {
- for(n = 0; n < Wcol->nt; n++) {
+ for(m = 0; m < Wcol.mt; m++) {
+ for(n = 0; n < Wcol.nt; n++) {
INSERT_TASK_dlaset(
&options,
- ChamUpperLower, Wcol->mb, Wcol->nb,
+ ChamUpperLower, Wcol.mb, Wcol.nb,
alpha, beta,
- Wcol(m,n), Wcol->mb );
+ W( &Wcol, m, n ), Wcol.mb );
}
}
}
- for(m = 0; m < Welt->mt; m++) {
- for(n = 0; n < Welt->nt; n++) {
+ for(m = 0; m < Welt.mt; m++) {
+ for(n = 0; n < Welt.nt; n++) {
INSERT_TASK_dlaset(
&options,
- ChamUpperLower, Welt->mb, Welt->nb,
+ ChamUpperLower, Welt.mb, Welt.nb,
alpha, beta,
- Welt(m,n), Welt->mb );
+ W( &Welt, m, n ), Welt.mb );
}
}
switch ( norm ) {
case ChamOneNorm:
case ChamInfNorm:
- chameleon_pzlansy_inf( uplo, A, Wcol, Welt, &options );
- CHAMELEON_Desc_Flush( Wcol, sequence );
+ chameleon_pzlansy_inf( uplo, A, &Wcol, &Welt, &options );
+ CHAMELEON_Desc_Flush( &Wcol, sequence );
break;
case ChamFrobeniusNorm:
- chameleon_pzlansy_frb( trans, uplo, A, Welt, &options );
+ chameleon_pzlansy_frb( trans, uplo, A, &Welt, &options );
break;
case ChamMaxNorm:
default:
- chameleon_pzlansy_max( trans, uplo, A, Welt, &options );
+ chameleon_pzlansy_max( trans, uplo, A, &Welt, &options );
}
/**
@@ -415,20 +419,20 @@ void chameleon_pzlansy_generic( cham_normtype_t norm, cham_uplo_t uplo, cham_tra
INSERT_TASK_dlacpy(
&options,
ChamUpperLower, 1, 1, 1,
- Welt(0,0), 1, Welt(m, n), 1);
+ W( &Welt, 0, 0 ), 1, W( &Welt, m, n ), 1);
}
}
}
- CHAMELEON_Desc_Flush( Welt, sequence );
+ CHAMELEON_Desc_Flush( &Welt, sequence );
RUNTIME_sequence_wait(chamctxt, sequence);
- *result = *(double *)Welt->get_blkaddr(Welt, A->myrank / A->q, A->myrank % A->q );
+ *result = *(double *)Welt.get_blkaddr( &Welt, A->myrank / A->q, A->myrank % A->q );
- if ( Wcol != NULL ) {
- CHAMELEON_Desc_Destroy( &Wcol );
+ if ( wcol_init ) {
+ chameleon_desc_destroy( &Wcol );
}
- CHAMELEON_Desc_Destroy( &Welt );
+ chameleon_desc_destroy( &Welt );
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, chamctxt);
diff --git a/compute/zgelqf.c b/compute/zgelqf.c
index 388fb71483bda011dd0aadfc102c29b828b739c1..8c0cc4bec038b8d787854331a87b3c8f33272a12 100644
--- a/compute/zgelqf.c
+++ b/compute/zgelqf.c
@@ -297,8 +297,7 @@ int CHAMELEON_zgelqf_Tile_Async( CHAM_desc_t *A, CHAM_desc_t *T,
CHAMELEON_Desc_Flush( T, sequence );
CHAMELEON_Desc_Flush( Dptr, sequence );
chameleon_sequence_wait( chamctxt, sequence );
- RUNTIME_desc_destroy( Dptr );
- chameleon_desc_mat_free( Dptr );
+ chameleon_desc_destroy( Dptr );
}
(void)D;
return CHAMELEON_SUCCESS;
diff --git a/compute/zgelqf_param.c b/compute/zgelqf_param.c
index 2330cad7862ff14a5d611656fa5e01af978afcf1..fecf0f20f31a8e1bff2314c9da0ba9cb6967a6cb 100644
--- a/compute/zgelqf_param.c
+++ b/compute/zgelqf_param.c
@@ -297,8 +297,7 @@ int CHAMELEON_zgelqf_param_Tile_Async( const libhqr_tree_t *qrtree, CHAM_desc_t
CHAMELEON_Desc_Flush( TT, sequence );
CHAMELEON_Desc_Flush( Dptr, sequence );
chameleon_sequence_wait( chamctxt, sequence );
- RUNTIME_desc_destroy( Dptr );
- chameleon_desc_mat_free( Dptr );
+ chameleon_desc_destroy( Dptr );
}
(void)D;
return CHAMELEON_SUCCESS;
diff --git a/compute/zgelqs.c b/compute/zgelqs.c
index 93db15cf40479feee90c75b7ffc41c2cc1d75636..08ec0dd3dcd37eb606db79e900464cda4686f5ad 100644
--- a/compute/zgelqs.c
+++ b/compute/zgelqs.c
@@ -339,8 +339,7 @@ int CHAMELEON_zgelqs_Tile_Async( CHAM_desc_t *A, CHAM_desc_t *T, CHAM_desc_t *B,
CHAMELEON_Desc_Flush( T, sequence );
CHAMELEON_Desc_Flush( Dptr, sequence );
chameleon_sequence_wait( chamctxt, sequence );
- RUNTIME_desc_destroy( Dptr );
- chameleon_desc_mat_free( Dptr );
+ chameleon_desc_destroy( Dptr );
}
(void)D;
return CHAMELEON_SUCCESS;
diff --git a/compute/zgelqs_param.c b/compute/zgelqs_param.c
index 11f08cc57362e09583cd5ea264a9e9e113777810..b594dd602962c779e6450134376570ce8feadc04 100644
--- a/compute/zgelqs_param.c
+++ b/compute/zgelqs_param.c
@@ -346,8 +346,7 @@ int CHAMELEON_zgelqs_param_Tile_Async( const libhqr_tree_t *qrtree, CHAM_desc_t
CHAMELEON_Desc_Flush( TT, sequence );
CHAMELEON_Desc_Flush( Dptr, sequence );
chameleon_sequence_wait( chamctxt, sequence );
- RUNTIME_desc_destroy( Dptr );
- chameleon_desc_mat_free( Dptr );
+ chameleon_desc_destroy( Dptr );
}
(void)D;
return CHAMELEON_SUCCESS;
diff --git a/compute/zgels.c b/compute/zgels.c
index 7118f3e58cfbb6e52eefba84c762adad905feb03..99d7914da27fd1cdc88463a70e07ba5fee7d4675 100644
--- a/compute/zgels.c
+++ b/compute/zgels.c
@@ -432,8 +432,7 @@ int CHAMELEON_zgels_Tile_Async( cham_trans_t trans, CHAM_desc_t *A,
CHAMELEON_Desc_Flush( B, sequence );
CHAMELEON_Desc_Flush( Dptr, sequence );
chameleon_sequence_wait( chamctxt, sequence );
- RUNTIME_desc_destroy( Dptr );
- chameleon_desc_mat_free( Dptr );
+ chameleon_desc_destroy( Dptr );
}
(void)D;
return CHAMELEON_SUCCESS;
diff --git a/compute/zgels_param.c b/compute/zgels_param.c
index 7b41e4cdfc196e3a2dd9a6433222e2a0819fc050..ea23c9a4919594174e511c4a3256b61c9233ba4d 100644
--- a/compute/zgels_param.c
+++ b/compute/zgels_param.c
@@ -428,8 +428,7 @@ int CHAMELEON_zgels_param_Tile_Async( const libhqr_tree_t *qrtree, cham_trans_t
CHAMELEON_Desc_Flush( TT, sequence );
CHAMELEON_Desc_Flush( Dptr, sequence );
chameleon_sequence_wait( chamctxt, sequence );
- RUNTIME_desc_destroy( Dptr );
- chameleon_desc_mat_free( Dptr );
+ chameleon_desc_destroy( Dptr );
}
(void)D;
return CHAMELEON_SUCCESS;
diff --git a/compute/zgeqrf.c b/compute/zgeqrf.c
index 8e0d642ed787504dc3e22d8cbdf38399d86bc931..f365fc5106713605e8a485b5193412494c95e51b 100644
--- a/compute/zgeqrf.c
+++ b/compute/zgeqrf.c
@@ -296,8 +296,7 @@ int CHAMELEON_zgeqrf_Tile_Async( CHAM_desc_t *A, CHAM_desc_t *T,
CHAMELEON_Desc_Flush( T, sequence );
CHAMELEON_Desc_Flush( Dptr, sequence );
chameleon_sequence_wait( chamctxt, sequence );
- RUNTIME_desc_destroy( Dptr );
- chameleon_desc_mat_free( Dptr );
+ chameleon_desc_destroy( Dptr );
}
(void)D;
return CHAMELEON_SUCCESS;
diff --git a/compute/zgeqrf_param.c b/compute/zgeqrf_param.c
index 34171ce2229bd0858387acab5b27d769edab8927..36cb65d2ad8e391e8b6cd16f425b18e93ea9bf98 100644
--- a/compute/zgeqrf_param.c
+++ b/compute/zgeqrf_param.c
@@ -319,8 +319,7 @@ int CHAMELEON_zgeqrf_param_Tile_Async( const libhqr_tree_t *qrtree, CHAM_desc_t
CHAMELEON_Desc_Flush( TT, sequence );
CHAMELEON_Desc_Flush( Dptr, sequence );
chameleon_sequence_wait( chamctxt, sequence );
- RUNTIME_desc_destroy( Dptr );
- chameleon_desc_mat_free( Dptr );
+ chameleon_desc_destroy( Dptr );
}
(void)D;
return CHAMELEON_SUCCESS;
diff --git a/compute/zgeqrs.c b/compute/zgeqrs.c
index f38b05dbc5e9425d4271fe1a2d66d5dcbb439db8..ee2d2bc6d43636e7da152fd1b39ed0e91ebaa803 100644
--- a/compute/zgeqrs.c
+++ b/compute/zgeqrs.c
@@ -334,8 +334,7 @@ int CHAMELEON_zgeqrs_Tile_Async( CHAM_desc_t *A, CHAM_desc_t *T, CHAM_desc_t *B,
CHAMELEON_Desc_Flush( T, sequence );
CHAMELEON_Desc_Flush( Dptr, sequence );
chameleon_sequence_wait( chamctxt, sequence );
- RUNTIME_desc_destroy( Dptr );
- chameleon_desc_mat_free( Dptr );
+ chameleon_desc_destroy( Dptr );
}
(void)D;
return CHAMELEON_SUCCESS;
diff --git a/compute/zgeqrs_param.c b/compute/zgeqrs_param.c
index 79a6280f63f7b13e8cfd44377d9461cea19e741b..15a5ff0d64df8e8dc001aee59844fef3fcf34e40 100644
--- a/compute/zgeqrs_param.c
+++ b/compute/zgeqrs_param.c
@@ -335,8 +335,7 @@ int CHAMELEON_zgeqrs_param_Tile_Async( const libhqr_tree_t *qrtree,
CHAMELEON_Desc_Flush( TT, sequence );
CHAMELEON_Desc_Flush( Dptr, sequence );
chameleon_sequence_wait( chamctxt, sequence );
- RUNTIME_desc_destroy( Dptr );
- chameleon_desc_mat_free( Dptr );
+ chameleon_desc_destroy( Dptr );
}
(void)D;
return CHAMELEON_SUCCESS;
diff --git a/compute/zgesv_incpiv.c b/compute/zgesv_incpiv.c
index 0e0d330f86ee53bf2bbf5a6ceff678c9b2ab6985..6dd3073f6016555191f353a3b30492edc10743a7 100644
--- a/compute/zgesv_incpiv.c
+++ b/compute/zgesv_incpiv.c
@@ -331,8 +331,7 @@ int CHAMELEON_zgesv_incpiv_Tile_Async( CHAM_desc_t *A, CHAM_desc_t *L, int *IPIV
CHAMELEON_Desc_Flush( Dptr, sequence );
CHAMELEON_Desc_Flush( B, sequence );
chameleon_sequence_wait( chamctxt, sequence );
- RUNTIME_desc_destroy( Dptr );
- chameleon_desc_mat_free( Dptr );
+ chameleon_desc_destroy( Dptr );
}
(void)D;
return CHAMELEON_SUCCESS;
diff --git a/compute/zgesvd.c b/compute/zgesvd.c
index 05f72f68c58839e99ae0fe2581602d82473c2771..a9ba03d418296d6db8f7c78c0277f58f93be690d 100644
--- a/compute/zgesvd.c
+++ b/compute/zgesvd.c
@@ -471,7 +471,7 @@ int CHAMELEON_zgesvd_Tile_Async( cham_job_t jobu, cham_job_t jobvt,
#if defined(CHAMELEON_COPY_DIAG)
{
- chameleon_zdesc_alloc(D, A->mb, A->nb, A->m, A->n, 0, 0, A->m, A->n, );
+ chameleon_zdesc_alloc_diag( &D, A->mb, A->m, A->n, A->p, A->q );
Dptr = &D;
}
#endif
@@ -480,12 +480,12 @@ int CHAMELEON_zgesvd_Tile_Async( cham_job_t jobu, cham_job_t jobvt,
sequence, request );
/* Allocate band structure */
- chameleon_zdesc_alloc_diag( descAB,
- LDAB, NB,
- LDAB, MINMN,
- 0, 0,
- LDAB, MINMN,
- 1, 1 );
+ chameleon_zdesc_alloc( descAB,
+ LDAB, NB, /* mb, nb */
+ LDAB, N, /* lm, ln */
+ 0, 0, /* i, j */
+ LDAB, N, /* m, n */
+ );
/* Convert matrix to band form */
chameleon_pztile2band( uplo,
@@ -559,7 +559,7 @@ int CHAMELEON_zgesvd_Tile_Async( cham_job_t jobu, cham_job_t jobvt,
chameleon_sequence_wait( chamctxt, sequence );
#endif /* !defined(CHAMELEON_SIMULATION) */
- chameleon_desc_mat_free( &descAB );
+ chameleon_desc_destroy( &descAB );
subA = NULL;
subT = NULL;
@@ -640,8 +640,7 @@ int CHAMELEON_zgesvd_Tile_Async( cham_job_t jobu, cham_job_t jobvt,
free(E);
if ( Dptr ) {
- RUNTIME_desc_destroy( Dptr );
- chameleon_desc_mat_free( Dptr );
+ chameleon_desc_destroy( Dptr );
}
(void)D;
return CHAMELEON_SUCCESS;
diff --git a/compute/zgetrf_incpiv.c b/compute/zgetrf_incpiv.c
index 92f3a6ddc2509089ff5cde9b7aebd1a135662090..f990126ebc1a351e7164c4c3cebb114aa3d4c422 100644
--- a/compute/zgetrf_incpiv.c
+++ b/compute/zgetrf_incpiv.c
@@ -298,8 +298,7 @@ int CHAMELEON_zgetrf_incpiv_Tile_Async( CHAM_desc_t *A, CHAM_desc_t *L, int *IPI
CHAMELEON_Desc_Flush( L, sequence );
CHAMELEON_Desc_Flush( Dptr, sequence );
chameleon_sequence_wait( chamctxt, sequence );
- RUNTIME_desc_destroy( Dptr );
- chameleon_desc_mat_free( Dptr );
+ chameleon_desc_destroy( Dptr );
}
(void)D;
return CHAMELEON_SUCCESS;
diff --git a/compute/zheevd.c b/compute/zheevd.c
index a640fe54f9634b1da661a313f0fee0f3b8520eea..7f1a8b497ae733f1364976b1fdd5ad061b4e7b01 100644
--- a/compute/zheevd.c
+++ b/compute/zheevd.c
@@ -534,9 +534,8 @@ int CHAMELEON_zheevd_Tile_Async( cham_job_t jobz, cham_uplo_t uplo,
free(Q2);
free(V);
free(E);
- if (Dptr != NULL) {
- RUNTIME_desc_destroy( Dptr );
- chameleon_desc_mat_free( Dptr );
+ if ( Dptr != NULL ) {
+ chameleon_desc_destroy( Dptr );
}
(void)D;
return CHAMELEON_SUCCESS;
diff --git a/compute/zhetrd.c b/compute/zhetrd.c
index 39deedc70ac1cd6dd467af5499ca421104e521d2..0815e1dd689c39e660a729c133efdade2487e92a 100644
--- a/compute/zhetrd.c
+++ b/compute/zhetrd.c
@@ -392,8 +392,7 @@ int CHAMELEON_zhetrd_Tile_Async( cham_job_t jobz,
NB = descA.mb;
#if defined(CHAMELEON_COPY_DIAG)
{
- chameleon_zdesc_alloc_diag( D, A->mb, A->nb, chameleon_min(A->m, A->n), A->nb,
- 0, 0, chameleon_min(A->m, A->n), A->nb, A->p, A->q );
+ chameleon_zdesc_alloc_diag( &D, A->mb, A->m, A->n, A->p, A->q );
Dptr = &D;
}
#endif
@@ -404,12 +403,12 @@ int CHAMELEON_zhetrd_Tile_Async( cham_job_t jobz,
LDAB = NB+1;
/* Allocate band structure */
- chameleon_zdesc_alloc_diag( descAB,
- LDAB, NB, /* mb, nb */
- LDAB, N, /* lm, ln */
- 0, 0, /* i, j */
- LDAB, N, /* m, n */
- 1, 1 );
+ chameleon_zdesc_alloc( descAB,
+ LDAB, NB, /* mb, nb */
+ LDAB, N, /* lm, ln */
+ 0, 0, /* i, j */
+ LDAB, N, /* m, n */
+ );
/* Copy data into band structure */
chameleon_pztile2band( uplo, A, &descAB,
@@ -432,10 +431,9 @@ int CHAMELEON_zhetrd_Tile_Async( cham_job_t jobz,
}
#endif /* !defined(CHAMELEON_SIMULATION) */
if (Dptr != NULL) {
- RUNTIME_desc_destroy( Dptr );
- chameleon_desc_mat_free( Dptr );
+ chameleon_desc_destroy( Dptr );
}
- chameleon_desc_mat_free( &descAB );
+ chameleon_desc_destroy( &descAB );
(void)D;
return CHAMELEON_SUCCESS;
}
diff --git a/compute/ztile.c b/compute/ztile.c
index 8214c4c6141373c443c0a25d979fca976084752c..d95a729b7bb430b4fe99fcee7275fe4ef2055252 100644
--- a/compute/ztile.c
+++ b/compute/ztile.c
@@ -77,8 +77,8 @@ int CHAMELEON_zLapack_to_Tile( CHAMELEON_Complex64_t *Af77, int LDA, CHAM_desc_t
/* Create the B descriptor to handle the Lapack format matrix */
CHAMELEON_Desc_Create_User( &B, Af77, ChamComplexDouble, A->mb, A->nb, A->bsiz,
- LDA, A->n, 0, 0, A->m, A->n, 1, 1,
- chameleon_getaddr_cm, chameleon_getblkldd_cm, NULL );
+ LDA, A->n, 0, 0, A->m, A->n, 1, 1,
+ chameleon_getaddr_cm, chameleon_getblkldd_cm, NULL );
/* Start the computation */
chameleon_sequence_create( chamctxt, &sequence );
@@ -154,8 +154,8 @@ int CHAMELEON_zTile_to_Lapack( CHAM_desc_t *A, CHAMELEON_Complex64_t *Af77, int
/* Create the B descriptor to handle the Lapack format matrix */
CHAMELEON_Desc_Create_User( &B, Af77, ChamComplexDouble, A->mb, A->nb, A->bsiz,
- LDA, A->n, 0, 0, A->m, A->n, 1, 1,
- chameleon_getaddr_cm, chameleon_getblkldd_cm, NULL );
+ LDA, A->n, 0, 0, A->m, A->n, 1, 1,
+ chameleon_getaddr_cm, chameleon_getblkldd_cm, NULL );
/* Start the computation */
chameleon_sequence_create( chamctxt, &sequence );
diff --git a/compute/ztpgqrt.c b/compute/ztpgqrt.c
index 0512caeda802c5cb0b299486b053270a7a002bc4..4be3c82932a45caa1861f42068ca2d2b53c73616 100644
--- a/compute/ztpgqrt.c
+++ b/compute/ztpgqrt.c
@@ -426,8 +426,7 @@ int CHAMELEON_ztpgqrt_Tile_Async( int L,
CHAMELEON_Desc_Flush( Q2, sequence );
CHAMELEON_Desc_Flush( Dptr, sequence );
chameleon_sequence_wait( chamctxt, sequence );
- RUNTIME_desc_destroy( Dptr );
- chameleon_desc_mat_free( Dptr );
+ chameleon_desc_destroy( Dptr );
}
(void)D;
return CHAMELEON_SUCCESS;
diff --git a/compute/zunglq.c b/compute/zunglq.c
index 1851f381a1ac51b1522a6c5ea567caeab6a1add5..41015464634f910376b21d6d3f9df91dd83d4371 100644
--- a/compute/zunglq.c
+++ b/compute/zunglq.c
@@ -322,8 +322,7 @@ int CHAMELEON_zunglq_Tile_Async( CHAM_desc_t *A, CHAM_desc_t *T, CHAM_desc_t *Q,
CHAMELEON_Desc_Flush( T, sequence );
CHAMELEON_Desc_Flush( Dptr, sequence );
chameleon_sequence_wait( chamctxt, sequence );
- RUNTIME_desc_destroy( Dptr );
- chameleon_desc_mat_free( Dptr );
+ chameleon_desc_destroy( Dptr );
}
(void)D;
return CHAMELEON_SUCCESS;
diff --git a/compute/zunglq_param.c b/compute/zunglq_param.c
index a0571b4fa48ce53fccf6f1d6313ebd3b44873d1e..e6d36954211c8a931a7d8e9aa2b96a55c1953557 100644
--- a/compute/zunglq_param.c
+++ b/compute/zunglq_param.c
@@ -321,8 +321,7 @@ int CHAMELEON_zunglq_param_Tile_Async( const libhqr_tree_t *qrtree, CHAM_desc_t
CHAMELEON_Desc_Flush( TT, sequence );
CHAMELEON_Desc_Flush( Dptr, sequence );
chameleon_sequence_wait( chamctxt, sequence );
- RUNTIME_desc_destroy( Dptr );
- chameleon_desc_mat_free( Dptr );
+ chameleon_desc_destroy( Dptr );
}
(void)D;
return CHAMELEON_SUCCESS;
diff --git a/compute/zungqr.c b/compute/zungqr.c
index 4f15a5a1695b97adf7f1fde0a34aef0750af10b0..c51539616760f3c45ceecf2b1f631846fc5443fc 100644
--- a/compute/zungqr.c
+++ b/compute/zungqr.c
@@ -319,8 +319,7 @@ int CHAMELEON_zungqr_Tile_Async( CHAM_desc_t *A, CHAM_desc_t *T, CHAM_desc_t *Q,
CHAMELEON_Desc_Flush( T, sequence );
CHAMELEON_Desc_Flush( Dptr, sequence );
chameleon_sequence_wait( chamctxt, sequence );
- RUNTIME_desc_destroy( Dptr );
- chameleon_desc_mat_free( Dptr );
+ chameleon_desc_destroy( Dptr );
}
(void)D;
return CHAMELEON_SUCCESS;
diff --git a/compute/zungqr_param.c b/compute/zungqr_param.c
index 0f17560c32b96be73b8cdba829afbb282e537433..5b46d66ff7626ee058fb7119d0ee344728509a3e 100644
--- a/compute/zungqr_param.c
+++ b/compute/zungqr_param.c
@@ -328,8 +328,7 @@ int CHAMELEON_zungqr_param_Tile_Async( const libhqr_tree_t *qrtree, CHAM_desc_t
CHAMELEON_Desc_Flush( TT, sequence );
CHAMELEON_Desc_Flush( Dptr, sequence );
chameleon_sequence_wait( chamctxt, sequence );
- RUNTIME_desc_destroy( Dptr );
- chameleon_desc_mat_free( Dptr );
+ chameleon_desc_destroy( Dptr );
}
(void)D;
return CHAMELEON_SUCCESS;
diff --git a/compute/zunmlq.c b/compute/zunmlq.c
index 896c5d47d34c32cf494e9b84a409eca7b155933b..f460e12e4b353fa1746f03a9bfd4b9706e8f564b 100644
--- a/compute/zunmlq.c
+++ b/compute/zunmlq.c
@@ -381,8 +381,7 @@ int CHAMELEON_zunmlq_Tile_Async( cham_side_t side, cham_trans_t trans,
CHAMELEON_Desc_Flush( T, sequence );
CHAMELEON_Desc_Flush( Dptr, sequence );
chameleon_sequence_wait( chamctxt, sequence );
- RUNTIME_desc_destroy( Dptr );
- chameleon_desc_mat_free( Dptr );
+ chameleon_desc_destroy( Dptr );
}
(void)D;
return CHAMELEON_SUCCESS;
diff --git a/compute/zunmlq_param.c b/compute/zunmlq_param.c
index 6ca36dff3e1872eca47a27111914dcbe465f77fa..4c0a72358b357b1c3e83253c9dc0844606e2bea4 100644
--- a/compute/zunmlq_param.c
+++ b/compute/zunmlq_param.c
@@ -382,8 +382,7 @@ int CHAMELEON_zunmlq_param_Tile_Async( const libhqr_tree_t *qrtree, cham_side_t
CHAMELEON_Desc_Flush( TT, sequence );
CHAMELEON_Desc_Flush( Dptr, sequence );
chameleon_sequence_wait( chamctxt, sequence );
- RUNTIME_desc_destroy( Dptr );
- chameleon_desc_mat_free( Dptr );
+ chameleon_desc_destroy( Dptr );
}
(void)D;
return CHAMELEON_SUCCESS;
diff --git a/compute/zunmqr.c b/compute/zunmqr.c
index 0ddbaa8537774f8640e46dab2b3b8badad3fbb0a..6271ed98eeedbcb23bb88ff909ad303fa8d70c42 100644
--- a/compute/zunmqr.c
+++ b/compute/zunmqr.c
@@ -384,8 +384,7 @@ int CHAMELEON_zunmqr_Tile_Async( cham_side_t side, cham_trans_t trans,
CHAMELEON_Desc_Flush( T, sequence );
CHAMELEON_Desc_Flush( Dptr, sequence );
chameleon_sequence_wait( chamctxt, sequence );
- RUNTIME_desc_destroy( Dptr );
- chameleon_desc_mat_free( Dptr );
+ chameleon_desc_destroy( Dptr );
}
(void)D;
return CHAMELEON_SUCCESS;
diff --git a/compute/zunmqr_param.c b/compute/zunmqr_param.c
index a8e6cc5920f972eb3d12a8e5804b9a521606e23b..5674ba090c9ceb0c581b383993b58a8ec0fb6573 100644
--- a/compute/zunmqr_param.c
+++ b/compute/zunmqr_param.c
@@ -389,8 +389,7 @@ int CHAMELEON_zunmqr_param_Tile_Async( const libhqr_tree_t *qrtree,
CHAMELEON_Desc_Flush( TT, sequence );
CHAMELEON_Desc_Flush( Dptr, sequence );
chameleon_sequence_wait( chamctxt, sequence );
- RUNTIME_desc_destroy( Dptr );
- chameleon_desc_mat_free( Dptr );
+ chameleon_desc_destroy( Dptr );
}
(void)D;
return CHAMELEON_SUCCESS;
diff --git a/control/compute_z.h b/control/compute_z.h
index 3682d99cc71356fd4ad5100ca7e91d0208a86356..3229f1389f86287bdd0fd355f3934c25f67a99ab 100644
--- a/control/compute_z.h
+++ b/control/compute_z.h
@@ -25,34 +25,6 @@
#ifndef _compute_z_h_
#define _compute_z_h_
-/**
- * LAPACK/Tile Descriptor accesses
- */
-#define ChamDescInput 1
-#define ChamDescOutput 2
-#define ChamDescInout (ChamDescInput | ChamDescOutput)
-
-/**
- * Macro for matrix conversion / Lapack interface
- */
-#define chameleon_zdesc_alloc_diag( descA, mb, nb, lm, ln, i, j, m, n, p, q) \
- descA = chameleon_desc_init_diag( \
- ChamComplexDouble, (mb), (nb), ((mb)*(nb)), \
- (m), (n), (i), (j), (m), (n), p, q); \
- chameleon_desc_mat_alloc( &(descA) ); \
- RUNTIME_desc_create( &(descA) );
-
-#define chameleon_zdesc_alloc( descA, mb, nb, lm, ln, i, j, m, n, free) \
- descA = chameleon_desc_init( \
- ChamComplexDouble, (mb), (nb), ((mb)*(nb)), \
- (m), (n), (i), (j), (m), (n), 1, 1); \
- if ( chameleon_desc_mat_alloc( &(descA) ) ) { \
- chameleon_error( __func__, "chameleon_desc_mat_alloc() failed"); \
- {free;}; \
- return CHAMELEON_ERR_OUT_OF_RESOURCES; \
- } \
- RUNTIME_desc_create( &(descA) );
-
/**
* Declarations of internal sequential functions
*/
@@ -148,6 +120,41 @@ void chameleon_pzungqr_param( int genD, int K, const libhqr_tree_t *qrtree,
RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
+
+/**
+ * LAPACK/Tile Descriptor accesses
+ */
+#define ChamDescInput 1
+#define ChamDescOutput 2
+#define ChamDescInout (ChamDescInput | ChamDescOutput)
+
+/**
+ * Macro for matrix conversion / Lapack interface
+ */
+static inline int
+chameleon_zdesc_alloc_diag( CHAM_desc_t *descA, int nb, int m, int n, int p, int q ) {
+ int diag_m = chameleon_min( m, n );
+ return chameleon_desc_init( descA, CHAMELEON_MAT_ALLOC_GLOBAL,
+ ChamComplexDouble, nb, nb, nb*nb,
+ diag_m, nb, 0, 0, diag_m, nb, p, q,
+ chameleon_getaddr_diag,
+ chameleon_getblkldd_ccrb,
+ chameleon_getrankof_2d_diag );
+}
+
+#define chameleon_zdesc_alloc( descA, mb, nb, lm, ln, i, j, m, n, free) \
+ { \
+ int rc; \
+ rc = chameleon_desc_init( &(descA), CHAMELEON_MAT_ALLOC_GLOBAL, \
+ ChamComplexDouble, (mb), (nb), ((mb)*(nb)), \
+ (m), (n), (i), (j), (m), (n), 1, 1, \
+ NULL, NULL, NULL ); \
+ if ( rc != CHAMELEON_SUCCESS ) { \
+ {free;} \
+ return rc; \
+ } \
+ }
+
/**
* @brief Internal function to convert the lapack format to tile format in
* LAPACK interface calls
@@ -160,35 +167,28 @@ chameleon_zlap2tile( CHAM_context_t *chamctxt,
RUNTIME_sequence_t *seq, RUNTIME_request_t *req )
{
/* Initialize the Lapack descriptor */
- *descAl = chameleon_desc_init_user( ChamComplexDouble, mb, nb, (mb)*(nb),
- lm, ln, 0, 0, m, n, 1, 1,
- chameleon_getaddr_cm, chameleon_getblkldd_cm, NULL );
- descAl->mat = A;
+ chameleon_desc_init( descAl, A, ChamComplexDouble, mb, nb, (mb)*(nb),
+ lm, ln, 0, 0, m, n, 1, 1,
+ chameleon_getaddr_cm, chameleon_getblkldd_cm, NULL );
descAl->styp = ChamCM;
- /* Initialize the tile descriptor */
- *descAt = chameleon_desc_init( ChamComplexDouble, mb, nb, (mb)*(nb),
- lm, ln, 0, 0, m, n, 1, 1 );
-
if ( CHAMELEON_TRANSLATION == ChamOutOfPlace ) {
- if ( chameleon_desc_mat_alloc( descAt ) ) {
- chameleon_error( "chameleon_zlap2tile", "chameleon_desc_mat_alloc() failed");
- return CHAMELEON_ERR_OUT_OF_RESOURCES;
- }
-
- RUNTIME_desc_create( descAl );
- RUNTIME_desc_create( descAt );
+ /* Initialize the tile descriptor */
+ chameleon_desc_init( descAt, CHAMELEON_MAT_ALLOC_GLOBAL, ChamComplexDouble, mb, nb, (mb)*(nb),
+ lm, ln, 0, 0, m, n, 1, 1,
+ chameleon_getaddr_ccrb, chameleon_getblkldd_ccrb, NULL );
if ( mode & ChamDescInput ) {
chameleon_pzlacpy( uplo, descAl, descAt, seq, req );
}
}
else {
- chameleon_fatal_error( "chameleon_zlap2tile", "INPLACE translation not supported yet");
- descAt->mat = A;
+ /* Initialize the tile descriptor */
+ chameleon_desc_init( descAt, A, ChamComplexDouble, mb, nb, (mb)*(nb),
+ lm, ln, 0, 0, m, n, 1, 1,
+ chameleon_getaddr_ccrb, chameleon_getblkldd_ccrb, NULL );
- RUNTIME_desc_create( descAl );
- RUNTIME_desc_create( descAt );
+ chameleon_fatal_error( "chameleon_zlap2tile", "INPLACE translation not supported yet");
if ( mode & ChamDescInput ) {
/* CHAMELEON_zgecfi_Async( lm, ln, A, ChamCM, mb, nb, */
@@ -235,11 +235,8 @@ chameleon_ztile2lap( CHAM_context_t *chamctxt, CHAM_desc_t *descAl, CHAM_desc_t
static inline void
chameleon_ztile2lap_cleanup( CHAM_context_t *chamctxt, CHAM_desc_t *descAl, CHAM_desc_t *descAt )
{
- if ( CHAMELEON_TRANSLATION == ChamOutOfPlace ) {
- chameleon_desc_mat_free( descAt );
- }
- RUNTIME_desc_destroy( descAl );
- RUNTIME_desc_destroy( descAt );
+ chameleon_desc_destroy( descAl );
+ chameleon_desc_destroy( descAt );
}
#endif /* _compute_z_h_ */
diff --git a/control/descriptor.c b/control/descriptor.c
index 7216518491ea24858167d648ab3ff69764b385ff..c27fe5749f2164bd558fedffc9d9f0ec9208552e 100644
--- a/control/descriptor.c
+++ b/control/descriptor.c
@@ -31,12 +31,49 @@
static int nbdesc = 0;
+/**
+ *
+ */
+int chameleon_desc_mat_alloc( CHAM_desc_t *desc )
+{
+ size_t size = (size_t)(desc->llm) * (size_t)(desc->lln)
+ * (size_t)CHAMELEON_Element_Size(desc->dtyp);
+ if ((desc->mat = RUNTIME_malloc(size)) == NULL) {
+ chameleon_error("chameleon_desc_mat_alloc", "malloc() failed");
+ return CHAMELEON_ERR_OUT_OF_RESOURCES;
+ }
+
+ /* The matrix has already been registered by the Runtime alloc */
+ desc->register_mat = 0;
+
+ return CHAMELEON_SUCCESS;
+}
+
+/**
+ *
+ */
+int chameleon_desc_mat_free( CHAM_desc_t *desc )
+{
+ if ( (desc->mat != NULL) &&
+ (desc->use_mat == 1 ) &&
+ (desc->alloc_mat == 1 ) )
+ {
+ size_t size = (size_t)(desc->llm) * (size_t)(desc->lln)
+ * (size_t)CHAMELEON_Element_Size(desc->dtyp);
+
+ RUNTIME_free(desc->mat, size);
+ desc->mat = NULL;
+ }
+
+ return CHAMELEON_SUCCESS;
+}
+
/**
******************************************************************************
*
* @ingroup Descriptor
*
- * chameleon_desc_init_user - Internal function to create tiled matrix descriptor
+ * @brief Internal function to create tiled matrix descriptor
* with generic function for data distribution and storage format.
*
******************************************************************************
@@ -96,121 +133,132 @@ static int nbdesc = 0;
* @return The descriptor with the matrix description parameters set.
*
*/
-CHAM_desc_t chameleon_desc_init_user(cham_flttype_t dtyp, int mb, int nb, int bsiz,
- int lm, int ln, int i, int j,
- int m, int n, int p, int q,
- void* (*get_blkaddr)( const CHAM_desc_t*, int, int ),
- int (*get_blkldd) ( const CHAM_desc_t*, int ),
- int (*get_rankof) ( const CHAM_desc_t*, int, int ))
+int chameleon_desc_init( CHAM_desc_t *desc, void *mat,
+ cham_flttype_t dtyp, int mb, int nb, int bsiz,
+ int lm, int ln, int i, int j,
+ int m, int n, int p, int q,
+ void* (*get_blkaddr)( const CHAM_desc_t*, int, int ),
+ int (*get_blkldd) ( const CHAM_desc_t*, int ),
+ int (*get_rankof) ( const CHAM_desc_t*, int, int ) )
{
CHAM_context_t *chamctxt;
- CHAM_desc_t desc;
+ int rc = CHAMELEON_SUCCESS;
- memset( &desc, 0, sizeof(CHAM_desc_t) );
+ memset( desc, 0, sizeof(CHAM_desc_t) );
chamctxt = chameleon_context_self();
if (chamctxt == NULL) {
chameleon_error("CHAMELEON_Desc_Create", "CHAMELEON not initialized");
- return desc;
+ return CHAMELEON_ERR_NOT_INITIALIZED;
}
// If one of the function get_* is NULL, we switch back to the default, like in chameleon_desc_init()
- desc.get_blkaddr = get_blkaddr ? get_blkaddr : chameleon_getaddr_ccrb;
- desc.get_blkldd = get_blkldd ? get_blkldd : chameleon_getblkldd_ccrb;
- desc.get_rankof = get_rankof ? get_rankof : chameleon_getrankof_2d;
+ desc->get_blkaddr = get_blkaddr ? get_blkaddr : chameleon_getaddr_ccrb;
+ desc->get_blkldd = get_blkldd ? get_blkldd : chameleon_getblkldd_ccrb;
+ desc->get_rankof = get_rankof ? get_rankof : chameleon_getrankof_2d;
// Matrix properties
- desc.dtyp = dtyp;
+ desc->dtyp = dtyp;
// Should be given as parameter to follow get_blkaddr (unused)
- desc.styp = ChamCCRB;
- desc.mb = mb;
- desc.nb = nb;
- desc.bsiz = bsiz;
+ desc->styp = ChamCCRB;
+ desc->mb = mb;
+ desc->nb = nb;
+ desc->bsiz = bsiz;
// Large matrix parameters
- desc.lm = lm;
- desc.ln = ln;
+ desc->lm = lm;
+ desc->ln = ln;
// Large matrix derived parameters
- desc.lmt = (lm%mb==0) ? (lm/mb) : (lm/mb+1);
- desc.lnt = (ln%nb==0) ? (ln/nb) : (ln/nb+1);
+ desc->lmt = (lm%mb==0) ? (lm/mb) : (lm/mb+1);
+ desc->lnt = (ln%nb==0) ? (ln/nb) : (ln/nb+1);
// Submatrix parameters
- desc.i = i;
- desc.j = j;
- desc.m = m;
- desc.n = n;
+ desc->i = i;
+ desc->j = j;
+ desc->m = m;
+ desc->n = n;
// Submatrix derived parameters
- desc.mt = (m == 0) ? 0 : (i+m-1)/mb - i/mb + 1;
- desc.nt = (n == 0) ? 0 : (j+n-1)/nb - j/nb + 1;
+ desc->mt = (m == 0) ? 0 : (i+m-1)/mb - i/mb + 1;
+ desc->nt = (n == 0) ? 0 : (j+n-1)/nb - j/nb + 1;
- desc.id = nbdesc; nbdesc++;
- desc.occurences = 0;
- desc.use_mat = 1;
- desc.alloc_mat = 1;
- desc.register_mat = (chamctxt->ncudas > 0) ? 1 : 0;
- desc.ooc = 0;
+ desc->id = nbdesc;
+ nbdesc++;
+ desc->occurences = 0;
- desc.myrank = RUNTIME_comm_rank( chamctxt );
+ desc->myrank = RUNTIME_comm_rank( chamctxt );
// Grid size
- desc.p = p;
- desc.q = q;
+ desc->p = p;
+ desc->q = q;
// Local dimensions in tiles
- if ( desc.myrank < (p*q) ) {
- desc.llmt = (desc.lmt + p - 1) / p;
- desc.llnt = (desc.lnt + q - 1) / q;
+ if ( desc->myrank < (p*q) ) {
+ desc->llmt = (desc->lmt + p - 1) / p;
+ desc->llnt = (desc->lnt + q - 1) / q;
// Local dimensions
- if ( ((desc.lmt-1) % p) == (desc.myrank / q) ) {
- desc.llm = ( desc.llmt - 1 ) * mb + ((lm%mb==0) ? mb : (lm%mb));
+ if ( ((desc->lmt-1) % p) == (desc->myrank / q) ) {
+ desc->llm = ( desc->llmt - 1 ) * mb + ((lm%mb==0) ? mb : (lm%mb));
} else {
- desc.llm = desc.llmt * mb;
+ desc->llm = desc->llmt * mb;
}
- if ( ((desc.lnt-1) % q) == (desc.myrank % q) ) {
- desc.lln = ( desc.llnt - 1 ) * nb + ((ln%nb==0) ? nb : (ln%nb));
+ if ( ((desc->lnt-1) % q) == (desc->myrank % q) ) {
+ desc->lln = ( desc->llnt - 1 ) * nb + ((ln%nb==0) ? nb : (ln%nb));
} else {
- desc.lln = desc.llnt * nb;
+ desc->lln = desc->llnt * nb;
}
- desc.llm1 = (desc.llm/mb);
- desc.lln1 = (desc.lln/nb);
+ desc->llm1 = (desc->llm/mb);
+ desc->lln1 = (desc->lln/nb);
} else {
- desc.llmt = 0;
- desc.llnt = 0;
- desc.llm = 0;
- desc.lln = 0;
- desc.llm1 = 0;
- desc.lln1 = 0;
+ desc->llmt = 0;
+ desc->llnt = 0;
+ desc->llm = 0;
+ desc->lln = 0;
+ desc->llm1 = 0;
+ desc->lln1 = 0;
}
+ /* memory of the matrix is handled by the user */
+ desc->alloc_mat = 0;
+ /* if the user gives a pointer to the overall data (tiles) we can use it */
+ desc->use_mat = 0;
+ /* users data can have multiple forms: let him register tiles */
+ desc->register_mat = 0;
+ /* The matrix is alocated tile by tile with out of core */
+ desc->ooc = 0;
+
// Matrix address
- desc.mat = NULL;
- desc.A21 = (size_t)(desc.llm - desc.llm%mb)*(size_t)(desc.lln - desc.lln%nb);
- desc.A12 = (size_t)( desc.llm%mb)*(size_t)(desc.lln - desc.lln%nb) + desc.A21;
- desc.A22 = (size_t)(desc.llm - desc.llm%mb)*(size_t)( desc.lln%nb) + desc.A12;
+ if ( mat == CHAMELEON_MAT_ALLOC_GLOBAL ) {
+ rc = chameleon_desc_mat_alloc( desc );
- return desc;
-}
+ desc->alloc_mat = 1;
+ desc->use_mat = 1;
+ }
+ else if ( mat == CHAMELEON_MAT_ALLOC_TILE ) {
+ //chameleon_error( "chameleon_desc_init", "CHAMELEON_MAT_ALLOC_TILE is not available yet" );
+ //desc->mat = NULL;
+ rc = chameleon_desc_mat_alloc( desc );
+ desc->use_mat = 1;
-/**
- * Internal static descriptor initializer
- */
-CHAM_desc_t chameleon_desc_init(cham_flttype_t dtyp, int mb, int nb, int bsiz,
- int lm, int ln, int i, int j,
- int m, int n, int p, int q)
-{
- return chameleon_desc_init_user(dtyp, mb, nb, bsiz, lm, ln, i, j, m, n, p, q,
- chameleon_getaddr_ccrb, chameleon_getblkldd_ccrb, chameleon_getrankof_2d);
-}
+ desc->alloc_mat = 1;
+ }
+ else if ( mat == CHAMELEON_MAT_OOC ) {
+ desc->mat = NULL;
+ desc->ooc = 1;
+ }
+ else {
+ /* memory of the matrix is handled by users */
+ desc->mat = mat;
+ desc->use_mat = 1;
+ }
-/**
- * Internal static descriptor initializer for a block diagonal matrix
- */
-CHAM_desc_t chameleon_desc_init_diag(cham_flttype_t dtyp, int mb, int nb, int bsiz,
- int lm, int ln, int i, int j,
- int m, int n, int p, int q)
-{
- return chameleon_desc_init_user(dtyp, mb, nb, bsiz, lm, ln, i, j, m, n, p, q,
- chameleon_getaddr_ccrb, chameleon_getblkldd_ccrb, chameleon_getrankof_2d_diag);
+ desc->A21 = (size_t)(desc->llm - desc->llm%mb)*(size_t)(desc->lln - desc->lln%nb);
+ desc->A12 = (size_t)( desc->llm%mb)*(size_t)(desc->lln - desc->lln%nb) + desc->A21;
+ desc->A22 = (size_t)(desc->llm - desc->llm%mb)*(size_t)( desc->lln%nb) + desc->A12;
+
+ /* Create runtime specific structure like registering data */
+ RUNTIME_desc_create( desc );
+
+ return rc;
}
/**
@@ -248,6 +296,12 @@ CHAM_desc_t* chameleon_desc_submatrix(CHAM_desc_t *descA, int i, int j, int m, i
return descB;
}
+void chameleon_desc_destroy( CHAM_desc_t *desc )
+{
+ RUNTIME_desc_destroy( desc );
+ chameleon_desc_mat_free( desc );
+}
+
/**
* Check for descriptor correctness
*/
@@ -295,42 +349,6 @@ int chameleon_desc_check(const CHAM_desc_t *desc)
return CHAMELEON_SUCCESS;
}
-/**
- *
- */
-int chameleon_desc_mat_alloc( CHAM_desc_t *desc )
-{
- size_t size = (size_t)(desc->llm) * (size_t)(desc->lln)
- * (size_t)CHAMELEON_Element_Size(desc->dtyp);
- if ((desc->mat = RUNTIME_malloc(size)) == NULL) {
- chameleon_error("chameleon_desc_mat_alloc", "malloc() failed");
- return CHAMELEON_ERR_OUT_OF_RESOURCES;
- }
-
- /* The matrix has already been registered by the Runtime alloc */
- desc->register_mat = 0;
-
- return CHAMELEON_SUCCESS;
-}
-
-/**
- *
- */
-int chameleon_desc_mat_free( CHAM_desc_t *desc )
-{
- if ( (desc->mat != NULL) &&
- (desc->use_mat == 1 ) &&
- (desc->alloc_mat == 1 ) )
- {
- size_t size = (size_t)(desc->llm) * (size_t)(desc->lln)
- * (size_t)CHAMELEON_Element_Size(desc->dtyp);
-
- RUNTIME_free(desc->mat, size);
- desc->mat = NULL;
- }
- return CHAMELEON_SUCCESS;
-}
-
/**
*****************************************************************************
*
@@ -393,63 +411,12 @@ int chameleon_desc_mat_free( CHAM_desc_t *desc )
* \retval CHAMELEON_SUCCESS successful exit
*
*/
-int CHAMELEON_Desc_Create(CHAM_desc_t **descptr, void *mat, cham_flttype_t dtyp, int mb, int nb, int bsiz,
- int lm, int ln, int i, int j, int m, int n, int p, int q)
+int CHAMELEON_Desc_Create( CHAM_desc_t **descptr, void *mat, cham_flttype_t dtyp, int mb, int nb, int bsiz,
+ int lm, int ln, int i, int j, int m, int n, int p, int q )
{
- CHAM_context_t *chamctxt;
- CHAM_desc_t *desc;
- int status;
-
- *descptr = NULL;
-
- chamctxt = chameleon_context_self();
- if (chamctxt == NULL) {
- chameleon_error("CHAMELEON_Desc_Create", "CHAMELEON not initialized");
- return CHAMELEON_ERR_NOT_INITIALIZED;
- }
-
- /* Allocate memory and initialize the descriptor */
- desc = (CHAM_desc_t*)malloc(sizeof(CHAM_desc_t));
- if (desc == NULL) {
- chameleon_error("CHAMELEON_Desc_Create", "malloc() failed");
- return CHAMELEON_ERR_OUT_OF_RESOURCES;
- }
- *desc = chameleon_desc_init(dtyp, mb, nb, bsiz, lm, ln, i, j, m, n, p, q);
-
- if (mat == NULL) {
-
- size_t size = (size_t)(desc->llm) * (size_t)(desc->lln)
- * (size_t)CHAMELEON_Element_Size(desc->dtyp);
-
- if ((desc->mat = RUNTIME_malloc(size)) == NULL) {
- chameleon_error("CHAMELEON_Desc_Create", "malloc() failed");
- free(desc);
- return CHAMELEON_ERR_OUT_OF_RESOURCES;
- }
- desc->use_mat = 1;
- desc->alloc_mat = 1;
- desc->register_mat = 0;
-
- } else {
- desc->mat = mat;
- /* memory of the matrix is handled by users */
- desc->alloc_mat = 0;
- desc->use_mat = 1;
- desc->register_mat = 0;
- }
-
- /* Create scheduler structure like registering data */
- RUNTIME_desc_create( desc );
-
- status = chameleon_desc_check( desc );
- if (status != CHAMELEON_SUCCESS) {
- chameleon_error("CHAMELEON_Desc_Create", "invalid descriptor");
- CHAMELEON_Desc_Destroy( &desc );
- return status;
- }
-
- *descptr = desc;
- return CHAMELEON_SUCCESS;
+ return CHAMELEON_Desc_Create_User( descptr, mat, dtyp, mb, nb, bsiz,
+ lm, ln, i, j, m, n, p, q,
+ NULL, NULL, NULL );
}
/**
@@ -507,11 +474,11 @@ int CHAMELEON_Desc_Create(CHAM_desc_t **descptr, void *mat, cham_flttype_t dtyp,
* \retval CHAMELEON_SUCCESS successful exit
*
*/
-int CHAMELEON_Desc_Create_User(CHAM_desc_t **descptr, void *mat, cham_flttype_t dtyp, int mb, int nb, int bsiz,
- int lm, int ln, int i, int j, int m, int n, int p, int q,
- void* (*get_blkaddr)( const CHAM_desc_t*, int, int ),
- int (*get_blkldd) ( const CHAM_desc_t*, int ),
- int (*get_rankof) ( const CHAM_desc_t*, int, int ))
+int CHAMELEON_Desc_Create_User( CHAM_desc_t **descptr, void *mat, cham_flttype_t dtyp, int mb, int nb, int bsiz,
+ int lm, int ln, int i, int j, int m, int n, int p, int q,
+ void* (*get_blkaddr)( const CHAM_desc_t*, int, int ),
+ int (*get_blkldd) ( const CHAM_desc_t*, int ),
+ int (*get_rankof) ( const CHAM_desc_t*, int, int ) )
{
CHAM_context_t *chamctxt;
CHAM_desc_t *desc;
@@ -532,22 +499,9 @@ int CHAMELEON_Desc_Create_User(CHAM_desc_t **descptr, void *mat, cham_flttype_t
return CHAMELEON_ERR_OUT_OF_RESOURCES;
}
- *desc = chameleon_desc_init_user(dtyp, mb, nb, bsiz, lm, ln, i, j, m, n, p, q,
- get_blkaddr, get_blkldd, get_rankof);
-
- /* if the user gives a pointer to the overall data (tiles) we can use it */
- desc->use_mat = (mat == NULL) ? 0 : 1;
-
- /* memory of the matrix is handled by the user */
- desc->alloc_mat = 0;
-
- /* users data can have multiple forms: let him register tiles */
- desc->register_mat = 0;
-
- desc->mat = mat;
-
- /* Create runtime specific structure like registering data */
- RUNTIME_desc_create( desc );
+ chameleon_desc_init( desc, mat, dtyp, mb, nb, bsiz,
+ lm, ln, i, j, m, n, p, q,
+ get_blkaddr, get_blkldd, get_rankof );
status = chameleon_desc_check( desc );
if (status != CHAMELEON_SUCCESS) {
@@ -616,47 +570,11 @@ int CHAMELEON_Desc_Create_OOC_User(CHAM_desc_t **descptr, cham_flttype_t dtyp, i
chameleon_error("CHAMELEON_Desc_Create_OOC_User", "Only StarPU supports on-demand tile allocation");
return CHAMELEON_ERR_NOT_SUPPORTED;
#else
- CHAM_context_t *chamctxt;
- CHAM_desc_t *desc;
- int status;
-
- *descptr = NULL;
-
- chamctxt = chameleon_context_self();
- if (chamctxt == NULL) {
- chameleon_error("CHAMELEON_Desc_Create_OOC_User", "CHAMELEON not initialized");
- return CHAMELEON_ERR_NOT_INITIALIZED;
- }
-
- /* Allocate memory and initialize the descriptor */
- desc = (CHAM_desc_t*)malloc(sizeof(CHAM_desc_t));
- if ( desc == NULL ) {
- chameleon_error("CHAMELEON_Desc_Create_OOC_User", "malloc() failed");
- return CHAMELEON_ERR_OUT_OF_RESOURCES;
- }
- *desc = chameleon_desc_init_user( dtyp, mb, nb, bsiz, lm, ln, i, j, m, n, p, q,
- chameleon_getaddr_null, NULL, get_rankof );
-
- /* memory of the matrix is completely handled by runtime */
- desc->use_mat = 0;
- desc->alloc_mat = 0;
- desc->register_mat = 0;
-
- desc->mat = NULL;
- desc->ooc = 1;
-
- /* Create scheduler structure like registering data */
- RUNTIME_desc_create( desc );
-
- status = chameleon_desc_check( desc );
- if (status != CHAMELEON_SUCCESS) {
- chameleon_error("CHAMELEON_Desc_Create_OOC_User", "invalid descriptor");
- CHAMELEON_Desc_Destroy( &desc );
- return status;
- }
-
- *descptr = desc;
- return CHAMELEON_SUCCESS;
+ int rc;
+ rc = CHAMELEON_Desc_Create_User( descptr, CHAMELEON_MAT_OOC, dtyp, mb, nb, bsiz,
+ lm, ln, i, j, m, n, p, q,
+ chameleon_getaddr_null, NULL, get_rankof );
+ return rc;
#endif
}
@@ -704,9 +622,9 @@ int CHAMELEON_Desc_Create_OOC_User(CHAM_desc_t **descptr, cham_flttype_t dtyp, i
int CHAMELEON_Desc_Create_OOC(CHAM_desc_t **descptr, cham_flttype_t dtyp, int mb, int nb, int bsiz,
int lm, int ln, int i, int j, int m, int n, int p, int q)
{
- return CHAMELEON_Desc_Create_OOC_User( descptr, dtyp, mb, nb, bsiz,
- lm, ln, i, j, m, n, p, q,
- chameleon_getrankof_2d );
+ return CHAMELEON_Desc_Create_User( descptr, CHAMELEON_MAT_OOC, dtyp, mb, nb, bsiz,
+ lm, ln, i, j, m, n, p, q,
+ chameleon_getaddr_null, NULL, NULL );
}
/**
@@ -741,8 +659,7 @@ int CHAMELEON_Desc_Destroy(CHAM_desc_t **desc)
return CHAMELEON_ERR_UNALLOCATED;
}
- RUNTIME_desc_destroy( *desc );
- chameleon_desc_mat_free( *desc );
+ chameleon_desc_destroy( *desc );
free(*desc);
*desc = NULL;
return CHAMELEON_SUCCESS;
diff --git a/control/descriptor.h b/control/descriptor.h
index dcf688df2eb0e85b564ca3bc2b0899ed4cee2f18..c3548c966ccb254287dc524dab723b12a705ad6d 100644
--- a/control/descriptor.h
+++ b/control/descriptor.h
@@ -37,6 +37,7 @@ inline static void* chameleon_geteltaddr(const CHAM_desc_t *A, int m, int n, int
inline static void* chameleon_getaddr_cm (const CHAM_desc_t *A, int m, int n);
inline static void* chameleon_getaddr_ccrb (const CHAM_desc_t *A, int m, int n);
inline static void* chameleon_getaddr_null (const CHAM_desc_t *A, int m, int n);
+inline static void* chameleon_getaddr_diag (const CHAM_desc_t *A, int m, int n);
inline static int chameleon_getblkldd_cm (const CHAM_desc_t *A, int m);
inline static int chameleon_getblkldd_ccrb(const CHAM_desc_t *A, int m);
@@ -46,21 +47,16 @@ inline static int chameleon_getblkldd_ccrb(const CHAM_desc_t *A, int m);
inline static int chameleon_getrankof_2d(const CHAM_desc_t *desc, int m, int n);
inline static int chameleon_getrankof_2d_diag(const CHAM_desc_t *desc, int m, int n);
-CHAM_desc_t chameleon_desc_init(cham_flttype_t dtyp, int mb, int nb, int bsiz,
- int lm, int ln, int i, int j, int m, int n, int p, int q);
-CHAM_desc_t chameleon_desc_init_diag(cham_flttype_t dtyp, int mb, int nb, int bsiz,
- int lm, int ln, int i, int j, int m, int n, int p, int q);
-CHAM_desc_t chameleon_desc_init_user(cham_flttype_t dtyp, int mb, int nb, int bsiz,
- int lm, int ln, int i, int j,
- int m, int n, int p, int q,
- void* (*get_blkaddr)( const CHAM_desc_t*, int, int ),
- int (*get_blkldd)( const CHAM_desc_t*, int ),
- int (*get_rankof)( const CHAM_desc_t*, int, int ));
-CHAM_desc_t* chameleon_desc_submatrix(CHAM_desc_t *descA, int i, int j, int m, int n);
-
-int chameleon_desc_check (const CHAM_desc_t *desc);
-int chameleon_desc_mat_alloc(CHAM_desc_t *desc);
-int chameleon_desc_mat_free (CHAM_desc_t *desc);
+int chameleon_desc_init ( CHAM_desc_t *desc, void *mat,
+ cham_flttype_t dtyp, int mb, int nb, int bsiz,
+ int lm, int ln, int i, int j,
+ int m, int n, int p, int q,
+ void* (*get_blkaddr)( const CHAM_desc_t*, int, int ),
+ int (*get_blkldd) ( const CHAM_desc_t*, int ),
+ int (*get_rankof) ( const CHAM_desc_t*, int, int ) );
+CHAM_desc_t* chameleon_desc_submatrix( CHAM_desc_t *descA, int i, int j, int m, int n );
+void chameleon_desc_destroy ( CHAM_desc_t *desc );
+int chameleon_desc_check ( const CHAM_desc_t *desc );
#define BLKLDD(A, k) A->get_blkldd( A, k )
@@ -116,6 +112,15 @@ inline static void *chameleon_getaddr_cm(const CHAM_desc_t *A, int m, int n)
return (void*)((intptr_t)A->mat + (offset*eltsize) );
}
+/**
+ * Internal function to return address of block (m,n) with m,n = block indices
+ */
+inline static void *chameleon_getaddr_diag( const CHAM_desc_t *A, int m, int n )
+{
+ assert( m == n );
+ return chameleon_getaddr_ccrb( A, m, 0 );
+}
+
/**
* Internal function to return address of block (m,n) with m,n = block indices
* This version lets the runtime allocate on-demand.
@@ -170,7 +175,6 @@ inline static int chameleon_getblkldd_cm(const CHAM_desc_t *A, int m) {
return A->llm;
}
-
/**
* Internal function to return MPI rank of element A(m,n) with m,n = block indices
*/
@@ -187,7 +191,7 @@ inline static int chameleon_getrankof_2d(const CHAM_desc_t *A, int m, int n)
inline static int chameleon_getrankof_2d_diag(const CHAM_desc_t *A, int m, int n)
{
int mm = m + A->i / A->mb;
- assert( n == 0 );
+ assert( m == n );
return (mm % A->p) * A->q + (mm % A->q);
}
diff --git a/control/workspace.c b/control/workspace.c
index 8f5814c318d154382b80fb01cee0bc7c812a566f..e743e33dbb8d40461a70bb0288d1cb824d656c21 100644
--- a/control/workspace.c
+++ b/control/workspace.c
@@ -74,32 +74,7 @@ int chameleon_alloc_ibnb_tile(int M, int N, cham_tasktype_t func, int type, CHAM
lm = IB * MT;
ln = NB * NT;
- /* Allocate and initialize descriptor */
- *desc = (CHAM_desc_t*)malloc(sizeof(CHAM_desc_t));
- if (*desc == NULL) {
- chameleon_error("chameleon_alloc_ibnb_tile", "malloc() failed");
- return CHAMELEON_ERR_OUT_OF_RESOURCES;
- }
- **desc = chameleon_desc_init(type, IB, NB, IB*NB, lm, ln, 0, 0, lm, ln, p, q);
-
- /* Allocate matrix */
- if (chameleon_desc_mat_alloc(*desc)) {
- chameleon_error("chameleon_alloc_ibnb_tile", "malloc() failed");
- free(*desc);
- return CHAMELEON_ERR_OUT_OF_RESOURCES;
- }
-
- RUNTIME_desc_create( *desc );
-
- /* Check that everything is ok */
- status = chameleon_desc_check(*desc);
- if (status != CHAMELEON_SUCCESS) {
- chameleon_error("chameleon_alloc_ibnb_tile", "invalid descriptor");
- free(*desc);
- return status;
- }
-
- return CHAMELEON_SUCCESS;
+ return CHAMELEON_Desc_Create( desc, NULL, type, IB, NB, IB*NB, lm, ln, 0, 0, lm, ln, p, q );
}
/**
@@ -144,18 +119,7 @@ int chameleon_alloc_ipiv(int M, int N, cham_tasktype_t func, int type, CHAM_desc
/* TODO: Fix the distribution for IPIV */
*IPIV = (int*)malloc( size );
- *desc = (CHAM_desc_t*)malloc(sizeof(CHAM_desc_t));
- **desc = chameleon_desc_init(type, IB, NB, IB*NB, lm, ln, 0, 0, lm, ln, p, q );
-
- if ( chameleon_desc_mat_alloc(*desc) ) {
- chameleon_error("chameleon_alloc_ipiv", "malloc() failed");
- free(*desc);
- return CHAMELEON_ERR_OUT_OF_RESOURCES;
- }
-
- RUNTIME_desc_create( *desc );
-
- return CHAMELEON_SUCCESS;
+ return CHAMELEON_Desc_Create( desc, NULL, type, IB, NB, IB*NB, lm, ln, 0, 0, lm, ln, p, q );
}
/**
@@ -193,8 +157,7 @@ int CHAMELEON_Dealloc_Workspace(CHAM_desc_t **desc)
chameleon_error("CHAMELEON_Dealloc_Worspace", "attempting to deallocate a NULL pointer");
return CHAMELEON_ERR_UNALLOCATED;
}
- chameleon_desc_mat_free( *desc );
- RUNTIME_desc_destroy( *desc );
+ chameleon_desc_destroy( *desc );
free(*desc);
*desc = NULL;
diff --git a/include/chameleon/constants.h b/include/chameleon/constants.h
index 4e7e0eaa202bae01f369954d87f683204552d07e..626a02de0989c56b5eafffef070bf107c13eb690 100644
--- a/include/chameleon/constants.h
+++ b/include/chameleon/constants.h
@@ -208,6 +208,13 @@ typedef enum chameleon_translation_e {
ChamOutOfPlace = 2,
} cham_translation_t;
+/**
+ * @brief Constant to describe how to initialize the mat pointer in descriptors
+ */
+#define CHAMELEON_MAT_ALLOC_GLOBAL NULL
+#define CHAMELEON_MAT_ALLOC_TILE ((void*)-1)
+#define CHAMELEON_MAT_OOC ((void*)-2)
+
/**
* CHAMELEON constants - success & error codes
*/