diff --git a/compute/pzungqr_param.c b/compute/pzungqr_param.c
index d16098dc98e13e2b0efa59b7a62305dc94fa0243..b1c41fead93a91692ed00224d787a9fabcc7f0fc 100644
--- a/compute/pzungqr_param.c
+++ b/compute/pzungqr_param.c
@@ -32,30 +32,30 @@
#define A(m,n) A, m, n
#define Q(m,n) Q, m, n
-#define T(m,n) T, m, n
+#define TS(m,n) TS, m, n
+#define TT(m,n) TT, m, n
#if defined(CHAMELEON_COPY_DIAG)
-#define DIAG(k) DIAG, k, 0
+#define D(m,n) D, m, n
#else
-#define DIAG(k) A, k, k
+#define D(m,n) A, m, n
#endif
/**
* Parallel construction of Q using tile V (application to identity) - dynamic scheduling
*/
void morse_pzungqr_param(const libhqr_tree_t *qrtree,
- MORSE_desc_t *A, MORSE_desc_t *Q, MORSE_desc_t *T,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+ MORSE_desc_t *A, MORSE_desc_t *Q, MORSE_desc_t *TS, MORSE_desc_t *TT,
+ MORSE_sequence_t *sequence, MORSE_request_t *request)
{
MORSE_context_t *morse;
MORSE_option_t options;
size_t ws_worker = 0;
size_t ws_host = 0;
- MORSE_desc_t *DIAG = NULL;
+ MORSE_desc_t *D = NULL;
int k, m, n, i, p;
- int ldak, ldqk, ldam, ldqm;
- int tempmm, tempnn, tempkmin, tempkm;
- int tempAkm, tempAkn;
+ int ldak, ldqk, ldam, ldqm, ldqp;
+ int tempmm, tempnn, tempkmin, tempkm, tempkn;
int ib, minMT;
int *tiles;
@@ -99,68 +99,96 @@ void morse_pzungqr_param(const libhqr_tree_t *qrtree,
RUNTIME_options_ws_alloc( &options, ws_worker, ws_host );
#if defined(CHAMELEON_COPY_DIAG)
- /* necessary to avoid dependencies between tasks regarding the diag tile */
- DIAG = (MORSE_desc_t*)malloc(sizeof(MORSE_desc_t));
- morse_zdesc_alloc_diag(*DIAG, A->mb, A->nb, minMT*A->mb, A->nb, 0, 0, minMT*A->mb, A->nb, A->p, A->q);
+ {
+ /* necessary to avoid dependencies between tasks regarding the diag tile */
+ D = (MORSE_desc_t*)malloc(sizeof(MORSE_desc_t));
+ morse_zdesc_alloc(*D, A->mb, A->nb, A->m, A->n, 0, 0, A->m, A->n, );
+ }
#endif
for (k = minMT-1; k >= 0; k--) {
RUNTIME_iteration_push(morse, k);
- tempAkm = k == A->mt-1 ? A->m-k*A->mb : A->mb;
- tempAkn = k == A->nt-1 ? A->n-k*A->nb : A->nb;
- tempkmin = chameleon_min( tempAkn, tempAkm );
- tempkm = k == Q->mt-1 ? Q->m-k*Q->mb : Q->mb;
+ tempkn = k == A->nt-1 ? A->n-k*A->nb : A->nb;
+ tempkm = k == A->mt-1 ? A->m-k*A->mb : A->mb;
ldak = BLKLDD(A, k);
ldqk = BLKLDD(Q, k);
/* Setting the order of tiles */
libhqr_treewalk(qrtree, k, tiles);
- for (i = Q->mt - 1; i > k; i--) {
+ for (i = Q->mt - 2; i >= k; i--) {
m = tiles[i];
p = qrtree->currpiv(qrtree, k, m);
tempmm = m == Q->mt-1 ? Q->m-m*Q->mb : Q->mb;
ldam = BLKLDD(A, m);
ldqm = BLKLDD(Q, m);
- for (n = k; n < Q->nt; n++) {
- tempnn = n == Q->nt-1 ? Q->n-n*Q->nb : Q->nb;
- MORSE_TASK_ztsmqr(
- &options,
- MorseLeft, MorseNoTrans,
- Q->mb, tempnn, tempmm, tempnn, tempAkn, ib, T->nb,
- Q(k, n), ldqk,
- Q(m, n), ldqm,
- A(m, k), ldam,
- T(m, k), T->mb);
+ ldqp = BLKLDD(Q, p);
+
+ /* TT or TS */
+
+ if(qrtree->gettype(qrtree, k , m) == 0){
+ for (n = k; n < Q->nt; n++) {
+ tempnn = n == Q->nt-1 ? Q->n-n*Q->nb : Q->nb;
+ MORSE_TASK_ztsmqr(
+ &options,
+ MorseLeft, MorseNoTrans,
+ Q->mb, tempnn, tempmm, tempnn, tempkn, ib, TS->nb,
+ Q(p, n), ldqp,
+ Q(m, n), ldqm,
+ A(m, k), ldam,
+ TS(m, k), TS->mb);
+ }
+ }
+ else {
+ for (n = k; n < Q->nt; n++) {
+ tempnn = n == Q->nt-1 ? Q->n-n*Q->nb : Q->nb;
+ MORSE_TASK_zttmqr(
+ &options,
+ MorseLeft, MorseNoTrans,
+ Q->mb, tempnn, tempmm, tempnn, tempkn, ib, TT->nb,
+ Q(p, n), ldqp,
+ Q(m, n), ldqm,
+ A(m, k), ldam,
+ TT(m, k), TT->mb);
+ }
}
}
+
+ for (i = 0; i < qrtree->getnbgeqrf(qrtree, k); i++) {
+ m = qrtree->getm(qrtree, k, i);
+
+ tempmm = m == A->mt-1 ? A->m-m*A->mb : A->mb;
+ tempkmin = chameleon_min(tempmm, tempkn);
+ ldam = BLKLDD(A, m);
+ ldqm = BLKLDD(Q, m);
+
#if defined(CHAMELEON_COPY_DIAG)
- MORSE_TASK_zlacpy(
- &options,
- MorseLower, tempkm, tempkmin, A->nb,
- A(k, k), ldak,
- DIAG(k), ldak );
+ MORSE_TASK_zlacpy(
+ &options,
+ MorseLower, tempmm, tempkmin, A->nb,
+ A(m, k), ldam,
+ D(m, k), ldam );
#if defined(CHAMELEON_USE_CUDA)
- MORSE_TASK_zlaset(
- &options,
- MorseUpper, tempkm, tempkmin,
- 0., 1.,
- DIAG(k), ldak );
+ MORSE_TASK_zlaset(
+ &options,
+ MorseUpper, tempmm, tempkmin,
+ 0., 1.,
+ D(m, k), ldam );
#endif
#endif
- for (n = k; n < Q->nt; n++) {
- tempnn = n == Q->nt-1 ? Q->n-n*Q->nb : Q->nb;
- MORSE_TASK_zunmqr(
- &options,
- MorseLeft, MorseNoTrans,
- tempkm, tempnn, tempkmin, ib, T->nb,
- DIAG(k), ldak,
- T(k, k), T->mb,
- Q(k, n), ldqk);
+ for (n = k; n < Q->nt; n++) {
+ tempnn = n == Q->nt-1 ? Q->n-n*Q->nb : Q->nb;
+ MORSE_TASK_zunmqr(
+ &options,
+ MorseLeft, MorseNoTrans,
+ tempmm, tempnn, tempkmin, ib, TS->nb,
+ D(m, k), ldam,
+ TS(m, k), TS->mb,
+ Q(m, n), ldqm);
+ }
}
-
RUNTIME_iteration_pop(morse);
}
RUNTIME_options_ws_free(&options);
@@ -169,8 +197,8 @@ void morse_pzungqr_param(const libhqr_tree_t *qrtree,
#if defined(CHAMELEON_COPY_DIAG)
MORSE_Sequence_Wait(sequence);
- morse_desc_mat_free(DIAG);
- free(DIAG);
+ morse_desc_mat_free(D);
+ free(D);
#endif
- (void)DIAG;
+ (void)D;
}
diff --git a/compute/pzunmqr_param.c b/compute/pzunmqr_param.c
index f18a979a3fafafd4b62564b9c20e3adbea3a6a36..17e9b2db4e3ac841ee033e32a2a81844b36fb2a8 100644
--- a/compute/pzunmqr_param.c
+++ b/compute/pzunmqr_param.c
@@ -181,6 +181,10 @@ void morse_pzunmqr_param(const libhqr_tree_t *qrtree,
tempkmin = k == minMT-1 ? minM-k*A->nb : A->nb;
ldak = BLKLDD(A, k);
ldbk = BLKLDD(B, k);
+
+ /* Setting the order of the tiles*/
+ libhqr_treewalk(qrtree, k, tiles);
+
for (i = B->mt-1; i > k; i--) {
m = tiles[i];
p = qrtree->currpiv(qrtree, k, m);
@@ -242,6 +246,10 @@ void morse_pzunmqr_param(const libhqr_tree_t *qrtree,
tempkmin = k == minMT-1 ? minM-k*A->nb : A->nb;
ldak = BLKLDD(A, k);
ldbk = BLKLDD(B, k);
+
+ /* Setting the order of tiles */
+ libhqr_treewalk(qrtree, k, tiles);
+
for (n = B->nt-1; n > k; n--) {
tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb;
ldan = BLKLDD(A, n);
@@ -297,6 +305,7 @@ void morse_pzunmqr_param(const libhqr_tree_t *qrtree,
tempkn = k == B->nt-1 ? B->n-k*B->nb : B->nb;
tempkmin = k == minMT-1 ? minM-k*A->nb : A->nb;
ldak = BLKLDD(A, k);
+
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
&options,
@@ -322,6 +331,10 @@ void morse_pzunmqr_param(const libhqr_tree_t *qrtree,
T(k, k), T->mb,
B(m, k), ldbm);
}
+
+ /* Setting the order of tiles */
+ libhqr_treewalk(qrtree, k, tiles);
+
for (n = k+1; n < B->nt; n++) {
tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb;
ldan = BLKLDD(A, n);
diff --git a/compute/zungqr_param.c b/compute/zungqr_param.c
index 6dddf1cd9382ab9ed7ee75f9033a1c57316b4f47..24bec0c623a53b11dceef761161a4882f6aef67e 100644
--- a/compute/zungqr_param.c
+++ b/compute/zungqr_param.c
@@ -40,6 +40,9 @@
*
*******************************************************************************
*
+ * @param[in] qrtree
+ * The tree used for the factorization
+ *
* @param[in] M
* The number of rows of the matrix Q. M >= 0.
*
@@ -84,7 +87,8 @@
int MORSE_zungqr_param(const libhqr_tree_t *qrtree,
int M, int N, int K,
MORSE_Complex64_t *A, int LDA,
- MORSE_desc_t *descT,
+ MORSE_desc_t *descTS,
+ MORSE_desc_t *descTT,
MORSE_Complex64_t *Q, int LDQ)
{
int NB;
@@ -149,7 +153,7 @@ int MORSE_zungqr_param(const libhqr_tree_t *qrtree,
/* }*/
/* Call the tile interface */
- MORSE_zungqr_param_Tile_Async(qrtree, &descA, descT, &descQ, sequence, &request);
+ MORSE_zungqr_param_Tile_Async(qrtree, &descA, descTS, descTT, &descQ, sequence, &request);
/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
morse_zooptile2lap(descQ, Q, NB, NB, LDQ, N, sequence, &request);
@@ -167,7 +171,8 @@ int MORSE_zungqr_param(const libhqr_tree_t *qrtree,
return status;
}
-/***************************************************************************//**
+/**
+ *******************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -201,7 +206,7 @@ int MORSE_zungqr_param(const libhqr_tree_t *qrtree,
* @sa MORSE_zgeqrf_Tile
*
******************************************************************************/
-int MORSE_zungqr_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *Q)
+int MORSE_zungqr_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *Q)
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -214,7 +219,7 @@ int MORSE_zungqr_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_
return MORSE_ERR_NOT_INITIALIZED;
}
morse_sequence_create(morse, &sequence);
- MORSE_zungqr_param_Tile_Async(qrtree, A, T, Q, sequence, &request);
+ MORSE_zungqr_param_Tile_Async(qrtree, A, TS, TT, Q, sequence, &request);
morse_sequence_wait(morse, sequence);
RUNTIME_desc_getoncpu(A);
RUNTIME_desc_getoncpu(Q);
@@ -224,7 +229,8 @@ int MORSE_zungqr_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_
return status;
}
-/***************************************************************************//**
+/**
+ *******************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
@@ -251,7 +257,7 @@ int MORSE_zungqr_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_
* @sa MORSE_zgeqrf_Tile_Async
*
******************************************************************************/
-int MORSE_zungqr_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *Q,
+int MORSE_zungqr_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *Q,
MORSE_sequence_t *sequence, MORSE_request_t *request)
{
MORSE_context_t *morse;
@@ -280,14 +286,18 @@ int MORSE_zungqr_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_desc_t *A,
morse_error("MORSE_zungqr_param_Tile", "invalid first descriptor");
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
- if (morse_desc_check(T) != MORSE_SUCCESS) {
+ if (morse_desc_check(TS) != MORSE_SUCCESS) {
morse_error("MORSE_zungqr_param_Tile", "invalid second descriptor");
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
- if (morse_desc_check(Q) != MORSE_SUCCESS) {
+ if (morse_desc_check(TT) != MORSE_SUCCESS) {
morse_error("MORSE_zungqr_param_Tile", "invalid third descriptor");
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
+ if (morse_desc_check(Q) != MORSE_SUCCESS) {
+ morse_error("MORSE_zungqr_param_Tile", "invalid fourth descriptor");
+ return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
+ }
/* Check input arguments */
if (A->nb != A->mb || Q->nb != Q->mb) {
morse_error("MORSE_zungqr_param_Tile", "only square tiles supported");
@@ -299,7 +309,7 @@ int MORSE_zungqr_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_desc_t *A,
return MORSE_SUCCESS;
*/
morse_pzlaset(MorseUpperLower, 0., 1., Q, sequence, request);
- morse_pzungqr_param(qrtree, A, Q, T, sequence, request);
+ morse_pzungqr_param(qrtree, A, Q, TS, TT, sequence, request);
return MORSE_SUCCESS;
}
diff --git a/control/compute_z.h b/control/compute_z.h
index 9340037118824f0cf0892e7158dd0945f5eb4d26..0601c8ffe1902971485b4f1e7e8507a644a056c9 100644
--- a/control/compute_z.h
+++ b/control/compute_z.h
@@ -168,6 +168,6 @@ void morse_pzunmqr_param(const libhqr_tree_t *qrtree, MORSE_enum side, MORSE_enu
MORSE_sequence_t *sequence, MORSE_request_t *request);
void morse_pzunglq_param(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *Q, MORSE_desc_t *T,
MORSE_sequence_t *sequence, MORSE_request_t *request);
-void morse_pzungqr_param(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *Q, MORSE_desc_t *T,
+void morse_pzungqr_param(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *Q, MORSE_desc_t *TS, MORSE_desc_t *TT,
MORSE_sequence_t *sequence, MORSE_request_t *request);
#endif /* defined(CHAMELEON_USE_LIBHQR) */
diff --git a/include/morse_z.h b/include/morse_z.h
index ac045809a3ecdb95ccfdfafe389be7d2bc88907a..1c7e6896c1e6eabf414c369451fe770c3372409d 100644
--- a/include/morse_z.h
+++ b/include/morse_z.h
@@ -280,9 +280,9 @@ int MORSE_zunmqr_Tile_Async(MORSE_enum side, MORSE_enum trans, MORSE_desc_t *A,
int MORSE_zgeqrf_param(const libhqr_tree_t *qrtree, int M, int N, MORSE_Complex64_t *A, int LDA, MORSE_desc_t *descTS, MORSE_desc_t *descTT);
int MORSE_zgeqrf_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT);
int MORSE_zgeqrf_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_sequence_t *sequence, MORSE_request_t *request);
-int MORSE_zungqr_param(const libhqr_tree_t *qrtree, int M, int N, int K, MORSE_Complex64_t *A, int LDA, MORSE_desc_t *descT, MORSE_Complex64_t *B, int LDB);
-int MORSE_zungqr_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *B);
-int MORSE_zungqr_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *B, MORSE_sequence_t *sequence, MORSE_request_t *request);
+ int MORSE_zungqr_param(const libhqr_tree_t *qrtree, int M, int N, int K, MORSE_Complex64_t *A, int LDA, MORSE_desc_t *descTS, MORSE_desc_t *descTT, MORSE_Complex64_t *B, int LDB);
+ int MORSE_zungqr_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *B);
+ int MORSE_zungqr_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *B, MORSE_sequence_t *sequence, MORSE_request_t *request);
int MORSE_zunmqr_param(const libhqr_tree_t *qrtree, MORSE_enum side, MORSE_enum trans, int M, int N, int K, MORSE_Complex64_t *A, int LDA, MORSE_desc_t *descT, MORSE_Complex64_t *B, int LDB);
int MORSE_zunmqr_param_Tile(const libhqr_tree_t *qrtree, MORSE_enum side, MORSE_enum trans, MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *B);
int MORSE_zunmqr_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_enum side, MORSE_enum trans, MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *B, MORSE_sequence_t *sequence, MORSE_request_t *request);
diff --git a/testing/testing_zgels_param.c b/testing/testing_zgels_param.c
index d87f776e34beec2f557bbde9c65a3ecf97395011..85e5b429bf848bcc345e9116dc12dedc9bac5326 100644
--- a/testing/testing_zgels_param.c
+++ b/testing/testing_zgels_param.c
@@ -33,6 +33,7 @@
#include <stdio.h>
#include <string.h>
#include <math.h>
+#include <assert.h>
#include <morse.h>
#include <coreblas/include/cblas.h>
@@ -115,9 +116,9 @@ int testing_zgels_param(int argc, char **argv)
matrix.p = 1;
/* Initialize qrtree */
- domino = 0; /* -1 */
- llvl = 0; /* -1 */
- hlvl = 0; /* -1 */
+ domino = -1; /* -1 */
+ llvl = -1; /* -1 */
+ hlvl = -1; /* -1 */
qr_a = TS->mt; /* -1 */
qr_p = 1; /* matrix.p */
tsrr = 0; /* 0 */
@@ -142,7 +143,7 @@ int testing_zgels_param(int argc, char **argv)
/* MORSE ZGELS */
if (M >= N)
/* Building the economy-size Q */
- MORSE_zungqr(M, N, K, A2, LDA, TS, Q, LDA);
+ MORSE_zungqr_param(&qrtree, M, N, K, A2, LDA, TS, TT, Q, LDA);
else
/* Building the economy-size Q */
MORSE_zunglq(M, N, K, A2, LDA, TS, Q, LDA);
@@ -197,7 +198,7 @@ int testing_zgels_param(int argc, char **argv)
/* Morse routines */
MORSE_zgeqrf_param( &qrtree, M, N, A2, LDA, TS, TT );
- MORSE_zungqr(M, N, K, A2, LDA, TS, Q, LDA);
+ MORSE_zungqr_param( &qrtree, M, N, K, A2, LDA, TS, TT, Q, LDA);
MORSE_zgeqrs(M, N, NRHS, A2, LDA, TS, B2, LDB);
/* Check the orthogonality, factorization and the solution */
@@ -276,7 +277,7 @@ int testing_zgels_param(int argc, char **argv)
printf(" Computational tests pass if scaled residuals are less than 60.\n");
MORSE_zgeqrf_param( &qrtree, M, N, A2, LDA, TS, TT );
- MORSE_zungqr(M, N, K, A2, LDA, TS, Q, LDA);
+ MORSE_zungqr_param( &qrtree, M, N, K, A2, LDA, TS, TT, Q, LDA);
MORSE_zunmqr(MorseLeft, MorseConjTrans, M, NRHS, N, A2, LDA, TS, B2, LDB);
MORSE_ztrsm(MorseLeft, MorseUpper, MorseNoTrans, MorseNonUnit, N, NRHS, 1.0, A2, LDA, B2, LDB);
}