diff --git a/compute/pzgeqrf_param.c b/compute/pzgeqrf_param.c
index 0831540e19f302dc8477787b3f48ec9ae798e72e..418d9ee8b797ab416109a57d90cc90a331d2fa72 100644
--- a/compute/pzgeqrf_param.c
+++ b/compute/pzgeqrf_param.c
@@ -30,8 +30,148 @@
* Parallel tile QR factorization (reduction Householder) - dynamic scheduling
*
* @param[in] genD
- * Indicate if the copies of the geqrt tiles must be done to speedup
- * computations in updates.
+ * Indicate if copies of the geqrt tiles must be done to speedup
+ * computations in updates. genD is considered only if D is not NULL.
+ *
+ * @param[in] uplo
+ * - ChamLower: Classic QR factorization of the matrix A.
+ * - ChamUpper: QR factorization of the TTQRT kernel.
+ * - ChamUpperLower: QR factorization of the TSQRT kernel.
+ */
+int chameleon_pzgeqrf_param_step( int genD, cham_uplo_t uplo, int k, int ib,
+ const libhqr_tree_t *qrtree, int *tiles,
+ CHAM_desc_t *A, CHAM_desc_t *TS, CHAM_desc_t *TT, CHAM_desc_t *D,
+ RUNTIME_option_t *options, RUNTIME_sequence_t *sequence )
+{
+ CHAM_desc_t *T;
+ int m, n, i, p;
+ int L, nbgeqrt;
+ int tempkmin, tempkn, tempnn, tempmm;
+ int node, nbtiles;
+
+ tempkn = k == A->nt-1 ? A->n-k*A->nb : A->nb;
+
+ /* The number of geqrt to apply */
+ nbgeqrt = qrtree->getnbgeqrf( qrtree, k );
+
+ T = TS;
+ for (i = 0; i < nbgeqrt; i++) {
+ m = qrtree->getm( qrtree, k, i );
+
+ /* We skip the QR factorization if this is the last diagonal tile */
+ if ( (uplo == ChamUpper) && (m == k) ) {
+ continue;
+ }
+
+ tempmm = m == A->mt-1 ? A->m-m*A->mb : A->mb;
+ tempkmin = chameleon_min(tempmm, tempkn);
+
+ INSERT_TASK_zgeqrt(
+ options,
+ tempmm, tempkn, ib, T->nb,
+ A(m, k), T(m, k) );
+
+ if ( genD ) {
+ int tempDmm = m == D->mt-1 ? D->m-m*D->mb : D->mb;
+ int tempDkn = k == D->nt-1 ? D->n-k*D->nb : D->nb;
+
+ INSERT_TASK_zlacpy(
+ options,
+ ChamLower, tempDmm, tempDkn, A->nb,
+ A(m, k), D(m, k) );
+#if defined(CHAMELEON_USE_CUDA)
+ INSERT_TASK_zlaset(
+ options,
+ ChamUpper, tempDmm, tempDkn,
+ 0., 1.,
+ D(m, k) );
+#endif
+ }
+
+ for (n = k+1; n < A->nt; n++) {
+ tempnn = n == A->nt-1 ? A->n-n*A->nb : A->nb;
+ INSERT_TASK_zunmqr(
+ options,
+ ChamLeft, ChamConjTrans,
+ tempmm, tempnn, tempkmin, ib, T->nb,
+ D(m, k),
+ T(m, k),
+ A(m, n));
+ }
+
+ if ( genD || ((k+1) < A->nt)) {
+ RUNTIME_data_flush( sequence, D(m, k) );
+ }
+ RUNTIME_data_flush( sequence, T(m, k) );
+ }
+
+ /* Setting the order of the tiles */
+ nbtiles = libhqr_walk_stepk( qrtree, k, tiles );
+
+ for (i = 0; i < nbtiles; i++) {
+ m = tiles[i];
+ p = qrtree->currpiv( qrtree, k, m );
+
+ tempmm = m == A->mt-1 ? A->m-m*A->mb : A->mb;
+
+ if ( qrtree->gettype( qrtree, k, m ) == LIBHQR_KILLED_BY_TS ) {
+ /* TS kernel */
+ T = TS;
+ L = 0;
+
+ /* Force TT kernel if this is the last diagonal tile */
+ if ( (uplo == ChamUpper) && (m == k) ) {
+ L = tempmm;
+ }
+ }
+ else {
+ /* TT kernel */
+ T = TT;
+ L = tempmm;
+ }
+
+ node = A->get_rankof( A, m, k );
+ RUNTIME_data_migrate( sequence, A(p, k), node );
+ RUNTIME_data_migrate( sequence, A(m, k), node );
+
+ INSERT_TASK_ztpqrt(
+ options,
+ tempmm, tempkn, chameleon_min(L, tempkn), ib, T->nb,
+ A(p, k),
+ A(m, k),
+ T(m, k));
+
+ for (n = k+1; n < A->nt; n++) {
+ tempnn = n == A->nt-1 ? A->n-n*A->nb : A->nb;
+
+ node = A->get_rankof( A, m, n );
+ RUNTIME_data_migrate( sequence, A(p, n), node );
+ RUNTIME_data_migrate( sequence, A(m, n), node );
+
+ INSERT_TASK_ztpmqrt(
+ options,
+ ChamLeft, ChamConjTrans,
+ tempmm, tempnn, A->nb, L, ib, T->nb,
+ A(m, k),
+ T(m, k),
+ A(p, n),
+ A(m, n));
+ }
+ RUNTIME_data_flush( sequence, A(m, k) );
+ RUNTIME_data_flush( sequence, T(m, k) );
+ }
+
+ return tiles[nbtiles];
+}
+
+
+/**
+ * Parallel tile QR factorization (reduction Householder) - dynamic scheduling
+ *
+ * @param[in] genD
+ * Indicate if copies of the geqrt tiles must be done to speedup
+ * computations in updates. genD is considered only if D is not NULL.
+ *
*/
void chameleon_pzgeqrf_param( int genD, int K,
const libhqr_tree_t *qrtree, CHAM_desc_t *A,
@@ -40,14 +180,11 @@ void chameleon_pzgeqrf_param( int genD, int K,
{
CHAM_context_t *chamctxt;
RUNTIME_option_t options;
- CHAM_desc_t *T;
size_t ws_worker = 0;
size_t ws_host = 0;
- int k, m, n, i, p;
- int L, nbgeqrt;
- int tempkmin, tempkn, tempnn, tempmm;
- int ib, node, nbtiles, *tiles;
+ int k, n;
+ int ib, *tiles;
chamctxt = chameleon_context_self();
if (sequence->status != CHAMELEON_SUCCESS) {
@@ -57,7 +194,7 @@ void chameleon_pzgeqrf_param( int genD, int K,
ib = CHAMELEON_IB;
- if ( D == NULL ) {
+ if ( (genD == 0) || (D == NULL) ) {
D = A;
genD = 0;
}
@@ -85,109 +222,13 @@ void chameleon_pzgeqrf_param( int genD, int K,
RUNTIME_options_ws_alloc( &options, ws_worker, ws_host );
/* Initialisation of temporary tiles array */
- tiles = (int*)calloc(qrtree->mt, sizeof(int));
+ tiles = (int*)calloc( qrtree->mt, sizeof(int) );
for (k = 0; k < K; k++) {
- RUNTIME_iteration_push(chamctxt, k);
- tempkn = k == A->nt-1 ? A->n-k*A->nb : A->nb;
-
- /* The number of geqrt to apply */
- nbgeqrt = qrtree->getnbgeqrf(qrtree, k);
-
- T = TS;
- for (i = 0; i < nbgeqrt; i++) {
- m = qrtree->getm(qrtree, k, i);
- tempmm = m == A->mt-1 ? A->m-m*A->mb : A->mb;
- tempkmin = chameleon_min(tempmm, tempkn);
-
- INSERT_TASK_zgeqrt(
- &options,
- tempmm, tempkn, ib, T->nb,
- A(m, k),
- T(m, k));
-
- if ( genD ) {
- int tempDmm = m == D->mt-1 ? D->m-m*D->mb : D->mb;
- int tempDkn = k == D->nt-1 ? D->n-k*D->nb : D->nb;
+ RUNTIME_iteration_push( chamctxt, k );
- INSERT_TASK_zlacpy(
- &options,
- ChamLower, tempDmm, tempDkn, A->nb,
- A(m, k),
- D(m, k) );
-#if defined(CHAMELEON_USE_CUDA)
- INSERT_TASK_zlaset(
- &options,
- ChamUpper, tempDmm, tempDkn,
- 0., 1.,
- D(m, k) );
-#endif
- }
-
- for (n = k+1; n < A->nt; n++) {
- tempnn = n == A->nt-1 ? A->n-n*A->nb : A->nb;
- INSERT_TASK_zunmqr(
- &options,
- ChamLeft, ChamConjTrans,
- tempmm, tempnn, tempkmin, ib, T->nb,
- D(m, k),
- T(m, k),
- A(m, n));
- }
- RUNTIME_data_flush( sequence, D(m, k) );
- RUNTIME_data_flush( sequence, T(m, k) );
- }
-
- /* Setting the order of the tiles */
- nbtiles = libhqr_walk_stepk( qrtree, k, tiles );
-
- for (i = 0; i < nbtiles; i++) {
- m = tiles[i];
- p = qrtree->currpiv(qrtree, k, m);
-
- tempmm = m == A->mt-1 ? A->m-m*A->mb : A->mb;
-
- if ( qrtree->gettype(qrtree, k, m) == LIBHQR_KILLED_BY_TS ) {
- /* TS kernel */
- T = TS;
- L = 0;
- }
- else {
- /* TT kernel */
- T = TT;
- L = tempmm;
- }
-
- node = A->get_rankof( A, m, k );
- RUNTIME_data_migrate( sequence, A(p, k), node );
- RUNTIME_data_migrate( sequence, A(m, k), node );
-
- INSERT_TASK_ztpqrt(
- &options,
- tempmm, tempkn, chameleon_min(L, tempkn), ib, T->nb,
- A(p, k),
- A(m, k),
- T(m, k));
-
- for (n = k+1; n < A->nt; n++) {
- tempnn = n == A->nt-1 ? A->n-n*A->nb : A->nb;
-
- node = A->get_rankof( A, m, n );
- RUNTIME_data_migrate( sequence, A(p, n), node );
- RUNTIME_data_migrate( sequence, A(m, n), node );
-
- INSERT_TASK_ztpmqrt(
- &options,
- ChamLeft, ChamConjTrans,
- tempmm, tempnn, A->nb, L, ib, T->nb,
- A(m, k),
- T(m, k),
- A(p, n),
- A(m, n));
- }
- RUNTIME_data_flush( sequence, A(m, k) );
- RUNTIME_data_flush( sequence, T(m, k) );
- }
+ chameleon_pzgeqrf_param_step( genD, ChamLower, k, ib, qrtree, tiles,
+ A, TS, TT, D, &options, sequence );
/* Restore the original location of the tiles */
for (n = k; n < A->nt; n++) {
@@ -195,10 +236,10 @@ void chameleon_pzgeqrf_param( int genD, int K,
A->get_rankof( A, k, n ) );
}
- RUNTIME_iteration_pop(chamctxt);
+ RUNTIME_iteration_pop( chamctxt );
}
- free(tiles);
- RUNTIME_options_ws_free(&options);
- RUNTIME_options_finalize(&options, chamctxt);
+ free( tiles );
+ RUNTIME_options_ws_free( &options );
+ RUNTIME_options_finalize( &options, chamctxt );
}
diff --git a/compute/pzungqr_param.c b/compute/pzungqr_param.c
index e9046ab30581d6f38e35f9514a9070fa36c89802..5ffb77ab7e61ac7dbbfdd80222bf245581f2d744 100644
--- a/compute/pzungqr_param.c
+++ b/compute/pzungqr_param.c
@@ -28,6 +28,131 @@
/**
* Parallel construction of Q using tile V (application to identity) - dynamic scheduling
+ *
+ * @param[in] genD
+ * Indicate if the copies of the A tiles must be done to speedup
+ * computations in updates.
+ *
+ * @param[in] uplo
+ * Indicate which kind of factorization has been performed on A to apply
+ * the respective Q generation.
+ * - ChamLower: Apply Classic QR factorization of the matrix A
+ * - ChamUpper: Apply the factorization of the upper part from a TT kernel.
+ * - ChamUpperLower: Apply the factorization of the full tile from a TS kernel.
+ */
+void chameleon_pzungqr_param_step( int genD, cham_uplo_t uplo, int k, int ib,
+ const libhqr_tree_t *qrtree, int nbtiles, int *tiles,
+ CHAM_desc_t *A, CHAM_desc_t *Q,
+ CHAM_desc_t *TS, CHAM_desc_t *TT, CHAM_desc_t *D,
+ RUNTIME_option_t *options, RUNTIME_sequence_t *sequence )
+{
+ CHAM_desc_t *T;
+ int m, n, i, p, L;
+ int tempmm, tempnn, tempkmin, tempkn;
+ int nbgeqrt, node;
+
+ tempkn = k == A->nt-1 ? A->n - k * A->nb : A->nb;
+
+ for (i = nbtiles-1; i >= 0; i--) {
+ m = tiles[i];
+ p = qrtree->currpiv( qrtree, k, m );
+
+ tempmm = m == Q->mt-1 ? Q->m-m*Q->mb : Q->mb;
+
+ if( qrtree->gettype( qrtree, k, m ) == LIBHQR_KILLED_BY_TS ) {
+ /* TS kernel */
+ T = TS;
+ L = 0;
+
+ /* Force TT kernel if this is the last diagonal tile */
+ if ( (uplo == ChamUpper) && (m == k) ) {
+ L = tempmm;
+ }
+ }
+ else {
+ /* TT kernel */
+ T = TT;
+ L = tempmm;
+ }
+
+ for (n = k; n < Q->nt; n++) {
+ tempnn = n == Q->nt-1 ? Q->n-n*Q->nb : Q->nb;
+
+ node = Q->get_rankof( Q, m, n );
+ RUNTIME_data_migrate( sequence, Q(p, n), node );
+ RUNTIME_data_migrate( sequence, Q(m, n), node );
+
+ INSERT_TASK_ztpmqrt(
+ options,
+ ChamLeft, ChamNoTrans,
+ tempmm, tempnn, tempkn, L, ib, T->nb,
+ A(m, k),
+ T(m, k),
+ Q(p, n),
+ Q(m, n));
+ }
+ RUNTIME_data_flush( sequence, A(m, k) );
+ RUNTIME_data_flush( sequence, T(m, k) );
+ }
+
+ T = TS;
+
+ /* The number of geqrt to apply */
+ nbgeqrt = qrtree->getnbgeqrf( qrtree, k );
+ for (i = 0; i < nbgeqrt; i++) {
+ m = qrtree->getm( qrtree, k, i );
+
+ /* We skip the QR factorization if this is the last diagonal tile */
+ if ( (uplo == ChamUpper) && (m == k) ) {
+ continue;
+ }
+
+ tempmm = m == A->mt-1 ? A->m-m*A->mb : A->mb;
+ tempkmin = chameleon_min( tempmm, tempkn );
+
+ if ( genD ) {
+ int tempDmm = m == D->mt-1 ? D->m - m * D->mb : D->mb;
+ INSERT_TASK_zlacpy(
+ options,
+ ChamLower, tempDmm, tempkmin, A->nb,
+ A(m, k),
+ D(m, k) );
+#if defined(CHAMELEON_USE_CUDA)
+ INSERT_TASK_zlaset(
+ options,
+ ChamUpper, tempDmm, tempkmin,
+ 0., 1.,
+ D(m, k) );
+#endif
+ }
+
+ for (n = k; n < Q->nt; n++) {
+ tempnn = n == Q->nt-1 ? Q->n-n*Q->nb : Q->nb;
+
+ /* Restore the original location of the tiles */
+ RUNTIME_data_migrate( sequence, Q(m, n),
+ Q->get_rankof( Q, m, n ) );
+
+ INSERT_TASK_zunmqr(
+ options,
+ ChamLeft, ChamNoTrans,
+ tempmm, tempnn, tempkmin, ib, T->nb,
+ D(m, k),
+ T(m, k),
+ Q(m, n));
+ }
+ RUNTIME_data_flush( sequence, D(m, k) );
+ RUNTIME_data_flush( sequence, T(m, k) );
+ }
+}
+
+/**
+ * Parallel construction of Q using tile V (application to identity) - dynamic scheduling
+ *
+ * @param[in] genD
+ * Indicate if the copies of the A tiles must be done to speedup
+ * computations in updates. genD is considered only if D is not NULL.
+ *
*/
void chameleon_pzungqr_param( int genD, int K,
const libhqr_tree_t *qrtree,
@@ -37,13 +162,9 @@ void chameleon_pzungqr_param( int genD, int K,
{
CHAM_context_t *chamctxt;
RUNTIME_option_t options;
- CHAM_desc_t *T;
size_t ws_worker = 0;
size_t ws_host = 0;
-
- int k, m, n, i, p, L;
- int tempmm, tempnn, tempkmin, tempkn;
- int ib, nbgeqrt, node, nbtiles, *tiles;
+ int k, ib, nbtiles, *tiles;
chamctxt = chameleon_context_self();
if (sequence->status != CHAMELEON_SUCCESS) {
@@ -53,7 +174,7 @@ void chameleon_pzungqr_param( int genD, int K,
ib = CHAMELEON_IB;
- if (D == NULL) {
+ if ( D == NULL ) {
D = A;
genD = 0;
}
@@ -83,92 +204,13 @@ void chameleon_pzungqr_param( int genD, int K,
for (k = K-1; k >=0; k--) {
RUNTIME_iteration_push(chamctxt, k);
- tempkn = k == A->nt-1 ? A->n-k*A->nb : A->nb;
-
/* Setting the order of tiles */
nbtiles = libhqr_walk_stepk( qrtree, k, tiles );
- for (i = nbtiles-1; i >= 0; i--) {
- m = tiles[i];
- p = qrtree->currpiv(qrtree, k, m);
-
- tempmm = m == Q->mt-1 ? Q->m-m*Q->mb : Q->mb;
-
- if( qrtree->gettype(qrtree, k, m) == LIBHQR_KILLED_BY_TS ) {
- /* TS kernel */
- T = TS;
- L = 0;
- }
- else {
- /* TT kernel */
- T = TT;
- L = tempmm;
- }
-
- for (n = k; n < Q->nt; n++) {
- tempnn = n == Q->nt-1 ? Q->n-n*Q->nb : Q->nb;
-
- node = Q->get_rankof( Q, m, n );
- RUNTIME_data_migrate( sequence, Q(p, n), node );
- RUNTIME_data_migrate( sequence, Q(m, n), node );
-
- INSERT_TASK_ztpmqrt(
- &options,
- ChamLeft, ChamNoTrans,
- tempmm, tempnn, tempkn, L, ib, T->nb,
- A(m, k),
- T(m, k),
- Q(p, n),
- Q(m, n));
- }
- RUNTIME_data_flush( sequence, A(m, k) );
- RUNTIME_data_flush( sequence, T(m, k) );
- }
-
- T = TS;
-
- /* The number of geqrt to apply */
- nbgeqrt = qrtree->getnbgeqrf(qrtree, k);
- for (i = 0; i < nbgeqrt; i++) {
- m = qrtree->getm(qrtree, k, i);
-
- tempmm = m == A->mt-1 ? A->m-m*A->mb : A->mb;
- tempkmin = chameleon_min(tempmm, tempkn);
-
- if ( genD ) {
- int tempDmm = m == D->mt-1 ? D->m-m*D->mb : D->mb;
- INSERT_TASK_zlacpy(
- &options,
- ChamLower, tempDmm, tempkmin, A->nb,
- A(m, k),
- D(m, k) );
-#if defined(CHAMELEON_USE_CUDA)
- INSERT_TASK_zlaset(
- &options,
- ChamUpper, tempDmm, tempkmin,
- 0., 1.,
- D(m, k) );
-#endif
- }
-
- for (n = k; n < Q->nt; n++) {
- tempnn = n == Q->nt-1 ? Q->n-n*Q->nb : Q->nb;
-
- /* Restore the original location of the tiles */
- RUNTIME_data_migrate( sequence, Q(m, n),
- Q->get_rankof( Q, m, n ) );
-
- INSERT_TASK_zunmqr(
- &options,
- ChamLeft, ChamNoTrans,
- tempmm, tempnn, tempkmin, ib, T->nb,
- D(m, k),
- T(m, k),
- Q(m, n));
- }
- RUNTIME_data_flush( sequence, D(m, k) );
- RUNTIME_data_flush( sequence, T(m, k) );
- }
+ chameleon_pzungqr_param_step( genD, ChamLower, k, ib,
+ qrtree, nbtiles, tiles,
+ A, Q, TS, TT, D,
+ &options, sequence );
RUNTIME_iteration_pop(chamctxt);
}
diff --git a/control/compute_z.h b/control/compute_z.h
index 82d3e99ca5858f37f507a32e442ac2504b4c3ae5..70e94325e4ad00cc0e7bdb826cd83bce2d458fe9 100644
--- a/control/compute_z.h
+++ b/control/compute_z.h
@@ -99,6 +99,15 @@ void chameleon_pzunmlq( int genD, cham_side_t side, cham_trans_t trans, CHAM_des
void chameleon_pzunmlqrh( int genD, int BS, cham_side_t side, cham_trans_t trans, CHAM_desc_t *A, CHAM_desc_t *B, CHAM_desc_t *T, CHAM_desc_t *D, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
void chameleon_pzbuild( 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_pzgeqrf_param_step( int genD, cham_uplo_t uplo, int k, int ib,
+ const libhqr_tree_t *qrtree, int *tiles,
+ CHAM_desc_t *A, CHAM_desc_t *TS, CHAM_desc_t *TT, CHAM_desc_t *D,
+ RUNTIME_option_t *options, RUNTIME_sequence_t *sequence );
+void chameleon_pzungqr_param_step( int genD, cham_uplo_t uplo, int k, int ib,
+ const libhqr_tree_t *qrtree, int nbtiles, int *tiles,
+ CHAM_desc_t *A, CHAM_desc_t *Q,
+ CHAM_desc_t *TS, CHAM_desc_t *TT, CHAM_desc_t *D,
+ RUNTIME_option_t *options, RUNTIME_sequence_t *sequence );
void chameleon_pzgelqf_param( int genD, const libhqr_tree_t *qrtree, CHAM_desc_t *A, CHAM_desc_t *TS, CHAM_desc_t *TT, CHAM_desc_t *D,
RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
void chameleon_pzgeqrf_param( int genD, int K, const libhqr_tree_t *qrtree,