From 52c188d947e4db2e76502bc2dda4739073dca8e8 Mon Sep 17 00:00:00 2001
From: Mathieu Faverge <mathieu.faverge@inria.fr>
Date: Wed, 17 Oct 2018 23:01:05 +0200
Subject: [PATCH] Factorize migrate calls, and fix tpm* kernels
---
compute/pzgelqf_param.c | 63 +++++++++++++++-----------------
compute/pzgelqfrh.c | 16 ++++----
compute/pzgeqrf_param.c | 53 ++++++++++++++-------------
compute/pzgeqrfrh.c | 16 ++++----
compute/pztpgqrt.c | 65 ++-------------------------------
compute/pztpqrt.c | 13 +++----
compute/pzunglq_param.c | 23 +++++-------
compute/pzunglqrh.c | 16 ++++----
compute/pzungqr_param.c | 49 +++++++++++--------------
compute/pzungqrrh.c | 18 ++++-----
compute/pzunmlq_param.c | 60 ++++++++++++++----------------
compute/pzunmlqrh.c | 58 +++++++++++++----------------
compute/pzunmqr_param.c | 62 +++++++++++++++----------------
compute/pzunmqrrh.c | 60 +++++++++++++-----------------
compute/zgels_param.c | 3 +-
compute/zgeqrf_param.c | 16 ++++++--
compute/ztpgqrt.c | 19 ++++++----
compute/zungqr_param.c | 21 ++++++++---
coreblas/compute/core_ztpmlqt.c | 2 +-
coreblas/compute/core_ztpmqrt.c | 2 +-
20 files changed, 277 insertions(+), 358 deletions(-)
diff --git a/compute/pzgelqf_param.c b/compute/pzgelqf_param.c
index 9f853ef7a..cfc343eb2 100644
--- a/compute/pzgelqf_param.c
+++ b/compute/pzgelqf_param.c
@@ -26,8 +26,7 @@
#define T(m,n) T, (m), (n)
#define D(m,n) D, (m), (n)
-
-/*
+/**
* Parallel tile LQ factorization (reduction Householder) - dynamic scheduling
*/
void chameleon_pzgelqf_param( int genD, const libhqr_tree_t *qrtree, CHAM_desc_t *A,
@@ -41,11 +40,10 @@ void chameleon_pzgelqf_param( int genD, const libhqr_tree_t *qrtree, CHAM_desc_t
size_t ws_host = 0;
int k, m, n, i, p;
- int K, L;
+ int K, L, nbgeqrt;
int ldak, ldam, lddk;
int tempkmin, tempkm, tempnn, tempmm, temppn;
- int ib;
- int *tiles;
+ int ib, node, nbtiles, *tiles;
chamctxt = chameleon_context_self();
if (sequence->status != CHAMELEON_SUCCESS)
@@ -60,37 +58,32 @@ void chameleon_pzgelqf_param( int genD, const libhqr_tree_t *qrtree, CHAM_desc_t
}
/*
- * zgelqt = A->nb * (ib+1)
- * zunmlq = A->nb * ib
- * ztslqt = A->nb * (ib+1)
- * zttlqt = A->nb * (ib+1)
- * ztsmlq = A->nb * ib
- * zttmlq = A->nb * ib
+ * zgelqt = A->nb * (ib+1)
+ * zunmlq = A->nb * ib
+ * ztplqt = A->nb * (ib+1)
+ * ztpmlqt = A->nb * ib
*/
ws_worker = A->nb * (ib+1);
/* Allocation of temporary (scratch) working space */
#if defined(CHAMELEON_USE_CUDA)
- /* Worker space
- *
- * zunmlq = A->nb * ib
- * ztsmlq = 2 * A->nb * ib
+ /*
+ * zunmqr = A->nb * ib
+ * ztpmqrt = 2 * A->nb * ib
*/
ws_worker = chameleon_max( ws_worker, ib * A->nb * 2 );
#endif
- /* Initialisation of tiles */
-
- tiles = (int*)calloc(qrtree->mt, sizeof(int));
-
ws_worker *= sizeof(CHAMELEON_Complex64_t);
ws_host *= sizeof(CHAMELEON_Complex64_t);
RUNTIME_options_ws_alloc( &options, ws_worker, ws_host );
+ /* Initialisation of temporary tiles array */
+ tiles = (int*)calloc(qrtree->mt, sizeof(int));
+
K = chameleon_min(A->mt, A->nt);
- /* The number of the factorization */
for (k = 0; k < K; k++) {
RUNTIME_iteration_push(chamctxt, k);
@@ -98,9 +91,11 @@ void chameleon_pzgelqf_param( int genD, const libhqr_tree_t *qrtree, CHAM_desc_t
ldak = BLKLDD(A, k);
lddk = BLKLDD(D, k);
- T = TS;
/* The number of geqrt to apply */
- for (i = 0; i < qrtree->getnbgeqrf(qrtree, k); i++) {
+ nbgeqrt = qrtree->getnbgeqrf(qrtree, k);
+
+ T = TS;
+ for (i = 0; i < nbgeqrt; i++) {
p = qrtree->getm(qrtree, k, i);
temppn = p == A->nt-1 ? A->n-p*A->nb : A->nb;
tempkmin = chameleon_min(tempkm, temppn);
@@ -110,6 +105,7 @@ void chameleon_pzgelqf_param( int genD, const libhqr_tree_t *qrtree, CHAM_desc_t
tempkm, temppn, ib, T->nb,
A( k, p), ldak,
T(k, p), T->mb);
+
if ( genD ) {
INSERT_TASK_zlacpy(
&options,
@@ -124,13 +120,14 @@ void chameleon_pzgelqf_param( int genD, const libhqr_tree_t *qrtree, CHAM_desc_t
D(k, p), lddk );
#endif
}
+
for (m = k+1; m < A->mt; m++) {
tempmm = m == A->mt-1 ? A->m-m*A->mb : A->mb;
ldam = BLKLDD(A, m);
INSERT_TASK_zunmlq(
&options,
ChamRight, ChamConjTrans,
- tempmm, temppn, tempkmin, ib, T->nb,
+ tempmm, temppn, tempkmin, ib, T->nb,
D(k, p), lddk,
T(k, p), T->mb,
A(m, p), ldam);
@@ -140,15 +137,15 @@ void chameleon_pzgelqf_param( int genD, const libhqr_tree_t *qrtree, CHAM_desc_t
}
/* Setting the order of the tiles */
- libhqr_walk_stepk( qrtree, k, tiles + (k+1) );
+ nbtiles = libhqr_walk_stepk( qrtree, k, tiles );
- for (i = k+1; i < A->nt; i++) {
+ for (i = 0; i < nbtiles; i++) {
n = tiles[i];
p = qrtree->currpiv(qrtree, k, n);
tempnn = n == A->nt-1 ? A->n-n*A->nb : A->nb;
- if (qrtree->gettype(qrtree, k, n) == 0) {
+ if ( qrtree->gettype(qrtree, k, n) == LIBHQR_KILLED_BY_TS ) {
/* TS kernel */
T = TS;
L = 0;
@@ -159,10 +156,9 @@ void chameleon_pzgelqf_param( int genD, const libhqr_tree_t *qrtree, CHAM_desc_t
L = tempnn;
}
- RUNTIME_data_migrate( sequence, A(k, p),
- A->get_rankof( A, k, n ) );
- RUNTIME_data_migrate( sequence, A(k, n),
- A->get_rankof( A, k, n ) );
+ node = A->get_rankof( A, k, n );
+ RUNTIME_data_migrate( sequence, A(k, p), node );
+ RUNTIME_data_migrate( sequence, A(k, n), node );
INSERT_TASK_ztplqt(
&options,
@@ -175,10 +171,9 @@ void chameleon_pzgelqf_param( int genD, const libhqr_tree_t *qrtree, CHAM_desc_t
tempmm = m == A->mt-1 ? A->m-m*A->mb : A->mb;
ldam = BLKLDD(A, m);
- RUNTIME_data_migrate( sequence, A(m, p),
- A->get_rankof( A, m, n ) );
- RUNTIME_data_migrate( sequence, A(m, n),
- A->get_rankof( A, m, n ) );
+ node = A->get_rankof( A, m, n );
+ RUNTIME_data_migrate( sequence, A(m, p), node );
+ RUNTIME_data_migrate( sequence, A(m, n), node );
INSERT_TASK_ztpmlqt(
&options,
diff --git a/compute/pzgelqfrh.c b/compute/pzgelqfrh.c
index 6dd19a90e..55c3121b2 100644
--- a/compute/pzgelqfrh.c
+++ b/compute/pzgelqfrh.c
@@ -46,7 +46,7 @@ void chameleon_pzgelqfrh( int genD, int BS, CHAM_desc_t *A, CHAM_desc_t *T, CHAM
int K, N, RD;
int ldak, ldam, lddk;
int tempkmin, tempkm, tempNn, tempnn, tempmm, tempNRDn;
- int ib;
+ int ib, node;
chamctxt = chameleon_context_self();
if (sequence->status != CHAMELEON_SUCCESS)
@@ -167,10 +167,9 @@ void chameleon_pzgelqfrh( int genD, int BS, CHAM_desc_t *A, CHAM_desc_t *T, CHAM
for (N = k; N+RD < A->nt; N += 2*RD) {
tempNRDn = N+RD == A->nt-1 ? A->n-(N+RD)*A->nb : A->nb;
- RUNTIME_data_migrate( sequence, A(k, N),
- A->get_rankof( A, k, N+RD ) );
- RUNTIME_data_migrate( sequence, A(k, N+RD),
- A->get_rankof( A, k, N+RD ) );
+ node = A->get_rankof( A, k, N+RD );
+ RUNTIME_data_migrate( sequence, A(k, N), node );
+ RUNTIME_data_migrate( sequence, A(k, N+RD), node );
/* TT kernel */
INSERT_TASK_ztplqt(
@@ -184,10 +183,9 @@ void chameleon_pzgelqfrh( int genD, int BS, CHAM_desc_t *A, CHAM_desc_t *T, CHAM
tempmm = m == A->mt-1 ? A->m-m*A->mb : A->mb;
ldam = BLKLDD(A, m );
- RUNTIME_data_migrate( sequence, A(m, N),
- A->get_rankof( A, m, N+RD ) );
- RUNTIME_data_migrate( sequence, A(m, N+RD),
- A->get_rankof( A, m, N+RD ) );
+ node = A->get_rankof( A, m, N+RD );
+ RUNTIME_data_migrate( sequence, A(m, N), node );
+ RUNTIME_data_migrate( sequence, A(m, N+RD), node );
INSERT_TASK_ztpmlqt(
&options,
diff --git a/compute/pzgeqrf_param.c b/compute/pzgeqrf_param.c
index 01c0a816f..cc63e3a18 100644
--- a/compute/pzgeqrf_param.c
+++ b/compute/pzgeqrf_param.c
@@ -22,16 +22,21 @@
#include <stdlib.h>
#include "libhqr.h"
-#define A(m,n) A, (m), (n)
-#define T(m,n) T, (m), (n)
-#define D(m,n) D, (m), (n)
+#define A(m,n) A, (m), (n)
+#define T(m,n) T, (m), (n)
+#define D(m,n) D, (m), (n)
/**
* 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.
*/
-void chameleon_pzgeqrf_param( int genD, const libhqr_tree_t *qrtree, CHAM_desc_t *A,
+void chameleon_pzgeqrf_param( int genD, int K,
+ 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)
+ RUNTIME_sequence_t *sequence, RUNTIME_request_t *request )
{
CHAM_context_t *chamctxt;
RUNTIME_option_t options;
@@ -40,11 +45,10 @@ void chameleon_pzgeqrf_param( int genD, const libhqr_tree_t *qrtree, CHAM_desc_t
size_t ws_host = 0;
int k, m, n, i, p;
- int K, L, nbgeqrt;
+ int L, nbgeqrt;
int ldap, ldam, lddm;
int tempkmin, tempkn, tempnn, tempmm;
- int ib;
- int *tiles;
+ int ib, node, nbtiles, *tiles;
chamctxt = chameleon_context_self();
if (sequence->status != CHAMELEON_SUCCESS)
@@ -75,23 +79,22 @@ void chameleon_pzgeqrf_param( int genD, const libhqr_tree_t *qrtree, CHAM_desc_t
ws_worker = chameleon_max( ws_worker, ib * A->nb * 2 );
#endif
- /* Initialisation of temporary tiles array */
- tiles = (int*)calloc(qrtree->mt, sizeof(int));
-
ws_worker *= sizeof(CHAMELEON_Complex64_t);
ws_host *= sizeof(CHAMELEON_Complex64_t);
RUNTIME_options_ws_alloc( &options, ws_worker, ws_host );
- K = chameleon_min(A->mt, A->nt);
+ /* Initialisation of temporary tiles array */
+ tiles = (int*)calloc(qrtree->mt, sizeof(int));
- /* The number of the factorization */
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;
@@ -99,13 +102,12 @@ void chameleon_pzgeqrf_param( int genD, const libhqr_tree_t *qrtree, CHAM_desc_t
ldam = BLKLDD(A, m);
lddm = BLKLDD(D, m);
- T = TS;
-
INSERT_TASK_zgeqrt(
&options,
tempmm, tempkn, ib, T->nb,
A(m, k), ldam,
T(m, k), T->mb);
+
if ( genD ) {
INSERT_TASK_zlacpy(
&options,
@@ -120,6 +122,7 @@ void chameleon_pzgeqrf_param( int genD, const libhqr_tree_t *qrtree, CHAM_desc_t
D(m, k), lddm );
#endif
}
+
for (n = k+1; n < A->nt; n++) {
tempnn = n == A->nt-1 ? A->n-n*A->nb : A->nb;
INSERT_TASK_zunmqr(
@@ -135,9 +138,9 @@ void chameleon_pzgeqrf_param( int genD, const libhqr_tree_t *qrtree, CHAM_desc_t
}
/* Setting the order of the tiles */
- libhqr_walk_stepk( qrtree, k, tiles + (k+1) );
+ nbtiles = libhqr_walk_stepk( qrtree, k, tiles );
- for (i = k+1; i < A->mt; i++) {
+ for (i = 0; i < nbtiles; i++) {
m = tiles[i];
p = qrtree->currpiv(qrtree, k, m);
@@ -145,7 +148,7 @@ void chameleon_pzgeqrf_param( int genD, const libhqr_tree_t *qrtree, CHAM_desc_t
ldap = BLKLDD(A, p);
ldam = BLKLDD(A, m);
- if (qrtree->gettype(qrtree, k, m) == 0) {
+ if ( qrtree->gettype(qrtree, k, m) == LIBHQR_KILLED_BY_TS ) {
/* TS kernel */
T = TS;
L = 0;
@@ -156,10 +159,9 @@ void chameleon_pzgeqrf_param( int genD, const libhqr_tree_t *qrtree, CHAM_desc_t
L = tempmm;
}
- RUNTIME_data_migrate( sequence, A(p, k),
- A->get_rankof( A, m, k ) );
- RUNTIME_data_migrate( sequence, A(m, k),
- A->get_rankof( A, m, k ) );
+ 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,
@@ -171,10 +173,9 @@ void chameleon_pzgeqrf_param( int genD, const libhqr_tree_t *qrtree, CHAM_desc_t
for (n = k+1; n < A->nt; n++) {
tempnn = n == A->nt-1 ? A->n-n*A->nb : A->nb;
- RUNTIME_data_migrate( sequence, A(p, n),
- A->get_rankof( A, m, n ) );
- RUNTIME_data_migrate( sequence, A(m, n),
- A->get_rankof( A, m, n ) );
+ 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,
diff --git a/compute/pzgeqrfrh.c b/compute/pzgeqrfrh.c
index cab40dd42..1eeb7cdbc 100644
--- a/compute/pzgeqrfrh.c
+++ b/compute/pzgeqrfrh.c
@@ -46,7 +46,7 @@ void chameleon_pzgeqrfrh( int genD, int BS, CHAM_desc_t *A, CHAM_desc_t *T, CHAM
int K, M, RD;
int ldaM, ldam, ldaMRD, lddM;
int tempkmin, tempkn, tempMm, tempnn, tempmm, tempMRDm;
- int ib;
+ int ib, node;
chamctxt = chameleon_context_self();
if (sequence->status != CHAMELEON_SUCCESS)
@@ -166,10 +166,9 @@ void chameleon_pzgeqrfrh( int genD, int BS, CHAM_desc_t *A, CHAM_desc_t *T, CHAM
ldaM = BLKLDD(A, M );
ldaMRD = BLKLDD(A, M+RD);
- RUNTIME_data_migrate( sequence, A(M, k),
- A->get_rankof( A, M+RD, k ) );
- RUNTIME_data_migrate( sequence, A(M+RD, k),
- A->get_rankof( A, M+RD, k ) );
+ node = A->get_rankof( A, M+RD, k );
+ RUNTIME_data_migrate( sequence, A(M, k), node );
+ RUNTIME_data_migrate( sequence, A(M+RD, k), node );
/* TT kernel */
INSERT_TASK_ztpqrt(
@@ -182,10 +181,9 @@ void chameleon_pzgeqrfrh( int genD, int BS, CHAM_desc_t *A, CHAM_desc_t *T, CHAM
for (n = k+1; n < A->nt; n++) {
tempnn = n == A->nt-1 ? A->n-n*A->nb : A->nb;
- RUNTIME_data_migrate( sequence, A(M, n),
- A->get_rankof( A, M+RD, n ) );
- RUNTIME_data_migrate( sequence, A(M+RD, n),
- A->get_rankof( A, M+RD, n ) );
+ node = A->get_rankof( A, M+RD, n );
+ RUNTIME_data_migrate( sequence, A(M, n), node );
+ RUNTIME_data_migrate( sequence, A(M+RD, n), node );
INSERT_TASK_ztpmqrt(
&options,
diff --git a/compute/pztpgqrt.c b/compute/pztpgqrt.c
index f2a68870a..f6062aa3d 100644
--- a/compute/pztpgqrt.c
+++ b/compute/pztpgqrt.c
@@ -20,8 +20,6 @@
*/
#include "control/common.h"
-#define V1(m,n) V1, m, n
-#define T1(m,n) T1, m, n
#define V2(m,n) V2, m, n
#define T2(m,n) T2, m, n
#define Q1(m,n) Q1, m, n
@@ -31,11 +29,9 @@
/**
* Parallel tile QR factorization - dynamic scheduling
*/
-void chameleon_pztpgqrt( int genD, int L,
- CHAM_desc_t *V1, CHAM_desc_t *T1,
+void chameleon_pztpgqrt( int KT, int L,
CHAM_desc_t *V2, CHAM_desc_t *T2,
CHAM_desc_t *Q1, CHAM_desc_t *Q2,
- CHAM_desc_t *D,
RUNTIME_sequence_t *sequence, RUNTIME_request_t *request )
{
CHAM_context_t *chamctxt;
@@ -46,7 +42,7 @@ void chameleon_pztpgqrt( int genD, int L,
int k, m, n;
int ldvk, ldvm, lddk;
int ldqk, ldqm;
- int tempkm, tempkn, tempkk, tempnn, tempmm, templm;
+ int tempkn, tempnn, tempmm, templm;
int ib;
/* Dimension of the first column */
@@ -61,11 +57,6 @@ void chameleon_pztpgqrt( int genD, int L,
ib = CHAMELEON_IB;
- if ( D == NULL ) {
- D = V1;
- genD = 0;
- }
-
/*
* ztpmqrt = Q1->nb * ib
*/
@@ -85,21 +76,17 @@ void chameleon_pztpgqrt( int genD, int L,
RUNTIME_options_ws_alloc( &options, ws_worker, ws_host );
- for (k = V1->nt-1; k >= 0; k--) {
+ for (k = KT-1; k >= 0; k--) {
RUNTIME_iteration_push(chamctxt, k);
- tempkm = k == V1->mt-1 ? V1->m-k*V1->mb : V1->mb;
- tempkk = k == V1->nt-1 ? V1->n-k*V1->nb : V1->nb;
tempkn = k == Q1->nt-1 ? Q1->n-k*Q1->nb : Q1->nb;
- ldvk = BLKLDD(V1, k);
- lddk = BLKLDD(D, k);
ldqk = BLKLDD(Q1, k);
/* Equivalent to the tsmqr step on Q1,Q2 */
maxmtk = chameleon_min( Q2->mt, maxmt+k ) - 1;
for (m = maxmtk; m > -1; m--) {
tempmm = m == Q2->mt-1 ? Q2->m-m*Q2->mb : Q2->mb;
- templm = m == maxmtk ? tempmm : 0;
+ templm = ((L > 0) && (m == maxmtk)) ? tempmm : 0;
ldvm = BLKLDD(V2, m);
ldqm = BLKLDD(Q2, m);
@@ -117,53 +104,9 @@ void chameleon_pztpgqrt( int genD, int L,
}
}
- for (m = Q1->mt - 1; m > k; m--) {
- tempmm = m == Q1->mt-1 ? Q1->m-m*Q1->mb : Q1->mb;
- ldvm = BLKLDD(V1, m);
- ldqm = BLKLDD(Q1, m);
- for (n = k; n < Q1->nt; n++) {
- tempnn = n == Q1->nt-1 ? Q1->n-n*Q1->nb : Q1->nb;
- /* TS kernel */
- INSERT_TASK_ztpmqrt(
- &options,
- ChamLeft, ChamNoTrans,
- tempmm, tempnn, tempkn, 0, ib, T1->nb,
- V1(m, k), ldvm,
- T1(m, k), T1->mb,
- Q1(k, n), ldqk,
- Q1(m, n), ldqm );
- }
- }
-
- if ( genD ) {
- INSERT_TASK_zlacpy(
- &options,
- ChamLower, tempkm, tempkk, V1->nb,
- V1(k, k), ldvk,
- D(k), lddk );
-#if defined(CHAMELEON_USE_CUDA)
- INSERT_TASK_zlaset(
- &options,
- ChamUpper, tempkm, tempkk,
- 0., 1.,
- D(k), lddk );
-#endif
- }
- for (n = k; n < Q1->nt; n++) {
- tempnn = n == Q1->nt-1 ? Q1->n-n*Q1->nb : Q1->nb;
- INSERT_TASK_zunmqr(
- &options,
- ChamLeft, ChamNoTrans,
- tempkm, tempnn, tempkk, ib, T1->nb,
- D(k), lddk,
- T1(k, k), T1->mb,
- Q1(k, n), ldqk);
- }
-
RUNTIME_iteration_pop(chamctxt);
}
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, chamctxt);
- (void)D;
}
diff --git a/compute/pztpqrt.c b/compute/pztpqrt.c
index fb2af03ec..37de659fe 100644
--- a/compute/pztpqrt.c
+++ b/compute/pztpqrt.c
@@ -20,9 +20,9 @@
*/
#include "control/common.h"
-#define A(m,n) A, m, n
-#define B(m,n) B, m, n
-#define T(m,n) T, m, n
+#define A(m,n) A, (m), (n)
+#define B(m,n) B, (m), (n)
+#define T(m,n) T, (m), (n)
/**
* Parallel tile QR factorization - dynamic scheduling
@@ -53,15 +53,14 @@ void chameleon_pztpqrt( int L, CHAM_desc_t *A, CHAM_desc_t *B, CHAM_desc_t *T,
ib = CHAMELEON_IB;
/*
- * ztsqrt = A->nb * (ib+1)
+ * ztpqrt = A->nb * (ib+1)
* ztpmqrt = A->nb * ib
*/
ws_worker = A->nb * (ib+1);
/* Allocation of temporary (scratch) working space */
#if defined(CHAMELEON_USE_CUDA)
- /* Worker space
- *
+ /*
* ztpmqrt = 2 * A->nb * ib
*/
ws_worker = chameleon_max( ws_worker, ib * A->nb * 2 );
@@ -81,7 +80,7 @@ void chameleon_pztpqrt( int L, CHAM_desc_t *A, CHAM_desc_t *B, CHAM_desc_t *T,
for (m = 0; m < maxmt; m++) {
tempmm = m == B->mt-1 ? B->m-m*B->mb : B->mb;
- templm = m == maxmt-1 ? tempmm : 0;
+ templm = ((L > 0) && (m == maxmt-1)) ? tempmm : 0;
ldbm = BLKLDD(B, m);
/* TT kernel */
INSERT_TASK_ztpqrt(
diff --git a/compute/pzunglq_param.c b/compute/pzunglq_param.c
index 4a82007a5..a6cc32e6f 100644
--- a/compute/pzunglq_param.c
+++ b/compute/pzunglq_param.c
@@ -43,8 +43,7 @@ void chameleon_pzunglq_param( int genD, const libhqr_tree_t *qrtree, CHAM_desc_t
int K, L;
int ldak, ldqm, lddk;
int tempkm, tempkmin, temppn, tempnn, tempmm;
- int ib;
- int *tiles;
+ int ib, node, nbtiles, *tiles;
chamctxt = chameleon_context_self();
if (sequence->status != CHAMELEON_SUCCESS)
@@ -77,15 +76,14 @@ void chameleon_pzunglq_param( int genD, const libhqr_tree_t *qrtree, CHAM_desc_t
ws_worker = chameleon_max( ws_worker, ib * A->nb * 2 );
#endif
- /* Initialisation of tiles */
-
- tiles = (int*)calloc( qrtree->mt, sizeof(int));
-
ws_worker *= sizeof(CHAMELEON_Complex64_t);
ws_host *= sizeof(CHAMELEON_Complex64_t);
RUNTIME_options_ws_alloc( &options, ws_worker, ws_host );
+ /* Initialisation of tiles */
+ tiles = (int*)calloc( qrtree->mt, sizeof(int));
+
K = chameleon_min(A->mt, A->nt);
for (k = K-1; k >= 0; k--) {
@@ -96,15 +94,15 @@ void chameleon_pzunglq_param( int genD, const libhqr_tree_t *qrtree, CHAM_desc_t
lddk = BLKLDD(D, k);
/* Setting the order of the tiles*/
- libhqr_walk_stepk(qrtree, k, tiles + (k+1));
+ nbtiles = libhqr_walk_stepk( qrtree, k, tiles );
- for (i = A->nt-1; i > k; i--) {
+ for (i = nbtiles-1; i >= 0; i--) {
n = tiles[i];
p = qrtree->currpiv(qrtree, k, n);
tempnn = n == Q->nt-1 ? Q->n-n*Q->nb : Q->nb;
- if(qrtree->gettype(qrtree, k, n) == 0){
+ if( qrtree->gettype(qrtree, k, n) == LIBHQR_KILLED_BY_TS ) {
/* TS kernel */
L = 0;
T = TS;
@@ -118,10 +116,9 @@ void chameleon_pzunglq_param( int genD, const libhqr_tree_t *qrtree, CHAM_desc_t
tempmm = m == Q->mt-1 ? Q->m-m*Q->mb : Q->mb;
ldqm = BLKLDD(Q, m);
- RUNTIME_data_migrate( sequence, Q(m, p),
- Q->get_rankof( Q, m, n ) );
- RUNTIME_data_migrate( sequence, Q(m, n),
- Q->get_rankof( Q, m, n ) );
+ node = Q->get_rankof( Q, m, n );
+ RUNTIME_data_migrate( sequence, Q(m, p), node );
+ RUNTIME_data_migrate( sequence, Q(m, n), node );
INSERT_TASK_ztpmlqt(
&options,
diff --git a/compute/pzunglqrh.c b/compute/pzunglqrh.c
index 4cde3dc14..e8a6bdac9 100644
--- a/compute/pzunglqrh.c
+++ b/compute/pzunglqrh.c
@@ -48,7 +48,7 @@ void chameleon_pzunglqrh( int genD, int BS,
int K, N, RD, lastRD;
int ldak, lddk, ldqm;
int tempkm, tempkmin, tempNn, tempnn, tempmm, tempNRDn;
- int ib;
+ int ib, node;
chamctxt = chameleon_context_self();
if (sequence->status != CHAMELEON_SUCCESS)
@@ -99,10 +99,9 @@ void chameleon_pzunglqrh( int genD, int BS,
tempmm = m == Q->mt-1 ? Q->m-m*Q->mb : Q->mb;
ldqm = BLKLDD(Q, m );
- RUNTIME_data_migrate( sequence, Q(m, N),
- Q->get_rankof( Q, m, N+RD ) );
- RUNTIME_data_migrate( sequence, Q(m, N+RD),
- Q->get_rankof( Q, m, N+RD ) );
+ node = Q->get_rankof( Q, m, N+RD );
+ RUNTIME_data_migrate( sequence, Q(m, N), node );
+ RUNTIME_data_migrate( sequence, Q(m, N+RD), node );
/* TT kernel */
INSERT_TASK_ztpmlqt(
@@ -129,10 +128,9 @@ void chameleon_pzunglqrh( int genD, int BS,
tempmm = m == Q->mt-1 ? Q->m-m*Q->mb : Q->mb;
ldqm = BLKLDD(Q, m);
- RUNTIME_data_migrate( sequence, Q(m, N),
- Q->get_rankof( Q, m, n ) );
- RUNTIME_data_migrate( sequence, Q(m, n),
- Q->get_rankof( Q, m, n ) );
+ node = Q->get_rankof( Q, m, n );
+ RUNTIME_data_migrate( sequence, Q(m, N), node );
+ RUNTIME_data_migrate( sequence, Q(m, n), node );
/* TS kernel */
INSERT_TASK_ztpmlqt(
diff --git a/compute/pzungqr_param.c b/compute/pzungqr_param.c
index 0bf5b2826..d46024821 100644
--- a/compute/pzungqr_param.c
+++ b/compute/pzungqr_param.c
@@ -29,7 +29,8 @@
/**
* Parallel construction of Q using tile V (application to identity) - dynamic scheduling
*/
-void chameleon_pzungqr_param( int genD, const libhqr_tree_t *qrtree,
+void chameleon_pzungqr_param( int genD, int K,
+ const libhqr_tree_t *qrtree,
CHAM_desc_t *A, CHAM_desc_t *Q,
CHAM_desc_t *TS, CHAM_desc_t *TT, CHAM_desc_t *D,
RUNTIME_sequence_t *sequence, RUNTIME_request_t *request )
@@ -43,8 +44,7 @@ void chameleon_pzungqr_param( int genD, const libhqr_tree_t *qrtree,
int k, m, n, i, p, L;
int ldam, ldqm, ldqp, lddm;
int tempmm, tempnn, tempkmin, tempkn;
- int ib, minMT;
- int *tiles;
+ int ib, nbgeqrt, node, nbtiles, *tiles;
chamctxt = chameleon_context_self();
if (sequence->status != CHAMELEON_SUCCESS)
@@ -53,12 +53,6 @@ void chameleon_pzungqr_param( int genD, const libhqr_tree_t *qrtree,
ib = CHAMELEON_IB;
- if (A->m > A->n) {
- minMT = A->nt;
- } else {
- minMT = A->mt;
- }
-
if (D == NULL) {
D = A;
genD = 0;
@@ -66,47 +60,44 @@ void chameleon_pzungqr_param( int genD, const libhqr_tree_t *qrtree,
/*
* zunmqr = A->nb * ib
- * ztsmqr = A->nb * ib
+ * ztpmqr = A->nb * ib
*/
ws_worker = A->nb * ib;
/* Allocation of temporary (scratch) working space */
#if defined(CHAMELEON_USE_CUDA)
- /* Worker space
- *
- * zunmqr = A->nb * ib
- * ztsmqr = 2 * A->nb * ib
+ /*
+ * ztpmqrt = 2 * A->nb * ib
*/
ws_worker = chameleon_max( ws_worker, ib * A->nb * 2 );
#endif
- /* Initialisation of tiles */
-
- tiles = (int*)calloc(qrtree->mt, sizeof(int));
-
ws_worker *= sizeof(CHAMELEON_Complex64_t);
ws_host *= sizeof(CHAMELEON_Complex64_t);
RUNTIME_options_ws_alloc( &options, ws_worker, ws_host );
- for (k = minMT-1; k >= 0; k--) {
+ /* Initialisation of temporary tiles array */
+ tiles = (int*)calloc(qrtree->mt, sizeof(int));
+
+ 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 */
- libhqr_walk_stepk(qrtree, k, tiles + (k+1));
+ nbtiles = libhqr_walk_stepk( qrtree, k, tiles );
- for (i = Q->mt-1; i > k; i--) {
+ 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;
+ ldqp = BLKLDD(Q, p);
ldam = BLKLDD(A, m);
ldqm = BLKLDD(Q, m);
- ldqp = BLKLDD(Q, p);
- if(qrtree->gettype(qrtree, k , m) == 0) {
+ if( qrtree->gettype(qrtree, k, m) == LIBHQR_KILLED_BY_TS ) {
/* TS kernel */
T = TS;
L = 0;
@@ -120,10 +111,9 @@ void chameleon_pzungqr_param( int genD, const libhqr_tree_t *qrtree,
for (n = k; n < Q->nt; n++) {
tempnn = n == Q->nt-1 ? Q->n-n*Q->nb : Q->nb;
- RUNTIME_data_migrate( sequence, Q(p, n),
- Q->get_rankof( Q, m, n ) );
- RUNTIME_data_migrate( sequence, Q(m, n),
- Q->get_rankof( Q, m, n ) );
+ 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,
@@ -139,7 +129,10 @@ void chameleon_pzungqr_param( int genD, const libhqr_tree_t *qrtree,
}
T = TS;
- for (i = 0; i < qrtree->getnbgeqrf(qrtree, k); i++) {
+
+ /* 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;
diff --git a/compute/pzungqrrh.c b/compute/pzungqrrh.c
index 78f218525..1b9036e6c 100644
--- a/compute/pzungqrrh.c
+++ b/compute/pzungqrrh.c
@@ -51,7 +51,7 @@ void chameleon_pzungqrrh( int genD, int BS,
int ldaM, ldam, ldaMRD, lddM;
int ldqM, ldqm, ldqMRD;
int tempkn, tempMm, tempnn, tempmm, tempMRDm, tempkmin;
- int ib;
+ int ib, node;
chamctxt = chameleon_context_self();
if (sequence->status != CHAMELEON_SUCCESS)
@@ -68,7 +68,7 @@ void chameleon_pzungqrrh( int genD, int BS,
/*
* zunmqr = A->nb * ib
* ztsmqr = A->nb * ib
- * zttmqr = A->nb * ib
+ * ztpmqrt = A->nb * ib
*/
ws_worker = A->nb * ib;
@@ -103,10 +103,9 @@ void chameleon_pzungqrrh( int genD, int BS,
for (n = k; n < Q->nt; n++) {
tempnn = n == Q->nt-1 ? Q->n-n*Q->nb : Q->nb;
- RUNTIME_data_migrate( sequence, Q(M, n),
- Q->get_rankof( Q, M+RD, n ) );
- RUNTIME_data_migrate( sequence, Q(M+RD, n),
- Q->get_rankof( Q, M+RD, n ) );
+ node = Q->get_rankof( Q, M+RD, n );
+ RUNTIME_data_migrate( sequence, Q(M, n), node );
+ RUNTIME_data_migrate( sequence, Q(M+RD, n), node );
/* TT kernel */
INSERT_TASK_ztpmqrt(
@@ -137,10 +136,9 @@ void chameleon_pzungqrrh( int genD, int BS,
for (n = k; n < Q->nt; n++) {
tempnn = n == Q->nt-1 ? Q->n-n*Q->nb : Q->nb;
- RUNTIME_data_migrate( sequence, Q(M, n),
- Q->get_rankof( Q, m, n ) );
- RUNTIME_data_migrate( sequence, Q(m, n),
- Q->get_rankof( Q, m, n ) );
+ node = Q->get_rankof( Q, m, n );
+ RUNTIME_data_migrate( sequence, Q(M, n), node );
+ RUNTIME_data_migrate( sequence, Q(m, n), node );
/* TS kernel */
INSERT_TASK_ztpmqrt(
diff --git a/compute/pzunmlq_param.c b/compute/pzunmlq_param.c
index 82cbe0c04..545431e82 100644
--- a/compute/pzunmlq_param.c
+++ b/compute/pzunmlq_param.c
@@ -45,7 +45,7 @@ void chameleon_pzunmlq_param( int genD, const libhqr_tree_t *qrtree,
int ldbm, ldak, ldbp, lddk;
int tempnn, temppn, tempkmin, tempmm, tempkm;
int ib, K, L;
- int *tiles;
+ int node, nbtiles, *tiles;
chamctxt = chameleon_context_self();
if (sequence->status != CHAMELEON_SUCCESS)
@@ -77,14 +77,14 @@ void chameleon_pzunmlq_param( int genD, const libhqr_tree_t *qrtree,
ws_worker = chameleon_max( ws_worker, ib * A->nb * 2 );
#endif
- /* Initialisation of tiles */
- tiles = (int*)calloc( qrtree->mt, sizeof(int) );
-
ws_worker *= sizeof(CHAMELEON_Complex64_t);
ws_host *= sizeof(CHAMELEON_Complex64_t);
RUNTIME_options_ws_alloc( &options, ws_worker, ws_host );
+ /* Initialisation of tiles */
+ tiles = (int*)calloc( qrtree->mt, sizeof(int) );
+
if (side == ChamLeft ) {
if (trans == ChamNoTrans) {
/*
@@ -135,9 +135,9 @@ void chameleon_pzunmlq_param( int genD, const libhqr_tree_t *qrtree,
}
/* Setting the order of the tiles*/
- libhqr_walk_stepk(qrtree, k, tiles + (k+1));
+ nbtiles = libhqr_walk_stepk( qrtree, k, tiles );
- for (i = k+1; i < A->nt; i++) {
+ for (i = 0; i < nbtiles; i++) {
m = tiles[i];
p = qrtree->currpiv(qrtree, k, m);
@@ -145,7 +145,7 @@ void chameleon_pzunmlq_param( int genD, const libhqr_tree_t *qrtree,
ldbp = BLKLDD(B, p);
ldbm = BLKLDD(B, m);
- if(qrtree->gettype(qrtree, k, m) == 0){
+ if( qrtree->gettype(qrtree, k, m) == LIBHQR_KILLED_BY_TS ) {
/* TS kernel */
L = 0;
T = TS;
@@ -158,10 +158,9 @@ void chameleon_pzunmlq_param( int genD, const libhqr_tree_t *qrtree,
for (n = 0; n < B->nt; n++) {
tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb;
- RUNTIME_data_migrate( sequence, B(p, n),
- B->get_rankof( B, m, n ) );
- RUNTIME_data_migrate( sequence, B(m, n),
- B->get_rankof( B, m, n ) );
+ node = B->get_rankof( B, m, n );
+ RUNTIME_data_migrate( sequence, B(p, n), node );
+ RUNTIME_data_migrate( sequence, B(m, n), node );
INSERT_TASK_ztpmlqt(
&options,
@@ -197,9 +196,9 @@ void chameleon_pzunmlq_param( int genD, const libhqr_tree_t *qrtree,
lddk = BLKLDD(D, k);
/* Setting the order of the tiles*/
- libhqr_walk_stepk(qrtree, k, tiles + (k+1));
+ nbtiles = libhqr_walk_stepk( qrtree, k, tiles );
- for (i = A->nt-1; i > k; i--) {
+ for (i = nbtiles-1; i >= 0; i--) {
m = tiles[i];
p = qrtree->currpiv(qrtree, k, m);
@@ -207,7 +206,7 @@ void chameleon_pzunmlq_param( int genD, const libhqr_tree_t *qrtree,
ldbp = BLKLDD(B, p);
ldbm = BLKLDD(B, m);
- if(qrtree->gettype(qrtree, k, m) == 0){
+ if( qrtree->gettype(qrtree, k, m) == LIBHQR_KILLED_BY_TS ) {
/* TS kernel */
L = 0;
T = TS;
@@ -220,10 +219,9 @@ void chameleon_pzunmlq_param( int genD, const libhqr_tree_t *qrtree,
for (n = 0; n < B->nt; n++) {
tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb;
- RUNTIME_data_migrate( sequence, B(p, n),
- B->get_rankof( B, m, n ) );
- RUNTIME_data_migrate( sequence, B(m, n),
- B->get_rankof( B, m, n ) );
+ node = B->get_rankof( B, m, n );
+ RUNTIME_data_migrate( sequence, B(p, n), node );
+ RUNTIME_data_migrate( sequence, B(m, n), node );
INSERT_TASK_ztpmlqt(
&options,
@@ -296,15 +294,15 @@ void chameleon_pzunmlq_param( int genD, const libhqr_tree_t *qrtree,
lddk = BLKLDD(D, k);
/* Setting the order of the tiles*/
- libhqr_walk_stepk(qrtree, k, tiles + (k+1));
+ nbtiles = libhqr_walk_stepk( qrtree, k, tiles );
- for (i = A->nt-1; i > k; i--) {
+ for (i = nbtiles-1; i >= 0; i--) {
n = tiles[i];
p = qrtree->currpiv(qrtree, k, n);
tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb;
- if(qrtree->gettype(qrtree, k, n) == 0){
+ if( qrtree->gettype(qrtree, k, n) == LIBHQR_KILLED_BY_TS ) {
/* TS kernel */
L = 0;
T = TS;
@@ -318,10 +316,9 @@ void chameleon_pzunmlq_param( int genD, const libhqr_tree_t *qrtree,
tempmm = m == B->mt-1 ? B->m-m*B->mb : B->mb;
ldbm = BLKLDD(B, m);
- RUNTIME_data_migrate( sequence, B(m, p),
- B->get_rankof( B, m, n ) );
- RUNTIME_data_migrate( sequence, B(m, n),
- B->get_rankof( B, m, n ) );
+ node = B->get_rankof( B, m, n );
+ RUNTIME_data_migrate( sequence, B(m, p), node );
+ RUNTIME_data_migrate( sequence, B(m, n), node );
INSERT_TASK_ztpmlqt(
&options,
@@ -429,15 +426,15 @@ void chameleon_pzunmlq_param( int genD, const libhqr_tree_t *qrtree,
}
/* Setting the order of tiles */
- libhqr_walk_stepk(qrtree, k, tiles + (k+1));
+ nbtiles = libhqr_walk_stepk( qrtree, k, tiles );
- for (i = k+1; i < A->nt; i++) {
+ for (i = 0; i < nbtiles; i++) {
n = tiles[i];
p = qrtree->currpiv(qrtree, k, n);
tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb;
- if(qrtree->gettype(qrtree, k, n) == 0){
+ if( qrtree->gettype(qrtree, k, n) == LIBHQR_KILLED_BY_TS ) {
/* TS kernel */
L = 0;
T = TS;
@@ -452,10 +449,9 @@ void chameleon_pzunmlq_param( int genD, const libhqr_tree_t *qrtree,
tempmm = m == B->mt-1 ? B->m-m*B->mb : B->mb;
ldbm = BLKLDD(B, m);
- RUNTIME_data_migrate( sequence, B(m, p),
- B->get_rankof( B, m, n ) );
- RUNTIME_data_migrate( sequence, B(m, n),
- B->get_rankof( B, m, n ) );
+ node = B->get_rankof( B, m, n );
+ RUNTIME_data_migrate( sequence, B(m, p), node );
+ RUNTIME_data_migrate( sequence, B(m, n), node );
INSERT_TASK_ztpmlqt(
&options,
diff --git a/compute/pzunmlqrh.c b/compute/pzunmlqrh.c
index 19f411ac4..3ad2c4eae 100644
--- a/compute/pzunmlqrh.c
+++ b/compute/pzunmlqrh.c
@@ -49,7 +49,7 @@ void chameleon_pzunmlqrh( int genD, int BS, cham_side_t side, cham_trans_t trans
int K, N, RD, lastRD;
int ldak, lddk, ldbN, ldbm, ldbNRD;
int tempNn, tempkm, tempnn, tempmm, tempNRDn, tempkmin;
- int ib;
+ int ib, node;
chamctxt = chameleon_context_self();
if (sequence->status != CHAMELEON_SUCCESS)
@@ -135,10 +135,9 @@ void chameleon_pzunmlqrh( int genD, int BS, cham_side_t side, cham_trans_t trans
for (n = 0; n < B->nt; n++) {
tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb;
- RUNTIME_data_migrate( sequence, B(N, n),
- B->get_rankof( B, m, n ) );
- RUNTIME_data_migrate( sequence, B(m, n),
- B->get_rankof( B, m, n ) );
+ node = B->get_rankof( B, m, n );
+ RUNTIME_data_migrate( sequence, B(N, n), node );
+ RUNTIME_data_migrate( sequence, B(m, n), node );
/* TS kernel */
INSERT_TASK_ztpmlqt(
@@ -161,10 +160,9 @@ void chameleon_pzunmlqrh( int genD, int BS, cham_side_t side, cham_trans_t trans
for (n = 0; n < B->nt; n++) {
tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb;
- RUNTIME_data_migrate( sequence, B(N, n),
- B->get_rankof( B, N+RD, n ) );
- RUNTIME_data_migrate( sequence, B(N+RD, n),
- B->get_rankof( B, N+RD, n ) );
+ node = B->get_rankof( B, N+RD, n );
+ RUNTIME_data_migrate( sequence, B(N, n), node );
+ RUNTIME_data_migrate( sequence, B(N+RD, n), node );
/* TT kernel */
INSERT_TASK_ztpmlqt(
@@ -210,10 +208,9 @@ void chameleon_pzunmlqrh( int genD, int BS, cham_side_t side, cham_trans_t trans
for (n = 0; n < B->nt; n++) {
tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb;
- RUNTIME_data_migrate( sequence, B(N, n),
- B->get_rankof( B, N+RD, n ) );
- RUNTIME_data_migrate( sequence, B(N+RD, n),
- B->get_rankof( B, N+RD, n ) );
+ node = B->get_rankof( B, N+RD, n );
+ RUNTIME_data_migrate( sequence, B(N, n), node );
+ RUNTIME_data_migrate( sequence, B(N+RD, n), node );
/* TT kernel */
INSERT_TASK_ztpmlqt(
@@ -239,10 +236,9 @@ void chameleon_pzunmlqrh( int genD, int BS, cham_side_t side, cham_trans_t trans
for (n = 0; n < B->nt; n++) {
tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb;
- RUNTIME_data_migrate( sequence, B(N, n),
- B->get_rankof( B, m, n ) );
- RUNTIME_data_migrate( sequence, B(m, n),
- B->get_rankof( B, m, n ) );
+ node = B->get_rankof( B, m, n );
+ RUNTIME_data_migrate( sequence, B(N, n), node );
+ RUNTIME_data_migrate( sequence, B(m, n), node );
/* TS kernel */
INSERT_TASK_ztpmlqt(
@@ -314,10 +310,9 @@ void chameleon_pzunmlqrh( int genD, int BS, cham_side_t side, cham_trans_t trans
ldbm = BLKLDD(B, m);
tempmm = m == B->mt-1 ? B->m-m*B->mb : B->mb;
- RUNTIME_data_migrate( sequence, B(m, N),
- B->get_rankof( B, m, N+RD ) );
- RUNTIME_data_migrate( sequence, B(m, N+RD),
- B->get_rankof( B, m, N+RD ) );
+ node = B->get_rankof( B, m, N+RD );
+ RUNTIME_data_migrate( sequence, B(m, N), node );
+ RUNTIME_data_migrate( sequence, B(m, N+RD), node );
/* TT kernel */
INSERT_TASK_ztpmlqt(
@@ -342,10 +337,9 @@ void chameleon_pzunmlqrh( int genD, int BS, cham_side_t side, cham_trans_t trans
tempmm = m == B->mt-1 ? B->m-m*B->mb : B->mb;
ldbm = BLKLDD(B, m);
- RUNTIME_data_migrate( sequence, B(m, N),
- B->get_rankof( B, m, n ) );
- RUNTIME_data_migrate( sequence, B(m, m),
- B->get_rankof( B, m, n ) );
+ node = B->get_rankof( B, m, n );
+ RUNTIME_data_migrate( sequence, B(m, N), node );
+ RUNTIME_data_migrate( sequence, B(m, m), node );
/* TS kernel */
INSERT_TASK_ztpmlqt(
@@ -444,10 +438,9 @@ void chameleon_pzunmlqrh( int genD, int BS, cham_side_t side, cham_trans_t trans
tempmm = m == B->mt-1 ? B->m-m*B->mb : B->mb;
ldbm = BLKLDD(B, m);
- RUNTIME_data_migrate( sequence, B(m, N),
- B->get_rankof( B, m, n ) );
- RUNTIME_data_migrate( sequence, B(m, n),
- B->get_rankof( B, m, n ) );
+ node = B->get_rankof( B, m, n );
+ RUNTIME_data_migrate( sequence, B(m, N), node );
+ RUNTIME_data_migrate( sequence, B(m, n), node );
/* TS kernel */
INSERT_TASK_ztpmlqt(
@@ -470,10 +463,9 @@ void chameleon_pzunmlqrh( int genD, int BS, cham_side_t side, cham_trans_t trans
tempmm = m == B->mt-1 ? B->m-m*B->mb : B->mb;
ldbm = BLKLDD(B, m);
- RUNTIME_data_migrate( sequence, B(m, N),
- B->get_rankof( B, m, N+RD ) );
- RUNTIME_data_migrate( sequence, B(m, N+RD),
- B->get_rankof( B, m, N+RD ) );
+ node = B->get_rankof( B, m, N+RD );
+ RUNTIME_data_migrate( sequence, B(m, N), node );
+ RUNTIME_data_migrate( sequence, B(m, N+RD), node );
/* TT kernel */
INSERT_TASK_ztpmlqt(
diff --git a/compute/pzunmqr_param.c b/compute/pzunmqr_param.c
index 27e8cb192..cd23e8114 100644
--- a/compute/pzunmqr_param.c
+++ b/compute/pzunmqr_param.c
@@ -45,7 +45,7 @@ void chameleon_pzunmqr_param( int genD, const libhqr_tree_t *qrtree,
int ldam, ldan, ldbm, ldbp, lddn, lddm;
int tempnn, tempkmin, tempmm, tempkn;
int ib, K, L;
- int *tiles;
+ int node, nbtiles, *tiles;
chamctxt = chameleon_context_self();
if (sequence->status != CHAMELEON_SUCCESS)
@@ -64,7 +64,7 @@ void chameleon_pzunmqr_param( int genD, const libhqr_tree_t *qrtree,
/*
* zunmqr = A->nb * ib
* ztsmqr = A->nb * ib
- * zttmqr = A->nb * ib
+ * ztpmqrt = A->nb * ib
*/
ws_worker = A->nb * ib;
@@ -77,14 +77,14 @@ void chameleon_pzunmqr_param( int genD, const libhqr_tree_t *qrtree,
ws_worker = chameleon_max( ws_worker, ib * A->nb * 2 );
#endif
- /* Initialisation of tiles */
- tiles = (int*)calloc( qrtree->mt, sizeof(int) );
-
ws_worker *= sizeof(CHAMELEON_Complex64_t);
ws_host *= sizeof(CHAMELEON_Complex64_t);
RUNTIME_options_ws_alloc( &options, ws_worker, ws_host );
+ /* Initialisation of tiles */
+ tiles = (int*)calloc( qrtree->mt, sizeof(int) );
+
if (side == ChamLeft ) {
if (trans == ChamConjTrans) {
/*
@@ -134,9 +134,9 @@ void chameleon_pzunmqr_param( int genD, const libhqr_tree_t *qrtree,
}
/* Setting the order of the tiles*/
- libhqr_walk_stepk(qrtree, k, tiles + (k+1));
+ nbtiles = libhqr_walk_stepk( qrtree, k, tiles );
- for (i = k+1; i < A->mt; i++) {
+ for (i = 0; i < nbtiles; i++) {
m = tiles[i];
p = qrtree->currpiv(qrtree, k, m);
@@ -145,7 +145,7 @@ void chameleon_pzunmqr_param( int genD, const libhqr_tree_t *qrtree,
ldbm = BLKLDD(B, m);
ldbp = BLKLDD(B, p);
- if(qrtree->gettype(qrtree, k, m) == 0){
+ if( qrtree->gettype(qrtree, k, m) == LIBHQR_KILLED_BY_TS ) {
/* TS kernel */
L = 0;
T = TS;
@@ -158,10 +158,9 @@ void chameleon_pzunmqr_param( int genD, const libhqr_tree_t *qrtree,
for (n = 0; n < B->nt; n++) {
tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb;
- RUNTIME_data_migrate( sequence, B(p, n),
- B->get_rankof( B, m, n ) );
- RUNTIME_data_migrate( sequence, B(m, n),
- B->get_rankof( B, m, n ) );
+ node = B->get_rankof( B, m, n );
+ RUNTIME_data_migrate( sequence, B(p, n), node );
+ RUNTIME_data_migrate( sequence, B(m, n), node );
INSERT_TASK_ztpmqrt(
&options,
@@ -195,9 +194,9 @@ void chameleon_pzunmqr_param( int genD, const libhqr_tree_t *qrtree,
tempkn = k == A->nt-1 ? A->n-k*A->nb : A->nb;
/* Setting the order of the tiles*/
- libhqr_walk_stepk(qrtree, k, tiles + (k+1));
+ nbtiles = libhqr_walk_stepk( qrtree, k, tiles );
- for (i = A->mt-1; i > k; i--) {
+ for (i = nbtiles-1; i >=0; i--) {
m = tiles[i];
p = qrtree->currpiv(qrtree, k, m);
@@ -206,7 +205,7 @@ void chameleon_pzunmqr_param( int genD, const libhqr_tree_t *qrtree,
ldbm = BLKLDD(B, m);
ldbp = BLKLDD(B, p);
- if(qrtree->gettype(qrtree, k, m) == 0){
+ if( qrtree->gettype(qrtree, k, m) == LIBHQR_KILLED_BY_TS ) {
/* TS kernel */
L = 0;
T = TS;
@@ -219,10 +218,9 @@ void chameleon_pzunmqr_param( int genD, const libhqr_tree_t *qrtree,
for (n = k; n < B->nt; n++) {
tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb;
- RUNTIME_data_migrate( sequence, B(p, n),
- B->get_rankof( B, m, n ) );
- RUNTIME_data_migrate( sequence, B(m, n),
- B->get_rankof( B, m, n ) );
+ node = B->get_rankof( B, m, n );
+ RUNTIME_data_migrate( sequence, B(p, n), node );
+ RUNTIME_data_migrate( sequence, B(m, n), node );
INSERT_TASK_ztpmqrt(
&options,
@@ -295,16 +293,16 @@ void chameleon_pzunmqr_param( int genD, const libhqr_tree_t *qrtree,
tempkn = k == A->nt-1 ? A->n - k*A->nb : A->nb;
/* Setting the order of the tiles*/
- libhqr_walk_stepk(qrtree, k, tiles + (k+1));
+ nbtiles = libhqr_walk_stepk( qrtree, k, tiles );
- for (i = A->nt-1; i > k; i--) {
+ for (i = nbtiles-1; i >= 0; i--) {
n = tiles[i];
p = qrtree->currpiv(qrtree, k, n);
tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb;
ldan = BLKLDD(A, n);
- if( qrtree->gettype(qrtree, k, n) == 0 ) {
+ if( qrtree->gettype(qrtree, k, n) == LIBHQR_KILLED_BY_TS ) {
/* TS kernel */
L = 0;
T = TS;
@@ -319,10 +317,9 @@ void chameleon_pzunmqr_param( int genD, const libhqr_tree_t *qrtree,
tempmm = m == B->mt-1 ? B->m-m*B->mb : B->mb;
ldbm = BLKLDD(B, m);
- RUNTIME_data_migrate( sequence, B(m, p),
- B->get_rankof( B, m, n ) );
- RUNTIME_data_migrate( sequence, B(m, n),
- B->get_rankof( B, m, n ) );
+ node = B->get_rankof( B, m, n );
+ RUNTIME_data_migrate( sequence, B(m, p), node );
+ RUNTIME_data_migrate( sequence, B(m, n), node );
INSERT_TASK_ztpmqrt(
&options,
@@ -429,16 +426,16 @@ void chameleon_pzunmqr_param( int genD, const libhqr_tree_t *qrtree,
}
/* Setting the order of tiles */
- libhqr_walk_stepk(qrtree, k, tiles + (k+1));
+ nbtiles = libhqr_walk_stepk( qrtree, k, tiles );
- for (i = k+1; i < A->nt; i++) {
+ for (i = 0; i < nbtiles; i++) {
n = tiles[i];
p = qrtree->currpiv(qrtree, k, n);
tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb;
ldan = BLKLDD(A, n);
- if( qrtree->gettype(qrtree, k, n) == 0 ) {
+ if( qrtree->gettype(qrtree, k, n) == LIBHQR_KILLED_BY_TS ) {
/* TS kernel */
L = 0;
T = TS;
@@ -453,10 +450,9 @@ void chameleon_pzunmqr_param( int genD, const libhqr_tree_t *qrtree,
tempmm = m == B->mt-1 ? B->m-m*B->mb : B->mb;
ldbm = BLKLDD(B, m);
- RUNTIME_data_migrate( sequence, B(m, p),
- B->get_rankof( B, m, n ) );
- RUNTIME_data_migrate( sequence, B(m, n),
- B->get_rankof( B, m, n ) );
+ node = B->get_rankof( B, m, n );
+ RUNTIME_data_migrate( sequence, B(m, p), node );
+ RUNTIME_data_migrate( sequence, B(m, n), node );
INSERT_TASK_ztpmqrt(
&options,
diff --git a/compute/pzunmqrrh.c b/compute/pzunmqrrh.c
index ab51157ba..efb0f13d5 100644
--- a/compute/pzunmqrrh.c
+++ b/compute/pzunmqrrh.c
@@ -50,7 +50,7 @@ void chameleon_pzunmqrrh( int genD, int BS, cham_side_t side, cham_trans_t trans
int ldaM, ldam, ldan, ldaMRD, lddM;
int ldbM, ldbm, ldbMRD;
int tempMm, tempkn, tempnn, tempmm, tempMRDm, tempkmin;
- int ib;
+ int ib, node;
chamctxt = chameleon_context_self();
if (sequence->status != CHAMELEON_SUCCESS)
@@ -67,7 +67,7 @@ void chameleon_pzunmqrrh( int genD, int BS, cham_side_t side, cham_trans_t trans
/*
* zunmqr = A->nb * ib
* ztsmqr = A->nb * ib
- * zttmqr = A->nb * ib
+ * ztpmqrt = A->nb * ib
*/
ws_worker = A->nb * ib;
@@ -135,10 +135,9 @@ void chameleon_pzunmqrrh( int genD, int BS, cham_side_t side, cham_trans_t trans
for (n = 0; n < B->nt; n++) {
tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb;
- RUNTIME_data_migrate( sequence, B(M, n),
- B->get_rankof( B, m, n ) );
- RUNTIME_data_migrate( sequence, B(m, n),
- B->get_rankof( B, m, n ) );
+ node = B->get_rankof( B, m, n );
+ RUNTIME_data_migrate( sequence, B(M, n), node );
+ RUNTIME_data_migrate( sequence, B(m, n), node );
/* TS kernel */
INSERT_TASK_ztpmqrt(
@@ -162,10 +161,9 @@ void chameleon_pzunmqrrh( int genD, int BS, cham_side_t side, cham_trans_t trans
for (n = 0; n < B->nt; n++) {
tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb;
- RUNTIME_data_migrate( sequence, B(M, n),
- B->get_rankof( B, M+RD, n ) );
- RUNTIME_data_migrate( sequence, B(M+RD, n),
- B->get_rankof( B, M+RD, n ) );
+ node = B->get_rankof( B, M+RD, n );
+ RUNTIME_data_migrate( sequence, B(M, n), node );
+ RUNTIME_data_migrate( sequence, B(M+RD, n), node );
/* TT kernel */
INSERT_TASK_ztpmqrt(
@@ -209,10 +207,9 @@ void chameleon_pzunmqrrh( int genD, int BS, cham_side_t side, cham_trans_t trans
for (n = 0; n < B->nt; n++) {
tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb;
- RUNTIME_data_migrate( sequence, B(M, n),
- B->get_rankof( B, M+RD, n ) );
- RUNTIME_data_migrate( sequence, B(M+RD, n),
- B->get_rankof( B, M+RD, n ) );
+ node = B->get_rankof( B, M+RD, n );
+ RUNTIME_data_migrate( sequence, B(M, n), node );
+ RUNTIME_data_migrate( sequence, B(M+RD, n), node );
/* TT kernel */
INSERT_TASK_ztpmqrt(
@@ -240,10 +237,9 @@ void chameleon_pzunmqrrh( int genD, int BS, cham_side_t side, cham_trans_t trans
for (n = 0; n < B->nt; n++) {
tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb;
- RUNTIME_data_migrate( sequence, B(M, n),
- B->get_rankof( B, m, n ) );
- RUNTIME_data_migrate( sequence, B(m, n),
- B->get_rankof( B, m, n ) );
+ node = B->get_rankof( B, m, n );
+ RUNTIME_data_migrate( sequence, B(M, n), node );
+ RUNTIME_data_migrate( sequence, B(m, n), node );
/* TS kernel */
INSERT_TASK_ztpmqrt(
@@ -311,10 +307,9 @@ void chameleon_pzunmqrrh( int genD, int BS, cham_side_t side, cham_trans_t trans
ldbm = BLKLDD(B, m);
tempmm = m == B->mt-1 ? B->m-m*B->mb : B->mb;
- RUNTIME_data_migrate( sequence, B(m, M),
- B->get_rankof( B, m, M+RD ) );
- RUNTIME_data_migrate( sequence, B(m, M+RD),
- B->get_rankof( B, m, M+RD ) );
+ node = B->get_rankof( B, m, M+RD );
+ RUNTIME_data_migrate( sequence, B(m, M), node );
+ RUNTIME_data_migrate( sequence, B(m, M+RD), node );
/* TT kernel */
INSERT_TASK_ztpmqrt(
@@ -341,10 +336,9 @@ void chameleon_pzunmqrrh( int genD, int BS, cham_side_t side, cham_trans_t trans
ldbm = BLKLDD(B, m);
tempmm = m == B->mt-1 ? B->m-m*B->mb : B->mb;
- RUNTIME_data_migrate( sequence, B(m, M),
- B->get_rankof( B, m, n ) );
- RUNTIME_data_migrate( sequence, B(m, m),
- B->get_rankof( B, m, n ) );
+ node = B->get_rankof( B, m, n );
+ RUNTIME_data_migrate( sequence, B(m, M), node );
+ RUNTIME_data_migrate( sequence, B(m, m), node );
/* TS kernel */
INSERT_TASK_ztpmqrt(
@@ -441,10 +435,9 @@ void chameleon_pzunmqrrh( int genD, int BS, cham_side_t side, cham_trans_t trans
tempmm = m == B->mt-1 ? B->m-m*B->mb : B->mb;
ldbm = BLKLDD(B, m);
- RUNTIME_data_migrate( sequence, B(m, M),
- B->get_rankof( B, m, n ) );
- RUNTIME_data_migrate( sequence, B(m, n),
- B->get_rankof( B, m, n ) );
+ node = B->get_rankof( B, m, n );
+ RUNTIME_data_migrate( sequence, B(m, M), node );
+ RUNTIME_data_migrate( sequence, B(m, n), node );
/* TS kernel */
INSERT_TASK_ztpmqrt(
@@ -467,10 +460,9 @@ void chameleon_pzunmqrrh( int genD, int BS, cham_side_t side, cham_trans_t trans
tempmm = m == B->mt-1 ? B->m-m*B->mb : B->mb;
ldbm = BLKLDD(B, m);
- RUNTIME_data_migrate( sequence, B(m, M),
- B->get_rankof( B, m, M+RD ) );
- RUNTIME_data_migrate( sequence, B(m, M+RD),
- B->get_rankof( B, m, M+RD ) );
+ node = B->get_rankof( B, m, M+RD );
+ RUNTIME_data_migrate( sequence, B(m, M), node );
+ RUNTIME_data_migrate( sequence, B(m, M+RD), node );
/* TT kernel */
INSERT_TASK_ztpmqrt(
diff --git a/compute/zgels_param.c b/compute/zgels_param.c
index b6beda808..7b41e4cdf 100644
--- a/compute/zgels_param.c
+++ b/compute/zgels_param.c
@@ -396,7 +396,8 @@ int CHAMELEON_zgels_param_Tile_Async( const libhqr_tree_t *qrtree, cham_trans_t
subB = chameleon_desc_submatrix(B, 0, 0, A->n, B->n);
subA = chameleon_desc_submatrix(A, 0, 0, A->n, A->n);
- chameleon_pzgeqrf_param( 1, qrtree, A, TS, TT, Dptr, sequence, request );
+ chameleon_pzgeqrf_param( 1, A->nt, qrtree, A,
+ TS, TT, Dptr, sequence, request );
chameleon_pzunmqr_param( 0, qrtree, ChamLeft, ChamConjTrans, A, B, TS, TT, Dptr, sequence, request );
chameleon_pztrsm( ChamLeft, ChamUpper, ChamNoTrans, ChamNonUnit, 1.0, subA, subB, sequence, request );
}
diff --git a/compute/zgeqrf_param.c b/compute/zgeqrf_param.c
index 20a24cfe9..34171ce22 100644
--- a/compute/zgeqrf_param.c
+++ b/compute/zgeqrf_param.c
@@ -122,14 +122,14 @@ int CHAMELEON_zgeqrf_param( const libhqr_tree_t *qrtree, int M, int N,
/* Submit the matrix conversion */
chameleon_zlap2tile( chamctxt, &descAl, &descAt, ChamDescInout, ChamUpperLower,
- A, NB, NB, LDA, N, M, N, sequence, &request );
+ A, NB, NB, LDA, N, M, N, sequence, &request );
/* Call the tile interface */
CHAMELEON_zgeqrf_param_Tile_Async( qrtree, &descAt, descTS, descTT, sequence, &request );
/* Submit the matrix conversion back */
chameleon_ztile2lap( chamctxt, &descAl, &descAt,
- ChamDescInout, ChamUpperLower, sequence, &request );
+ ChamDescInout, ChamUpperLower, sequence, &request );
CHAMELEON_Desc_Flush( descTS, sequence );
CHAMELEON_Desc_Flush( descTT, sequence );
@@ -248,6 +248,7 @@ int CHAMELEON_zgeqrf_param_Tile_Async( const libhqr_tree_t *qrtree, CHAM_desc_t
{
CHAM_context_t *chamctxt;
CHAM_desc_t D, *Dptr = NULL;
+ int KT;
chamctxt = chameleon_context_self();
if (chamctxt == NULL) {
@@ -293,6 +294,14 @@ int CHAMELEON_zgeqrf_param_Tile_Async( const libhqr_tree_t *qrtree, CHAM_desc_t
if (chameleon_min(M, N) == 0)
return CHAMELEON_SUCCESS;
*/
+
+ if ( A->m < A->n ) {
+ KT = A->mt;
+ }
+ else {
+ KT = A->nt;
+ }
+
#if defined(CHAMELEON_COPY_DIAG)
{
int n = chameleon_min(A->m, A->n);
@@ -301,7 +310,8 @@ int CHAMELEON_zgeqrf_param_Tile_Async( const libhqr_tree_t *qrtree, CHAM_desc_t
}
#endif
- chameleon_pzgeqrf_param( 1, qrtree, A, TS, TT, Dptr, sequence, request );
+ chameleon_pzgeqrf_param( 1, KT, qrtree, A,
+ TS, TT, Dptr, sequence, request );
if (Dptr != NULL) {
CHAMELEON_Desc_Flush( A, sequence );
diff --git a/compute/ztpgqrt.c b/compute/ztpgqrt.c
index f28a20191..0512caeda 100644
--- a/compute/ztpgqrt.c
+++ b/compute/ztpgqrt.c
@@ -340,6 +340,7 @@ int CHAMELEON_ztpgqrt_Tile_Async( int L,
{
CHAM_context_t *chamctxt;
CHAM_desc_t D, *Dptr = NULL;
+ int KT;
chamctxt = chameleon_context_self();
if (chamctxt == NULL) {
@@ -396,15 +397,16 @@ int CHAMELEON_ztpgqrt_Tile_Async( int L,
chameleon_error("CHAMELEON_ztpgqrt_Tile", "Triangular part must be aligned with tiles");
return chameleon_request_fail(sequence, request, CHAMELEON_ERR_ILLEGAL_VALUE);
}
+
+ if (V1->m > V1->n) {
+ KT = V1->nt;
+ } else {
+ KT = V1->mt;
+ }
+
#if defined(CHAMELEON_COPY_DIAG)
{
- int minMT;
- if (V1->m > V1->n) {
- minMT = V1->nt;
- } else {
- minMT = V1->mt;
- }
- chameleon_zdesc_alloc_diag(D, V1->mb, V1->nb, minMT*V1->mb, V1->nb, 0, 0, minMT*V1->mb, V1->nb, V1->p, V1->q);
+ chameleon_zdesc_alloc(D, V1->mb, V1->nb, V1->m, KT*V1->nb, 0, 0, V1->m, KT*V1->nb, );
Dptr = &D;
}
#endif
@@ -412,7 +414,8 @@ int CHAMELEON_ztpgqrt_Tile_Async( int L,
/* if (chamctxt->householder == ChamFlatHouseholder) { */
chameleon_pzlaset( ChamUpperLower, 0., 1., Q1, sequence, request );
chameleon_pzlaset( ChamUpperLower, 0., 0., Q2, sequence, request );
- chameleon_pztpgqrt( 1, L, V1, T1, V2, T2, Q1, Q2, Dptr, sequence, request );
+ chameleon_pztpgqrt( KT, L, V2, T2, Q1, Q2, sequence, request );
+ chameleon_pzungqr( 1, V1, Q1, T1, Dptr, sequence, request );
if (Dptr != NULL) {
CHAMELEON_Desc_Flush( V1, sequence );
diff --git a/compute/zungqr_param.c b/compute/zungqr_param.c
index faf5f6db9..0f17560c3 100644
--- a/compute/zungqr_param.c
+++ b/compute/zungqr_param.c
@@ -75,11 +75,11 @@
*
*/
int CHAMELEON_zungqr_param( const libhqr_tree_t *qrtree,
- int M, int N, int K,
- CHAMELEON_Complex64_t *A, int LDA,
- CHAM_desc_t *descTS,
- CHAM_desc_t *descTT,
- CHAMELEON_Complex64_t *Q, int LDQ )
+ int M, int N, int K,
+ CHAMELEON_Complex64_t *A, int LDA,
+ CHAM_desc_t *descTS,
+ CHAM_desc_t *descTT,
+ CHAMELEON_Complex64_t *Q, int LDQ )
{
int NB;
int status;
@@ -253,6 +253,8 @@ int CHAMELEON_zungqr_param_Tile_Async( const libhqr_tree_t *qrtree, CHAM_desc_t
{
CHAM_context_t *chamctxt;
CHAM_desc_t D, *Dptr = NULL;
+ int KT;
+
chamctxt = chameleon_context_self();
if (chamctxt == NULL) {
chameleon_fatal_error("CHAMELEON_zungqr_param_Tile", "CHAMELEON not initialized");
@@ -301,6 +303,13 @@ int CHAMELEON_zungqr_param_Tile_Async( const libhqr_tree_t *qrtree, CHAM_desc_t
if (N <= 0)
return CHAMELEON_SUCCESS;
*/
+ if ( A->m < A->n ) {
+ KT = A->mt;
+ }
+ else {
+ KT = A->nt;
+ }
+
#if defined(CHAMELEON_COPY_DIAG)
{
int n = chameleon_min(A->m, A->n);
@@ -310,7 +319,7 @@ int CHAMELEON_zungqr_param_Tile_Async( const libhqr_tree_t *qrtree, CHAM_desc_t
#endif
chameleon_pzlaset( ChamUpperLower, 0., 1., Q, sequence, request );
- chameleon_pzungqr_param( 1, qrtree, A, Q, TS, TT, Dptr, sequence, request );
+ chameleon_pzungqr_param( 1, KT, qrtree, A, Q, TS, TT, Dptr, sequence, request );
if (Dptr != NULL) {
CHAMELEON_Desc_Flush( A, sequence );
diff --git a/coreblas/compute/core_ztpmlqt.c b/coreblas/compute/core_ztpmlqt.c
index 9eca91bd3..72e54b1ad 100644
--- a/coreblas/compute/core_ztpmlqt.c
+++ b/coreblas/compute/core_ztpmlqt.c
@@ -158,7 +158,7 @@ int CORE_ztpmlqt( cham_side_t side, cham_trans_t trans,
else {
m1 = M;
n1 = K;
- ldwork = chameleon_max( n1, N );
+ ldwork = chameleon_max( K, chameleon_max( M, N ) );
}
/* TS case */
diff --git a/coreblas/compute/core_ztpmqrt.c b/coreblas/compute/core_ztpmqrt.c
index 3b5721fc7..5909f19ee 100644
--- a/coreblas/compute/core_ztpmqrt.c
+++ b/coreblas/compute/core_ztpmqrt.c
@@ -159,7 +159,7 @@ int CORE_ztpmqrt( cham_side_t side, cham_trans_t trans,
else {
m1 = M;
n1 = K;
- ldwork = m1;
+ ldwork = chameleon_max( K, chameleon_max( M, N ) );
}
/* TS case */
--
GitLab