From fc2d01f133d92fbcaddbab53ba0960212fea6377 Mon Sep 17 00:00:00 2001
From: Raphael Boucherie <raphael.boucherie@inria.fr>
Date: Fri, 25 Aug 2017 12:32:10 +0200
Subject: [PATCH] add flush for unmqr and unmlq
---
compute/pzunmlq.c | 25 +++++++++++++++++++++++++
compute/pzunmlq_param.c | 21 +++++++++++++++++++++
compute/pzunmqr.c | 25 +++++++++++++++++++++++++
compute/pzunmqr_param.c | 21 ++++++++++++++++++++-
4 files changed, 91 insertions(+), 1 deletion(-)
diff --git a/compute/pzunmlq.c b/compute/pzunmlq.c
index df686642b..84475c878 100644
--- a/compute/pzunmlq.c
+++ b/compute/pzunmlq.c
@@ -127,6 +127,10 @@ void morse_pzunmlq(MORSE_enum side, MORSE_enum trans,
T(k, k), T->mb,
B(k, n), ldbk);
}
+
+ MORSE_TASK_dataflush( &options, D(k) );
+ MORSE_TASK_dataflush( &options, T(k, k) );
+
for (m = k+1; m < B->mt; m++) {
tempmm = m == B->mt-1 ? B->m-m*B->mb : B->mb;
ldbm = BLKLDD(B, m);
@@ -146,6 +150,9 @@ void morse_pzunmlq(MORSE_enum side, MORSE_enum trans,
B(k, n), ldbk,
B(m, n), ldbm);
}
+
+ MORSE_TASK_dataflush( &options, A(k, m) );
+ MORSE_TASK_dataflush( &options, T(k, m) );
}
/* Restore the original location of the tiles */
@@ -187,6 +194,9 @@ void morse_pzunmlq(MORSE_enum side, MORSE_enum trans,
B(k, n), ldbk,
B(m, n), ldbm);
}
+
+ MORSE_TASK_dataflush( &options, A(k, m) );
+ MORSE_TASK_dataflush( &options, T(k, m) );
}
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
@@ -216,6 +226,8 @@ void morse_pzunmlq(MORSE_enum side, MORSE_enum trans,
T(k, k), T->mb,
B(k, n), ldbk);
}
+ MORSE_TASK_dataflush( &options, D(k) );
+ MORSE_TASK_dataflush( &options, T(k, k) );
RUNTIME_iteration_pop(morse);
}
}
@@ -250,6 +262,9 @@ void morse_pzunmlq(MORSE_enum side, MORSE_enum trans,
B(m, k), ldbm,
B(m, n), ldbm);
}
+
+ MORSE_TASK_dataflush( &options, A(k, n) );
+ MORSE_TASK_dataflush( &options, T(k, n) );
}
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
@@ -281,6 +296,9 @@ void morse_pzunmlq(MORSE_enum side, MORSE_enum trans,
B(m, k), ldbm);
}
+ MORSE_TASK_dataflush( &options, D(k) );
+ MORSE_TASK_dataflush( &options, T(k, k) );
+
RUNTIME_iteration_pop(morse);
}
}
@@ -319,6 +337,10 @@ void morse_pzunmlq(MORSE_enum side, MORSE_enum trans,
T(k, k), T->mb,
B(m, k), ldbm);
}
+
+ MORSE_TASK_dataflush( &options, D(k) );
+ MORSE_TASK_dataflush( &options, T(k, k) );
+
for (n = k+1; n < B->nt; n++) {
tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb;
for (m = 0; m < B->mt; m++) {
@@ -338,6 +360,9 @@ void morse_pzunmlq(MORSE_enum side, MORSE_enum trans,
B(m, k), ldbm,
B(m, n), ldbm);
}
+
+ MORSE_TASK_dataflush( &options, A(k, n) );
+ MORSE_TASK_dataflush( &options, T(k, n) );
}
/* Restore the original location of the tiles */
diff --git a/compute/pzunmlq_param.c b/compute/pzunmlq_param.c
index c37ea1b59..1bee53e27 100644
--- a/compute/pzunmlq_param.c
+++ b/compute/pzunmlq_param.c
@@ -129,6 +129,8 @@ void morse_pzunmlq_param(const libhqr_tree_t *qrtree,
}
}
+ MORSE_TASK_dataflush( &options, D(k, p) );
+ MORSE_TASK_dataflush( &options, T(k, p) );
/* Setting the order of the tiles*/
libhqr_walk_stepk(qrtree, k, tiles + (k+1));
@@ -167,6 +169,8 @@ void morse_pzunmlq_param(const libhqr_tree_t *qrtree,
B(p, n), ldbp,
B(m, n), ldbm);
}
+ MORSE_TASK_dataflush( &options, A(k, m) );
+ MORSE_TASK_dataflush( &options, T(k, m) );
}
/* Restore the original location of the tiles */
@@ -226,6 +230,8 @@ void morse_pzunmlq_param(const libhqr_tree_t *qrtree,
B(p, n), ldbp,
B(m, n), ldbm);
}
+ MORSE_TASK_dataflush( &options, A(k, m) );
+ MORSE_TASK_dataflush( &options, T(k, m) );
}
T = TS;
@@ -265,6 +271,10 @@ void morse_pzunmlq_param(const libhqr_tree_t *qrtree,
B(p, n), ldbp);
}
}
+
+ MORSE_TASK_dataflush( &options, D(k, p) );
+ MORSE_TASK_dataflush( &options, T(k, p) );
+
RUNTIME_iteration_pop(morse);
}
}
@@ -318,6 +328,8 @@ void morse_pzunmlq_param(const libhqr_tree_t *qrtree,
B(m, p), ldbm,
B(m, n), ldbm);
}
+ MORSE_TASK_dataflush( &options, A(k, n) );
+ MORSE_TASK_dataflush( &options, T(k, n) );
}
T = TS;
@@ -357,6 +369,10 @@ void morse_pzunmlq_param(const libhqr_tree_t *qrtree,
B(m, p), ldbm);
}
}
+
+ MORSE_TASK_dataflush( &options, D(k, p) );
+ MORSE_TASK_dataflush( &options, T(k, p) );
+
RUNTIME_iteration_pop(morse);
}
}
@@ -404,6 +420,9 @@ void morse_pzunmlq_param(const libhqr_tree_t *qrtree,
B(m, p), ldbm);
}
}
+
+ MORSE_TASK_dataflush( &options, D(k, p) );
+ MORSE_TASK_dataflush( &options, T(k, p) );
/* Setting the order of tiles */
libhqr_walk_stepk(qrtree, k, tiles + (k+1));
@@ -443,6 +462,8 @@ void morse_pzunmlq_param(const libhqr_tree_t *qrtree,
B(m, p), ldbm,
B(m, n), ldbm);
}
+ MORSE_TASK_dataflush( &options, A( k, n) );
+ MORSE_TASK_dataflush( &options, TT(k, n) );
}
RUNTIME_iteration_pop(morse);
diff --git a/compute/pzunmqr.c b/compute/pzunmqr.c
index 3a0f3c18d..72b7b62d4 100644
--- a/compute/pzunmqr.c
+++ b/compute/pzunmqr.c
@@ -127,6 +127,10 @@ void morse_pzunmqr(MORSE_enum side, MORSE_enum trans,
T(k, k), T->mb,
B(k, n), ldbk);
}
+
+ MORSE_TASK_dataflush( &options, D(k) );
+ MORSE_TASK_dataflush( &options, T(k, k) );
+
for (m = k+1; m < B->mt; m++) {
tempmm = m == B->mt-1 ? B->m-m*B->mb : B->mb;
ldam = BLKLDD(A, m);
@@ -147,6 +151,9 @@ void morse_pzunmqr(MORSE_enum side, MORSE_enum trans,
B(k, n), ldbk,
B(m, n), ldbm);
}
+
+ MORSE_TASK_dataflush( &options, A(m, k) );
+ MORSE_TASK_dataflush( &options, T(m, k) );
}
/* Restore the original location of the tiles */
@@ -189,6 +196,9 @@ void morse_pzunmqr(MORSE_enum side, MORSE_enum trans,
B(k, n), ldbk,
B(m, n), ldbm);
}
+
+ MORSE_TASK_dataflush( &options, A(m, k) );
+ MORSE_TASK_dataflush( &options, T(m, k) );
}
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
@@ -218,6 +228,8 @@ void morse_pzunmqr(MORSE_enum side, MORSE_enum trans,
T(k, k), T->mb,
B(k, n), ldbk);
}
+ MORSE_TASK_dataflush( &options, D(k) );
+ MORSE_TASK_dataflush( &options, T(k, k) );
RUNTIME_iteration_pop(morse);
}
}
@@ -254,6 +266,9 @@ void morse_pzunmqr(MORSE_enum side, MORSE_enum trans,
B(m, k), ldbm,
B(m, n), ldbm);
}
+
+ MORSE_TASK_dataflush( &options, A(n, k) );
+ MORSE_TASK_dataflush( &options, T(n, k) );
}
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
@@ -285,6 +300,9 @@ void morse_pzunmqr(MORSE_enum side, MORSE_enum trans,
B(m, k), ldbm);
}
+ MORSE_TASK_dataflush( &options, D(k) );
+ MORSE_TASK_dataflush( &options, T(k, k) );
+
RUNTIME_iteration_pop(morse);
}
}
@@ -323,6 +341,10 @@ void morse_pzunmqr(MORSE_enum side, MORSE_enum trans,
T(k, k), T->mb,
B(m, k), ldbm);
}
+
+ MORSE_TASK_dataflush( &options, D(k) );
+ MORSE_TASK_dataflush( &options, T(k, k) );
+
for (n = k+1; n < B->nt; n++) {
tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb;
ldan = BLKLDD(A, n);
@@ -343,6 +365,9 @@ void morse_pzunmqr(MORSE_enum side, MORSE_enum trans,
B(m, k), ldbm,
B(m, n), ldbm);
}
+
+ MORSE_TASK_dataflush( &options, A(n, k) );
+ MORSE_TASK_dataflush( &options, T(n, k) );
}
/* Restore the original location of the tiles */
diff --git a/compute/pzunmqr_param.c b/compute/pzunmqr_param.c
index 0ad3b1f87..4b415ff3d 100644
--- a/compute/pzunmqr_param.c
+++ b/compute/pzunmqr_param.c
@@ -127,6 +127,8 @@ void morse_pzunmqr_param(const libhqr_tree_t *qrtree,
T(m, k), T->mb,
B(m, n), ldbm);
}
+ MORSE_TASK_dataflush( &options, D(m, k) );
+ MORSE_TASK_dataflush( &options, T(m, k) );
}
/* Setting the order of the tiles*/
@@ -168,7 +170,9 @@ void morse_pzunmqr_param(const libhqr_tree_t *qrtree,
B(p, n), ldbp,
B(m, n), ldbm);
}
- }
+ MORSE_TASK_dataflush( &options, A(m, k) );
+ MORSE_TASK_dataflush( &options, T(m, k) );
+ }
/* Restore the original location of the tiles */
for (n = 0; n < B->nt; n++) {
@@ -227,6 +231,8 @@ void morse_pzunmqr_param(const libhqr_tree_t *qrtree,
B(p, n), ldbp,
B(m, n), ldbm);
}
+ MORSE_TASK_dataflush( &options, A(m, k) );
+ MORSE_TASK_dataflush( &options, T(m, k) );
}
T = TS;
@@ -267,6 +273,10 @@ void morse_pzunmqr_param(const libhqr_tree_t *qrtree,
B(m, n), ldbm);
}
}
+
+ MORSE_TASK_dataflush( &options, D(m, k) );
+ MORSE_TASK_dataflush( &options, T(m, k) );
+
RUNTIME_iteration_pop(morse);
}
}
@@ -320,6 +330,8 @@ void morse_pzunmqr_param(const libhqr_tree_t *qrtree,
B(m, p), ldbm,
B(m, n), ldbm);
}
+ MORSE_TASK_dataflush( &options, A(n, k) );
+ MORSE_TASK_dataflush( &options, T(n, k) );
}
T = TS;
@@ -359,6 +371,8 @@ void morse_pzunmqr_param(const libhqr_tree_t *qrtree,
T(n, k), T->mb,
B(m, n), ldbm);
}
+ MORSE_TASK_dataflush( &options, D(n, k) );
+ MORSE_TASK_dataflush( &options, T(n, k) );
}
RUNTIME_iteration_pop(morse);
}
@@ -405,7 +419,10 @@ void morse_pzunmqr_param(const libhqr_tree_t *qrtree,
T(n, k), T->mb,
B(m, n), ldbm);
}
+ MORSE_TASK_dataflush( &options, D(n, k) );
+ MORSE_TASK_dataflush( &options, T(n, k) );
}
+
/* Setting the order of tiles */
libhqr_walk_stepk(qrtree, k, tiles + (k+1));
@@ -446,6 +463,8 @@ void morse_pzunmqr_param(const libhqr_tree_t *qrtree,
B(m, p), ldbm,
B(m, n), ldbm);
}
+ MORSE_TASK_dataflush( &options, A(n, k) );
+ MORSE_TASK_dataflush( &options, T(n, k) );
}
RUNTIME_iteration_pop(morse);
--
GitLab