From 21bc5a7f9379f56efb49ffd38adedb5f93052649 Mon Sep 17 00:00:00 2001
From: Mathieu Faverge <mathieu.faverge@inria.fr>
Date: Wed, 31 Jan 2018 20:20:16 +0100
Subject: [PATCH] Update quark codelets
---
runtime/quark/codelets/codelet_zgelqt.c | 18 ++---
runtime/quark/codelets/codelet_zgeqrt.c | 18 ++---
runtime/quark/codelets/codelet_ztplqt.c | 76 ++++++++++++++++++++
runtime/quark/codelets/codelet_ztpmlqt.c | 88 ++++++++++++++++++++++++
runtime/quark/codelets/codelet_ztpmqrt.c | 6 +-
runtime/quark/codelets/codelet_ztpqrt.c | 4 +-
runtime/quark/codelets/codelet_ztslqt.c | 4 +-
runtime/quark/codelets/codelet_ztsmlq.c | 36 +++++-----
runtime/quark/codelets/codelet_ztsmqr.c | 36 +++++-----
runtime/quark/codelets/codelet_ztsqrt.c | 4 +-
runtime/quark/codelets/codelet_zttlqt.c | 22 +++---
runtime/quark/codelets/codelet_zttmlq.c | 36 +++++-----
runtime/quark/codelets/codelet_zttmqr.c | 39 ++++++-----
runtime/quark/codelets/codelet_zttqrt.c | 22 +++---
14 files changed, 289 insertions(+), 120 deletions(-)
create mode 100644 runtime/quark/codelets/codelet_ztplqt.c
create mode 100644 runtime/quark/codelets/codelet_ztpmlqt.c
diff --git a/runtime/quark/codelets/codelet_zgelqt.c b/runtime/quark/codelets/codelet_zgelqt.c
index 16919df2d..32983f643 100644
--- a/runtime/quark/codelets/codelet_zgelqt.c
+++ b/runtime/quark/codelets/codelet_zgelqt.c
@@ -118,14 +118,14 @@ void MORSE_TASK_zgelqt(const MORSE_option_t *options,
quark_option_t *opt = (quark_option_t*)(options->schedopt);
DAG_CORE_GELQT;
QUARK_Insert_Task(opt->quark, CORE_zgelqt_quark, (Quark_Task_Flags*)opt,
- sizeof(int), &m, VALUE,
- sizeof(int), &n, VALUE,
- sizeof(int), &ib, VALUE,
- sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A, MORSE_Complex64_t, Am, An), INOUT,
- sizeof(int), &lda, VALUE,
- sizeof(MORSE_Complex64_t)*ib*nb, RTBLKADDR(T, MORSE_Complex64_t, Tm, Tn), OUTPUT,
- sizeof(int), &ldt, VALUE,
- sizeof(MORSE_Complex64_t)*nb, NULL, SCRATCH,
- sizeof(MORSE_Complex64_t)*ib*nb, NULL, SCRATCH,
+ sizeof(int), &m, VALUE,
+ sizeof(int), &n, VALUE,
+ sizeof(int), &ib, VALUE,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A, MORSE_Complex64_t, Am, An), INOUT,
+ sizeof(int), &lda, VALUE,
+ sizeof(MORSE_Complex64_t)*ib*nb, RTBLKADDR(T, MORSE_Complex64_t, Tm, Tn), OUTPUT,
+ sizeof(int), &ldt, VALUE,
+ sizeof(MORSE_Complex64_t)*nb, NULL, SCRATCH,
+ sizeof(MORSE_Complex64_t)*ib*nb, NULL, SCRATCH,
0);
}
diff --git a/runtime/quark/codelets/codelet_zgeqrt.c b/runtime/quark/codelets/codelet_zgeqrt.c
index c2575a706..f18033e27 100644
--- a/runtime/quark/codelets/codelet_zgeqrt.c
+++ b/runtime/quark/codelets/codelet_zgeqrt.c
@@ -118,14 +118,14 @@ void MORSE_TASK_zgeqrt(const MORSE_option_t *options,
quark_option_t *opt = (quark_option_t*)(options->schedopt);
DAG_CORE_GEQRT;
QUARK_Insert_Task(opt->quark, CORE_zgeqrt_quark, (Quark_Task_Flags*)opt,
- sizeof(int), &m, VALUE,
- sizeof(int), &n, VALUE,
- sizeof(int), &ib, VALUE,
- sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A, MORSE_Complex64_t, Am, An), INOUT,
- sizeof(int), &lda, VALUE,
- sizeof(MORSE_Complex64_t)*ib*nb, RTBLKADDR(T, MORSE_Complex64_t, Tm, Tn), OUTPUT,
- sizeof(int), &ldt, VALUE,
- sizeof(MORSE_Complex64_t)*nb, NULL, SCRATCH,
- sizeof(MORSE_Complex64_t)*ib*nb, NULL, SCRATCH,
+ sizeof(int), &m, VALUE,
+ sizeof(int), &n, VALUE,
+ sizeof(int), &ib, VALUE,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A, MORSE_Complex64_t, Am, An), INOUT,
+ sizeof(int), &lda, VALUE,
+ sizeof(MORSE_Complex64_t)*ib*nb, RTBLKADDR(T, MORSE_Complex64_t, Tm, Tn), OUTPUT,
+ sizeof(int), &ldt, VALUE,
+ sizeof(MORSE_Complex64_t)*nb, NULL, SCRATCH,
+ sizeof(MORSE_Complex64_t)*ib*nb, NULL, SCRATCH,
0);
}
diff --git a/runtime/quark/codelets/codelet_ztplqt.c b/runtime/quark/codelets/codelet_ztplqt.c
new file mode 100644
index 000000000..b1723f8fd
--- /dev/null
+++ b/runtime/quark/codelets/codelet_ztplqt.c
@@ -0,0 +1,76 @@
+/**
+ *
+ * @copyright (c) 2009-2016 The University of Tennessee and The University
+ * of Tennessee Research Foundation.
+ * All rights reserved.
+ * @copyright (c) 2012-2016 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria,
+ * Univ. Bordeaux. All rights reserved.
+ *
+ **/
+
+/**
+ *
+ * @file codelet_ztplqt.c
+ *
+ * MORSE codelets kernel
+ * MORSE is a software package provided by Univ. of Tennessee,
+ * Univ. of California Berkeley and Univ. of Colorado Denver
+ *
+ * @version 0.9.0
+ * @author Mathieu Faverge
+ * @date 2016-12-15
+ * @precisions normal z -> s d c
+ *
+ **/
+#include "chameleon_quark.h"
+#include "chameleon/morse_tasks_z.h"
+#include "coreblas/coreblas_z.h"
+
+static void
+CORE_ztplqt_quark( Quark *quark )
+{
+ int M;
+ int N;
+ int L;
+ int ib;
+ MORSE_Complex64_t *A;
+ int lda;
+ MORSE_Complex64_t *B;
+ int ldb;
+ MORSE_Complex64_t *T;
+ int ldt;
+ MORSE_Complex64_t *WORK;
+
+ quark_unpack_args_11( quark, M, N, L, ib,
+ A, lda, B, ldb, T, ldt, WORK );
+
+ CORE_ztplqt( M, N, L, ib,
+ A, lda, B, ldb, T, ldt, WORK );
+}
+
+void MORSE_TASK_ztplqt( const MORSE_option_t *options,
+ int M, int N, int L, int ib, int nb,
+ const MORSE_desc_t *A, int Am, int An, int lda,
+ const MORSE_desc_t *B, int Bm, int Bn, int ldb,
+ const MORSE_desc_t *T, int Tm, int Tn, int ldt )
+{
+ quark_option_t *opt = (quark_option_t*)(options->schedopt);
+ DAG_CORE_TSLQT;
+
+ int shapeB = ( L == 0 ) ? 0 : (QUARK_REGION_L | QUARK_REGION_D);
+
+ QUARK_Insert_Task(
+ opt->quark, CORE_ztplqt_quark, (Quark_Task_Flags*)opt,
+ sizeof(int), &M, VALUE,
+ sizeof(int), &N, VALUE,
+ sizeof(int), &L, VALUE,
+ sizeof(int), &ib, VALUE,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR( A, MORSE_Complex64_t, Am, An ), INOUT | QUARK_REGION_L | QUARK_REGION_D,
+ sizeof(int), &lda, VALUE,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR( B, MORSE_Complex64_t, Bm, Bn ), INOUT | shapeB | LOCALITY,
+ sizeof(int), &ldb, VALUE,
+ sizeof(MORSE_Complex64_t)*nb*ib, RTBLKADDR( T, MORSE_Complex64_t, Tm, Tn ), OUTPUT,
+ sizeof(int), &ldt, VALUE,
+ sizeof(MORSE_Complex64_t)*(ib+1)*nb, NULL, SCRATCH,
+ 0);
+}
diff --git a/runtime/quark/codelets/codelet_ztpmlqt.c b/runtime/quark/codelets/codelet_ztpmlqt.c
new file mode 100644
index 000000000..0d6f2e6f6
--- /dev/null
+++ b/runtime/quark/codelets/codelet_ztpmlqt.c
@@ -0,0 +1,88 @@
+/**
+ *
+ * @copyright (c) 2009-2016 The University of Tennessee and The University
+ * of Tennessee Research Foundation.
+ * All rights reserved.
+ * @copyright (c) 2012-2016 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria,
+ * Univ. Bordeaux. All rights reserved.
+ *
+ **/
+
+/**
+ *
+ * @file codelet_ztpmlqt.c
+ *
+ * MORSE codelets kernel
+ * MORSE is a software package provided by Univ. of Tennessee,
+ * Univ. of California Berkeley and Univ. of Colorado Denver
+ *
+ * @version 0.9.0
+ * @author Mathieu Faverge
+ * @date 2016-12-15
+ * @precisions normal z -> s d c
+ *
+ **/
+#include "chameleon_quark.h"
+#include "chameleon/morse_tasks_z.h"
+#include "coreblas/coreblas_z.h"
+
+static void
+CORE_ztpmlqt_quark( Quark *quark )
+{
+ MORSE_enum side;
+ MORSE_enum trans;
+ int M;
+ int N;
+ int K;
+ int L;
+ int ib;
+ const MORSE_Complex64_t *V;
+ int ldv;
+ const MORSE_Complex64_t *T;
+ int ldt;
+ MORSE_Complex64_t *A;
+ int lda;
+ MORSE_Complex64_t *B;
+ int ldb;
+ MORSE_Complex64_t *WORK;
+
+ quark_unpack_args_16( quark, side, trans, M, N, K, L, ib,
+ V, ldv, T, ldt, A, lda, B, ldb, WORK );
+
+ CORE_ztpmlqt( side, trans, M, N, K, L, ib,
+ V, ldv, T, ldt, A, lda, B, ldb, WORK );
+}
+
+void MORSE_TASK_ztpmlqt( const MORSE_option_t *options,
+ MORSE_enum side, MORSE_enum trans,
+ int M, int N, int K, int L, int ib, int nb,
+ const MORSE_desc_t *V, int Vm, int Vn, int ldv,
+ const MORSE_desc_t *T, int Tm, int Tn, int ldt,
+ const MORSE_desc_t *A, int Am, int An, int lda,
+ const MORSE_desc_t *B, int Bm, int Bn, int ldb )
+{
+ quark_option_t *opt = (quark_option_t*)(options->schedopt);
+ DAG_CORE_TSMQR;
+
+ int shapeV = ( L == 0 ) ? 0 : (QUARK_REGION_L | QUARK_REGION_D);
+
+ QUARK_Insert_Task(
+ opt->quark, CORE_ztpmlqt_quark, (Quark_Task_Flags*)opt,
+ sizeof(MORSE_enum), &side, VALUE,
+ sizeof(MORSE_enum), &trans, VALUE,
+ sizeof(int), &M, VALUE,
+ sizeof(int), &N, VALUE,
+ sizeof(int), &K, VALUE,
+ sizeof(int), &L, VALUE,
+ sizeof(int), &ib, VALUE,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR( V, MORSE_Complex64_t, Vm, Vn ), INPUT | shapeV,
+ sizeof(int), &ldv, VALUE,
+ sizeof(MORSE_Complex64_t)*ib*nb, RTBLKADDR( T, MORSE_Complex64_t, Tm, Tn ), INPUT,
+ sizeof(int), &ldt, VALUE,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR( A, MORSE_Complex64_t, Am, An ), INOUT,
+ sizeof(int), &lda, VALUE,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR( B, MORSE_Complex64_t, Bm, Bn ), INOUT | LOCALITY,
+ sizeof(int), &ldb, VALUE,
+ sizeof(MORSE_Complex64_t)*ib*nb, NULL, SCRATCH,
+ 0);
+}
diff --git a/runtime/quark/codelets/codelet_ztpmqrt.c b/runtime/quark/codelets/codelet_ztpmqrt.c
index d0efd49bd..16b1bbff6 100644
--- a/runtime/quark/codelets/codelet_ztpmqrt.c
+++ b/runtime/quark/codelets/codelet_ztpmqrt.c
@@ -64,6 +64,8 @@ void MORSE_TASK_ztpmqrt( const MORSE_option_t *options,
quark_option_t *opt = (quark_option_t*)(options->schedopt);
DAG_CORE_TSMQR;
+ int shapeV = ( L == 0 ) ? 0 : (QUARK_REGION_U | QUARK_REGION_D);
+
QUARK_Insert_Task(
opt->quark, CORE_ztpmqrt_quark, (Quark_Task_Flags*)opt,
sizeof(MORSE_enum), &side, VALUE,
@@ -73,13 +75,13 @@ void MORSE_TASK_ztpmqrt( const MORSE_option_t *options,
sizeof(int), &K, VALUE,
sizeof(int), &L, VALUE,
sizeof(int), &ib, VALUE,
- sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR( V, MORSE_Complex64_t, Vm, Vn ), INPUT,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR( V, MORSE_Complex64_t, Vm, Vn ), INPUT | shapeV,
sizeof(int), &ldv, VALUE,
sizeof(MORSE_Complex64_t)*ib*nb, RTBLKADDR( T, MORSE_Complex64_t, Tm, Tn ), INPUT,
sizeof(int), &ldt, VALUE,
sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR( A, MORSE_Complex64_t, Am, An ), INOUT,
sizeof(int), &lda, VALUE,
- sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR( B, MORSE_Complex64_t, Bm, Bn ), INOUT,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR( B, MORSE_Complex64_t, Bm, Bn ), INOUT | LOCALITY,
sizeof(int), &ldb, VALUE,
sizeof(MORSE_Complex64_t)*ib*nb, NULL, SCRATCH,
0);
diff --git a/runtime/quark/codelets/codelet_ztpqrt.c b/runtime/quark/codelets/codelet_ztpqrt.c
index 72c763a66..a7f1067fd 100644
--- a/runtime/quark/codelets/codelet_ztpqrt.c
+++ b/runtime/quark/codelets/codelet_ztpqrt.c
@@ -57,6 +57,8 @@ void MORSE_TASK_ztpqrt( const MORSE_option_t *options,
quark_option_t *opt = (quark_option_t*)(options->schedopt);
DAG_CORE_TSQRT;
+ int shapeB = ( L == 0 ) ? 0 : (QUARK_REGION_U | QUARK_REGION_D);
+
QUARK_Insert_Task(
opt->quark, CORE_ztpqrt_quark, (Quark_Task_Flags*)opt,
sizeof(int), &M, VALUE,
@@ -65,7 +67,7 @@ void MORSE_TASK_ztpqrt( const MORSE_option_t *options,
sizeof(int), &ib, VALUE,
sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR( A, MORSE_Complex64_t, Am, An ), INOUT | QUARK_REGION_U | QUARK_REGION_D,
sizeof(int), &lda, VALUE,
- sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR( B, MORSE_Complex64_t, Bm, Bn ), INOUT,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR( B, MORSE_Complex64_t, Bm, Bn ), INOUT | shapeB | LOCALITY,
sizeof(int), &ldb, VALUE,
sizeof(MORSE_Complex64_t)*nb*ib, RTBLKADDR( T, MORSE_Complex64_t, Tm, Tn ), OUTPUT,
sizeof(int), &ldt, VALUE,
diff --git a/runtime/quark/codelets/codelet_ztslqt.c b/runtime/quark/codelets/codelet_ztslqt.c
index 766d3ff87..ed30b43b2 100644
--- a/runtime/quark/codelets/codelet_ztslqt.c
+++ b/runtime/quark/codelets/codelet_ztslqt.c
@@ -136,9 +136,9 @@ void MORSE_TASK_ztslqt(const MORSE_option_t *options,
sizeof(int), &m, VALUE,
sizeof(int), &n, VALUE,
sizeof(int), &ib, VALUE,
- sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A1, MORSE_Complex64_t, A1m, A1n), INOUT | QUARK_REGION_D | QUARK_REGION_L | LOCALITY,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A1, MORSE_Complex64_t, A1m, A1n), INOUT | QUARK_REGION_L | QUARK_REGION_D,
sizeof(int), &lda1, VALUE,
- sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A2, MORSE_Complex64_t, A2m, A2n), INOUT,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A2, MORSE_Complex64_t, A2m, A2n), INOUT | LOCALITY,
sizeof(int), &lda2, VALUE,
sizeof(MORSE_Complex64_t)*ib*nb, RTBLKADDR(T, MORSE_Complex64_t, Tm, Tn), OUTPUT,
sizeof(int), &ldt, VALUE,
diff --git a/runtime/quark/codelets/codelet_ztsmlq.c b/runtime/quark/codelets/codelet_ztsmlq.c
index a59d5b4e0..9020a57f5 100644
--- a/runtime/quark/codelets/codelet_ztsmlq.c
+++ b/runtime/quark/codelets/codelet_ztsmlq.c
@@ -173,23 +173,23 @@ void MORSE_TASK_ztsmlq(const MORSE_option_t *options,
quark_option_t *opt = (quark_option_t*)(options->schedopt);
DAG_CORE_TSMLQ;
QUARK_Insert_Task(opt->quark, CORE_ztsmlq_quark, (Quark_Task_Flags*)opt,
- sizeof(MORSE_enum), &side, VALUE,
- sizeof(MORSE_enum), &trans, VALUE,
- sizeof(int), &m1, VALUE,
- sizeof(int), &n1, VALUE,
- sizeof(int), &m2, VALUE,
- sizeof(int), &n2, VALUE,
- sizeof(int), &k, VALUE,
- sizeof(int), &ib, VALUE,
- sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A1, MORSE_Complex64_t, A1m, A1n), INOUT,
- sizeof(int), &lda1, VALUE,
- sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A2, MORSE_Complex64_t, A2m, A2n), INOUT | LOCALITY,
- sizeof(int), &lda2, VALUE,
- sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(V, MORSE_Complex64_t, Vm, Vn), INPUT,
- sizeof(int), &ldv, VALUE,
- sizeof(MORSE_Complex64_t)*ib*nb, RTBLKADDR(T, MORSE_Complex64_t, Tm, Tn), INPUT,
- sizeof(int), &ldt, VALUE,
- sizeof(MORSE_Complex64_t)*ib*nb, NULL, SCRATCH,
- sizeof(int), &ldwork, VALUE,
+ sizeof(MORSE_enum), &side, VALUE,
+ sizeof(MORSE_enum), &trans, VALUE,
+ sizeof(int), &m1, VALUE,
+ sizeof(int), &n1, VALUE,
+ sizeof(int), &m2, VALUE,
+ sizeof(int), &n2, VALUE,
+ sizeof(int), &k, VALUE,
+ sizeof(int), &ib, VALUE,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A1, MORSE_Complex64_t, A1m, A1n), INOUT,
+ sizeof(int), &lda1, VALUE,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A2, MORSE_Complex64_t, A2m, A2n), INOUT | LOCALITY,
+ sizeof(int), &lda2, VALUE,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(V, MORSE_Complex64_t, Vm, Vn), INPUT,
+ sizeof(int), &ldv, VALUE,
+ sizeof(MORSE_Complex64_t)*ib*nb, RTBLKADDR(T, MORSE_Complex64_t, Tm, Tn), INPUT,
+ sizeof(int), &ldt, VALUE,
+ sizeof(MORSE_Complex64_t)*ib*nb, NULL, SCRATCH,
+ sizeof(int), &ldwork, VALUE,
0);
}
diff --git a/runtime/quark/codelets/codelet_ztsmqr.c b/runtime/quark/codelets/codelet_ztsmqr.c
index de57a8aa8..ed68791bf 100644
--- a/runtime/quark/codelets/codelet_ztsmqr.c
+++ b/runtime/quark/codelets/codelet_ztsmqr.c
@@ -173,23 +173,23 @@ void MORSE_TASK_ztsmqr(const MORSE_option_t *options,
quark_option_t *opt = (quark_option_t*)(options->schedopt);
DAG_CORE_TSMQR;
QUARK_Insert_Task(opt->quark, CORE_ztsmqr_quark, (Quark_Task_Flags*)opt,
- sizeof(MORSE_enum), &side, VALUE,
- sizeof(MORSE_enum), &trans, VALUE,
- sizeof(int), &m1, VALUE,
- sizeof(int), &n1, VALUE,
- sizeof(int), &m2, VALUE,
- sizeof(int), &n2, VALUE,
- sizeof(int), &k, VALUE,
- sizeof(int), &ib, VALUE,
- sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A1, MORSE_Complex64_t, A1m, A1n), INOUT,
- sizeof(int), &lda1, VALUE,
- sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A2, MORSE_Complex64_t, A2m, A2n), INOUT | LOCALITY,
- sizeof(int), &lda2, VALUE,
- sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(V, MORSE_Complex64_t, Vm, Vn), INPUT,
- sizeof(int), &ldv, VALUE,
- sizeof(MORSE_Complex64_t)*ib*nb, RTBLKADDR(T, MORSE_Complex64_t, Tm, Tn), INPUT,
- sizeof(int), &ldt, VALUE,
- sizeof(MORSE_Complex64_t)*ib*nb, NULL, SCRATCH,
- sizeof(int), &ldwork, VALUE,
+ sizeof(MORSE_enum), &side, VALUE,
+ sizeof(MORSE_enum), &trans, VALUE,
+ sizeof(int), &m1, VALUE,
+ sizeof(int), &n1, VALUE,
+ sizeof(int), &m2, VALUE,
+ sizeof(int), &n2, VALUE,
+ sizeof(int), &k, VALUE,
+ sizeof(int), &ib, VALUE,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A1, MORSE_Complex64_t, A1m, A1n), INOUT,
+ sizeof(int), &lda1, VALUE,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A2, MORSE_Complex64_t, A2m, A2n), INOUT | LOCALITY,
+ sizeof(int), &lda2, VALUE,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(V, MORSE_Complex64_t, Vm, Vn), INPUT,
+ sizeof(int), &ldv, VALUE,
+ sizeof(MORSE_Complex64_t)*ib*nb, RTBLKADDR(T, MORSE_Complex64_t, Tm, Tn), INPUT,
+ sizeof(int), &ldt, VALUE,
+ sizeof(MORSE_Complex64_t)*ib*nb, NULL, SCRATCH,
+ sizeof(int), &ldwork, VALUE,
0);
}
diff --git a/runtime/quark/codelets/codelet_ztsqrt.c b/runtime/quark/codelets/codelet_ztsqrt.c
index f3f09a3ea..eb960b0d5 100644
--- a/runtime/quark/codelets/codelet_ztsqrt.c
+++ b/runtime/quark/codelets/codelet_ztsqrt.c
@@ -124,9 +124,9 @@ void MORSE_TASK_ztsqrt(const MORSE_option_t *options,
sizeof(int), &m, VALUE,
sizeof(int), &n, VALUE,
sizeof(int), &ib, VALUE,
- sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A1, MORSE_Complex64_t, A1m, A1n), INOUT | QUARK_REGION_D | QUARK_REGION_U | LOCALITY,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A1, MORSE_Complex64_t, A1m, A1n), INOUT | QUARK_REGION_U | QUARK_REGION_D,
sizeof(int), &lda1, VALUE,
- sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A2, MORSE_Complex64_t, A2m, A2n), INOUT,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A2, MORSE_Complex64_t, A2m, A2n), INOUT | LOCALITY,
sizeof(int), &lda2, VALUE,
sizeof(MORSE_Complex64_t)*ib*nb, RTBLKADDR(T, MORSE_Complex64_t, Tm, Tn), OUTPUT,
sizeof(int), &ldt, VALUE,
diff --git a/runtime/quark/codelets/codelet_zttlqt.c b/runtime/quark/codelets/codelet_zttlqt.c
index 9ec251332..89020dc0a 100644
--- a/runtime/quark/codelets/codelet_zttlqt.c
+++ b/runtime/quark/codelets/codelet_zttlqt.c
@@ -133,16 +133,16 @@ void MORSE_TASK_zttlqt(const MORSE_option_t *options,
quark_option_t *opt = (quark_option_t*)(options->schedopt);
DAG_CORE_TTLQT;
QUARK_Insert_Task(opt->quark, CORE_zttlqt_quark, (Quark_Task_Flags*)opt,
- sizeof(int), &m, VALUE,
- sizeof(int), &n, VALUE,
- sizeof(int), &ib, VALUE,
- sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A1, MORSE_Complex64_t, A1m, A1n), INOUT,
- sizeof(int), &lda1, VALUE,
- sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A2, MORSE_Complex64_t, A2m, A2n), INOUT | LOCALITY,
- sizeof(int), &lda2, VALUE,
- sizeof(MORSE_Complex64_t)*ib*nb, RTBLKADDR(T, MORSE_Complex64_t, Tm, Tn), OUTPUT,
- sizeof(int), &ldt, VALUE,
- sizeof(MORSE_Complex64_t)*nb, NULL, SCRATCH,
- sizeof(MORSE_Complex64_t)*ib*nb, NULL, SCRATCH,
+ sizeof(int), &m, VALUE,
+ sizeof(int), &n, VALUE,
+ sizeof(int), &ib, VALUE,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A1, MORSE_Complex64_t, A1m, A1n), INOUT | QUARK_REGION_L | QUARK_REGION_D,
+ sizeof(int), &lda1, VALUE,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A2, MORSE_Complex64_t, A2m, A2n), INOUT | QUARK_REGION_L | QUARK_REGION_D | LOCALITY,
+ sizeof(int), &lda2, VALUE,
+ sizeof(MORSE_Complex64_t)*ib*nb, RTBLKADDR(T, MORSE_Complex64_t, Tm, Tn), OUTPUT,
+ sizeof(int), &ldt, VALUE,
+ sizeof(MORSE_Complex64_t)*nb, NULL, SCRATCH,
+ sizeof(MORSE_Complex64_t)*ib*nb, NULL, SCRATCH,
0);
}
diff --git a/runtime/quark/codelets/codelet_zttmlq.c b/runtime/quark/codelets/codelet_zttmlq.c
index e2de20109..f7a258e6b 100644
--- a/runtime/quark/codelets/codelet_zttmlq.c
+++ b/runtime/quark/codelets/codelet_zttmlq.c
@@ -165,23 +165,23 @@ void MORSE_TASK_zttmlq(const MORSE_option_t *options,
quark_option_t *opt = (quark_option_t*)(options->schedopt);
DAG_CORE_TTMLQ;
QUARK_Insert_Task(opt->quark, CORE_zttmlq_quark, (Quark_Task_Flags*)opt,
- sizeof(MORSE_enum), &side, VALUE,
- sizeof(MORSE_enum), &trans, VALUE,
- sizeof(int), &m1, VALUE,
- sizeof(int), &n1, VALUE,
- sizeof(int), &m2, VALUE,
- sizeof(int), &n2, VALUE,
- sizeof(int), &k, VALUE,
- sizeof(int), &ib, VALUE,
- sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A1, MORSE_Complex64_t, A1m, A1n), INOUT,
- sizeof(int), &lda1, VALUE,
- sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A2, MORSE_Complex64_t, A2m, A2n), INOUT,
- sizeof(int), &lda2, VALUE,
- sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(V, MORSE_Complex64_t, Vm, Vn), INPUT|QUARK_REGION_D|QUARK_REGION_L,
- sizeof(int), &ldv, VALUE,
- sizeof(MORSE_Complex64_t)*ib*nb, RTBLKADDR(T, MORSE_Complex64_t, Tm, Tn), INPUT,
- sizeof(int), &ldt, VALUE,
- sizeof(MORSE_Complex64_t)*ib*nb, NULL, SCRATCH,
- sizeof(int), &ldwork, VALUE,
+ sizeof(MORSE_enum), &side, VALUE,
+ sizeof(MORSE_enum), &trans, VALUE,
+ sizeof(int), &m1, VALUE,
+ sizeof(int), &n1, VALUE,
+ sizeof(int), &m2, VALUE,
+ sizeof(int), &n2, VALUE,
+ sizeof(int), &k, VALUE,
+ sizeof(int), &ib, VALUE,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A1, MORSE_Complex64_t, A1m, A1n), INOUT,
+ sizeof(int), &lda1, VALUE,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A2, MORSE_Complex64_t, A2m, A2n), INOUT | LOCALITY,
+ sizeof(int), &lda2, VALUE,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(V, MORSE_Complex64_t, Vm, Vn), INPUT | QUARK_REGION_L | QUARK_REGION_D,
+ sizeof(int), &ldv, VALUE,
+ sizeof(MORSE_Complex64_t)*ib*nb, RTBLKADDR(T, MORSE_Complex64_t, Tm, Tn), INPUT,
+ sizeof(int), &ldt, VALUE,
+ sizeof(MORSE_Complex64_t)*ib*nb, NULL, SCRATCH,
+ sizeof(int), &ldwork, VALUE,
0);
}
diff --git a/runtime/quark/codelets/codelet_zttmqr.c b/runtime/quark/codelets/codelet_zttmqr.c
index ed4311d99..89402c127 100644
--- a/runtime/quark/codelets/codelet_zttmqr.c
+++ b/runtime/quark/codelets/codelet_zttmqr.c
@@ -32,7 +32,8 @@
#include "chameleon/morse_tasks_z.h"
#include "coreblas/coreblas_z.h"
-void CORE_zttmqr_quark(Quark *quark)
+static void
+CORE_zttmqr_quark( Quark *quark )
{
MORSE_enum side;
MORSE_enum trans;
@@ -165,23 +166,23 @@ void MORSE_TASK_zttmqr(const MORSE_option_t *options,
quark_option_t *opt = (quark_option_t*)(options->schedopt);
DAG_CORE_TTMQR;
QUARK_Insert_Task(opt->quark, CORE_zttmqr_quark, (Quark_Task_Flags*)opt,
- sizeof(MORSE_enum), &side, VALUE,
- sizeof(MORSE_enum), &trans, VALUE,
- sizeof(int), &m1, VALUE,
- sizeof(int), &n1, VALUE,
- sizeof(int), &m2, VALUE,
- sizeof(int), &n2, VALUE,
- sizeof(int), &k, VALUE,
- sizeof(int), &ib, VALUE,
- sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A1, MORSE_Complex64_t, A1m, A1n), INOUT,
- sizeof(int), &lda1, VALUE,
- sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A2, MORSE_Complex64_t, A2m, A2n), INOUT,
- sizeof(int), &lda2, VALUE,
- sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(V, MORSE_Complex64_t, Vm, Vn), INPUT|QUARK_REGION_D|QUARK_REGION_U,
- sizeof(int), &ldv, VALUE,
- sizeof(MORSE_Complex64_t)*ib*nb, RTBLKADDR(T, MORSE_Complex64_t, Tm, Tn), INPUT,
- sizeof(int), &ldt, VALUE,
- sizeof(MORSE_Complex64_t)*ib*nb, NULL, SCRATCH,
- sizeof(int), &ldwork, VALUE,
+ sizeof(MORSE_enum), &side, VALUE,
+ sizeof(MORSE_enum), &trans, VALUE,
+ sizeof(int), &m1, VALUE,
+ sizeof(int), &n1, VALUE,
+ sizeof(int), &m2, VALUE,
+ sizeof(int), &n2, VALUE,
+ sizeof(int), &k, VALUE,
+ sizeof(int), &ib, VALUE,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A1, MORSE_Complex64_t, A1m, A1n), INOUT,
+ sizeof(int), &lda1, VALUE,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A2, MORSE_Complex64_t, A2m, A2n), INOUT | LOCALITY,
+ sizeof(int), &lda2, VALUE,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(V, MORSE_Complex64_t, Vm, Vn), INPUT | QUARK_REGION_U | QUARK_REGION_D,
+ sizeof(int), &ldv, VALUE,
+ sizeof(MORSE_Complex64_t)*ib*nb, RTBLKADDR(T, MORSE_Complex64_t, Tm, Tn), INPUT,
+ sizeof(int), &ldt, VALUE,
+ sizeof(MORSE_Complex64_t)*ib*nb, NULL, SCRATCH,
+ sizeof(int), &ldwork, VALUE,
0);
}
diff --git a/runtime/quark/codelets/codelet_zttqrt.c b/runtime/quark/codelets/codelet_zttqrt.c
index c7397d924..45284f08c 100644
--- a/runtime/quark/codelets/codelet_zttqrt.c
+++ b/runtime/quark/codelets/codelet_zttqrt.c
@@ -133,16 +133,16 @@ void MORSE_TASK_zttqrt(const MORSE_option_t *options,
quark_option_t *opt = (quark_option_t*)(options->schedopt);
DAG_CORE_TTQRT;
QUARK_Insert_Task(opt->quark, CORE_zttqrt_quark, (Quark_Task_Flags*)opt,
- sizeof(int), &m, VALUE,
- sizeof(int), &n, VALUE,
- sizeof(int), &ib, VALUE,
- sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A1, MORSE_Complex64_t, A1m, A1n), INOUT,
- sizeof(int), &lda1, VALUE,
- sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A2, MORSE_Complex64_t, A2m, A2n), INOUT | LOCALITY,
- sizeof(int), &lda2, VALUE,
- sizeof(MORSE_Complex64_t)*ib*nb, RTBLKADDR(T, MORSE_Complex64_t, Tm, Tn), OUTPUT,
- sizeof(int), &ldt, VALUE,
- sizeof(MORSE_Complex64_t)*nb, NULL, SCRATCH,
- sizeof(MORSE_Complex64_t)*ib*nb, NULL, SCRATCH,
+ sizeof(int), &m, VALUE,
+ sizeof(int), &n, VALUE,
+ sizeof(int), &ib, VALUE,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A1, MORSE_Complex64_t, A1m, A1n), INOUT | QUARK_REGION_U | QUARK_REGION_D,
+ sizeof(int), &lda1, VALUE,
+ sizeof(MORSE_Complex64_t)*nb*nb, RTBLKADDR(A2, MORSE_Complex64_t, A2m, A2n), INOUT | QUARK_REGION_U | QUARK_REGION_D | LOCALITY,
+ sizeof(int), &lda2, VALUE,
+ sizeof(MORSE_Complex64_t)*ib*nb, RTBLKADDR(T, MORSE_Complex64_t, Tm, Tn), OUTPUT,
+ sizeof(int), &ldt, VALUE,
+ sizeof(MORSE_Complex64_t)*nb, NULL, SCRATCH,
+ sizeof(MORSE_Complex64_t)*ib*nb, NULL, SCRATCH,
0);
}
--
GitLab