summaryrefslogtreecommitdiff
path: root/src/routines/level3/xherk.cc
diff options
context:
space:
mode:
authorCedric Nugteren <web@cedricnugteren.nl>2016-06-16 18:07:46 +0200
committerCedric Nugteren <web@cedricnugteren.nl>2016-06-16 18:07:46 +0200
commit52ccaf5b25e14c9ce032315e5e96b1f27886d481 (patch)
tree087288b7aebf2a06ffc4e7dcbcd4353f7a3be6a7 /src/routines/level3/xherk.cc
parent39b7dbc5e37829abfbcfb77852b9138b31540b42 (diff)
Added XOMATCOPY routines to perform out-of-place matrix scaling, copying, and/or transposing
Diffstat (limited to 'src/routines/level3/xherk.cc')
-rw-r--r--src/routines/level3/xherk.cc12
1 files changed, 8 insertions, 4 deletions
diff --git a/src/routines/level3/xherk.cc b/src/routines/level3/xherk.cc
index 9ab50dd2..df97a94f 100644
--- a/src/routines/level3/xherk.cc
+++ b/src/routines/level3/xherk.cc
@@ -124,7 +124,8 @@ StatusCode Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, cons
status = PadCopyTransposeMatrix(eventProcessA.pointer(), emptyEventList,
a_one, a_two, a_ld, a_offset, a_buffer,
n_ceiled, k_ceiled, n_ceiled, 0, a_temp,
- program, true, a_rotated, a_conjugate);
+ ConstantOne<T>(), program,
+ true, a_rotated, a_conjugate);
eventWaitList.push_back(eventProcessA);
if (ErrorIn(status)) { return status; }
}
@@ -133,7 +134,8 @@ StatusCode Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, cons
status = PadCopyTransposeMatrix(eventProcessB.pointer(), emptyEventList,
a_one, a_two, a_ld, a_offset, a_buffer,
n_ceiled, k_ceiled, n_ceiled, 0, b_temp,
- program, true, a_rotated, b_conjugate);
+ ConstantOne<T>(), program,
+ true, a_rotated, b_conjugate);
eventWaitList.push_back(eventProcessB);
if (ErrorIn(status)) { return status; }
}
@@ -144,7 +146,8 @@ StatusCode Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, cons
status = PadCopyTransposeMatrix(eventProcessC.pointer(), emptyEventList,
n, n, c_ld, c_offset, c_buffer,
n_ceiled, n_ceiled, n_ceiled, 0, c_temp,
- program, true, c_rotated, false);
+ ConstantOne<T>(), program,
+ true, c_rotated, false);
eventWaitList.push_back(eventProcessC);
if (ErrorIn(status)) { return status; }
@@ -180,7 +183,8 @@ StatusCode Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, cons
status = PadCopyTransposeMatrix(event_, eventWaitList,
n_ceiled, n_ceiled, n_ceiled, 0, c_temp,
n, n, c_ld, c_offset, c_buffer,
- program, false, c_rotated, false, upper, lower, true);
+ ConstantOne<T>(), program,
+ false, c_rotated, false, upper, lower, true);
if (ErrorIn(status)) { return status; }
// Successfully finished the computation