diff options
Diffstat (limited to 'src/routines/level3/xherk.cpp')
-rw-r--r-- | src/routines/level3/xherk.cpp | 8 |
1 files changed, 4 insertions, 4 deletions
diff --git a/src/routines/level3/xherk.cpp b/src/routines/level3/xherk.cpp index cc87e3e9..0fa1b7b1 100644 --- a/src/routines/level3/xherk.cpp +++ b/src/routines/level3/xherk.cpp @@ -111,7 +111,7 @@ StatusCode Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, cons // case nothing has to be done, these kernels can be skipped. Two copies are created. if (!a_no_temp) { auto eventProcessA = Event(); - status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessA.pointer(), emptyEventList, + status = PadCopyTransposeMatrix(queue_, device_, db_, eventProcessA.pointer(), emptyEventList, a_one, a_two, a_ld, a_offset, a_buffer, n_ceiled, k_ceiled, n_ceiled, 0, a_temp, ConstantOne<T>(), program, @@ -121,7 +121,7 @@ StatusCode Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, cons } if (!b_no_temp) { auto eventProcessB = Event(); - status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessB.pointer(), emptyEventList, + status = PadCopyTransposeMatrix(queue_, device_, db_, eventProcessB.pointer(), emptyEventList, a_one, a_two, a_ld, a_offset, a_buffer, n_ceiled, k_ceiled, n_ceiled, 0, b_temp, ConstantOne<T>(), program, @@ -133,7 +133,7 @@ StatusCode Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, cons // Furthermore, also creates a (possibly padded) copy of matrix C, since it is not allowed to // modify the other triangle. auto eventProcessC = Event(); - status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessC.pointer(), emptyEventList, + status = PadCopyTransposeMatrix(queue_, device_, db_, eventProcessC.pointer(), emptyEventList, n, n, c_ld, c_offset, c_buffer, n_ceiled, n_ceiled, n_ceiled, 0, c_temp, ConstantOne<T>(), program, @@ -170,7 +170,7 @@ StatusCode Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, cons // Runs the post-processing kernel auto upper = (triangle == Triangle::kUpper); auto lower = (triangle == Triangle::kLower); - status = PadCopyTransposeMatrix(queue_, device_, context_, db_, event_, eventWaitList, + status = PadCopyTransposeMatrix(queue_, device_, db_, event_, eventWaitList, n_ceiled, n_ceiled, n_ceiled, 0, c_temp, n, n, c_ld, c_offset, c_buffer, ConstantOne<T>(), program, |