diff options
Diffstat (limited to 'src/routines/level3/xher2k.cc')
-rw-r--r-- | src/routines/level3/xher2k.cc | 16 |
1 files changed, 8 insertions, 8 deletions
diff --git a/src/routines/level3/xher2k.cc b/src/routines/level3/xher2k.cc index c891c202..43f7bb76 100644 --- a/src/routines/level3/xher2k.cc +++ b/src/routines/level3/xher2k.cc @@ -129,7 +129,7 @@ StatusCode Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, co // case nothing has to be done, these kernels can be skipped. if (!a1_no_temp) { auto eventProcessA1 = Event(); - status = PadCopyTransposeMatrix(eventProcessA1.pointer(), emptyEventList, + status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessA1.pointer(), emptyEventList, ab_one, ab_two, a_ld, a_offset, a_buffer, n_ceiled, k_ceiled, n_ceiled, 0, a1_temp, ConstantOne<T>(), program, @@ -139,7 +139,7 @@ StatusCode Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, co } if (!a2_no_temp) { auto eventProcessA2 = Event(); - status = PadCopyTransposeMatrix(eventProcessA2.pointer(), emptyEventList, + status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessA2.pointer(), emptyEventList, ab_one, ab_two, a_ld, a_offset, a_buffer, n_ceiled, k_ceiled, n_ceiled, 0, a2_temp, ConstantOne<T>(), program, @@ -149,7 +149,7 @@ StatusCode Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, co } if (!b1_no_temp) { auto eventProcessB1 = Event(); - status = PadCopyTransposeMatrix(eventProcessB1.pointer(), emptyEventList, + status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessB1.pointer(), emptyEventList, ab_one, ab_two, b_ld, b_offset, b_buffer, n_ceiled, k_ceiled, n_ceiled, 0, b1_temp, ConstantOne<T>(), program, @@ -159,7 +159,7 @@ StatusCode Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, co } if (!b2_no_temp) { auto eventProcessB2 = Event(); - status = PadCopyTransposeMatrix(eventProcessB2.pointer(), emptyEventList, + status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessB2.pointer(), emptyEventList, ab_one, ab_two, b_ld, b_offset, b_buffer, n_ceiled, k_ceiled, n_ceiled, 0, b2_temp, ConstantOne<T>(), program, @@ -171,7 +171,7 @@ StatusCode Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, co // 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(eventProcessC.pointer(), emptyEventList, + status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessC.pointer(), emptyEventList, n, n, c_ld, c_offset, c_buffer, n_ceiled, n_ceiled, n_ceiled, 0, c_temp, ConstantOne<T>(), program, @@ -201,7 +201,7 @@ StatusCode Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, co // Launches the kernel auto eventKernel1 = Event(); - status = RunKernel(kernel, global, local, eventKernel1.pointer(), eventWaitList); + status = RunKernel(kernel, queue_, device_, global, local, eventKernel1.pointer(), eventWaitList); if (ErrorIn(status)) { return status; } eventWaitList.push_back(eventKernel1); @@ -217,14 +217,14 @@ StatusCode Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, co // Runs the kernel again auto eventKernel2 = Event(); - status = RunKernel(kernel, global, local, eventKernel2.pointer(), eventWaitList); + status = RunKernel(kernel, queue_, device_, global, local, eventKernel2.pointer(), eventWaitList); if (ErrorIn(status)) { return status; } eventWaitList.push_back(eventKernel2); // Runs the post-processing kernel auto upper = (triangle == Triangle::kUpper); auto lower = (triangle == Triangle::kLower); - status = PadCopyTransposeMatrix(event_, eventWaitList, + status = PadCopyTransposeMatrix(queue_, device_, context_, db_, event_, eventWaitList, n_ceiled, n_ceiled, n_ceiled, 0, c_temp, n, n, c_ld, c_offset, c_buffer, ConstantOne<T>(), program, |