Add atomic_fence to triage

Author: oleksandr-pavlyk

dpctl/tensor/libtensor/include/kernels/sorting/sort_utils.hpp

@@ -111,6 +111,8 @@ sycl::event map_back_impl(sycl::queue &exec_q,
         cgh.parallel_for<KernelName>(
             sycl::range<1>(nelems), [=](sycl::id<1> id) {
                 const IndexTy linear_index = flat_index_data[id];
+                sycl::atomic_fence(sycl::memory_order::acq_rel,
+                                   sycl::memory_scope::device);
                 reduced_index_data[id] = (linear_index % row_size);
             });
     });

dpctl/tensor/libtensor/include/kernels/sorting/topk.hpp

@@ -122,6 +122,9 @@ sycl::event write_out_impl(sycl::queue &exec_q,
                     const IndexTy res_ind = index_data[src_idx];
                     const argTy v = arg_tp[res_ind];
 
+                    sycl::atomic_fence(sycl::memory_order::acq_rel,
+                                       sycl::memory_scope::device);
+
                     const std::size_t dst_idx = data_id;
                     vals_tp[dst_idx] = v;
                     inds_tp[dst_idx] = (res_ind % axis_nelems);