Fix some bugs.

device/selections/Hlt1/src/GatherSelections.cu

@@ -176,7 +176,7 @@ void gather_selections::gather_selections_t::operator()(
   Allen::copy_async<dev_fns_parameters_t, host_fns_parameters_t>(arguments, context);
 
   // * Run all selections in one go
-  global_function(gather_selections::run_lines)(first<host_number_of_events_t>(arguments), dim3(32, 4), context)(
+  global_function(gather_selections::run_lines)(first<host_number_of_events_t>(arguments), dim3(32, 8), context)(
     data<dev_fn_indices_t>(arguments), data<dev_fns_parameters_t>(arguments), 
     data<dev_odin_raw_input_t>(arguments),
     data<dev_odin_raw_input_offsets_t>(arguments),

device/selections/line_types/include/Line.cuh

@@ -150,10 +150,10 @@ process_line(char* input, unsigned run_no, unsigned evt_hi, unsigned evt_lo, uns
 
     // Check if blockIdx.x (event_number) is in dev_event_list
     unsigned mask = 0;
-    for (unsigned i = threadIdx.x; i < number_of_events; i += blockDim.x) {
-      mask |= __ballot_sync(0xFFFFFFFF, parameters.dev_event_list[i]);
+    for (unsigned i = 0; i < (number_of_events + 31) / 32; ++i) {
+      const auto index = i * 32 + threadIdx.x;
+      mask |= __ballot_sync(0xFFFFFFFF, index < number_of_events ? threadIdx.x == parameters.dev_event_list[i] : false);
     }
-    __syncwarp();
 
     if (mask) {
       const auto pre_scaler_hash = std::get<3>(type_casted_input);