sgemm_wg: Output CPU data to binary

kernel/src/vx_spawn.c

@@ -258,7 +258,6 @@ void vx_spawn_tasks_cluster(int num_tasks, vx_spawn_tasks_cb callback, void *arg
   // threads, handle this in the last wave amongst other full warps.
   if (rem_threads_in_last_warp != 0 && core_id_in_cluster == 0) {
     // adjust offset
-    // FIXME: consider cluster_id here
     // FIXME: use rem_threads_in_last_warp_this_core
     wspawn_args.offset += (num_tasks_this_cluster - rem_threads_in_last_warp);
 

tests/regression/sgemm_wg/main.cpp

@@ -119,6 +119,14 @@ int run_test(const kernel_arg_t& kernel_arg,
   file.write(reinterpret_cast<char *>(staging_buf.data()), buf_size);
   file.close();
 
+  std::ofstream ref_file("reference.c.bin", std::ios::binary | std::ios::out);
+  if (!ref_file) {
+    std::cerr << "error: failed to open reference.c.bin for writing\n";
+    exit(EXIT_FAILURE);
+  }
+  ref_file.write(reinterpret_cast<char *>(ref_data.data()), buf_size);
+  ref_file.close();
+
   // verify result
   std::cout << "verify result" << std::endl;
   {