Commit 3af28bf7 authored by Giovanni La Mura's avatar Giovanni La Mura
Browse files

Bracket nvtxRange calls into compiler flags

parent f2018157
Loading
Loading
Loading
Loading
+12 −0
Original line number Diff line number Diff line
@@ -544,13 +544,17 @@ void frfme(string data_file, string output_path) {
	  t_start = chrono::high_resolution_clock::now();
	  message = "INFO: Mapping data to device.\n";
	  logger.log(message);
#ifdef USE_NVTX
	  nvtxRangePush("Mapping to device");
#endif
	  map_data(
            vec_wsum, size_vec_wsum, global_vec_w, size_global_vec_w, vec_tt1_wk,
	    size_vec_tt1_wk, vkv, _xv, nxv, _yv, nyv, _zv, nzv, vec_vkzm, jlmf, jlml,
	    nkv, nlmmt
          );
#ifdef USE_NVTX
	  nvtxRangePop();
#endif
	  t_end = chrono::high_resolution_clock::now();
	  elapsed = t_end - t_start;
	  frfme_duration += elapsed;
@@ -560,13 +564,17 @@ void frfme(string data_file, string output_path) {
	  t_start = chrono::high_resolution_clock::now();
	  message = "INFO: computing loop.\n";
	  logger.log(message);
#ifdef USE_NVTX
	  nvtxRangePush("Offloaded loop");
#endif
	  offload_loop(
            vec_wsum, size_vec_wsum, global_vec_w, size_global_vec_w, vec_tt1_wk,
	    size_vec_tt1_wk, vkv, _xv, nxv, _yv, nyv, _zv, nzv, vec_vkzm, jlmf, jlml,
	    nkv, nlmmt, delks, frsh
          );
#ifdef USE_NVTX
	  nvtxRangePop();
#endif
	  t_end = chrono::high_resolution_clock::now();
	  elapsed = t_end - t_start;
	  frfme_duration += elapsed;
@@ -576,13 +584,17 @@ void frfme(string data_file, string output_path) {
	  t_start = chrono::high_resolution_clock::now();
	  message = "INFO: cleaning device memory.\n";
	  logger.log(message);
#ifdef USE_NVTX
	  nvtxRangePush("Cleaning device");
#endif
	  unmap_data(
            vec_wsum, size_vec_wsum, global_vec_w, size_global_vec_w, vec_tt1_wk,
	    size_vec_tt1_wk, vkv, _xv, nxv, _yv, nyv, _zv, nzv, vec_vkzm, jlmf, jlml,
	    nkv, nlmmt
          );
#ifdef USE_NVTX
	  nvtxRangePop();
#endif
	  t_end = chrono::high_resolution_clock::now();
	  elapsed = t_end - t_start;
	  frfme_duration += elapsed;