Commit 3641144c authored by Giovanni La Mura's avatar Giovanni La Mura
Browse files

Compute the ram overhead estimate at the beginning of inclusion()

parent 437d1a74

src/inclusion/inclusion.cpp

+60 −2
Original line number Diff line number Diff line
@@ -146,10 +146,12 @@ void inclusion(const string& config_file, const string& data_file, const string& 
  }

  Logger *logger = new Logger(LOG_DEBG);

  int device_count = 0;

  const double ram_overhead_factor = get_ram_overhead();



#ifdef USE_MAGMA

  //===========

  // Initialise MAGMA

  //===========

#ifdef USE_MAGMA

  const magma_int_t d_array_max_size = 32; // TEMPORARY: can become configurable parameter

  magma_device_t *device_array = new magma_device_t[d_array_max_size];

  magma_int_t num_devices;
@@ -228,8 +230,64 @@ void inclusion(const string& config_file, const string& data_file, const string& 
    int nsph = gconf->number_of_spheres;

    // Sanity check on number of sphere consistency, should always be verified

    if (s_nsph == nsph) {

      // Shortcuts to variables stored in configuration objects

      char virtual_line[256];

      sprintf(virtual_line, "%.5lg.\n", sconf->get_particle_radius(gconf));

      message = "INFO: particle radius is " + (string)virtual_line;

      logger->log(message);

      // Memory requirements test

      long cid_size_bytes = InclusionIterationData::get_size(gconf, sconf);

      double cid_size_gb = cid_size_bytes / 1024.0 / 1024.0 / 1024.0;

      sprintf(virtual_line, "%.5lg", cid_size_gb);

      message = "INFO: iteration data requires " + (string)virtual_line + "GiB of RAM.\n";

      logger->log(message);

      int omp_wavelength_threads = 1;

#ifdef _OPENMP

#pragma omp parallel

      {

	int threadId = omp_get_thread_num();

	if (threadId == 0) {

	  omp_wavelength_threads = omp_get_num_threads();

	}

      }

#endif //_OPENMP

      double requested_ram_gb = (ram_overhead_factor + omp_wavelength_threads) * cid_size_gb;

      sprintf(virtual_line, "%.5lg", requested_ram_gb);

      message = "INFO: code execution needs " + (string)virtual_line + "GiB of RAM.\n";

      logger->log(message);

      if (gconf->host_ram_gb > 0.0) {

	if (requested_ram_gb > gconf->host_ram_gb) {

	  // ERROR: host system does not have the necessary RAM

	  message = "ERROR: the requested model saturates the system RAM!\n";

	  logger->log(message);

	  fclose(timing_file);

	  delete time_logger;

	  delete logger;

	  delete sconf;

	  delete gconf;

	  exit(1);

	}

      }

      if (gconf->gpu_ram_gb > 0.0) {

	// mat_size_bytes = sizeof(dcomplex) * 2 * 2 * NSPH * NSPH * LI * LI * (LI + 2) * (LI + 2)

	long matrix_size_bytes = sizeof(dcomplex)

	  * 4 * gconf->number_of_spheres * gconf->number_of_spheres

	  * gconf->li * gconf->li * (gconf->li + 2) * (gconf->li + 2);

	double matrix_size_gb = matrix_size_bytes / 1024.0 / 1024.0 / 1024.0;

	int refinement_factor = (gconf->refine_flag) ? 3 : 1;

	if (refinement_factor * omp_wavelength_threads * matrix_size_gb > gconf->gpu_ram_gb) {

	  // ERROR: GPU does not have the necessary RAM

	  message = "ERROR: the requested model saturates the GPU RAM!\n";

	  logger->log(message);

	  fclose(timing_file);

	  delete time_logger;

	  delete logger;

	  delete sconf;

	  delete gconf;

	  exit(1);

	}

      }

      ScatteringAngles *p_scattering_angles = new ScatteringAngles(gconf);

      // Shortcuts to variables stored in configuration objects

      double wp = sconf->wp;

      // Open empty virtual ascii file for output

      InclusionOutputInfo *p_output = new InclusionOutputInfo(sconf, gconf, mpidata);