int mmm1d_tune(char **log)
{
char buffer[32 + 2*ES_DOUBLE_SPACE + ES_INTEGER_SPACE];
double int_time, min_time=1e200, min_rad = -1;
double maxrad = box_l[2]; /* N_psi = 2, theta=2/3 maximum for rho */
double switch_radius;
if (mmm1d_params.far_switch_radius_2 < 0) {
/* determine besselcutoff and optimal switching radius. Should be around 0.33 */
for (switch_radius = 0.2*maxrad;
switch_radius < 0.4*maxrad;
switch_radius += 0.025*maxrad) {
if (switch_radius <= bessel_radii[MAXIMAL_B_CUT - 1]) {
// this switching radius is too small for our Bessel series
continue;
}
mmm1d_params.far_switch_radius_2 = SQR(switch_radius);
coulomb.method = COULOMB_MMM1D;
/* initialize mmm1d temporary structures */
mpi_bcast_coulomb_params();
/* perform force calculation test */
int_time = time_force_calc(TEST_INTEGRATIONS);
/* exit on errors */
if (int_time < 0)
return ES_ERROR;
sprintf(buffer, "r= %f t= %f ms\n", switch_radius, int_time);
*log = strcat_alloc(*log, buffer);
if (int_time < min_time) {
min_time = int_time;
min_rad = switch_radius;
}
/* stop if all hope is vain... */
else if (int_time > 2*min_time)
break;
}
switch_radius = min_rad;
mmm1d_params.far_switch_radius_2 = SQR(switch_radius);
}
else {
if (mmm1d_params.far_switch_radius_2 <= SQR(bessel_radii[MAXIMAL_B_CUT - 1])) {
// this switching radius is too small for our Bessel series
*log = strcat_alloc(*log, "could not find reasonable bessel cutoff");
return ES_ERROR;
}
}
coulomb.method = COULOMB_MMM1D;
mpi_bcast_coulomb_params();
return ES_OK;
}
/** Determine runtime for a specific cutoff */
double time_r_cut(double r_cut) {
double t;
/** Set cutoff to time */
scafacos->set_r_cut(r_cut);
/** Tune other parameters */
scafacos->tune(particles.charges, particles.positions);
on_coulomb_change();
return time_force_calc(10);
}
int tclcommand_time_integration(ClientData data, Tcl_Interp *interp, int argc, char *argv[]) {
char buffer[10+TCL_DOUBLE_SPACE];
double t;
int n = 1;
if(argc > 2) {
Tcl_AppendResult(interp, "time_integration expects zero or one argument.", (char *)NULL);
return TCL_ERROR;
}
if(argc == 2) {
if(!(ARG1_IS_I(n) && (n > 1))) {
return TCL_ERROR;
}
}
t = time_force_calc(n);
sprintf(buffer, "%lf", t);
Tcl_AppendResult(interp, buffer, (char *)NULL);
return TCL_OK;
}
//Tuning
int EwaldgpuForce::adaptive_tune(char **log,SystemInterface &s)
{
ewaldgpu_params.isTuned = false;
int Kmax = ewaldgpu_params.K_max;
double alpha_array[Kmax]; //All computed alpha in dependence of K
double rcut_array[Kmax]; //All computed r_cut in dependence of all computed alpha
//Squared charge
Particle *particle;
particle = (Particle*)malloc(n_part*sizeof(Particle));
mpi_get_particles(particle, NULL);
double q_sqr = compute_q_sqare(particle);
char b[3*ES_INTEGER_SPACE + 3*ES_DOUBLE_SPACE + 128];
if (skin == -1) {
*log = strcat_alloc(*log, "ewaldgpu cannot be tuned, since the skin is not yet set");
return ES_ERROR;
}
//Compute alpha for all reciprocal k-sphere radius K
for(int K = 0; K < Kmax ;K++)
{
alpha_array[K] = tune_alpha(ewaldgpu_params.accuracy/sqrt(2), ewaldgpu_params.precision, K+1, box_l[0]*box_l[1]*box_l[2], q_sqr, n_part);
//printf("K:%i alpha:%f\n",K+1,alpha_array[K]);
}
//Compute r_cut for all computed alpha
for(int K = 0; K < Kmax ;K++)
{
rcut_array[K] = tune_rcut(ewaldgpu_params.accuracy/sqrt(2), ewaldgpu_params.precision, alpha_array[K], box_l[0]*box_l[1]*box_l[2], q_sqr, n_part);
//printf("K:%i rcut:%f \n",K+1,rcut_array[K]);
}
//Test if accuracy was reached
if(rcut_array[Kmax-1]<0)
{
return ES_ERROR;
}
/***********************************************************************************
PERFORMANCE TIME
***********************************************************************************/
//Test performance time for the diverent (K, rcut, alpha)
double int_time_best = 1E30;
int K_best = Kmax;
for(int K = 0; K < Kmax ;K++)
{
if(alpha_array[K]>0 and rcut_array[K]>0 and rcut_array[K]<(std::min(box_l[0],std::min(box_l[1],box_l[2])))/2.0-skin)
{
set_params(rcut_array[K], K+1, K+1, K+1, alpha_array[K]);
mpi_bcast_coulomb_params();
double int_time = time_force_calc(ewaldgpu_params.time_calc_steps);
if(int_time<int_time_best)
{
int_time_best = int_time;
K_best = K;
}
//printf("TIME K:%i int_time:%f\n",K+1,int_time);
}
}
set_params(rcut_array[K_best], K_best+1, K_best+1, K_best+1, alpha_array[K_best]);
ewaldgpu_params.isTuned = true;
mpi_bcast_coulomb_params();
//Print Status
sprintf(b, "ewaldgpu tune parameters: Accuracy goal = %f\n", ewaldgpu_params.accuracy);
*log = strcat_alloc(*log, b);
sprintf(b, "ewaldgpu tune parameters: Alpha = %f\n", ewaldgpu_params.alpha);
*log = strcat_alloc(*log, b);
sprintf(b, "ewaldgpu tune parameters: r_cut = %f\n", ewaldgpu_params.rcut);
*log = strcat_alloc(*log, b);
sprintf(b, "ewaldgpu tune parameters: num_kx = %i\n", ewaldgpu_params.num_kx);
*log = strcat_alloc(*log, b);
sprintf(b, "ewaldgpu tune parameters: num_ky = %i\n", ewaldgpu_params.num_ky);
*log = strcat_alloc(*log, b);
sprintf(b, "ewaldgpu tune parameters: num_kz = %i\n", ewaldgpu_params.num_kz);
*log = strcat_alloc(*log, b);
return ES_OK;
}
int mmm1d_tune(char **log)
{
char buffer[32 + 2*ES_DOUBLE_SPACE + ES_INTEGER_SPACE];
double int_time, min_time=1e200, min_rad = -1;
double maxrad = box_l[2]; /* N_psi = 2, theta=2/3 maximum for rho */
double switch_radius;
if (mmm1d_params.bessel_cutoff < 0 && mmm1d_params.far_switch_radius_2 < 0) {
/* determine besselcutoff and optimal switching radius */
for (switch_radius = RAD_STEPPING*maxrad;
switch_radius < maxrad;
switch_radius += RAD_STEPPING*maxrad) {
mmm1d_params.bessel_cutoff = determine_bessel_cutoff(switch_radius, mmm1d_params.maxPWerror, MAXIMAL_B_CUT);
/* no reasonable cutoff possible */
if (mmm1d_params.bessel_cutoff == MAXIMAL_B_CUT)
continue;
mmm1d_params.far_switch_radius_2 = SQR(switch_radius);
coulomb.method = COULOMB_MMM1D;
/* initialize mmm1d temporary structures */
mpi_bcast_coulomb_params();
/* perform force calculation test */
int_time = time_force_calc(TEST_INTEGRATIONS);
/* exit on errors */
if (int_time < 0)
return ES_ERROR;
sprintf(buffer, "r= %f c= %d t= %f ms\n",
switch_radius, mmm1d_params.bessel_cutoff, int_time);
*log = strcat_alloc(*log, buffer);
if (int_time < min_time) {
min_time = int_time;
min_rad = switch_radius;
}
/* stop if all hope is vain... */
else if (int_time > 2*min_time)
break;
}
switch_radius = min_rad;
mmm1d_params.far_switch_radius_2 = SQR(switch_radius);
mmm1d_params.bessel_cutoff = determine_bessel_cutoff(switch_radius, mmm1d_params.maxPWerror, MAXIMAL_B_CUT);
mmm1d_params.bessel_calculated = 1;
}
else if (mmm1d_params.bessel_cutoff < 0) {
/* determine besselcutoff to achieve at least the given pairwise error */
mmm1d_params.bessel_cutoff = determine_bessel_cutoff(sqrt(mmm1d_params.far_switch_radius_2),
mmm1d_params.maxPWerror, MAXIMAL_B_CUT);
if (mmm1d_params.bessel_cutoff == MAXIMAL_B_CUT) {
*log = strcat_alloc(*log, "could not find reasonable bessel cutoff");
return ES_ERROR;
}
mmm1d_params.bessel_calculated = 1;
}
else
mmm1d_params.bessel_calculated = 0;
coulomb.method = COULOMB_MMM1D;
mpi_bcast_coulomb_params();
return ES_OK;
}
int scafacos_p3m_tuning(){
int mesh[3] = {0, 0, 0}, tmp_mesh_points;
int tmp_mesh[3];
double r_cut_iL_min, r_cut_iL_max, r_cut_iL = -1, tmp_r_cut_iL=0.0;
int cao;
double alpha_L = -1, tmp_alpha_L=0.0;
double accuracy = -1, tmp_accuracy=0.0;
double time_best=1e20, tmp_time;
char b[3*ES_INTEGER_SPACE + 3*ES_DOUBLE_SPACE + 128];
int buffer_cao, buffer_grid;
double buffer_alpha_L, buffer_r_cut_iL, buffer_tolerance_field;
if (skin == -1) {
// *log = strcat_alloc(*log, "p3m cannot be tuned, since the skin is not yet set");
return ES_ERROR;
}
/* preparation */
mpi_bcast_event(FCS_P3M_COUNT_CHARGES);
/* Print Status */
sprintf(b, "P3M tune parameters: Accuracy goal = %.5e\n", scafacos_p3m.params.tolerance_field);
// *log = strcat_alloc(*log, b);
sprintf(b, "System: box_l = %.5e # charged part = %d Sum[q_i^2] = %.5e\n", box_l[0], scafacos_p3m.sum_qpart, scafacos_p3m.sum_q2);
// *log = strcat_alloc(*log, b);
if (scafacos_p3m.sum_qpart == 0) {
// *log = strcat_alloc(*log, "no charged particles in the system, cannot tune P3M");
return ES_ERROR;
}
double density, r_cut;
density = n_total_particles /(box_l[0]* box_l[1]* box_l[2]);
int number = 100;
r_cut= pow(number/ density/3.14 *3 /4 , 1/3.0);
if(scafacos_p3m.params.r_cut_iL == 0.0) {
// WARNING: r_cut_iL_min should not be small !
r_cut_iL_min = dmin(min_local_box_l, min_box_l/2);
r_cut_iL_min *= 0.18;
r_cut_iL_min = r_cut *0.8;
r_cut_iL_max = dmin(min_local_box_l, min_box_l/2);
r_cut_iL_max *= 0.28;
r_cut_iL_max = r_cut *1.5;
if(r_cut_iL_max >= 4.5)
r_cut_iL_max = 4.5;
// r_cut_iL_max = dmin(min_local_box_l, min_box_l/2);// - skin;
// r_cut_iL_min *= box_l_i[0];
// r_cut_iL_max *= box_l_i[0];
fprintf(stderr, "r_cut_iL_max: %f r_cut_iL_min: %f \n", r_cut_iL_max, r_cut_iL_min);
fprintf(stderr, "box_l is: %f \n", box_l[0]);
}
else{
fprintf(stderr, "Scafacos_p3m tuning varies the cutoff radius; you have to leave cutoff unspecified");
return ES_ERROR;
}
// *log = strcat_alloc(*log, "mesh cao r_cut_iL alpha_L err rs_err ks_err time [ms]\n");
for(tmp_r_cut_iL = r_cut_iL_min; tmp_r_cut_iL <= r_cut_iL_max; tmp_r_cut_iL += 0.1){
scafacos_p3m.params.cutoff = tmp_r_cut_iL;
tmp_time = time_force_calc(1);
// *log = strcat_alloc(*log, " %f ms ");
fprintf(stderr, "time: %f \n", tmp_time);
// *log = strcat_alloc(*log, b);
if ((tmp_time < time_best) && (tmp_time > 0)) {
time_best = tmp_time;
fcs_p3m_get_grid(fcs_handle, &buffer_grid);
fcs_p3m_get_cao(fcs_handle, &buffer_cao);
fcs_p3m_get_r_cut(fcs_handle, &buffer_r_cut_iL);
fcs_p3m_get_alpha(fcs_handle, &buffer_alpha_L);
fcs_p3m_get_tolerance_field(fcs_handle, &buffer_tolerance_field);
// target accuracy remains the same
}
/* no hope of further optimisation */
//else if (tmp_time > time_best + P3M_TIME_GRAN) {
// break;
//}
}
if(time_best == 1e20) {
fprintf(stderr, "failed to tune P3M parameters to required accuracy\n");
return ES_ERROR;
}
/* prepare tuned p3m parameters for broadcast*/
scafacos_p3m.params.cao = buffer_cao;
scafacos_p3m.params.alpha_L = buffer_alpha_L;
scafacos_p3m.params.grid = buffer_grid;
scafacos_p3m.params.cutoff = buffer_r_cut_iL;
scafacos_p3m.params.alpha_L = buffer_alpha_L;
/* Tell the user about the outcome */
fprintf(stderr, "\nresulting parameters:\n%-4d %-3d %.5e %.5e %.5e %-8d\n", buffer_grid, buffer_cao, buffer_r_cut_iL, buffer_alpha_L, buffer_tolerance_field, (int)time_best);
return ES_OK;
}