/*@T
*
* The [[place_particle]] routine determines the initial particle
* placement, but not the desired mass. We want the fluid in the
* initial configuration to exist roughly at the reference density.
* One way to do this is to take the volume in the indicated body of
* fluid, multiply by the mass density, and divide by the number of
* particles; but that requires that we be able to compute the volume
* of the fluid region. Alternately, we can simply compute the
* average mass density assuming each particle has mass one, then use
* that to compute the particle mass necessary in order to achieve the
* desired reference density. We do this with [[normalize_mass]].
*
* @c*/
void normalize_mass(sim_state_t* s, proc_info* pInfo, sim_param_t* param)
{
if (pInfo->proc == 0) {
s->mass = 1; // Set mass with only one processor
//hash_particles_parallel(s, pInfo, param->h);
hash_particles(s, param->h); // Hashing with only one processor
}
float rho0 = param->rho0;
float rho2s = 0;
float rhos = 0;
#pragma omp barrier // Barrier because need hashing information before computing density
compute_density(s, pInfo, param);
#pragma omp barrier // Want all processor to finish updating their own densities
//printf("Starting: %d\n", pInfo->proc);
#pragma omp parallel for reduction(+:rhos, rho2s)
for (int i = 0; i < s->n; ++i) {
rho2s += (s->part[i].rho)*(s->part[i].rho);
rhos += s->part[i].rho;
}
#pragma omp single // Only one processor to update this
{
s->mass *= ( rho0*rhos / rho2s );
}
}
/*@T
*
* The [[place_particle]] routine determines the initial particle
* placement, but not the desired mass. We want the fluid in the
* initial configuration to exist roughly at the reference density.
* One way to do this is to take the volume in the indicated body of
* fluid, multiply by the mass density, and divide by the number of
* particles; but that requires that we be able to compute the volume
* of the fluid region. Alternately, we can simply compute the
* average mass density assuming each particle has mass one, then use
* that to compute the particle mass necessary in order to achieve the
* desired reference density. We do this with [[normalize_mass]].
*
* @c*/
void normalize_mass(sim_state_t* s, sim_param_t* param)
{
s->mass = 1;
hash_particles(s, param->h);
compute_density(s, param);
float rho0 = param->rho0;
float rho2s = 0;
float rhos = 0;
for (int i = 0; i < s->n; ++i) {
rho2s += (s->part[i].rho)*(s->part[i].rho);
rhos += s->part[i].rho;
}
s->mass *= ( rho0*rhos / rho2s );
}
int main(int argc, char** argv)
{
sim_param_t params;
if (get_params(argc, argv, ¶ms) != 0)
exit(-1);
// Create global
sim_state_t* globalState = init_particles(¶ms);
#pragma omp parallel shared(globalState, params)
{
int proc = omp_get_thread_num();
int nproc = omp_get_num_threads();
FILE* fp = fopen(params.fname, "w");
int nframes = params.nframes;
int npframe = params.npframe;
float dt = params.dt;
int n = globalState->n;
// Processor information and holder
proc_info* pInfo = malloc(sizeof(proc_info));
pInfo->proc = proc;
pInfo->nproc = nproc;
pInfo->beg = round((proc/(double)nproc)*n);
pInfo->end = round(((proc+1)/(double)nproc)*n);
pInfo->forceAccu = calloc(3*n, sizeof(float)); // Never used this...
if (proc == 0) {
printf("Running in parallel with %d processor\n", nproc);
}
normalize_mass(globalState, pInfo, ¶ms);
double t_start = omp_get_wtime();
if (proc == 0) { // We only write for one processor
write_header(fp, n, nframes, params.h);
write_frame_data(fp, n, globalState, NULL);
}
if (proc == 0) {
hash_particles(globalState, params.h);
}
//hash_particles_parallel(globalState, pInfo, params.h);
#pragma omp barrier // Need the hashing to be done
compute_accel(globalState, pInfo, ¶ms);
#pragma omp barrier
leapfrog_start(globalState, pInfo, dt);
check_state(globalState, pInfo);
for (int frame = 1; frame < nframes; ++frame) {
// We sort according to Z-Morton to ensure locality, need to implement paralle qsort
if (frame % 5 == 0) {
// Dividing into chunks of sorting each chunk
// This alone turned out to better than sorting the entire array
qsort(globalState->part+pInfo->beg, pInfo->end-pInfo->beg ,sizeof(particle_t),compPart);
// Sorting the array consisting of sorted chunks
// This turned out to actually lower the performance. That's why
// I commented it.
// #pragma omp barrier
// if( pInfo->nproc >1 ) arraymerge(globalState->part, globalState->n, pInfo);
//#pragma omp barrier*/
// Serial version
/*#pragma omp single // Implied barrier
qsort(globalState->part, n, sizeof(particle_t), compPart);*/
}
/*else if (frame % 49) {*/
/*if (proc == 0) {*/
/*}*/
/*}*/
#pragma omp barrier // Need sort to finish
for (int i = 0; i < npframe; ++i) {
if (proc == 0 && npframe % 4 == 0) { // Ammortize hashing cost
hash_particles(globalState, params.h);
}
#pragma omp barrier
compute_accel(globalState, pInfo, ¶ms);
leapfrog_step(globalState, pInfo, dt);
check_state(globalState, pInfo);
#pragma omp barrier
}
if (proc == 0) {
printf("Frame: %d of %d - %2.1f%%\n",frame, nframes,
100*(float)frame/nframes);
write_frame_data(fp, n, globalState, NULL);
}
}
double t_end = omp_get_wtime();
if (proc == 0) {
printf("Ran in %g seconds\n", t_end-t_start);
}
free(pInfo);
fclose(fp);
}
free_state(globalState);
}
void compute_accel(sim_state_t* state, sim_param_t* params) {
// Unpack basic parameters
const float h = params->h;
const float rho0 = params->rho0;
// const float k = params->k;
// const float mu = params->mu;
const float g = params->g;
// const float mass = state->mass;
const float h2 = params->h2;
// Unpack system state
particle_t* p = state->part;
particle_t** hash = state->hash;
const int n = state->n;
// Rehash the particles
hash_particles(state, h);
// Compute density and color
compute_density(state, params);
// Constants for interaction term
const float C0 = params->C0;
const float Cp = params->Cp;
const float Cv = params->Cv;
// Start with gravity and surface forces
for (int i = 0; i < n; ++i)
{
vec3_set(p[i].a, 0, -g, 0);
}
// Accumulate forces
#ifdef USE_BUCKETING
/* BEGIN TASK */
// Start multi-threaded
// Iterate over each bucket, and within each bucket each particle
#pragma omp parallel
{
// Get the thread ID and total threads
int thread_id = omp_get_thread_num();
int total_threads = omp_get_num_threads();
// Get the appropriate forces vector
float* forces = forces_all + thread_id * (state->n * 3);
memset(forces, 0, sizeof(float) * state->n * 3);
// Get the dedupe buffer for this thread
char* usedBinID = used_bin_id_flags + (thread_id * HASH_SIZE);
// Create storage for neighbor set
unsigned buckets[MAX_NBR_BINS];
unsigned numbins;
// Process all buckets that the thread is responsible for
for (int iter_bucket = thread_id; iter_bucket < HASH_SIZE; iter_bucket += total_threads)
{
for (particle_t* pi = hash[iter_bucket]; pi != NULL ; pi = pi->next)
{
// Get the position of the pi force accumulator
unsigned diffPosI = pi - state->part;
float* pia = forces + 3 * diffPosI;
// Compute equal and opposite forces for the particle,
// first get neighbors
numbins = particle_neighborhood(buckets, pi, h, usedBinID);
for (int j = 0; j < numbins; ++j)
{
// Get the neighbor particles
unsigned bucketid = buckets[j];
for (particle_t* pj = hash[bucketid]; pj != NULL ; pj = pj->next) {
// Compute forces only if appropriate
if (pi < pj && abs(pi->ix - pj->ix) <= 1
&& abs(pi->iy - pj->iy) <= 1
&& abs(pi->iz - pj->iz) <= 1)
{
// Get the position of the pj force accumulator
unsigned diffPosJ = pj - state->part;
float* pja = forces + 3 * diffPosJ;
// Accumulate forces
update_forces(pi, pj, h2, rho0, C0, Cp, Cv, pia, pja);
}
}
}
}
}
// Accumulate values in vector
for (int iter_particle = 0; iter_particle < state->n; ++iter_particle)
{
particle_t* cur_particle = state->part + iter_particle;
float* pia = forces + 3 * iter_particle;
if(pia[0] != 0 || pia[1] != 0 || pia[2] != 0)
{
omp_set_lock(&cur_particle->lock);
vec3_saxpy(cur_particle->a, 1, pia);
omp_unset_lock(&cur_particle->lock);
}
}
}
/* END TASK */
#else
for (int i = 0; i < n; ++i) {
particle_t* pi = p+i;
for (int j = i+1; j < n; ++j) {
particle_t* pj = p+j;
update_forces(pi, pj, h2, rho0, C0, Cp, Cv, pi->a, pj->a);
}
}
#endif
}
