guilhem_lavaux/cosmotool
0
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Tweaks for speed

Browse Source
This commit is contained in:
Guilhem Lavaux 2022-01-29 08:22:08 +01:00
parent 8ab094ad3d
commit d875416200
2 changed files with 128 additions and 135 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

163
sample/simple3DFilter.cpp
View File

@ -1,22 +1,22 @@
#include "openmp.hpp"
#include "omptl/algorithm"
#include <cassert>
#include "yorick.hpp"
#include "sphSmooth.hpp"
#include "mykdtree.hpp"
#include "miniargs.hpp"
#include <H5Cpp.h>
#include "hdf5_array.hpp"
#include <iostream>
#include <boost/format.hpp>
#include "miniargs.hpp"
#include "mykdtree.hpp"
#include "omptl/algorithm"
#include "openmp.hpp"
#include "sphSmooth.hpp"
#include "yorick.hpp"
#include <H5Cpp.h>
#include <boost/bind.hpp>
#include <boost/format.hpp>
#include <cassert>
#include <iostream>
using namespace std;
using namespace CosmoTool;
#define N_SPH 32
struct VCoord{
struct VCoord {
float v[3];
float mass;
};
@ -27,46 +27,42 @@ typedef boost::multi_array<float, 2> array_type;
typedef boost::multi_array<float, 3> array3_type;
typedef boost::multi_array<float, 4> array4_type;
ComputePrecision getVelocity(const VCoord& v, int i)
{
return v.mass * v.v[i];
}
ComputePrecision getVelocity(const VCoord &v, int i) { return v.mass * v.v[i]; }
ComputePrecision getMass(const VCoord& v)
{
return v.mass;
}
ComputePrecision getMass(const VCoord &v) { return v.mass; }
typedef SPHSmooth<VCoord> MySmooth;
typedef MySmooth::SPHTree MyTree;
typedef MyTree::Cell MyCell;
template<typename FuncT>
void computeInterpolatedField(MyTree *tree1, double boxsize, int Nres, double cx, double cy, double cz,
array3_type& bins, array3_type& arr, FuncT func, double rLimit2)
{
#pragma omp parallel
template <typename FuncT>
void computeInterpolatedField(MyTree *tree1, double boxsize, int Nres,
double cx, double cy, double cz,
array3_type &bins, array3_type &arr, FuncT func,
double rLimit2) {
int rz_max = 0;
#pragma omp parallel shared(rz_max)
{
MySmooth smooth1(tree1, N_SPH);
#pragma omp for schedule(dynamic)
for (int rz = 0; rz < Nres; rz++)
{
double pz = (rz)*boxsize/Nres-cz;
#pragma omp for collapse(3) schedule(dynamic)
for (int rz = 0; rz < Nres; rz++) {
cout << format("[%d] %d / %d") % smp_get_thread_id() % rz % Nres << endl;
for (int ry = 0; ry < Nres; ry++)
{
double py = (ry)*boxsize/Nres-cy;
for (int rx = 0; rx < Nres; rx++)
{
double px = (rx)*boxsize/Nres-cx;
for (int ry = 0; ry < Nres; ry++) {
for (int rx = 0; rx < Nres; rx++) {
if (rz > rz_max) {
rz_max = rz;
cout << format("[%d] %d / %d") % smp_get_thread_id() % rz % Nres
<< endl;
}
double px = (rx)*boxsize / Nres - cx;
double py = (ry)*boxsize / Nres - cy;
double pz = (rz)*boxsize / Nres - cz;
MyTree::coords c = { float(px), float(py), float(pz) };
MyTree::coords c = {float(px), float(py), float(pz)};
double r2 = c[0]*c[0]+c[1]*c[1]+c[2]*c[2];
if (r2 > rLimit2)
{
double r2 = c[0] * c[0] + c[1] * c[1] + c[2] * c[2];
if (r2 > rLimit2) {
arr[rx][ry][rz] = 0;
continue;
}
@ -84,23 +80,20 @@ void computeInterpolatedField(MyTree *tree1, double boxsize, int Nres, double cx
}
}
int main(int argc, char **argv)
{
int main(int argc, char **argv) {
char *fname1, *outFile;
double rLimit, boxsize, rLimit2, cx, cy, cz;
int Nres;
MiniArgDesc args[] = {
{ "INPUT DATA1", &fname1, MINIARG_STRING },
{ "RADIUS LIMIT", &rLimit, MINIARG_DOUBLE },
{ "BOXSIZE", &boxsize, MINIARG_DOUBLE },
{ "RESOLUTION", &Nres, MINIARG_INT },
{ "CX", &cx, MINIARG_DOUBLE },
{ "CY", &cy, MINIARG_DOUBLE },
{ "CZ", &cz, MINIARG_DOUBLE },
{ "OUTPUT FILE", &outFile, MINIARG_STRING },
{ 0, 0, MINIARG_NULL }
};
MiniArgDesc args[] = {{"INPUT DATA1", &fname1, MINIARG_STRING},
{"RADIUS LIMIT", &rLimit, MINIARG_DOUBLE},
{"BOXSIZE", &boxsize, MINIARG_DOUBLE},
{"RESOLUTION", &Nres, MINIARG_INT},
{"CX", &cx, MINIARG_DOUBLE},
{"CY", &cy, MINIARG_DOUBLE},
{"CZ", &cz, MINIARG_DOUBLE},
{"OUTPUT FILE", &outFile, MINIARG_STRING},
{0, 0, MINIARG_NULL}};
if (!parseMiniArgs(argc, argv, args))
return 1;
@ -112,7 +105,7 @@ int main(int argc, char **argv)
array3_type bins(boost::extents[Nres][Nres][Nres]);
rLimit2 = rLimit*rLimit;
rLimit2 = rLimit * rLimit;
hdf5_read_array(in_f, "particles", v1_data);
assert(v1_data.shape()[1] == 7);
@ -124,30 +117,30 @@ int main(int argc, char **argv)
MyCell *allCells_1 = new MyCell[N1_points];
#pragma omp parallel for schedule(static)
for (uint32_t i = 0; i < Nres*Nres*Nres; i++)
for (uint32_t i = 0; i < Nres * Nres * Nres; i++)
bins.data()[i] = 0;
cout << "Shuffling data in cells..." << endl;
#pragma omp parallel for schedule(static)
for (uint64_t i = 0 ; i < N1_points; i++)
{
for (uint64_t i = 0; i < N1_points; i++) {
for (int j = 0; j < 3; j++)
allCells_1[i].coord[j] = v1_data[i][j];
for (int k = 0; k < 3; k++)
allCells_1[i].val.pValue.v[k] = v1_data[i][3+k];
allCells_1[i].val.pValue.v[k] = v1_data[i][3 + k];
allCells_1[i].val.pValue.mass = v1_data[i][6];
allCells_1[i].active = true;
allCells_1[i].val.weight = 0.0;
long rx = floor((allCells_1[i].coord[0]+cx)*Nres/boxsize+0.5);
long ry = floor((allCells_1[i].coord[1]+cy)*Nres/boxsize+0.5);
long rz = floor((allCells_1[i].coord[2]+cz)*Nres/boxsize+0.5);
long rx = floor((allCells_1[i].coord[0] + cx) * Nres / boxsize + 0.5);
long ry = floor((allCells_1[i].coord[1] + cy) * Nres / boxsize + 0.5);
long rz = floor((allCells_1[i].coord[2] + cz) * Nres / boxsize + 0.5);
if (rx < 0 || rx >= Nres || ry < 0 || ry >= Nres || rz < 0 || rz >= Nres)
continue;
//#pragma omp atomic update
bins[rx][ry][rz]++;
auto &b = bins[rx][ry][rz];
#pragma omp atomic
b++;
}
v1_data.resize(boost::extents[1][1]);
@ -158,32 +151,31 @@ int main(int argc, char **argv)
cout << "Creating smoothing filter..." << endl;
// array3_type out_rad_1(boost::extents[Nres][Nres][Nres]);
// array3_type out_rad_1(boost::extents[Nres][Nres][Nres]);
cout << "Weighing..." << endl;
#pragma omp parallel
int rz_max = 0;
#pragma omp parallel shared(rz_max)
{
MySmooth smooth1(&tree1, N_SPH);
#pragma omp for schedule(dynamic)
for (int rz = 0; rz < Nres; rz++)
{
double pz = (rz)*boxsize/Nres-cz;
#pragma omp for collapse(3) schedule(dynamic, 8)
for (int rz = 0; rz < Nres; rz++) {
for (int ry = 0; ry < Nres; ry++) {
for (int rx = 0; rx < Nres; rx++) {
if (rz > rz_max) {
rz_max = rz;
(cout << rz << " / " << Nres << endl).flush();
for (int ry = 0; ry < Nres; ry++)
{
double py = (ry)*boxsize/Nres-cy;
for (int rx = 0; rx < Nres; rx++)
{
double px = (rx)*boxsize/Nres-cx;
}
double pz = (rz)*boxsize / Nres - cz;
double py = (ry)*boxsize / Nres - cy;
double px = (rx)*boxsize / Nres - cx;
MyTree::coords c = { float(px), float(py), float(pz) };
MyTree::coords c = {float(px), float(py), float(pz)};
double r2 = c[0]*c[0]+c[1]*c[1]+c[2]*c[2];
if (r2 > rLimit2)
{
double r2 = c[0] * c[0] + c[1] * c[1] + c[2] * c[2];
if (r2 > rLimit2) {
continue;
}
@ -192,11 +184,9 @@ int main(int argc, char **argv)
smooth1.fetchNeighbours(c, numInCell);
else
smooth1.fetchNeighbours(c);
#pragma omp critical
smooth1.addGridSite(c);
}
}
(cout << " Done " << rz << endl).flush();
}
}
@ -204,13 +194,14 @@ int main(int argc, char **argv)
array3_type interpolated(boost::extents[Nres][Nres][Nres]);
computeInterpolatedField(&tree1, boxsize, Nres, cx, cy, cz,
bins, interpolated, getMass, rLimit2);
computeInterpolatedField(&tree1, boxsize, Nres, cx, cy, cz, bins,
interpolated, getMass, rLimit2);
hdf5_write_array(out_f, "density", interpolated);
//out_f.flush();
// out_f.flush();
for (int i = 0; i < 3; i++) {
computeInterpolatedField(&tree1, boxsize, Nres, cx, cy, cz,
bins, interpolated, boost::bind(getVelocity, _1, i), rLimit2);
computeInterpolatedField(&tree1, boxsize, Nres, cx, cy, cz, bins,
interpolated, boost::bind(getVelocity, _1, i),
rLimit2);
hdf5_write_array(out_f, str(format("p%d") % i), interpolated);
}

4
src/sphSmooth.tcc
View File

@ -192,7 +192,9 @@ void SPHSmooth<ValType,Ndims>::addGridSite(const typename SPHTree::coords& c)
{
ComputePrecision d = internal.distances[i];
SPHCell& cell = *(internal.ngb[i]);
cell.val.weight += getKernel(d/internal.smoothRadius) / r3;
double kernel_value = getKernel(d/internal.smoothRadius) / r3;
#pragma omp atomic
cell.val.weight += kernel_value;
}
}