PrusaSlicer/src/libslic3r/SLA/Rotfinder.cpp
tamasmeszaros c193d7c930 Brute force optimization code, buggy yet
wip


wip


wip refactor
2020-09-10 14:03:30 +02:00

181 lines
6.1 KiB
C++

#include <limits>
#include <exception>
//#include <libnest2d/optimizers/nlopt/genetic.hpp>
#include <libslic3r/Optimize/BruteforceOptimizer.hpp>
#include <libslic3r/SLA/Rotfinder.hpp>
#include <libslic3r/SLA/Concurrency.hpp>
#include <libslic3r/SLA/SupportTree.hpp>
#include <libslic3r/SLA/SupportPointGenerator.hpp>
#include <libslic3r/SimplifyMesh.hpp>
#include "Model.hpp"
#include <thread>
namespace Slic3r {
namespace sla {
using VertexFaceMap = std::vector<std::vector<size_t>>;
VertexFaceMap create_vertex_face_map(const TriangleMesh &mesh) {
std::vector<std::vector<size_t>> vmap(mesh.its.vertices.size());
size_t fi = 0;
for (const Vec3i &tri : mesh.its.indices) {
for (int vi = 0; vi < tri.size(); ++vi) {
auto from = vmap[tri(vi)].begin(), to = vmap[tri(vi)].end();
vmap[tri(vi)].insert(std::lower_bound(from, to, fi), fi);
}
}
return vmap;
}
// Find transformed mesh ground level without copy and with parallell reduce.
double find_ground_level(const TriangleMesh &mesh,
const Transform3d & tr,
size_t threads)
{
size_t vsize = mesh.its.vertices.size();
auto minfn = [](double a, double b) { return std::min(a, b); };
auto findminz = [&mesh, &tr] (size_t vi, double submin) {
Vec3d v = tr * mesh.its.vertices[vi].template cast<double>();
return std::min(submin, v.z());
};
double zmin = mesh.its.vertices.front().z();
return ccr_par::reduce(size_t(0), vsize, zmin, findminz, minfn,
vsize / threads);
}
// Try to guess the number of support points needed to support a mesh
double calculate_model_supportedness(const TriangleMesh & mesh,
// const VertexFaceMap &vmap,
const Transform3d & tr)
{
static constexpr double POINTS_PER_UNIT_AREA = 1.;
if (mesh.its.vertices.empty()) return std::nan("");
size_t Nthr = std::thread::hardware_concurrency();
size_t facesize = mesh.its.indices.size();
double zmin = find_ground_level(mesh, tr, Nthr);
auto score_mergefn = [&mesh, &tr, zmin](size_t fi, double subscore) {
static const Vec3d DOWN = {0., 0., -1.};
const auto &face = mesh.its.indices[fi];
Vec3d p1 = tr * mesh.its.vertices[face(0)].template cast<double>();
Vec3d p2 = tr * mesh.its.vertices[face(1)].template cast<double>();
Vec3d p3 = tr * mesh.its.vertices[face(2)].template cast<double>();
Vec3d U = p2 - p1;
Vec3d V = p3 - p1;
Vec3d C = U.cross(V);
Vec3d N = C.normalized();
double area = 0.5 * C.norm();
double zlvl = zmin + 0.1;
if (p1.z() <= zlvl && p2.z() <= zlvl && p3.z() <= zlvl) {
// score += area * POINTS_PER_UNIT_AREA;
return subscore;
}
double phi = 1. - std::acos(N.dot(DOWN)) / PI;
phi = phi * (phi > 0.5);
// std::cout << "area: " << area << std::endl;
subscore += area * POINTS_PER_UNIT_AREA * phi;
return subscore;
};
double score = ccr_seq::reduce(size_t(0), facesize, 0., score_mergefn,
std::plus<double>{}, facesize / Nthr);
return score / mesh.its.indices.size();
}
std::array<double, 2> find_best_rotation(const ModelObject& modelobj,
float accuracy,
std::function<void(unsigned)> statuscb,
std::function<bool()> stopcond)
{
static const unsigned MAX_TRIES = 100;
// return value
std::array<double, 2> rot;
// We will use only one instance of this converted mesh to examine different
// rotations
TriangleMesh mesh = modelobj.raw_mesh();
mesh.require_shared_vertices();
// auto vmap = create_vertex_face_map(mesh);
// simplify_mesh(mesh);
// For current iteration number
unsigned status = 0;
// The maximum number of iterations
auto max_tries = unsigned(accuracy * MAX_TRIES);
// call status callback with zero, because we are at the start
statuscb(status);
// So this is the object function which is called by the solver many times
// It has to yield a single value representing the current score. We will
// call the status callback in each iteration but the actual value may be
// the same for subsequent iterations (status goes from 0 to 100 but
// iterations can be many more)
auto objfunc = [&mesh, &status, &statuscb, &stopcond, max_tries]
(const opt::Input<2> &in)
{
std::cout << "in: " << in[0] << " " << in[1] << std::endl;
// prepare the rotation transformation
Transform3d rt = Transform3d::Identity();
rt.rotate(Eigen::AngleAxisd(in[1], Vec3d::UnitY()));
rt.rotate(Eigen::AngleAxisd(in[0], Vec3d::UnitX()));
double score = sla::calculate_model_supportedness(mesh, rt);
std::cout << score << std::endl;
// report status
if(!stopcond()) statuscb( unsigned(++status * 100.0/max_tries) );
return score;
};
// Firing up the genetic optimizer. For now it uses the nlopt library.
opt::Optimizer<opt::AlgBruteForce> solver(opt::StopCriteria{}
.max_iterations(max_tries)
.rel_score_diff(1e-6)
.stop_condition(stopcond),
10 /*grid size*/);
// We are searching rotations around the three axes x, y, z. Thus the
// problem becomes a 3 dimensional optimization task.
// We can specify the bounds for a dimension in the following way:
auto b = opt::Bound{-PI, PI};
// Now we start the optimization process with initial angles (0, 0, 0)
auto result = solver.to_min().optimize(objfunc, opt::initvals({0.0, 0.0}),
opt::bounds({b, b}));
// Save the result and fck off
rot[0] = std::get<0>(result.optimum);
rot[1] = std::get<1>(result.optimum);
return rot;
}
}
}