TEST(generateExpression, algebra_solver) {
static const bool user_facing = true;
std::stringstream msgs;
stan::lang::algebra_solver so; // null ctor should work and not raise error
std::string system_function_name = "bronzino";
stan::lang::variable y("y_var_name");
y.set_type(stan::lang::vector_type());
stan::lang::variable theta("theta_var_name");
theta.set_type(stan::lang::vector_type());
stan::lang::variable x_r("x_r_r_var_name");
x_r.set_type(stan::lang::bare_array_type(stan::lang::double_type()));
stan::lang::variable x_i("x_i_var_name");
x_i.set_type(stan::lang::bare_array_type(stan::lang::int_type()));
stan::lang::algebra_solver so2(system_function_name, y, theta, x_r, x_i);
stan::lang::expression e1 = so2;
generate_expression(e1, user_facing, msgs);
EXPECT_EQ(msgs.str(),
"algebra_solver(bronzino_functor__(), y_var_name, "
"theta_var_name, x_r_r_var_name, x_i_var_name, pstream__)");
}
double EigenSolverPoissonImageEditing::solve(const NamedParameters& solverParameters, const NamedParameters& problemParameters, bool profileSolve, std::vector<SolverIteration>& iters)
{
int numUnknowns = 0;
std::unordered_map<vec2i, int, vec2iHash> pixelLocationsToIndex;
std::vector<vec2i> pixelLocations;
size_t pixelCount = m_dims[0] * m_dims[1];
std::vector<float4> h_unknownFloat(pixelCount);
std::vector<float4> h_target(pixelCount);
std::vector<float> h_mask(pixelCount);
findAndCopyArrayToCPU("X", h_unknownFloat, problemParameters);
findAndCopyArrayToCPU("T", h_target, problemParameters);
findAndCopyArrayToCPU("M", h_mask, problemParameters);
for (int y = 0; y < (int)m_dims[1]; ++y) {
for (int x = 0; x < (int)m_dims[0]; ++x) {
if (h_mask[y*m_dims[0] + x] == 0.0f) {
++numUnknowns;
vec2i p(x, y);
pixelLocationsToIndex[p] =(int)pixelLocations.size();
pixelLocations.push_back(p);
}
}
}
printf("# Unknowns: %d\n", numUnknowns);
int numResiduals = (int)pixelLocations.size() * 4;
Eigen::VectorXf x_r(numUnknowns), b_r(numResiduals);
Eigen::VectorXf x_g(numUnknowns), b_g(numResiduals);
Eigen::VectorXf x_b(numUnknowns), b_b(numResiduals);
Eigen::VectorXf x_a(numUnknowns), b_a(numResiduals);
b_r.setZero();
b_g.setZero();
b_b.setZero();
b_a.setZero();
for (int i = 0; i < pixelLocations.size(); ++i) {
vec2i p = pixelLocations[i];
float4 color = sampleImage(h_unknownFloat.data(), p, m_dims[0]);
x_r[i] = color.x;
//printf("%f\n", color.x);
x_g[i] = color.y;
x_b[i] = color.z;
x_a[i] = color.w;
}
SpMatrixf A(numResiduals, numUnknowns);
A.setZero();
printf("Constructing Matrix\n");
std::vector<Tripf> entriesA;
std::vector<vec2i> offsets;
offsets.push_back(vec2i(-1, 0));
offsets.push_back(vec2i(1, 0));
offsets.push_back(vec2i(0, -1));
offsets.push_back(vec2i(0, 1));
for (int i = 0; i < pixelLocations.size(); ++i) {
vec2i p = pixelLocations[i];
int numInternalNeighbors = 0;
float4 g_p = sampleImage(h_target.data(), p, m_dims[0]);
int j = 0;
for (vec2i off : offsets) {
vec2i q = p + off;
if (q.x >= 0 && q.y >= 0 && q.x < (int)m_dims[0] && q.y < (int)m_dims[1]) {
auto it = pixelLocationsToIndex.find(q);
int row = 4 * i + j;
if (it == pixelLocationsToIndex.end()) {
float4 f_q = sampleImage(h_unknownFloat.data(), q, m_dims[0]);
b_r[row] += f_q.x;
b_g[row] += f_q.y;
b_b[row] += f_q.z;
b_a[row] += f_q.w;
}
else {
entriesA.push_back(Tripf(row, it->second, -1.0f));
}
entriesA.push_back(Tripf(row, i, 1.0f));
float4 g_q = sampleImage(h_target.data(), q, m_dims[0]);
b_r[row] += (g_p.x - g_q.x);
b_g[row] += (g_p.y - g_q.y);
b_b[row] += (g_p.z - g_q.z);
b_a[row] += (g_p.w - g_q.w);
}
++j;
}
}
printf("Entries Set\n");
A.setFromTriplets(entriesA.begin(), entriesA.end());
printf("Sparse Matrix Constructed\n");
A.makeCompressed();
printf("Matrix Compressed\n");
{
float totalCost = 0.0f;
float cost_r = (A*x_r - b_r).squaredNorm();
float cost_g = (A*x_g - b_g).squaredNorm();
float cost_b = (A*x_b - b_b).squaredNorm();
float cost_a = (A*x_a - b_a).squaredNorm();
totalCost = cost_r + cost_g + cost_b + cost_a;
printf("Initial Cost: %f : (%f, %f, %f, %f)\n", totalCost, cost_r, cost_g, cost_b, cost_a);
}
AxEqBSolver solver;
solver.setMaxIterations(97);
printf("Solvers Initialized\n");
clock_t start = clock(), diff;
solver.compute(A);
//printf("solver.compute(A)\n");
solveAxEqb(solver, b_r, x_r);
//printf("Red solve done\n");
solveAxEqb(solver, b_g, x_g);
//printf("Green solve done\n");
solveAxEqb(solver, b_b, x_b);
//printf("Blue solve done\n");
solveAxEqb(solver, b_a, x_a);
diff = clock() - start;
printf("Time taken %f ms\n", diff*1000.0 / double(CLOCKS_PER_SEC));
float totalCost = 0.0f;
float cost_r = (A*x_r - b_r).squaredNorm();
float cost_g = (A*x_g - b_g).squaredNorm();
float cost_b = (A*x_b - b_b).squaredNorm();
float cost_a = (A*x_a - b_a).squaredNorm();
totalCost = cost_r + cost_g + cost_b + cost_a;
printf("Final Cost: %f : (%f, %f, %f, %f)\n", totalCost, cost_r, cost_g, cost_b, cost_a);
for (int i = 0; i < pixelLocations.size(); ++i) {
setPixel(h_unknownFloat.data(), pixelLocations[i], m_dims[0], x_r[i], x_g[i], x_b[i]);
}
findAndCopyToArrayFromCPU("X", h_unknownFloat, problemParameters);;
return (double)totalCost;
}
