Jungfraujoch/rugnux/RigidBodyGPUEngine.h

// SPDX-FileCopyrightText: 2026 Filip Leonarski, Paul Scherrer Institute <filip.leonarski@psi.ch>
// SPDX-License-Identifier: GPL-3.0-only

#pragma once

// The device half of RigidBodyTargetGPU (RigidBodyGPU.h): one engine is one CUDA stream and the buffers
// for one resolution zone of one fit. The host half works out everything that depends only on the
// model, the cell and the zone - with gemmi, which stays out of nvcc - and hands it over in the plain
// structures below; the engine then does, per evaluation, what depends on the placement. CUDA builds
// only.

#include <array>
#include <cstddef>
#include <memory>
#include <stdexcept>
#include <vector>

// One atom's density on one zone's grid, as PutModelDensityOnGrid() (ModelGrid.cpp) sets it up: gemmi's
// precalculated five-Gaussian sum and the radius it cuts the sum at.
struct RigidBodyGPUAtom {
    float a[5];
    float b[5][6];      // isotropic: b[k][0] multiplies r^2; anisotropic: the matrix, u11 u22 u33 u12 u13 u23
    float occ;
    float radius;
    int aniso;
};

// One term of SymmetryComposition: F1 at k = hR, times the phase of the operator's translation.
struct RigidBodyGPUTerm {
    int index;          // of F1(k) in the half-u transform, or of F1(-k) where k is outside the stored half
    int conj;           // read as the Friedel mate, conj F1(-k)
    double phase[2];    // exp(+2 pi i h.t), real and imaginary
    double s[3];        // k as a Cartesian reciprocal vector
};

// Everything a zone needs that does not depend on the placement.
struct RigidBodyGPUZone {
    int nu = 0, nv = 0, nw = 0;           // the zone's grid, u fastest
    double orth[9] = {}, frac[9] = {};    // row-major
    double volume = 0;
    double blur = 0;                      // DensityCalculator's, which the rows' unblur undoes
    std::vector<RigidBodyGPUAtom> atoms;  // model order

    // The bulk-solvent mask's atoms: the model's index of each, and its radius (probe included).
    std::vector<int> mask_atom;
    std::vector<float> mask_radius;
    // Every image: the group's operators, each with each centring vector, fractional.
    std::vector<std::array<double, 12>> images;   // rot[9] row-major, tran[3]

    // The composition: rows (composed indices) and Ops() terms per row, row-major.
    std::vector<std::array<int, 3>> row_hkl;
    std::vector<double> row_scale;        // n_cen * unblur
    std::vector<double> row_stol2;
    size_t ops = 1;
    std::vector<RigidBodyGPUTerm> terms;

    // The observations the residuals are over: each one's row (-1 without one) and amplitude.
    std::vector<int> obs_row;
    std::vector<float> obs_fobs, obs_sigma;
    double f_mean = 1;
    // The scale's points: the observations gemmi's prepare_points() would take, in order.
    std::vector<int> point_obs;
    std::vector<std::array<int, 3>> point_hkl;
    std::vector<double> point_stol2;
    std::vector<float> point_fobs, point_sigma;
    std::vector<std::array<double, 6>> constraints;   // adp_symmetry_constraints()
};

// Upper bounds an engine is sized for.
struct RigidBodyGPUCapacity {
    size_t atoms = 0;
    size_t grid_points = 0;                // nu * nv * nw
    size_t complex_points = 0;             // (nu / 2 + 1) * nv * nw
    size_t bricks = 0;
    size_t pairs = 0;                      // (brick, atom) pairs of the gather
    size_t rows = 0, terms = 0;
    size_t observations = 0;
    size_t fft_work_bytes = 0;
};

// FitModelScale()'s solvent grid on five or fewer strong reflections, where its fits chain from one grid
// point to the next and the device does not reproduce it (ModelScaleGPU::FitSolvent).
class ModelScaleGPUTooFewReflections : public std::runtime_error {
public:
    ModelScaleGPUTooFewReflections() : std::runtime_error("ModelScaleGPU::FitSolvent: too few reflections for independent grid points") {}
};

struct RigidBodyGPUEngineImpl;

class RigidBodyGPUEngine {
public:
    // The bytes an engine of this capacity reserves on the device.
    static size_t DeviceBytes(const RigidBodyGPUCapacity &capacity);
    // The largest cuFFT work area a (nu, nv, nw) grid needs, batch 1 or 3.
    static size_t FFTWorkBytes(int nu, int nv, int nw);
    // (brick, atom) pairs of the gather over at most, for a zone's atoms on its grid.
    static size_t PairBound(const RigidBodyGPUZone &zone);
    static size_t Bricks(int nu, int nv, int nw);
    // The most gather bricks the 2 d + 1 points of an atom's box can fall in along an axis of n points.
    static size_t AxisBrickBound(int d, int n);
    // Whether the gather reproduces gemmi's box walk on this zone: every atom's box narrower than the cell,
    // so that no point is reached by two images of one atom.
    static bool Supports(const RigidBodyGPUZone &zone);
    static void MemoryInfo(size_t &free, size_t &total);
    static int CurrentDevice();

    RigidBodyGPUEngine(const RigidBodyGPUCapacity &capacity, int device);
    ~RigidBodyGPUEngine();

    // Per fit: each atom's position relative to the model centroid, which the placements rotate.
    void SetBody(const std::vector<std::array<double, 3>> &relative);
    // Per zone. Throws if the zone does not fit the capacity.
    void SetZone(const RigidBodyGPUZone &zone);

    // Per evaluation, at the placement x -> R x_rel + t: Fcalc and dF/dt at the rows, then the bulk-
    // solvent mask of the same placement (ModelMaskGPU) and Fmask.
    void Fcalc(const double rotation[9], const double translation[3]);
    void Fmask();
    // The scale at this evaluation's Fcalc and Fmask (ModelScaleGPU): FitModelScale()'s solvent grid, and
    // the overall scale and anisotropic B at a fixed solvent. FitSolvent() throws ModelScaleGPUTooFewReflections where
    // the grid's fits would chain (too few reflections for the isotropic start), which the host then does.
    void FitSolvent(double &k_sol, double &b_sol);
    void FitScale(double k_sol, double b_sol, double &k_overall, double b_star[6]);
    // The scale's points: fcmol (Fcalc as complex<float>) and fmask, downloaded for the host's fit.
    void DownloadPoints(std::vector<std::array<float, 2>> &fcmol, std::vector<std::array<float, 2>> &fmask);
    // The residuals at the scale given, NumObservations() of them, into `residuals` (host).
    void Residuals(double k_overall, const double b_star[6], double k_sol, double b_sol, double *residuals);

    // The Jacobian at the last evaluation: `rotation[j]` and `translation[j]` place the body one step
    // along rotation axis j. Writes J (fixed scale) and J_k (the scale parameters' columns) on the
    // device and returns J_k^T J_k (p x p) and J_k^T J (p x 6), p = 1 + constraints, row-major.
    void Jacobian(const double rotation[3][9], const double translation[3][3], double step, double k_overall,
                  const double b_star[6], double k_sol, double b_sol, std::vector<double> &jtj,
                  std::vector<double> &jtq);
    // J - J_k x, x p x 6 row-major, into `jacobian` (host, NumObservations() x 6).
    void ProjectJacobian(const std::vector<double> &x, double *jacobian);

private:
    std::unique_ptr<RigidBodyGPUEngineImpl> impl_;
};