Jungfraujoch/image_analysis/structure_refinement/ModelMaskGPU.cu

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

#include "ModelMaskGPU.h"

#include <cmath>
#include <cstdint>
#include <string>

#include "../../common/JFJochException.h"

namespace {

void cuda_err(cudaError_t val) {
    if (val != cudaSuccess)
        throw JFJochException(JFJochExceptionCategory::GPUCUDAError, cudaGetErrorString(val));
}

constexpr int THREADS = 256;
constexpr int IMAGE_THREADS = 64;

// gemmi SolventMasker(AtomicRadiiSet::Refmac): rshrink and island_min_volume.
constexpr double R_SHRINK = 0.8;
constexpr double ISLAND_MIN_VOLUME = 50.0;

__device__ int wrap(int a, int n) {
    const int r = a % n;
    return r < 0 ? r + n : r;
}

__global__ void fill_solvent(float *mask, size_t n) {
    const size_t i = blockIdx.x * static_cast<size_t>(blockDim.x) + threadIdx.x;
    if (i < n)
        mask[i] = 1.0f;
}

// One block per image (atom, operator): gemmi's use_points_around<true>(fpos, radius) around the image,
// setting to 0 every point of the box with !(dist_sq > radius^2). The box is sized from the interplanar
// spacing, clamped to n - 1 and centred on the rounded grid position, as gemmi does it.
__global__ void mask_images(const ModelMaskAtom *atoms, const ModelMaskOp *ops, int n_ops,
                            ModelMaskGPU::Geometry g, float *mask) {
    const ModelMaskAtom a = atoms[blockIdx.x / n_ops];
    // The image's box, one axis per thread: double arithmetic is slow on the GPU, so it is done once
    // per image rather than once per point.
    __shared__ int lo[3], len[3];
    __shared__ float c[3];   // the image's grid position less the box's first point
    if (threadIdx.x < 3) {
        const int k = threadIdx.x;
        const ModelMaskOp &op = ops[blockIdx.x % n_ops];
        const int n = k == 0 ? g.nu : k == 1 ? g.nv : g.nw;
        const double x = op.rot[3 * k] * a.x + op.rot[3 * k + 1] * a.y + op.rot[3 * k + 2] * a.z + op.tran[k];
        const double nx = x * n;
        const int centre = static_cast<int>(round(nx));
        const int d = min(static_cast<int>(ceil(a.radius / g.spacing[k])), n - 1);
        lo[k] = centre - d;
        len[k] = 2 * d + 1;
        c[k] = static_cast<float>(nx - centre) + static_cast<float>(d);
    }
    __syncthreads();
    const float r2 = static_cast<float>(a.radius * a.radius);
    const int total = len[0] * len[1] * len[2];
    for (int t = threadIdx.x; t < total; t += blockDim.x) {
        const int iu = t % len[0], iv = (t / len[0]) % len[1], iw = t / (len[0] * len[1]);
        const float fu = c[0] - iu, fv = c[1] - iv, fw = c[2] - iw;
        const float x = g.orth_n[0] * fu + g.orth_n[1] * fv + g.orth_n[2] * fw;
        const float y = g.orth_n[3] * fu + g.orth_n[4] * fv + g.orth_n[5] * fw;
        const float z = g.orth_n[6] * fu + g.orth_n[7] * fv + g.orth_n[8] * fw;
        if (x * x + y * y + z * z > r2)
            continue;
        const size_t idx = wrap(lo[0] + iu, g.nu)
                           + static_cast<size_t>(g.nu) * (wrap(lo[1] + iv, g.nv)
                                                          + static_cast<size_t>(g.nv) * wrap(lo[2] + iw, g.nw));
        mask[idx] = 0.0f;
    }
}

// Union-find labelling of the solvent after Playne & Hawick (2018) IEEE TPDS 29, 1217-1230: every
// solvent point starts as its own set, neighbouring points are united with atomicCAS, always linking the
// larger root under the smaller, so each component ends up rooted at its lowest index whatever the
// order of the unions.

// Path halving: a non-root entry is pointed at its grandparent. Entries only ever point to a lower
// index of the same set, so the racing writes of other threads are harmless and a root is changed
// only by the atomicCAS in unite().
__device__ int find_root(int *label, int i) {
    int p = label[i];
    while (p != i) {
        const int gp = label[p];
        if (gp != p)
            label[i] = gp;
        i = p;
        p = gp;
    }
    return i;
}

__device__ void unite(int *label, int a, int b) {
    while (true) {
        a = find_root(label, a);
        b = find_root(label, b);
        if (a == b)
            return;
        if (a < b) {
            const int t = a;
            a = b;
            b = t;
        }
        const int old = atomicCAS(&label[a], a, b);
        if (old == a)
            return;
        a = old;
    }
}

__global__ void uf_init(const float *mask, int *label, size_t n) {
    const size_t i = blockIdx.x * static_cast<size_t>(blockDim.x) + threadIdx.x;
    if (i < n)
        label[i] = mask[i] == 1.0f ? static_cast<int>(i) : -1;
}

// gemmi's FloodFill joins a run of points to the runs of the rows v +- 1, w +- 1 over u - 1 .. u + len,
// with wrap-around: 26-connectivity on the periodic grid. Each point unites with the 13 neighbours
// ahead of it; the other 13 unite with it from their side.
__global__ void uf_merge(const float *mask, int *label, int nu, int nv, int nw) {
    const size_t n = static_cast<size_t>(nu) * nv * nw;
    const size_t i = blockIdx.x * static_cast<size_t>(blockDim.x) + threadIdx.x;
    if (i >= n || mask[i] != 1.0f)
        return;
    const int u = static_cast<int>(i % nu);
    const int v = static_cast<int>((i / nu) % nv);
    const int w = static_cast<int>(i / (static_cast<size_t>(nu) * nv));
    const int off[13][3] = {{1, 0, 0}, {-1, 1, 0}, {0, 1, 0}, {1, 1, 0},
                            {-1, -1, 1}, {0, -1, 1}, {1, -1, 1}, {-1, 0, 1}, {0, 0, 1},
                            {1, 0, 1}, {-1, 1, 1}, {0, 1, 1}, {1, 1, 1}};
    for (int k = 0; k < 13; k++) {
        const size_t j = wrap(u + off[k][0], nu)
                         + static_cast<size_t>(nu) * (wrap(v + off[k][1], nv)
                                                      + static_cast<size_t>(nv) * wrap(w + off[k][2], nw));
        if (mask[j] == 1.0f)
            unite(label, static_cast<int>(i), static_cast<int>(j));
    }
}

// The roots, read-only and into a separate array: writing them over the labels while other threads
// still walk the labels would let a path-halving write replace a written root with a mere ancestor.
__global__ void uf_roots(const int *label, int *root, size_t n) {
    const size_t i = blockIdx.x * static_cast<size_t>(blockDim.x) + threadIdx.x;
    if (i >= n)
        return;
    int r = label[i];
    if (r >= 0)
        while (label[r] != r)
            r = label[r];
    root[i] = r;
}

// Component sizes, counted only while they are at most limit: every island is counted in full, and the
// one large component stops being counted almost at once instead of queueing millions of atomics on one
// address. Only size <= limit is ever read, which the capping does not change.
__global__ void uf_count(const int *root, int *size, int limit, size_t n) {
    const size_t i = blockIdx.x * static_cast<size_t>(blockDim.x) + threadIdx.x;
    if (i >= n)
        return;
    const int r = root[i];
    if (r >= 0 && reinterpret_cast<volatile int *>(size)[r] <= limit)
        atomicAdd(&size[r], 1);
}

__global__ void uf_remove(float *mask, const int *root, const int *size, int limit, size_t n) {
    const size_t i = blockIdx.x * static_cast<size_t>(blockDim.x) + threadIdx.x;
    if (i < n && root[i] >= 0 && size[root[i]] <= limit)
        mask[i] = 0.0f;
}

unsigned int Blocks(size_t n) {
    return static_cast<unsigned int>((n + THREADS - 1) / THREADS);
}

}  // namespace

size_t ModelMaskGPU::DeviceBytes(size_t max_points) {
    return 2 * max_points * sizeof(int) + MAX_OPS * sizeof(ModelMaskOp);
}

ModelMaskGPU::ModelMaskGPU(cudaStream_t stream, size_t max_points)
        : stream(stream), max_points(max_points), ops_d(MAX_OPS), label(max_points), root(max_points) {
    if (max_points > INT32_MAX)
        throw JFJochException(JFJochExceptionCategory::InputParameterInvalid,
                              "Solvent mask grid too large for int32 labels");
}

void ModelMaskGPU::SetGrid(const ModelMaskGrid &grid, const std::vector<ModelMaskOp> &ops) {
    npoints = static_cast<size_t>(grid.nu) * grid.nv * grid.nw;
    if (npoints > max_points)
        throw JFJochException(JFJochExceptionCategory::InputParameterInvalid,
                              "Solvent mask grid larger than the buffers");
    if (ops.empty() || ops.size() > MAX_OPS)
        throw JFJochException(JFJochExceptionCategory::InputParameterInvalid,
                              "Solvent mask needs 1 to " + std::to_string(MAX_OPS) + " operators");

    const int n[3] = {grid.nu, grid.nv, grid.nw};
    geom.nu = grid.nu;
    geom.nv = grid.nv;
    geom.nw = grid.nw;
    for (int i = 0; i < 3; i++)
        for (int j = 0; j < 3; j++)
            geom.orth_n[3 * i + j] = static_cast<float>(grid.orth[3 * i + j] / n[j]);

    // gemmi Grid::calculate_spacing(): 1 / (n |a*|), where a* is row k of orth's inverse, which is the
    // cross product of the other two columns over the determinant (= volume).
    const double *o = grid.orth;
    const double col[3][3] = {{o[0], o[3], o[6]}, {o[1], o[4], o[7]}, {o[2], o[5], o[8]}};
    for (int k = 0; k < 3; k++) {
        const double *p = col[(k + 1) % 3], *q = col[(k + 2) % 3];
        const double cx = p[1] * q[2] - p[2] * q[1], cy = p[2] * q[0] - p[0] * q[2], cz = p[0] * q[1] - p[1] * q[0];
        geom.spacing[k] = grid.volume / (n[k] * std::sqrt(cx * cx + cy * cy + cz * cz));
    }

    // gemmi's shrink, set_margin_around(r_shrink), only ever looks at the grid offsets other than 0
    // within r_shrink, in the box of floor(r_shrink / spacing). No such offset, no change.
    const int d[3] = {static_cast<int>(std::floor(R_SHRINK / geom.spacing[0])),
                      static_cast<int>(std::floor(R_SHRINK / geom.spacing[1])),
                      static_cast<int>(std::floor(R_SHRINK / geom.spacing[2]))};
    for (int w = -d[2]; w <= d[2]; w++)
        for (int v = -d[1]; v <= d[1]; v++)
            for (int u = -d[0]; u <= d[0]; u++) {
                const double f[3] = {static_cast<double>(u) / n[0], static_cast<double>(v) / n[1],
                                     static_cast<double>(w) / n[2]};
                double r2 = 0;
                for (int i = 0; i < 3; i++) {
                    const double x = o[3 * i] * f[0] + o[3 * i + 1] * f[1] + o[3 * i + 2] * f[2];
                    r2 += x * x;
                }
                if (r2 <= R_SHRINK * R_SHRINK && r2 != 0.0)
                    throw JFJochException(JFJochExceptionCategory::InputParameterInvalid,
                                          "Solvent mask grid too fine: the GPU mask has no shrink step");
            }

    // gemmi SolventMasker::remove_islands(), rounded the same way.
    island_limit = static_cast<int>(static_cast<size_t>(ISLAND_MIN_VOLUME * npoints / grid.volume));

    n_ops = static_cast<int>(ops.size());
    cuda_err(cudaMemcpyAsync(ops_d, ops.data(), ops.size() * sizeof(ModelMaskOp), cudaMemcpyHostToDevice, stream));
}

void ModelMaskGPU::Compute(const ModelMaskAtom *d_atoms, int n_atoms, float *d_mask) {
    fill_solvent<<<Blocks(npoints), THREADS, 0, stream>>>(d_mask, npoints);
    cuda_err(cudaGetLastError());
    if (n_atoms > 0) {
        mask_images<<<static_cast<unsigned int>(n_atoms) * n_ops, IMAGE_THREADS, 0, stream>>>(
                d_atoms, ops_d, n_ops, geom, d_mask);
        cuda_err(cudaGetLastError());
    }
    RemoveIslands(d_mask);
}

// gemmi SolventMasker::remove_islands(): every component of solvent with at most island_limit points
// becomes macromolecule.
void ModelMaskGPU::RemoveIslands(float *d_mask) {
    uf_init<<<Blocks(npoints), THREADS, 0, stream>>>(d_mask, label, npoints);
    cuda_err(cudaGetLastError());
    uf_merge<<<Blocks(npoints), THREADS, 0, stream>>>(d_mask, label, geom.nu, geom.nv, geom.nw);
    cuda_err(cudaGetLastError());
    uf_roots<<<Blocks(npoints), THREADS, 0, stream>>>(label, root, npoints);
    cuda_err(cudaGetLastError());
    int *size = label;   // the parents are no longer needed
    cuda_err(cudaMemsetAsync(size, 0, npoints * sizeof(int), stream));
    uf_count<<<Blocks(npoints), THREADS, 0, stream>>>(root, size, island_limit, npoints);
    cuda_err(cudaGetLastError());
    uf_remove<<<Blocks(npoints), THREADS, 0, stream>>>(d_mask, root, size, island_limit, npoints);
    cuda_err(cudaGetLastError());
}