argument/gpu_sdf.py

"""GPU voxel -> SDF pipeline for run_strand2surface (CuPy / CUDA).

Mirrors the CPU ``rasterise -> interior-fill -> EDT-SDF -> Gaussian-smooth``
path, but runs the dense regular-grid work on the GPU. Returns a host (NumPy)
SDF so marching cubes + remeshing continue on the CPU unchanged.

These stages are the part that scales with voxel count (the two Euclidean
distance transforms dominate as ``radius`` shrinks), and they map almost 1:1
onto ``cupyx.scipy.ndimage`` -- which is why CuPy beats hand-written kernels
here. The isotropic remesh is *not* here: it is serial topological editing and
does not belong on the GPU.
"""

from __future__ import annotations

import numpy as np


def cupy_available() -> bool:
    """True if CuPy imports and a CUDA device is visible."""
    try:
        import cupy as cp

        return cp.cuda.runtime.getDeviceCount() > 0
    except Exception:  # noqa: BLE001 - any failure means "no usable GPU"
        return False


def _rasterise_gpu(cp, strands, bbox_min, voxel_size, grid_dims):
    """Boolean occupancy grid from strand polylines, on the GPU.

    Same bucket-by-step-count DDA as the CPU path: every voxel along every edge
    is visited so no voxel is skipped. Setting True is idempotent, so the
    scattered writes need no atomics.
    """
    gx, gy, gz = int(grid_dims[0]), int(grid_dims[1]), int(grid_dims[2])
    occ = cp.zeros((gx, gy, gz), dtype=cp.bool_)
    inv_vs = 1.0 / voxel_size
    bmin = cp.asarray(np.asarray(bbox_min, dtype=np.float64))

    def splat(pts):  # pts: (M, 3) float64 on device
        vi = cp.floor((pts - bmin) * inv_vs).astype(cp.int32)
        vx = cp.clip(vi[:, 0], 0, gx - 1)
        vy = cp.clip(vi[:, 1], 0, gy - 1)
        vz = cp.clip(vi[:, 2], 0, gz - 1)
        occ[vx, vy, vz] = True

    singles = [s for s in strands if len(s) == 1]
    if singles:
        splat(cp.asarray(np.concatenate(singles, axis=0), dtype=cp.float64))

    p0_list = [s[:-1] for s in strands if len(s) >= 2]
    p1_list = [s[1:] for s in strands if len(s) >= 2]
    if not p0_list:
        return occ

    p0 = cp.asarray(np.concatenate(p0_list, axis=0), dtype=cp.float64)
    p1 = cp.asarray(np.concatenate(p1_list, axis=0), dtype=cp.float64)
    diffs = p1 - p0
    edge_len_vox = cp.abs(diffs * inv_vs).max(axis=1)
    n_steps = cp.maximum(cp.ceil(edge_len_vox).astype(cp.int64), 1)
    max_ns = int(n_steps.max())
    chunk_pts = 20_000_000

    for ns in range(1, max_ns + 1):
        mask = n_steps == ns
        count = int(mask.sum())
        if count == 0:
            continue
        sub_p0 = p0[mask]
        sub_diffs = diffs[mask]
        pts_per_edge = ns + 1
        edges_per_chunk = max(chunk_pts // pts_per_edge, 1)
        t = cp.linspace(0.0, 1.0, pts_per_edge, dtype=cp.float64)
        for i in range(0, count, edges_per_chunk):
            cp0 = sub_p0[i:i + edges_per_chunk]
            cd = sub_diffs[i:i + edges_per_chunk]
            pts = cp0[:, None, :] + t[None, :, None] * cd[:, None, :]
            splat(pts.reshape(-1, 3))
    return occ


def _fill_interior_gpu(cp, cndi, struct, occ, closing_radius, eff_radius_vox):
    """Morphological close + flood-fill interior chambers, on the GPU."""
    if closing_radius > 0:
        # cupyx only implements iterated morphology with brute_force=True.
        occ = cndi.binary_closing(
            occ, structure=struct, iterations=closing_radius, brute_force=True,
        )
    edt = cndi.distance_transform_edt(~occ)
    solid = edt <= eff_radius_vox
    empty = ~solid
    labelled, n_labels = cndi.label(empty, structure=struct)
    if n_labels <= 1:
        return occ

    boundary: set[int] = set()
    for ax in range(3):
        for s in (0, solid.shape[ax] - 1):
            sl = [slice(None)] * 3
            sl[ax] = s
            boundary.update(cp.asnumpy(cp.unique(labelled[tuple(sl)])).tolist())
    boundary.discard(0)

    interior_labels = sorted(set(range(1, n_labels + 1)) - boundary)
    if not interior_labels:
        return occ
    interior_mask = cp.isin(labelled, cp.asarray(interior_labels))
    return occ | interior_mask


def occupancy_to_sdf(
    strands,
    bbox_min,
    voxel_size,
    grid_dims,
    eff_radius_vox,
    closing_radius,
    smooth_iterations,
    radius,
) -> np.ndarray:
    """Run rasterise -> fill -> EDT/SDF -> smooth on the GPU; return host SDF.

    Output is a float32 NumPy array matching the CPU ``build_sdf_edt`` +
    Gaussian-smooth result, ready for ``skimage.marching_cubes``.
    """
    import cupy as cp
    import cupyx.scipy.ndimage as cndi
    from scipy.ndimage import generate_binary_structure

    struct = cp.asarray(generate_binary_structure(3, 1))

    occ = _rasterise_gpu(cp, strands, bbox_min, voxel_size, grid_dims)
    if closing_radius > 0:
        occ = _fill_interior_gpu(
            cp, cndi, struct, occ, closing_radius, eff_radius_vox,
        )

    edt = cndi.distance_transform_edt(~occ).astype(cp.float32)
    sdf = edt * voxel_size - (eff_radius_vox * voxel_size)
    del occ, edt

    sigma = (radius * 0.5) / voxel_size
    if smooth_iterations > 0 and sigma > 0:
        for _ in range(smooth_iterations):
            sdf = cndi.gaussian_filter(sdf, sigma=sigma)

    cp.cuda.Stream.null.synchronize()
    out = cp.asnumpy(sdf)
    del sdf
    cp.get_default_memory_pool().free_all_blocks()
    return out