pyfracval.cca_kernels#

CCA-specific JIT kernels for PyFracVAL.

JIT-compiled helper functions used during CCA sticking and retry operations.

These kernels accelerate retry and overlap stages in the CCA procedure from [Morán et al., 2019].

Constants#

_GOLDEN_RATIO

Golden ratio constant used for Fibonacci spiral rotations.

_TWO_PI

2 * pi constant for angle calculations.

Functions#

_cca_reintento_kernel

JIT kernel for the CCA “reintento” (retry) overlap check.

batch_check_overlaps_cca

JIT parallel overlap checker for CCA batch rotation.

batch_rotate_cluster_cca

Rotate all positions in a CCA cluster around its centre of mass.

Module Contents#

pyfracval.cca_kernels.batch_check_overlaps_cca(coords1, radii1, coords2_batch, radii2, tolerance)[source]#

Check overlap for batch of cluster2 configurations (CCA).

Uses Numba parallel loops to evaluate multiple cluster configurations simultaneously.

Parameters:

  • coords1 (np.ndarray) – Cluster 1 coordinates (n1, 3)

  • radii1 (np.ndarray) – Cluster 1 radii (n1,)

  • coords2_batch (np.ndarray) – Batch of cluster 2 configurations (n_batch, n2, 3)

  • radii2 (np.ndarray) – Cluster 2 radii (n2,) - same for all configurations

  • tolerance (float) – Overlap tolerance

Returns:

(n_batch,) array of max overlap values for each configuration

Return type:

np.ndarray

pyfracval.cca_kernels.batch_rotate_cluster_cca(coords2_in, cm2, cand2_idx, vec_0, i_vec, j_vec, angles)[source]#

Batch rotate cluster2 for multiple angles (CCA).

For each angle, calculates the target position on the intersection circle, then rotates the entire cluster to align the candidate particle with that target.

Parameters:

  • coords2_in (np.ndarray) – Cluster 2 coordinates (n2, 3)

  • cm2 (np.ndarray) – Center of mass of cluster 2 (3,)

  • cand2_idx (int) – Index of candidate particle in cluster 2

  • vec_0 (np.ndarray) – [x0, y0, z0, r0] - center and radius of intersection circle

  • i_vec (np.ndarray) – First basis vector (3,)

  • j_vec (np.ndarray) – Second basis vector (3,)

  • angles (np.ndarray) – Array of rotation angles (n_angles,)

Returns:

(n_angles, n2, 3) array of rotated cluster configurations

Return type:

np.ndarray

pyfracval.cca_kernels.cca_sticking_v1_kernel(coords1_in, radii1, cm1, coords2_in, radii2, cm2_in, cand1_idx, cand2_idx, gamma_pc, theta_a, theta_b)[source]#

Fused ext_case=0 sticking placement.

The interpreted version of this spends most of its time on numpy dispatch for 3-vectors rather than on arithmetic: per N=1024 aggregate it drove ~30k norm calls, ~24k array allocations and ~20k zeros. Fusing the whole placement into one compiled function removes all of that, and lets the two cluster transforms run as tight loops.

Randomness is hoisted out: the two angles the sphere-sphere intersections would have sampled are passed in, so this stays pure and the caller keeps sole ownership of the RNG stream (which is what makes runs reproducible).

Returns (coords1_out, coords2_out, cm2_out, vec0, i_vec, j_vec, ok); ok is False when either intersection has no solution, in which case the arrays are meaningless and the caller must skip them.