pytc.kmat.contract_K1_minus_K2_isdf_streaming¶

pytc.kmat.contract_K1_minus_K2_isdf_streaming(phi_p, phi_q, phi_r, phi_s, grad_phi_p, grad_phi_q, U1, rank_block_size=128, panel_size=None)[source]¶

Streaming-capable wrapper around contract_K1_minus_K2_isdf_jit().

U1 can be a device jax.Array or a host numpy ndarray.

panel_size=None or panel_size >= U1.shape[1] → delegate to the JIT once with the full K1; behaviour is bit-identical to the resident fast-path.
Otherwise, iterate over axis-1 (rank-column) panels of panel_size; each panel is jax.device_put just before its call, phi_r/phi_s sliced to the matching slab, and partial contributions summed on device. Axis 0 of U1 (the k axis) is untouched.

The last panel is zero-padded to panel_size so the JIT compiles once for the whole loop. Zero-padded rows/columns contribute 0 to the sum.

Note: a function named contract_K1_minus_K2_isdf (without the _streaming suffix) already exists as a range-based wrapper that slices from a full phi_piv / grad_phi_piv — keep the names distinct.