tree_d_3d.kernels.fmm_m2l_host.fmm_m2l_host_pyccel
The Pyccel version (CPU) of the fmm_m2l_host kernel of the FMM.
fmm_m2l_host_numpy.py next to this file is the version for machines without Pyccel,
with the same arguments; the folder’s __init__.py declares the kernel (see
tree_d_3d.kernels). The flat quadtree is described in tree_d_3d.flat_tree.
Functions
| Name | Description |
|---|---|
fmm_m2l_host | Add the multipoles of cells sources[row_start[n]:row_start[n + 1]] to cell n. |
fmm_m2l_hostfunction#
def fmm_m2l_host(row_start: int64[:], sources: int64[:], dst_center: float[:, :], src_center: float[:, :], moment_re: float[:, :], moment_im: float[:, :], coeff: float[:, :], local_re: float[:, :], local_im: float[:, :]) -> NoneAdd the multipoles of cells sources[row_start[n]:row_start[n + 1]] to cell n.
The destination cells’ centres are dst_center, the sources’ src_center
(the same array within one tree).
With t = c_dst - c_src, w = 1 / t and the scaled moments
a'_k = a_k w**k (a'_0 = a_0), the translation is a real matrix product,
s_l = sum_k coeff[k, l] a'_k, with coeff[0, l] = 1 / l,
coeff[k, 0] = 1 / k and coeff[k, l] = C(k + l - 1, l) / k (zero beyond the
truncation): b_0 += a_0 ln t - s_0 and b_l += -(-1)**l s_l w**l. Each cell
is written by its own row only, so the rows are independent (chunks of 64 cells
share their scratch).