Tracked centrally in fastfields/fastfields-lib#17 (item C3).
The trailing-_ variants sym_addmatvec_ / sym_submatvec_ advertise in-place accumulation (out0 += H @ vec), but the implementation materialises a private contiguous copy (out = np.array(out_b, ...)), writes the binding's result into that copy, and returns it — leaving the caller's accumulator untouched.
This silently drops the update and diverges from the cupy backend (fastfields-cupy/.../_sym.py), whose _ variants truly write through the caller's buffer.
Fix: operate in place on the caller's out0 (write through its DLPack buffer), matching the _ contract and the cupy backend.
Fixed in PR referencing this issue.
Tracked centrally in fastfields/fastfields-lib#17 (item C3).
The trailing-
_variantssym_addmatvec_/sym_submatvec_advertise in-place accumulation (out0 += H @ vec), but the implementation materialises a private contiguous copy (out = np.array(out_b, ...)), writes the binding's result into that copy, and returns it — leaving the caller's accumulator untouched.This silently drops the update and diverges from the cupy backend (
fastfields-cupy/.../_sym.py), whose_variants truly write through the caller's buffer.Fix: operate in place on the caller's
out0(write through its DLPack buffer), matching the_contract and the cupy backend.Fixed in PR referencing this issue.