def _(out_y):
self.Y[0][out_y].assign_vector(
_unreduced_squant(
unreduced[0][out_y].get_vector(),
(self.input_squant.params, self.weight_squant.params),
self.output_squant.params,
n_summands).reduce_after_mul())
def _(i):
res = _unreduced_squant(
sint.load_mem(unreduced.address + i * n_per_thread,
size=n_per_thread),
(self.input_squant.params, self.weight_squant.params),
self.output_squant.params, n_summands).reduce_after_mul()
res.store_in_mem(self.Y.address + i * n_per_thread)
