42 ms·

It takes a different code path for efficiency. e.g if (batch_size > 1024): kernel_x else: kernel_y