#if 0
void convert_neon(struct image_in const * __restrict__ input,
		  struct image_out const * __restrict__ output)
#endif
{
	do_prefetch(&in[-1*stride - 1], 0);
	do_prefetch(&in[ 0*stride - 1], 0);
	do_prefetch(&in[+1*stride - 1], (rows_per_loop) < 3 ? 3 : 0);
	do_prefetch(&in[+2*stride - 1], (rows_per_loop) < 3 ? 3 : 0);

	if (rows_per_loop > 2) {
		do_prefetch(&in[+3*stride - 1], 0);
		do_prefetch(&in[+4*stride - 1], 3);
	}

	do {
		unsigned int const	offs_r = offsetof(rgb_pixel_t, r);
		unsigned int const	offs_g = offsetof(rgb_pixel_t, g);
		unsigned int const	offs_b = offsetof(rgb_pixel_t, b);

		uint8x8x2_t	rowp = vld2_u8(&in[-1*stride - 1]);
		uint8x8x2_t	row0 = vld2_u8(&in[ 0*stride - 1]);
		uint8x8x2_t	row1 = vld2_u8(&in[+1*stride - 1]);

		convert_fn0(&(out[0*out_stride]),
			    offs_r, offs_g, offs_b,
			    rowp.val[0], rowp.val[1],
			    row0.val[0], row0.val[1],
			    row1.val[0], row1.val[1]);

		uint8x8x2_t	row2 = vld2_u8(&in[+2*stride - 1]);

		convert_fn1(&(out[1*out_stride]),
			    offs_r, offs_g, offs_b,
			    row0.val[0], row0.val[1],
			    row1.val[0], row1.val[1],
			    row2.val[0], row2.val[1]);

		if (rows_per_loop < 3)
			break;

		uint8x8x2_t	row3 = vld2_u8(&in[+3*stride - 1]);

		convert_fn0(&(out[2*out_stride]),
			    offs_r, offs_g, offs_b,
			    row1.val[0], row1.val[1],
			    row2.val[0], row2.val[1],
			    row3.val[0], row3.val[1]);

		uint8x8x2_t	row4 = vld2_u8(&in[+4*stride - 1]);

		convert_fn1(&(out[3*out_stride]),
			    offs_r, offs_g, offs_b,
			    row2.val[0], row2.val[1],
			    row3.val[0], row3.val[1],
			    row4.val[0], row4.val[1]);
	} while (0);
}