void vp9_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride,
uint8_t *dst, ptrdiff_t dst_stride,
const int16_t *filter_x, int x_step_q4,
const int16_t *filter_y, int y_step_q4,
int w, int h) {
/* Fixed size intermediate buffer places limits on parameters. */
DECLARE_ALIGNED_ARRAY(16, uint8_t, temp, 64 * 64);
vp9_convolve8_c(src, src_stride, temp, 64,
filter_x, x_step_q4, filter_y, y_step_q4, w, h);
vp9_convolve_avg_c(temp, 64, dst, dst_stride, NULL, 0, NULL, 0, w, h);
}
void vp9_convolve8_neon(const uint8_t *src, ptrdiff_t src_stride,
uint8_t *dst, ptrdiff_t dst_stride,
const int16_t *filter_x, int x_step_q4,
const int16_t *filter_y, int y_step_q4,
int w, int h) {
/* Given our constraints: w <= 64, h <= 64, taps == 8 we can reduce the
* maximum buffer size to 64 * 64 + 7 (+ 1 to make it divisible by 4).
*/
DECLARE_ALIGNED(8, uint8_t, temp[64 * 72]);
// Account for the vertical phase needing 3 lines prior and 4 lines post
int intermediate_height = h + 7;
if (x_step_q4 != 16 || y_step_q4 != 16) {
vp9_convolve8_c(src, src_stride,
dst, dst_stride,
filter_x, x_step_q4,
filter_y, y_step_q4,
w, h);
return;
}
/* Filter starting 3 lines back. The neon implementation will ignore the
* given height and filter a multiple of 4 lines. Since this goes in to
* the temp buffer which has lots of extra room and is subsequently discarded
* this is safe if somewhat less than ideal.
*/
vp9_convolve8_horiz_neon(src - src_stride * 3, src_stride,
temp, 64,
filter_x, x_step_q4, filter_y, y_step_q4,
w, intermediate_height);
/* Step into the temp buffer 3 lines to get the actual frame data */
vp9_convolve8_vert_neon(temp + 64 * 3, 64,
dst, dst_stride,
filter_x, x_step_q4, filter_y, y_step_q4,
w, h);
}
