1/*
2 *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
3 *
4 *  Use of this source code is governed by a BSD-style license
5 *  that can be found in the LICENSE file in the root of the source
6 *  tree. An additional intellectual property rights grant can be found
7 *  in the file PATENTS.  All contributing project authors may
8 *  be found in the AUTHORS file in the root of the source tree.
9 */
10
11#include "./vp9_rtcd.h"
12#include "vp9/common/vp9_common.h"
13#include "vpx_ports/mem.h"
14
15void vp9_convolve8_neon(const uint8_t *src, ptrdiff_t src_stride,
16                        uint8_t *dst, ptrdiff_t dst_stride,
17                        const int16_t *filter_x, int x_step_q4,
18                        const int16_t *filter_y, int y_step_q4,
19                        int w, int h) {
20  /* Given our constraints: w <= 64, h <= 64, taps == 8 we can reduce the
21   * maximum buffer size to 64 * 64 + 7 (+ 1 to make it divisible by 4).
22   */
23  DECLARE_ALIGNED_ARRAY(8, uint8_t, temp, 64 * 72);
24
25  // Account for the vertical phase needing 3 lines prior and 4 lines post
26  int intermediate_height = h + 7;
27
28  if (x_step_q4 != 16 || y_step_q4 != 16)
29    return vp9_convolve8_c(src, src_stride,
30                           dst, dst_stride,
31                           filter_x, x_step_q4,
32                           filter_y, y_step_q4,
33                           w, h);
34
35  /* Filter starting 3 lines back. The neon implementation will ignore the
36   * given height and filter a multiple of 4 lines. Since this goes in to
37   * the temp buffer which has lots of extra room and is subsequently discarded
38   * this is safe if somewhat less than ideal.
39   */
40  vp9_convolve8_horiz_neon(src - src_stride * 3, src_stride,
41                           temp, 64,
42                           filter_x, x_step_q4, filter_y, y_step_q4,
43                           w, intermediate_height);
44
45  /* Step into the temp buffer 3 lines to get the actual frame data */
46  vp9_convolve8_vert_neon(temp + 64 * 3, 64,
47                          dst, dst_stride,
48                          filter_x, x_step_q4, filter_y, y_step_q4,
49                          w, h);
50}
51
52void vp9_convolve8_avg_neon(const uint8_t *src, ptrdiff_t src_stride,
53                            uint8_t *dst, ptrdiff_t dst_stride,
54                            const int16_t *filter_x, int x_step_q4,
55                            const int16_t *filter_y, int y_step_q4,
56                            int w, int h) {
57  DECLARE_ALIGNED_ARRAY(8, uint8_t, temp, 64 * 72);
58  int intermediate_height = h + 7;
59
60  if (x_step_q4 != 16 || y_step_q4 != 16)
61    return vp9_convolve8_avg_c(src, src_stride,
62                               dst, dst_stride,
63                               filter_x, x_step_q4,
64                               filter_y, y_step_q4,
65                               w, h);
66
67  /* This implementation has the same issues as above. In addition, we only want
68   * to average the values after both passes.
69   */
70  vp9_convolve8_horiz_neon(src - src_stride * 3, src_stride,
71                           temp, 64,
72                           filter_x, x_step_q4, filter_y, y_step_q4,
73                           w, intermediate_height);
74  vp9_convolve8_avg_vert_neon(temp + 64 * 3,
75                              64, dst, dst_stride,
76                              filter_x, x_step_q4, filter_y, y_step_q4,
77                              w, h);
78}
79