vp9_convolve_neon.c 3.27 KB
Newer Older
Johann's avatar
Johann committed
1
2
3
4
5
6
7
8
9
10
11
12
/*
 *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
 *
 *  Use of this source code is governed by a BSD-style license
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
 *  in the file PATENTS.  All contributing project authors may
 *  be found in the AUTHORS file in the root of the source tree.
 */

#include "./vp9_rtcd.h"
#include "vp9/common/vp9_common.h"
13
#include "vpx_ports/mem.h"
Johann's avatar
Johann committed
14
15
16
17
18
19
20
21
22

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).
   */
23
  DECLARE_ALIGNED_ARRAY(8, uint8_t, temp, 64 * 72);
Johann's avatar
Johann committed
24
25
26
27

  // Account for the vertical phase needing 3 lines prior and 4 lines post
  int intermediate_height = h + 7;

Johann's avatar
Johann committed
28
29
30
31
32
33
34
35
  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;
  }
Johann's avatar
Johann committed
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58

  /* 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);
}

void vp9_convolve8_avg_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) {
59
  DECLARE_ALIGNED_ARRAY(8, uint8_t, temp, 64 * 72);
Johann's avatar
Johann committed
60
61
  int intermediate_height = h + 7;

Johann's avatar
Johann committed
62
63
64
65
66
67
68
69
  if (x_step_q4 != 16 || y_step_q4 != 16) {
    vp9_convolve8_avg_c(src, src_stride,
                        dst, dst_stride,
                        filter_x, x_step_q4,
                        filter_y, y_step_q4,
                        w, h);
    return;
  }
Johann's avatar
Johann committed
70
71
72
73
74
75
76
77
78
79
80
81
82

  /* This implementation has the same issues as above. In addition, we only want
   * to average the values after both passes.
   */
  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);
  vp9_convolve8_avg_vert_neon(temp + 64 * 3,
                              64, dst, dst_stride,
                              filter_x, x_step_q4, filter_y, y_step_q4,
                              w, h);
}