vpx_convolve_neon.c revision 7ce0a1d1337c01056ba24006efab21f00e179e04 
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 <assert.h>
12
13#include "./vpx_dsp_rtcd.h"
14#include "vpx_dsp/vpx_dsp_common.h"
15#include "vpx_ports/mem.h"
16
17void vpx_convolve8_neon(const uint8_t *src, ptrdiff_t src_stride,
18                        uint8_t *dst, ptrdiff_t dst_stride,
19                        const int16_t *filter_x, int x_step_q4,
20                        const int16_t *filter_y, int y_step_q4,
21                        int w, int h) {
22  /* Given our constraints: w <= 64, h <= 64, taps == 8 we can reduce the
23   * maximum buffer size to 64 * 64 + 7 (+ 1 to make it divisible by 4).
24   */
25  DECLARE_ALIGNED(8, uint8_t, temp[64 * 72]);
26
27  // Account for the vertical phase needing 3 lines prior and 4 lines post
28  int intermediate_height = h + 7;
29
30  assert(y_step_q4 == 16);
31  assert(x_step_q4 == 16);
32
33  /* Filter starting 3 lines back. The neon implementation will ignore the
34   * given height and filter a multiple of 4 lines. Since this goes in to
35   * the temp buffer which has lots of extra room and is subsequently discarded
36   * this is safe if somewhat less than ideal.
37   */
38  vpx_convolve8_horiz_neon(src - src_stride * 3, src_stride,
39                           temp, 64,
40                           filter_x, x_step_q4, filter_y, y_step_q4,
41                           w, intermediate_height);
42
43  /* Step into the temp buffer 3 lines to get the actual frame data */
44  vpx_convolve8_vert_neon(temp + 64 * 3, 64,
45                          dst, dst_stride,
46                          filter_x, x_step_q4, filter_y, y_step_q4,
47                          w, h);
48}
49
50void vpx_convolve8_avg_neon(const uint8_t *src, ptrdiff_t src_stride,
51                            uint8_t *dst, ptrdiff_t dst_stride,
52                            const int16_t *filter_x, int x_step_q4,
53                            const int16_t *filter_y, int y_step_q4,
54                            int w, int h) {
55  DECLARE_ALIGNED(8, uint8_t, temp[64 * 72]);
56  int intermediate_height = h + 7;
57
58  assert(y_step_q4 == 16);
59  assert(x_step_q4 == 16);
60
61  /* This implementation has the same issues as above. In addition, we only want
62   * to average the values after both passes.
63   */
64  vpx_convolve8_horiz_neon(src - src_stride * 3, src_stride,
65                           temp, 64,
66                           filter_x, x_step_q4, filter_y, y_step_q4,
67                           w, intermediate_height);
68  vpx_convolve8_avg_vert_neon(temp + 64 * 3,
69                              64, dst, dst_stride,
70                              filter_x, x_step_q4, filter_y, y_step_q4,
71                              w, h);
72}
73