2 * Copyright (c) 2013 The WebM project authors. All Rights Reserved.
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.
11 #include "./vp9_rtcd.h"
12 #include "vp9/common/vp9_common.h"
13 #include "vpx_ports/mem.h"
15 void 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,
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).
23 DECLARE_ALIGNED_ARRAY(8, uint8_t, temp, 64 * 72);
25 // Account for the vertical phase needing 3 lines prior and 4 lines post
26 int intermediate_height = h + 7;
28 if (x_step_q4 != 16 || y_step_q4 != 16) {
29 vp9_convolve8_c(src, src_stride,
37 /* Filter starting 3 lines back. The neon implementation will ignore the
38 * given height and filter a multiple of 4 lines. Since this goes in to
39 * the temp buffer which has lots of extra room and is subsequently discarded
40 * this is safe if somewhat less than ideal.
42 vp9_convolve8_horiz_neon(src - src_stride * 3, src_stride,
44 filter_x, x_step_q4, filter_y, y_step_q4,
45 w, intermediate_height);
47 /* Step into the temp buffer 3 lines to get the actual frame data */
48 vp9_convolve8_vert_neon(temp + 64 * 3, 64,
50 filter_x, x_step_q4, filter_y, y_step_q4,
54 void vp9_convolve8_avg_neon(const uint8_t *src, ptrdiff_t src_stride,
55 uint8_t *dst, ptrdiff_t dst_stride,
56 const int16_t *filter_x, int x_step_q4,
57 const int16_t *filter_y, int y_step_q4,
59 DECLARE_ALIGNED_ARRAY(8, uint8_t, temp, 64 * 72);
60 int intermediate_height = h + 7;
62 if (x_step_q4 != 16 || y_step_q4 != 16) {
63 vp9_convolve8_avg_c(src, src_stride,
71 /* This implementation has the same issues as above. In addition, we only want
72 * to average the values after both passes.
74 vp9_convolve8_horiz_neon(src - src_stride * 3, src_stride,
76 filter_x, x_step_q4, filter_y, y_step_q4,
77 w, intermediate_height);
78 vp9_convolve8_avg_vert_neon(temp + 64 * 3,
80 filter_x, x_step_q4, filter_y, y_step_q4,