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
vpx_convolve8_neon(const uint8_t * src,ptrdiff_t src_stride,uint8_t * dst,ptrdiff_t dst_stride,const InterpKernel * filter,int x0_q4,int x_step_q4,int y0_q4,int y_step_q4,int w,int h)17 void vpx_convolve8_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
18 ptrdiff_t dst_stride, const InterpKernel *filter,
19 int x0_q4, int x_step_q4, int y0_q4, int y_step_q4,
20 int w, int h) {
21 /* Given our constraints: w <= 64, h <= 64, taps == 8 we can reduce the
22 * maximum buffer size to 64 * 64 + 7 (+ 1 to make it divisible by 4).
23 */
24 uint8_t temp[64 * 72];
25
26 // Account for the vertical phase needing 3 lines prior and 4 lines post
27 // (+ 1 to make it divisible by 4).
28 const int intermediate_height = h + 8;
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 given
34 * height and filter a multiple of 4 lines. Since this goes in to the temp
35 * buffer which has lots of extra room and is subsequently discarded this is
36 * safe if somewhat less than ideal. */
37 vpx_convolve8_horiz_neon(src - src_stride * 3, src_stride, temp, w, filter,
38 x0_q4, x_step_q4, y0_q4, y_step_q4, w,
39 intermediate_height);
40
41 /* Step into the temp buffer 3 lines to get the actual frame data */
42 vpx_convolve8_vert_neon(temp + w * 3, w, dst, dst_stride, filter, x0_q4,
43 x_step_q4, y0_q4, y_step_q4, w, h);
44 }
45
vpx_convolve8_avg_neon(const uint8_t * src,ptrdiff_t src_stride,uint8_t * dst,ptrdiff_t dst_stride,const InterpKernel * filter,int x0_q4,int x_step_q4,int y0_q4,int y_step_q4,int w,int h)46 void vpx_convolve8_avg_neon(const uint8_t *src, ptrdiff_t src_stride,
47 uint8_t *dst, ptrdiff_t dst_stride,
48 const InterpKernel *filter, int x0_q4,
49 int x_step_q4, int y0_q4, int y_step_q4, int w,
50 int h) {
51 uint8_t temp[64 * 72];
52 const int intermediate_height = h + 8;
53
54 assert(y_step_q4 == 16);
55 assert(x_step_q4 == 16);
56
57 /* This implementation has the same issues as above. In addition, we only want
58 * to average the values after both passes.
59 */
60 vpx_convolve8_horiz_neon(src - src_stride * 3, src_stride, temp, w, filter,
61 x0_q4, x_step_q4, y0_q4, y_step_q4, w,
62 intermediate_height);
63 vpx_convolve8_avg_vert_neon(temp + w * 3, w, dst, dst_stride, filter, x0_q4,
64 x_step_q4, y0_q4, y_step_q4, w, h);
65 }
66