xref: /aosp_15_r20/external/libvpx/vpx_dsp/arm/vpx_convolve_neon.c (revision fb1b10ab9aebc7c7068eedab379b749d7e3900be)
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_dsp/vpx_filter.h"
16 #include "vpx_ports/mem.h"
17 
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)18 void vpx_convolve8_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
19                         ptrdiff_t dst_stride, const InterpKernel *filter,
20                         int x0_q4, int x_step_q4, int y0_q4, 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 row to make it divisible by 4).
24   DECLARE_ALIGNED(32, uint8_t, im_block[64 * 72]);
25   const int im_stride = 64;
26 
27   const int vert_filter_taps = vpx_get_filter_taps(filter[y0_q4]) <= 4 ? 4 : 8;
28   // Account for the vertical phase needing vert_filter_taps / 2 - 1 lines prior
29   // and vert_filter_taps / 2 lines post. (+1 to make total divisible by 4.)
30   const int im_height = h + vert_filter_taps;
31   const ptrdiff_t border_offset = vert_filter_taps / 2 - 1;
32 
33   assert(y_step_q4 == 16);
34   assert(x_step_q4 == 16);
35 
36   // Filter starting border_offset rows back. The Neon implementation will
37   // ignore the given height and filter a multiple of 4 lines. Since this goes
38   // into the temporary buffer which has lots of extra room and is subsequently
39   // discarded this is safe if somewhat less than ideal.
40   vpx_convolve8_horiz_neon(src - src_stride * border_offset, src_stride,
41                            im_block, im_stride, filter, x0_q4, x_step_q4, y0_q4,
42                            y_step_q4, w, im_height);
43 
44   // Step into the temporary buffer border_offset rows to get actual frame data.
45   vpx_convolve8_vert_neon(im_block + im_stride * border_offset, im_stride, dst,
46                           dst_stride, filter, x0_q4, x_step_q4, y0_q4,
47                           y_step_q4, w, h);
48 }
49 
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)50 void vpx_convolve8_avg_neon(const uint8_t *src, ptrdiff_t src_stride,
51                             uint8_t *dst, ptrdiff_t dst_stride,
52                             const InterpKernel *filter, int x0_q4,
53                             int x_step_q4, int y0_q4, int y_step_q4, int w,
54                             int h) {
55   DECLARE_ALIGNED(32, uint8_t, im_block[64 * 72]);
56   const int im_stride = 64;
57   const int im_height = h + SUBPEL_TAPS;
58   const ptrdiff_t border_offset = SUBPEL_TAPS / 2 - 1;
59 
60   assert(y_step_q4 == 16);
61   assert(x_step_q4 == 16);
62 
63   // This implementation has the same issues as above. In addition, we only want
64   // to average the values after both passes.
65   vpx_convolve8_horiz_neon(src - src_stride * border_offset, src_stride,
66                            im_block, im_stride, filter, x0_q4, x_step_q4, y0_q4,
67                            y_step_q4, w, im_height);
68 
69   vpx_convolve8_avg_vert_neon(im_block + im_stride * border_offset, im_stride,
70                               dst, dst_stride, filter, x0_q4, x_step_q4, y0_q4,
71                               y_step_q4, w, h);
72 }
73