xref: /aosp_15_r20/external/libaom/av1/common/resize.c (revision 77c1e3ccc04c968bd2bc212e87364f250e820521)
1 /*
2  * Copyright (c) 2016, Alliance for Open Media. All rights reserved.
3  *
4  * This source code is subject to the terms of the BSD 2 Clause License and
5  * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6  * was not distributed with this source code in the LICENSE file, you can
7  * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8  * Media Patent License 1.0 was not distributed with this source code in the
9  * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10  */
11 
12 #include <assert.h>
13 #include <limits.h>
14 #include <math.h>
15 #include <stdbool.h>
16 #include <stdio.h>
17 #include <stdlib.h>
18 #include <string.h>
19 
20 #include "config/aom_config.h"
21 #include "config/av1_rtcd.h"
22 
23 #include "aom_dsp/aom_dsp_common.h"
24 #include "aom_dsp/flow_estimation/corner_detect.h"
25 #include "aom_ports/mem.h"
26 #include "av1/common/common.h"
27 #include "av1/common/resize.h"
28 
29 #include "config/aom_dsp_rtcd.h"
30 #include "config/aom_scale_rtcd.h"
31 
32 // Filters for interpolation (0.5-band) - note this also filters integer pels.
33 static const InterpKernel filteredinterp_filters500[(1 << RS_SUBPEL_BITS)] = {
34   { -3, 0, 35, 64, 35, 0, -3, 0 },    { -3, 0, 34, 64, 36, 0, -3, 0 },
35   { -3, -1, 34, 64, 36, 1, -3, 0 },   { -3, -1, 33, 64, 37, 1, -3, 0 },
36   { -3, -1, 32, 64, 38, 1, -3, 0 },   { -3, -1, 31, 64, 39, 1, -3, 0 },
37   { -3, -1, 31, 63, 39, 2, -3, 0 },   { -2, -2, 30, 63, 40, 2, -3, 0 },
38   { -2, -2, 29, 63, 41, 2, -3, 0 },   { -2, -2, 29, 63, 41, 3, -4, 0 },
39   { -2, -2, 28, 63, 42, 3, -4, 0 },   { -2, -2, 27, 63, 43, 3, -4, 0 },
40   { -2, -3, 27, 63, 43, 4, -4, 0 },   { -2, -3, 26, 62, 44, 5, -4, 0 },
41   { -2, -3, 25, 62, 45, 5, -4, 0 },   { -2, -3, 25, 62, 45, 5, -4, 0 },
42   { -2, -3, 24, 62, 46, 5, -4, 0 },   { -2, -3, 23, 61, 47, 6, -4, 0 },
43   { -2, -3, 23, 61, 47, 6, -4, 0 },   { -2, -3, 22, 61, 48, 7, -4, -1 },
44   { -2, -3, 21, 60, 49, 7, -4, 0 },   { -1, -4, 20, 60, 49, 8, -4, 0 },
45   { -1, -4, 20, 60, 50, 8, -4, -1 },  { -1, -4, 19, 59, 51, 9, -4, -1 },
46   { -1, -4, 19, 59, 51, 9, -4, -1 },  { -1, -4, 18, 58, 52, 10, -4, -1 },
47   { -1, -4, 17, 58, 52, 11, -4, -1 }, { -1, -4, 16, 58, 53, 11, -4, -1 },
48   { -1, -4, 16, 57, 53, 12, -4, -1 }, { -1, -4, 15, 57, 54, 12, -4, -1 },
49   { -1, -4, 15, 56, 54, 13, -4, -1 }, { -1, -4, 14, 56, 55, 13, -4, -1 },
50   { -1, -4, 14, 55, 55, 14, -4, -1 }, { -1, -4, 13, 55, 56, 14, -4, -1 },
51   { -1, -4, 13, 54, 56, 15, -4, -1 }, { -1, -4, 12, 54, 57, 15, -4, -1 },
52   { -1, -4, 12, 53, 57, 16, -4, -1 }, { -1, -4, 11, 53, 58, 16, -4, -1 },
53   { -1, -4, 11, 52, 58, 17, -4, -1 }, { -1, -4, 10, 52, 58, 18, -4, -1 },
54   { -1, -4, 9, 51, 59, 19, -4, -1 },  { -1, -4, 9, 51, 59, 19, -4, -1 },
55   { -1, -4, 8, 50, 60, 20, -4, -1 },  { 0, -4, 8, 49, 60, 20, -4, -1 },
56   { 0, -4, 7, 49, 60, 21, -3, -2 },   { -1, -4, 7, 48, 61, 22, -3, -2 },
57   { 0, -4, 6, 47, 61, 23, -3, -2 },   { 0, -4, 6, 47, 61, 23, -3, -2 },
58   { 0, -4, 5, 46, 62, 24, -3, -2 },   { 0, -4, 5, 45, 62, 25, -3, -2 },
59   { 0, -4, 5, 45, 62, 25, -3, -2 },   { 0, -4, 5, 44, 62, 26, -3, -2 },
60   { 0, -4, 4, 43, 63, 27, -3, -2 },   { 0, -4, 3, 43, 63, 27, -2, -2 },
61   { 0, -4, 3, 42, 63, 28, -2, -2 },   { 0, -4, 3, 41, 63, 29, -2, -2 },
62   { 0, -3, 2, 41, 63, 29, -2, -2 },   { 0, -3, 2, 40, 63, 30, -2, -2 },
63   { 0, -3, 2, 39, 63, 31, -1, -3 },   { 0, -3, 1, 39, 64, 31, -1, -3 },
64   { 0, -3, 1, 38, 64, 32, -1, -3 },   { 0, -3, 1, 37, 64, 33, -1, -3 },
65   { 0, -3, 1, 36, 64, 34, -1, -3 },   { 0, -3, 0, 36, 64, 34, 0, -3 },
66 };
67 
68 // Filters for interpolation (0.625-band) - note this also filters integer pels.
69 static const InterpKernel filteredinterp_filters625[(1 << RS_SUBPEL_BITS)] = {
70   { -1, -8, 33, 80, 33, -8, -1, 0 }, { -1, -8, 31, 80, 34, -8, -1, 1 },
71   { -1, -8, 30, 80, 35, -8, -1, 1 }, { -1, -8, 29, 80, 36, -7, -2, 1 },
72   { -1, -8, 28, 80, 37, -7, -2, 1 }, { -1, -8, 27, 80, 38, -7, -2, 1 },
73   { 0, -8, 26, 79, 39, -7, -2, 1 },  { 0, -8, 25, 79, 40, -7, -2, 1 },
74   { 0, -8, 24, 79, 41, -7, -2, 1 },  { 0, -8, 23, 78, 42, -6, -2, 1 },
75   { 0, -8, 22, 78, 43, -6, -2, 1 },  { 0, -8, 21, 78, 44, -6, -2, 1 },
76   { 0, -8, 20, 78, 45, -5, -3, 1 },  { 0, -8, 19, 77, 47, -5, -3, 1 },
77   { 0, -8, 18, 77, 48, -5, -3, 1 },  { 0, -8, 17, 77, 49, -5, -3, 1 },
78   { 0, -8, 16, 76, 50, -4, -3, 1 },  { 0, -8, 15, 76, 51, -4, -3, 1 },
79   { 0, -8, 15, 75, 52, -3, -4, 1 },  { 0, -7, 14, 74, 53, -3, -4, 1 },
80   { 0, -7, 13, 74, 54, -3, -4, 1 },  { 0, -7, 12, 73, 55, -2, -4, 1 },
81   { 0, -7, 11, 73, 56, -2, -4, 1 },  { 0, -7, 10, 72, 57, -1, -4, 1 },
82   { 1, -7, 10, 71, 58, -1, -5, 1 },  { 0, -7, 9, 71, 59, 0, -5, 1 },
83   { 1, -7, 8, 70, 60, 0, -5, 1 },    { 1, -7, 7, 69, 61, 1, -5, 1 },
84   { 1, -6, 6, 68, 62, 1, -5, 1 },    { 0, -6, 6, 68, 62, 2, -5, 1 },
85   { 1, -6, 5, 67, 63, 2, -5, 1 },    { 1, -6, 5, 66, 64, 3, -6, 1 },
86   { 1, -6, 4, 65, 65, 4, -6, 1 },    { 1, -6, 3, 64, 66, 5, -6, 1 },
87   { 1, -5, 2, 63, 67, 5, -6, 1 },    { 1, -5, 2, 62, 68, 6, -6, 0 },
88   { 1, -5, 1, 62, 68, 6, -6, 1 },    { 1, -5, 1, 61, 69, 7, -7, 1 },
89   { 1, -5, 0, 60, 70, 8, -7, 1 },    { 1, -5, 0, 59, 71, 9, -7, 0 },
90   { 1, -5, -1, 58, 71, 10, -7, 1 },  { 1, -4, -1, 57, 72, 10, -7, 0 },
91   { 1, -4, -2, 56, 73, 11, -7, 0 },  { 1, -4, -2, 55, 73, 12, -7, 0 },
92   { 1, -4, -3, 54, 74, 13, -7, 0 },  { 1, -4, -3, 53, 74, 14, -7, 0 },
93   { 1, -4, -3, 52, 75, 15, -8, 0 },  { 1, -3, -4, 51, 76, 15, -8, 0 },
94   { 1, -3, -4, 50, 76, 16, -8, 0 },  { 1, -3, -5, 49, 77, 17, -8, 0 },
95   { 1, -3, -5, 48, 77, 18, -8, 0 },  { 1, -3, -5, 47, 77, 19, -8, 0 },
96   { 1, -3, -5, 45, 78, 20, -8, 0 },  { 1, -2, -6, 44, 78, 21, -8, 0 },
97   { 1, -2, -6, 43, 78, 22, -8, 0 },  { 1, -2, -6, 42, 78, 23, -8, 0 },
98   { 1, -2, -7, 41, 79, 24, -8, 0 },  { 1, -2, -7, 40, 79, 25, -8, 0 },
99   { 1, -2, -7, 39, 79, 26, -8, 0 },  { 1, -2, -7, 38, 80, 27, -8, -1 },
100   { 1, -2, -7, 37, 80, 28, -8, -1 }, { 1, -2, -7, 36, 80, 29, -8, -1 },
101   { 1, -1, -8, 35, 80, 30, -8, -1 }, { 1, -1, -8, 34, 80, 31, -8, -1 },
102 };
103 
104 // Filters for interpolation (0.75-band) - note this also filters integer pels.
105 static const InterpKernel filteredinterp_filters750[(1 << RS_SUBPEL_BITS)] = {
106   { 2, -11, 25, 96, 25, -11, 2, 0 }, { 2, -11, 24, 96, 26, -11, 2, 0 },
107   { 2, -11, 22, 96, 28, -11, 2, 0 }, { 2, -10, 21, 96, 29, -12, 2, 0 },
108   { 2, -10, 19, 96, 31, -12, 2, 0 }, { 2, -10, 18, 95, 32, -11, 2, 0 },
109   { 2, -10, 17, 95, 34, -12, 2, 0 }, { 2, -9, 15, 95, 35, -12, 2, 0 },
110   { 2, -9, 14, 94, 37, -12, 2, 0 },  { 2, -9, 13, 94, 38, -12, 2, 0 },
111   { 2, -8, 12, 93, 40, -12, 1, 0 },  { 2, -8, 11, 93, 41, -12, 1, 0 },
112   { 2, -8, 9, 92, 43, -12, 1, 1 },   { 2, -8, 8, 92, 44, -12, 1, 1 },
113   { 2, -7, 7, 91, 46, -12, 1, 0 },   { 2, -7, 6, 90, 47, -12, 1, 1 },
114   { 2, -7, 5, 90, 49, -12, 1, 0 },   { 2, -6, 4, 89, 50, -12, 1, 0 },
115   { 2, -6, 3, 88, 52, -12, 0, 1 },   { 2, -6, 2, 87, 54, -12, 0, 1 },
116   { 2, -5, 1, 86, 55, -12, 0, 1 },   { 2, -5, 0, 85, 57, -12, 0, 1 },
117   { 2, -5, -1, 84, 58, -11, 0, 1 },  { 2, -5, -2, 83, 60, -11, 0, 1 },
118   { 2, -4, -2, 82, 61, -11, -1, 1 }, { 1, -4, -3, 81, 63, -10, -1, 1 },
119   { 2, -4, -4, 80, 64, -10, -1, 1 }, { 1, -4, -4, 79, 66, -10, -1, 1 },
120   { 1, -3, -5, 77, 67, -9, -1, 1 },  { 1, -3, -6, 76, 69, -9, -1, 1 },
121   { 1, -3, -6, 75, 70, -8, -2, 1 },  { 1, -2, -7, 74, 71, -8, -2, 1 },
122   { 1, -2, -7, 72, 72, -7, -2, 1 },  { 1, -2, -8, 71, 74, -7, -2, 1 },
123   { 1, -2, -8, 70, 75, -6, -3, 1 },  { 1, -1, -9, 69, 76, -6, -3, 1 },
124   { 1, -1, -9, 67, 77, -5, -3, 1 },  { 1, -1, -10, 66, 79, -4, -4, 1 },
125   { 1, -1, -10, 64, 80, -4, -4, 2 }, { 1, -1, -10, 63, 81, -3, -4, 1 },
126   { 1, -1, -11, 61, 82, -2, -4, 2 }, { 1, 0, -11, 60, 83, -2, -5, 2 },
127   { 1, 0, -11, 58, 84, -1, -5, 2 },  { 1, 0, -12, 57, 85, 0, -5, 2 },
128   { 1, 0, -12, 55, 86, 1, -5, 2 },   { 1, 0, -12, 54, 87, 2, -6, 2 },
129   { 1, 0, -12, 52, 88, 3, -6, 2 },   { 0, 1, -12, 50, 89, 4, -6, 2 },
130   { 0, 1, -12, 49, 90, 5, -7, 2 },   { 1, 1, -12, 47, 90, 6, -7, 2 },
131   { 0, 1, -12, 46, 91, 7, -7, 2 },   { 1, 1, -12, 44, 92, 8, -8, 2 },
132   { 1, 1, -12, 43, 92, 9, -8, 2 },   { 0, 1, -12, 41, 93, 11, -8, 2 },
133   { 0, 1, -12, 40, 93, 12, -8, 2 },  { 0, 2, -12, 38, 94, 13, -9, 2 },
134   { 0, 2, -12, 37, 94, 14, -9, 2 },  { 0, 2, -12, 35, 95, 15, -9, 2 },
135   { 0, 2, -12, 34, 95, 17, -10, 2 }, { 0, 2, -11, 32, 95, 18, -10, 2 },
136   { 0, 2, -12, 31, 96, 19, -10, 2 }, { 0, 2, -12, 29, 96, 21, -10, 2 },
137   { 0, 2, -11, 28, 96, 22, -11, 2 }, { 0, 2, -11, 26, 96, 24, -11, 2 },
138 };
139 
140 // Filters for interpolation (0.875-band) - note this also filters integer pels.
141 static const InterpKernel filteredinterp_filters875[(1 << RS_SUBPEL_BITS)] = {
142   { 3, -8, 13, 112, 13, -8, 3, 0 },   { 2, -7, 12, 112, 15, -8, 3, -1 },
143   { 3, -7, 10, 112, 17, -9, 3, -1 },  { 2, -6, 8, 112, 19, -9, 3, -1 },
144   { 2, -6, 7, 112, 21, -10, 3, -1 },  { 2, -5, 6, 111, 22, -10, 3, -1 },
145   { 2, -5, 4, 111, 24, -10, 3, -1 },  { 2, -4, 3, 110, 26, -11, 3, -1 },
146   { 2, -4, 1, 110, 28, -11, 3, -1 },  { 2, -4, 0, 109, 30, -12, 4, -1 },
147   { 1, -3, -1, 108, 32, -12, 4, -1 }, { 1, -3, -2, 108, 34, -13, 4, -1 },
148   { 1, -2, -4, 107, 36, -13, 4, -1 }, { 1, -2, -5, 106, 38, -13, 4, -1 },
149   { 1, -1, -6, 105, 40, -14, 4, -1 }, { 1, -1, -7, 104, 42, -14, 4, -1 },
150   { 1, -1, -7, 103, 44, -15, 4, -1 }, { 1, 0, -8, 101, 46, -15, 4, -1 },
151   { 1, 0, -9, 100, 48, -15, 4, -1 },  { 1, 0, -10, 99, 50, -15, 4, -1 },
152   { 1, 1, -11, 97, 53, -16, 4, -1 },  { 0, 1, -11, 96, 55, -16, 4, -1 },
153   { 0, 1, -12, 95, 57, -16, 4, -1 },  { 0, 2, -13, 93, 59, -16, 4, -1 },
154   { 0, 2, -13, 91, 61, -16, 4, -1 },  { 0, 2, -14, 90, 63, -16, 4, -1 },
155   { 0, 2, -14, 88, 65, -16, 4, -1 },  { 0, 2, -15, 86, 67, -16, 4, 0 },
156   { 0, 3, -15, 84, 69, -17, 4, 0 },   { 0, 3, -16, 83, 71, -17, 4, 0 },
157   { 0, 3, -16, 81, 73, -16, 3, 0 },   { 0, 3, -16, 79, 75, -16, 3, 0 },
158   { 0, 3, -16, 77, 77, -16, 3, 0 },   { 0, 3, -16, 75, 79, -16, 3, 0 },
159   { 0, 3, -16, 73, 81, -16, 3, 0 },   { 0, 4, -17, 71, 83, -16, 3, 0 },
160   { 0, 4, -17, 69, 84, -15, 3, 0 },   { 0, 4, -16, 67, 86, -15, 2, 0 },
161   { -1, 4, -16, 65, 88, -14, 2, 0 },  { -1, 4, -16, 63, 90, -14, 2, 0 },
162   { -1, 4, -16, 61, 91, -13, 2, 0 },  { -1, 4, -16, 59, 93, -13, 2, 0 },
163   { -1, 4, -16, 57, 95, -12, 1, 0 },  { -1, 4, -16, 55, 96, -11, 1, 0 },
164   { -1, 4, -16, 53, 97, -11, 1, 1 },  { -1, 4, -15, 50, 99, -10, 0, 1 },
165   { -1, 4, -15, 48, 100, -9, 0, 1 },  { -1, 4, -15, 46, 101, -8, 0, 1 },
166   { -1, 4, -15, 44, 103, -7, -1, 1 }, { -1, 4, -14, 42, 104, -7, -1, 1 },
167   { -1, 4, -14, 40, 105, -6, -1, 1 }, { -1, 4, -13, 38, 106, -5, -2, 1 },
168   { -1, 4, -13, 36, 107, -4, -2, 1 }, { -1, 4, -13, 34, 108, -2, -3, 1 },
169   { -1, 4, -12, 32, 108, -1, -3, 1 }, { -1, 4, -12, 30, 109, 0, -4, 2 },
170   { -1, 3, -11, 28, 110, 1, -4, 2 },  { -1, 3, -11, 26, 110, 3, -4, 2 },
171   { -1, 3, -10, 24, 111, 4, -5, 2 },  { -1, 3, -10, 22, 111, 6, -5, 2 },
172   { -1, 3, -10, 21, 112, 7, -6, 2 },  { -1, 3, -9, 19, 112, 8, -6, 2 },
173   { -1, 3, -9, 17, 112, 10, -7, 3 },  { -1, 3, -8, 15, 112, 12, -7, 2 },
174 };
175 
176 const int16_t av1_resize_filter_normative[(
177     1 << RS_SUBPEL_BITS)][UPSCALE_NORMATIVE_TAPS] = {
178 #if UPSCALE_NORMATIVE_TAPS == 8
179   { 0, 0, 0, 128, 0, 0, 0, 0 },        { 0, 0, -1, 128, 2, -1, 0, 0 },
180   { 0, 1, -3, 127, 4, -2, 1, 0 },      { 0, 1, -4, 127, 6, -3, 1, 0 },
181   { 0, 2, -6, 126, 8, -3, 1, 0 },      { 0, 2, -7, 125, 11, -4, 1, 0 },
182   { -1, 2, -8, 125, 13, -5, 2, 0 },    { -1, 3, -9, 124, 15, -6, 2, 0 },
183   { -1, 3, -10, 123, 18, -6, 2, -1 },  { -1, 3, -11, 122, 20, -7, 3, -1 },
184   { -1, 4, -12, 121, 22, -8, 3, -1 },  { -1, 4, -13, 120, 25, -9, 3, -1 },
185   { -1, 4, -14, 118, 28, -9, 3, -1 },  { -1, 4, -15, 117, 30, -10, 4, -1 },
186   { -1, 5, -16, 116, 32, -11, 4, -1 }, { -1, 5, -16, 114, 35, -12, 4, -1 },
187   { -1, 5, -17, 112, 38, -12, 4, -1 }, { -1, 5, -18, 111, 40, -13, 5, -1 },
188   { -1, 5, -18, 109, 43, -14, 5, -1 }, { -1, 6, -19, 107, 45, -14, 5, -1 },
189   { -1, 6, -19, 105, 48, -15, 5, -1 }, { -1, 6, -19, 103, 51, -16, 5, -1 },
190   { -1, 6, -20, 101, 53, -16, 6, -1 }, { -1, 6, -20, 99, 56, -17, 6, -1 },
191   { -1, 6, -20, 97, 58, -17, 6, -1 },  { -1, 6, -20, 95, 61, -18, 6, -1 },
192   { -2, 7, -20, 93, 64, -18, 6, -2 },  { -2, 7, -20, 91, 66, -19, 6, -1 },
193   { -2, 7, -20, 88, 69, -19, 6, -1 },  { -2, 7, -20, 86, 71, -19, 6, -1 },
194   { -2, 7, -20, 84, 74, -20, 7, -2 },  { -2, 7, -20, 81, 76, -20, 7, -1 },
195   { -2, 7, -20, 79, 79, -20, 7, -2 },  { -1, 7, -20, 76, 81, -20, 7, -2 },
196   { -2, 7, -20, 74, 84, -20, 7, -2 },  { -1, 6, -19, 71, 86, -20, 7, -2 },
197   { -1, 6, -19, 69, 88, -20, 7, -2 },  { -1, 6, -19, 66, 91, -20, 7, -2 },
198   { -2, 6, -18, 64, 93, -20, 7, -2 },  { -1, 6, -18, 61, 95, -20, 6, -1 },
199   { -1, 6, -17, 58, 97, -20, 6, -1 },  { -1, 6, -17, 56, 99, -20, 6, -1 },
200   { -1, 6, -16, 53, 101, -20, 6, -1 }, { -1, 5, -16, 51, 103, -19, 6, -1 },
201   { -1, 5, -15, 48, 105, -19, 6, -1 }, { -1, 5, -14, 45, 107, -19, 6, -1 },
202   { -1, 5, -14, 43, 109, -18, 5, -1 }, { -1, 5, -13, 40, 111, -18, 5, -1 },
203   { -1, 4, -12, 38, 112, -17, 5, -1 }, { -1, 4, -12, 35, 114, -16, 5, -1 },
204   { -1, 4, -11, 32, 116, -16, 5, -1 }, { -1, 4, -10, 30, 117, -15, 4, -1 },
205   { -1, 3, -9, 28, 118, -14, 4, -1 },  { -1, 3, -9, 25, 120, -13, 4, -1 },
206   { -1, 3, -8, 22, 121, -12, 4, -1 },  { -1, 3, -7, 20, 122, -11, 3, -1 },
207   { -1, 2, -6, 18, 123, -10, 3, -1 },  { 0, 2, -6, 15, 124, -9, 3, -1 },
208   { 0, 2, -5, 13, 125, -8, 2, -1 },    { 0, 1, -4, 11, 125, -7, 2, 0 },
209   { 0, 1, -3, 8, 126, -6, 2, 0 },      { 0, 1, -3, 6, 127, -4, 1, 0 },
210   { 0, 1, -2, 4, 127, -3, 1, 0 },      { 0, 0, -1, 2, 128, -1, 0, 0 },
211 #else
212 #error "Invalid value of UPSCALE_NORMATIVE_TAPS"
213 #endif  // UPSCALE_NORMATIVE_TAPS == 8
214 };
215 
216 // Filters for interpolation (full-band) - no filtering for integer pixels
217 #define filteredinterp_filters1000 av1_resize_filter_normative
218 
choose_interp_filter(int in_length,int out_length)219 static const InterpKernel *choose_interp_filter(int in_length, int out_length) {
220   int out_length16 = out_length * 16;
221   if (out_length16 >= in_length * 16)
222     return filteredinterp_filters1000;
223   else if (out_length16 >= in_length * 13)
224     return filteredinterp_filters875;
225   else if (out_length16 >= in_length * 11)
226     return filteredinterp_filters750;
227   else if (out_length16 >= in_length * 9)
228     return filteredinterp_filters625;
229   else
230     return filteredinterp_filters500;
231 }
232 
interpolate_core(const uint8_t * const input,int in_length,uint8_t * output,int out_length,const int16_t * interp_filters,int interp_taps)233 static void interpolate_core(const uint8_t *const input, int in_length,
234                              uint8_t *output, int out_length,
235                              const int16_t *interp_filters, int interp_taps) {
236   const int32_t delta =
237       (((uint32_t)in_length << RS_SCALE_SUBPEL_BITS) + out_length / 2) /
238       out_length;
239   const int32_t offset =
240       in_length > out_length
241           ? (((int32_t)(in_length - out_length) << (RS_SCALE_SUBPEL_BITS - 1)) +
242              out_length / 2) /
243                 out_length
244           : -(((int32_t)(out_length - in_length)
245                << (RS_SCALE_SUBPEL_BITS - 1)) +
246               out_length / 2) /
247                 out_length;
248   uint8_t *optr = output;
249   int x, x1, x2, sum, k, int_pel, sub_pel;
250   int32_t y;
251 
252   x = 0;
253   y = offset + RS_SCALE_EXTRA_OFF;
254   while ((y >> RS_SCALE_SUBPEL_BITS) < (interp_taps / 2 - 1)) {
255     x++;
256     y += delta;
257   }
258   x1 = x;
259   x = out_length - 1;
260   y = delta * x + offset + RS_SCALE_EXTRA_OFF;
261   while ((y >> RS_SCALE_SUBPEL_BITS) + (int32_t)(interp_taps / 2) >=
262          in_length) {
263     x--;
264     y -= delta;
265   }
266   x2 = x;
267   if (x1 > x2) {
268     for (x = 0, y = offset + RS_SCALE_EXTRA_OFF; x < out_length;
269          ++x, y += delta) {
270       int_pel = y >> RS_SCALE_SUBPEL_BITS;
271       sub_pel = (y >> RS_SCALE_EXTRA_BITS) & RS_SUBPEL_MASK;
272       const int16_t *filter = &interp_filters[sub_pel * interp_taps];
273       sum = 0;
274       for (k = 0; k < interp_taps; ++k) {
275         const int pk = int_pel - interp_taps / 2 + 1 + k;
276         sum += filter[k] * input[AOMMAX(AOMMIN(pk, in_length - 1), 0)];
277       }
278       *optr++ = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
279     }
280   } else {
281     // Initial part.
282     for (x = 0, y = offset + RS_SCALE_EXTRA_OFF; x < x1; ++x, y += delta) {
283       int_pel = y >> RS_SCALE_SUBPEL_BITS;
284       sub_pel = (y >> RS_SCALE_EXTRA_BITS) & RS_SUBPEL_MASK;
285       const int16_t *filter = &interp_filters[sub_pel * interp_taps];
286       sum = 0;
287       for (k = 0; k < interp_taps; ++k)
288         sum += filter[k] * input[AOMMAX(int_pel - interp_taps / 2 + 1 + k, 0)];
289       *optr++ = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
290     }
291     // Middle part.
292     for (; x <= x2; ++x, y += delta) {
293       int_pel = y >> RS_SCALE_SUBPEL_BITS;
294       sub_pel = (y >> RS_SCALE_EXTRA_BITS) & RS_SUBPEL_MASK;
295       const int16_t *filter = &interp_filters[sub_pel * interp_taps];
296       sum = 0;
297       for (k = 0; k < interp_taps; ++k)
298         sum += filter[k] * input[int_pel - interp_taps / 2 + 1 + k];
299       *optr++ = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
300     }
301     // End part.
302     for (; x < out_length; ++x, y += delta) {
303       int_pel = y >> RS_SCALE_SUBPEL_BITS;
304       sub_pel = (y >> RS_SCALE_EXTRA_BITS) & RS_SUBPEL_MASK;
305       const int16_t *filter = &interp_filters[sub_pel * interp_taps];
306       sum = 0;
307       for (k = 0; k < interp_taps; ++k)
308         sum += filter[k] *
309                input[AOMMIN(int_pel - interp_taps / 2 + 1 + k, in_length - 1)];
310       *optr++ = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
311     }
312   }
313 }
314 
interpolate(const uint8_t * const input,int in_length,uint8_t * output,int out_length)315 static void interpolate(const uint8_t *const input, int in_length,
316                         uint8_t *output, int out_length) {
317   const InterpKernel *interp_filters =
318       choose_interp_filter(in_length, out_length);
319 
320   interpolate_core(input, in_length, output, out_length, &interp_filters[0][0],
321                    SUBPEL_TAPS);
322 }
323 
av1_get_upscale_convolve_step(int in_length,int out_length)324 int32_t av1_get_upscale_convolve_step(int in_length, int out_length) {
325   return ((in_length << RS_SCALE_SUBPEL_BITS) + out_length / 2) / out_length;
326 }
327 
get_upscale_convolve_x0(int in_length,int out_length,int32_t x_step_qn)328 static int32_t get_upscale_convolve_x0(int in_length, int out_length,
329                                        int32_t x_step_qn) {
330   const int err = out_length * x_step_qn - (in_length << RS_SCALE_SUBPEL_BITS);
331   const int32_t x0 =
332       (-((out_length - in_length) << (RS_SCALE_SUBPEL_BITS - 1)) +
333        out_length / 2) /
334           out_length +
335       RS_SCALE_EXTRA_OFF - err / 2;
336   return (int32_t)((uint32_t)x0 & RS_SCALE_SUBPEL_MASK);
337 }
338 
down2_symeven(const uint8_t * const input,int length,uint8_t * output,int start_offset)339 void down2_symeven(const uint8_t *const input, int length, uint8_t *output,
340                    int start_offset) {
341   // Actual filter len = 2 * filter_len_half.
342   const int16_t *filter = av1_down2_symeven_half_filter;
343   const int filter_len_half = sizeof(av1_down2_symeven_half_filter) / 2;
344   int i, j;
345   uint8_t *optr = output;
346   int l1 = filter_len_half;
347   int l2 = (length - filter_len_half);
348   l1 += (l1 & 1);
349   l2 += (l2 & 1);
350   if (l1 > l2) {
351     // Short input length.
352     for (i = start_offset; i < length; i += 2) {
353       int sum = (1 << (FILTER_BITS - 1));
354       for (j = 0; j < filter_len_half; ++j) {
355         sum +=
356             (input[AOMMAX(i - j, 0)] + input[AOMMIN(i + 1 + j, length - 1)]) *
357             filter[j];
358       }
359       sum >>= FILTER_BITS;
360       *optr++ = clip_pixel(sum);
361     }
362   } else {
363     // Initial part.
364     for (i = start_offset; i < l1; i += 2) {
365       int sum = (1 << (FILTER_BITS - 1));
366       for (j = 0; j < filter_len_half; ++j) {
367         sum += (input[AOMMAX(i - j, 0)] + input[i + 1 + j]) * filter[j];
368       }
369       sum >>= FILTER_BITS;
370       *optr++ = clip_pixel(sum);
371     }
372     // Middle part.
373     for (; i < l2; i += 2) {
374       int sum = (1 << (FILTER_BITS - 1));
375       for (j = 0; j < filter_len_half; ++j) {
376         sum += (input[i - j] + input[i + 1 + j]) * filter[j];
377       }
378       sum >>= FILTER_BITS;
379       *optr++ = clip_pixel(sum);
380     }
381     // End part.
382     for (; i < length; i += 2) {
383       int sum = (1 << (FILTER_BITS - 1));
384       for (j = 0; j < filter_len_half; ++j) {
385         sum +=
386             (input[i - j] + input[AOMMIN(i + 1 + j, length - 1)]) * filter[j];
387       }
388       sum >>= FILTER_BITS;
389       *optr++ = clip_pixel(sum);
390     }
391   }
392 }
393 
down2_symodd(const uint8_t * const input,int length,uint8_t * output)394 static void down2_symodd(const uint8_t *const input, int length,
395                          uint8_t *output) {
396   // Actual filter len = 2 * filter_len_half - 1.
397   const int16_t *filter = av1_down2_symodd_half_filter;
398   const int filter_len_half = sizeof(av1_down2_symodd_half_filter) / 2;
399   int i, j;
400   uint8_t *optr = output;
401   int l1 = filter_len_half - 1;
402   int l2 = (length - filter_len_half + 1);
403   l1 += (l1 & 1);
404   l2 += (l2 & 1);
405   if (l1 > l2) {
406     // Short input length.
407     for (i = 0; i < length; i += 2) {
408       int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
409       for (j = 1; j < filter_len_half; ++j) {
410         sum += (input[(i - j < 0 ? 0 : i - j)] +
411                 input[(i + j >= length ? length - 1 : i + j)]) *
412                filter[j];
413       }
414       sum >>= FILTER_BITS;
415       *optr++ = clip_pixel(sum);
416     }
417   } else {
418     // Initial part.
419     for (i = 0; i < l1; i += 2) {
420       int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
421       for (j = 1; j < filter_len_half; ++j) {
422         sum += (input[(i - j < 0 ? 0 : i - j)] + input[i + j]) * filter[j];
423       }
424       sum >>= FILTER_BITS;
425       *optr++ = clip_pixel(sum);
426     }
427     // Middle part.
428     for (; i < l2; i += 2) {
429       int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
430       for (j = 1; j < filter_len_half; ++j) {
431         sum += (input[i - j] + input[i + j]) * filter[j];
432       }
433       sum >>= FILTER_BITS;
434       *optr++ = clip_pixel(sum);
435     }
436     // End part.
437     for (; i < length; i += 2) {
438       int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
439       for (j = 1; j < filter_len_half; ++j) {
440         sum += (input[i - j] + input[(i + j >= length ? length - 1 : i + j)]) *
441                filter[j];
442       }
443       sum >>= FILTER_BITS;
444       *optr++ = clip_pixel(sum);
445     }
446   }
447 }
448 
get_down2_length(int length,int steps)449 static int get_down2_length(int length, int steps) {
450   for (int s = 0; s < steps; ++s) length = (length + 1) >> 1;
451   return length;
452 }
453 
get_down2_steps(int in_length,int out_length)454 static int get_down2_steps(int in_length, int out_length) {
455   int steps = 0;
456   int proj_in_length;
457   while ((proj_in_length = get_down2_length(in_length, 1)) >= out_length) {
458     ++steps;
459     in_length = proj_in_length;
460     if (in_length == 1) {
461       // Special case: we break because any further calls to get_down2_length()
462       // with be with length == 1, which return 1, resulting in an infinite
463       // loop.
464       break;
465     }
466   }
467   return steps;
468 }
469 
resize_multistep(const uint8_t * const input,int length,uint8_t * output,int olength,uint8_t * otmp)470 static void resize_multistep(const uint8_t *const input, int length,
471                              uint8_t *output, int olength, uint8_t *otmp) {
472   if (length == olength) {
473     memcpy(output, input, sizeof(output[0]) * length);
474     return;
475   }
476   const int steps = get_down2_steps(length, olength);
477 
478   if (steps > 0) {
479     uint8_t *out = NULL;
480     int filteredlength = length;
481 
482     assert(otmp != NULL);
483     uint8_t *otmp2 = otmp + get_down2_length(length, 1);
484     for (int s = 0; s < steps; ++s) {
485       const int proj_filteredlength = get_down2_length(filteredlength, 1);
486       const uint8_t *const in = (s == 0 ? input : out);
487       if (s == steps - 1 && proj_filteredlength == olength)
488         out = output;
489       else
490         out = (s & 1 ? otmp2 : otmp);
491       if (filteredlength & 1)
492         down2_symodd(in, filteredlength, out);
493       else
494         down2_symeven(in, filteredlength, out, 0);
495       filteredlength = proj_filteredlength;
496     }
497     if (filteredlength != olength) {
498       interpolate(out, filteredlength, output, olength);
499     }
500   } else {
501     interpolate(input, length, output, olength);
502   }
503 }
504 
fill_col_to_arr(uint8_t * img,int stride,int len,uint8_t * arr)505 static void fill_col_to_arr(uint8_t *img, int stride, int len, uint8_t *arr) {
506   int i;
507   uint8_t *iptr = img;
508   uint8_t *aptr = arr;
509   for (i = 0; i < len; ++i, iptr += stride) {
510     *aptr++ = *iptr;
511   }
512 }
513 
fill_arr_to_col(uint8_t * img,int stride,int len,uint8_t * arr)514 static void fill_arr_to_col(uint8_t *img, int stride, int len, uint8_t *arr) {
515   int i;
516   uint8_t *iptr = img;
517   uint8_t *aptr = arr;
518   for (i = 0; i < len; ++i, iptr += stride) {
519     *iptr = *aptr++;
520   }
521 }
522 
av1_resize_vert_dir_c(uint8_t * intbuf,uint8_t * output,int out_stride,int height,int height2,int width2,int start_col)523 bool av1_resize_vert_dir_c(uint8_t *intbuf, uint8_t *output, int out_stride,
524                            int height, int height2, int width2, int start_col) {
525   bool mem_status = true;
526   uint8_t *arrbuf = (uint8_t *)aom_malloc(sizeof(*arrbuf) * height);
527   uint8_t *arrbuf2 = (uint8_t *)aom_malloc(sizeof(*arrbuf2) * height2);
528   if (arrbuf == NULL || arrbuf2 == NULL) {
529     mem_status = false;
530     goto Error;
531   }
532 
533   for (int i = start_col; i < width2; ++i) {
534     fill_col_to_arr(intbuf + i, width2, height, arrbuf);
535     down2_symeven(arrbuf, height, arrbuf2, 0);
536     fill_arr_to_col(output + i, out_stride, height2, arrbuf2);
537   }
538 
539 Error:
540   aom_free(arrbuf);
541   aom_free(arrbuf2);
542   return mem_status;
543 }
544 
av1_resize_horz_dir_c(const uint8_t * const input,int in_stride,uint8_t * intbuf,int height,int filtered_length,int width2)545 void av1_resize_horz_dir_c(const uint8_t *const input, int in_stride,
546                            uint8_t *intbuf, int height, int filtered_length,
547                            int width2) {
548   for (int i = 0; i < height; ++i)
549     down2_symeven(input + in_stride * i, filtered_length, intbuf + width2 * i,
550                   0);
551 }
552 
av1_resize_plane_to_half(const uint8_t * const input,int height,int width,int in_stride,uint8_t * output,int height2,int width2,int out_stride)553 bool av1_resize_plane_to_half(const uint8_t *const input, int height, int width,
554                               int in_stride, uint8_t *output, int height2,
555                               int width2, int out_stride) {
556   uint8_t *intbuf = (uint8_t *)aom_malloc(sizeof(*intbuf) * width2 * height);
557   if (intbuf == NULL) {
558     return false;
559   }
560 
561   // Resize in the horizontal direction
562   av1_resize_horz_dir(input, in_stride, intbuf, height, width, width2);
563   // Resize in the vertical direction
564   bool mem_status = av1_resize_vert_dir(intbuf, output, out_stride, height,
565                                         height2, width2, 0 /*start_col*/);
566   aom_free(intbuf);
567   return mem_status;
568 }
569 
570 // Check if both the output width and height are half of input width and
571 // height respectively.
should_resize_by_half(int height,int width,int height2,int width2)572 bool should_resize_by_half(int height, int width, int height2, int width2) {
573   const bool is_width_by_2 = get_down2_length(width, 1) == width2;
574   const bool is_height_by_2 = get_down2_length(height, 1) == height2;
575   return (is_width_by_2 && is_height_by_2);
576 }
577 
av1_resize_plane(const uint8_t * input,int height,int width,int in_stride,uint8_t * output,int height2,int width2,int out_stride)578 bool av1_resize_plane(const uint8_t *input, int height, int width,
579                       int in_stride, uint8_t *output, int height2, int width2,
580                       int out_stride) {
581   int i;
582   bool mem_status = true;
583   uint8_t *intbuf = (uint8_t *)aom_malloc(sizeof(uint8_t) * width2 * height);
584   uint8_t *tmpbuf =
585       (uint8_t *)aom_malloc(sizeof(uint8_t) * AOMMAX(width, height));
586   uint8_t *arrbuf = (uint8_t *)aom_malloc(sizeof(uint8_t) * height);
587   uint8_t *arrbuf2 = (uint8_t *)aom_malloc(sizeof(uint8_t) * height2);
588   if (intbuf == NULL || tmpbuf == NULL || arrbuf == NULL || arrbuf2 == NULL) {
589     mem_status = false;
590     goto Error;
591   }
592   assert(width > 0);
593   assert(height > 0);
594   assert(width2 > 0);
595   assert(height2 > 0);
596   for (i = 0; i < height; ++i)
597     resize_multistep(input + in_stride * i, width, intbuf + width2 * i, width2,
598                      tmpbuf);
599   for (i = 0; i < width2; ++i) {
600     fill_col_to_arr(intbuf + i, width2, height, arrbuf);
601     resize_multistep(arrbuf, height, arrbuf2, height2, tmpbuf);
602     fill_arr_to_col(output + i, out_stride, height2, arrbuf2);
603   }
604 
605 Error:
606   aom_free(intbuf);
607   aom_free(tmpbuf);
608   aom_free(arrbuf);
609   aom_free(arrbuf2);
610   return mem_status;
611 }
612 
upscale_normative_rect(const uint8_t * const input,int height,int width,int in_stride,uint8_t * output,int height2,int width2,int out_stride,int x_step_qn,int x0_qn,int pad_left,int pad_right)613 static bool upscale_normative_rect(const uint8_t *const input, int height,
614                                    int width, int in_stride, uint8_t *output,
615                                    int height2, int width2, int out_stride,
616                                    int x_step_qn, int x0_qn, int pad_left,
617                                    int pad_right) {
618   assert(width > 0);
619   assert(height > 0);
620   assert(width2 > 0);
621   assert(height2 > 0);
622   assert(height2 == height);
623 
624   // Extend the left/right pixels of the tile column if needed
625   // (either because we can't sample from other tiles, or because we're at
626   // a frame edge).
627   // Save the overwritten pixels into tmp_left and tmp_right.
628   // Note: Because we pass input-1 to av1_convolve_horiz_rs, we need one extra
629   // column of border pixels compared to what we'd naively think.
630   const int border_cols = UPSCALE_NORMATIVE_TAPS / 2 + 1;
631   uint8_t *tmp_left =
632       NULL;  // Silence spurious "may be used uninitialized" warnings
633   uint8_t *tmp_right = NULL;
634   uint8_t *const in_tl = (uint8_t *)(input - border_cols);  // Cast off 'const'
635   uint8_t *const in_tr = (uint8_t *)(input + width);
636   if (pad_left) {
637     tmp_left = (uint8_t *)aom_malloc(sizeof(*tmp_left) * border_cols * height);
638     if (!tmp_left) return false;
639     for (int i = 0; i < height; i++) {
640       memcpy(tmp_left + i * border_cols, in_tl + i * in_stride, border_cols);
641       memset(in_tl + i * in_stride, input[i * in_stride], border_cols);
642     }
643   }
644   if (pad_right) {
645     tmp_right =
646         (uint8_t *)aom_malloc(sizeof(*tmp_right) * border_cols * height);
647     if (!tmp_right) {
648       aom_free(tmp_left);
649       return false;
650     }
651     for (int i = 0; i < height; i++) {
652       memcpy(tmp_right + i * border_cols, in_tr + i * in_stride, border_cols);
653       memset(in_tr + i * in_stride, input[i * in_stride + width - 1],
654              border_cols);
655     }
656   }
657 
658   av1_convolve_horiz_rs(input - 1, in_stride, output, out_stride, width2,
659                         height2, &av1_resize_filter_normative[0][0], x0_qn,
660                         x_step_qn);
661 
662   // Restore the left/right border pixels
663   if (pad_left) {
664     for (int i = 0; i < height; i++) {
665       memcpy(in_tl + i * in_stride, tmp_left + i * border_cols, border_cols);
666     }
667     aom_free(tmp_left);
668   }
669   if (pad_right) {
670     for (int i = 0; i < height; i++) {
671       memcpy(in_tr + i * in_stride, tmp_right + i * border_cols, border_cols);
672     }
673     aom_free(tmp_right);
674   }
675   return true;
676 }
677 
678 #if CONFIG_AV1_HIGHBITDEPTH
highbd_interpolate_core(const uint16_t * const input,int in_length,uint16_t * output,int out_length,int bd,const int16_t * interp_filters,int interp_taps)679 static void highbd_interpolate_core(const uint16_t *const input, int in_length,
680                                     uint16_t *output, int out_length, int bd,
681                                     const int16_t *interp_filters,
682                                     int interp_taps) {
683   const int32_t delta =
684       (((uint32_t)in_length << RS_SCALE_SUBPEL_BITS) + out_length / 2) /
685       out_length;
686   const int32_t offset =
687       in_length > out_length
688           ? (((int32_t)(in_length - out_length) << (RS_SCALE_SUBPEL_BITS - 1)) +
689              out_length / 2) /
690                 out_length
691           : -(((int32_t)(out_length - in_length)
692                << (RS_SCALE_SUBPEL_BITS - 1)) +
693               out_length / 2) /
694                 out_length;
695   uint16_t *optr = output;
696   int x, x1, x2, sum, k, int_pel, sub_pel;
697   int32_t y;
698 
699   x = 0;
700   y = offset + RS_SCALE_EXTRA_OFF;
701   while ((y >> RS_SCALE_SUBPEL_BITS) < (interp_taps / 2 - 1)) {
702     x++;
703     y += delta;
704   }
705   x1 = x;
706   x = out_length - 1;
707   y = delta * x + offset + RS_SCALE_EXTRA_OFF;
708   while ((y >> RS_SCALE_SUBPEL_BITS) + (int32_t)(interp_taps / 2) >=
709          in_length) {
710     x--;
711     y -= delta;
712   }
713   x2 = x;
714   if (x1 > x2) {
715     for (x = 0, y = offset + RS_SCALE_EXTRA_OFF; x < out_length;
716          ++x, y += delta) {
717       int_pel = y >> RS_SCALE_SUBPEL_BITS;
718       sub_pel = (y >> RS_SCALE_EXTRA_BITS) & RS_SUBPEL_MASK;
719       const int16_t *filter = &interp_filters[sub_pel * interp_taps];
720       sum = 0;
721       for (k = 0; k < interp_taps; ++k) {
722         const int pk = int_pel - interp_taps / 2 + 1 + k;
723         sum += filter[k] * input[AOMMAX(AOMMIN(pk, in_length - 1), 0)];
724       }
725       *optr++ = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
726     }
727   } else {
728     // Initial part.
729     for (x = 0, y = offset + RS_SCALE_EXTRA_OFF; x < x1; ++x, y += delta) {
730       int_pel = y >> RS_SCALE_SUBPEL_BITS;
731       sub_pel = (y >> RS_SCALE_EXTRA_BITS) & RS_SUBPEL_MASK;
732       const int16_t *filter = &interp_filters[sub_pel * interp_taps];
733       sum = 0;
734       for (k = 0; k < interp_taps; ++k)
735         sum += filter[k] * input[AOMMAX(int_pel - interp_taps / 2 + 1 + k, 0)];
736       *optr++ = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
737     }
738     // Middle part.
739     for (; x <= x2; ++x, y += delta) {
740       int_pel = y >> RS_SCALE_SUBPEL_BITS;
741       sub_pel = (y >> RS_SCALE_EXTRA_BITS) & RS_SUBPEL_MASK;
742       const int16_t *filter = &interp_filters[sub_pel * interp_taps];
743       sum = 0;
744       for (k = 0; k < interp_taps; ++k)
745         sum += filter[k] * input[int_pel - interp_taps / 2 + 1 + k];
746       *optr++ = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
747     }
748     // End part.
749     for (; x < out_length; ++x, y += delta) {
750       int_pel = y >> RS_SCALE_SUBPEL_BITS;
751       sub_pel = (y >> RS_SCALE_EXTRA_BITS) & RS_SUBPEL_MASK;
752       const int16_t *filter = &interp_filters[sub_pel * interp_taps];
753       sum = 0;
754       for (k = 0; k < interp_taps; ++k)
755         sum += filter[k] *
756                input[AOMMIN(int_pel - interp_taps / 2 + 1 + k, in_length - 1)];
757       *optr++ = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
758     }
759   }
760 }
761 
highbd_interpolate(const uint16_t * const input,int in_length,uint16_t * output,int out_length,int bd)762 static void highbd_interpolate(const uint16_t *const input, int in_length,
763                                uint16_t *output, int out_length, int bd) {
764   const InterpKernel *interp_filters =
765       choose_interp_filter(in_length, out_length);
766 
767   highbd_interpolate_core(input, in_length, output, out_length, bd,
768                           &interp_filters[0][0], SUBPEL_TAPS);
769 }
770 
highbd_down2_symeven(const uint16_t * const input,int length,uint16_t * output,int bd)771 static void highbd_down2_symeven(const uint16_t *const input, int length,
772                                  uint16_t *output, int bd) {
773   // Actual filter len = 2 * filter_len_half.
774   static const int16_t *filter = av1_down2_symeven_half_filter;
775   const int filter_len_half = sizeof(av1_down2_symeven_half_filter) / 2;
776   int i, j;
777   uint16_t *optr = output;
778   int l1 = filter_len_half;
779   int l2 = (length - filter_len_half);
780   l1 += (l1 & 1);
781   l2 += (l2 & 1);
782   if (l1 > l2) {
783     // Short input length.
784     for (i = 0; i < length; i += 2) {
785       int sum = (1 << (FILTER_BITS - 1));
786       for (j = 0; j < filter_len_half; ++j) {
787         sum +=
788             (input[AOMMAX(0, i - j)] + input[AOMMIN(i + 1 + j, length - 1)]) *
789             filter[j];
790       }
791       sum >>= FILTER_BITS;
792       *optr++ = clip_pixel_highbd(sum, bd);
793     }
794   } else {
795     // Initial part.
796     for (i = 0; i < l1; i += 2) {
797       int sum = (1 << (FILTER_BITS - 1));
798       for (j = 0; j < filter_len_half; ++j) {
799         sum += (input[AOMMAX(0, i - j)] + input[i + 1 + j]) * filter[j];
800       }
801       sum >>= FILTER_BITS;
802       *optr++ = clip_pixel_highbd(sum, bd);
803     }
804     // Middle part.
805     for (; i < l2; i += 2) {
806       int sum = (1 << (FILTER_BITS - 1));
807       for (j = 0; j < filter_len_half; ++j) {
808         sum += (input[i - j] + input[i + 1 + j]) * filter[j];
809       }
810       sum >>= FILTER_BITS;
811       *optr++ = clip_pixel_highbd(sum, bd);
812     }
813     // End part.
814     for (; i < length; i += 2) {
815       int sum = (1 << (FILTER_BITS - 1));
816       for (j = 0; j < filter_len_half; ++j) {
817         sum +=
818             (input[i - j] + input[AOMMIN(i + 1 + j, length - 1)]) * filter[j];
819       }
820       sum >>= FILTER_BITS;
821       *optr++ = clip_pixel_highbd(sum, bd);
822     }
823   }
824 }
825 
highbd_down2_symodd(const uint16_t * const input,int length,uint16_t * output,int bd)826 static void highbd_down2_symodd(const uint16_t *const input, int length,
827                                 uint16_t *output, int bd) {
828   // Actual filter len = 2 * filter_len_half - 1.
829   static const int16_t *filter = av1_down2_symodd_half_filter;
830   const int filter_len_half = sizeof(av1_down2_symodd_half_filter) / 2;
831   int i, j;
832   uint16_t *optr = output;
833   int l1 = filter_len_half - 1;
834   int l2 = (length - filter_len_half + 1);
835   l1 += (l1 & 1);
836   l2 += (l2 & 1);
837   if (l1 > l2) {
838     // Short input length.
839     for (i = 0; i < length; i += 2) {
840       int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
841       for (j = 1; j < filter_len_half; ++j) {
842         sum += (input[AOMMAX(i - j, 0)] + input[AOMMIN(i + j, length - 1)]) *
843                filter[j];
844       }
845       sum >>= FILTER_BITS;
846       *optr++ = clip_pixel_highbd(sum, bd);
847     }
848   } else {
849     // Initial part.
850     for (i = 0; i < l1; i += 2) {
851       int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
852       for (j = 1; j < filter_len_half; ++j) {
853         sum += (input[AOMMAX(i - j, 0)] + input[i + j]) * filter[j];
854       }
855       sum >>= FILTER_BITS;
856       *optr++ = clip_pixel_highbd(sum, bd);
857     }
858     // Middle part.
859     for (; i < l2; i += 2) {
860       int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
861       for (j = 1; j < filter_len_half; ++j) {
862         sum += (input[i - j] + input[i + j]) * filter[j];
863       }
864       sum >>= FILTER_BITS;
865       *optr++ = clip_pixel_highbd(sum, bd);
866     }
867     // End part.
868     for (; i < length; i += 2) {
869       int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
870       for (j = 1; j < filter_len_half; ++j) {
871         sum += (input[i - j] + input[AOMMIN(i + j, length - 1)]) * filter[j];
872       }
873       sum >>= FILTER_BITS;
874       *optr++ = clip_pixel_highbd(sum, bd);
875     }
876   }
877 }
878 
highbd_resize_multistep(const uint16_t * const input,int length,uint16_t * output,int olength,uint16_t * otmp,int bd)879 static void highbd_resize_multistep(const uint16_t *const input, int length,
880                                     uint16_t *output, int olength,
881                                     uint16_t *otmp, int bd) {
882   if (length == olength) {
883     memcpy(output, input, sizeof(output[0]) * length);
884     return;
885   }
886   const int steps = get_down2_steps(length, olength);
887 
888   if (steps > 0) {
889     uint16_t *out = NULL;
890     int filteredlength = length;
891 
892     assert(otmp != NULL);
893     uint16_t *otmp2 = otmp + get_down2_length(length, 1);
894     for (int s = 0; s < steps; ++s) {
895       const int proj_filteredlength = get_down2_length(filteredlength, 1);
896       const uint16_t *const in = (s == 0 ? input : out);
897       if (s == steps - 1 && proj_filteredlength == olength)
898         out = output;
899       else
900         out = (s & 1 ? otmp2 : otmp);
901       if (filteredlength & 1)
902         highbd_down2_symodd(in, filteredlength, out, bd);
903       else
904         highbd_down2_symeven(in, filteredlength, out, bd);
905       filteredlength = proj_filteredlength;
906     }
907     if (filteredlength != olength) {
908       highbd_interpolate(out, filteredlength, output, olength, bd);
909     }
910   } else {
911     highbd_interpolate(input, length, output, olength, bd);
912   }
913 }
914 
highbd_fill_col_to_arr(uint16_t * img,int stride,int len,uint16_t * arr)915 static void highbd_fill_col_to_arr(uint16_t *img, int stride, int len,
916                                    uint16_t *arr) {
917   int i;
918   uint16_t *iptr = img;
919   uint16_t *aptr = arr;
920   for (i = 0; i < len; ++i, iptr += stride) {
921     *aptr++ = *iptr;
922   }
923 }
924 
highbd_fill_arr_to_col(uint16_t * img,int stride,int len,uint16_t * arr)925 static void highbd_fill_arr_to_col(uint16_t *img, int stride, int len,
926                                    uint16_t *arr) {
927   int i;
928   uint16_t *iptr = img;
929   uint16_t *aptr = arr;
930   for (i = 0; i < len; ++i, iptr += stride) {
931     *iptr = *aptr++;
932   }
933 }
934 
av1_highbd_resize_plane(const uint8_t * input,int height,int width,int in_stride,uint8_t * output,int height2,int width2,int out_stride,int bd)935 void av1_highbd_resize_plane(const uint8_t *input, int height, int width,
936                              int in_stride, uint8_t *output, int height2,
937                              int width2, int out_stride, int bd) {
938   int i;
939   uint16_t *intbuf = (uint16_t *)aom_malloc(sizeof(uint16_t) * width2 * height);
940   uint16_t *tmpbuf =
941       (uint16_t *)aom_malloc(sizeof(uint16_t) * AOMMAX(width, height));
942   uint16_t *arrbuf = (uint16_t *)aom_malloc(sizeof(uint16_t) * height);
943   uint16_t *arrbuf2 = (uint16_t *)aom_malloc(sizeof(uint16_t) * height2);
944   if (intbuf == NULL || tmpbuf == NULL || arrbuf == NULL || arrbuf2 == NULL)
945     goto Error;
946   for (i = 0; i < height; ++i) {
947     highbd_resize_multistep(CONVERT_TO_SHORTPTR(input + in_stride * i), width,
948                             intbuf + width2 * i, width2, tmpbuf, bd);
949   }
950   for (i = 0; i < width2; ++i) {
951     highbd_fill_col_to_arr(intbuf + i, width2, height, arrbuf);
952     highbd_resize_multistep(arrbuf, height, arrbuf2, height2, tmpbuf, bd);
953     highbd_fill_arr_to_col(CONVERT_TO_SHORTPTR(output + i), out_stride, height2,
954                            arrbuf2);
955   }
956 
957 Error:
958   aom_free(intbuf);
959   aom_free(tmpbuf);
960   aom_free(arrbuf);
961   aom_free(arrbuf2);
962 }
963 
highbd_upscale_normative_rect(const uint8_t * const input,int height,int width,int in_stride,uint8_t * output,int height2,int width2,int out_stride,int x_step_qn,int x0_qn,int pad_left,int pad_right,int bd)964 static bool highbd_upscale_normative_rect(const uint8_t *const input,
965                                           int height, int width, int in_stride,
966                                           uint8_t *output, int height2,
967                                           int width2, int out_stride,
968                                           int x_step_qn, int x0_qn,
969                                           int pad_left, int pad_right, int bd) {
970   assert(width > 0);
971   assert(height > 0);
972   assert(width2 > 0);
973   assert(height2 > 0);
974   assert(height2 == height);
975 
976   // Extend the left/right pixels of the tile column if needed
977   // (either because we can't sample from other tiles, or because we're at
978   // a frame edge).
979   // Save the overwritten pixels into tmp_left and tmp_right.
980   // Note: Because we pass input-1 to av1_convolve_horiz_rs, we need one extra
981   // column of border pixels compared to what we'd naively think.
982   const int border_cols = UPSCALE_NORMATIVE_TAPS / 2 + 1;
983   const int border_size = border_cols * sizeof(uint16_t);
984   uint16_t *tmp_left =
985       NULL;  // Silence spurious "may be used uninitialized" warnings
986   uint16_t *tmp_right = NULL;
987   uint16_t *const input16 = CONVERT_TO_SHORTPTR(input);
988   uint16_t *const in_tl = input16 - border_cols;
989   uint16_t *const in_tr = input16 + width;
990   if (pad_left) {
991     tmp_left = (uint16_t *)aom_malloc(sizeof(*tmp_left) * border_cols * height);
992     if (!tmp_left) return false;
993     for (int i = 0; i < height; i++) {
994       memcpy(tmp_left + i * border_cols, in_tl + i * in_stride, border_size);
995       aom_memset16(in_tl + i * in_stride, input16[i * in_stride], border_cols);
996     }
997   }
998   if (pad_right) {
999     tmp_right =
1000         (uint16_t *)aom_malloc(sizeof(*tmp_right) * border_cols * height);
1001     if (!tmp_right) {
1002       aom_free(tmp_left);
1003       return false;
1004     }
1005     for (int i = 0; i < height; i++) {
1006       memcpy(tmp_right + i * border_cols, in_tr + i * in_stride, border_size);
1007       aom_memset16(in_tr + i * in_stride, input16[i * in_stride + width - 1],
1008                    border_cols);
1009     }
1010   }
1011 
1012   av1_highbd_convolve_horiz_rs(CONVERT_TO_SHORTPTR(input - 1), in_stride,
1013                                CONVERT_TO_SHORTPTR(output), out_stride, width2,
1014                                height2, &av1_resize_filter_normative[0][0],
1015                                x0_qn, x_step_qn, bd);
1016 
1017   // Restore the left/right border pixels
1018   if (pad_left) {
1019     for (int i = 0; i < height; i++) {
1020       memcpy(in_tl + i * in_stride, tmp_left + i * border_cols, border_size);
1021     }
1022     aom_free(tmp_left);
1023   }
1024   if (pad_right) {
1025     for (int i = 0; i < height; i++) {
1026       memcpy(in_tr + i * in_stride, tmp_right + i * border_cols, border_size);
1027     }
1028     aom_free(tmp_right);
1029   }
1030   return true;
1031 }
1032 #endif  // CONFIG_AV1_HIGHBITDEPTH
1033 
av1_resize_and_extend_frame_c(const YV12_BUFFER_CONFIG * src,YV12_BUFFER_CONFIG * dst,const InterpFilter filter,const int phase_scaler,const int num_planes)1034 void av1_resize_and_extend_frame_c(const YV12_BUFFER_CONFIG *src,
1035                                    YV12_BUFFER_CONFIG *dst,
1036                                    const InterpFilter filter,
1037                                    const int phase_scaler,
1038                                    const int num_planes) {
1039   assert(filter == BILINEAR || filter == EIGHTTAP_SMOOTH ||
1040          filter == EIGHTTAP_REGULAR);
1041   const InterpKernel *const kernel =
1042       (const InterpKernel *)av1_interp_filter_params_list[filter].filter_ptr;
1043 
1044   for (int i = 0; i < AOMMIN(num_planes, MAX_MB_PLANE); ++i) {
1045     const int is_uv = i > 0;
1046     const int src_w = src->crop_widths[is_uv];
1047     const int src_h = src->crop_heights[is_uv];
1048     const uint8_t *src_buffer = src->buffers[i];
1049     const int src_stride = src->strides[is_uv];
1050     const int dst_w = dst->crop_widths[is_uv];
1051     const int dst_h = dst->crop_heights[is_uv];
1052     uint8_t *dst_buffer = dst->buffers[i];
1053     const int dst_stride = dst->strides[is_uv];
1054     for (int y = 0; y < dst_h; y += 16) {
1055       const int y_q4 =
1056           src_h == dst_h ? 0 : y * 16 * src_h / dst_h + phase_scaler;
1057       for (int x = 0; x < dst_w; x += 16) {
1058         const int x_q4 =
1059             src_w == dst_w ? 0 : x * 16 * src_w / dst_w + phase_scaler;
1060         const uint8_t *src_ptr =
1061             src_buffer + y * src_h / dst_h * src_stride + x * src_w / dst_w;
1062         uint8_t *dst_ptr = dst_buffer + y * dst_stride + x;
1063 
1064         // Width and height of the actual working area.
1065         const int work_w = AOMMIN(16, dst_w - x);
1066         const int work_h = AOMMIN(16, dst_h - y);
1067         // SIMD versions of aom_scaled_2d() have some trouble handling
1068         // nonstandard sizes, so fall back on the C version to handle borders.
1069         if (work_w != 16 || work_h != 16) {
1070           aom_scaled_2d_c(src_ptr, src_stride, dst_ptr, dst_stride, kernel,
1071                           x_q4 & 0xf, 16 * src_w / dst_w, y_q4 & 0xf,
1072                           16 * src_h / dst_h, work_w, work_h);
1073         } else {
1074           aom_scaled_2d(src_ptr, src_stride, dst_ptr, dst_stride, kernel,
1075                         x_q4 & 0xf, 16 * src_w / dst_w, y_q4 & 0xf,
1076                         16 * src_h / dst_h, 16, 16);
1077         }
1078       }
1079     }
1080   }
1081   aom_extend_frame_borders(dst, num_planes);
1082 }
1083 
av1_resize_and_extend_frame_nonnormative(const YV12_BUFFER_CONFIG * src,YV12_BUFFER_CONFIG * dst,int bd,int num_planes)1084 bool av1_resize_and_extend_frame_nonnormative(const YV12_BUFFER_CONFIG *src,
1085                                               YV12_BUFFER_CONFIG *dst, int bd,
1086                                               int num_planes) {
1087   // TODO(dkovalev): replace YV12_BUFFER_CONFIG with aom_image_t
1088 
1089   // We use AOMMIN(num_planes, MAX_MB_PLANE) instead of num_planes to quiet
1090   // the static analysis warnings.
1091   for (int i = 0; i < AOMMIN(num_planes, MAX_MB_PLANE); ++i) {
1092     const int is_uv = i > 0;
1093 #if CONFIG_AV1_HIGHBITDEPTH
1094     if (src->flags & YV12_FLAG_HIGHBITDEPTH) {
1095       av1_highbd_resize_plane(src->buffers[i], src->crop_heights[is_uv],
1096                               src->crop_widths[is_uv], src->strides[is_uv],
1097                               dst->buffers[i], dst->crop_heights[is_uv],
1098                               dst->crop_widths[is_uv], dst->strides[is_uv], bd);
1099     } else if (!av1_resize_plane(src->buffers[i], src->crop_heights[is_uv],
1100                                  src->crop_widths[is_uv], src->strides[is_uv],
1101                                  dst->buffers[i], dst->crop_heights[is_uv],
1102                                  dst->crop_widths[is_uv],
1103                                  dst->strides[is_uv])) {
1104       return false;
1105     }
1106 #else
1107     (void)bd;
1108     if (!av1_resize_plane(src->buffers[i], src->crop_heights[is_uv],
1109                           src->crop_widths[is_uv], src->strides[is_uv],
1110                           dst->buffers[i], dst->crop_heights[is_uv],
1111                           dst->crop_widths[is_uv], dst->strides[is_uv]))
1112       return false;
1113 #endif
1114   }
1115   aom_extend_frame_borders(dst, num_planes);
1116   return true;
1117 }
1118 
av1_upscale_normative_rows(const AV1_COMMON * cm,const uint8_t * src,int src_stride,uint8_t * dst,int dst_stride,int plane,int rows)1119 void av1_upscale_normative_rows(const AV1_COMMON *cm, const uint8_t *src,
1120                                 int src_stride, uint8_t *dst, int dst_stride,
1121                                 int plane, int rows) {
1122   const int is_uv = (plane > 0);
1123   const int ss_x = is_uv && cm->seq_params->subsampling_x;
1124   const int downscaled_plane_width = ROUND_POWER_OF_TWO(cm->width, ss_x);
1125   const int upscaled_plane_width =
1126       ROUND_POWER_OF_TWO(cm->superres_upscaled_width, ss_x);
1127   const int superres_denom = cm->superres_scale_denominator;
1128 
1129   TileInfo tile_col;
1130   const int32_t x_step_qn = av1_get_upscale_convolve_step(
1131       downscaled_plane_width, upscaled_plane_width);
1132   int32_t x0_qn = get_upscale_convolve_x0(downscaled_plane_width,
1133                                           upscaled_plane_width, x_step_qn);
1134 
1135   for (int j = 0; j < cm->tiles.cols; j++) {
1136     av1_tile_set_col(&tile_col, cm, j);
1137     // Determine the limits of this tile column in both the source
1138     // and destination images.
1139     // Note: The actual location which we start sampling from is
1140     // (downscaled_x0 - 1 + (x0_qn/2^14)), and this quantity increases
1141     // by exactly dst_width * (x_step_qn/2^14) pixels each iteration.
1142     const int downscaled_x0 = tile_col.mi_col_start << (MI_SIZE_LOG2 - ss_x);
1143     const int downscaled_x1 = tile_col.mi_col_end << (MI_SIZE_LOG2 - ss_x);
1144     const int src_width = downscaled_x1 - downscaled_x0;
1145 
1146     const int upscaled_x0 = (downscaled_x0 * superres_denom) / SCALE_NUMERATOR;
1147     int upscaled_x1;
1148     if (j == cm->tiles.cols - 1) {
1149       // Note that we can't just use AOMMIN here - due to rounding,
1150       // (downscaled_x1 * superres_denom) / SCALE_NUMERATOR may be less than
1151       // upscaled_plane_width.
1152       upscaled_x1 = upscaled_plane_width;
1153     } else {
1154       upscaled_x1 = (downscaled_x1 * superres_denom) / SCALE_NUMERATOR;
1155     }
1156 
1157     const uint8_t *const src_ptr = src + downscaled_x0;
1158     uint8_t *const dst_ptr = dst + upscaled_x0;
1159     const int dst_width = upscaled_x1 - upscaled_x0;
1160 
1161     const int pad_left = (j == 0);
1162     const int pad_right = (j == cm->tiles.cols - 1);
1163 
1164     bool success;
1165 #if CONFIG_AV1_HIGHBITDEPTH
1166     if (cm->seq_params->use_highbitdepth)
1167       success = highbd_upscale_normative_rect(
1168           src_ptr, rows, src_width, src_stride, dst_ptr, rows, dst_width,
1169           dst_stride, x_step_qn, x0_qn, pad_left, pad_right,
1170           cm->seq_params->bit_depth);
1171     else
1172       success = upscale_normative_rect(src_ptr, rows, src_width, src_stride,
1173                                        dst_ptr, rows, dst_width, dst_stride,
1174                                        x_step_qn, x0_qn, pad_left, pad_right);
1175 #else
1176     success = upscale_normative_rect(src_ptr, rows, src_width, src_stride,
1177                                      dst_ptr, rows, dst_width, dst_stride,
1178                                      x_step_qn, x0_qn, pad_left, pad_right);
1179 #endif
1180     if (!success) {
1181       aom_internal_error(cm->error, AOM_CODEC_MEM_ERROR,
1182                          "Error upscaling frame");
1183     }
1184     // Update the fractional pixel offset to prepare for the next tile column.
1185     x0_qn += (dst_width * x_step_qn) - (src_width << RS_SCALE_SUBPEL_BITS);
1186   }
1187 }
1188 
upscale_normative_and_extend_frame(const AV1_COMMON * cm,const YV12_BUFFER_CONFIG * src,YV12_BUFFER_CONFIG * dst)1189 static void upscale_normative_and_extend_frame(const AV1_COMMON *cm,
1190                                                const YV12_BUFFER_CONFIG *src,
1191                                                YV12_BUFFER_CONFIG *dst) {
1192   const int num_planes = av1_num_planes(cm);
1193   for (int i = 0; i < num_planes; ++i) {
1194     const int is_uv = (i > 0);
1195     av1_upscale_normative_rows(cm, src->buffers[i], src->strides[is_uv],
1196                                dst->buffers[i], dst->strides[is_uv], i,
1197                                src->crop_heights[is_uv]);
1198   }
1199 
1200   aom_extend_frame_borders(dst, num_planes);
1201 }
1202 
av1_realloc_and_scale_if_required(AV1_COMMON * cm,YV12_BUFFER_CONFIG * unscaled,YV12_BUFFER_CONFIG * scaled,const InterpFilter filter,const int phase,const bool use_optimized_scaler,const bool for_psnr,const int border_in_pixels,const bool alloc_pyramid)1203 YV12_BUFFER_CONFIG *av1_realloc_and_scale_if_required(
1204     AV1_COMMON *cm, YV12_BUFFER_CONFIG *unscaled, YV12_BUFFER_CONFIG *scaled,
1205     const InterpFilter filter, const int phase, const bool use_optimized_scaler,
1206     const bool for_psnr, const int border_in_pixels, const bool alloc_pyramid) {
1207   // If scaling is performed for the sole purpose of calculating PSNR, then our
1208   // target dimensions are superres upscaled width/height. Otherwise our target
1209   // dimensions are coded width/height.
1210   const int scaled_width = for_psnr ? cm->superres_upscaled_width : cm->width;
1211   const int scaled_height =
1212       for_psnr ? cm->superres_upscaled_height : cm->height;
1213   const bool scaling_required = (scaled_width != unscaled->y_crop_width) ||
1214                                 (scaled_height != unscaled->y_crop_height);
1215 
1216   if (scaling_required) {
1217     const int num_planes = av1_num_planes(cm);
1218     const SequenceHeader *seq_params = cm->seq_params;
1219 
1220     // Reallocate the frame buffer based on the target dimensions when scaling
1221     // is required.
1222     if (aom_realloc_frame_buffer(
1223             scaled, scaled_width, scaled_height, seq_params->subsampling_x,
1224             seq_params->subsampling_y, seq_params->use_highbitdepth,
1225             border_in_pixels, cm->features.byte_alignment, NULL, NULL, NULL,
1226             alloc_pyramid, 0))
1227       aom_internal_error(cm->error, AOM_CODEC_MEM_ERROR,
1228                          "Failed to allocate scaled buffer");
1229 
1230     bool has_optimized_scaler = av1_has_optimized_scaler(
1231         unscaled->y_crop_width, unscaled->y_crop_height, scaled_width,
1232         scaled_height);
1233     if (num_planes > 1) {
1234       has_optimized_scaler = has_optimized_scaler &&
1235                              av1_has_optimized_scaler(unscaled->uv_crop_width,
1236                                                       unscaled->uv_crop_height,
1237                                                       scaled->uv_crop_width,
1238                                                       scaled->uv_crop_height);
1239     }
1240 
1241 #if CONFIG_AV1_HIGHBITDEPTH
1242     if (use_optimized_scaler && has_optimized_scaler &&
1243         cm->seq_params->bit_depth == AOM_BITS_8) {
1244       av1_resize_and_extend_frame(unscaled, scaled, filter, phase, num_planes);
1245     } else {
1246       if (!av1_resize_and_extend_frame_nonnormative(
1247               unscaled, scaled, (int)cm->seq_params->bit_depth, num_planes))
1248         aom_internal_error(cm->error, AOM_CODEC_MEM_ERROR,
1249                            "Failed to allocate buffers during resize");
1250     }
1251 #else
1252     if (use_optimized_scaler && has_optimized_scaler) {
1253       av1_resize_and_extend_frame(unscaled, scaled, filter, phase, num_planes);
1254     } else {
1255       if (!av1_resize_and_extend_frame_nonnormative(
1256               unscaled, scaled, (int)cm->seq_params->bit_depth, num_planes))
1257         aom_internal_error(cm->error, AOM_CODEC_MEM_ERROR,
1258                            "Failed to allocate buffers during resize");
1259     }
1260 #endif
1261     return scaled;
1262   }
1263   return unscaled;
1264 }
1265 
1266 // Calculates the scaled dimension given the original dimension and the scale
1267 // denominator.
calculate_scaled_size_helper(int * dim,int denom)1268 static void calculate_scaled_size_helper(int *dim, int denom) {
1269   if (denom != SCALE_NUMERATOR) {
1270     // We need to ensure the constraint in "Appendix A" of the spec:
1271     // * FrameWidth is greater than or equal to 16
1272     // * FrameHeight is greater than or equal to 16
1273     // For this, we clamp the downscaled dimension to at least 16. One
1274     // exception: if original dimension itself was < 16, then we keep the
1275     // downscaled dimension to be same as the original, to ensure that resizing
1276     // is valid.
1277     const int min_dim = AOMMIN(16, *dim);
1278     // Use this version if we need *dim to be even
1279     // *width = (*width * SCALE_NUMERATOR + denom) / (2 * denom);
1280     // *width <<= 1;
1281     *dim = (*dim * SCALE_NUMERATOR + denom / 2) / (denom);
1282     *dim = AOMMAX(*dim, min_dim);
1283   }
1284 }
1285 
av1_calculate_scaled_size(int * width,int * height,int resize_denom)1286 void av1_calculate_scaled_size(int *width, int *height, int resize_denom) {
1287   calculate_scaled_size_helper(width, resize_denom);
1288   calculate_scaled_size_helper(height, resize_denom);
1289 }
1290 
av1_calculate_scaled_superres_size(int * width,int * height,int superres_denom)1291 void av1_calculate_scaled_superres_size(int *width, int *height,
1292                                         int superres_denom) {
1293   (void)height;
1294   calculate_scaled_size_helper(width, superres_denom);
1295 }
1296 
av1_calculate_unscaled_superres_size(int * width,int * height,int denom)1297 void av1_calculate_unscaled_superres_size(int *width, int *height, int denom) {
1298   if (denom != SCALE_NUMERATOR) {
1299     // Note: av1_calculate_scaled_superres_size() rounds *up* after division
1300     // when the resulting dimensions are odd. So here, we round *down*.
1301     *width = *width * denom / SCALE_NUMERATOR;
1302     (void)height;
1303   }
1304 }
1305 
1306 // Copy only the config data from 'src' to 'dst'.
copy_buffer_config(const YV12_BUFFER_CONFIG * const src,YV12_BUFFER_CONFIG * const dst)1307 static void copy_buffer_config(const YV12_BUFFER_CONFIG *const src,
1308                                YV12_BUFFER_CONFIG *const dst) {
1309   dst->bit_depth = src->bit_depth;
1310   dst->color_primaries = src->color_primaries;
1311   dst->transfer_characteristics = src->transfer_characteristics;
1312   dst->matrix_coefficients = src->matrix_coefficients;
1313   dst->monochrome = src->monochrome;
1314   dst->chroma_sample_position = src->chroma_sample_position;
1315   dst->color_range = src->color_range;
1316 }
1317 
1318 // TODO(afergs): Look for in-place upscaling
1319 // TODO(afergs): aom_ vs av1_ functions? Which can I use?
1320 // Upscale decoded image.
av1_superres_upscale(AV1_COMMON * cm,BufferPool * const pool,bool alloc_pyramid)1321 void av1_superres_upscale(AV1_COMMON *cm, BufferPool *const pool,
1322                           bool alloc_pyramid) {
1323   const int num_planes = av1_num_planes(cm);
1324   if (!av1_superres_scaled(cm)) return;
1325   const SequenceHeader *const seq_params = cm->seq_params;
1326   const int byte_alignment = cm->features.byte_alignment;
1327 
1328   YV12_BUFFER_CONFIG copy_buffer;
1329   memset(&copy_buffer, 0, sizeof(copy_buffer));
1330 
1331   YV12_BUFFER_CONFIG *const frame_to_show = &cm->cur_frame->buf;
1332 
1333   const int aligned_width = ALIGN_POWER_OF_TWO(cm->width, 3);
1334   if (aom_alloc_frame_buffer(
1335           &copy_buffer, aligned_width, cm->height, seq_params->subsampling_x,
1336           seq_params->subsampling_y, seq_params->use_highbitdepth,
1337           AOM_BORDER_IN_PIXELS, byte_alignment, false, 0))
1338     aom_internal_error(cm->error, AOM_CODEC_MEM_ERROR,
1339                        "Failed to allocate copy buffer for superres upscaling");
1340 
1341   // Copy function assumes the frames are the same size.
1342   // Note that it does not copy YV12_BUFFER_CONFIG config data.
1343   aom_yv12_copy_frame(frame_to_show, &copy_buffer, num_planes);
1344 
1345   assert(copy_buffer.y_crop_width == aligned_width);
1346   assert(copy_buffer.y_crop_height == cm->height);
1347 
1348   // Realloc the current frame buffer at a higher resolution in place.
1349   if (pool != NULL) {
1350     // Use callbacks if on the decoder.
1351     aom_codec_frame_buffer_t *fb = &cm->cur_frame->raw_frame_buffer;
1352     aom_release_frame_buffer_cb_fn_t release_fb_cb = pool->release_fb_cb;
1353     aom_get_frame_buffer_cb_fn_t cb = pool->get_fb_cb;
1354     void *cb_priv = pool->cb_priv;
1355 
1356     lock_buffer_pool(pool);
1357     // Realloc with callback does not release the frame buffer - release first.
1358     if (release_fb_cb(cb_priv, fb)) {
1359       unlock_buffer_pool(pool);
1360       aom_internal_error(
1361           cm->error, AOM_CODEC_MEM_ERROR,
1362           "Failed to free current frame buffer before superres upscaling");
1363     }
1364     // aom_realloc_frame_buffer() leaves config data for frame_to_show intact
1365     if (aom_realloc_frame_buffer(
1366             frame_to_show, cm->superres_upscaled_width,
1367             cm->superres_upscaled_height, seq_params->subsampling_x,
1368             seq_params->subsampling_y, seq_params->use_highbitdepth,
1369             AOM_BORDER_IN_PIXELS, byte_alignment, fb, cb, cb_priv,
1370             alloc_pyramid, 0)) {
1371       unlock_buffer_pool(pool);
1372       aom_internal_error(
1373           cm->error, AOM_CODEC_MEM_ERROR,
1374           "Failed to allocate current frame buffer for superres upscaling");
1375     }
1376     unlock_buffer_pool(pool);
1377   } else {
1378     // Make a copy of the config data for frame_to_show in copy_buffer
1379     copy_buffer_config(frame_to_show, &copy_buffer);
1380 
1381     // Don't use callbacks on the encoder.
1382     // aom_alloc_frame_buffer() clears the config data for frame_to_show
1383     if (aom_alloc_frame_buffer(
1384             frame_to_show, cm->superres_upscaled_width,
1385             cm->superres_upscaled_height, seq_params->subsampling_x,
1386             seq_params->subsampling_y, seq_params->use_highbitdepth,
1387             AOM_BORDER_IN_PIXELS, byte_alignment, alloc_pyramid, 0))
1388       aom_internal_error(
1389           cm->error, AOM_CODEC_MEM_ERROR,
1390           "Failed to reallocate current frame buffer for superres upscaling");
1391 
1392     // Restore config data back to frame_to_show
1393     copy_buffer_config(&copy_buffer, frame_to_show);
1394   }
1395   // TODO(afergs): verify frame_to_show is correct after realloc
1396   //               encoder:
1397   //               decoder:
1398 
1399   assert(frame_to_show->y_crop_width == cm->superres_upscaled_width);
1400   assert(frame_to_show->y_crop_height == cm->superres_upscaled_height);
1401 
1402   // Scale up and back into frame_to_show.
1403   assert(frame_to_show->y_crop_width != cm->width);
1404   upscale_normative_and_extend_frame(cm, &copy_buffer, frame_to_show);
1405 
1406   // Free the copy buffer
1407   aom_free_frame_buffer(&copy_buffer);
1408 }
1409