Line data Source code
1 : /*
2 : * Copyright (c) 2017, 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 <tmmintrin.h>
13 :
14 : #include "./av1_rtcd.h"
15 : #include "av1/common/warped_motion.h"
16 :
17 : /* This is a modified version of 'warped_filter' from warped_motion.c:
18 : * Each coefficient is stored in 8 bits instead of 16 bits
19 : * The coefficients are rearranged in the column order 0, 2, 4, 6, 1, 3, 5, 7
20 :
21 : This is done in order to avoid overflow: Since the tap with the largest
22 : coefficient could be any of taps 2, 3, 4 or 5, we can't use the summation
23 : order ((0 + 1) + (4 + 5)) + ((2 + 3) + (6 + 7)) used in the regular
24 : convolve functions.
25 :
26 : Instead, we use the summation order
27 : ((0 + 2) + (4 + 6)) + ((1 + 3) + (5 + 7)).
28 : The rearrangement of coefficients in this table is so that we can get the
29 : coefficients into the correct order more quickly.
30 : */
31 : /* clang-format off */
32 : DECLARE_ALIGNED(8, static const int8_t,
33 : filter_8bit[WARPEDPIXEL_PREC_SHIFTS * 3 + 1][8]) = {
34 : #if WARPEDPIXEL_PREC_BITS == 6
35 : // [-1, 0)
36 : { 0, 127, 0, 0, 0, 1, 0, 0}, { 0, 127, 0, 0, -1, 2, 0, 0},
37 : { 1, 127, -1, 0, -3, 4, 0, 0}, { 1, 126, -2, 0, -4, 6, 1, 0},
38 : { 1, 126, -3, 0, -5, 8, 1, 0}, { 1, 125, -4, 0, -6, 11, 1, 0},
39 : { 1, 124, -4, 0, -7, 13, 1, 0}, { 2, 123, -5, 0, -8, 15, 1, 0},
40 : { 2, 122, -6, 0, -9, 18, 1, 0}, { 2, 121, -6, 0, -10, 20, 1, 0},
41 : { 2, 120, -7, 0, -11, 22, 2, 0}, { 2, 119, -8, 0, -12, 25, 2, 0},
42 : { 3, 117, -8, 0, -13, 27, 2, 0}, { 3, 116, -9, 0, -13, 29, 2, 0},
43 : { 3, 114, -10, 0, -14, 32, 3, 0}, { 3, 113, -10, 0, -15, 35, 2, 0},
44 : { 3, 111, -11, 0, -15, 37, 3, 0}, { 3, 109, -11, 0, -16, 40, 3, 0},
45 : { 3, 108, -12, 0, -16, 42, 3, 0}, { 4, 106, -13, 0, -17, 45, 3, 0},
46 : { 4, 104, -13, 0, -17, 47, 3, 0}, { 4, 102, -14, 0, -17, 50, 3, 0},
47 : { 4, 100, -14, 0, -17, 52, 3, 0}, { 4, 98, -15, 0, -18, 55, 4, 0},
48 : { 4, 96, -15, 0, -18, 58, 3, 0}, { 4, 94, -16, 0, -18, 60, 4, 0},
49 : { 4, 91, -16, 0, -18, 63, 4, 0}, { 4, 89, -16, 0, -18, 65, 4, 0},
50 : { 4, 87, -17, 0, -18, 68, 4, 0}, { 4, 85, -17, 0, -18, 70, 4, 0},
51 : { 4, 82, -17, 0, -18, 73, 4, 0}, { 4, 80, -17, 0, -18, 75, 4, 0},
52 : { 4, 78, -18, 0, -18, 78, 4, 0}, { 4, 75, -18, 0, -17, 80, 4, 0},
53 : { 4, 73, -18, 0, -17, 82, 4, 0}, { 4, 70, -18, 0, -17, 85, 4, 0},
54 : { 4, 68, -18, 0, -17, 87, 4, 0}, { 4, 65, -18, 0, -16, 89, 4, 0},
55 : { 4, 63, -18, 0, -16, 91, 4, 0}, { 4, 60, -18, 0, -16, 94, 4, 0},
56 : { 3, 58, -18, 0, -15, 96, 4, 0}, { 4, 55, -18, 0, -15, 98, 4, 0},
57 : { 3, 52, -17, 0, -14, 100, 4, 0}, { 3, 50, -17, 0, -14, 102, 4, 0},
58 : { 3, 47, -17, 0, -13, 104, 4, 0}, { 3, 45, -17, 0, -13, 106, 4, 0},
59 : { 3, 42, -16, 0, -12, 108, 3, 0}, { 3, 40, -16, 0, -11, 109, 3, 0},
60 : { 3, 37, -15, 0, -11, 111, 3, 0}, { 2, 35, -15, 0, -10, 113, 3, 0},
61 : { 3, 32, -14, 0, -10, 114, 3, 0}, { 2, 29, -13, 0, -9, 116, 3, 0},
62 : { 2, 27, -13, 0, -8, 117, 3, 0}, { 2, 25, -12, 0, -8, 119, 2, 0},
63 : { 2, 22, -11, 0, -7, 120, 2, 0}, { 1, 20, -10, 0, -6, 121, 2, 0},
64 : { 1, 18, -9, 0, -6, 122, 2, 0}, { 1, 15, -8, 0, -5, 123, 2, 0},
65 : { 1, 13, -7, 0, -4, 124, 1, 0}, { 1, 11, -6, 0, -4, 125, 1, 0},
66 : { 1, 8, -5, 0, -3, 126, 1, 0}, { 1, 6, -4, 0, -2, 126, 1, 0},
67 : { 0, 4, -3, 0, -1, 127, 1, 0}, { 0, 2, -1, 0, 0, 127, 0, 0},
68 : // [0, 1)
69 : { 0, 0, 1, 0, 0, 127, 0, 0}, { 0, -1, 2, 0, 0, 127, 0, 0},
70 : { 0, -3, 4, 1, 1, 127, -2, 0}, { 0, -5, 6, 1, 1, 127, -2, 0},
71 : { 0, -6, 8, 1, 2, 126, -3, 0}, {-1, -7, 11, 2, 2, 126, -4, -1},
72 : {-1, -8, 13, 2, 3, 125, -5, -1}, {-1, -10, 16, 3, 3, 124, -6, -1},
73 : {-1, -11, 18, 3, 4, 123, -7, -1}, {-1, -12, 20, 3, 4, 122, -7, -1},
74 : {-1, -13, 23, 3, 4, 121, -8, -1}, {-2, -14, 25, 4, 5, 120, -9, -1},
75 : {-1, -15, 27, 4, 5, 119, -10, -1}, {-1, -16, 30, 4, 5, 118, -11, -1},
76 : {-2, -17, 33, 5, 6, 116, -12, -1}, {-2, -17, 35, 5, 6, 114, -12, -1},
77 : {-2, -18, 38, 5, 6, 113, -13, -1}, {-2, -19, 41, 6, 7, 111, -14, -2},
78 : {-2, -19, 43, 6, 7, 110, -15, -2}, {-2, -20, 46, 6, 7, 108, -15, -2},
79 : {-2, -20, 49, 6, 7, 106, -16, -2}, {-2, -21, 51, 7, 7, 104, -16, -2},
80 : {-2, -21, 54, 7, 7, 102, -17, -2}, {-2, -21, 56, 7, 8, 100, -18, -2},
81 : {-2, -22, 59, 7, 8, 98, -18, -2}, {-2, -22, 62, 7, 8, 96, -19, -2},
82 : {-2, -22, 64, 7, 8, 94, -19, -2}, {-2, -22, 67, 8, 8, 91, -20, -2},
83 : {-2, -22, 69, 8, 8, 89, -20, -2}, {-2, -22, 72, 8, 8, 87, -21, -2},
84 : {-2, -21, 74, 8, 8, 84, -21, -2}, {-2, -22, 77, 8, 8, 82, -21, -2},
85 : {-2, -21, 79, 8, 8, 79, -21, -2}, {-2, -21, 82, 8, 8, 77, -22, -2},
86 : {-2, -21, 84, 8, 8, 74, -21, -2}, {-2, -21, 87, 8, 8, 72, -22, -2},
87 : {-2, -20, 89, 8, 8, 69, -22, -2}, {-2, -20, 91, 8, 8, 67, -22, -2},
88 : {-2, -19, 94, 8, 7, 64, -22, -2}, {-2, -19, 96, 8, 7, 62, -22, -2},
89 : {-2, -18, 98, 8, 7, 59, -22, -2}, {-2, -18, 100, 8, 7, 56, -21, -2},
90 : {-2, -17, 102, 7, 7, 54, -21, -2}, {-2, -16, 104, 7, 7, 51, -21, -2},
91 : {-2, -16, 106, 7, 6, 49, -20, -2}, {-2, -15, 108, 7, 6, 46, -20, -2},
92 : {-2, -15, 110, 7, 6, 43, -19, -2}, {-2, -14, 111, 7, 6, 41, -19, -2},
93 : {-1, -13, 113, 6, 5, 38, -18, -2}, {-1, -12, 114, 6, 5, 35, -17, -2},
94 : {-1, -12, 116, 6, 5, 33, -17, -2}, {-1, -11, 118, 5, 4, 30, -16, -1},
95 : {-1, -10, 119, 5, 4, 27, -15, -1}, {-1, -9, 120, 5, 4, 25, -14, -2},
96 : {-1, -8, 121, 4, 3, 23, -13, -1}, {-1, -7, 122, 4, 3, 20, -12, -1},
97 : {-1, -7, 123, 4, 3, 18, -11, -1}, {-1, -6, 124, 3, 3, 16, -10, -1},
98 : {-1, -5, 125, 3, 2, 13, -8, -1}, {-1, -4, 126, 2, 2, 11, -7, -1},
99 : { 0, -3, 126, 2, 1, 8, -6, 0}, { 0, -2, 127, 1, 1, 6, -5, 0},
100 : { 0, -2, 127, 1, 1, 4, -3, 0}, { 0, 0, 127, 0, 0, 2, -1, 0},
101 : // [1, 2)
102 : { 0, 0, 127, 0, 0, 1, 0, 0}, { 0, 0, 127, 0, 0, -1, 2, 0},
103 : { 0, 1, 127, -1, 0, -3, 4, 0}, { 0, 1, 126, -2, 0, -4, 6, 1},
104 : { 0, 1, 126, -3, 0, -5, 8, 1}, { 0, 1, 125, -4, 0, -6, 11, 1},
105 : { 0, 1, 124, -4, 0, -7, 13, 1}, { 0, 2, 123, -5, 0, -8, 15, 1},
106 : { 0, 2, 122, -6, 0, -9, 18, 1}, { 0, 2, 121, -6, 0, -10, 20, 1},
107 : { 0, 2, 120, -7, 0, -11, 22, 2}, { 0, 2, 119, -8, 0, -12, 25, 2},
108 : { 0, 3, 117, -8, 0, -13, 27, 2}, { 0, 3, 116, -9, 0, -13, 29, 2},
109 : { 0, 3, 114, -10, 0, -14, 32, 3}, { 0, 3, 113, -10, 0, -15, 35, 2},
110 : { 0, 3, 111, -11, 0, -15, 37, 3}, { 0, 3, 109, -11, 0, -16, 40, 3},
111 : { 0, 3, 108, -12, 0, -16, 42, 3}, { 0, 4, 106, -13, 0, -17, 45, 3},
112 : { 0, 4, 104, -13, 0, -17, 47, 3}, { 0, 4, 102, -14, 0, -17, 50, 3},
113 : { 0, 4, 100, -14, 0, -17, 52, 3}, { 0, 4, 98, -15, 0, -18, 55, 4},
114 : { 0, 4, 96, -15, 0, -18, 58, 3}, { 0, 4, 94, -16, 0, -18, 60, 4},
115 : { 0, 4, 91, -16, 0, -18, 63, 4}, { 0, 4, 89, -16, 0, -18, 65, 4},
116 : { 0, 4, 87, -17, 0, -18, 68, 4}, { 0, 4, 85, -17, 0, -18, 70, 4},
117 : { 0, 4, 82, -17, 0, -18, 73, 4}, { 0, 4, 80, -17, 0, -18, 75, 4},
118 : { 0, 4, 78, -18, 0, -18, 78, 4}, { 0, 4, 75, -18, 0, -17, 80, 4},
119 : { 0, 4, 73, -18, 0, -17, 82, 4}, { 0, 4, 70, -18, 0, -17, 85, 4},
120 : { 0, 4, 68, -18, 0, -17, 87, 4}, { 0, 4, 65, -18, 0, -16, 89, 4},
121 : { 0, 4, 63, -18, 0, -16, 91, 4}, { 0, 4, 60, -18, 0, -16, 94, 4},
122 : { 0, 3, 58, -18, 0, -15, 96, 4}, { 0, 4, 55, -18, 0, -15, 98, 4},
123 : { 0, 3, 52, -17, 0, -14, 100, 4}, { 0, 3, 50, -17, 0, -14, 102, 4},
124 : { 0, 3, 47, -17, 0, -13, 104, 4}, { 0, 3, 45, -17, 0, -13, 106, 4},
125 : { 0, 3, 42, -16, 0, -12, 108, 3}, { 0, 3, 40, -16, 0, -11, 109, 3},
126 : { 0, 3, 37, -15, 0, -11, 111, 3}, { 0, 2, 35, -15, 0, -10, 113, 3},
127 : { 0, 3, 32, -14, 0, -10, 114, 3}, { 0, 2, 29, -13, 0, -9, 116, 3},
128 : { 0, 2, 27, -13, 0, -8, 117, 3}, { 0, 2, 25, -12, 0, -8, 119, 2},
129 : { 0, 2, 22, -11, 0, -7, 120, 2}, { 0, 1, 20, -10, 0, -6, 121, 2},
130 : { 0, 1, 18, -9, 0, -6, 122, 2}, { 0, 1, 15, -8, 0, -5, 123, 2},
131 : { 0, 1, 13, -7, 0, -4, 124, 1}, { 0, 1, 11, -6, 0, -4, 125, 1},
132 : { 0, 1, 8, -5, 0, -3, 126, 1}, { 0, 1, 6, -4, 0, -2, 126, 1},
133 : { 0, 0, 4, -3, 0, -1, 127, 1}, { 0, 0, 2, -1, 0, 0, 127, 0},
134 : // dummy (replicate row index 191)
135 : { 0, 0, 2, -1, 0, 0, 127, 0},
136 :
137 : #else
138 : // [-1, 0)
139 : { 0, 127, 0, 0, 0, 1, 0, 0}, { 1, 127, -1, 0, -3, 4, 0, 0},
140 : { 1, 126, -3, 0, -5, 8, 1, 0}, { 1, 124, -4, 0, -7, 13, 1, 0},
141 : { 2, 122, -6, 0, -9, 18, 1, 0}, { 2, 120, -7, 0, -11, 22, 2, 0},
142 : { 3, 117, -8, 0, -13, 27, 2, 0}, { 3, 114, -10, 0, -14, 32, 3, 0},
143 : { 3, 111, -11, 0, -15, 37, 3, 0}, { 3, 108, -12, 0, -16, 42, 3, 0},
144 : { 4, 104, -13, 0, -17, 47, 3, 0}, { 4, 100, -14, 0, -17, 52, 3, 0},
145 : { 4, 96, -15, 0, -18, 58, 3, 0}, { 4, 91, -16, 0, -18, 63, 4, 0},
146 : { 4, 87, -17, 0, -18, 68, 4, 0}, { 4, 82, -17, 0, -18, 73, 4, 0},
147 : { 4, 78, -18, 0, -18, 78, 4, 0}, { 4, 73, -18, 0, -17, 82, 4, 0},
148 : { 4, 68, -18, 0, -17, 87, 4, 0}, { 4, 63, -18, 0, -16, 91, 4, 0},
149 : { 3, 58, -18, 0, -15, 96, 4, 0}, { 3, 52, -17, 0, -14, 100, 4, 0},
150 : { 3, 47, -17, 0, -13, 104, 4, 0}, { 3, 42, -16, 0, -12, 108, 3, 0},
151 : { 3, 37, -15, 0, -11, 111, 3, 0}, { 3, 32, -14, 0, -10, 114, 3, 0},
152 : { 2, 27, -13, 0, -8, 117, 3, 0}, { 2, 22, -11, 0, -7, 120, 2, 0},
153 : { 1, 18, -9, 0, -6, 122, 2, 0}, { 1, 13, -7, 0, -4, 124, 1, 0},
154 : { 1, 8, -5, 0, -3, 126, 1, 0}, { 0, 4, -3, 0, -1, 127, 1, 0},
155 : // [0, 1)
156 : { 0, 0, 1, 0, 0, 127, 0, 0}, { 0, -3, 4, 1, 1, 127, -2, 0},
157 : { 0, -6, 8, 1, 2, 126, -3, 0}, {-1, -8, 13, 2, 3, 125, -5, -1},
158 : {-1, -11, 18, 3, 4, 123, -7, -1}, {-1, -13, 23, 3, 4, 121, -8, -1},
159 : {-1, -15, 27, 4, 5, 119, -10, -1}, {-2, -17, 33, 5, 6, 116, -12, -1},
160 : {-2, -18, 38, 5, 6, 113, -13, -1}, {-2, -19, 43, 6, 7, 110, -15, -2},
161 : {-2, -20, 49, 6, 7, 106, -16, -2}, {-2, -21, 54, 7, 7, 102, -17, -2},
162 : {-2, -22, 59, 7, 8, 98, -18, -2}, {-2, -22, 64, 7, 8, 94, -19, -2},
163 : {-2, -22, 69, 8, 8, 89, -20, -2}, {-2, -21, 74, 8, 8, 84, -21, -2},
164 : {-2, -21, 79, 8, 8, 79, -21, -2}, {-2, -21, 84, 8, 8, 74, -21, -2},
165 : {-2, -20, 89, 8, 8, 69, -22, -2}, {-2, -19, 94, 8, 7, 64, -22, -2},
166 : {-2, -18, 98, 8, 7, 59, -22, -2}, {-2, -17, 102, 7, 7, 54, -21, -2},
167 : {-2, -16, 106, 7, 6, 49, -20, -2}, {-2, -15, 110, 7, 6, 43, -19, -2},
168 : {-1, -13, 113, 6, 5, 38, -18, -2}, {-1, -12, 116, 6, 5, 33, -17, -2},
169 : {-1, -10, 119, 5, 4, 27, -15, -1}, {-1, -8, 121, 4, 3, 23, -13, -1},
170 : {-1, -7, 123, 4, 3, 18, -11, -1}, {-1, -5, 125, 3, 2, 13, -8, -1},
171 : { 0, -3, 126, 2, 1, 8, -6, 0}, { 0, -2, 127, 1, 1, 4, -3, 0},
172 : // [1, 2)
173 : { 0, 0, 127, 0, 0, 1, 0, 0}, { 0, 1, 127, -1, 0, -3, 4, 0},
174 : { 0, 1, 126, -3, 0, -5, 8, 1}, { 0, 1, 124, -4, 0, -7, 13, 1},
175 : { 0, 2, 122, -6, 0, -9, 18, 1}, { 0, 2, 120, -7, 0, -11, 22, 2},
176 : { 0, 3, 117, -8, 0, -13, 27, 2}, { 0, 3, 114, -10, 0, -14, 32, 3},
177 : { 0, 3, 111, -11, 0, -15, 37, 3}, { 0, 3, 108, -12, 0, -16, 42, 3},
178 : { 0, 4, 104, -13, 0, -17, 47, 3}, { 0, 4, 100, -14, 0, -17, 52, 3},
179 : { 0, 4, 96, -15, 0, -18, 58, 3}, { 0, 4, 91, -16, 0, -18, 63, 4},
180 : { 0, 4, 87, -17, 0, -18, 68, 4}, { 0, 4, 82, -17, 0, -18, 73, 4},
181 : { 0, 4, 78, -18, 0, -18, 78, 4}, { 0, 4, 73, -18, 0, -17, 82, 4},
182 : { 0, 4, 68, -18, 0, -17, 87, 4}, { 0, 4, 63, -18, 0, -16, 91, 4},
183 : { 0, 3, 58, -18, 0, -15, 96, 4}, { 0, 3, 52, -17, 0, -14, 100, 4},
184 : { 0, 3, 47, -17, 0, -13, 104, 4}, { 0, 3, 42, -16, 0, -12, 108, 3},
185 : { 0, 3, 37, -15, 0, -11, 111, 3}, { 0, 3, 32, -14, 0, -10, 114, 3},
186 : { 0, 2, 27, -13, 0, -8, 117, 3}, { 0, 2, 22, -11, 0, -7, 120, 2},
187 : { 0, 1, 18, -9, 0, -6, 122, 2}, { 0, 1, 13, -7, 0, -4, 124, 1},
188 : { 0, 1, 8, -5, 0, -3, 126, 1}, { 0, 0, 4, -3, 0, -1, 127, 1},
189 : // dummy (replicate row index 95)
190 : { 0, 0, 4, -3, 0, -1, 127, 1},
191 : #endif // WARPEDPIXEL_PREC_BITS == 6
192 : };
193 : /* clang-format on */
194 :
195 : // Shuffle masks: we want to convert a sequence of bytes 0, 1, 2, ..., 15
196 : // in an SSE register into two sequences:
197 : // 0, 2, 2, 4, ..., 12, 12, 14, <don't care>
198 : // 1, 3, 3, 5, ..., 13, 13, 15, <don't care>
199 : static const uint8_t even_mask[16] = { 0, 2, 2, 4, 4, 6, 6, 8,
200 : 8, 10, 10, 12, 12, 14, 14, 0 };
201 : static const uint8_t odd_mask[16] = { 1, 3, 3, 5, 5, 7, 7, 9,
202 : 9, 11, 11, 13, 13, 15, 15, 0 };
203 :
204 0 : void av1_warp_affine_ssse3(const int32_t *mat, const uint8_t *ref, int width,
205 : int height, int stride, uint8_t *pred, int p_col,
206 : int p_row, int p_width, int p_height, int p_stride,
207 : int subsampling_x, int subsampling_y, int comp_avg,
208 : int16_t alpha, int16_t beta, int16_t gamma,
209 : int16_t delta) {
210 : __m128i tmp[15];
211 : int i, j, k;
212 0 : const int bd = 8;
213 :
214 : /* Note: For this code to work, the left/right frame borders need to be
215 : extended by at least 13 pixels each. By the time we get here, other
216 : code will have set up this border, but we allow an explicit check
217 : for debugging purposes.
218 : */
219 : /*for (i = 0; i < height; ++i) {
220 : for (j = 0; j < 13; ++j) {
221 : assert(ref[i * stride - 13 + j] == ref[i * stride]);
222 : assert(ref[i * stride + width + j] == ref[i * stride + (width - 1)]);
223 : }
224 : }*/
225 :
226 0 : for (i = 0; i < p_height; i += 8) {
227 0 : for (j = 0; j < p_width; j += 8) {
228 : // (x, y) coordinates of the center of this block in the destination
229 : // image
230 0 : const int32_t dst_x = p_col + j + 4;
231 0 : const int32_t dst_y = p_row + i + 4;
232 :
233 : int32_t x4, y4, ix4, sx4, iy4, sy4;
234 0 : if (subsampling_x)
235 0 : x4 = (mat[2] * 4 * dst_x + mat[3] * 4 * dst_y + mat[0] * 2 +
236 0 : (mat[2] + mat[3] - (1 << WARPEDMODEL_PREC_BITS))) /
237 : 4;
238 : else
239 0 : x4 = mat[2] * dst_x + mat[3] * dst_y + mat[0];
240 :
241 0 : if (subsampling_y)
242 0 : y4 = (mat[4] * 4 * dst_x + mat[5] * 4 * dst_y + mat[1] * 2 +
243 0 : (mat[4] + mat[5] - (1 << WARPEDMODEL_PREC_BITS))) /
244 : 4;
245 : else
246 0 : y4 = mat[4] * dst_x + mat[5] * dst_y + mat[1];
247 :
248 0 : ix4 = x4 >> WARPEDMODEL_PREC_BITS;
249 0 : sx4 = x4 & ((1 << WARPEDMODEL_PREC_BITS) - 1);
250 0 : iy4 = y4 >> WARPEDMODEL_PREC_BITS;
251 0 : sy4 = y4 & ((1 << WARPEDMODEL_PREC_BITS) - 1);
252 :
253 : // Add in all the constant terms, including rounding and offset
254 0 : sx4 += alpha * (-4) + beta * (-4) + (1 << (WARPEDDIFF_PREC_BITS - 1)) +
255 : (WARPEDPIXEL_PREC_SHIFTS << WARPEDDIFF_PREC_BITS);
256 0 : sy4 += gamma * (-4) + delta * (-4) + (1 << (WARPEDDIFF_PREC_BITS - 1)) +
257 : (WARPEDPIXEL_PREC_SHIFTS << WARPEDDIFF_PREC_BITS);
258 :
259 0 : sx4 &= ~((1 << WARP_PARAM_REDUCE_BITS) - 1);
260 0 : sy4 &= ~((1 << WARP_PARAM_REDUCE_BITS) - 1);
261 :
262 : // Horizontal filter
263 : // If the block is aligned such that, after clamping, every sample
264 : // would be taken from the leftmost/rightmost column, then we can
265 : // skip the expensive horizontal filter.
266 0 : if (ix4 <= -7) {
267 0 : for (k = -7; k < AOMMIN(8, p_height - i); ++k) {
268 0 : int iy = iy4 + k;
269 0 : if (iy < 0)
270 0 : iy = 0;
271 0 : else if (iy > height - 1)
272 0 : iy = height - 1;
273 0 : tmp[k + 7] = _mm_set1_epi16(
274 0 : (1 << (bd + WARPEDPIXEL_FILTER_BITS - HORSHEAR_REDUCE_PREC_BITS -
275 0 : 1)) +
276 0 : ref[iy * stride] *
277 : (1 << (WARPEDPIXEL_FILTER_BITS - HORSHEAR_REDUCE_PREC_BITS)));
278 : }
279 0 : } else if (ix4 >= width + 6) {
280 0 : for (k = -7; k < AOMMIN(8, p_height - i); ++k) {
281 0 : int iy = iy4 + k;
282 0 : if (iy < 0)
283 0 : iy = 0;
284 0 : else if (iy > height - 1)
285 0 : iy = height - 1;
286 0 : tmp[k + 7] = _mm_set1_epi16(
287 0 : (1 << (bd + WARPEDPIXEL_FILTER_BITS - HORSHEAR_REDUCE_PREC_BITS -
288 0 : 1)) +
289 0 : ref[iy * stride + (width - 1)] *
290 : (1 << (WARPEDPIXEL_FILTER_BITS - HORSHEAR_REDUCE_PREC_BITS)));
291 : }
292 : } else {
293 0 : for (k = -7; k < AOMMIN(8, p_height - i); ++k) {
294 0 : int iy = iy4 + k;
295 0 : if (iy < 0)
296 0 : iy = 0;
297 0 : else if (iy > height - 1)
298 0 : iy = height - 1;
299 0 : int sx = sx4 + beta * (k + 4);
300 :
301 : // Load source pixels
302 0 : const __m128i src =
303 0 : _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
304 0 : const __m128i src_even =
305 0 : _mm_shuffle_epi8(src, _mm_loadu_si128((__m128i *)even_mask));
306 0 : const __m128i src_odd =
307 0 : _mm_shuffle_epi8(src, _mm_loadu_si128((__m128i *)odd_mask));
308 :
309 : // Filter even-index pixels
310 0 : const __m128i tmp_0 = _mm_loadl_epi64((
311 0 : __m128i *)&filter_8bit[(sx + 0 * alpha) >> WARPEDDIFF_PREC_BITS]);
312 0 : const __m128i tmp_1 = _mm_loadl_epi64((
313 0 : __m128i *)&filter_8bit[(sx + 1 * alpha) >> WARPEDDIFF_PREC_BITS]);
314 0 : const __m128i tmp_2 = _mm_loadl_epi64((
315 0 : __m128i *)&filter_8bit[(sx + 2 * alpha) >> WARPEDDIFF_PREC_BITS]);
316 0 : const __m128i tmp_3 = _mm_loadl_epi64((
317 0 : __m128i *)&filter_8bit[(sx + 3 * alpha) >> WARPEDDIFF_PREC_BITS]);
318 0 : const __m128i tmp_4 = _mm_loadl_epi64((
319 0 : __m128i *)&filter_8bit[(sx + 4 * alpha) >> WARPEDDIFF_PREC_BITS]);
320 0 : const __m128i tmp_5 = _mm_loadl_epi64((
321 0 : __m128i *)&filter_8bit[(sx + 5 * alpha) >> WARPEDDIFF_PREC_BITS]);
322 0 : const __m128i tmp_6 = _mm_loadl_epi64((
323 0 : __m128i *)&filter_8bit[(sx + 6 * alpha) >> WARPEDDIFF_PREC_BITS]);
324 0 : const __m128i tmp_7 = _mm_loadl_epi64((
325 0 : __m128i *)&filter_8bit[(sx + 7 * alpha) >> WARPEDDIFF_PREC_BITS]);
326 :
327 : // Coeffs 0 2 0 2 4 6 4 6 1 3 1 3 5 7 5 7 for pixels 0 2
328 0 : const __m128i tmp_8 = _mm_unpacklo_epi16(tmp_0, tmp_2);
329 : // Coeffs 0 2 0 2 4 6 4 6 1 3 1 3 5 7 5 7 for pixels 1 3
330 0 : const __m128i tmp_9 = _mm_unpacklo_epi16(tmp_1, tmp_3);
331 : // Coeffs 0 2 0 2 4 6 4 6 1 3 1 3 5 7 5 7 for pixels 4 6
332 0 : const __m128i tmp_10 = _mm_unpacklo_epi16(tmp_4, tmp_6);
333 : // Coeffs 0 2 0 2 4 6 4 6 1 3 1 3 5 7 5 7 for pixels 5 7
334 0 : const __m128i tmp_11 = _mm_unpacklo_epi16(tmp_5, tmp_7);
335 :
336 : // Coeffs 0 2 0 2 0 2 0 2 4 6 4 6 4 6 4 6 for pixels 0 2 4 6
337 0 : const __m128i tmp_12 = _mm_unpacklo_epi32(tmp_8, tmp_10);
338 : // Coeffs 1 3 1 3 1 3 1 3 5 7 5 7 5 7 5 7 for pixels 0 2 4 6
339 0 : const __m128i tmp_13 = _mm_unpackhi_epi32(tmp_8, tmp_10);
340 : // Coeffs 0 2 0 2 0 2 0 2 4 6 4 6 4 6 4 6 for pixels 1 3 5 7
341 0 : const __m128i tmp_14 = _mm_unpacklo_epi32(tmp_9, tmp_11);
342 : // Coeffs 1 3 1 3 1 3 1 3 5 7 5 7 5 7 5 7 for pixels 1 3 5 7
343 0 : const __m128i tmp_15 = _mm_unpackhi_epi32(tmp_9, tmp_11);
344 :
345 : // Coeffs 0 2 for pixels 0 2 4 6 1 3 5 7
346 0 : const __m128i coeff_02 = _mm_unpacklo_epi64(tmp_12, tmp_14);
347 : // Coeffs 4 6 for pixels 0 2 4 6 1 3 5 7
348 0 : const __m128i coeff_46 = _mm_unpackhi_epi64(tmp_12, tmp_14);
349 : // Coeffs 1 3 for pixels 0 2 4 6 1 3 5 7
350 0 : const __m128i coeff_13 = _mm_unpacklo_epi64(tmp_13, tmp_15);
351 : // Coeffs 5 7 for pixels 0 2 4 6 1 3 5 7
352 0 : const __m128i coeff_57 = _mm_unpackhi_epi64(tmp_13, tmp_15);
353 :
354 : // The pixel order we need for 'src' is:
355 : // 0 2 2 4 4 6 6 8 1 3 3 5 5 7 7 9
356 0 : const __m128i src_02 = _mm_unpacklo_epi64(src_even, src_odd);
357 0 : const __m128i res_02 = _mm_maddubs_epi16(src_02, coeff_02);
358 : // 4 6 6 8 8 10 10 12 5 7 7 9 9 11 11 13
359 0 : const __m128i src_46 = _mm_unpacklo_epi64(_mm_srli_si128(src_even, 4),
360 0 : _mm_srli_si128(src_odd, 4));
361 0 : const __m128i res_46 = _mm_maddubs_epi16(src_46, coeff_46);
362 : // 1 3 3 5 5 7 7 9 2 4 4 6 6 8 8 10
363 0 : const __m128i src_13 =
364 0 : _mm_unpacklo_epi64(src_odd, _mm_srli_si128(src_even, 2));
365 0 : const __m128i res_13 = _mm_maddubs_epi16(src_13, coeff_13);
366 : // 5 7 7 9 9 11 11 13 6 8 8 10 10 12 12 14
367 0 : const __m128i src_57 = _mm_unpacklo_epi64(
368 0 : _mm_srli_si128(src_odd, 4), _mm_srli_si128(src_even, 6));
369 0 : const __m128i res_57 = _mm_maddubs_epi16(src_57, coeff_57);
370 :
371 0 : const __m128i round_const =
372 0 : _mm_set1_epi16((1 << (bd + WARPEDPIXEL_FILTER_BITS - 1)) +
373 : ((1 << HORSHEAR_REDUCE_PREC_BITS) >> 1));
374 :
375 : // Note: The values res_02 + res_46 and res_13 + res_57 both
376 : // fit into int16s at this point, but their sum may be too wide to fit
377 : // into an int16. However, once we also add round_const, the sum of
378 : // all of these fits into a uint16.
379 : //
380 : // The wrapping behaviour of _mm_add_* is used here to make sure we
381 : // get the correct result despite converting between different
382 : // (implicit) types.
383 0 : const __m128i res_even = _mm_add_epi16(res_02, res_46);
384 0 : const __m128i res_odd = _mm_add_epi16(res_13, res_57);
385 0 : const __m128i res =
386 0 : _mm_add_epi16(_mm_add_epi16(res_even, res_odd), round_const);
387 0 : tmp[k + 7] = _mm_srli_epi16(res, HORSHEAR_REDUCE_PREC_BITS);
388 : }
389 : }
390 :
391 : // Vertical filter
392 0 : for (k = -4; k < AOMMIN(4, p_height - i - 4); ++k) {
393 0 : int sy = sy4 + delta * (k + 4);
394 :
395 : // Load from tmp and rearrange pairs of consecutive rows into the
396 : // column order 0 0 2 2 4 4 6 6; 1 1 3 3 5 5 7 7
397 0 : const __m128i *src = tmp + (k + 4);
398 0 : const __m128i src_0 = _mm_unpacklo_epi16(src[0], src[1]);
399 0 : const __m128i src_2 = _mm_unpacklo_epi16(src[2], src[3]);
400 0 : const __m128i src_4 = _mm_unpacklo_epi16(src[4], src[5]);
401 0 : const __m128i src_6 = _mm_unpacklo_epi16(src[6], src[7]);
402 :
403 : // Filter even-index pixels
404 0 : const __m128i tmp_0 = _mm_loadu_si128(
405 0 : (__m128i *)(warped_filter +
406 0 : ((sy + 0 * gamma) >> WARPEDDIFF_PREC_BITS)));
407 0 : const __m128i tmp_2 = _mm_loadu_si128(
408 0 : (__m128i *)(warped_filter +
409 0 : ((sy + 2 * gamma) >> WARPEDDIFF_PREC_BITS)));
410 0 : const __m128i tmp_4 = _mm_loadu_si128(
411 0 : (__m128i *)(warped_filter +
412 0 : ((sy + 4 * gamma) >> WARPEDDIFF_PREC_BITS)));
413 0 : const __m128i tmp_6 = _mm_loadu_si128(
414 0 : (__m128i *)(warped_filter +
415 0 : ((sy + 6 * gamma) >> WARPEDDIFF_PREC_BITS)));
416 :
417 0 : const __m128i tmp_8 = _mm_unpacklo_epi32(tmp_0, tmp_2);
418 0 : const __m128i tmp_10 = _mm_unpacklo_epi32(tmp_4, tmp_6);
419 0 : const __m128i tmp_12 = _mm_unpackhi_epi32(tmp_0, tmp_2);
420 0 : const __m128i tmp_14 = _mm_unpackhi_epi32(tmp_4, tmp_6);
421 :
422 0 : const __m128i coeff_0 = _mm_unpacklo_epi64(tmp_8, tmp_10);
423 0 : const __m128i coeff_2 = _mm_unpackhi_epi64(tmp_8, tmp_10);
424 0 : const __m128i coeff_4 = _mm_unpacklo_epi64(tmp_12, tmp_14);
425 0 : const __m128i coeff_6 = _mm_unpackhi_epi64(tmp_12, tmp_14);
426 :
427 0 : const __m128i res_0 = _mm_madd_epi16(src_0, coeff_0);
428 0 : const __m128i res_2 = _mm_madd_epi16(src_2, coeff_2);
429 0 : const __m128i res_4 = _mm_madd_epi16(src_4, coeff_4);
430 0 : const __m128i res_6 = _mm_madd_epi16(src_6, coeff_6);
431 :
432 0 : const __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_2),
433 : _mm_add_epi32(res_4, res_6));
434 :
435 : // Filter odd-index pixels
436 0 : const __m128i src_1 = _mm_unpackhi_epi16(src[0], src[1]);
437 0 : const __m128i src_3 = _mm_unpackhi_epi16(src[2], src[3]);
438 0 : const __m128i src_5 = _mm_unpackhi_epi16(src[4], src[5]);
439 0 : const __m128i src_7 = _mm_unpackhi_epi16(src[6], src[7]);
440 :
441 0 : const __m128i tmp_1 = _mm_loadu_si128(
442 0 : (__m128i *)(warped_filter +
443 0 : ((sy + 1 * gamma) >> WARPEDDIFF_PREC_BITS)));
444 0 : const __m128i tmp_3 = _mm_loadu_si128(
445 0 : (__m128i *)(warped_filter +
446 0 : ((sy + 3 * gamma) >> WARPEDDIFF_PREC_BITS)));
447 0 : const __m128i tmp_5 = _mm_loadu_si128(
448 0 : (__m128i *)(warped_filter +
449 0 : ((sy + 5 * gamma) >> WARPEDDIFF_PREC_BITS)));
450 0 : const __m128i tmp_7 = _mm_loadu_si128(
451 0 : (__m128i *)(warped_filter +
452 0 : ((sy + 7 * gamma) >> WARPEDDIFF_PREC_BITS)));
453 :
454 0 : const __m128i tmp_9 = _mm_unpacklo_epi32(tmp_1, tmp_3);
455 0 : const __m128i tmp_11 = _mm_unpacklo_epi32(tmp_5, tmp_7);
456 0 : const __m128i tmp_13 = _mm_unpackhi_epi32(tmp_1, tmp_3);
457 0 : const __m128i tmp_15 = _mm_unpackhi_epi32(tmp_5, tmp_7);
458 :
459 0 : const __m128i coeff_1 = _mm_unpacklo_epi64(tmp_9, tmp_11);
460 0 : const __m128i coeff_3 = _mm_unpackhi_epi64(tmp_9, tmp_11);
461 0 : const __m128i coeff_5 = _mm_unpacklo_epi64(tmp_13, tmp_15);
462 0 : const __m128i coeff_7 = _mm_unpackhi_epi64(tmp_13, tmp_15);
463 :
464 0 : const __m128i res_1 = _mm_madd_epi16(src_1, coeff_1);
465 0 : const __m128i res_3 = _mm_madd_epi16(src_3, coeff_3);
466 0 : const __m128i res_5 = _mm_madd_epi16(src_5, coeff_5);
467 0 : const __m128i res_7 = _mm_madd_epi16(src_7, coeff_7);
468 :
469 0 : const __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_3),
470 : _mm_add_epi32(res_5, res_7));
471 :
472 : // Rearrange pixels back into the order 0 ... 7
473 0 : const __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd);
474 0 : const __m128i res_hi = _mm_unpackhi_epi32(res_even, res_odd);
475 :
476 : // Round and pack into 8 bits
477 0 : const __m128i round_const =
478 0 : _mm_set1_epi32(-(1 << (bd + VERSHEAR_REDUCE_PREC_BITS - 1)) +
479 : ((1 << VERSHEAR_REDUCE_PREC_BITS) >> 1));
480 :
481 0 : const __m128i res_lo_round = _mm_srai_epi32(
482 : _mm_add_epi32(res_lo, round_const), VERSHEAR_REDUCE_PREC_BITS);
483 0 : const __m128i res_hi_round = _mm_srai_epi32(
484 : _mm_add_epi32(res_hi, round_const), VERSHEAR_REDUCE_PREC_BITS);
485 :
486 0 : const __m128i res_16bit = _mm_packs_epi32(res_lo_round, res_hi_round);
487 0 : __m128i res_8bit = _mm_packus_epi16(res_16bit, res_16bit);
488 :
489 : // Store, blending with 'pred' if needed
490 0 : __m128i *const p = (__m128i *)&pred[(i + k + 4) * p_stride + j];
491 :
492 : // Note: If we're outputting a 4x4 block, we need to be very careful
493 : // to only output 4 pixels at this point, to avoid encode/decode
494 : // mismatches when encoding with multiple threads.
495 0 : if (p_width == 4) {
496 0 : if (comp_avg) {
497 0 : const __m128i orig = _mm_cvtsi32_si128(*(uint32_t *)p);
498 0 : res_8bit = _mm_avg_epu8(res_8bit, orig);
499 : }
500 0 : *(uint32_t *)p = _mm_cvtsi128_si32(res_8bit);
501 : } else {
502 0 : if (comp_avg) res_8bit = _mm_avg_epu8(res_8bit, _mm_loadl_epi64(p));
503 : _mm_storel_epi64(p, res_8bit);
504 : }
505 : }
506 : }
507 : }
508 0 : }
|
