Line data Source code
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 <math.h>
13 :
14 : #include "aom_ports/mem.h"
15 :
16 : #include "av1/encoder/aq_variance.h"
17 :
18 : #include "av1/common/seg_common.h"
19 : #include "av1/encoder/ratectrl.h"
20 : #include "av1/encoder/rd.h"
21 : #include "av1/encoder/segmentation.h"
22 : #include "aom_ports/system_state.h"
23 :
24 : #define ENERGY_MIN (-4)
25 : #define ENERGY_MAX (1)
26 : #define ENERGY_SPAN (ENERGY_MAX - ENERGY_MIN + 1)
27 : #define ENERGY_IN_BOUNDS(energy) \
28 : assert((energy) >= ENERGY_MIN && (energy) <= ENERGY_MAX)
29 :
30 : static const double rate_ratio[MAX_SEGMENTS] = { 2.5, 2.0, 1.5, 1.0,
31 : 0.75, 1.0, 1.0, 1.0 };
32 : static const int segment_id[ENERGY_SPAN] = { 0, 1, 1, 2, 3, 4 };
33 :
34 : #define SEGMENT_ID(i) segment_id[(i)-ENERGY_MIN]
35 :
36 : DECLARE_ALIGNED(16, static const uint8_t, av1_all_zeros[MAX_SB_SIZE]) = { 0 };
37 : #if CONFIG_HIGHBITDEPTH
38 : DECLARE_ALIGNED(16, static const uint16_t,
39 : av1_highbd_all_zeros[MAX_SB_SIZE]) = { 0 };
40 : #endif
41 :
42 0 : unsigned int av1_vaq_segment_id(int energy) {
43 0 : ENERGY_IN_BOUNDS(energy);
44 0 : return SEGMENT_ID(energy);
45 : }
46 :
47 0 : void av1_vaq_frame_setup(AV1_COMP *cpi) {
48 0 : AV1_COMMON *cm = &cpi->common;
49 0 : struct segmentation *seg = &cm->seg;
50 : int i;
51 :
52 0 : if (frame_is_intra_only(cm) || cm->error_resilient_mode ||
53 0 : cpi->refresh_alt_ref_frame ||
54 0 : (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref)) {
55 0 : cpi->vaq_refresh = 1;
56 :
57 0 : av1_enable_segmentation(seg);
58 0 : av1_clearall_segfeatures(seg);
59 :
60 0 : seg->abs_delta = SEGMENT_DELTADATA;
61 :
62 0 : aom_clear_system_state();
63 :
64 0 : for (i = 0; i < MAX_SEGMENTS; ++i) {
65 0 : int qindex_delta =
66 0 : av1_compute_qdelta_by_rate(&cpi->rc, cm->frame_type, cm->base_qindex,
67 : rate_ratio[i], cm->bit_depth);
68 :
69 : // We don't allow qindex 0 in a segment if the base value is not 0.
70 : // Q index 0 (lossless) implies 4x4 encoding only and in AQ mode a segment
71 : // Q delta is sometimes applied without going back around the rd loop.
72 : // This could lead to an illegal combination of partition size and q.
73 0 : if ((cm->base_qindex != 0) && ((cm->base_qindex + qindex_delta) == 0)) {
74 0 : qindex_delta = -cm->base_qindex + 1;
75 : }
76 :
77 : // No need to enable SEG_LVL_ALT_Q for this segment.
78 0 : if (rate_ratio[i] == 1.0) {
79 0 : continue;
80 : }
81 :
82 0 : av1_set_segdata(seg, i, SEG_LVL_ALT_Q, qindex_delta);
83 0 : av1_enable_segfeature(seg, i, SEG_LVL_ALT_Q);
84 : }
85 : }
86 0 : }
87 :
88 : /* TODO(agrange, paulwilkins): The block_variance calls the unoptimized versions
89 : * of variance() and highbd_8_variance(). It should not.
90 : */
91 0 : static void aq_variance(const uint8_t *a, int a_stride, const uint8_t *b,
92 : int b_stride, int w, int h, unsigned int *sse,
93 : int *sum) {
94 : int i, j;
95 :
96 0 : *sum = 0;
97 0 : *sse = 0;
98 :
99 0 : for (i = 0; i < h; i++) {
100 0 : for (j = 0; j < w; j++) {
101 0 : const int diff = a[j] - b[j];
102 0 : *sum += diff;
103 0 : *sse += diff * diff;
104 : }
105 :
106 0 : a += a_stride;
107 0 : b += b_stride;
108 : }
109 0 : }
110 :
111 : #if CONFIG_HIGHBITDEPTH
112 0 : static void aq_highbd_variance64(const uint8_t *a8, int a_stride,
113 : const uint8_t *b8, int b_stride, int w, int h,
114 : uint64_t *sse, uint64_t *sum) {
115 : int i, j;
116 :
117 0 : uint16_t *a = CONVERT_TO_SHORTPTR(a8);
118 0 : uint16_t *b = CONVERT_TO_SHORTPTR(b8);
119 0 : *sum = 0;
120 0 : *sse = 0;
121 :
122 0 : for (i = 0; i < h; i++) {
123 0 : for (j = 0; j < w; j++) {
124 0 : const int diff = a[j] - b[j];
125 0 : *sum += diff;
126 0 : *sse += diff * diff;
127 : }
128 0 : a += a_stride;
129 0 : b += b_stride;
130 : }
131 0 : }
132 :
133 0 : static void aq_highbd_8_variance(const uint8_t *a8, int a_stride,
134 : const uint8_t *b8, int b_stride, int w, int h,
135 : unsigned int *sse, int *sum) {
136 0 : uint64_t sse_long = 0;
137 0 : uint64_t sum_long = 0;
138 0 : aq_highbd_variance64(a8, a_stride, b8, b_stride, w, h, &sse_long, &sum_long);
139 0 : *sse = (unsigned int)sse_long;
140 0 : *sum = (int)sum_long;
141 0 : }
142 : #endif // CONFIG_HIGHBITDEPTH
143 :
144 0 : static unsigned int block_variance(const AV1_COMP *const cpi, MACROBLOCK *x,
145 : BLOCK_SIZE bs) {
146 0 : MACROBLOCKD *xd = &x->e_mbd;
147 : unsigned int var, sse;
148 0 : int right_overflow =
149 0 : (xd->mb_to_right_edge < 0) ? ((-xd->mb_to_right_edge) >> 3) : 0;
150 0 : int bottom_overflow =
151 0 : (xd->mb_to_bottom_edge < 0) ? ((-xd->mb_to_bottom_edge) >> 3) : 0;
152 :
153 0 : if (right_overflow || bottom_overflow) {
154 0 : const int bw = 8 * mi_size_wide[bs] - right_overflow;
155 0 : const int bh = 8 * mi_size_high[bs] - bottom_overflow;
156 : int avg;
157 : #if CONFIG_HIGHBITDEPTH
158 0 : if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
159 0 : aq_highbd_8_variance(x->plane[0].src.buf, x->plane[0].src.stride,
160 0 : CONVERT_TO_BYTEPTR(av1_highbd_all_zeros), 0, bw, bh,
161 : &sse, &avg);
162 0 : sse >>= 2 * (xd->bd - 8);
163 0 : avg >>= (xd->bd - 8);
164 : } else {
165 0 : aq_variance(x->plane[0].src.buf, x->plane[0].src.stride, av1_all_zeros, 0,
166 : bw, bh, &sse, &avg);
167 : }
168 : #else
169 : aq_variance(x->plane[0].src.buf, x->plane[0].src.stride, av1_all_zeros, 0,
170 : bw, bh, &sse, &avg);
171 : #endif // CONFIG_HIGHBITDEPTH
172 0 : var = sse - (unsigned int)(((int64_t)avg * avg) / (bw * bh));
173 0 : return (unsigned int)((uint64_t)var * 256) / (bw * bh);
174 : } else {
175 : #if CONFIG_HIGHBITDEPTH
176 0 : if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
177 0 : var =
178 0 : cpi->fn_ptr[bs].vf(x->plane[0].src.buf, x->plane[0].src.stride,
179 0 : CONVERT_TO_BYTEPTR(av1_highbd_all_zeros), 0, &sse);
180 : } else {
181 0 : var = cpi->fn_ptr[bs].vf(x->plane[0].src.buf, x->plane[0].src.stride,
182 : av1_all_zeros, 0, &sse);
183 : }
184 : #else
185 : var = cpi->fn_ptr[bs].vf(x->plane[0].src.buf, x->plane[0].src.stride,
186 : av1_all_zeros, 0, &sse);
187 : #endif // CONFIG_HIGHBITDEPTH
188 0 : return (unsigned int)((uint64_t)var * 256) >> num_pels_log2_lookup[bs];
189 : }
190 : }
191 :
192 0 : double av1_log_block_var(const AV1_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bs) {
193 0 : unsigned int var = block_variance(cpi, x, bs);
194 0 : aom_clear_system_state();
195 0 : return log(var + 1.0);
196 : }
197 :
198 : #define DEFAULT_E_MIDPOINT 10.0
199 0 : int av1_block_energy(const AV1_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bs) {
200 : double energy;
201 : double energy_midpoint;
202 0 : aom_clear_system_state();
203 0 : energy_midpoint =
204 0 : (cpi->oxcf.pass == 2) ? cpi->twopass.mb_av_energy : DEFAULT_E_MIDPOINT;
205 0 : energy = av1_log_block_var(cpi, x, bs) - energy_midpoint;
206 0 : return clamp((int)round(energy), ENERGY_MIN, ENERGY_MAX);
207 : }
|
