1
// Copyright 2011 Google Inc. All Rights Reserved.
3
// Use of this source code is governed by a BSD-style license
4
// that can be found in the COPYING file in the root of the source
5
// tree. An additional intellectual property rights grant can be found
6
// in the file PATENTS. All contributing project authors may
7
// be found in the AUTHORS file in the root of the source tree.
8
// -----------------------------------------------------------------------------
10
// Selecting filter level
12
// Author: somnath@google.com (Somnath Banerjee)
14
#include "./vp8enci.h"
16
#if defined(__cplusplus) || defined(c_plusplus)
20
// NOTE: clip1, tables and InitTables are repeated entries of dsp.c
21
static uint8_t abs0[255 + 255 + 1]; // abs(i)
22
static uint8_t abs1[255 + 255 + 1]; // abs(i)>>1
23
static int8_t sclip1[1020 + 1020 + 1]; // clips [-1020, 1020] to [-128, 127]
24
static int8_t sclip2[112 + 112 + 1]; // clips [-112, 112] to [-16, 15]
25
static uint8_t clip1[255 + 510 + 1]; // clips [-255,510] to [0,255]
27
static int tables_ok = 0;
29
static void InitTables(void) {
32
for (i = -255; i <= 255; ++i) {
33
abs0[255 + i] = (i < 0) ? -i : i;
34
abs1[255 + i] = abs0[255 + i] >> 1;
36
for (i = -1020; i <= 1020; ++i) {
37
sclip1[1020 + i] = (i < -128) ? -128 : (i > 127) ? 127 : i;
39
for (i = -112; i <= 112; ++i) {
40
sclip2[112 + i] = (i < -16) ? -16 : (i > 15) ? 15 : i;
42
for (i = -255; i <= 255 + 255; ++i) {
43
clip1[255 + i] = (i < 0) ? 0 : (i > 255) ? 255 : i;
49
//------------------------------------------------------------------------------
50
// Edge filtering functions
52
// 4 pixels in, 2 pixels out
53
static WEBP_INLINE void do_filter2(uint8_t* p, int step) {
54
const int p1 = p[-2*step], p0 = p[-step], q0 = p[0], q1 = p[step];
55
const int a = 3 * (q0 - p0) + sclip1[1020 + p1 - q1];
56
const int a1 = sclip2[112 + ((a + 4) >> 3)];
57
const int a2 = sclip2[112 + ((a + 3) >> 3)];
58
p[-step] = clip1[255 + p0 + a2];
59
p[ 0] = clip1[255 + q0 - a1];
62
// 4 pixels in, 4 pixels out
63
static WEBP_INLINE void do_filter4(uint8_t* p, int step) {
64
const int p1 = p[-2*step], p0 = p[-step], q0 = p[0], q1 = p[step];
65
const int a = 3 * (q0 - p0);
66
const int a1 = sclip2[112 + ((a + 4) >> 3)];
67
const int a2 = sclip2[112 + ((a + 3) >> 3)];
68
const int a3 = (a1 + 1) >> 1;
69
p[-2*step] = clip1[255 + p1 + a3];
70
p[- step] = clip1[255 + p0 + a2];
71
p[ 0] = clip1[255 + q0 - a1];
72
p[ step] = clip1[255 + q1 - a3];
76
static WEBP_INLINE int hev(const uint8_t* p, int step, int thresh) {
77
const int p1 = p[-2*step], p0 = p[-step], q0 = p[0], q1 = p[step];
78
return (abs0[255 + p1 - p0] > thresh) || (abs0[255 + q1 - q0] > thresh);
81
static WEBP_INLINE int needs_filter(const uint8_t* p, int step, int thresh) {
82
const int p1 = p[-2*step], p0 = p[-step], q0 = p[0], q1 = p[step];
83
return (2 * abs0[255 + p0 - q0] + abs1[255 + p1 - q1]) <= thresh;
86
static WEBP_INLINE int needs_filter2(const uint8_t* p,
87
int step, int t, int it) {
88
const int p3 = p[-4*step], p2 = p[-3*step], p1 = p[-2*step], p0 = p[-step];
89
const int q0 = p[0], q1 = p[step], q2 = p[2*step], q3 = p[3*step];
90
if ((2 * abs0[255 + p0 - q0] + abs1[255 + p1 - q1]) > t)
92
return abs0[255 + p3 - p2] <= it && abs0[255 + p2 - p1] <= it &&
93
abs0[255 + p1 - p0] <= it && abs0[255 + q3 - q2] <= it &&
94
abs0[255 + q2 - q1] <= it && abs0[255 + q1 - q0] <= it;
97
//------------------------------------------------------------------------------
98
// Simple In-loop filtering (Paragraph 15.2)
100
static void SimpleVFilter16(uint8_t* p, int stride, int thresh) {
102
for (i = 0; i < 16; ++i) {
103
if (needs_filter(p + i, stride, thresh)) {
104
do_filter2(p + i, stride);
109
static void SimpleHFilter16(uint8_t* p, int stride, int thresh) {
111
for (i = 0; i < 16; ++i) {
112
if (needs_filter(p + i * stride, 1, thresh)) {
113
do_filter2(p + i * stride, 1);
118
static void SimpleVFilter16i(uint8_t* p, int stride, int thresh) {
120
for (k = 3; k > 0; --k) {
122
SimpleVFilter16(p, stride, thresh);
126
static void SimpleHFilter16i(uint8_t* p, int stride, int thresh) {
128
for (k = 3; k > 0; --k) {
130
SimpleHFilter16(p, stride, thresh);
134
//------------------------------------------------------------------------------
135
// Complex In-loop filtering (Paragraph 15.3)
137
static WEBP_INLINE void FilterLoop24(uint8_t* p,
138
int hstride, int vstride, int size,
139
int thresh, int ithresh, int hev_thresh) {
141
if (needs_filter2(p, hstride, thresh, ithresh)) {
142
if (hev(p, hstride, hev_thresh)) {
143
do_filter2(p, hstride);
145
do_filter4(p, hstride);
152
// on three inner edges
153
static void VFilter16i(uint8_t* p, int stride,
154
int thresh, int ithresh, int hev_thresh) {
156
for (k = 3; k > 0; --k) {
158
FilterLoop24(p, stride, 1, 16, thresh, ithresh, hev_thresh);
162
static void HFilter16i(uint8_t* p, int stride,
163
int thresh, int ithresh, int hev_thresh) {
165
for (k = 3; k > 0; --k) {
167
FilterLoop24(p, 1, stride, 16, thresh, ithresh, hev_thresh);
171
static void VFilter8i(uint8_t* u, uint8_t* v, int stride,
172
int thresh, int ithresh, int hev_thresh) {
173
FilterLoop24(u + 4 * stride, stride, 1, 8, thresh, ithresh, hev_thresh);
174
FilterLoop24(v + 4 * stride, stride, 1, 8, thresh, ithresh, hev_thresh);
177
static void HFilter8i(uint8_t* u, uint8_t* v, int stride,
178
int thresh, int ithresh, int hev_thresh) {
179
FilterLoop24(u + 4, 1, stride, 8, thresh, ithresh, hev_thresh);
180
FilterLoop24(v + 4, 1, stride, 8, thresh, ithresh, hev_thresh);
183
//------------------------------------------------------------------------------
185
void (*VP8EncVFilter16i)(uint8_t*, int, int, int, int) = VFilter16i;
186
void (*VP8EncHFilter16i)(uint8_t*, int, int, int, int) = HFilter16i;
187
void (*VP8EncVFilter8i)(uint8_t*, uint8_t*, int, int, int, int) = VFilter8i;
188
void (*VP8EncHFilter8i)(uint8_t*, uint8_t*, int, int, int, int) = HFilter8i;
190
void (*VP8EncSimpleVFilter16i)(uint8_t*, int, int) = SimpleVFilter16i;
191
void (*VP8EncSimpleHFilter16i)(uint8_t*, int, int) = SimpleHFilter16i;
193
//------------------------------------------------------------------------------
194
// Paragraph 15.4: compute the inner-edge filtering strength
196
static int GetILevel(int sharpness, int level) {
203
if (level > 9 - sharpness) {
204
level = 9 - sharpness;
207
if (level < 1) level = 1;
211
static void DoFilter(const VP8EncIterator* const it, int level) {
212
const VP8Encoder* const enc = it->enc_;
213
const int ilevel = GetILevel(enc->config_->filter_sharpness, level);
214
const int limit = 2 * level + ilevel;
216
uint8_t* const y_dst = it->yuv_out2_ + Y_OFF;
217
uint8_t* const u_dst = it->yuv_out2_ + U_OFF;
218
uint8_t* const v_dst = it->yuv_out2_ + V_OFF;
220
// copy current block to yuv_out2_
221
memcpy(y_dst, it->yuv_out_, YUV_SIZE * sizeof(uint8_t));
223
if (enc->filter_hdr_.simple_ == 1) { // simple
224
VP8EncSimpleHFilter16i(y_dst, BPS, limit);
225
VP8EncSimpleVFilter16i(y_dst, BPS, limit);
227
const int hev_thresh = (level >= 40) ? 2 : (level >= 15) ? 1 : 0;
228
VP8EncHFilter16i(y_dst, BPS, limit, ilevel, hev_thresh);
229
VP8EncHFilter8i(u_dst, v_dst, BPS, limit, ilevel, hev_thresh);
230
VP8EncVFilter16i(y_dst, BPS, limit, ilevel, hev_thresh);
231
VP8EncVFilter8i(u_dst, v_dst, BPS, limit, ilevel, hev_thresh);
235
//------------------------------------------------------------------------------
239
static const double kMinValue = 1.e-10; // minimal threshold
241
void VP8SSIMAddStats(const DistoStats* const src, DistoStats* const dst) {
245
dst->xxm += src->xxm;
246
dst->xym += src->xym;
247
dst->yym += src->yym;
250
static void VP8SSIMAccumulate(const uint8_t* src1, int stride1,
251
const uint8_t* src2, int stride2,
252
int xo, int yo, int W, int H,
253
DistoStats* const stats) {
254
const int ymin = (yo - KERNEL < 0) ? 0 : yo - KERNEL;
255
const int ymax = (yo + KERNEL > H - 1) ? H - 1 : yo + KERNEL;
256
const int xmin = (xo - KERNEL < 0) ? 0 : xo - KERNEL;
257
const int xmax = (xo + KERNEL > W - 1) ? W - 1 : xo + KERNEL;
259
src1 += ymin * stride1;
260
src2 += ymin * stride2;
261
for (y = ymin; y <= ymax; ++y, src1 += stride1, src2 += stride2) {
262
for (x = xmin; x <= xmax; ++x) {
263
const int s1 = src1[x];
264
const int s2 = src2[x];
268
stats->xxm += s1 * s1;
269
stats->xym += s1 * s2;
270
stats->yym += s2 * s2;
275
double VP8SSIMGet(const DistoStats* const stats) {
276
const double xmxm = stats->xm * stats->xm;
277
const double ymym = stats->ym * stats->ym;
278
const double xmym = stats->xm * stats->ym;
279
const double w2 = stats->w * stats->w;
280
double sxx = stats->xxm * stats->w - xmxm;
281
double syy = stats->yym * stats->w - ymym;
282
double sxy = stats->xym * stats->w - xmym;
286
// small errors are possible, due to rounding. Clamp to zero.
287
if (sxx < 0.) sxx = 0.;
288
if (syy < 0.) syy = 0.;
291
fnum = (2 * xmym + C1) * (2 * sxy + C2);
292
fden = (xmxm + ymym + C1) * (sxx + syy + C2);
293
return (fden != 0.) ? fnum / fden : kMinValue;
296
double VP8SSIMGetSquaredError(const DistoStats* const s) {
298
const double iw2 = 1. / (s->w * s->w);
299
const double sxx = s->xxm * s->w - s->xm * s->xm;
300
const double syy = s->yym * s->w - s->ym * s->ym;
301
const double sxy = s->xym * s->w - s->xm * s->ym;
302
const double SSE = iw2 * (sxx + syy - 2. * sxy);
303
if (SSE > kMinValue) return SSE;
308
void VP8SSIMAccumulatePlane(const uint8_t* src1, int stride1,
309
const uint8_t* src2, int stride2,
310
int W, int H, DistoStats* const stats) {
312
for (y = 0; y < H; ++y) {
313
for (x = 0; x < W; ++x) {
314
VP8SSIMAccumulate(src1, stride1, src2, stride2, x, y, W, H, stats);
319
static double GetMBSSIM(const uint8_t* yuv1, const uint8_t* yuv2) {
321
DistoStats s = { .0, .0, .0, .0, .0, .0 };
323
// compute SSIM in a 10 x 10 window
324
for (x = 3; x < 13; x++) {
325
for (y = 3; y < 13; y++) {
326
VP8SSIMAccumulate(yuv1 + Y_OFF, BPS, yuv2 + Y_OFF, BPS, x, y, 16, 16, &s);
329
for (x = 1; x < 7; x++) {
330
for (y = 1; y < 7; y++) {
331
VP8SSIMAccumulate(yuv1 + U_OFF, BPS, yuv2 + U_OFF, BPS, x, y, 8, 8, &s);
332
VP8SSIMAccumulate(yuv1 + V_OFF, BPS, yuv2 + V_OFF, BPS, x, y, 8, 8, &s);
335
return VP8SSIMGet(&s);
338
//------------------------------------------------------------------------------
339
// Exposed APIs: Encoder should call the following 3 functions to adjust
340
// loop filter strength
342
void VP8InitFilter(VP8EncIterator* const it) {
344
if (!it->lf_stats_) return;
347
for (s = 0; s < NUM_MB_SEGMENTS; s++) {
348
for (i = 0; i < MAX_LF_LEVELS; i++) {
349
(*it->lf_stats_)[s][i] = 0;
354
void VP8StoreFilterStats(VP8EncIterator* const it) {
356
const int s = it->mb_->segment_;
357
const int level0 = it->enc_->dqm_[s].fstrength_; // TODO: ref_lf_delta[]
359
// explore +/-quant range of values around level0
360
const int delta_min = -it->enc_->dqm_[s].quant_;
361
const int delta_max = it->enc_->dqm_[s].quant_;
362
const int step_size = (delta_max - delta_min >= 4) ? 4 : 1;
364
if (!it->lf_stats_) return;
366
// NOTE: Currently we are applying filter only across the sublock edges
367
// There are two reasons for that.
368
// 1. Applying filter on macro block edges will change the pixels in
369
// the left and top macro blocks. That will be hard to restore
370
// 2. Macro Blocks on the bottom and right are not yet compressed. So we
371
// cannot apply filter on the right and bottom macro block edges.
372
if (it->mb_->type_ == 1 && it->mb_->skip_) return;
374
// Always try filter level zero
375
(*it->lf_stats_)[s][0] += GetMBSSIM(it->yuv_in_, it->yuv_out_);
377
for (d = delta_min; d <= delta_max; d += step_size) {
378
const int level = level0 + d;
379
if (level <= 0 || level >= MAX_LF_LEVELS) {
383
(*it->lf_stats_)[s][level] += GetMBSSIM(it->yuv_in_, it->yuv_out2_);
387
void VP8AdjustFilterStrength(VP8EncIterator* const it) {
389
VP8Encoder* const enc = it->enc_;
391
if (!it->lf_stats_) {
394
for (s = 0; s < NUM_MB_SEGMENTS; s++) {
395
int i, best_level = 0;
396
// Improvement over filter level 0 should be at least 1e-5 (relatively)
397
double best_v = 1.00001 * (*it->lf_stats_)[s][0];
398
for (i = 1; i < MAX_LF_LEVELS; i++) {
399
const double v = (*it->lf_stats_)[s][i];
405
enc->dqm_[s].fstrength_ = best_level;
409
#if defined(__cplusplus) || defined(c_plusplus)