1 // Copyright 2011 Google Inc. All Rights Reserved. 2 // 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 // ----------------------------------------------------------------------------- 9 // 10 // Macroblock analysis 11 // 12 // Author: Skal (pascal.massimino (at) gmail.com) 13 14 #include <stdlib.h> 15 #include <string.h> 16 #include <assert.h> 17 18 #include "./vp8enci.h" 19 #include "./cost.h" 20 #include "../utils/utils.h" 21 22 #if defined(__cplusplus) || defined(c_plusplus) 23 extern "C" { 24 #endif 25 26 #define MAX_ITERS_K_MEANS 6 27 28 //------------------------------------------------------------------------------ 29 // Smooth the segment map by replacing isolated block by the majority of its 30 // neighbours. 31 32 static void SmoothSegmentMap(VP8Encoder* const enc) { 33 int n, x, y; 34 const int w = enc->mb_w_; 35 const int h = enc->mb_h_; 36 const int majority_cnt_3_x_3_grid = 5; 37 uint8_t* const tmp = (uint8_t*)WebPSafeMalloc((uint64_t)w * h, sizeof(*tmp)); 38 assert((uint64_t)(w * h) == (uint64_t)w * h); // no overflow, as per spec 39 40 if (tmp == NULL) return; 41 for (y = 1; y < h - 1; ++y) { 42 for (x = 1; x < w - 1; ++x) { 43 int cnt[NUM_MB_SEGMENTS] = { 0 }; 44 const VP8MBInfo* const mb = &enc->mb_info_[x + w * y]; 45 int majority_seg = mb->segment_; 46 // Check the 8 neighbouring segment values. 47 cnt[mb[-w - 1].segment_]++; // top-left 48 cnt[mb[-w + 0].segment_]++; // top 49 cnt[mb[-w + 1].segment_]++; // top-right 50 cnt[mb[ - 1].segment_]++; // left 51 cnt[mb[ + 1].segment_]++; // right 52 cnt[mb[ w - 1].segment_]++; // bottom-left 53 cnt[mb[ w + 0].segment_]++; // bottom 54 cnt[mb[ w + 1].segment_]++; // bottom-right 55 for (n = 0; n < NUM_MB_SEGMENTS; ++n) { 56 if (cnt[n] >= majority_cnt_3_x_3_grid) { 57 majority_seg = n; 58 } 59 } 60 tmp[x + y * w] = majority_seg; 61 } 62 } 63 for (y = 1; y < h - 1; ++y) { 64 for (x = 1; x < w - 1; ++x) { 65 VP8MBInfo* const mb = &enc->mb_info_[x + w * y]; 66 mb->segment_ = tmp[x + y * w]; 67 } 68 } 69 free(tmp); 70 } 71 72 //------------------------------------------------------------------------------ 73 // set segment susceptibility alpha_ / beta_ 74 75 static WEBP_INLINE int clip(int v, int m, int M) { 76 return (v < m) ? m : (v > M) ? M : v; 77 } 78 79 static void SetSegmentAlphas(VP8Encoder* const enc, 80 const int centers[NUM_MB_SEGMENTS], 81 int mid) { 82 const int nb = enc->segment_hdr_.num_segments_; 83 int min = centers[0], max = centers[0]; 84 int n; 85 86 if (nb > 1) { 87 for (n = 0; n < nb; ++n) { 88 if (min > centers[n]) min = centers[n]; 89 if (max < centers[n]) max = centers[n]; 90 } 91 } 92 if (max == min) max = min + 1; 93 assert(mid <= max && mid >= min); 94 for (n = 0; n < nb; ++n) { 95 const int alpha = 255 * (centers[n] - mid) / (max - min); 96 const int beta = 255 * (centers[n] - min) / (max - min); 97 enc->dqm_[n].alpha_ = clip(alpha, -127, 127); 98 enc->dqm_[n].beta_ = clip(beta, 0, 255); 99 } 100 } 101 102 //------------------------------------------------------------------------------ 103 // Compute susceptibility based on DCT-coeff histograms: 104 // the higher, the "easier" the macroblock is to compress. 105 106 #define MAX_ALPHA 255 // 8b of precision for susceptibilities. 107 #define ALPHA_SCALE (2 * MAX_ALPHA) // scaling factor for alpha. 108 #define DEFAULT_ALPHA (-1) 109 #define IS_BETTER_ALPHA(alpha, best_alpha) ((alpha) > (best_alpha)) 110 111 static int FinalAlphaValue(int alpha) { 112 alpha = MAX_ALPHA - alpha; 113 return clip(alpha, 0, MAX_ALPHA); 114 } 115 116 static int GetAlpha(const VP8Histogram* const histo) { 117 int max_value = 0, last_non_zero = 1; 118 int k; 119 int alpha; 120 for (k = 0; k <= MAX_COEFF_THRESH; ++k) { 121 const int value = histo->distribution[k]; 122 if (value > 0) { 123 if (value > max_value) max_value = value; 124 last_non_zero = k; 125 } 126 } 127 // 'alpha' will later be clipped to [0..MAX_ALPHA] range, clamping outer 128 // values which happen to be mostly noise. This leaves the maximum precision 129 // for handling the useful small values which contribute most. 130 alpha = (max_value > 1) ? ALPHA_SCALE * last_non_zero / max_value : 0; 131 return alpha; 132 } 133 134 static void MergeHistograms(const VP8Histogram* const in, 135 VP8Histogram* const out) { 136 int i; 137 for (i = 0; i <= MAX_COEFF_THRESH; ++i) { 138 out->distribution[i] += in->distribution[i]; 139 } 140 } 141 142 //------------------------------------------------------------------------------ 143 // Simplified k-Means, to assign Nb segments based on alpha-histogram 144 145 static void AssignSegments(VP8Encoder* const enc, 146 const int alphas[MAX_ALPHA + 1]) { 147 const int nb = enc->segment_hdr_.num_segments_; 148 int centers[NUM_MB_SEGMENTS]; 149 int weighted_average = 0; 150 int map[MAX_ALPHA + 1]; 151 int a, n, k; 152 int min_a = 0, max_a = MAX_ALPHA, range_a; 153 // 'int' type is ok for histo, and won't overflow 154 int accum[NUM_MB_SEGMENTS], dist_accum[NUM_MB_SEGMENTS]; 155 156 // bracket the input 157 for (n = 0; n <= MAX_ALPHA && alphas[n] == 0; ++n) {} 158 min_a = n; 159 for (n = MAX_ALPHA; n > min_a && alphas[n] == 0; --n) {} 160 max_a = n; 161 range_a = max_a - min_a; 162 163 // Spread initial centers evenly 164 for (n = 1, k = 0; n < 2 * nb; n += 2) { 165 centers[k++] = min_a + (n * range_a) / (2 * nb); 166 } 167 168 for (k = 0; k < MAX_ITERS_K_MEANS; ++k) { // few iters are enough 169 int total_weight; 170 int displaced; 171 // Reset stats 172 for (n = 0; n < nb; ++n) { 173 accum[n] = 0; 174 dist_accum[n] = 0; 175 } 176 // Assign nearest center for each 'a' 177 n = 0; // track the nearest center for current 'a' 178 for (a = min_a; a <= max_a; ++a) { 179 if (alphas[a]) { 180 while (n < nb - 1 && abs(a - centers[n + 1]) < abs(a - centers[n])) { 181 n++; 182 } 183 map[a] = n; 184 // accumulate contribution into best centroid 185 dist_accum[n] += a * alphas[a]; 186 accum[n] += alphas[a]; 187 } 188 } 189 // All point are classified. Move the centroids to the 190 // center of their respective cloud. 191 displaced = 0; 192 weighted_average = 0; 193 total_weight = 0; 194 for (n = 0; n < nb; ++n) { 195 if (accum[n]) { 196 const int new_center = (dist_accum[n] + accum[n] / 2) / accum[n]; 197 displaced += abs(centers[n] - new_center); 198 centers[n] = new_center; 199 weighted_average += new_center * accum[n]; 200 total_weight += accum[n]; 201 } 202 } 203 weighted_average = (weighted_average + total_weight / 2) / total_weight; 204 if (displaced < 5) break; // no need to keep on looping... 205 } 206 207 // Map each original value to the closest centroid 208 for (n = 0; n < enc->mb_w_ * enc->mb_h_; ++n) { 209 VP8MBInfo* const mb = &enc->mb_info_[n]; 210 const int alpha = mb->alpha_; 211 mb->segment_ = map[alpha]; 212 mb->alpha_ = centers[map[alpha]]; // for the record. 213 } 214 215 if (nb > 1) { 216 const int smooth = (enc->config_->preprocessing & 1); 217 if (smooth) SmoothSegmentMap(enc); 218 } 219 220 SetSegmentAlphas(enc, centers, weighted_average); // pick some alphas. 221 } 222 223 //------------------------------------------------------------------------------ 224 // Macroblock analysis: collect histogram for each mode, deduce the maximal 225 // susceptibility and set best modes for this macroblock. 226 // Segment assignment is done later. 227 228 // Number of modes to inspect for alpha_ evaluation. For high-quality settings 229 // (method >= FAST_ANALYSIS_METHOD) we don't need to test all the possible modes 230 // during the analysis phase. 231 #define FAST_ANALYSIS_METHOD 4 // method above which we do partial analysis 232 #define MAX_INTRA16_MODE 2 233 #define MAX_INTRA4_MODE 2 234 #define MAX_UV_MODE 2 235 236 static int MBAnalyzeBestIntra16Mode(VP8EncIterator* const it) { 237 const int max_mode = 238 (it->enc_->method_ >= FAST_ANALYSIS_METHOD) ? MAX_INTRA16_MODE 239 : NUM_PRED_MODES; 240 int mode; 241 int best_alpha = DEFAULT_ALPHA; 242 int best_mode = 0; 243 244 VP8MakeLuma16Preds(it); 245 for (mode = 0; mode < max_mode; ++mode) { 246 VP8Histogram histo = { { 0 } }; 247 int alpha; 248 249 VP8CollectHistogram(it->yuv_in_ + Y_OFF, 250 it->yuv_p_ + VP8I16ModeOffsets[mode], 251 0, 16, &histo); 252 alpha = GetAlpha(&histo); 253 if (IS_BETTER_ALPHA(alpha, best_alpha)) { 254 best_alpha = alpha; 255 best_mode = mode; 256 } 257 } 258 VP8SetIntra16Mode(it, best_mode); 259 return best_alpha; 260 } 261 262 static int MBAnalyzeBestIntra4Mode(VP8EncIterator* const it, 263 int best_alpha) { 264 uint8_t modes[16]; 265 const int max_mode = 266 (it->enc_->method_ >= FAST_ANALYSIS_METHOD) ? MAX_INTRA4_MODE 267 : NUM_BMODES; 268 int i4_alpha; 269 VP8Histogram total_histo = { { 0 } }; 270 int cur_histo = 0; 271 272 VP8IteratorStartI4(it); 273 do { 274 int mode; 275 int best_mode_alpha = DEFAULT_ALPHA; 276 VP8Histogram histos[2]; 277 const uint8_t* const src = it->yuv_in_ + Y_OFF + VP8Scan[it->i4_]; 278 279 VP8MakeIntra4Preds(it); 280 for (mode = 0; mode < max_mode; ++mode) { 281 int alpha; 282 283 memset(&histos[cur_histo], 0, sizeof(histos[cur_histo])); 284 VP8CollectHistogram(src, it->yuv_p_ + VP8I4ModeOffsets[mode], 285 0, 1, &histos[cur_histo]); 286 alpha = GetAlpha(&histos[cur_histo]); 287 if (IS_BETTER_ALPHA(alpha, best_mode_alpha)) { 288 best_mode_alpha = alpha; 289 modes[it->i4_] = mode; 290 cur_histo ^= 1; // keep track of best histo so far. 291 } 292 } 293 // accumulate best histogram 294 MergeHistograms(&histos[cur_histo ^ 1], &total_histo); 295 // Note: we reuse the original samples for predictors 296 } while (VP8IteratorRotateI4(it, it->yuv_in_ + Y_OFF)); 297 298 i4_alpha = GetAlpha(&total_histo); 299 if (IS_BETTER_ALPHA(i4_alpha, best_alpha)) { 300 VP8SetIntra4Mode(it, modes); 301 best_alpha = i4_alpha; 302 } 303 return best_alpha; 304 } 305 306 static int MBAnalyzeBestUVMode(VP8EncIterator* const it) { 307 int best_alpha = DEFAULT_ALPHA; 308 int best_mode = 0; 309 const int max_mode = 310 (it->enc_->method_ >= FAST_ANALYSIS_METHOD) ? MAX_UV_MODE 311 : NUM_PRED_MODES; 312 int mode; 313 VP8MakeChroma8Preds(it); 314 for (mode = 0; mode < max_mode; ++mode) { 315 VP8Histogram histo = { { 0 } }; 316 int alpha; 317 VP8CollectHistogram(it->yuv_in_ + U_OFF, 318 it->yuv_p_ + VP8UVModeOffsets[mode], 319 16, 16 + 4 + 4, &histo); 320 alpha = GetAlpha(&histo); 321 if (IS_BETTER_ALPHA(alpha, best_alpha)) { 322 best_alpha = alpha; 323 best_mode = mode; 324 } 325 } 326 VP8SetIntraUVMode(it, best_mode); 327 return best_alpha; 328 } 329 330 static void MBAnalyze(VP8EncIterator* const it, 331 int alphas[MAX_ALPHA + 1], 332 int* const alpha, int* const uv_alpha) { 333 const VP8Encoder* const enc = it->enc_; 334 int best_alpha, best_uv_alpha; 335 336 VP8SetIntra16Mode(it, 0); // default: Intra16, DC_PRED 337 VP8SetSkip(it, 0); // not skipped 338 VP8SetSegment(it, 0); // default segment, spec-wise. 339 340 best_alpha = MBAnalyzeBestIntra16Mode(it); 341 if (enc->method_ >= 5) { 342 // We go and make a fast decision for intra4/intra16. 343 // It's usually not a good and definitive pick, but helps seeding the stats 344 // about level bit-cost. 345 // TODO(skal): improve criterion. 346 best_alpha = MBAnalyzeBestIntra4Mode(it, best_alpha); 347 } 348 best_uv_alpha = MBAnalyzeBestUVMode(it); 349 350 // Final susceptibility mix 351 best_alpha = (3 * best_alpha + best_uv_alpha + 2) >> 2; 352 best_alpha = FinalAlphaValue(best_alpha); 353 alphas[best_alpha]++; 354 it->mb_->alpha_ = best_alpha; // for later remapping. 355 356 // Accumulate for later complexity analysis. 357 *alpha += best_alpha; // mixed susceptibility (not just luma) 358 *uv_alpha += best_uv_alpha; 359 } 360 361 static void DefaultMBInfo(VP8MBInfo* const mb) { 362 mb->type_ = 1; // I16x16 363 mb->uv_mode_ = 0; 364 mb->skip_ = 0; // not skipped 365 mb->segment_ = 0; // default segment 366 mb->alpha_ = 0; 367 } 368 369 //------------------------------------------------------------------------------ 370 // Main analysis loop: 371 // Collect all susceptibilities for each macroblock and record their 372 // distribution in alphas[]. Segments is assigned a-posteriori, based on 373 // this histogram. 374 // We also pick an intra16 prediction mode, which shouldn't be considered 375 // final except for fast-encode settings. We can also pick some intra4 modes 376 // and decide intra4/intra16, but that's usually almost always a bad choice at 377 // this stage. 378 379 static void ResetAllMBInfo(VP8Encoder* const enc) { 380 int n; 381 for (n = 0; n < enc->mb_w_ * enc->mb_h_; ++n) { 382 DefaultMBInfo(&enc->mb_info_[n]); 383 } 384 // Default susceptibilities. 385 enc->dqm_[0].alpha_ = 0; 386 enc->dqm_[0].beta_ = 0; 387 // Note: we can't compute this alpha_ / uv_alpha_. 388 WebPReportProgress(enc->pic_, enc->percent_ + 20, &enc->percent_); 389 } 390 391 int VP8EncAnalyze(VP8Encoder* const enc) { 392 int ok = 1; 393 const int do_segments = 394 enc->config_->emulate_jpeg_size || // We need the complexity evaluation. 395 (enc->segment_hdr_.num_segments_ > 1) || 396 (enc->method_ == 0); // for method 0, we need preds_[] to be filled. 397 enc->alpha_ = 0; 398 enc->uv_alpha_ = 0; 399 if (do_segments) { 400 int alphas[MAX_ALPHA + 1] = { 0 }; 401 VP8EncIterator it; 402 403 VP8IteratorInit(enc, &it); 404 do { 405 VP8IteratorImport(&it); 406 MBAnalyze(&it, alphas, &enc->alpha_, &enc->uv_alpha_); 407 ok = VP8IteratorProgress(&it, 20); 408 // Let's pretend we have perfect lossless reconstruction. 409 } while (ok && VP8IteratorNext(&it, it.yuv_in_)); 410 enc->alpha_ /= enc->mb_w_ * enc->mb_h_; 411 enc->uv_alpha_ /= enc->mb_w_ * enc->mb_h_; 412 if (ok) AssignSegments(enc, alphas); 413 } else { // Use only one default segment. 414 ResetAllMBInfo(enc); 415 } 416 return ok; 417 } 418 419 #if defined(__cplusplus) || defined(c_plusplus) 420 } // extern "C" 421 #endif 422