analysis.c 13.7 KB
Newer Older
a  
Kai Westerkamp committed
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421
// Copyright 2011 Google Inc. All Rights Reserved.
//
// Use of this source code is governed by a BSD-style license
// that can be found in the COPYING file in the root of the source
// tree. An additional intellectual property rights grant can be found
// in the file PATENTS. All contributing project authors may
// be found in the AUTHORS file in the root of the source tree.
// -----------------------------------------------------------------------------
//
// Macroblock analysis
//
// Author: Skal (pascal.massimino@gmail.com)

#include <stdlib.h>
#include <string.h>
#include <assert.h>

#include "./vp8enci.h"
#include "./cost.h"
#include "../utils/utils.h"

#if defined(__cplusplus) || defined(c_plusplus)
extern "C" {
#endif

#define MAX_ITERS_K_MEANS  6

//------------------------------------------------------------------------------
// Smooth the segment map by replacing isolated block by the majority of its
// neighbours.

static void SmoothSegmentMap(VP8Encoder* const enc) {
  int n, x, y;
  const int w = enc->mb_w_;
  const int h = enc->mb_h_;
  const int majority_cnt_3_x_3_grid = 5;
  uint8_t* const tmp = (uint8_t*)WebPSafeMalloc((uint64_t)w * h, sizeof(*tmp));
  assert((uint64_t)(w * h) == (uint64_t)w * h);   // no overflow, as per spec

  if (tmp == NULL) return;
  for (y = 1; y < h - 1; ++y) {
    for (x = 1; x < w - 1; ++x) {
      int cnt[NUM_MB_SEGMENTS] = { 0 };
      const VP8MBInfo* const mb = &enc->mb_info_[x + w * y];
      int majority_seg = mb->segment_;
      // Check the 8 neighbouring segment values.
      cnt[mb[-w - 1].segment_]++;  // top-left
      cnt[mb[-w + 0].segment_]++;  // top
      cnt[mb[-w + 1].segment_]++;  // top-right
      cnt[mb[   - 1].segment_]++;  // left
      cnt[mb[   + 1].segment_]++;  // right
      cnt[mb[ w - 1].segment_]++;  // bottom-left
      cnt[mb[ w + 0].segment_]++;  // bottom
      cnt[mb[ w + 1].segment_]++;  // bottom-right
      for (n = 0; n < NUM_MB_SEGMENTS; ++n) {
        if (cnt[n] >= majority_cnt_3_x_3_grid) {
          majority_seg = n;
        }
      }
      tmp[x + y * w] = majority_seg;
    }
  }
  for (y = 1; y < h - 1; ++y) {
    for (x = 1; x < w - 1; ++x) {
      VP8MBInfo* const mb = &enc->mb_info_[x + w * y];
      mb->segment_ = tmp[x + y * w];
    }
  }
  free(tmp);
}

//------------------------------------------------------------------------------
// set segment susceptibility alpha_ / beta_

static WEBP_INLINE int clip(int v, int m, int M) {
  return (v < m) ? m : (v > M) ? M : v;
}

static void SetSegmentAlphas(VP8Encoder* const enc,
                             const int centers[NUM_MB_SEGMENTS],
                             int mid) {
  const int nb = enc->segment_hdr_.num_segments_;
  int min = centers[0], max = centers[0];
  int n;

  if (nb > 1) {
    for (n = 0; n < nb; ++n) {
      if (min > centers[n]) min = centers[n];
      if (max < centers[n]) max = centers[n];
    }
  }
  if (max == min) max = min + 1;
  assert(mid <= max && mid >= min);
  for (n = 0; n < nb; ++n) {
    const int alpha = 255 * (centers[n] - mid) / (max - min);
    const int beta = 255 * (centers[n] - min) / (max - min);
    enc->dqm_[n].alpha_ = clip(alpha, -127, 127);
    enc->dqm_[n].beta_ = clip(beta, 0, 255);
  }
}

//------------------------------------------------------------------------------
// Compute susceptibility based on DCT-coeff histograms:
// the higher, the "easier" the macroblock is to compress.

#define MAX_ALPHA 255                // 8b of precision for susceptibilities.
#define ALPHA_SCALE (2 * MAX_ALPHA)  // scaling factor for alpha.
#define DEFAULT_ALPHA (-1)
#define IS_BETTER_ALPHA(alpha, best_alpha) ((alpha) > (best_alpha))

static int FinalAlphaValue(int alpha) {
  alpha = MAX_ALPHA - alpha;
  return clip(alpha, 0, MAX_ALPHA);
}

static int GetAlpha(const VP8Histogram* const histo) {
  int max_value = 0, last_non_zero = 1;
  int k;
  int alpha;
  for (k = 0; k <= MAX_COEFF_THRESH; ++k) {
    const int value = histo->distribution[k];
    if (value > 0) {
      if (value > max_value) max_value = value;
      last_non_zero = k;
    }
  }
  // 'alpha' will later be clipped to [0..MAX_ALPHA] range, clamping outer
  // values which happen to be mostly noise. This leaves the maximum precision
  // for handling the useful small values which contribute most.
  alpha = (max_value > 1) ? ALPHA_SCALE * last_non_zero / max_value : 0;
  return alpha;
}

static void MergeHistograms(const VP8Histogram* const in,
                            VP8Histogram* const out) {
  int i;
  for (i = 0; i <= MAX_COEFF_THRESH; ++i) {
    out->distribution[i] += in->distribution[i];
  }
}

//------------------------------------------------------------------------------
// Simplified k-Means, to assign Nb segments based on alpha-histogram

static void AssignSegments(VP8Encoder* const enc,
                           const int alphas[MAX_ALPHA + 1]) {
  const int nb = enc->segment_hdr_.num_segments_;
  int centers[NUM_MB_SEGMENTS];
  int weighted_average = 0;
  int map[MAX_ALPHA + 1];
  int a, n, k;
  int min_a = 0, max_a = MAX_ALPHA, range_a;
  // 'int' type is ok for histo, and won't overflow
  int accum[NUM_MB_SEGMENTS], dist_accum[NUM_MB_SEGMENTS];

  // bracket the input
  for (n = 0; n <= MAX_ALPHA && alphas[n] == 0; ++n) {}
  min_a = n;
  for (n = MAX_ALPHA; n > min_a && alphas[n] == 0; --n) {}
  max_a = n;
  range_a = max_a - min_a;

  // Spread initial centers evenly
  for (n = 1, k = 0; n < 2 * nb; n += 2) {
    centers[k++] = min_a + (n * range_a) / (2 * nb);
  }

  for (k = 0; k < MAX_ITERS_K_MEANS; ++k) {     // few iters are enough
    int total_weight;
    int displaced;
    // Reset stats
    for (n = 0; n < nb; ++n) {
      accum[n] = 0;
      dist_accum[n] = 0;
    }
    // Assign nearest center for each 'a'
    n = 0;    // track the nearest center for current 'a'
    for (a = min_a; a <= max_a; ++a) {
      if (alphas[a]) {
        while (n < nb - 1 && abs(a - centers[n + 1]) < abs(a - centers[n])) {
          n++;
        }
        map[a] = n;
        // accumulate contribution into best centroid
        dist_accum[n] += a * alphas[a];
        accum[n] += alphas[a];
      }
    }
    // All point are classified. Move the centroids to the
    // center of their respective cloud.
    displaced = 0;
    weighted_average = 0;
    total_weight = 0;
    for (n = 0; n < nb; ++n) {
      if (accum[n]) {
        const int new_center = (dist_accum[n] + accum[n] / 2) / accum[n];
        displaced += abs(centers[n] - new_center);
        centers[n] = new_center;
        weighted_average += new_center * accum[n];
        total_weight += accum[n];
      }
    }
    weighted_average = (weighted_average + total_weight / 2) / total_weight;
    if (displaced < 5) break;   // no need to keep on looping...
  }

  // Map each original value to the closest centroid
  for (n = 0; n < enc->mb_w_ * enc->mb_h_; ++n) {
    VP8MBInfo* const mb = &enc->mb_info_[n];
    const int alpha = mb->alpha_;
    mb->segment_ = map[alpha];
    mb->alpha_ = centers[map[alpha]];  // for the record.
  }

  if (nb > 1) {
    const int smooth = (enc->config_->preprocessing & 1);
    if (smooth) SmoothSegmentMap(enc);
  }

  SetSegmentAlphas(enc, centers, weighted_average);  // pick some alphas.
}

//------------------------------------------------------------------------------
// Macroblock analysis: collect histogram for each mode, deduce the maximal
// susceptibility and set best modes for this macroblock.
// Segment assignment is done later.

// Number of modes to inspect for alpha_ evaluation. For high-quality settings
// (method >= FAST_ANALYSIS_METHOD) we don't need to test all the possible modes
// during the analysis phase.
#define FAST_ANALYSIS_METHOD 4  // method above which we do partial analysis
#define MAX_INTRA16_MODE 2
#define MAX_INTRA4_MODE  2
#define MAX_UV_MODE      2

static int MBAnalyzeBestIntra16Mode(VP8EncIterator* const it) {
  const int max_mode =
      (it->enc_->method_ >= FAST_ANALYSIS_METHOD) ? MAX_INTRA16_MODE
                                                  : NUM_PRED_MODES;
  int mode;
  int best_alpha = DEFAULT_ALPHA;
  int best_mode = 0;

  VP8MakeLuma16Preds(it);
  for (mode = 0; mode < max_mode; ++mode) {
    VP8Histogram histo = { { 0 } };
    int alpha;

    VP8CollectHistogram(it->yuv_in_ + Y_OFF,
                        it->yuv_p_ + VP8I16ModeOffsets[mode],
                        0, 16, &histo);
    alpha = GetAlpha(&histo);
    if (IS_BETTER_ALPHA(alpha, best_alpha)) {
      best_alpha = alpha;
      best_mode = mode;
    }
  }
  VP8SetIntra16Mode(it, best_mode);
  return best_alpha;
}

static int MBAnalyzeBestIntra4Mode(VP8EncIterator* const it,
                                   int best_alpha) {
  uint8_t modes[16];
  const int max_mode =
      (it->enc_->method_ >= FAST_ANALYSIS_METHOD) ? MAX_INTRA4_MODE
                                                  : NUM_BMODES;
  int i4_alpha;
  VP8Histogram total_histo = { { 0 } };
  int cur_histo = 0;

  VP8IteratorStartI4(it);
  do {
    int mode;
    int best_mode_alpha = DEFAULT_ALPHA;
    VP8Histogram histos[2];
    const uint8_t* const src = it->yuv_in_ + Y_OFF + VP8Scan[it->i4_];

    VP8MakeIntra4Preds(it);
    for (mode = 0; mode < max_mode; ++mode) {
      int alpha;

      memset(&histos[cur_histo], 0, sizeof(histos[cur_histo]));
      VP8CollectHistogram(src, it->yuv_p_ + VP8I4ModeOffsets[mode],
                          0, 1, &histos[cur_histo]);
      alpha = GetAlpha(&histos[cur_histo]);
      if (IS_BETTER_ALPHA(alpha, best_mode_alpha)) {
        best_mode_alpha = alpha;
        modes[it->i4_] = mode;
        cur_histo ^= 1;   // keep track of best histo so far.
      }
    }
    // accumulate best histogram
    MergeHistograms(&histos[cur_histo ^ 1], &total_histo);
    // Note: we reuse the original samples for predictors
  } while (VP8IteratorRotateI4(it, it->yuv_in_ + Y_OFF));

  i4_alpha = GetAlpha(&total_histo);
  if (IS_BETTER_ALPHA(i4_alpha, best_alpha)) {
    VP8SetIntra4Mode(it, modes);
    best_alpha = i4_alpha;
  }
  return best_alpha;
}

static int MBAnalyzeBestUVMode(VP8EncIterator* const it) {
  int best_alpha = DEFAULT_ALPHA;
  int best_mode = 0;
  const int max_mode =
      (it->enc_->method_ >= FAST_ANALYSIS_METHOD) ? MAX_UV_MODE
                                                  : NUM_PRED_MODES;
  int mode;
  VP8MakeChroma8Preds(it);
  for (mode = 0; mode < max_mode; ++mode) {
    VP8Histogram histo = { { 0 } };
    int alpha;
    VP8CollectHistogram(it->yuv_in_ + U_OFF,
                        it->yuv_p_ + VP8UVModeOffsets[mode],
                        16, 16 + 4 + 4, &histo);
    alpha = GetAlpha(&histo);
    if (IS_BETTER_ALPHA(alpha, best_alpha)) {
      best_alpha = alpha;
      best_mode = mode;
    }
  }
  VP8SetIntraUVMode(it, best_mode);
  return best_alpha;
}

static void MBAnalyze(VP8EncIterator* const it,
                      int alphas[MAX_ALPHA + 1],
                      int* const alpha, int* const uv_alpha) {
  const VP8Encoder* const enc = it->enc_;
  int best_alpha, best_uv_alpha;

  VP8SetIntra16Mode(it, 0);  // default: Intra16, DC_PRED
  VP8SetSkip(it, 0);         // not skipped
  VP8SetSegment(it, 0);      // default segment, spec-wise.

  best_alpha = MBAnalyzeBestIntra16Mode(it);
  if (enc->method_ >= 5) {
    // We go and make a fast decision for intra4/intra16.
    // It's usually not a good and definitive pick, but helps seeding the stats
    // about level bit-cost.
    // TODO(skal): improve criterion.
    best_alpha = MBAnalyzeBestIntra4Mode(it, best_alpha);
  }
  best_uv_alpha = MBAnalyzeBestUVMode(it);

  // Final susceptibility mix
  best_alpha = (3 * best_alpha + best_uv_alpha + 2) >> 2;
  best_alpha = FinalAlphaValue(best_alpha);
  alphas[best_alpha]++;
  it->mb_->alpha_ = best_alpha;   // for later remapping.

  // Accumulate for later complexity analysis.
  *alpha += best_alpha;   // mixed susceptibility (not just luma)
  *uv_alpha += best_uv_alpha;
}

static void DefaultMBInfo(VP8MBInfo* const mb) {
  mb->type_ = 1;     // I16x16
  mb->uv_mode_ = 0;
  mb->skip_ = 0;     // not skipped
  mb->segment_ = 0;  // default segment
  mb->alpha_ = 0;
}

//------------------------------------------------------------------------------
// Main analysis loop:
// Collect all susceptibilities for each macroblock and record their
// distribution in alphas[]. Segments is assigned a-posteriori, based on
// this histogram.
// We also pick an intra16 prediction mode, which shouldn't be considered
// final except for fast-encode settings. We can also pick some intra4 modes
// and decide intra4/intra16, but that's usually almost always a bad choice at
// this stage.

static void ResetAllMBInfo(VP8Encoder* const enc) {
  int n;
  for (n = 0; n < enc->mb_w_ * enc->mb_h_; ++n) {
    DefaultMBInfo(&enc->mb_info_[n]);
  }
  // Default susceptibilities.
  enc->dqm_[0].alpha_ = 0;
  enc->dqm_[0].beta_ = 0;
  // Note: we can't compute this alpha_ / uv_alpha_.
  WebPReportProgress(enc->pic_, enc->percent_ + 20, &enc->percent_);
}

int VP8EncAnalyze(VP8Encoder* const enc) {
  int ok = 1;
  const int do_segments =
      enc->config_->emulate_jpeg_size ||   // We need the complexity evaluation.
      (enc->segment_hdr_.num_segments_ > 1) ||
      (enc->method_ == 0);  // for method 0, we need preds_[] to be filled.
  enc->alpha_ = 0;
  enc->uv_alpha_ = 0;
  if (do_segments) {
    int alphas[MAX_ALPHA + 1] = { 0 };
    VP8EncIterator it;

    VP8IteratorInit(enc, &it);
    do {
      VP8IteratorImport(&it);
      MBAnalyze(&it, alphas, &enc->alpha_, &enc->uv_alpha_);
      ok = VP8IteratorProgress(&it, 20);
      // Let's pretend we have perfect lossless reconstruction.
    } while (ok && VP8IteratorNext(&it, it.yuv_in_));
    enc->alpha_ /= enc->mb_w_ * enc->mb_h_;
    enc->uv_alpha_ /= enc->mb_w_ * enc->mb_h_;
    if (ok) AssignSegments(enc, alphas);
  } else {   // Use only one default segment.
    ResetAllMBInfo(enc);
  }
  return ok;
}

#if defined(__cplusplus) || defined(c_plusplus)
}    // extern "C"
#endif