1// Copyright 2016 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// Image transform methods for lossless encoder.
11//
12// Authors: Vikas Arora (vikaas.arora@gmail.com)
13// Jyrki Alakuijala (jyrki@google.com)
14// Urvang Joshi (urvang@google.com)
15// Vincent Rabaud (vrabaud@google.com)
16
17#include "../dsp/lossless.h"
18#include "../dsp/lossless_common.h"
19#include "./vp8li_enc.h"
20
21#define MAX_DIFF_COST (1e30f)
22
23static const float kSpatialPredictorBias = 15.f;
24static const int kPredLowEffort = 11;
25static const uint32_t kMaskAlpha = 0xff000000;
26
27// Mostly used to reduce code size + readability
28static WEBP_INLINE int GetMin(int a, int b) { return (a > b) ? b : a; }
29static WEBP_INLINE int GetMax(int a, int b) { return (a < b) ? b : a; }
30
31//------------------------------------------------------------------------------
32// Methods to calculate Entropy (Shannon).
33
34static float PredictionCostSpatial(const int counts[256], int weight_0,
35 double exp_val) {
36 const int significant_symbols = 256 >> 4;
37 const double exp_decay_factor = 0.6;
38 double bits = weight_0 * counts[0];
39 int i;
40 for (i = 1; i < significant_symbols; ++i) {
41 bits += exp_val * (counts[i] + counts[256 - i]);
42 exp_val *= exp_decay_factor;
43 }
44 return (float)(-0.1 * bits);
45}
46
47static float PredictionCostSpatialHistogram(const int accumulated[4][256],
48 const int tile[4][256]) {
49 int i;
50 double retval = 0;
51 for (i = 0; i < 4; ++i) {
52 const double kExpValue = 0.94;
53 retval += PredictionCostSpatial(tile[i], 1, kExpValue);
54 retval += VP8LCombinedShannonEntropy(tile[i], accumulated[i]);
55 }
56 return (float)retval;
57}
58
59static WEBP_INLINE void UpdateHisto(int histo_argb[4][256], uint32_t argb) {
60 ++histo_argb[0][argb >> 24];
61 ++histo_argb[1][(argb >> 16) & 0xff];
62 ++histo_argb[2][(argb >> 8) & 0xff];
63 ++histo_argb[3][argb & 0xff];
64}
65
66//------------------------------------------------------------------------------
67// Spatial transform functions.
68
69static WEBP_INLINE void PredictBatch(int mode, int x_start, int y,
70 int num_pixels, const uint32_t* current,
71 const uint32_t* upper, uint32_t* out) {
72 if (x_start == 0) {
73 if (y == 0) {
74 // ARGB_BLACK.
75 VP8LPredictorsSub[0](current, NULL, 1, out);
76 } else {
77 // Top one.
78 VP8LPredictorsSub[2](current, upper, 1, out);
79 }
80 ++x_start;
81 ++out;
82 --num_pixels;
83 }
84 if (y == 0) {
85 // Left one.
86 VP8LPredictorsSub[1](current + x_start, NULL, num_pixels, out);
87 } else {
88 VP8LPredictorsSub[mode](current + x_start, upper + x_start, num_pixels,
89 out);
90 }
91}
92
93static int MaxDiffBetweenPixels(uint32_t p1, uint32_t p2) {
94 const int diff_a = abs((int)(p1 >> 24) - (int)(p2 >> 24));
95 const int diff_r = abs((int)((p1 >> 16) & 0xff) - (int)((p2 >> 16) & 0xff));
96 const int diff_g = abs((int)((p1 >> 8) & 0xff) - (int)((p2 >> 8) & 0xff));
97 const int diff_b = abs((int)(p1 & 0xff) - (int)(p2 & 0xff));
98 return GetMax(GetMax(diff_a, diff_r), GetMax(diff_g, diff_b));
99}
100
101static int MaxDiffAroundPixel(uint32_t current, uint32_t up, uint32_t down,
102 uint32_t left, uint32_t right) {
103 const int diff_up = MaxDiffBetweenPixels(current, up);
104 const int diff_down = MaxDiffBetweenPixels(current, down);
105 const int diff_left = MaxDiffBetweenPixels(current, left);
106 const int diff_right = MaxDiffBetweenPixels(current, right);
107 return GetMax(GetMax(diff_up, diff_down), GetMax(diff_left, diff_right));
108}
109
110static uint32_t AddGreenToBlueAndRed(uint32_t argb) {
111 const uint32_t green = (argb >> 8) & 0xff;
112 uint32_t red_blue = argb & 0x00ff00ffu;
113 red_blue += (green << 16) | green;
114 red_blue &= 0x00ff00ffu;
115 return (argb & 0xff00ff00u) | red_blue;
116}
117
118static void MaxDiffsForRow(int width, int stride, const uint32_t* const argb,
119 uint8_t* const max_diffs, int used_subtract_green) {
120 uint32_t current, up, down, left, right;
121 int x;
122 if (width <= 2) return;
123 current = argb[0];
124 right = argb[1];
125 if (used_subtract_green) {
126 current = AddGreenToBlueAndRed(current);
127 right = AddGreenToBlueAndRed(right);
128 }
129 // max_diffs[0] and max_diffs[width - 1] are never used.
130 for (x = 1; x < width - 1; ++x) {
131 up = argb[-stride + x];
132 down = argb[stride + x];
133 left = current;
134 current = right;
135 right = argb[x + 1];
136 if (used_subtract_green) {
137 up = AddGreenToBlueAndRed(up);
138 down = AddGreenToBlueAndRed(down);
139 right = AddGreenToBlueAndRed(right);
140 }
141 max_diffs[x] = MaxDiffAroundPixel(current, up, down, left, right);
142 }
143}
144
145// Quantize the difference between the actual component value and its prediction
146// to a multiple of quantization, working modulo 256, taking care not to cross
147// a boundary (inclusive upper limit).
148static uint8_t NearLosslessComponent(uint8_t value, uint8_t predict,
149 uint8_t boundary, int quantization) {
150 const int residual = (value - predict) & 0xff;
151 const int boundary_residual = (boundary - predict) & 0xff;
152 const int lower = residual & ~(quantization - 1);
153 const int upper = lower + quantization;
154 // Resolve ties towards a value closer to the prediction (i.e. towards lower
155 // if value comes after prediction and towards upper otherwise).
156 const int bias = ((boundary - value) & 0xff) < boundary_residual;
157 if (residual - lower < upper - residual + bias) {
158 // lower is closer to residual than upper.
159 if (residual > boundary_residual && lower <= boundary_residual) {
160 // Halve quantization step to avoid crossing boundary. This midpoint is
161 // on the same side of boundary as residual because midpoint >= residual
162 // (since lower is closer than upper) and residual is above the boundary.
163 return lower + (quantization >> 1);
164 }
165 return lower;
166 } else {
167 // upper is closer to residual than lower.
168 if (residual <= boundary_residual && upper > boundary_residual) {
169 // Halve quantization step to avoid crossing boundary. This midpoint is
170 // on the same side of boundary as residual because midpoint <= residual
171 // (since upper is closer than lower) and residual is below the boundary.
172 return lower + (quantization >> 1);
173 }
174 return upper & 0xff;
175 }
176}
177
178// Quantize every component of the difference between the actual pixel value and
179// its prediction to a multiple of a quantization (a power of 2, not larger than
180// max_quantization which is a power of 2, smaller than max_diff). Take care if
181// value and predict have undergone subtract green, which means that red and
182// blue are represented as offsets from green.
183static uint32_t NearLossless(uint32_t value, uint32_t predict,
184 int max_quantization, int max_diff,
185 int used_subtract_green) {
186 int quantization;
187 uint8_t new_green = 0;
188 uint8_t green_diff = 0;
189 uint8_t a, r, g, b;
190 if (max_diff <= 2) {
191 return VP8LSubPixels(value, predict);
192 }
193 quantization = max_quantization;
194 while (quantization >= max_diff) {
195 quantization >>= 1;
196 }
197 if ((value >> 24) == 0 || (value >> 24) == 0xff) {
198 // Preserve transparency of fully transparent or fully opaque pixels.
199 a = ((value >> 24) - (predict >> 24)) & 0xff;
200 } else {
201 a = NearLosslessComponent(value >> 24, predict >> 24, 0xff, quantization);
202 }
203 g = NearLosslessComponent((value >> 8) & 0xff, (predict >> 8) & 0xff, 0xff,
204 quantization);
205 if (used_subtract_green) {
206 // The green offset will be added to red and blue components during decoding
207 // to obtain the actual red and blue values.
208 new_green = ((predict >> 8) + g) & 0xff;
209 // The amount by which green has been adjusted during quantization. It is
210 // subtracted from red and blue for compensation, to avoid accumulating two
211 // quantization errors in them.
212 green_diff = (new_green - (value >> 8)) & 0xff;
213 }
214 r = NearLosslessComponent(((value >> 16) - green_diff) & 0xff,
215 (predict >> 16) & 0xff, 0xff - new_green,
216 quantization);
217 b = NearLosslessComponent((value - green_diff) & 0xff, predict & 0xff,
218 0xff - new_green, quantization);
219 return ((uint32_t)a << 24) | ((uint32_t)r << 16) | ((uint32_t)g << 8) | b;
220}
221
222// Stores the difference between the pixel and its prediction in "out".
223// In case of a lossy encoding, updates the source image to avoid propagating
224// the deviation further to pixels which depend on the current pixel for their
225// predictions.
226static WEBP_INLINE void GetResidual(
227 int width, int height, uint32_t* const upper_row,
228 uint32_t* const current_row, const uint8_t* const max_diffs, int mode,
229 int x_start, int x_end, int y, int max_quantization, int exact,
230 int used_subtract_green, uint32_t* const out) {
231 if (exact) {
232 PredictBatch(mode, x_start, y, x_end - x_start, current_row, upper_row,
233 out);
234 } else {
235 const VP8LPredictorFunc pred_func = VP8LPredictors[mode];
236 int x;
237 for (x = x_start; x < x_end; ++x) {
238 uint32_t predict;
239 uint32_t residual;
240 if (y == 0) {
241 predict = (x == 0) ? ARGB_BLACK : current_row[x - 1]; // Left.
242 } else if (x == 0) {
243 predict = upper_row[x]; // Top.
244 } else {
245 predict = pred_func(current_row[x - 1], upper_row + x);
246 }
247 if (max_quantization == 1 || mode == 0 || y == 0 || y == height - 1 ||
248 x == 0 || x == width - 1) {
249 residual = VP8LSubPixels(current_row[x], predict);
250 } else {
251 residual = NearLossless(current_row[x], predict, max_quantization,
252 max_diffs[x], used_subtract_green);
253 // Update the source image.
254 current_row[x] = VP8LAddPixels(predict, residual);
255 // x is never 0 here so we do not need to update upper_row like below.
256 }
257 if ((current_row[x] & kMaskAlpha) == 0) {
258 // If alpha is 0, cleanup RGB. We can choose the RGB values of the
259 // residual for best compression. The prediction of alpha itself can be
260 // non-zero and must be kept though. We choose RGB of the residual to be
261 // 0.
262 residual &= kMaskAlpha;
263 // Update the source image.
264 current_row[x] = predict & ~kMaskAlpha;
265 // The prediction for the rightmost pixel in a row uses the leftmost
266 // pixel
267 // in that row as its top-right context pixel. Hence if we change the
268 // leftmost pixel of current_row, the corresponding change must be
269 // applied
270 // to upper_row as well where top-right context is being read from.
271 if (x == 0 && y != 0) upper_row[width] = current_row[0];
272 }
273 out[x - x_start] = residual;
274 }
275 }
276}
277
278// Returns best predictor and updates the accumulated histogram.
279// If max_quantization > 1, assumes that near lossless processing will be
280// applied, quantizing residuals to multiples of quantization levels up to
281// max_quantization (the actual quantization level depends on smoothness near
282// the given pixel).
283static int GetBestPredictorForTile(int width, int height,
284 int tile_x, int tile_y, int bits,
285 int accumulated[4][256],
286 uint32_t* const argb_scratch,
287 const uint32_t* const argb,
288 int max_quantization,
289 int exact, int used_subtract_green,
290 const uint32_t* const modes) {
291 const int kNumPredModes = 14;
292 const int start_x = tile_x << bits;
293 const int start_y = tile_y << bits;
294 const int tile_size = 1 << bits;
295 const int max_y = GetMin(tile_size, height - start_y);
296 const int max_x = GetMin(tile_size, width - start_x);
297 // Whether there exist columns just outside the tile.
298 const int have_left = (start_x > 0);
299 const int have_right = (max_x < width - start_x);
300 // Position and size of the strip covering the tile and adjacent columns if
301 // they exist.
302 const int context_start_x = start_x - have_left;
303 const int context_width = max_x + have_left + have_right;
304 const int tiles_per_row = VP8LSubSampleSize(width, bits);
305 // Prediction modes of the left and above neighbor tiles.
306 const int left_mode = (tile_x > 0) ?
307 (modes[tile_y * tiles_per_row + tile_x - 1] >> 8) & 0xff : 0xff;
308 const int above_mode = (tile_y > 0) ?
309 (modes[(tile_y - 1) * tiles_per_row + tile_x] >> 8) & 0xff : 0xff;
310 // The width of upper_row and current_row is one pixel larger than image width
311 // to allow the top right pixel to point to the leftmost pixel of the next row
312 // when at the right edge.
313 uint32_t* upper_row = argb_scratch;
314 uint32_t* current_row = upper_row + width + 1;
315 uint8_t* const max_diffs = (uint8_t*)(current_row + width + 1);
316 float best_diff = MAX_DIFF_COST;
317 int best_mode = 0;
318 int mode;
319 int histo_stack_1[4][256];
320 int histo_stack_2[4][256];
321 // Need pointers to be able to swap arrays.
322 int (*histo_argb)[256] = histo_stack_1;
323 int (*best_histo)[256] = histo_stack_2;
324 int i, j;
325 uint32_t residuals[1 << MAX_TRANSFORM_BITS];
326 assert(bits <= MAX_TRANSFORM_BITS);
327 assert(max_x <= (1 << MAX_TRANSFORM_BITS));
328
329 for (mode = 0; mode < kNumPredModes; ++mode) {
330 float cur_diff;
331 int relative_y;
332 memset(histo_argb, 0, sizeof(histo_stack_1));
333 if (start_y > 0) {
334 // Read the row above the tile which will become the first upper_row.
335 // Include a pixel to the left if it exists; include a pixel to the right
336 // in all cases (wrapping to the leftmost pixel of the next row if it does
337 // not exist).
338 memcpy(current_row + context_start_x,
339 argb + (start_y - 1) * width + context_start_x,
340 sizeof(*argb) * (max_x + have_left + 1));
341 }
342 for (relative_y = 0; relative_y < max_y; ++relative_y) {
343 const int y = start_y + relative_y;
344 int relative_x;
345 uint32_t* tmp = upper_row;
346 upper_row = current_row;
347 current_row = tmp;
348 // Read current_row. Include a pixel to the left if it exists; include a
349 // pixel to the right in all cases except at the bottom right corner of
350 // the image (wrapping to the leftmost pixel of the next row if it does
351 // not exist in the current row).
352 memcpy(current_row + context_start_x,
353 argb + y * width + context_start_x,
354 sizeof(*argb) * (max_x + have_left + (y + 1 < height)));
355 if (max_quantization > 1 && y >= 1 && y + 1 < height) {
356 MaxDiffsForRow(context_width, width, argb + y * width + context_start_x,
357 max_diffs + context_start_x, used_subtract_green);
358 }
359
360 GetResidual(width, height, upper_row, current_row, max_diffs, mode,
361 start_x, start_x + max_x, y, max_quantization, exact,
362 used_subtract_green, residuals);
363 for (relative_x = 0; relative_x < max_x; ++relative_x) {
364 UpdateHisto(histo_argb, residuals[relative_x]);
365 }
366 }
367 cur_diff = PredictionCostSpatialHistogram(
368 (const int (*)[256])accumulated, (const int (*)[256])histo_argb);
369 // Favor keeping the areas locally similar.
370 if (mode == left_mode) cur_diff -= kSpatialPredictorBias;
371 if (mode == above_mode) cur_diff -= kSpatialPredictorBias;
372
373 if (cur_diff < best_diff) {
374 int (*tmp)[256] = histo_argb;
375 histo_argb = best_histo;
376 best_histo = tmp;
377 best_diff = cur_diff;
378 best_mode = mode;
379 }
380 }
381
382 for (i = 0; i < 4; i++) {
383 for (j = 0; j < 256; j++) {
384 accumulated[i][j] += best_histo[i][j];
385 }
386 }
387
388 return best_mode;
389}
390
391// Converts pixels of the image to residuals with respect to predictions.
392// If max_quantization > 1, applies near lossless processing, quantizing
393// residuals to multiples of quantization levels up to max_quantization
394// (the actual quantization level depends on smoothness near the given pixel).
395static void CopyImageWithPrediction(int width, int height,
396 int bits, uint32_t* const modes,
397 uint32_t* const argb_scratch,
398 uint32_t* const argb,
399 int low_effort, int max_quantization,
400 int exact, int used_subtract_green) {
401 const int tiles_per_row = VP8LSubSampleSize(width, bits);
402 // The width of upper_row and current_row is one pixel larger than image width
403 // to allow the top right pixel to point to the leftmost pixel of the next row
404 // when at the right edge.
405 uint32_t* upper_row = argb_scratch;
406 uint32_t* current_row = upper_row + width + 1;
407 uint8_t* current_max_diffs = (uint8_t*)(current_row + width + 1);
408 uint8_t* lower_max_diffs = current_max_diffs + width;
409 int y;
410
411 for (y = 0; y < height; ++y) {
412 int x;
413 uint32_t* const tmp32 = upper_row;
414 upper_row = current_row;
415 current_row = tmp32;
416 memcpy(current_row, argb + y * width,
417 sizeof(*argb) * (width + (y + 1 < height)));
418
419 if (low_effort) {
420 PredictBatch(kPredLowEffort, 0, y, width, current_row, upper_row,
421 argb + y * width);
422 } else {
423 if (max_quantization > 1) {
424 // Compute max_diffs for the lower row now, because that needs the
425 // contents of argb for the current row, which we will overwrite with
426 // residuals before proceeding with the next row.
427 uint8_t* const tmp8 = current_max_diffs;
428 current_max_diffs = lower_max_diffs;
429 lower_max_diffs = tmp8;
430 if (y + 2 < height) {
431 MaxDiffsForRow(width, width, argb + (y + 1) * width, lower_max_diffs,
432 used_subtract_green);
433 }
434 }
435 for (x = 0; x < width;) {
436 const int mode =
437 (modes[(y >> bits) * tiles_per_row + (x >> bits)] >> 8) & 0xff;
438 int x_end = x + (1 << bits);
439 if (x_end > width) x_end = width;
440 GetResidual(width, height, upper_row, current_row, current_max_diffs,
441 mode, x, x_end, y, max_quantization, exact,
442 used_subtract_green, argb + y * width + x);
443 x = x_end;
444 }
445 }
446 }
447}
448
449// Finds the best predictor for each tile, and converts the image to residuals
450// with respect to predictions. If near_lossless_quality < 100, applies
451// near lossless processing, shaving off more bits of residuals for lower
452// qualities.
453void VP8LResidualImage(int width, int height, int bits, int low_effort,
454 uint32_t* const argb, uint32_t* const argb_scratch,
455 uint32_t* const image, int near_lossless_quality,
456 int exact, int used_subtract_green) {
457 const int tiles_per_row = VP8LSubSampleSize(width, bits);
458 const int tiles_per_col = VP8LSubSampleSize(height, bits);
459 int tile_y;
460 int histo[4][256];
461 const int max_quantization = 1 << VP8LNearLosslessBits(near_lossless_quality);
462 if (low_effort) {
463 int i;
464 for (i = 0; i < tiles_per_row * tiles_per_col; ++i) {
465 image[i] = ARGB_BLACK | (kPredLowEffort << 8);
466 }
467 } else {
468 memset(histo, 0, sizeof(histo));
469 for (tile_y = 0; tile_y < tiles_per_col; ++tile_y) {
470 int tile_x;
471 for (tile_x = 0; tile_x < tiles_per_row; ++tile_x) {
472 const int pred = GetBestPredictorForTile(width, height, tile_x, tile_y,
473 bits, histo, argb_scratch, argb, max_quantization, exact,
474 used_subtract_green, image);
475 image[tile_y * tiles_per_row + tile_x] = ARGB_BLACK | (pred << 8);
476 }
477 }
478 }
479
480 CopyImageWithPrediction(width, height, bits, image, argb_scratch, argb,
481 low_effort, max_quantization, exact,
482 used_subtract_green);
483}
484
485//------------------------------------------------------------------------------
486// Color transform functions.
487
488static WEBP_INLINE void MultipliersClear(VP8LMultipliers* const m) {
489 m->green_to_red_ = 0;
490 m->green_to_blue_ = 0;
491 m->red_to_blue_ = 0;
492}
493
494static WEBP_INLINE void ColorCodeToMultipliers(uint32_t color_code,
495 VP8LMultipliers* const m) {
496 m->green_to_red_ = (color_code >> 0) & 0xff;
497 m->green_to_blue_ = (color_code >> 8) & 0xff;
498 m->red_to_blue_ = (color_code >> 16) & 0xff;
499}
500
501static WEBP_INLINE uint32_t MultipliersToColorCode(
502 const VP8LMultipliers* const m) {
503 return 0xff000000u |
504 ((uint32_t)(m->red_to_blue_) << 16) |
505 ((uint32_t)(m->green_to_blue_) << 8) |
506 m->green_to_red_;
507}
508
509static float PredictionCostCrossColor(const int accumulated[256],
510 const int counts[256]) {
511 // Favor low entropy, locally and globally.
512 // Favor small absolute values for PredictionCostSpatial
513 static const double kExpValue = 2.4;
514 return VP8LCombinedShannonEntropy(counts, accumulated) +
515 PredictionCostSpatial(counts, 3, kExpValue);
516}
517
518static float GetPredictionCostCrossColorRed(
519 const uint32_t* argb, int stride, int tile_width, int tile_height,
520 VP8LMultipliers prev_x, VP8LMultipliers prev_y, int green_to_red,
521 const int accumulated_red_histo[256]) {
522 int histo[256] = { 0 };
523 float cur_diff;
524
525 VP8LCollectColorRedTransforms(argb, stride, tile_width, tile_height,
526 green_to_red, histo);
527
528 cur_diff = PredictionCostCrossColor(accumulated_red_histo, histo);
529 if ((uint8_t)green_to_red == prev_x.green_to_red_) {
530 cur_diff -= 3; // favor keeping the areas locally similar
531 }
532 if ((uint8_t)green_to_red == prev_y.green_to_red_) {
533 cur_diff -= 3; // favor keeping the areas locally similar
534 }
535 if (green_to_red == 0) {
536 cur_diff -= 3;
537 }
538 return cur_diff;
539}
540
541static void GetBestGreenToRed(
542 const uint32_t* argb, int stride, int tile_width, int tile_height,
543 VP8LMultipliers prev_x, VP8LMultipliers prev_y, int quality,
544 const int accumulated_red_histo[256], VP8LMultipliers* const best_tx) {
545 const int kMaxIters = 4 + ((7 * quality) >> 8); // in range [4..6]
546 int green_to_red_best = 0;
547 int iter, offset;
548 float best_diff = GetPredictionCostCrossColorRed(
549 argb, stride, tile_width, tile_height, prev_x, prev_y,
550 green_to_red_best, accumulated_red_histo);
551 for (iter = 0; iter < kMaxIters; ++iter) {
552 // ColorTransformDelta is a 3.5 bit fixed point, so 32 is equal to
553 // one in color computation. Having initial delta here as 1 is sufficient
554 // to explore the range of (-2, 2).
555 const int delta = 32 >> iter;
556 // Try a negative and a positive delta from the best known value.
557 for (offset = -delta; offset <= delta; offset += 2 * delta) {
558 const int green_to_red_cur = offset + green_to_red_best;
559 const float cur_diff = GetPredictionCostCrossColorRed(
560 argb, stride, tile_width, tile_height, prev_x, prev_y,
561 green_to_red_cur, accumulated_red_histo);
562 if (cur_diff < best_diff) {
563 best_diff = cur_diff;
564 green_to_red_best = green_to_red_cur;
565 }
566 }
567 }
568 best_tx->green_to_red_ = green_to_red_best;
569}
570
571static float GetPredictionCostCrossColorBlue(
572 const uint32_t* argb, int stride, int tile_width, int tile_height,
573 VP8LMultipliers prev_x, VP8LMultipliers prev_y,
574 int green_to_blue, int red_to_blue, const int accumulated_blue_histo[256]) {
575 int histo[256] = { 0 };
576 float cur_diff;
577
578 VP8LCollectColorBlueTransforms(argb, stride, tile_width, tile_height,
579 green_to_blue, red_to_blue, histo);
580
581 cur_diff = PredictionCostCrossColor(accumulated_blue_histo, histo);
582 if ((uint8_t)green_to_blue == prev_x.green_to_blue_) {
583 cur_diff -= 3; // favor keeping the areas locally similar
584 }
585 if ((uint8_t)green_to_blue == prev_y.green_to_blue_) {
586 cur_diff -= 3; // favor keeping the areas locally similar
587 }
588 if ((uint8_t)red_to_blue == prev_x.red_to_blue_) {
589 cur_diff -= 3; // favor keeping the areas locally similar
590 }
591 if ((uint8_t)red_to_blue == prev_y.red_to_blue_) {
592 cur_diff -= 3; // favor keeping the areas locally similar
593 }
594 if (green_to_blue == 0) {
595 cur_diff -= 3;
596 }
597 if (red_to_blue == 0) {
598 cur_diff -= 3;
599 }
600 return cur_diff;
601}
602
603#define kGreenRedToBlueNumAxis 8
604#define kGreenRedToBlueMaxIters 7
605static void GetBestGreenRedToBlue(
606 const uint32_t* argb, int stride, int tile_width, int tile_height,
607 VP8LMultipliers prev_x, VP8LMultipliers prev_y, int quality,
608 const int accumulated_blue_histo[256],
609 VP8LMultipliers* const best_tx) {
610 const int8_t offset[kGreenRedToBlueNumAxis][2] =
611 {{0, -1}, {0, 1}, {-1, 0}, {1, 0}, {-1, -1}, {-1, 1}, {1, -1}, {1, 1}};
612 const int8_t delta_lut[kGreenRedToBlueMaxIters] = { 16, 16, 8, 4, 2, 2, 2 };
613 const int iters =
614 (quality < 25) ? 1 : (quality > 50) ? kGreenRedToBlueMaxIters : 4;
615 int green_to_blue_best = 0;
616 int red_to_blue_best = 0;
617 int iter;
618 // Initial value at origin:
619 float best_diff = GetPredictionCostCrossColorBlue(
620 argb, stride, tile_width, tile_height, prev_x, prev_y,
621 green_to_blue_best, red_to_blue_best, accumulated_blue_histo);
622 for (iter = 0; iter < iters; ++iter) {
623 const int delta = delta_lut[iter];
624 int axis;
625 for (axis = 0; axis < kGreenRedToBlueNumAxis; ++axis) {
626 const int green_to_blue_cur =
627 offset[axis][0] * delta + green_to_blue_best;
628 const int red_to_blue_cur = offset[axis][1] * delta + red_to_blue_best;
629 const float cur_diff = GetPredictionCostCrossColorBlue(
630 argb, stride, tile_width, tile_height, prev_x, prev_y,
631 green_to_blue_cur, red_to_blue_cur, accumulated_blue_histo);
632 if (cur_diff < best_diff) {
633 best_diff = cur_diff;
634 green_to_blue_best = green_to_blue_cur;
635 red_to_blue_best = red_to_blue_cur;
636 }
637 if (quality < 25 && iter == 4) {
638 // Only axis aligned diffs for lower quality.
639 break; // next iter.
640 }
641 }
642 if (delta == 2 && green_to_blue_best == 0 && red_to_blue_best == 0) {
643 // Further iterations would not help.
644 break; // out of iter-loop.
645 }
646 }
647 best_tx->green_to_blue_ = green_to_blue_best;
648 best_tx->red_to_blue_ = red_to_blue_best;
649}
650#undef kGreenRedToBlueMaxIters
651#undef kGreenRedToBlueNumAxis
652
653static VP8LMultipliers GetBestColorTransformForTile(
654 int tile_x, int tile_y, int bits,
655 VP8LMultipliers prev_x,
656 VP8LMultipliers prev_y,
657 int quality, int xsize, int ysize,
658 const int accumulated_red_histo[256],
659 const int accumulated_blue_histo[256],
660 const uint32_t* const argb) {
661 const int max_tile_size = 1 << bits;
662 const int tile_y_offset = tile_y * max_tile_size;
663 const int tile_x_offset = tile_x * max_tile_size;
664 const int all_x_max = GetMin(tile_x_offset + max_tile_size, xsize);
665 const int all_y_max = GetMin(tile_y_offset + max_tile_size, ysize);
666 const int tile_width = all_x_max - tile_x_offset;
667 const int tile_height = all_y_max - tile_y_offset;
668 const uint32_t* const tile_argb = argb + tile_y_offset * xsize
669 + tile_x_offset;
670 VP8LMultipliers best_tx;
671 MultipliersClear(&best_tx);
672
673 GetBestGreenToRed(tile_argb, xsize, tile_width, tile_height,
674 prev_x, prev_y, quality, accumulated_red_histo, &best_tx);
675 GetBestGreenRedToBlue(tile_argb, xsize, tile_width, tile_height,
676 prev_x, prev_y, quality, accumulated_blue_histo,
677 &best_tx);
678 return best_tx;
679}
680
681static void CopyTileWithColorTransform(int xsize, int ysize,
682 int tile_x, int tile_y,
683 int max_tile_size,
684 VP8LMultipliers color_transform,
685 uint32_t* argb) {
686 const int xscan = GetMin(max_tile_size, xsize - tile_x);
687 int yscan = GetMin(max_tile_size, ysize - tile_y);
688 argb += tile_y * xsize + tile_x;
689 while (yscan-- > 0) {
690 VP8LTransformColor(&color_transform, argb, xscan);
691 argb += xsize;
692 }
693}
694
695void VP8LColorSpaceTransform(int width, int height, int bits, int quality,
696 uint32_t* const argb, uint32_t* image) {
697 const int max_tile_size = 1 << bits;
698 const int tile_xsize = VP8LSubSampleSize(width, bits);
699 const int tile_ysize = VP8LSubSampleSize(height, bits);
700 int accumulated_red_histo[256] = { 0 };
701 int accumulated_blue_histo[256] = { 0 };
702 int tile_x, tile_y;
703 VP8LMultipliers prev_x, prev_y;
704 MultipliersClear(&prev_y);
705 MultipliersClear(&prev_x);
706 for (tile_y = 0; tile_y < tile_ysize; ++tile_y) {
707 for (tile_x = 0; tile_x < tile_xsize; ++tile_x) {
708 int y;
709 const int tile_x_offset = tile_x * max_tile_size;
710 const int tile_y_offset = tile_y * max_tile_size;
711 const int all_x_max = GetMin(tile_x_offset + max_tile_size, width);
712 const int all_y_max = GetMin(tile_y_offset + max_tile_size, height);
713 const int offset = tile_y * tile_xsize + tile_x;
714 if (tile_y != 0) {
715 ColorCodeToMultipliers(image[offset - tile_xsize], &prev_y);
716 }
717 prev_x = GetBestColorTransformForTile(tile_x, tile_y, bits,
718 prev_x, prev_y,
719 quality, width, height,
720 accumulated_red_histo,
721 accumulated_blue_histo,
722 argb);
723 image[offset] = MultipliersToColorCode(&prev_x);
724 CopyTileWithColorTransform(width, height, tile_x_offset, tile_y_offset,
725 max_tile_size, prev_x, argb);
726
727 // Gather accumulated histogram data.
728 for (y = tile_y_offset; y < all_y_max; ++y) {
729 int ix = y * width + tile_x_offset;
730 const int ix_end = ix + all_x_max - tile_x_offset;
731 for (; ix < ix_end; ++ix) {
732 const uint32_t pix = argb[ix];
733 if (ix >= 2 &&
734 pix == argb[ix - 2] &&
735 pix == argb[ix - 1]) {
736 continue; // repeated pixels are handled by backward references
737 }
738 if (ix >= width + 2 &&
739 argb[ix - 2] == argb[ix - width - 2] &&
740 argb[ix - 1] == argb[ix - width - 1] &&
741 pix == argb[ix - width]) {
742 continue; // repeated pixels are handled by backward references
743 }
744 ++accumulated_red_histo[(pix >> 16) & 0xff];
745 ++accumulated_blue_histo[(pix >> 0) & 0xff];
746 }
747 }
748 }
749 }
750}
751