diff options
Diffstat (limited to 'thirdparty/libwebp/src/enc/histogram_enc.c')
| -rw-r--r-- | thirdparty/libwebp/src/enc/histogram_enc.c | 550 |
1 files changed, 381 insertions, 169 deletions
diff --git a/thirdparty/libwebp/src/enc/histogram_enc.c b/thirdparty/libwebp/src/enc/histogram_enc.c index 9fdbc627a1..8ac6fa8e02 100644 --- a/thirdparty/libwebp/src/enc/histogram_enc.c +++ b/thirdparty/libwebp/src/enc/histogram_enc.c @@ -51,10 +51,12 @@ static void HistogramCopy(const VP8LHistogram* const src, VP8LHistogram* const dst) { uint32_t* const dst_literal = dst->literal_; const int dst_cache_bits = dst->palette_code_bits_; + const int literal_size = VP8LHistogramNumCodes(dst_cache_bits); const int histo_size = VP8LGetHistogramSize(dst_cache_bits); assert(src->palette_code_bits_ == dst_cache_bits); memcpy(dst, src, histo_size); dst->literal_ = dst_literal; + memcpy(dst->literal_, src->literal_, literal_size * sizeof(*dst->literal_)); } int VP8LGetHistogramSize(int cache_bits) { @@ -91,9 +93,19 @@ void VP8LHistogramCreate(VP8LHistogram* const p, VP8LHistogramStoreRefs(refs, p); } -void VP8LHistogramInit(VP8LHistogram* const p, int palette_code_bits) { +void VP8LHistogramInit(VP8LHistogram* const p, int palette_code_bits, + int init_arrays) { p->palette_code_bits_ = palette_code_bits; - HistogramClear(p); + if (init_arrays) { + HistogramClear(p); + } else { + p->trivial_symbol_ = 0; + p->bit_cost_ = 0.; + p->literal_cost_ = 0.; + p->red_cost_ = 0.; + p->blue_cost_ = 0.; + memset(p->is_used_, 0, sizeof(p->is_used_)); + } } VP8LHistogram* VP8LAllocateHistogram(int cache_bits) { @@ -104,37 +116,84 @@ VP8LHistogram* VP8LAllocateHistogram(int cache_bits) { histo = (VP8LHistogram*)memory; // literal_ won't necessary be aligned. histo->literal_ = (uint32_t*)(memory + sizeof(VP8LHistogram)); - VP8LHistogramInit(histo, cache_bits); + VP8LHistogramInit(histo, cache_bits, /*init_arrays=*/ 0); return histo; } +// Resets the pointers of the histograms to point to the bit buffer in the set. +static void HistogramSetResetPointers(VP8LHistogramSet* const set, + int cache_bits) { + int i; + const int histo_size = VP8LGetHistogramSize(cache_bits); + uint8_t* memory = (uint8_t*) (set->histograms); + memory += set->max_size * sizeof(*set->histograms); + for (i = 0; i < set->max_size; ++i) { + memory = (uint8_t*) WEBP_ALIGN(memory); + set->histograms[i] = (VP8LHistogram*) memory; + // literal_ won't necessary be aligned. + set->histograms[i]->literal_ = (uint32_t*)(memory + sizeof(VP8LHistogram)); + memory += histo_size; + } +} + +// Returns the total size of the VP8LHistogramSet. +static size_t HistogramSetTotalSize(int size, int cache_bits) { + const int histo_size = VP8LGetHistogramSize(cache_bits); + return (sizeof(VP8LHistogramSet) + size * (sizeof(VP8LHistogram*) + + histo_size + WEBP_ALIGN_CST)); +} + VP8LHistogramSet* VP8LAllocateHistogramSet(int size, int cache_bits) { int i; VP8LHistogramSet* set; - const int histo_size = VP8LGetHistogramSize(cache_bits); - const size_t total_size = - sizeof(*set) + size * (sizeof(*set->histograms) + - histo_size + WEBP_ALIGN_CST); + const size_t total_size = HistogramSetTotalSize(size, cache_bits); uint8_t* memory = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*memory)); if (memory == NULL) return NULL; set = (VP8LHistogramSet*)memory; memory += sizeof(*set); set->histograms = (VP8LHistogram**)memory; - memory += size * sizeof(*set->histograms); set->max_size = size; set->size = size; + HistogramSetResetPointers(set, cache_bits); for (i = 0; i < size; ++i) { - memory = (uint8_t*)WEBP_ALIGN(memory); - set->histograms[i] = (VP8LHistogram*)memory; - // literal_ won't necessary be aligned. - set->histograms[i]->literal_ = (uint32_t*)(memory + sizeof(VP8LHistogram)); - VP8LHistogramInit(set->histograms[i], cache_bits); - memory += histo_size; + VP8LHistogramInit(set->histograms[i], cache_bits, /*init_arrays=*/ 0); } return set; } +void VP8LHistogramSetClear(VP8LHistogramSet* const set) { + int i; + const int cache_bits = set->histograms[0]->palette_code_bits_; + const int size = set->max_size; + const size_t total_size = HistogramSetTotalSize(size, cache_bits); + uint8_t* memory = (uint8_t*)set; + + memset(memory, 0, total_size); + memory += sizeof(*set); + set->histograms = (VP8LHistogram**)memory; + set->max_size = size; + set->size = size; + HistogramSetResetPointers(set, cache_bits); + for (i = 0; i < size; ++i) { + set->histograms[i]->palette_code_bits_ = cache_bits; + } +} + +// Removes the histogram 'i' from 'set' by setting it to NULL. +static void HistogramSetRemoveHistogram(VP8LHistogramSet* const set, int i, + int* const num_used) { + assert(set->histograms[i] != NULL); + set->histograms[i] = NULL; + --*num_used; + // If we remove the last valid one, shrink until the next valid one. + if (i == set->size - 1) { + while (set->size >= 1 && set->histograms[set->size - 1] == NULL) { + --set->size; + } + } +} + // ----------------------------------------------------------------------------- void VP8LHistogramAddSinglePixOrCopy(VP8LHistogram* const histo, @@ -237,7 +296,8 @@ static double FinalHuffmanCost(const VP8LStreaks* const stats) { // Get the symbol entropy for the distribution 'population'. // Set 'trivial_sym', if there's only one symbol present in the distribution. static double PopulationCost(const uint32_t* const population, int length, - uint32_t* const trivial_sym) { + uint32_t* const trivial_sym, + uint8_t* const is_used) { VP8LBitEntropy bit_entropy; VP8LStreaks stats; VP8LGetEntropyUnrefined(population, length, &bit_entropy, &stats); @@ -245,6 +305,8 @@ static double PopulationCost(const uint32_t* const population, int length, *trivial_sym = (bit_entropy.nonzeros == 1) ? bit_entropy.nonzero_code : VP8L_NON_TRIVIAL_SYM; } + // The histogram is used if there is at least one non-zero streak. + *is_used = (stats.streaks[1][0] != 0 || stats.streaks[1][1] != 0); return BitsEntropyRefine(&bit_entropy) + FinalHuffmanCost(&stats); } @@ -253,7 +315,9 @@ static double PopulationCost(const uint32_t* const population, int length, // non-zero: both the zero-th one, or both the last one. static WEBP_INLINE double GetCombinedEntropy(const uint32_t* const X, const uint32_t* const Y, - int length, int trivial_at_end) { + int length, int is_X_used, + int is_Y_used, + int trivial_at_end) { VP8LStreaks stats; if (trivial_at_end) { // This configuration is due to palettization that transforms an indexed @@ -262,28 +326,43 @@ static WEBP_INLINE double GetCombinedEntropy(const uint32_t* const X, // Only FinalHuffmanCost needs to be evaluated. memset(&stats, 0, sizeof(stats)); // Deal with the non-zero value at index 0 or length-1. - stats.streaks[1][0] += 1; + stats.streaks[1][0] = 1; // Deal with the following/previous zero streak. - stats.counts[0] += 1; - stats.streaks[0][1] += length - 1; + stats.counts[0] = 1; + stats.streaks[0][1] = length - 1; return FinalHuffmanCost(&stats); } else { VP8LBitEntropy bit_entropy; - VP8LGetCombinedEntropyUnrefined(X, Y, length, &bit_entropy, &stats); + if (is_X_used) { + if (is_Y_used) { + VP8LGetCombinedEntropyUnrefined(X, Y, length, &bit_entropy, &stats); + } else { + VP8LGetEntropyUnrefined(X, length, &bit_entropy, &stats); + } + } else { + if (is_Y_used) { + VP8LGetEntropyUnrefined(Y, length, &bit_entropy, &stats); + } else { + memset(&stats, 0, sizeof(stats)); + stats.counts[0] = 1; + stats.streaks[0][length > 3] = length; + VP8LBitEntropyInit(&bit_entropy); + } + } return BitsEntropyRefine(&bit_entropy) + FinalHuffmanCost(&stats); } } // Estimates the Entropy + Huffman + other block overhead size cost. -double VP8LHistogramEstimateBits(const VP8LHistogram* const p) { +double VP8LHistogramEstimateBits(VP8LHistogram* const p) { return - PopulationCost( - p->literal_, VP8LHistogramNumCodes(p->palette_code_bits_), NULL) - + PopulationCost(p->red_, NUM_LITERAL_CODES, NULL) - + PopulationCost(p->blue_, NUM_LITERAL_CODES, NULL) - + PopulationCost(p->alpha_, NUM_LITERAL_CODES, NULL) - + PopulationCost(p->distance_, NUM_DISTANCE_CODES, NULL) + PopulationCost(p->literal_, VP8LHistogramNumCodes(p->palette_code_bits_), + NULL, &p->is_used_[0]) + + PopulationCost(p->red_, NUM_LITERAL_CODES, NULL, &p->is_used_[1]) + + PopulationCost(p->blue_, NUM_LITERAL_CODES, NULL, &p->is_used_[2]) + + PopulationCost(p->alpha_, NUM_LITERAL_CODES, NULL, &p->is_used_[3]) + + PopulationCost(p->distance_, NUM_DISTANCE_CODES, NULL, &p->is_used_[4]) + VP8LExtraCost(p->literal_ + NUM_LITERAL_CODES, NUM_LENGTH_CODES) + VP8LExtraCost(p->distance_, NUM_DISTANCE_CODES); } @@ -299,7 +378,8 @@ static int GetCombinedHistogramEntropy(const VP8LHistogram* const a, int trivial_at_end = 0; assert(a->palette_code_bits_ == b->palette_code_bits_); *cost += GetCombinedEntropy(a->literal_, b->literal_, - VP8LHistogramNumCodes(palette_code_bits), 0); + VP8LHistogramNumCodes(palette_code_bits), + a->is_used_[0], b->is_used_[0], 0); *cost += VP8LExtraCostCombined(a->literal_ + NUM_LITERAL_CODES, b->literal_ + NUM_LITERAL_CODES, NUM_LENGTH_CODES); @@ -319,19 +399,23 @@ static int GetCombinedHistogramEntropy(const VP8LHistogram* const a, } *cost += - GetCombinedEntropy(a->red_, b->red_, NUM_LITERAL_CODES, trivial_at_end); + GetCombinedEntropy(a->red_, b->red_, NUM_LITERAL_CODES, a->is_used_[1], + b->is_used_[1], trivial_at_end); if (*cost > cost_threshold) return 0; *cost += - GetCombinedEntropy(a->blue_, b->blue_, NUM_LITERAL_CODES, trivial_at_end); + GetCombinedEntropy(a->blue_, b->blue_, NUM_LITERAL_CODES, a->is_used_[2], + b->is_used_[2], trivial_at_end); if (*cost > cost_threshold) return 0; - *cost += GetCombinedEntropy(a->alpha_, b->alpha_, NUM_LITERAL_CODES, - trivial_at_end); + *cost += + GetCombinedEntropy(a->alpha_, b->alpha_, NUM_LITERAL_CODES, + a->is_used_[3], b->is_used_[3], trivial_at_end); if (*cost > cost_threshold) return 0; *cost += - GetCombinedEntropy(a->distance_, b->distance_, NUM_DISTANCE_CODES, 0); + GetCombinedEntropy(a->distance_, b->distance_, NUM_DISTANCE_CODES, + a->is_used_[4], b->is_used_[4], 0); *cost += VP8LExtraCostCombined(a->distance_, b->distance_, NUM_DISTANCE_CODES); if (*cost > cost_threshold) return 0; @@ -377,7 +461,9 @@ static double HistogramAddEval(const VP8LHistogram* const a, static double HistogramAddThresh(const VP8LHistogram* const a, const VP8LHistogram* const b, double cost_threshold) { - double cost = -a->bit_cost_; + double cost; + assert(a != NULL && b != NULL); + cost = -a->bit_cost_; GetCombinedHistogramEntropy(a, b, cost_threshold, &cost); return cost; } @@ -419,16 +505,19 @@ static void UpdateDominantCostRange( static void UpdateHistogramCost(VP8LHistogram* const h) { uint32_t alpha_sym, red_sym, blue_sym; const double alpha_cost = - PopulationCost(h->alpha_, NUM_LITERAL_CODES, &alpha_sym); + PopulationCost(h->alpha_, NUM_LITERAL_CODES, &alpha_sym, + &h->is_used_[3]); const double distance_cost = - PopulationCost(h->distance_, NUM_DISTANCE_CODES, NULL) + + PopulationCost(h->distance_, NUM_DISTANCE_CODES, NULL, &h->is_used_[4]) + VP8LExtraCost(h->distance_, NUM_DISTANCE_CODES); const int num_codes = VP8LHistogramNumCodes(h->palette_code_bits_); - h->literal_cost_ = PopulationCost(h->literal_, num_codes, NULL) + - VP8LExtraCost(h->literal_ + NUM_LITERAL_CODES, - NUM_LENGTH_CODES); - h->red_cost_ = PopulationCost(h->red_, NUM_LITERAL_CODES, &red_sym); - h->blue_cost_ = PopulationCost(h->blue_, NUM_LITERAL_CODES, &blue_sym); + h->literal_cost_ = + PopulationCost(h->literal_, num_codes, NULL, &h->is_used_[0]) + + VP8LExtraCost(h->literal_ + NUM_LITERAL_CODES, NUM_LENGTH_CODES); + h->red_cost_ = + PopulationCost(h->red_, NUM_LITERAL_CODES, &red_sym, &h->is_used_[1]); + h->blue_cost_ = + PopulationCost(h->blue_, NUM_LITERAL_CODES, &blue_sym, &h->is_used_[2]); h->bit_cost_ = h->literal_cost_ + h->red_cost_ + h->blue_cost_ + alpha_cost + distance_cost; if ((alpha_sym | red_sym | blue_sym) == VP8L_NON_TRIVIAL_SYM) { @@ -473,6 +562,7 @@ static void HistogramBuild( VP8LHistogram** const histograms = image_histo->histograms; VP8LRefsCursor c = VP8LRefsCursorInit(backward_refs); assert(histo_bits > 0); + VP8LHistogramSetClear(image_histo); while (VP8LRefsCursorOk(&c)) { const PixOrCopy* const v = c.cur_pos; const int ix = (y >> histo_bits) * histo_xsize + (x >> histo_bits); @@ -487,17 +577,37 @@ static void HistogramBuild( } // Copies the histograms and computes its bit_cost. -static void HistogramCopyAndAnalyze( - VP8LHistogramSet* const orig_histo, VP8LHistogramSet* const image_histo) { - int i; - const int histo_size = orig_histo->size; +static const uint16_t kInvalidHistogramSymbol = (uint16_t)(-1); +static void HistogramCopyAndAnalyze(VP8LHistogramSet* const orig_histo, + VP8LHistogramSet* const image_histo, + int* const num_used, + uint16_t* const histogram_symbols) { + int i, cluster_id; + int num_used_orig = *num_used; VP8LHistogram** const orig_histograms = orig_histo->histograms; VP8LHistogram** const histograms = image_histo->histograms; - for (i = 0; i < histo_size; ++i) { + assert(image_histo->max_size == orig_histo->max_size); + for (cluster_id = 0, i = 0; i < orig_histo->max_size; ++i) { VP8LHistogram* const histo = orig_histograms[i]; UpdateHistogramCost(histo); - // Copy histograms from orig_histo[] to image_histo[]. - HistogramCopy(histo, histograms[i]); + + // Skip the histogram if it is completely empty, which can happen for tiles + // with no information (when they are skipped because of LZ77). + if (!histo->is_used_[0] && !histo->is_used_[1] && !histo->is_used_[2] + && !histo->is_used_[3] && !histo->is_used_[4]) { + // The first histogram is always used. If an histogram is empty, we set + // its id to be the same as the previous one: this will improve + // compressibility for later LZ77. + assert(i > 0); + HistogramSetRemoveHistogram(image_histo, i, num_used); + HistogramSetRemoveHistogram(orig_histo, i, &num_used_orig); + histogram_symbols[i] = kInvalidHistogramSymbol; + } else { + // Copy histograms from orig_histo[] to image_histo[]. + HistogramCopy(histo, histograms[i]); + histogram_symbols[i] = cluster_id++; + assert(cluster_id <= image_histo->max_size); + } } } @@ -514,29 +624,33 @@ static void HistogramAnalyzeEntropyBin(VP8LHistogramSet* const image_histo, // Analyze the dominant (literal, red and blue) entropy costs. for (i = 0; i < histo_size; ++i) { + if (histograms[i] == NULL) continue; UpdateDominantCostRange(histograms[i], &cost_range); } // bin-hash histograms on three of the dominant (literal, red and blue) // symbol costs and store the resulting bin_id for each histogram. for (i = 0; i < histo_size; ++i) { + // bin_map[i] is not set to a special value as its use will later be guarded + // by another (histograms[i] == NULL). + if (histograms[i] == NULL) continue; bin_map[i] = GetHistoBinIndex(histograms[i], &cost_range, low_effort); } } -// Compact image_histo[] by merging some histograms with same bin_id together if -// it's advantageous. +// Merges some histograms with same bin_id together if it's advantageous. +// Sets the remaining histograms to NULL. static void HistogramCombineEntropyBin(VP8LHistogramSet* const image_histo, + int *num_used, + const uint16_t* const clusters, + uint16_t* const cluster_mappings, VP8LHistogram* cur_combo, const uint16_t* const bin_map, - int bin_map_size, int num_bins, + int num_bins, double combine_cost_factor, int low_effort) { VP8LHistogram** const histograms = image_histo->histograms; int idx; - // Work in-place: processed histograms are put at the beginning of - // image_histo[]. At the end, we just have to truncate the array. - int size = 0; struct { int16_t first; // position of the histogram that accumulates all // histograms with the same bin_id @@ -549,16 +663,19 @@ static void HistogramCombineEntropyBin(VP8LHistogramSet* const image_histo, bin_info[idx].num_combine_failures = 0; } - for (idx = 0; idx < bin_map_size; ++idx) { - const int bin_id = bin_map[idx]; - const int first = bin_info[bin_id].first; - assert(size <= idx); + // By default, a cluster matches itself. + for (idx = 0; idx < *num_used; ++idx) cluster_mappings[idx] = idx; + for (idx = 0; idx < image_histo->size; ++idx) { + int bin_id, first; + if (histograms[idx] == NULL) continue; + bin_id = bin_map[idx]; + first = bin_info[bin_id].first; if (first == -1) { - // just move histogram #idx to its final position - histograms[size] = histograms[idx]; - bin_info[bin_id].first = size++; + bin_info[bin_id].first = idx; } else if (low_effort) { HistogramAdd(histograms[idx], histograms[first], histograms[first]); + HistogramSetRemoveHistogram(image_histo, idx, num_used); + cluster_mappings[clusters[idx]] = clusters[first]; } else { // try to merge #idx into #first (both share the same bin_id) const double bit_cost = histograms[idx]->bit_cost_; @@ -581,19 +698,18 @@ static void HistogramCombineEntropyBin(VP8LHistogramSet* const image_histo, bin_info[bin_id].num_combine_failures >= max_combine_failures) { // move the (better) merged histogram to its final slot HistogramSwap(&cur_combo, &histograms[first]); + HistogramSetRemoveHistogram(image_histo, idx, num_used); + cluster_mappings[clusters[idx]] = clusters[first]; } else { - histograms[size++] = histograms[idx]; ++bin_info[bin_id].num_combine_failures; } - } else { - histograms[size++] = histograms[idx]; } } } - image_histo->size = size; if (low_effort) { // for low_effort case, update the final cost when everything is merged - for (idx = 0; idx < size; ++idx) { + for (idx = 0; idx < image_histo->size; ++idx) { + if (histograms[idx] == NULL) continue; UpdateHistogramCost(histograms[idx]); } } @@ -624,16 +740,9 @@ typedef struct { int max_size; } HistoQueue; -static int HistoQueueInit(HistoQueue* const histo_queue, const int max_index) { +static int HistoQueueInit(HistoQueue* const histo_queue, const int max_size) { histo_queue->size = 0; - // max_index^2 for the queue size is safe. If you look at - // HistogramCombineGreedy, and imagine that UpdateQueueFront always pushes - // data to the queue, you insert at most: - // - max_index*(max_index-1)/2 (the first two for loops) - // - max_index - 1 in the last for loop at the first iteration of the while - // loop, max_index - 2 at the second iteration ... therefore - // max_index*(max_index-1)/2 overall too - histo_queue->max_size = max_index * max_index; + histo_queue->max_size = max_size; // We allocate max_size + 1 because the last element at index "size" is // used as temporary data (and it could be up to max_size). histo_queue->queue = (HistogramPair*)WebPSafeMalloc( @@ -674,6 +783,18 @@ static void HistoQueueUpdateHead(HistoQueue* const histo_queue, } } +// Update the cost diff and combo of a pair of histograms. This needs to be +// called when the the histograms have been merged with a third one. +static void HistoQueueUpdatePair(const VP8LHistogram* const h1, + const VP8LHistogram* const h2, + double threshold, + HistogramPair* const pair) { + const double sum_cost = h1->bit_cost_ + h2->bit_cost_; + pair->cost_combo = 0.; + GetCombinedHistogramEntropy(h1, h2, sum_cost + threshold, &pair->cost_combo); + pair->cost_diff = pair->cost_combo - sum_cost; +} + // Create a pair from indices "idx1" and "idx2" provided its cost // is inferior to "threshold", a negative entropy. // It returns the cost of the pair, or 0. if it superior to threshold. @@ -683,8 +804,9 @@ static double HistoQueuePush(HistoQueue* const histo_queue, const VP8LHistogram* h1; const VP8LHistogram* h2; HistogramPair pair; - double sum_cost; + // Stop here if the queue is full. + if (histo_queue->size == histo_queue->max_size) return 0.; assert(threshold <= 0.); if (idx1 > idx2) { const int tmp = idx2; @@ -695,16 +817,12 @@ static double HistoQueuePush(HistoQueue* const histo_queue, pair.idx2 = idx2; h1 = histograms[idx1]; h2 = histograms[idx2]; - sum_cost = h1->bit_cost_ + h2->bit_cost_; - pair.cost_combo = 0.; - GetCombinedHistogramEntropy(h1, h2, sum_cost + threshold, &pair.cost_combo); - pair.cost_diff = pair.cost_combo - sum_cost; + + HistoQueueUpdatePair(h1, h2, threshold, &pair); // Do not even consider the pair if it does not improve the entropy. if (pair.cost_diff >= threshold) return 0.; - // We cannot add more elements than the capacity. - assert(histo_queue->size < histo_queue->max_size); histo_queue->queue[histo_queue->size++] = pair; HistoQueueUpdateHead(histo_queue, &histo_queue->queue[histo_queue->size - 1]); @@ -715,42 +833,43 @@ static double HistoQueuePush(HistoQueue* const histo_queue, // Combines histograms by continuously choosing the one with the highest cost // reduction. -static int HistogramCombineGreedy(VP8LHistogramSet* const image_histo) { +static int HistogramCombineGreedy(VP8LHistogramSet* const image_histo, + int* const num_used) { int ok = 0; - int image_histo_size = image_histo->size; + const int image_histo_size = image_histo->size; int i, j; VP8LHistogram** const histograms = image_histo->histograms; - // Indexes of remaining histograms. - int* const clusters = - (int*)WebPSafeMalloc(image_histo_size, sizeof(*clusters)); // Priority queue of histogram pairs. HistoQueue histo_queue; - if (!HistoQueueInit(&histo_queue, image_histo_size) || clusters == NULL) { + // image_histo_size^2 for the queue size is safe. If you look at + // HistogramCombineGreedy, and imagine that UpdateQueueFront always pushes + // data to the queue, you insert at most: + // - image_histo_size*(image_histo_size-1)/2 (the first two for loops) + // - image_histo_size - 1 in the last for loop at the first iteration of + // the while loop, image_histo_size - 2 at the second iteration ... + // therefore image_histo_size*(image_histo_size-1)/2 overall too + if (!HistoQueueInit(&histo_queue, image_histo_size * image_histo_size)) { goto End; } for (i = 0; i < image_histo_size; ++i) { - // Initialize clusters indexes. - clusters[i] = i; + if (image_histo->histograms[i] == NULL) continue; for (j = i + 1; j < image_histo_size; ++j) { - // Initialize positions array. + // Initialize queue. + if (image_histo->histograms[j] == NULL) continue; HistoQueuePush(&histo_queue, histograms, i, j, 0.); } } - while (image_histo_size > 1 && histo_queue.size > 0) { + while (histo_queue.size > 0) { const int idx1 = histo_queue.queue[0].idx1; const int idx2 = histo_queue.queue[0].idx2; HistogramAdd(histograms[idx2], histograms[idx1], histograms[idx1]); histograms[idx1]->bit_cost_ = histo_queue.queue[0].cost_combo; + // Remove merged histogram. - for (i = 0; i + 1 < image_histo_size; ++i) { - if (clusters[i] >= idx2) { - clusters[i] = clusters[i + 1]; - } - } - --image_histo_size; + HistogramSetRemoveHistogram(image_histo, idx2, num_used); // Remove pairs intersecting the just combined best pair. for (i = 0; i < histo_queue.size;) { @@ -765,24 +884,15 @@ static int HistogramCombineGreedy(VP8LHistogramSet* const image_histo) { } // Push new pairs formed with combined histogram to the queue. - for (i = 0; i < image_histo_size; ++i) { - if (clusters[i] != idx1) { - HistoQueuePush(&histo_queue, histograms, idx1, clusters[i], 0.); - } - } - } - // Move remaining histograms to the beginning of the array. - for (i = 0; i < image_histo_size; ++i) { - if (i != clusters[i]) { // swap the two histograms - HistogramSwap(&histograms[i], &histograms[clusters[i]]); + for (i = 0; i < image_histo->size; ++i) { + if (i == idx1 || image_histo->histograms[i] == NULL) continue; + HistoQueuePush(&histo_queue, image_histo->histograms, idx1, i, 0.); } } - image_histo->size = image_histo_size; ok = 1; End: - WebPSafeFree(clusters); HistoQueueClear(&histo_queue); return ok; } @@ -790,47 +900,69 @@ static int HistogramCombineGreedy(VP8LHistogramSet* const image_histo) { // Perform histogram aggregation using a stochastic approach. // 'do_greedy' is set to 1 if a greedy approach needs to be performed // afterwards, 0 otherwise. +static int PairComparison(const void* idx1, const void* idx2) { + // To be used with bsearch: <0 when *idx1<*idx2, >0 if >, 0 when ==. + return (*(int*) idx1 - *(int*) idx2); +} static int HistogramCombineStochastic(VP8LHistogramSet* const image_histo, - int min_cluster_size, + int* const num_used, int min_cluster_size, int* const do_greedy) { - int iter; + int j, iter; uint32_t seed = 1; int tries_with_no_success = 0; - int image_histo_size = image_histo->size; - const int outer_iters = image_histo_size; + const int outer_iters = *num_used; const int num_tries_no_success = outer_iters / 2; VP8LHistogram** const histograms = image_histo->histograms; - // Priority queue of histogram pairs. Its size of "kCostHeapSizeSqrt"^2 + // Priority queue of histogram pairs. Its size of 'kHistoQueueSize' // impacts the quality of the compression and the speed: the smaller the // faster but the worse for the compression. HistoQueue histo_queue; - const int kHistoQueueSizeSqrt = 3; + const int kHistoQueueSize = 9; int ok = 0; + // mapping from an index in image_histo with no NULL histogram to the full + // blown image_histo. + int* mappings; + + if (*num_used < min_cluster_size) { + *do_greedy = 1; + return 1; + } - if (!HistoQueueInit(&histo_queue, kHistoQueueSizeSqrt)) { + mappings = (int*) WebPSafeMalloc(*num_used, sizeof(*mappings)); + if (mappings == NULL || !HistoQueueInit(&histo_queue, kHistoQueueSize)) { goto End; } + // Fill the initial mapping. + for (j = 0, iter = 0; iter < image_histo->size; ++iter) { + if (histograms[iter] == NULL) continue; + mappings[j++] = iter; + } + assert(j == *num_used); + // Collapse similar histograms in 'image_histo'. - ++min_cluster_size; - for (iter = 0; iter < outer_iters && image_histo_size >= min_cluster_size && - ++tries_with_no_success < num_tries_no_success; + for (iter = 0; + iter < outer_iters && *num_used >= min_cluster_size && + ++tries_with_no_success < num_tries_no_success; ++iter) { + int* mapping_index; double best_cost = (histo_queue.size == 0) ? 0. : histo_queue.queue[0].cost_diff; int best_idx1 = -1, best_idx2 = 1; - int j; - const uint32_t rand_range = (image_histo_size - 1) * image_histo_size; - // image_histo_size / 2 was chosen empirically. Less means faster but worse + const uint32_t rand_range = (*num_used - 1) * (*num_used); + // (*num_used) / 2 was chosen empirically. Less means faster but worse // compression. - const int num_tries = image_histo_size / 2; + const int num_tries = (*num_used) / 2; - for (j = 0; j < num_tries; ++j) { + // Pick random samples. + for (j = 0; *num_used >= 2 && j < num_tries; ++j) { double curr_cost; // Choose two different histograms at random and try to combine them. const uint32_t tmp = MyRand(&seed) % rand_range; - const uint32_t idx1 = tmp / (image_histo_size - 1); - uint32_t idx2 = tmp % (image_histo_size - 1); + uint32_t idx1 = tmp / (*num_used - 1); + uint32_t idx2 = tmp % (*num_used - 1); if (idx2 >= idx1) ++idx2; + idx1 = mappings[idx1]; + idx2 = mappings[idx2]; // Calculate cost reduction on combination. curr_cost = @@ -843,18 +975,21 @@ static int HistogramCombineStochastic(VP8LHistogramSet* const image_histo, } if (histo_queue.size == 0) continue; - // Merge the two best histograms. + // Get the best histograms. best_idx1 = histo_queue.queue[0].idx1; best_idx2 = histo_queue.queue[0].idx2; assert(best_idx1 < best_idx2); - HistogramAddEval(histograms[best_idx1], histograms[best_idx2], - histograms[best_idx1], 0); - // Swap the best_idx2 histogram with the last one (which is now unused). - --image_histo_size; - if (best_idx2 != image_histo_size) { - HistogramSwap(&histograms[image_histo_size], &histograms[best_idx2]); - } - histograms[image_histo_size] = NULL; + // Pop best_idx2 from mappings. + mapping_index = (int*) bsearch(&best_idx2, mappings, *num_used, + sizeof(best_idx2), &PairComparison); + assert(mapping_index != NULL); + memmove(mapping_index, mapping_index + 1, sizeof(*mapping_index) * + ((*num_used) - (mapping_index - mappings) - 1)); + // Merge the histograms and remove best_idx2 from the queue. + HistogramAdd(histograms[best_idx2], histograms[best_idx1], + histograms[best_idx1]); + histograms[best_idx1]->bit_cost_ = histo_queue.queue[0].cost_combo; + HistogramSetRemoveHistogram(image_histo, best_idx2, num_used); // Parse the queue and update each pair that deals with best_idx1, // best_idx2 or image_histo_size. for (j = 0; j < histo_queue.size;) { @@ -877,12 +1012,6 @@ static int HistogramCombineStochastic(VP8LHistogramSet* const image_histo, p->idx2 = best_idx1; do_eval = 1; } - if (p->idx2 == image_histo_size) { - // No need to re-evaluate here as it does not involve a pair - // containing best_idx1 or best_idx2. - p->idx2 = best_idx2; - } - assert(p->idx2 < image_histo_size); // Make sure the index order is respected. if (p->idx1 > p->idx2) { const int tmp = p->idx2; @@ -891,8 +1020,7 @@ static int HistogramCombineStochastic(VP8LHistogramSet* const image_histo, } if (do_eval) { // Re-evaluate the cost of an updated pair. - GetCombinedHistogramEntropy(histograms[p->idx1], histograms[p->idx2], 0, - &p->cost_diff); + HistoQueueUpdatePair(histograms[p->idx1], histograms[p->idx2], 0., p); if (p->cost_diff >= 0.) { HistoQueuePopPair(&histo_queue, p); continue; @@ -901,15 +1029,14 @@ static int HistogramCombineStochastic(VP8LHistogramSet* const image_histo, HistoQueueUpdateHead(&histo_queue, p); ++j; } - tries_with_no_success = 0; } - image_histo->size = image_histo_size; - *do_greedy = (image_histo->size <= min_cluster_size); + *do_greedy = (*num_used <= min_cluster_size); ok = 1; End: HistoQueueClear(&histo_queue); + WebPSafeFree(mappings); return ok; } @@ -917,23 +1044,29 @@ End: // Histogram refinement // Find the best 'out' histogram for each of the 'in' histograms. +// At call-time, 'out' contains the histograms of the clusters. // Note: we assume that out[]->bit_cost_ is already up-to-date. static void HistogramRemap(const VP8LHistogramSet* const in, - const VP8LHistogramSet* const out, + VP8LHistogramSet* const out, uint16_t* const symbols) { int i; VP8LHistogram** const in_histo = in->histograms; VP8LHistogram** const out_histo = out->histograms; - const int in_size = in->size; + const int in_size = out->max_size; const int out_size = out->size; if (out_size > 1) { for (i = 0; i < in_size; ++i) { int best_out = 0; double best_bits = MAX_COST; int k; + if (in_histo[i] == NULL) { + // Arbitrarily set to the previous value if unused to help future LZ77. + symbols[i] = symbols[i - 1]; + continue; + } for (k = 0; k < out_size; ++k) { - const double cur_bits = - HistogramAddThresh(out_histo[k], in_histo[i], best_bits); + double cur_bits; + cur_bits = HistogramAddThresh(out_histo[k], in_histo[i], best_bits); if (k == 0 || cur_bits < best_bits) { best_bits = cur_bits; best_out = k; @@ -949,12 +1082,13 @@ static void HistogramRemap(const VP8LHistogramSet* const in, } // Recompute each out based on raw and symbols. - for (i = 0; i < out_size; ++i) { - HistogramClear(out_histo[i]); - } + VP8LHistogramSetClear(out); + out->size = out_size; for (i = 0; i < in_size; ++i) { - const int idx = symbols[i]; + int idx; + if (in_histo[i] == NULL) continue; + idx = symbols[i]; HistogramAdd(in_histo[i], out_histo[idx], out_histo[idx]); } } @@ -970,6 +1104,70 @@ static double GetCombineCostFactor(int histo_size, int quality) { return combine_cost_factor; } +// Given a HistogramSet 'set', the mapping of clusters 'cluster_mapping' and the +// current assignment of the cells in 'symbols', merge the clusters and +// assign the smallest possible clusters values. +static void OptimizeHistogramSymbols(const VP8LHistogramSet* const set, + uint16_t* const cluster_mappings, + int num_clusters, + uint16_t* const cluster_mappings_tmp, + uint16_t* const symbols) { + int i, cluster_max; + int do_continue = 1; + // First, assign the lowest cluster to each pixel. + while (do_continue) { + do_continue = 0; + for (i = 0; i < num_clusters; ++i) { + int k; + k = cluster_mappings[i]; + while (k != cluster_mappings[k]) { + cluster_mappings[k] = cluster_mappings[cluster_mappings[k]]; + k = cluster_mappings[k]; + } + if (k != cluster_mappings[i]) { + do_continue = 1; + cluster_mappings[i] = k; + } + } + } + // Create a mapping from a cluster id to its minimal version. + cluster_max = 0; + memset(cluster_mappings_tmp, 0, + set->max_size * sizeof(*cluster_mappings_tmp)); + assert(cluster_mappings[0] == 0); + // Re-map the ids. + for (i = 0; i < set->max_size; ++i) { + int cluster; + if (symbols[i] == kInvalidHistogramSymbol) continue; + cluster = cluster_mappings[symbols[i]]; + assert(symbols[i] < num_clusters); + if (cluster > 0 && cluster_mappings_tmp[cluster] == 0) { + ++cluster_max; + cluster_mappings_tmp[cluster] = cluster_max; + } + symbols[i] = cluster_mappings_tmp[cluster]; + } + + // Make sure all cluster values are used. + cluster_max = 0; + for (i = 0; i < set->max_size; ++i) { + if (symbols[i] == kInvalidHistogramSymbol) continue; + if (symbols[i] <= cluster_max) continue; + ++cluster_max; + assert(symbols[i] == cluster_max); + } +} + +static void RemoveEmptyHistograms(VP8LHistogramSet* const image_histo) { + uint32_t size; + int i; + for (i = 0, size = 0; i < image_histo->size; ++i) { + if (image_histo->histograms[i] == NULL) continue; + image_histo->histograms[size++] = image_histo->histograms[i]; + } + image_histo->size = size; +} + int VP8LGetHistoImageSymbols(int xsize, int ysize, const VP8LBackwardRefs* const refs, int quality, int low_effort, @@ -987,28 +1185,37 @@ int VP8LGetHistoImageSymbols(int xsize, int ysize, // histograms of small sizes (as bin_map will be very sparse) and // maximum quality q==100 (to preserve the compression gains at that level). const int entropy_combine_num_bins = low_effort ? NUM_PARTITIONS : BIN_SIZE; - const int entropy_combine = - (orig_histo->size > entropy_combine_num_bins * 2) && (quality < 100); - - if (orig_histo == NULL) goto Error; + int entropy_combine; + uint16_t* const map_tmp = + WebPSafeMalloc(2 * image_histo_raw_size, sizeof(map_tmp)); + uint16_t* const cluster_mappings = map_tmp + image_histo_raw_size; + int num_used = image_histo_raw_size; + if (orig_histo == NULL || map_tmp == NULL) goto Error; // Construct the histograms from backward references. HistogramBuild(xsize, histo_bits, refs, orig_histo); // Copies the histograms and computes its bit_cost. - HistogramCopyAndAnalyze(orig_histo, image_histo); + // histogram_symbols is optimized + HistogramCopyAndAnalyze(orig_histo, image_histo, &num_used, + histogram_symbols); + + entropy_combine = + (num_used > entropy_combine_num_bins * 2) && (quality < 100); if (entropy_combine) { - const int bin_map_size = orig_histo->size; - // Reuse histogram_symbols storage. By definition, it's guaranteed to be ok. - uint16_t* const bin_map = histogram_symbols; + uint16_t* const bin_map = map_tmp; const double combine_cost_factor = GetCombineCostFactor(image_histo_raw_size, quality); + const uint32_t num_clusters = num_used; - HistogramAnalyzeEntropyBin(orig_histo, bin_map, low_effort); + HistogramAnalyzeEntropyBin(image_histo, bin_map, low_effort); // Collapse histograms with similar entropy. - HistogramCombineEntropyBin(image_histo, tmp_histo, bin_map, bin_map_size, + HistogramCombineEntropyBin(image_histo, &num_used, histogram_symbols, + cluster_mappings, tmp_histo, bin_map, entropy_combine_num_bins, combine_cost_factor, low_effort); + OptimizeHistogramSymbols(image_histo, cluster_mappings, num_clusters, + map_tmp, histogram_symbols); } // Don't combine the histograms using stochastic and greedy heuristics for @@ -1018,21 +1225,26 @@ int VP8LGetHistoImageSymbols(int xsize, int ysize, // cubic ramp between 1 and MAX_HISTO_GREEDY: const int threshold_size = (int)(1 + (x * x * x) * (MAX_HISTO_GREEDY - 1)); int do_greedy; - if (!HistogramCombineStochastic(image_histo, threshold_size, &do_greedy)) { + if (!HistogramCombineStochastic(image_histo, &num_used, threshold_size, + &do_greedy)) { goto Error; } - if (do_greedy && !HistogramCombineGreedy(image_histo)) { - goto Error; + if (do_greedy) { + RemoveEmptyHistograms(image_histo); + if (!HistogramCombineGreedy(image_histo, &num_used)) { + goto Error; + } } } - // TODO(vrabaud): Optimize HistogramRemap for low-effort compression mode. // Find the optimal map from original histograms to the final ones. + RemoveEmptyHistograms(image_histo); HistogramRemap(orig_histo, image_histo, histogram_symbols); ok = 1; Error: VP8LFreeHistogramSet(orig_histo); + WebPSafeFree(map_tmp); return ok; } |