/* Find the best 'out_' histogram for each of the 'in' histograms. * When called, clusters[0..num_clusters) contains the unique values from * symbols[0..in_size), but this property is not preserved in this function. * Note: we assume that out_[]->bit_cost_ is already up-to-date. */ public static void BrotliHistogramRemap(HistogramDistance *in_, size_t in_size, uint *clusters, size_t num_clusters, HistogramDistance *out_, uint *symbols) { size_t i; for (i = 0; i < in_size; ++i) { uint best_out = i == 0 ? symbols[0] : symbols[i - 1]; double best_bits = BrotliHistogramBitCostDistance(&in_[i], &out_[best_out]); size_t j; for (j = 0; j < num_clusters; ++j) { double cur_bits = BrotliHistogramBitCostDistance(&in_[i], &out_[clusters[j]]); if (cur_bits < best_bits) { best_bits = cur_bits; best_out = clusters[j]; } } symbols[i] = best_out; } /* Recompute each out_ based on raw and symbols. */ for (i = 0; i < num_clusters; ++i) { HistogramDistance.HistogramClear(&out_[clusters[i]]); } for (i = 0; i < in_size; ++i) { HistogramDistance.HistogramAddHistogram(&out_[symbols[i]], &in_[i]); } }
public static void HistogramAddVector(HistogramDistance *self, ushort *p, size_t n) { self->total_count_ += n; n += 1; while (--n != 0) { ++self->data_[*p++]; } }
public static void ClearHistograms(HistogramDistance *array, size_t length) { size_t i; for (i = 0; i < length; ++i) { HistogramClear(array + i); } }
private static unsafe void BrotliBuildHistogramsWithContext( Command *cmds, size_t num_commands, BlockSplit *literal_split, BlockSplit *insert_and_copy_split, BlockSplit *dist_split, byte *ringbuffer, size_t start_pos, size_t mask, byte prev_byte, byte prev_byte2, ContextType *context_modes, HistogramLiteral *literal_histograms, HistogramCommand *insert_and_copy_histograms, HistogramDistance *copy_dist_histograms) { size_t pos = start_pos; BlockSplitIterator literal_it; BlockSplitIterator insert_and_copy_it; BlockSplitIterator dist_it; size_t i; InitBlockSplitIterator(&literal_it, literal_split); InitBlockSplitIterator(&insert_and_copy_it, insert_and_copy_split); InitBlockSplitIterator(&dist_it, dist_split); for (i = 0; i < num_commands; ++i) { Command *cmd = &cmds[i]; size_t j; BlockSplitIteratorNext(&insert_and_copy_it); HistogramCommand.HistogramAdd(&insert_and_copy_histograms[insert_and_copy_it.type_], cmd->cmd_prefix_); for (j = cmd->insert_len_; j != 0; --j) { size_t context; BlockSplitIteratorNext(&literal_it); context = context_modes != null ? ((literal_it.type_ << BROTLI_LITERAL_CONTEXT_BITS) + Context(prev_byte, prev_byte2, context_modes[literal_it.type_])) : literal_it.type_; HistogramLiteral.HistogramAdd(&literal_histograms[context], ringbuffer[pos & mask]); prev_byte2 = prev_byte; prev_byte = ringbuffer[pos & mask]; ++pos; } pos += CommandCopyLen(cmd); if (CommandCopyLen(cmd) != 0) { prev_byte2 = ringbuffer[(pos - 2) & mask]; prev_byte = ringbuffer[(pos - 1) & mask]; if (cmd->cmd_prefix_ >= 128) { size_t context; BlockSplitIteratorNext(&dist_it); context = (dist_it.type_ << BROTLI_DISTANCE_CONTEXT_BITS) + CommandDistanceContext(cmd); HistogramDistance.HistogramAdd(©_dist_histograms[context], cmd->dist_prefix_); } } } }
public static void HistogramAddHistogram(HistogramDistance *self, HistogramDistance *v) { size_t i; self->total_count_ += v->total_count_; for (i = 0; i < DATA_SIZE; ++i) { self->data_[i] += v->data_[i]; } }
/* What is the bit cost of moving histogram from cur_symbol to candidate. */ public static double BrotliHistogramBitCostDistance( HistogramDistance *histogram, HistogramDistance *candidate) { if (histogram->total_count_ == 0) { return(0.0); } else { HistogramDistance tmp = *histogram; HistogramDistance.HistogramAddHistogram(&tmp, candidate); return(BitCostDistance.BrotliPopulationCost(&tmp) - candidate->bit_cost_); } }
/* Reorders elements of the out_[0..length) array and changes values in * symbols[0..length) array in the following way: * when called, symbols[] contains indexes into out_[], and has N unique * values (possibly N < length) * on return, symbols'[i] = f(symbols[i]) and * out_'[symbols'[i]] = out_[symbols[i]], for each 0 <= i < length, * where f is a bijection between the range of symbols[] and [0..N), and * the first occurrences of values in symbols'[i] come in consecutive * increasing order. * Returns N, the number of unique values in symbols[]. */ public static size_t BrotliHistogramReindex(ref MemoryManager m, HistogramDistance *out_, uint *symbols, size_t length) { const uint kInvalidIndex = uint.MaxValue; uint * new_index = (uint *)BrotliAllocate(ref m, length * sizeof(uint)); uint next_index; HistogramDistance *tmp; size_t i; for (i = 0; i < length; ++i) { new_index[i] = kInvalidIndex; } next_index = 0; for (i = 0; i < length; ++i) { if (new_index[symbols[i]] == kInvalidIndex) { new_index[symbols[i]] = next_index; ++next_index; } } /* TODO: by using idea of "cycle-sort" we can avoid allocation of * tmp and reduce the number of copying by the factor of 2. */ tmp = (HistogramDistance *)BrotliAllocate(ref m, next_index * sizeof(HistogramDistance)); next_index = 0; for (i = 0; i < length; ++i) { if (new_index[symbols[i]] == next_index) { tmp[next_index] = out_[symbols[i]]; ++next_index; } symbols[i] = new_index[symbols[i]]; } BrotliFree(ref m, new_index); for (i = 0; i < next_index; ++i) { out_[i] = tmp[i]; } BrotliFree(ref m, tmp); return(next_index); }
public static unsafe void BuildAndStoreEntropyCodes(ref MemoryManager m, BlockEncoder *self, HistogramDistance *histograms, size_t histograms_size, HuffmanTree *tree, size_t *storage_ix, byte *storage) { size_t alphabet_size = self->alphabet_size_; size_t table_size = histograms_size * alphabet_size; self->depths_ = (byte *)BrotliAllocate(ref m, table_size * sizeof(byte)); self->bits_ = (ushort *)BrotliAllocate(ref m, table_size * sizeof(ushort)); { size_t i; for (i = 0; i < histograms_size; ++i) { size_t ix = i * alphabet_size; BuildAndStoreHuffmanTree(&histograms[i].data_[0], alphabet_size, tree, &self->depths_[ix], &self->bits_[ix], storage_ix, storage); } } }
public static void BrotliClusterHistograms( ref MemoryManager m, HistogramDistance *in_, size_t in_size, size_t max_histograms, HistogramDistance *out_, size_t *out_size, uint *histogram_symbols) { uint * cluster_size = (uint *)BrotliAllocate(ref m, in_size * sizeof(uint)); uint * clusters = (uint *)BrotliAllocate(ref m, in_size * sizeof(uint)); size_t num_clusters = 0; size_t max_input_histograms = 64; size_t pairs_capacity = max_input_histograms * max_input_histograms / 2; /* For the first pass of clustering, we allow all pairs. */ HistogramPair *pairs = (HistogramPair *)BrotliAllocate(ref m, (pairs_capacity + 1) * sizeof(HistogramPair)); size_t i; for (i = 0; i < in_size; ++i) { cluster_size[i] = 1; } for (i = 0; i < in_size; ++i) { out_[i] = in_[i]; out_[i].bit_cost_ = BitCostDistance.BrotliPopulationCost(&in_[i]); histogram_symbols[i] = (uint)i; } for (i = 0; i < in_size; i += max_input_histograms) { size_t num_to_combine = Math.Min(in_size - i, max_input_histograms); size_t num_new_clusters; size_t j; for (j = 0; j < num_to_combine; ++j) { clusters[num_clusters + j] = (uint)(i + j); } num_new_clusters = BrotliHistogramCombine(out_, cluster_size, &histogram_symbols[i], &clusters[num_clusters], pairs, num_to_combine, num_to_combine, max_histograms, pairs_capacity); num_clusters += num_new_clusters; } { /* For the second pass, we limit the total number of histogram pairs. * After this limit is reached, we only keep searching for the best pair. */ size_t max_num_pairs = Math.Min( 64 * num_clusters, (num_clusters / 2) * num_clusters); BrotliEnsureCapacity(ref m, sizeof(HistogramPair), (void **)&pairs, &pairs_capacity, max_num_pairs + 1); /* Collapse similar histograms. */ num_clusters = BrotliHistogramCombine(out_, cluster_size, histogram_symbols, clusters, pairs, num_clusters, in_size, max_histograms, max_num_pairs); } BrotliFree(ref m, pairs); BrotliFree(ref m, cluster_size); /* Find the optimal map from original histograms to the final ones. */ BrotliHistogramRemap(in_, in_size, clusters, num_clusters, out_, histogram_symbols); BrotliFree(ref m, clusters); /* Convert the context map to a canonical form. */ *out_size = BrotliHistogramReindex(ref m, out_, histogram_symbols, in_size); }
public static size_t BrotliHistogramCombine(HistogramDistance *out_, uint *cluster_size, uint *symbols, uint *clusters, HistogramPair *pairs, size_t num_clusters, size_t symbols_size, size_t max_clusters, size_t max_num_pairs) { double cost_diff_threshold = 0.0; size_t min_cluster_size = 1; size_t num_pairs = 0; { /* We maintain a vector of histogram pairs, with the property that the pair * with the maximum bit cost reduction is the first. */ size_t idx1; for (idx1 = 0; idx1 < num_clusters; ++idx1) { size_t idx2; for (idx2 = idx1 + 1; idx2 < num_clusters; ++idx2) { BrotliCompareAndPushToQueue(out_, cluster_size, clusters[idx1], clusters[idx2], max_num_pairs, &pairs[0], &num_pairs); } } } while (num_clusters > min_cluster_size) { uint best_idx1; uint best_idx2; size_t i; if (pairs[0].cost_diff >= cost_diff_threshold) { cost_diff_threshold = 1e99; min_cluster_size = max_clusters; continue; } /* Take the best pair from the top of heap. */ best_idx1 = pairs[0].idx1; best_idx2 = pairs[0].idx2; HistogramDistance.HistogramAddHistogram(&out_[best_idx1], &out_[best_idx2]); out_[best_idx1].bit_cost_ = pairs[0].cost_combo; cluster_size[best_idx1] += cluster_size[best_idx2]; for (i = 0; i < symbols_size; ++i) { if (symbols[i] == best_idx2) { symbols[i] = best_idx1; } } for (i = 0; i < num_clusters; ++i) { if (clusters[i] == best_idx2) { memmove(&clusters[i], &clusters[i + 1], (num_clusters - i - 1) * sizeof(uint)); break; } } --num_clusters; { /* Remove pairs intersecting the just combined best pair. */ size_t copy_to_idx = 0; for (i = 0; i < num_pairs; ++i) { HistogramPair *p = &pairs[i]; if (p->idx1 == best_idx1 || p->idx2 == best_idx1 || p->idx1 == best_idx2 || p->idx2 == best_idx2) { /* Remove invalid pair from the queue. */ continue; } if (HistogramPairIsLess(&pairs[0], p)) { /* Replace the top of the queue if needed. */ HistogramPair front = pairs[0]; pairs[0] = *p; pairs[copy_to_idx] = front; } else { pairs[copy_to_idx] = *p; } ++copy_to_idx; } num_pairs = copy_to_idx; } /* Push new pairs formed with the combined histogram to the heap. */ for (i = 0; i < num_clusters; ++i) { BrotliCompareAndPushToQueue(out_, cluster_size, best_idx1, clusters[i], max_num_pairs, &pairs[0], &num_pairs); } } return(num_clusters); }
static void BrotliCompareAndPushToQueue( HistogramDistance *out_, uint *cluster_size, uint idx1, uint idx2, size_t max_num_pairs, HistogramPair *pairs, size_t *num_pairs) { bool is_good_pair = false; HistogramPair p = new HistogramPair(); if (idx1 == idx2) { return; } if (idx2 < idx1) { uint t = idx2; idx2 = idx1; idx1 = t; } p.idx1 = idx1; p.idx2 = idx2; p.cost_diff = 0.5 * ClusterCostDiff(cluster_size[idx1], cluster_size[idx2]); p.cost_diff -= out_[idx1].bit_cost_; p.cost_diff -= out_[idx2].bit_cost_; if (out_[idx1].total_count_ == 0) { p.cost_combo = out_[idx2].bit_cost_; is_good_pair = true; } else if (out_[idx2].total_count_ == 0) { p.cost_combo = out_[idx1].bit_cost_; is_good_pair = true; } else { double threshold = *num_pairs == 0 ? 1e99 : Math.Max(0.0, pairs[0].cost_diff); HistogramDistance combo = out_[idx1]; double cost_combo; HistogramDistance.HistogramAddHistogram(&combo, &out_[idx2]); cost_combo = BitCostDistance.BrotliPopulationCost(&combo); if (cost_combo < threshold - p.cost_diff) { p.cost_combo = cost_combo; is_good_pair = true; } } if (is_good_pair) { p.cost_diff += p.cost_combo; if (*num_pairs > 0 && HistogramPairIsLess(&pairs[0], &p)) { /* Replace the top of the queue if needed. */ if (*num_pairs < max_num_pairs) { pairs[*num_pairs] = pairs[0]; ++(*num_pairs); } pairs[0] = p; } else if (*num_pairs < max_num_pairs) { pairs[*num_pairs] = p; ++(*num_pairs); } } }
public static void HistogramAdd(HistogramDistance *self, size_t val) { ++self->data_[val]; ++self->total_count_; }
public static void HistogramClear(HistogramDistance *self) { memset(self->data_, 0, DATA_SIZE * sizeof(uint)); self->total_count_ = 0; self->bit_cost_ = double.MaxValue; }
public static double BrotliPopulationCost(HistogramDistance *histogram) { const double kOneSymbolHistogramCost = 12; const double kTwoSymbolHistogramCost = 20; const double kThreeSymbolHistogramCost = 28; const double kFourSymbolHistogramCost = 37; size_t data_size = HistogramDistance.HistogramDataSize(); int count = 0; size_t *s = stackalloc size_t[5]; double bits = 0.0; size_t i; if (histogram->total_count_ == 0) { return(kOneSymbolHistogramCost); } for (i = 0; i < data_size; ++i) { if (histogram->data_[i] > 0) { s[count] = i; ++count; if (count > 4) { break; } } } if (count == 1) { return(kOneSymbolHistogramCost); } if (count == 2) { return(kTwoSymbolHistogramCost + (double)histogram->total_count_); } if (count == 3) { uint histo0 = histogram->data_[s[0]]; uint histo1 = histogram->data_[s[1]]; uint histo2 = histogram->data_[s[2]]; uint histomax = Math.Max(histo0, Math.Max(histo1, histo2)); return(kThreeSymbolHistogramCost + 2 * (histo0 + histo1 + histo2) - histomax); } if (count == 4) { uint *histo = stackalloc uint[4]; uint h23; uint histomax; for (i = 0; i < 4; ++i) { histo[i] = histogram->data_[s[i]]; } /* Sort */ for (i = 0; i < 4; ++i) { size_t j; for (j = i + 1; j < 4; ++j) { if (histo[j] > histo[i]) { uint tmp = histo[j]; histo[j] = histo[i]; histo[i] = tmp; } } } h23 = histo[2] + histo[3]; histomax = Math.Max(h23, histo[0]); return(kFourSymbolHistogramCost + 3 * h23 + 2 * (histo[0] + histo[1]) - histomax); } { /* In this loop we compute the entropy of the histogram and simultaneously * build a simplified histogram of the code length codes where we use the * zero repeat code 17, but we don't use the non-zero repeat code 16. */ size_t max_depth = 1; uint * depth_histo = stackalloc uint[BROTLI_CODE_LENGTH_CODES]; memset(depth_histo, 0, BROTLI_CODE_LENGTH_CODES * sizeof(uint)); double log2total = FastLog2(histogram->total_count_); for (i = 0; i < data_size;) { if (histogram->data_[i] > 0) { /* Compute -log2(P(symbol)) = -log2(count(symbol)/total_count) = * = log2(total_count) - log2(count(symbol)) */ double log2p = log2total - FastLog2(histogram->data_[i]); /* Approximate the bit depth by round(-log2(P(symbol))) */ size_t depth = (size_t)(log2p + 0.5); bits += histogram->data_[i] * log2p; if (depth > 15) { depth = 15; } if (depth > max_depth) { max_depth = depth; } ++depth_histo[depth]; ++i; } else { /* Compute the run length of zeros and add the appropriate number of 0 * and 17 code length codes to the code length code histogram. */ uint reps = 1; size_t k; for (k = i + 1; k < data_size && histogram->data_[k] == 0; ++k) { ++reps; } i += reps; if (i == data_size) { /* Don't add any cost for the last zero run, since these are encoded * only implicitly. */ break; } if (reps < 3) { depth_histo[0] += reps; } else { reps -= 2; while (reps > 0) { ++depth_histo[BROTLI_REPEAT_ZERO_CODE_LENGTH]; /* Add the 3 extra bits for the 17 code length code. */ bits += 3; reps >>= 3; } } } } /* Add the estimated encoding cost of the code length code histogram. */ bits += (double)(18 + 2 * max_depth); /* Add the entropy of the code length code histogram. */ bits += BitsEntropy(depth_histo, BROTLI_CODE_LENGTH_CODES); } return(bits); }
/* Does either of three things: * (1) emits the current block with a new block type; * (2) emits the current block with the type of the second last block; * (3) merges the current block with the last block. */ public static unsafe void BlockSplitterFinishBlock( BlockSplitterDistance *self, bool is_final) { BlockSplit * split = self->split_; double * last_entropy = self->last_entropy_; HistogramDistance *histograms = self->histograms_; self->block_size_ = Math.Max(self->block_size_, self->min_block_size_); if (self->num_blocks_ == 0) { /* Create first block. */ split->lengths[0] = (uint)self->block_size_; split->types[0] = 0; last_entropy[0] = BitsEntropy(histograms[0].data_, self->alphabet_size_); last_entropy[1] = last_entropy[0]; ++self->num_blocks_; ++split->num_types; ++self->curr_histogram_ix_; if (self->curr_histogram_ix_ < *self->histograms_size_) { HistogramDistance.HistogramClear(&histograms[self->curr_histogram_ix_]); } self->block_size_ = 0; } else if (self->block_size_ > 0) { double entropy = BitsEntropy(histograms[self->curr_histogram_ix_].data_, self->alphabet_size_); HistogramDistance *combined_histo = stackalloc HistogramDistance[2]; double * combined_entropy = stackalloc double[2]; double * diff = stackalloc double[2]; size_t j; for (j = 0; j < 2; ++j) { size_t last_histogram_ix = j == 0 ? self->last_histogram_ix_0 : self->last_histogram_ix_1; combined_histo[j] = histograms[self->curr_histogram_ix_]; HistogramDistance.HistogramAddHistogram(&combined_histo[j], &histograms[last_histogram_ix]); combined_entropy[j] = BitsEntropy( &combined_histo[j].data_[0], self->alphabet_size_); diff[j] = combined_entropy[j] - entropy - last_entropy[j]; } if (split->num_types < BROTLI_MAX_NUMBER_OF_BLOCK_TYPES && diff[0] > self->split_threshold_ && diff[1] > self->split_threshold_) { /* Create new block. */ split->lengths[self->num_blocks_] = (uint)self->block_size_; split->types[self->num_blocks_] = (byte)split->num_types; self->last_histogram_ix_1 = self->last_histogram_ix_0; self->last_histogram_ix_0 = (byte)split->num_types; last_entropy[1] = last_entropy[0]; last_entropy[0] = entropy; ++self->num_blocks_; ++split->num_types; ++self->curr_histogram_ix_; if (self->curr_histogram_ix_ < *self->histograms_size_) { HistogramDistance.HistogramClear(&histograms[self->curr_histogram_ix_]); } self->block_size_ = 0; self->merge_last_count_ = 0; self->target_block_size_ = self->min_block_size_; } else if (diff[1] < diff[0] - 20.0) { /* Combine this block with second last block. */ split->lengths[self->num_blocks_] = (uint)self->block_size_; split->types[self->num_blocks_] = split->types[self->num_blocks_ - 2]; size_t tmp = self->last_histogram_ix_0; self->last_histogram_ix_0 = self->last_histogram_ix_1; self->last_histogram_ix_1 = tmp; histograms[self->last_histogram_ix_0] = combined_histo[1]; last_entropy[1] = last_entropy[0]; last_entropy[0] = combined_entropy[1]; ++self->num_blocks_; self->block_size_ = 0; HistogramDistance.HistogramClear(&histograms[self->curr_histogram_ix_]); self->merge_last_count_ = 0; self->target_block_size_ = self->min_block_size_; } else { /* Combine this block with last block. */ split->lengths[self->num_blocks_ - 1] += (uint)self->block_size_; histograms[self->last_histogram_ix_0] = combined_histo[0]; last_entropy[0] = combined_entropy[0]; if (split->num_types == 1) { last_entropy[1] = last_entropy[0]; } self->block_size_ = 0; HistogramDistance.HistogramClear(&histograms[self->curr_histogram_ix_]); if (++self->merge_last_count_ > 1) { self->target_block_size_ += self->min_block_size_; } } } if (is_final) { *self->histograms_size_ = split->num_types; split->num_blocks = self->num_blocks_; } }