// SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause /* ****************************************************************** * Huffman encoder, part of New Generation Entropy library * Copyright (c) Meta Platforms, Inc. and affiliates. * * You can contact the author at : * - FSE+HUF source repository : https://github.com/Cyan4973/FiniteStateEntropy * - Public forum : https://groups.google.com/forum/#!forum/lz4c * * This source code is licensed under both the BSD-style license (found in the * LICENSE file in the root directory of this source tree) and the GPLv2 (found * in the COPYING file in the root directory of this source tree). * You may select, at your option, one of the above-listed licenses.
****************************************************************** */
/* HUF_compressWeights() : * Same as FSE_compress(), but dedicated to huff0's weights compression. * The use case needs much less stack memory. * Note : all elements within weightTable are supposed to be <= HUF_TABLELOG_MAX.
*/ #define MAX_FSE_TABLELOG_FOR_HUFF_HEADER 6
if (workspaceSize < sizeof(HUF_CompressWeightsWksp)) return ERROR(GENERIC);
/* init conditions */ if (wtSize <= 1) return 0; /* Not compressible */
/* Scan input and build symbol stats */
{ unsignedconst maxCount = HIST_count_simple(wksp->count, &maxSymbolValue, weightTable, wtSize); /* never fails */ if (maxCount == wtSize) return 1; /* only a single symbol in src : rle */ if (maxCount == 1) return 0; /* each symbol present maximum once => not compressible */
}
/* write raw values as 4-bits (max : 15) */ if (maxSymbolValue > (256-128)) return ERROR(GENERIC); /* should not happen : likely means source cannot be compressed */ if (((maxSymbolValue+1)/2) + 1 > maxDstSize) return ERROR(dstSize_tooSmall); /* not enough space within dst buffer */
op[0] = (BYTE)(128 /*special case*/ + (maxSymbolValue-1));
wksp->huffWeight[maxSymbolValue] = 0; /* to be sure it doesn't cause msan issue in final combination */ for (n=0; n<maxSymbolValue; n+=2)
op[(n/2)+1] = (BYTE)((wksp->huffWeight[n] << 4) + wksp->huffWeight[n+1]); return ((maxSymbolValue+1)/2) + 1;
}
size_t HUF_readCTable (HUF_CElt* CTable, unsigned* maxSymbolValuePtr, constvoid* src, size_t srcSize, unsigned* hasZeroWeights)
{
BYTE huffWeight[HUF_SYMBOLVALUE_MAX + 1]; /* init not required, even though some static analyzer may complain */
U32 rankVal[HUF_TABLELOG_ABSOLUTEMAX + 1]; /* large enough for values from 0 to 16 */
U32 tableLog = 0;
U32 nbSymbols = 0;
HUF_CElt* const ct = CTable + 1;
/* get symbol weights */
CHECK_V_F(readSize, HUF_readStats(huffWeight, HUF_SYMBOLVALUE_MAX+1, rankVal, &nbSymbols, &tableLog, src, srcSize));
*hasZeroWeights = (rankVal[0] > 0);
/* check result */ if (tableLog > HUF_TABLELOG_MAX) return ERROR(tableLog_tooLarge); if (nbSymbols > *maxSymbolValuePtr+1) return ERROR(maxSymbolValue_tooSmall);
/* Prepare base value per rank */
{ U32 n, nextRankStart = 0; for (n=1; n<=tableLog; n++) {
U32 curr = nextRankStart;
nextRankStart += (rankVal[n] << (n-1));
rankVal[n] = curr;
} }
/* fill nbBits */
{ U32 n; for (n=0; n<nbSymbols; n++) { const U32 w = huffWeight[n];
HUF_setNbBits(ct + n, (BYTE)(tableLog + 1 - w) & -(w != 0));
} }
/* fill val */
{ U16 nbPerRank[HUF_TABLELOG_MAX+2] = {0}; /* support w=0=>n=tableLog+1 */
U16 valPerRank[HUF_TABLELOG_MAX+2] = {0};
{ U32 n; for (n=0; n<nbSymbols; n++) nbPerRank[HUF_getNbBits(ct[n])]++; } /* determine stating value per rank */
valPerRank[tableLog+1] = 0; /* for w==0 */
{ U16 min = 0;
U32 n; for (n=tableLog; n>0; n--) { /* start at n=tablelog <-> w=1 */
valPerRank[n] = min; /* get starting value within each rank */
min += nbPerRank[n];
min >>= 1;
} } /* assign value within rank, symbol order */
{ U32 n; for (n=0; n<nbSymbols; n++) HUF_setValue(ct + n, valPerRank[HUF_getNbBits(ct[n])]++); }
}
/* * HUF_setMaxHeight(): * Try to enforce @targetNbBits on the Huffman tree described in @huffNode. * * It attempts to convert all nodes with nbBits > @targetNbBits * to employ @targetNbBits instead. Then it adjusts the tree * so that it remains a valid canonical Huffman tree. * * @pre The sum of the ranks of each symbol == 2^largestBits, * where largestBits == huffNode[lastNonNull].nbBits. * @post The sum of the ranks of each symbol == 2^largestBits, * where largestBits is the return value (expected <= targetNbBits). * * @param huffNode The Huffman tree modified in place to enforce targetNbBits. * It's presumed sorted, from most frequent to rarest symbol. * @param lastNonNull The symbol with the lowest count in the Huffman tree. * @param targetNbBits The allowed number of bits, which the Huffman tree * may not respect. After this function the Huffman tree will * respect targetNbBits. * @return The maximum number of bits of the Huffman tree after adjustment.
*/ static U32 HUF_setMaxHeight(nodeElt* huffNode, U32 lastNonNull, U32 targetNbBits)
{ const U32 largestBits = huffNode[lastNonNull].nbBits; /* early exit : no elt > targetNbBits, so the tree is already valid. */ if (largestBits <= targetNbBits) return largestBits;
/* there are several too large elements (at least >= 2) */
{ int totalCost = 0; const U32 baseCost = 1 << (largestBits - targetNbBits); int n = (int)lastNonNull;
/* Adjust any ranks > targetNbBits to targetNbBits. * Compute totalCost, which is how far the sum of the ranks is * we are over 2^largestBits after adjust the offending ranks.
*/ while (huffNode[n].nbBits > targetNbBits) {
totalCost += baseCost - (1 << (largestBits - huffNode[n].nbBits));
huffNode[n].nbBits = (BYTE)targetNbBits;
n--;
} /* n stops at huffNode[n].nbBits <= targetNbBits */
assert(huffNode[n].nbBits <= targetNbBits); /* n end at index of smallest symbol using < targetNbBits */ while (huffNode[n].nbBits == targetNbBits) --n;
/* renorm totalCost from 2^largestBits to 2^targetNbBits
* note : totalCost is necessarily a multiple of baseCost */
assert(((U32)totalCost & (baseCost - 1)) == 0);
totalCost >>= (largestBits - targetNbBits);
assert(totalCost > 0);
/* Get pos of last (smallest = lowest cum. count) symbol per rank */
ZSTD_memset(rankLast, 0xF0, sizeof(rankLast));
{ U32 currentNbBits = targetNbBits; int pos; for (pos=n ; pos >= 0; pos--) { if (huffNode[pos].nbBits >= currentNbBits) continue;
currentNbBits = huffNode[pos].nbBits; /* < targetNbBits */
rankLast[targetNbBits-currentNbBits] = (U32)pos;
} }
while (totalCost > 0) { /* Try to reduce the next power of 2 above totalCost because we * gain back half the rank.
*/
U32 nBitsToDecrease = ZSTD_highbit32((U32)totalCost) + 1; for ( ; nBitsToDecrease > 1; nBitsToDecrease--) {
U32 const highPos = rankLast[nBitsToDecrease];
U32 const lowPos = rankLast[nBitsToDecrease-1]; if (highPos == noSymbol) continue; /* Decrease highPos if no symbols of lowPos or if it is * not cheaper to remove 2 lowPos than highPos.
*/ if (lowPos == noSymbol) break;
{ U32 const highTotal = huffNode[highPos].count;
U32 const lowTotal = 2 * huffNode[lowPos].count; if (highTotal <= lowTotal) break;
} } /* only triggered when no more rank 1 symbol left => find closest one (note : there is necessarily at least one !) */
assert(rankLast[nBitsToDecrease] != noSymbol || nBitsToDecrease == 1); /* HUF_MAX_TABLELOG test just to please gcc 5+; but it should not be necessary */ while ((nBitsToDecrease<=HUF_TABLELOG_MAX) && (rankLast[nBitsToDecrease] == noSymbol))
nBitsToDecrease++;
assert(rankLast[nBitsToDecrease] != noSymbol); /* Increase the number of bits to gain back half the rank cost. */
totalCost -= 1 << (nBitsToDecrease-1);
huffNode[rankLast[nBitsToDecrease]].nbBits++;
/* Fix up the new rank. * If the new rank was empty, this symbol is now its smallest. * Otherwise, this symbol will be the largest in the new rank so no adjustment.
*/ if (rankLast[nBitsToDecrease-1] == noSymbol)
rankLast[nBitsToDecrease-1] = rankLast[nBitsToDecrease]; /* Fix up the old rank. * If the symbol was at position 0, meaning it was the highest weight symbol in the tree, * it must be the only symbol in its rank, so the old rank now has no symbols. * Otherwise, since the Huffman nodes are sorted by count, the previous position is now * the smallest node in the rank. If the previous position belongs to a different rank, * then the rank is now empty.
*/ if (rankLast[nBitsToDecrease] == 0) /* special case, reached largest symbol */
rankLast[nBitsToDecrease] = noSymbol; else {
rankLast[nBitsToDecrease]--; if (huffNode[rankLast[nBitsToDecrease]].nbBits != targetNbBits-nBitsToDecrease)
rankLast[nBitsToDecrease] = noSymbol; /* this rank is now empty */
}
} /* while (totalCost > 0) */
/* If we've removed too much weight, then we have to add it back. * To avoid overshooting again, we only adjust the smallest rank. * We take the largest nodes from the lowest rank 0 and move them * to rank 1. There's guaranteed to be enough rank 0 symbols because * TODO.
*/ while (totalCost < 0) { /* Sometimes, cost correction overshoot */ /* special case : no rank 1 symbol (using targetNbBits-1); * let's create one from largest rank 0 (using targetNbBits).
*/ if (rankLast[1] == noSymbol) { while (huffNode[n].nbBits == targetNbBits) n--;
huffNode[n+1].nbBits--;
assert(n >= 0);
rankLast[1] = (U32)(n+1);
totalCost++; continue;
}
huffNode[ rankLast[1] + 1 ].nbBits--;
rankLast[1]++;
totalCost ++;
}
} /* repay normalized cost */
} /* there are several too large elements (at least >= 2) */
/* RANK_POSITION_DISTINCT_COUNT_CUTOFF == Cutoff point in HUF_sort() buckets for which we use log2 bucketing. * Strategy is to use as many buckets as possible for representing distinct * counts while using the remainder to represent all "large" counts. * * To satisfy this requirement for 192 buckets, we can do the following: * Let buckets 0-166 represent distinct counts of [0, 166] * Let buckets 166 to 192 represent all remaining counts up to RANK_POSITION_MAX_COUNT_LOG using log2 bucketing.
*/ #define RANK_POSITION_MAX_COUNT_LOG 32 #define RANK_POSITION_LOG_BUCKETS_BEGIN ((RANK_POSITION_TABLE_SIZE - 1) - RANK_POSITION_MAX_COUNT_LOG - 1 /* == 158 */) #define RANK_POSITION_DISTINCT_COUNT_CUTOFF (RANK_POSITION_LOG_BUCKETS_BEGIN + ZSTD_highbit32(RANK_POSITION_LOG_BUCKETS_BEGIN) /* == 166 */)
/* Return the appropriate bucket index for a given count. See definition of * RANK_POSITION_DISTINCT_COUNT_CUTOFF for explanation of bucketing strategy.
*/ static U32 HUF_getIndex(U32 const count) { return (count < RANK_POSITION_DISTINCT_COUNT_CUTOFF)
? count
: ZSTD_highbit32(count) + RANK_POSITION_LOG_BUCKETS_BEGIN;
}
/* Helper swap function for HUF_quickSortPartition() */ staticvoid HUF_swapNodes(nodeElt* a, nodeElt* b) {
nodeElt tmp = *a;
*a = *b;
*b = tmp;
}
/* Returns 0 if the huffNode array is not sorted by descending count */
MEM_STATIC int HUF_isSorted(nodeElt huffNode[], U32 const maxSymbolValue1) {
U32 i; for (i = 1; i < maxSymbolValue1; ++i) { if (huffNode[i].count > huffNode[i-1].count) { return 0;
}
} return 1;
}
/* Insertion sort by descending order */
HINT_INLINE void HUF_insertionSort(nodeElt huffNode[], intconst low, intconst high) { int i; intconst size = high-low+1;
huffNode += low; for (i = 1; i < size; ++i) {
nodeElt const key = huffNode[i]; int j = i - 1; while (j >= 0 && huffNode[j].count < key.count) {
huffNode[j + 1] = huffNode[j];
j--;
}
huffNode[j + 1] = key;
}
}
/* Pivot helper function for quicksort. */ staticint HUF_quickSortPartition(nodeElt arr[], intconst low, intconst high) { /* Simply select rightmost element as pivot. "Better" selectors like * median-of-three don't experimentally appear to have any benefit.
*/
U32 const pivot = arr[high].count; int i = low - 1; int j = low; for ( ; j < high; j++) { if (arr[j].count > pivot) {
i++;
HUF_swapNodes(&arr[i], &arr[j]);
}
}
HUF_swapNodes(&arr[i + 1], &arr[high]); return i + 1;
}
/* Classic quicksort by descending with partially iterative calls * to reduce worst case callstack size.
*/ staticvoid HUF_simpleQuickSort(nodeElt arr[], int low, int high) { intconst kInsertionSortThreshold = 8; if (high - low < kInsertionSortThreshold) {
HUF_insertionSort(arr, low, high); return;
} while (low < high) { intconst idx = HUF_quickSortPartition(arr, low, high); if (idx - low < high - idx) {
HUF_simpleQuickSort(arr, low, idx - 1);
low = idx + 1;
} else {
HUF_simpleQuickSort(arr, idx + 1, high);
high = idx - 1;
}
}
}
/* * HUF_sort(): * Sorts the symbols [0, maxSymbolValue] by count[symbol] in decreasing order. * This is a typical bucket sorting strategy that uses either quicksort or insertion sort to sort each bucket. * * @param[out] huffNode Sorted symbols by decreasing count. Only members `.count` and `.byte` are filled. * Must have (maxSymbolValue + 1) entries. * @param[in] count Histogram of the symbols. * @param[in] maxSymbolValue Maximum symbol value. * @param rankPosition This is a scratch workspace. Must have RANK_POSITION_TABLE_SIZE entries.
*/ staticvoid HUF_sort(nodeElt huffNode[], constunsigned count[], U32 const maxSymbolValue, rankPos rankPosition[]) {
U32 n;
U32 const maxSymbolValue1 = maxSymbolValue+1;
/* Compute base and set curr to base. * For symbol s let lowerRank = HUF_getIndex(count[n]) and rank = lowerRank + 1. * See HUF_getIndex to see bucketing strategy. * We attribute each symbol to lowerRank's base value, because we want to know where * each rank begins in the output, so for rank R we want to count ranks R+1 and above.
*/
ZSTD_memset(rankPosition, 0, sizeof(*rankPosition) * RANK_POSITION_TABLE_SIZE); for (n = 0; n < maxSymbolValue1; ++n) {
U32 lowerRank = HUF_getIndex(count[n]);
assert(lowerRank < RANK_POSITION_TABLE_SIZE - 1);
rankPosition[lowerRank].base++;
}
assert(rankPosition[RANK_POSITION_TABLE_SIZE - 1].base == 0); /* Set up the rankPosition table */ for (n = RANK_POSITION_TABLE_SIZE - 1; n > 0; --n) {
rankPosition[n-1].base += rankPosition[n].base;
rankPosition[n-1].curr = rankPosition[n-1].base;
}
/* Insert each symbol into their appropriate bucket, setting up rankPosition table. */ for (n = 0; n < maxSymbolValue1; ++n) {
U32 const c = count[n];
U32 const r = HUF_getIndex(c) + 1;
U32 const pos = rankPosition[r].curr++;
assert(pos < maxSymbolValue1);
huffNode[pos].count = c;
huffNode[pos].byte = (BYTE)n;
}
/* HUF_buildCTable_wksp() : * Same as HUF_buildCTable(), but using externally allocated scratch buffer. * `workSpace` must be aligned on 4-bytes boundaries, and be at least as large as sizeof(HUF_buildCTable_wksp_tables).
*/ #define STARTNODE (HUF_SYMBOLVALUE_MAX+1)
/* HUF_buildTree(): * Takes the huffNode array sorted by HUF_sort() and builds an unlimited-depth Huffman tree. * * @param huffNode The array sorted by HUF_sort(). Builds the Huffman tree in this array. * @param maxSymbolValue The maximum symbol value. * @return The smallest node in the Huffman tree (by count).
*/ staticint HUF_buildTree(nodeElt* huffNode, U32 maxSymbolValue)
{
nodeElt* const huffNode0 = huffNode - 1; int nonNullRank; int lowS, lowN; int nodeNb = STARTNODE; int n, nodeRoot;
DEBUGLOG(5, "HUF_buildTree (alphabet size = %u)", maxSymbolValue + 1); /* init for parents */
nonNullRank = (int)maxSymbolValue; while(huffNode[nonNullRank].count == 0) nonNullRank--;
lowS = nonNullRank; nodeRoot = nodeNb + lowS - 1; lowN = nodeNb;
huffNode[nodeNb].count = huffNode[lowS].count + huffNode[lowS-1].count;
huffNode[lowS].parent = huffNode[lowS-1].parent = (U16)nodeNb;
nodeNb++; lowS-=2; for (n=nodeNb; n<=nodeRoot; n++) huffNode[n].count = (U32)(1U<<30);
huffNode0[0].count = (U32)(1U<<31); /* fake entry, strong barrier */
DEBUGLOG(6, "Initial distribution of bits completed (%zu sorted symbols)", showHNodeBits(huffNode, maxSymbolValue+1));
return nonNullRank;
}
/* * HUF_buildCTableFromTree(): * Build the CTable given the Huffman tree in huffNode. * * @param[out] CTable The output Huffman CTable. * @param huffNode The Huffman tree. * @param nonNullRank The last and smallest node in the Huffman tree. * @param maxSymbolValue The maximum symbol value. * @param maxNbBits The exact maximum number of bits used in the Huffman tree.
*/ staticvoid HUF_buildCTableFromTree(HUF_CElt* CTable, nodeElt const* huffNode, int nonNullRank, U32 maxSymbolValue, U32 maxNbBits)
{
HUF_CElt* const ct = CTable + 1; /* fill result into ctable (val, nbBits) */ int n;
U16 nbPerRank[HUF_TABLELOG_MAX+1] = {0};
U16 valPerRank[HUF_TABLELOG_MAX+1] = {0}; intconst alphabetSize = (int)(maxSymbolValue + 1); for (n=0; n<=nonNullRank; n++)
nbPerRank[huffNode[n].nbBits]++; /* determine starting value per rank */
{ U16 min = 0; for (n=(int)maxNbBits; n>0; n--) {
valPerRank[n] = min; /* get starting value within each rank */
min += nbPerRank[n];
min >>= 1;
} } for (n=0; n<alphabetSize; n++)
HUF_setNbBits(ct + huffNode[n].byte, huffNode[n].nbBits); /* push nbBits per symbol, symbol order */ for (n=0; n<alphabetSize; n++)
HUF_setValue(ct + n, valPerRank[HUF_getNbBits(ct[n])]++); /* assign value within rank, symbol order */
/* HUF_CStream_t: * Huffman uses its own BIT_CStream_t implementation. * There are three major differences from BIT_CStream_t: * 1. HUF_addBits() takes a HUF_CElt (size_t) which is * the pair (nbBits, value) in the format: * format: * - Bits [0, 4) = nbBits * - Bits [4, 64 - nbBits) = 0 * - Bits [64 - nbBits, 64) = value * 2. The bitContainer is built from the upper bits and * right shifted. E.g. to add a new value of N bits * you right shift the bitContainer by N, then or in * the new value into the N upper bits. * 3. The bitstream has two bit containers. You can add * bits to the second container and merge them into * the first container.
*/
/*! HUF_addBits(): * Adds the symbol stored in HUF_CElt elt to the bitstream. * * @param elt The element we're adding. This is a (nbBits, value) pair. * See the HUF_CStream_t docs for the format. * @param idx Insert into the bitstream at this idx. * @param kFast This is a template parameter. If the bitstream is guaranteed * to have at least 4 unused bits after this call it may be 1, * otherwise it must be 0. HUF_addBits() is faster when fast is set.
*/
FORCE_INLINE_TEMPLATE void HUF_addBits(HUF_CStream_t* bitC, HUF_CElt elt, int idx, int kFast)
{
assert(idx <= 1);
assert(HUF_getNbBits(elt) <= HUF_TABLELOG_ABSOLUTEMAX); /* This is efficient on x86-64 with BMI2 because shrx * only reads the low 6 bits of the register. The compiler * knows this and elides the mask. When fast is set, * every operation can use the same value loaded from elt.
*/
bitC->bitContainer[idx] >>= HUF_getNbBits(elt);
bitC->bitContainer[idx] |= kFast ? HUF_getValueFast(elt) : HUF_getValue(elt); /* We only read the low 8 bits of bitC->bitPos[idx] so it * doesn't matter that the high bits have noise from the value.
*/
bitC->bitPos[idx] += HUF_getNbBitsFast(elt);
assert((bitC->bitPos[idx] & 0xFF) <= HUF_BITS_IN_CONTAINER); /* The last 4-bits of elt are dirty if fast is set, * so we must not be overwriting bits that have already been * inserted into the bit container.
*/ #if DEBUGLEVEL >= 1
{
size_t const nbBits = HUF_getNbBits(elt);
size_t const dirtyBits = nbBits == 0 ? 0 : ZSTD_highbit32((U32)nbBits) + 1;
(void)dirtyBits; /* Middle bits are 0. */
assert(((elt >> dirtyBits) << (dirtyBits + nbBits)) == 0); /* We didn't overwrite any bits in the bit container. */
assert(!kFast || (bitC->bitPos[idx] & 0xFF) <= HUF_BITS_IN_CONTAINER);
(void)dirtyBits;
} #endif
}
/*! HUF_mergeIndex1() : * Merges the bit container @ index 1 into the bit container @ index 0 * and zeros the bit container @ index 1.
*/
FORCE_INLINE_TEMPLATE void HUF_mergeIndex1(HUF_CStream_t* bitC)
{
assert((bitC->bitPos[1] & 0xFF) < HUF_BITS_IN_CONTAINER);
bitC->bitContainer[0] >>= (bitC->bitPos[1] & 0xFF);
bitC->bitContainer[0] |= bitC->bitContainer[1];
bitC->bitPos[0] += bitC->bitPos[1];
assert((bitC->bitPos[0] & 0xFF) <= HUF_BITS_IN_CONTAINER);
}
/*! HUF_flushBits() : * Flushes the bits in the bit container @ index 0. * * @post bitPos will be < 8. * @param kFast If kFast is set then we must know a-priori that * the bit container will not overflow.
*/
FORCE_INLINE_TEMPLATE void HUF_flushBits(HUF_CStream_t* bitC, int kFast)
{ /* The upper bits of bitPos are noisy, so we must mask by 0xFF. */
size_t const nbBits = bitC->bitPos[0] & 0xFF;
size_t const nbBytes = nbBits >> 3; /* The top nbBits bits of bitContainer are the ones we need. */
size_t const bitContainer = bitC->bitContainer[0] >> (HUF_BITS_IN_CONTAINER - nbBits); /* Mask bitPos to account for the bytes we consumed. */
bitC->bitPos[0] &= 7;
assert(nbBits > 0);
assert(nbBits <= sizeof(bitC->bitContainer[0]) * 8);
assert(bitC->ptr <= bitC->endPtr);
MEM_writeLEST(bitC->ptr, bitContainer);
bitC->ptr += nbBytes;
assert(!kFast || bitC->ptr <= bitC->endPtr); if (!kFast && bitC->ptr > bitC->endPtr) bitC->ptr = bitC->endPtr; /* bitContainer doesn't need to be modified because the leftover * bits are already the top bitPos bits. And we don't care about * noise in the lower values.
*/
}
/*! HUF_endMark() * @returns The Huffman stream end mark: A 1-bit value = 1.
*/ static HUF_CElt HUF_endMark(void)
{
HUF_CElt endMark;
HUF_setNbBits(&endMark, 1);
HUF_setValue(&endMark, 1); return endMark;
}
/*! HUF_closeCStream() : * @return Size of CStream, in bytes,
* or 0 if it could not fit into dstBuffer */ static size_t HUF_closeCStream(HUF_CStream_t* bitC)
{
HUF_addBits(bitC, HUF_endMark(), /* idx */ 0, /* kFast */ 0);
HUF_flushBits(bitC, /* kFast */ 0);
{
size_t const nbBits = bitC->bitPos[0] & 0xFF; if (bitC->ptr >= bitC->endPtr) return 0; /* overflow detected */ return (size_t)(bitC->ptr - bitC->startPtr) + (nbBits > 0);
}
}
FORCE_INLINE_TEMPLATE void
HUF_compress1X_usingCTable_internal_body_loop(HUF_CStream_t* bitC, const BYTE* ip, size_t srcSize, const HUF_CElt* ct, int kUnroll, int kFastFlush, int kLastFast)
{ /* Join to kUnroll */ int n = (int)srcSize; int rem = n % kUnroll; if (rem > 0) { for (; rem > 0; --rem) {
HUF_encodeSymbol(bitC, ip[--n], ct, 0, /* fast */ 0);
}
HUF_flushBits(bitC, kFastFlush);
}
assert(n % kUnroll == 0);
/* Join to 2 * kUnroll */ if (n % (2 * kUnroll)) { int u; for (u = 1; u < kUnroll; ++u) {
HUF_encodeSymbol(bitC, ip[n - u], ct, 0, 1);
}
HUF_encodeSymbol(bitC, ip[n - kUnroll], ct, 0, kLastFast);
HUF_flushBits(bitC, kFastFlush);
n -= kUnroll;
}
assert(n % (2 * kUnroll) == 0);
for (; n>0; n-= 2 * kUnroll) { /* Encode kUnroll symbols into the bitstream @ index 0. */ int u; for (u = 1; u < kUnroll; ++u) {
HUF_encodeSymbol(bitC, ip[n - u], ct, /* idx */ 0, /* fast */ 1);
}
HUF_encodeSymbol(bitC, ip[n - kUnroll], ct, /* idx */ 0, /* fast */ kLastFast);
HUF_flushBits(bitC, kFastFlush); /* Encode kUnroll symbols into the bitstream @ index 1. * This allows us to start filling the bit container * without any data dependencies.
*/
HUF_zeroIndex1(bitC); for (u = 1; u < kUnroll; ++u) {
HUF_encodeSymbol(bitC, ip[n - kUnroll - u], ct, /* idx */ 1, /* fast */ 1);
}
HUF_encodeSymbol(bitC, ip[n - kUnroll - kUnroll], ct, /* idx */ 1, /* fast */ kLastFast); /* Merge bitstream @ index 1 into the bitstream @ index 0 */
HUF_mergeIndex1(bitC);
HUF_flushBits(bitC, kFastFlush);
}
assert(n == 0);
}
/* * Returns a tight upper bound on the output space needed by Huffman * with 8 bytes buffer to handle over-writes. If the output is at least * this large we don't need to do bounds checks during Huffman encoding.
*/ static size_t HUF_tightCompressBound(size_t srcSize, size_t tableLog)
{ return ((srcSize * tableLog) >> 3) + 8;
}
if (dstSize < 6 + 1 + 1 + 1 + 8) return 0; /* minimum space to compress successfully */ if (srcSize < 12) return 0; /* no saving possible : too small input */
op += 6; /* jumpTable */
/* checks & inits */ if (wkspSize < sizeof(*table)) return ERROR(workSpace_tooSmall); if (!srcSize) return 0; /* Uncompressed */ if (!dstSize) return 0; /* cannot fit anything within dst budget */ if (srcSize > HUF_BLOCKSIZE_MAX) return ERROR(srcSize_wrong); /* current block size limit */ if (huffLog > HUF_TABLELOG_MAX) return ERROR(tableLog_tooLarge); if (maxSymbolValue > HUF_SYMBOLVALUE_MAX) return ERROR(maxSymbolValue_tooLarge); if (!maxSymbolValue) maxSymbolValue = HUF_SYMBOLVALUE_MAX; if (!huffLog) huffLog = HUF_TABLELOG_DEFAULT;
/* Heuristic : If old table is valid, use it for small inputs */ if ((flags & HUF_flags_preferRepeat) && repeat && *repeat == HUF_repeat_valid) { return HUF_compressCTable_internal(ostart, op, oend,
src, srcSize,
nbStreams, oldHufTable, flags);
}
/* If uncompressible data is suspected, do a smaller sampling first */
DEBUG_STATIC_ASSERT(SUSPECT_INCOMPRESSIBLE_SAMPLE_RATIO >= 2); if ((flags & HUF_flags_suspectUncompressible) && srcSize >= (SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE * SUSPECT_INCOMPRESSIBLE_SAMPLE_RATIO)) {
size_t largestTotal = 0;
DEBUGLOG(5, "input suspected incompressible : sampling to check");
{ unsigned maxSymbolValueBegin = maxSymbolValue;
CHECK_V_F(largestBegin, HIST_count_simple (table->count, &maxSymbolValueBegin, (const BYTE*)src, SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE) );
largestTotal += largestBegin;
}
{ unsigned maxSymbolValueEnd = maxSymbolValue;
CHECK_V_F(largestEnd, HIST_count_simple (table->count, &maxSymbolValueEnd, (const BYTE*)src + srcSize - SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE, SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE) );
largestTotal += largestEnd;
} if (largestTotal <= ((2 * SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE) >> 7)+4) return 0; /* heuristic : probably not compressible enough */
}
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.
