ext/zstdruby/libzstd/compress/zstd_compress_superblock.c in zstd-ruby-1.5.5.1 vs ext/zstdruby/libzstd/compress/zstd_compress_superblock.c in zstd-ruby-1.5.6.0

- old
+ new

@@ -74,12 +74,12 @@ cLitSize += hufMetadata->hufDesSize; DEBUGLOG(5, "ZSTD_compressSubBlock_literal (hSize=%zu)", hufMetadata->hufDesSize); } { int const flags = bmi2 ? HUF_flags_bmi2 : 0; - const size_t cSize = singleStream ? HUF_compress1X_usingCTable(op, oend-op, literals, litSize, hufTable, flags) - : HUF_compress4X_usingCTable(op, oend-op, literals, litSize, hufTable, flags); + const size_t cSize = singleStream ? HUF_compress1X_usingCTable(op, (size_t)(oend-op), literals, litSize, hufTable, flags) + : HUF_compress4X_usingCTable(op, (size_t)(oend-op), literals, litSize, hufTable, flags); op += cSize; cLitSize += cSize; if (cSize == 0 || ERR_isError(cSize)) { DEBUGLOG(5, "Failed to write entropy tables %s", ZSTD_getErrorName(cSize)); return 0; @@ -100,11 +100,11 @@ /* Build header */ switch(lhSize) { case 3: /* 2 - 2 - 10 - 10 */ - { U32 const lhc = hType + ((!singleStream) << 2) + ((U32)litSize<<4) + ((U32)cLitSize<<14); + { U32 const lhc = hType + ((U32)(!singleStream) << 2) + ((U32)litSize<<4) + ((U32)cLitSize<<14); MEM_writeLE24(ostart, lhc); break; } case 4: /* 2 - 2 - 14 - 14 */ { U32 const lhc = hType + (2 << 2) + ((U32)litSize<<4) + ((U32)cLitSize<<18); @@ -120,34 +120,34 @@ default: /* not possible : lhSize is {3,4,5} */ assert(0); } *entropyWritten = 1; DEBUGLOG(5, "Compressed literals: %u -> %u", (U32)litSize, (U32)(op-ostart)); - return op-ostart; + return (size_t)(op-ostart); } static size_t ZSTD_seqDecompressedSize(seqStore_t const* seqStore, - const seqDef* sequences, size_t nbSeq, - size_t litSize, int lastSequence) + const seqDef* sequences, size_t nbSeqs, + size_t litSize, int lastSubBlock) { - const seqDef* const sstart = sequences; - const seqDef* const send = sequences + nbSeq; - const seqDef* sp = sstart; size_t matchLengthSum = 0; size_t litLengthSum = 0; - (void)(litLengthSum); /* suppress unused variable warning on some environments */ - while (send-sp > 0) { - ZSTD_sequenceLength const seqLen = ZSTD_getSequenceLength(seqStore, sp); + size_t n; + for (n=0; n<nbSeqs; n++) { + const ZSTD_sequenceLength seqLen = ZSTD_getSequenceLength(seqStore, sequences+n); litLengthSum += seqLen.litLength; matchLengthSum += seqLen.matchLength; - sp++; } - assert(litLengthSum <= litSize); - if (!lastSequence) { + DEBUGLOG(5, "ZSTD_seqDecompressedSize: %u sequences from %p: %u literals + %u matchlength", + (unsigned)nbSeqs, (const void*)sequences, + (unsigned)litLengthSum, (unsigned)matchLengthSum); + if (!lastSubBlock) assert(litLengthSum == litSize); - } + else + assert(litLengthSum <= litSize); + (void)litLengthSum; return matchLengthSum + litSize; } /** ZSTD_compressSubBlock_sequences() : * Compresses sequences section for a sub-block. @@ -178,18 +178,18 @@ *entropyWritten = 0; /* Sequences Header */ RETURN_ERROR_IF((oend-op) < 3 /*max nbSeq Size*/ + 1 /*seqHead*/, dstSize_tooSmall, ""); - if (nbSeq < 0x7F) + if (nbSeq < 128) *op++ = (BYTE)nbSeq; else if (nbSeq < LONGNBSEQ) op[0] = (BYTE)((nbSeq>>8) + 0x80), op[1] = (BYTE)nbSeq, op+=2; else op[0]=0xFF, MEM_writeLE16(op+1, (U16)(nbSeq - LONGNBSEQ)), op+=3; if (nbSeq==0) { - return op - ostart; + return (size_t)(op - ostart); } /* seqHead : flags for FSE encoding type */ seqHead = op++; @@ -207,11 +207,11 @@ const U32 repeat = set_repeat; *seqHead = (BYTE)((repeat<<6) + (repeat<<4) + (repeat<<2)); } { size_t const bitstreamSize = ZSTD_encodeSequences( - op, oend - op, + op, (size_t)(oend - op), fseTables->matchlengthCTable, mlCode, fseTables->offcodeCTable, ofCode, fseTables->litlengthCTable, llCode, sequences, nbSeq, longOffsets, bmi2); @@ -251,11 +251,11 @@ return 0; } #endif *entropyWritten = 1; - return op - ostart; + return (size_t)(op - ostart); } /** ZSTD_compressSubBlock() : * Compresses a single sub-block. * @return : compressed size of the sub-block @@ -277,32 +277,33 @@ BYTE* op = ostart + ZSTD_blockHeaderSize; DEBUGLOG(5, "ZSTD_compressSubBlock (litSize=%zu, nbSeq=%zu, writeLitEntropy=%d, writeSeqEntropy=%d, lastBlock=%d)", litSize, nbSeq, writeLitEntropy, writeSeqEntropy, lastBlock); { size_t cLitSize = ZSTD_compressSubBlock_literal((const HUF_CElt*)entropy->huf.CTable, &entropyMetadata->hufMetadata, literals, litSize, - op, oend-op, bmi2, writeLitEntropy, litEntropyWritten); + op, (size_t)(oend-op), + bmi2, writeLitEntropy, litEntropyWritten); FORWARD_IF_ERROR(cLitSize, "ZSTD_compressSubBlock_literal failed"); if (cLitSize == 0) return 0; op += cLitSize; } { size_t cSeqSize = ZSTD_compressSubBlock_sequences(&entropy->fse, &entropyMetadata->fseMetadata, sequences, nbSeq, llCode, mlCode, ofCode, cctxParams, - op, oend-op, + op, (size_t)(oend-op), bmi2, writeSeqEntropy, seqEntropyWritten); FORWARD_IF_ERROR(cSeqSize, "ZSTD_compressSubBlock_sequences failed"); if (cSeqSize == 0) return 0; op += cSeqSize; } /* Write block header */ - { size_t cSize = (op-ostart)-ZSTD_blockHeaderSize; + { size_t cSize = (size_t)(op-ostart) - ZSTD_blockHeaderSize; U32 const cBlockHeader24 = lastBlock + (((U32)bt_compressed)<<1) + (U32)(cSize << 3); MEM_writeLE24(ostart, cBlockHeader24); } - return op-ostart; + return (size_t)(op-ostart); } static size_t ZSTD_estimateSubBlockSize_literal(const BYTE* literals, size_t litSize, const ZSTD_hufCTables_t* huf, const ZSTD_hufCTablesMetadata_t* hufMetadata, @@ -387,27 +388,33 @@ workspace, wkspSize); if (writeEntropy) cSeqSizeEstimate += fseMetadata->fseTablesSize; return cSeqSizeEstimate + sequencesSectionHeaderSize; } -static size_t ZSTD_estimateSubBlockSize(const BYTE* literals, size_t litSize, +typedef struct { + size_t estLitSize; + size_t estBlockSize; +} EstimatedBlockSize; +static EstimatedBlockSize ZSTD_estimateSubBlockSize(const BYTE* literals, size_t litSize, const BYTE* ofCodeTable, const BYTE* llCodeTable, const BYTE* mlCodeTable, size_t nbSeq, const ZSTD_entropyCTables_t* entropy, const ZSTD_entropyCTablesMetadata_t* entropyMetadata, void* workspace, size_t wkspSize, - int writeLitEntropy, int writeSeqEntropy) { - size_t cSizeEstimate = 0; - cSizeEstimate += ZSTD_estimateSubBlockSize_literal(literals, litSize, - &entropy->huf, &entropyMetadata->hufMetadata, - workspace, wkspSize, writeLitEntropy); - cSizeEstimate += ZSTD_estimateSubBlockSize_sequences(ofCodeTable, llCodeTable, mlCodeTable, + int writeLitEntropy, int writeSeqEntropy) +{ + EstimatedBlockSize ebs; + ebs.estLitSize = ZSTD_estimateSubBlockSize_literal(literals, litSize, + &entropy->huf, &entropyMetadata->hufMetadata, + workspace, wkspSize, writeLitEntropy); + ebs.estBlockSize = ZSTD_estimateSubBlockSize_sequences(ofCodeTable, llCodeTable, mlCodeTable, nbSeq, &entropy->fse, &entropyMetadata->fseMetadata, workspace, wkspSize, writeSeqEntropy); - return cSizeEstimate + ZSTD_blockHeaderSize; + ebs.estBlockSize += ebs.estLitSize + ZSTD_blockHeaderSize; + return ebs; } static int ZSTD_needSequenceEntropyTables(ZSTD_fseCTablesMetadata_t const* fseMetadata) { if (fseMetadata->llType == set_compressed || fseMetadata->llType == set_rle) @@ -417,17 +424,60 @@ if (fseMetadata->ofType == set_compressed || fseMetadata->ofType == set_rle) return 1; return 0; } +static size_t countLiterals(seqStore_t const* seqStore, const seqDef* sp, size_t seqCount) +{ + size_t n, total = 0; + assert(sp != NULL); + for (n=0; n<seqCount; n++) { + total += ZSTD_getSequenceLength(seqStore, sp+n).litLength; + } + DEBUGLOG(6, "countLiterals for %zu sequences from %p => %zu bytes", seqCount, (const void*)sp, total); + return total; +} + +#define BYTESCALE 256 + +static size_t sizeBlockSequences(const seqDef* sp, size_t nbSeqs, + size_t targetBudget, size_t avgLitCost, size_t avgSeqCost, + int firstSubBlock) +{ + size_t n, budget = 0, inSize=0; + /* entropy headers */ + size_t const headerSize = (size_t)firstSubBlock * 120 * BYTESCALE; /* generous estimate */ + assert(firstSubBlock==0 || firstSubBlock==1); + budget += headerSize; + + /* first sequence => at least one sequence*/ + budget += sp[0].litLength * avgLitCost + avgSeqCost; + if (budget > targetBudget) return 1; + inSize = sp[0].litLength + (sp[0].mlBase+MINMATCH); + + /* loop over sequences */ + for (n=1; n<nbSeqs; n++) { + size_t currentCost = sp[n].litLength * avgLitCost + avgSeqCost; + budget += currentCost; + inSize += sp[n].litLength + (sp[n].mlBase+MINMATCH); + /* stop when sub-block budget is reached */ + if ( (budget > targetBudget) + /* though continue to expand until the sub-block is deemed compressible */ + && (budget < inSize * BYTESCALE) ) + break; + } + + return n; +} + /** ZSTD_compressSubBlock_multi() : * Breaks super-block into multiple sub-blocks and compresses them. - * Entropy will be written to the first block. - * The following blocks will use repeat mode to compress. - * All sub-blocks are compressed blocks (no raw or rle blocks). - * @return : compressed size of the super block (which is multiple ZSTD blocks) - * Or 0 if it failed to compress. */ + * Entropy will be written into the first block. + * The following blocks use repeat_mode to compress. + * Sub-blocks are all compressed, except the last one when beneficial. + * @return : compressed size of the super block (which features multiple ZSTD blocks) + * or 0 if it failed to compress. */ static size_t ZSTD_compressSubBlock_multi(const seqStore_t* seqStorePtr, const ZSTD_compressedBlockState_t* prevCBlock, ZSTD_compressedBlockState_t* nextCBlock, const ZSTD_entropyCTablesMetadata_t* entropyMetadata, const ZSTD_CCtx_params* cctxParams, @@ -436,127 +486,188 @@ const int bmi2, U32 lastBlock, void* workspace, size_t wkspSize) { const seqDef* const sstart = seqStorePtr->sequencesStart; const seqDef* const send = seqStorePtr->sequences; - const seqDef* sp = sstart; + const seqDef* sp = sstart; /* tracks progresses within seqStorePtr->sequences */ + size_t const nbSeqs = (size_t)(send - sstart); const BYTE* const lstart = seqStorePtr->litStart; const BYTE* const lend = seqStorePtr->lit; const BYTE* lp = lstart; + size_t const nbLiterals = (size_t)(lend - lstart); BYTE const* ip = (BYTE const*)src; BYTE const* const iend = ip + srcSize; BYTE* const ostart = (BYTE*)dst; BYTE* const oend = ostart + dstCapacity; BYTE* op = ostart; const BYTE* llCodePtr = seqStorePtr->llCode; const BYTE* mlCodePtr = seqStorePtr->mlCode; const BYTE* ofCodePtr = seqStorePtr->ofCode; - size_t targetCBlockSize = cctxParams->targetCBlockSize; - size_t litSize, seqCount; - int writeLitEntropy = entropyMetadata->hufMetadata.hType == set_compressed; + size_t const minTarget = ZSTD_TARGETCBLOCKSIZE_MIN; /* enforce minimum size, to reduce undesirable side effects */ + size_t const targetCBlockSize = MAX(minTarget, cctxParams->targetCBlockSize); + int writeLitEntropy = (entropyMetadata->hufMetadata.hType == set_compressed); int writeSeqEntropy = 1; - int lastSequence = 0; - DEBUGLOG(5, "ZSTD_compressSubBlock_multi (litSize=%u, nbSeq=%u)", - (unsigned)(lend-lp), (unsigned)(send-sstart)); + DEBUGLOG(5, "ZSTD_compressSubBlock_multi (srcSize=%u, litSize=%u, nbSeq=%u)", + (unsigned)srcSize, (unsigned)(lend-lstart), (unsigned)(send-sstart)); - litSize = 0; - seqCount = 0; - do { - size_t cBlockSizeEstimate = 0; - if (sstart == send) { - lastSequence = 1; - } else { - const seqDef* const sequence = sp + seqCount; - lastSequence = sequence == send - 1; - litSize += ZSTD_getSequenceLength(seqStorePtr, sequence).litLength; - seqCount++; + /* let's start by a general estimation for the full block */ + if (nbSeqs > 0) { + EstimatedBlockSize const ebs = + ZSTD_estimateSubBlockSize(lp, nbLiterals, + ofCodePtr, llCodePtr, mlCodePtr, nbSeqs, + &nextCBlock->entropy, entropyMetadata, + workspace, wkspSize, + writeLitEntropy, writeSeqEntropy); + /* quick estimation */ + size_t const avgLitCost = nbLiterals ? (ebs.estLitSize * BYTESCALE) / nbLiterals : BYTESCALE; + size_t const avgSeqCost = ((ebs.estBlockSize - ebs.estLitSize) * BYTESCALE) / nbSeqs; + const size_t nbSubBlocks = MAX((ebs.estBlockSize + (targetCBlockSize/2)) / targetCBlockSize, 1); + size_t n, avgBlockBudget, blockBudgetSupp=0; + avgBlockBudget = (ebs.estBlockSize * BYTESCALE) / nbSubBlocks; + DEBUGLOG(5, "estimated fullblock size=%u bytes ; avgLitCost=%.2f ; avgSeqCost=%.2f ; targetCBlockSize=%u, nbSubBlocks=%u ; avgBlockBudget=%.0f bytes", + (unsigned)ebs.estBlockSize, (double)avgLitCost/BYTESCALE, (double)avgSeqCost/BYTESCALE, + (unsigned)targetCBlockSize, (unsigned)nbSubBlocks, (double)avgBlockBudget/BYTESCALE); + /* simplification: if estimates states that the full superblock doesn't compress, just bail out immediately + * this will result in the production of a single uncompressed block covering @srcSize.*/ + if (ebs.estBlockSize > srcSize) return 0; + + /* compress and write sub-blocks */ + assert(nbSubBlocks>0); + for (n=0; n < nbSubBlocks-1; n++) { + /* determine nb of sequences for current sub-block + nbLiterals from next sequence */ + size_t const seqCount = sizeBlockSequences(sp, (size_t)(send-sp), + avgBlockBudget + blockBudgetSupp, avgLitCost, avgSeqCost, n==0); + /* if reached last sequence : break to last sub-block (simplification) */ + assert(seqCount <= (size_t)(send-sp)); + if (sp + seqCount == send) break; + assert(seqCount > 0); + /* compress sub-block */ + { int litEntropyWritten = 0; + int seqEntropyWritten = 0; + size_t litSize = countLiterals(seqStorePtr, sp, seqCount); + const size_t decompressedSize = + ZSTD_seqDecompressedSize(seqStorePtr, sp, seqCount, litSize, 0); + size_t const cSize = ZSTD_compressSubBlock(&nextCBlock->entropy, entropyMetadata, + sp, seqCount, + lp, litSize, + llCodePtr, mlCodePtr, ofCodePtr, + cctxParams, + op, (size_t)(oend-op), + bmi2, writeLitEntropy, writeSeqEntropy, + &litEntropyWritten, &seqEntropyWritten, + 0); + FORWARD_IF_ERROR(cSize, "ZSTD_compressSubBlock failed"); + + /* check compressibility, update state components */ + if (cSize > 0 && cSize < decompressedSize) { + DEBUGLOG(5, "Committed sub-block compressing %u bytes => %u bytes", + (unsigned)decompressedSize, (unsigned)cSize); + assert(ip + decompressedSize <= iend); + ip += decompressedSize; + lp += litSize; + op += cSize; + llCodePtr += seqCount; + mlCodePtr += seqCount; + ofCodePtr += seqCount; + /* Entropy only needs to be written once */ + if (litEntropyWritten) { + writeLitEntropy = 0; + } + if (seqEntropyWritten) { + writeSeqEntropy = 0; + } + sp += seqCount; + blockBudgetSupp = 0; + } } + /* otherwise : do not compress yet, coalesce current sub-block with following one */ } - if (lastSequence) { - assert(lp <= lend); - assert(litSize <= (size_t)(lend - lp)); - litSize = (size_t)(lend - lp); - } - /* I think there is an optimization opportunity here. - * Calling ZSTD_estimateSubBlockSize for every sequence can be wasteful - * since it recalculates estimate from scratch. - * For example, it would recount literal distribution and symbol codes every time. - */ - cBlockSizeEstimate = ZSTD_estimateSubBlockSize(lp, litSize, ofCodePtr, llCodePtr, mlCodePtr, seqCount, - &nextCBlock->entropy, entropyMetadata, - workspace, wkspSize, writeLitEntropy, writeSeqEntropy); - if (cBlockSizeEstimate > targetCBlockSize || lastSequence) { - int litEntropyWritten = 0; - int seqEntropyWritten = 0; - const size_t decompressedSize = ZSTD_seqDecompressedSize(seqStorePtr, sp, seqCount, litSize, lastSequence); - const size_t cSize = ZSTD_compressSubBlock(&nextCBlock->entropy, entropyMetadata, - sp, seqCount, - lp, litSize, - llCodePtr, mlCodePtr, ofCodePtr, - cctxParams, - op, oend-op, - bmi2, writeLitEntropy, writeSeqEntropy, - &litEntropyWritten, &seqEntropyWritten, - lastBlock && lastSequence); - FORWARD_IF_ERROR(cSize, "ZSTD_compressSubBlock failed"); - if (cSize > 0 && cSize < decompressedSize) { - DEBUGLOG(5, "Committed the sub-block"); - assert(ip + decompressedSize <= iend); - ip += decompressedSize; - sp += seqCount; - lp += litSize; - op += cSize; - llCodePtr += seqCount; - mlCodePtr += seqCount; - ofCodePtr += seqCount; - litSize = 0; - seqCount = 0; - /* Entropy only needs to be written once */ - if (litEntropyWritten) { - writeLitEntropy = 0; - } - if (seqEntropyWritten) { - writeSeqEntropy = 0; - } + } /* if (nbSeqs > 0) */ + + /* write last block */ + DEBUGLOG(5, "Generate last sub-block: %u sequences remaining", (unsigned)(send - sp)); + { int litEntropyWritten = 0; + int seqEntropyWritten = 0; + size_t litSize = (size_t)(lend - lp); + size_t seqCount = (size_t)(send - sp); + const size_t decompressedSize = + ZSTD_seqDecompressedSize(seqStorePtr, sp, seqCount, litSize, 1); + size_t const cSize = ZSTD_compressSubBlock(&nextCBlock->entropy, entropyMetadata, + sp, seqCount, + lp, litSize, + llCodePtr, mlCodePtr, ofCodePtr, + cctxParams, + op, (size_t)(oend-op), + bmi2, writeLitEntropy, writeSeqEntropy, + &litEntropyWritten, &seqEntropyWritten, + lastBlock); + FORWARD_IF_ERROR(cSize, "ZSTD_compressSubBlock failed"); + + /* update pointers, the nb of literals borrowed from next sequence must be preserved */ + if (cSize > 0 && cSize < decompressedSize) { + DEBUGLOG(5, "Last sub-block compressed %u bytes => %u bytes", + (unsigned)decompressedSize, (unsigned)cSize); + assert(ip + decompressedSize <= iend); + ip += decompressedSize; + lp += litSize; + op += cSize; + llCodePtr += seqCount; + mlCodePtr += seqCount; + ofCodePtr += seqCount; + /* Entropy only needs to be written once */ + if (litEntropyWritten) { + writeLitEntropy = 0; } + if (seqEntropyWritten) { + writeSeqEntropy = 0; + } + sp += seqCount; } - } while (!lastSequence); + } + + if (writeLitEntropy) { - DEBUGLOG(5, "ZSTD_compressSubBlock_multi has literal entropy tables unwritten"); + DEBUGLOG(5, "Literal entropy tables were never written"); ZSTD_memcpy(&nextCBlock->entropy.huf, &prevCBlock->entropy.huf, sizeof(prevCBlock->entropy.huf)); } if (writeSeqEntropy && ZSTD_needSequenceEntropyTables(&entropyMetadata->fseMetadata)) { /* If we haven't written our entropy tables, then we've violated our contract and * must emit an uncompressed block. */ - DEBUGLOG(5, "ZSTD_compressSubBlock_multi has sequence entropy tables unwritten"); + DEBUGLOG(5, "Sequence entropy tables were never written => cancel, emit an uncompressed block"); return 0; } + if (ip < iend) { - size_t const cSize = ZSTD_noCompressBlock(op, oend - op, ip, iend - ip, lastBlock); - DEBUGLOG(5, "ZSTD_compressSubBlock_multi last sub-block uncompressed, %zu bytes", (size_t)(iend - ip)); + /* some data left : last part of the block sent uncompressed */ + size_t const rSize = (size_t)((iend - ip)); + size_t const cSize = ZSTD_noCompressBlock(op, (size_t)(oend - op), ip, rSize, lastBlock); + DEBUGLOG(5, "Generate last uncompressed sub-block of %u bytes", (unsigned)(rSize)); FORWARD_IF_ERROR(cSize, "ZSTD_noCompressBlock failed"); assert(cSize != 0); op += cSize; /* We have to regenerate the repcodes because we've skipped some sequences */ if (sp < send) { - seqDef const* seq; + const seqDef* seq; repcodes_t rep; ZSTD_memcpy(&rep, prevCBlock->rep, sizeof(rep)); for (seq = sstart; seq < sp; ++seq) { ZSTD_updateRep(rep.rep, seq->offBase, ZSTD_getSequenceLength(seqStorePtr, seq).litLength == 0); } ZSTD_memcpy(nextCBlock->rep, &rep, sizeof(rep)); } } - DEBUGLOG(5, "ZSTD_compressSubBlock_multi compressed"); - return op-ostart; + + DEBUGLOG(5, "ZSTD_compressSubBlock_multi compressed all subBlocks: total compressed size = %u", + (unsigned)(op-ostart)); + return (size_t)(op-ostart); } size_t ZSTD_compressSuperBlock(ZSTD_CCtx* zc, void* dst, size_t dstCapacity, - void const* src, size_t srcSize, - unsigned lastBlock) { + const void* src, size_t srcSize, + unsigned lastBlock) +{ ZSTD_entropyCTablesMetadata_t entropyMetadata; FORWARD_IF_ERROR(ZSTD_buildBlockEntropyStats(&zc->seqStore, &zc->blockState.prevCBlock->entropy, &zc->blockState.nextCBlock->entropy,