From: Nick Terrell Date: Tue, 3 Nov 2020 01:52:29 +0000 (-0800) Subject: Add and integrate lazy row hash strategy X-Git-Tag: v1.5.0^2~60^2~1 X-Git-Url: http://git.ipfire.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=4694423c4f73e7d6101e44c949cfaaa0205f1500;p=thirdparty%2Fzstd.git Add and integrate lazy row hash strategy --- diff --git a/.gitignore b/.gitignore index ae277e932..ea574d747 100644 --- a/.gitignore +++ b/.gitignore @@ -50,3 +50,5 @@ googletest/ *.code-workspace compile_commands.json .clangd +perf.data +perf.data.old diff --git a/lib/compress/zstd_compress.c b/lib/compress/zstd_compress.c index 2eb9cb7ac..cb58ca294 100644 --- a/lib/compress/zstd_compress.c +++ b/lib/compress/zstd_compress.c @@ -72,6 +72,10 @@ struct ZSTD_CDict_s { ZSTD_customMem customMem; U32 dictID; int compressionLevel; /* 0 indicates that advanced API was used to select CDict params */ + ZSTD_useRowMatchFinderMode_e useRowMatchFinder; /* Indicates whether the CDict was created with params that would use + * row-based matchfinder. Unless the cdict is reloaded, we will use + * the same greedy/lazy matchfinder at compression time. + */ }; /* typedef'd to ZSTD_CDict within "zstd.h" */ ZSTD_CCtx* ZSTD_createCCtx(void) @@ -202,6 +206,49 @@ size_t ZSTD_sizeof_CStream(const ZSTD_CStream* zcs) /* private API call, for dictBuilder only */ const seqStore_t* ZSTD_getSeqStore(const ZSTD_CCtx* ctx) { return &(ctx->seqStore); } +/* Returns true if the strategy supports using a row based matchfinder */ +static int ZSTD_rowMatchFinderSupported(const ZSTD_strategy strategy) { + return (strategy >= ZSTD_greedy && strategy <= ZSTD_lazy2); +} + +/* Returns true if the strategy and useRowMatchFinder mode indicate that we will use the row based matchfinder + * for this compression. + */ +static int ZSTD_rowMatchFinderUsed(const ZSTD_strategy strategy, const ZSTD_useRowMatchFinderMode_e mode) { + assert(mode != ZSTD_urm_auto); + return ZSTD_rowMatchFinderSupported(strategy) && (mode == ZSTD_urm_enableRowMatchFinder); +} + +/* Returns row matchfinder usage enum given an initial mode and cParams */ +static ZSTD_useRowMatchFinderMode_e ZSTD_resolveRowMatchFinderMode(ZSTD_useRowMatchFinderMode_e mode, + const ZSTD_compressionParameters* const cParams) { +#if !defined(ZSTD_NO_INTRINSICS) && (defined(__SSE2__) || defined(__ARM_NEON)) + int const kHasSIMD128 = 1; +#else + int const kHasSIMD128 = 0; +#endif + if (mode != ZSTD_urm_auto) return mode; /* if requested enabled, but no SIMD, we still will use row matchfinder */ + mode = ZSTD_urm_disableRowMatchFinder; + if (!ZSTD_rowMatchFinderSupported(cParams->strategy)) return mode; + if (kHasSIMD128) { + if (cParams->windowLog > 14) mode = ZSTD_urm_enableRowMatchFinder; + } else { + if (cParams->windowLog > 17) mode = ZSTD_urm_enableRowMatchFinder; + } + return mode; +} + +/* Returns 1 if the arguments indicate that we should allocate a chainTable, 0 otherwise */ +static int ZSTD_allocateChainTable(const ZSTD_strategy strategy, + const ZSTD_useRowMatchFinderMode_e useRowMatchFinder, + const U32 forDDSDict) { + assert(useRowMatchFinder != ZSTD_urm_auto); + /* We always should allocate a chaintable if we are allocating a matchstate for a DDS dictionary matchstate. + * We do not allocate a chaintable if we are using ZSTD_fast, or are using the row-based matchfinder. + */ + return forDDSDict || ((strategy != ZSTD_fast) && !ZSTD_rowMatchFinderUsed(strategy, useRowMatchFinder)); +} + /* Returns 1 if compression parameters are such that we should * enable long distance matching (wlog >= 27, strategy >= btopt). * Returns 0 otherwise. @@ -241,6 +288,7 @@ static ZSTD_CCtx_params ZSTD_makeCCtxParamsFromCParams( cctxParams.splitBlocks = 1; } + cctxParams.useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(cctxParams.useRowMatchFinder, &cParams); assert(!ZSTD_checkCParams(cParams)); return cctxParams; } @@ -299,6 +347,8 @@ static void ZSTD_CCtxParams_init_internal(ZSTD_CCtx_params* cctxParams, ZSTD_par * But, set it for tracing anyway. */ cctxParams->compressionLevel = compressionLevel; + cctxParams->useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(cctxParams->useRowMatchFinder, ¶ms->cParams); + DEBUGLOG(4, "ZSTD_CCtxParams_init_internal: useRowMatchFinder=%d", cctxParams->useRowMatchFinder); } size_t ZSTD_CCtxParams_init_advanced(ZSTD_CCtx_params* cctxParams, ZSTD_parameters params) @@ -498,12 +548,17 @@ ZSTD_bounds ZSTD_cParam_getBounds(ZSTD_cParameter param) bounds.lowerBound = 0; bounds.upperBound = 1; return bounds; - + case ZSTD_c_splitBlocks: bounds.lowerBound = 0; bounds.upperBound = 1; return bounds; + case ZSTD_c_useRowMatchFinder: + bounds.lowerBound = (int)ZSTD_urm_auto; + bounds.upperBound = (int)ZSTD_urm_enableRowMatchFinder; + return bounds; + default: bounds.error = ERROR(parameter_unsupported); return bounds; @@ -566,6 +621,7 @@ static int ZSTD_isUpdateAuthorized(ZSTD_cParameter param) case ZSTD_c_blockDelimiters: case ZSTD_c_validateSequences: case ZSTD_c_splitBlocks: + case ZSTD_c_useRowMatchFinder: default: return 0; } @@ -619,6 +675,7 @@ size_t ZSTD_CCtx_setParameter(ZSTD_CCtx* cctx, ZSTD_cParameter param, int value) case ZSTD_c_blockDelimiters: case ZSTD_c_validateSequences: case ZSTD_c_splitBlocks: + case ZSTD_c_useRowMatchFinder: break; default: RETURN_ERROR(parameter_unsupported, "unknown parameter"); @@ -835,6 +892,11 @@ size_t ZSTD_CCtxParams_setParameter(ZSTD_CCtx_params* CCtxParams, CCtxParams->splitBlocks = value; return CCtxParams->splitBlocks; + case ZSTD_c_useRowMatchFinder: + BOUNDCHECK(ZSTD_c_useRowMatchFinder, value); + CCtxParams->useRowMatchFinder = (ZSTD_useRowMatchFinderMode_e)value; + return CCtxParams->useRowMatchFinder; + default: RETURN_ERROR(parameter_unsupported, "unknown parameter"); } } @@ -961,6 +1023,9 @@ size_t ZSTD_CCtxParams_getParameter( case ZSTD_c_splitBlocks : *value = (int)CCtxParams->splitBlocks; break; + case ZSTD_c_useRowMatchFinder : + *value = (int)CCtxParams->useRowMatchFinder; + break; default: RETURN_ERROR(parameter_unsupported, "unknown parameter"); } return 0; @@ -1327,9 +1392,14 @@ ZSTD_compressionParameters ZSTD_getCParamsFromCCtxParams( static size_t ZSTD_sizeof_matchState(const ZSTD_compressionParameters* const cParams, + const ZSTD_useRowMatchFinderMode_e useRowMatchFinder, + const U32 enableDedicatedDictSearch, const U32 forCCtx) { - size_t const chainSize = (cParams->strategy == ZSTD_fast) ? 0 : ((size_t)1 << cParams->chainLog); + /* chain table size should be 0 for fast or row-hash strategies */ + size_t const chainSize = ZSTD_allocateChainTable(cParams->strategy, useRowMatchFinder, enableDedicatedDictSearch && !forCCtx) + ? ((size_t)1 << cParams->chainLog) + : 0; size_t const hSize = ((size_t)1) << cParams->hashLog; U32 const hashLog3 = (forCCtx && cParams->minMatch==3) ? MIN(ZSTD_HASHLOG3_MAX, cParams->windowLog) : 0; size_t const h3Size = hashLog3 ? ((size_t)1) << hashLog3 : 0; @@ -1345,6 +1415,9 @@ ZSTD_sizeof_matchState(const ZSTD_compressionParameters* const cParams, + ZSTD_cwksp_aligned_alloc_size((1<strategy, useRowMatchFinder) + ? ZSTD_cwksp_aligned_alloc_size(hSize*sizeof(U16)) + : 0; size_t const optSpace = (forCCtx && (cParams->strategy >= ZSTD_btopt)) ? optPotentialSpace : 0; @@ -1352,16 +1425,18 @@ ZSTD_sizeof_matchState(const ZSTD_compressionParameters* const cParams, /* tables are guaranteed to be sized in multiples of 64 bytes (or 16 uint32_t) */ ZSTD_STATIC_ASSERT(ZSTD_HASHLOG_MIN >= 4 && ZSTD_WINDOWLOG_MIN >= 4 && ZSTD_CHAINLOG_MIN >= 4); + assert(useRowMatchFinder != ZSTD_urm_auto); DEBUGLOG(4, "chainSize: %u - hSize: %u - h3Size: %u", (U32)chainSize, (U32)hSize, (U32)h3Size); - return tableSpace + optSpace + slackSpace; + return tableSpace + optSpace + slackSpace + lazyAdditionalSpace; } static size_t ZSTD_estimateCCtxSize_usingCCtxParams_internal( const ZSTD_compressionParameters* cParams, const ldmParams_t* ldmParams, const int isStatic, + const ZSTD_useRowMatchFinderMode_e useRowMatchFinder, const size_t buffInSize, const size_t buffOutSize, const U64 pledgedSrcSize) @@ -1375,7 +1450,7 @@ static size_t ZSTD_estimateCCtxSize_usingCCtxParams_internal( + 3 * ZSTD_cwksp_alloc_size(maxNbSeq * sizeof(BYTE)); size_t const entropySpace = ZSTD_cwksp_alloc_size(ENTROPY_WORKSPACE_SIZE); size_t const blockStateSpace = 2 * ZSTD_cwksp_alloc_size(sizeof(ZSTD_compressedBlockState_t)); - size_t const matchStateSize = ZSTD_sizeof_matchState(cParams, /* forCCtx */ 1); + size_t const matchStateSize = ZSTD_sizeof_matchState(cParams, useRowMatchFinder, /* enableDedicatedDictSearch */ 0, /* forCCtx */ 1); size_t const ldmSpace = ZSTD_ldm_getTableSize(*ldmParams); size_t const maxNbLdmSeq = ZSTD_ldm_getMaxNbSeq(*ldmParams, blockSize); @@ -1406,19 +1481,32 @@ size_t ZSTD_estimateCCtxSize_usingCCtxParams(const ZSTD_CCtx_params* params) { ZSTD_compressionParameters const cParams = ZSTD_getCParamsFromCCtxParams(params, ZSTD_CONTENTSIZE_UNKNOWN, 0, ZSTD_cpm_noAttachDict); + ZSTD_useRowMatchFinderMode_e const useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(params->useRowMatchFinder, + &cParams); RETURN_ERROR_IF(params->nbWorkers > 0, GENERIC, "Estimate CCtx size is supported for single-threaded compression only."); /* estimateCCtxSize is for one-shot compression. So no buffers should * be needed. However, we still allocate two 0-sized buffers, which can * take space under ASAN. */ return ZSTD_estimateCCtxSize_usingCCtxParams_internal( - &cParams, ¶ms->ldmParams, 1, 0, 0, ZSTD_CONTENTSIZE_UNKNOWN); + &cParams, ¶ms->ldmParams, 1, useRowMatchFinder, 0, 0, ZSTD_CONTENTSIZE_UNKNOWN); } size_t ZSTD_estimateCCtxSize_usingCParams(ZSTD_compressionParameters cParams) { - ZSTD_CCtx_params const params = ZSTD_makeCCtxParamsFromCParams(cParams); - return ZSTD_estimateCCtxSize_usingCCtxParams(¶ms); + ZSTD_CCtx_params initialParams = ZSTD_makeCCtxParamsFromCParams(cParams); + if (ZSTD_rowMatchFinderSupported(cParams.strategy)) { + /* Pick bigger of not using and using row-based matchfinder for greedy and lazy strategies */ + size_t noRowCCtxSize; + size_t rowCCtxSize; + initialParams.useRowMatchFinder = ZSTD_urm_disableRowMatchFinder; + noRowCCtxSize = ZSTD_estimateCCtxSize_usingCCtxParams(&initialParams); + initialParams.useRowMatchFinder = ZSTD_urm_enableRowMatchFinder; + rowCCtxSize = ZSTD_estimateCCtxSize_usingCCtxParams(&initialParams); + return MAX(noRowCCtxSize, rowCCtxSize); + } else { + return ZSTD_estimateCCtxSize_usingCCtxParams(&initialParams); + } } static size_t ZSTD_estimateCCtxSize_internal(int compressionLevel) @@ -1458,17 +1546,29 @@ size_t ZSTD_estimateCStreamSize_usingCCtxParams(const ZSTD_CCtx_params* params) size_t const outBuffSize = (params->outBufferMode == ZSTD_bm_buffered) ? ZSTD_compressBound(blockSize) + 1 : 0; + ZSTD_useRowMatchFinderMode_e const useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(params->useRowMatchFinder, ¶ms->cParams); return ZSTD_estimateCCtxSize_usingCCtxParams_internal( - &cParams, ¶ms->ldmParams, 1, inBuffSize, outBuffSize, + &cParams, ¶ms->ldmParams, 1, useRowMatchFinder, inBuffSize, outBuffSize, ZSTD_CONTENTSIZE_UNKNOWN); } } size_t ZSTD_estimateCStreamSize_usingCParams(ZSTD_compressionParameters cParams) { - ZSTD_CCtx_params const params = ZSTD_makeCCtxParamsFromCParams(cParams); - return ZSTD_estimateCStreamSize_usingCCtxParams(¶ms); + ZSTD_CCtx_params initialParams = ZSTD_makeCCtxParamsFromCParams(cParams); + if (ZSTD_rowMatchFinderSupported(cParams.strategy)) { + /* Pick bigger of not using and using row-based matchfinder for greedy and lazy strategies */ + size_t noRowCCtxSize; + size_t rowCCtxSize; + initialParams.useRowMatchFinder = ZSTD_urm_disableRowMatchFinder; + noRowCCtxSize = ZSTD_estimateCStreamSize_usingCCtxParams(&initialParams); + initialParams.useRowMatchFinder = ZSTD_urm_enableRowMatchFinder; + rowCCtxSize = ZSTD_estimateCStreamSize_usingCCtxParams(&initialParams); + return MAX(noRowCCtxSize, rowCCtxSize); + } else { + return ZSTD_estimateCStreamSize_usingCCtxParams(&initialParams); + } } static size_t ZSTD_estimateCStreamSize_internal(int compressionLevel) @@ -1593,20 +1693,27 @@ typedef enum { ZSTD_resetTarget_CCtx } ZSTD_resetTarget_e; + static size_t ZSTD_reset_matchState(ZSTD_matchState_t* ms, ZSTD_cwksp* ws, const ZSTD_compressionParameters* cParams, + const ZSTD_useRowMatchFinderMode_e useRowMatchFinder, const ZSTD_compResetPolicy_e crp, const ZSTD_indexResetPolicy_e forceResetIndex, const ZSTD_resetTarget_e forWho) { - size_t const chainSize = (cParams->strategy == ZSTD_fast) ? 0 : ((size_t)1 << cParams->chainLog); + /* disable chain table allocation for fast or row-based strategies */ + size_t const chainSize = ZSTD_allocateChainTable(cParams->strategy, useRowMatchFinder, + ms->dedicatedDictSearch && (forWho == ZSTD_resetTarget_CDict)) + ? ((size_t)1 << cParams->chainLog) + : 0; size_t const hSize = ((size_t)1) << cParams->hashLog; U32 const hashLog3 = ((forWho == ZSTD_resetTarget_CCtx) && cParams->minMatch==3) ? MIN(ZSTD_HASHLOG3_MAX, cParams->windowLog) : 0; size_t const h3Size = hashLog3 ? ((size_t)1) << hashLog3 : 0; DEBUGLOG(4, "reset indices : %u", forceResetIndex == ZSTDirp_reset); + assert(useRowMatchFinder != ZSTD_urm_auto); if (forceResetIndex == ZSTDirp_reset) { ZSTD_window_init(&ms->window); ZSTD_cwksp_mark_tables_dirty(ws); @@ -1645,11 +1752,23 @@ ZSTD_reset_matchState(ZSTD_matchState_t* ms, ms->opt.priceTable = (ZSTD_optimal_t*)ZSTD_cwksp_reserve_aligned(ws, (ZSTD_OPT_NUM+1) * sizeof(ZSTD_optimal_t)); } + if (ZSTD_rowMatchFinderUsed(cParams->strategy, useRowMatchFinder)) { + { /* Row match finder needs an additional table of hashes ("tags") */ + size_t const tagTableSize = hSize*sizeof(U16); + ms->tagTable = (U16*)ZSTD_cwksp_reserve_aligned(ws, tagTableSize); + if (ms->tagTable) ZSTD_memset(ms->tagTable, 0, tagTableSize); + } + { /* Switch to 32-entry rows if searchLog is 5 (or more) */ + U32 const rowLog = cParams->searchLog < 5 ? 4 : 5; + assert(cParams->hashLog > rowLog); + ms->rowHashLog = ZSTD_highbit32((U32)1 << (cParams->hashLog - rowLog)); + } + } + ms->cParams = *cParams; RETURN_ERROR_IF(ZSTD_cwksp_reserve_failed(ws), memory_allocation, "failed a workspace allocation in ZSTD_reset_matchState"); - return 0; } @@ -1675,12 +1794,13 @@ static size_t ZSTD_resetCCtx_internal(ZSTD_CCtx* zc, ZSTD_buffered_policy_e const zbuff) { ZSTD_cwksp* const ws = &zc->workspace; - DEBUGLOG(4, "ZSTD_resetCCtx_internal: pledgedSrcSize=%u, wlog=%u", - (U32)pledgedSrcSize, params.cParams.windowLog); + DEBUGLOG(4, "ZSTD_resetCCtx_internal: pledgedSrcSize=%u, wlog=%u, useRowMatchFinder=%d", + (U32)pledgedSrcSize, params.cParams.windowLog, (int)params.useRowMatchFinder); assert(!ZSTD_isError(ZSTD_checkCParams(params.cParams))); zc->isFirstBlock = 1; + assert(params.useRowMatchFinder != ZSTD_urm_auto); if (params.ldmParams.enableLdm) { /* Adjust long distance matching parameters */ ZSTD_ldm_adjustParameters(¶ms.ldmParams, ¶ms.cParams); @@ -1706,7 +1826,7 @@ static size_t ZSTD_resetCCtx_internal(ZSTD_CCtx* zc, size_t const neededSpace = ZSTD_estimateCCtxSize_usingCCtxParams_internal( - ¶ms.cParams, ¶ms.ldmParams, zc->staticSize != 0, + ¶ms.cParams, ¶ms.ldmParams, zc->staticSize != 0, params.useRowMatchFinder, buffInSize, buffOutSize, pledgedSrcSize); int resizeWorkspace; @@ -1802,6 +1922,7 @@ static size_t ZSTD_resetCCtx_internal(ZSTD_CCtx* zc, &zc->blockState.matchState, ws, ¶ms.cParams, + params.useRowMatchFinder, crp, needsIndexReset, ZSTD_resetTarget_CCtx), ""); @@ -1878,6 +1999,7 @@ ZSTD_resetCCtx_byAttachingCDict(ZSTD_CCtx* cctx, U64 pledgedSrcSize, ZSTD_buffered_policy_e zbuff) { + DEBUGLOG(4, "ZSTD_resetCCtx_byAttachingCDict() pledgedSrcSize=%zu", pledgedSrcSize); { ZSTD_compressionParameters adjusted_cdict_cParams = cdict->matchState.cParams; unsigned const windowLog = params.cParams.windowLog; @@ -1893,6 +2015,7 @@ ZSTD_resetCCtx_byAttachingCDict(ZSTD_CCtx* cctx, params.cParams = ZSTD_adjustCParams_internal(adjusted_cdict_cParams, pledgedSrcSize, cdict->dictContentSize, ZSTD_cpm_attachDict); params.cParams.windowLog = windowLog; + params.useRowMatchFinder = cdict->useRowMatchFinder; /* cdict overrides */ FORWARD_IF_ERROR(ZSTD_resetCCtx_internal(cctx, params, pledgedSrcSize, ZSTDcrp_makeClean, zbuff), ""); assert(cctx->appliedParams.cParams.strategy == adjusted_cdict_cParams.strategy); @@ -1937,14 +2060,14 @@ static size_t ZSTD_resetCCtx_byCopyingCDict(ZSTD_CCtx* cctx, const ZSTD_compressionParameters *cdict_cParams = &cdict->matchState.cParams; assert(!cdict->matchState.dedicatedDictSearch); - - DEBUGLOG(4, "copying dictionary into context"); + DEBUGLOG(4, "ZSTD_resetCCtx_byCopyingCDict() pledgedSrcSize=%zu", pledgedSrcSize); { unsigned const windowLog = params.cParams.windowLog; assert(windowLog != 0); /* Copy only compression parameters related to tables. */ params.cParams = *cdict_cParams; params.cParams.windowLog = windowLog; + params.useRowMatchFinder = cdict->useRowMatchFinder; FORWARD_IF_ERROR(ZSTD_resetCCtx_internal(cctx, params, pledgedSrcSize, ZSTDcrp_leaveDirty, zbuff), ""); assert(cctx->appliedParams.cParams.strategy == cdict_cParams->strategy); @@ -1953,17 +2076,30 @@ static size_t ZSTD_resetCCtx_byCopyingCDict(ZSTD_CCtx* cctx, } ZSTD_cwksp_mark_tables_dirty(&cctx->workspace); + assert(params.useRowMatchFinder != ZSTD_urm_auto); /* copy tables */ - { size_t const chainSize = (cdict_cParams->strategy == ZSTD_fast) ? 0 : ((size_t)1 << cdict_cParams->chainLog); + { size_t const chainSize = ZSTD_allocateChainTable(cdict_cParams->strategy, cdict->useRowMatchFinder, 0 /* DDS guaranteed disabled */) + ? ((size_t)1 << cdict_cParams->chainLog) + : 0; size_t const hSize = (size_t)1 << cdict_cParams->hashLog; ZSTD_memcpy(cctx->blockState.matchState.hashTable, cdict->matchState.hashTable, hSize * sizeof(U32)); - ZSTD_memcpy(cctx->blockState.matchState.chainTable, + /* Do not copy cdict's chainTable if cctx has parameters such that it would not use chainTable */ + if (ZSTD_allocateChainTable(cctx->appliedParams.cParams.strategy, cctx->appliedParams.useRowMatchFinder, 0 /* forDDSDict */)) { + ZSTD_memcpy(cctx->blockState.matchState.chainTable, cdict->matchState.chainTable, chainSize * sizeof(U32)); + } + /* copy tag table */ + if (ZSTD_rowMatchFinderUsed(cdict_cParams->strategy, cdict->useRowMatchFinder)) { + size_t const tagTableSize = hSize*sizeof(U16); + ZSTD_memcpy(cctx->blockState.matchState.tagTable, + cdict->matchState.tagTable, + tagTableSize); + } } /* Zero the hashTable3, since the cdict never fills it */ @@ -2027,14 +2163,15 @@ static size_t ZSTD_copyCCtx_internal(ZSTD_CCtx* dstCCtx, U64 pledgedSrcSize, ZSTD_buffered_policy_e zbuff) { - DEBUGLOG(5, "ZSTD_copyCCtx_internal"); RETURN_ERROR_IF(srcCCtx->stage!=ZSTDcs_init, stage_wrong, "Can't copy a ctx that's not in init stage."); - + DEBUGLOG(5, "ZSTD_copyCCtx_internal"); ZSTD_memcpy(&dstCCtx->customMem, &srcCCtx->customMem, sizeof(ZSTD_customMem)); { ZSTD_CCtx_params params = dstCCtx->requestedParams; /* Copy only compression parameters related to tables. */ params.cParams = srcCCtx->appliedParams.cParams; + assert(srcCCtx->appliedParams.useRowMatchFinder != ZSTD_urm_auto); + params.useRowMatchFinder = srcCCtx->appliedParams.useRowMatchFinder; params.fParams = fParams; ZSTD_resetCCtx_internal(dstCCtx, params, pledgedSrcSize, ZSTDcrp_leaveDirty, zbuff); @@ -2048,7 +2185,11 @@ static size_t ZSTD_copyCCtx_internal(ZSTD_CCtx* dstCCtx, ZSTD_cwksp_mark_tables_dirty(&dstCCtx->workspace); /* copy tables */ - { size_t const chainSize = (srcCCtx->appliedParams.cParams.strategy == ZSTD_fast) ? 0 : ((size_t)1 << srcCCtx->appliedParams.cParams.chainLog); + { size_t const chainSize = ZSTD_allocateChainTable(srcCCtx->appliedParams.cParams.strategy, + srcCCtx->appliedParams.useRowMatchFinder, + 0 /* forDDSDict */) + ? ((size_t)1 << srcCCtx->appliedParams.cParams.chainLog) + : 0; size_t const hSize = (size_t)1 << srcCCtx->appliedParams.cParams.hashLog; int const h3log = srcCCtx->blockState.matchState.hashLog3; size_t const h3Size = h3log ? ((size_t)1 << h3log) : 0; @@ -2239,7 +2380,7 @@ typedef struct { /* ZSTD_buildSequencesStatistics(): * Returns the size of the statistics for a given set of sequences, or a ZSTD error code, * Also modifies LLtype, Offtype, MLtype, and lastNCount to the appropriate values. - * + * * entropyWkspSize must be of size at least ENTROPY_WORKSPACE_SIZE - (MaxSeq + 1)*sizeof(U32) */ static ZSTD_symbolEncodingTypeStats_t @@ -2510,7 +2651,7 @@ ZSTD_entropyCompressSeqStore(seqStore_t* seqStorePtr, /* ZSTD_selectBlockCompressor() : * Not static, but internal use only (used by long distance matcher) * assumption : strat is a valid strategy */ -ZSTD_blockCompressor ZSTD_selectBlockCompressor(ZSTD_strategy strat, ZSTD_dictMode_e dictMode) +ZSTD_blockCompressor ZSTD_selectBlockCompressor(ZSTD_strategy strat, ZSTD_useRowMatchFinderMode_e useRowMatchFinder, ZSTD_dictMode_e dictMode) { static const ZSTD_blockCompressor blockCompressor[4][ZSTD_STRATEGY_MAX+1] = { { ZSTD_compressBlock_fast /* default for 0 */, @@ -2558,7 +2699,28 @@ ZSTD_blockCompressor ZSTD_selectBlockCompressor(ZSTD_strategy strat, ZSTD_dictMo ZSTD_STATIC_ASSERT((unsigned)ZSTD_fast == 1); assert(ZSTD_cParam_withinBounds(ZSTD_c_strategy, strat)); - selectedCompressor = blockCompressor[(int)dictMode][(int)strat]; + DEBUGLOG(4, "Selected block compressor: dictMode=%d strat=%d rowMatchfinder=%d", (int)dictMode, (int)strat, (int)useRowMatchFinder); + if (ZSTD_rowMatchFinderUsed(strat, useRowMatchFinder)) { + static const ZSTD_blockCompressor rowBasedBlockCompressors[4][3] = { + { ZSTD_compressBlock_greedy_row, + ZSTD_compressBlock_lazy_row, + ZSTD_compressBlock_lazy2_row }, + { ZSTD_compressBlock_greedy_extDict_row, + ZSTD_compressBlock_lazy_extDict_row, + ZSTD_compressBlock_lazy2_extDict_row }, + { ZSTD_compressBlock_greedy_dictMatchState_row, + ZSTD_compressBlock_lazy_dictMatchState_row, + ZSTD_compressBlock_lazy2_dictMatchState_row }, + { ZSTD_compressBlock_greedy_dedicatedDictSearch_row, + ZSTD_compressBlock_lazy_dedicatedDictSearch_row, + ZSTD_compressBlock_lazy2_dedicatedDictSearch_row } + }; + DEBUGLOG(4, "Selecting a row-based matchfinder"); + assert(useRowMatchFinder != ZSTD_urm_auto); + selectedCompressor = rowBasedBlockCompressors[(int)dictMode][(int)strat - (int)ZSTD_greedy]; + } else { + selectedCompressor = blockCompressor[(int)dictMode][(int)strat]; + } assert(selectedCompressor != NULL); return selectedCompressor; } @@ -2627,6 +2789,7 @@ static size_t ZSTD_buildSeqStore(ZSTD_CCtx* zc, const void* src, size_t srcSize) ZSTD_ldm_blockCompress(&zc->externSeqStore, ms, &zc->seqStore, zc->blockState.nextCBlock->rep, + zc->appliedParams.useRowMatchFinder, src, srcSize); assert(zc->externSeqStore.pos <= zc->externSeqStore.size); } else if (zc->appliedParams.ldmParams.enableLdm) { @@ -2643,10 +2806,13 @@ static size_t ZSTD_buildSeqStore(ZSTD_CCtx* zc, const void* src, size_t srcSize) ZSTD_ldm_blockCompress(&ldmSeqStore, ms, &zc->seqStore, zc->blockState.nextCBlock->rep, + zc->appliedParams.useRowMatchFinder, src, srcSize); assert(ldmSeqStore.pos == ldmSeqStore.size); } else { /* not long range mode */ - ZSTD_blockCompressor const blockCompressor = ZSTD_selectBlockCompressor(zc->appliedParams.cParams.strategy, dictMode); + ZSTD_blockCompressor const blockCompressor = ZSTD_selectBlockCompressor(zc->appliedParams.cParams.strategy, + zc->appliedParams.useRowMatchFinder, + dictMode); ms->ldmSeqStore = NULL; lastLLSize = blockCompressor(ms, &zc->seqStore, zc->blockState.nextCBlock->rep, src, srcSize); } @@ -2954,7 +3120,7 @@ static size_t ZSTD_buildBlockEntropyStats_sequences(seqStore_t* seqStorePtr, /** ZSTD_buildBlockEntropyStats() : * Builds entropy for the block. * Requires workspace size ENTROPY_WORKSPACE_SIZE - * + * * @return : 0 on success or error code */ size_t ZSTD_buildBlockEntropyStats(seqStore_t* seqStorePtr, @@ -3219,7 +3385,7 @@ static void ZSTD_seqStore_resolveOffCodes(repcodes_t* const dRepcodes, repcodes_ /* ZSTD_compressSeqStore_singleBlock(): * Compresses a seqStore into a block with a block header, into the buffer dst. - * + * * Returns the total size of that block (including header) or a ZSTD error code. */ static size_t ZSTD_compressSeqStore_singleBlock(ZSTD_CCtx* zc, seqStore_t* const seqStore, @@ -3298,11 +3464,11 @@ typedef struct { * Estimates the cost of seqStore prior to split, and estimates the cost of splitting the sequences in half. * If advantageous to split, then we recurse down the two sub-blocks. If not, or if an error occurred in estimation, then * we do not recurse. - * + * * Note: The recursion depth is capped by a heuristic minimum number of sequences, defined by MIN_SEQUENCES_BLOCK_SPLITTING. * In theory, this means the absolute largest recursion depth is 10 == log2(maxNbSeqInBlock/MIN_SEQUENCES_BLOCK_SPLITTING). * In practice, recursion depth usually doesn't go beyond 4. - * + * * Furthermore, the number of splits is capped by MAX_NB_SPLITS. At MAX_NB_SPLITS == 196 with the current existing blockSize * maximum of 128 KB, this value is actually impossible to reach. */ @@ -3357,7 +3523,7 @@ static size_t ZSTD_deriveBlockSplits(ZSTD_CCtx* zc, U32 partitions[], U32 nbSeq) /* ZSTD_compressBlock_splitBlock(): * Attempts to split a given block into multiple blocks to improve compression ratio. - * + * * Returns combined size of all blocks (which includes headers), or a ZSTD error code. */ static size_t ZSTD_compressBlock_splitBlock_internal(ZSTD_CCtx* zc, void* dst, size_t dstCapacity, @@ -3921,6 +4087,7 @@ static size_t ZSTD_loadDictionaryContent(ZSTD_matchState_t* ms, const BYTE* ip = (const BYTE*) src; const BYTE* const iend = ip + srcSize; + DEBUGLOG(4, "ZSTD_loadDictionaryContent(): useRowMatchFinder=%d", (int)params->useRowMatchFinder); ZSTD_window_update(&ms->window, src, srcSize); ms->loadedDictEnd = params->forceWindow ? 0 : (U32)(iend - ms->window.base); @@ -3956,11 +4123,24 @@ static size_t ZSTD_loadDictionaryContent(ZSTD_matchState_t* ms, case ZSTD_greedy: case ZSTD_lazy: case ZSTD_lazy2: - if (chunk >= HASH_READ_SIZE && ms->dedicatedDictSearch) { - assert(chunk == remaining); /* must load everything in one go */ - ZSTD_dedicatedDictSearch_lazy_loadDictionary(ms, ichunk-HASH_READ_SIZE); - } else if (chunk >= HASH_READ_SIZE) { - ZSTD_insertAndFindFirstIndex(ms, ichunk-HASH_READ_SIZE); + if (chunk >= HASH_READ_SIZE) { + if (ms->dedicatedDictSearch) { + assert(chunk == remaining); /* must load everything in one go */ + assert(ms->chainTable != NULL); + ZSTD_dedicatedDictSearch_lazy_loadDictionary(ms, ichunk-HASH_READ_SIZE); + } else { + assert(params->useRowMatchFinder != ZSTD_urm_auto); + if (params->useRowMatchFinder == ZSTD_urm_enableRowMatchFinder) { + size_t const tagTableSize = ((size_t)1 << params->cParams.hashLog) * sizeof(U16); + if (ip == src) + ZSTD_memset(ms->tagTable, 0, tagTableSize); + ZSTD_row_update(ms, ichunk-HASH_READ_SIZE); + DEBUGLOG(4, "Using row-based hash table for lazy dict"); + } else { + ZSTD_insertAndFindFirstIndex(ms, ichunk-HASH_READ_SIZE); + DEBUGLOG(4, "Using chain-based hash table for lazy dict"); + } + } } break; @@ -4115,7 +4295,6 @@ static size_t ZSTD_loadZstdDictionary(ZSTD_compressedBlockState_t* bs, const BYTE* const dictEnd = dictPtr + dictSize; size_t dictID; size_t eSize; - ZSTD_STATIC_ASSERT(HUF_WORKSPACE_SIZE >= (1<= 8); assert(MEM_readLE32(dictPtr) == ZSTD_MAGIC_DICTIONARY); @@ -4454,7 +4633,10 @@ size_t ZSTD_estimateCDictSize_advanced( DEBUGLOG(5, "sizeof(ZSTD_CDict) : %u", (unsigned)sizeof(ZSTD_CDict)); return ZSTD_cwksp_alloc_size(sizeof(ZSTD_CDict)) + ZSTD_cwksp_alloc_size(HUF_WORKSPACE_SIZE) - + ZSTD_sizeof_matchState(&cParams, /* forCCtx */ 0) + /* enableDedicatedDictSearch == 1 ensures that CDict estimation will not be too small + * in case we are using DDS with row-hash. */ + + ZSTD_sizeof_matchState(&cParams, ZSTD_resolveRowMatchFinderMode(ZSTD_urm_auto, &cParams), + /* enableDedicatedDictSearch */ 1, /* forCCtx */ 0) + (dictLoadMethod == ZSTD_dlm_byRef ? 0 : ZSTD_cwksp_alloc_size(ZSTD_cwksp_align(dictSize, sizeof(void *)))); } @@ -4508,6 +4690,7 @@ static size_t ZSTD_initCDict_internal( &cdict->matchState, &cdict->workspace, ¶ms.cParams, + params.useRowMatchFinder, ZSTDcrp_makeClean, ZSTDirp_reset, ZSTD_resetTarget_CDict), ""); @@ -4531,14 +4714,17 @@ static size_t ZSTD_initCDict_internal( static ZSTD_CDict* ZSTD_createCDict_advanced_internal(size_t dictSize, ZSTD_dictLoadMethod_e dictLoadMethod, - ZSTD_compressionParameters cParams, ZSTD_customMem customMem) + ZSTD_compressionParameters cParams, + ZSTD_useRowMatchFinderMode_e useRowMatchFinder, + U32 enableDedicatedDictSearch, + ZSTD_customMem customMem) { if ((!customMem.customAlloc) ^ (!customMem.customFree)) return NULL; { size_t const workspaceSize = ZSTD_cwksp_alloc_size(sizeof(ZSTD_CDict)) + ZSTD_cwksp_alloc_size(HUF_WORKSPACE_SIZE) + - ZSTD_sizeof_matchState(&cParams, /* forCCtx */ 0) + + ZSTD_sizeof_matchState(&cParams, useRowMatchFinder, enableDedicatedDictSearch, /* forCCtx */ 0) + (dictLoadMethod == ZSTD_dlm_byRef ? 0 : ZSTD_cwksp_alloc_size(ZSTD_cwksp_align(dictSize, sizeof(void*)))); void* const workspace = ZSTD_customMalloc(workspaceSize, customMem); @@ -4557,7 +4743,7 @@ static ZSTD_CDict* ZSTD_createCDict_advanced_internal(size_t dictSize, ZSTD_cwksp_move(&cdict->workspace, &ws); cdict->customMem = customMem; cdict->compressionLevel = ZSTD_NO_CLEVEL; /* signals advanced API usage */ - + cdict->useRowMatchFinder = useRowMatchFinder; return cdict; } } @@ -4609,10 +4795,13 @@ ZSTDLIB_API ZSTD_CDict* ZSTD_createCDict_advanced2( &cctxParams, ZSTD_CONTENTSIZE_UNKNOWN, dictSize, ZSTD_cpm_createCDict); } + DEBUGLOG(3, "ZSTD_createCDict_advanced2: DDS: %u", cctxParams.enableDedicatedDictSearch); cctxParams.cParams = cParams; + cctxParams.useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(cctxParams.useRowMatchFinder, &cParams); cdict = ZSTD_createCDict_advanced_internal(dictSize, dictLoadMethod, cctxParams.cParams, + cctxParams.useRowMatchFinder, cctxParams.enableDedicatedDictSearch, customMem); if (ZSTD_isError( ZSTD_initCDict_internal(cdict, @@ -4681,7 +4870,9 @@ const ZSTD_CDict* ZSTD_initStaticCDict( ZSTD_dictContentType_e dictContentType, ZSTD_compressionParameters cParams) { - size_t const matchStateSize = ZSTD_sizeof_matchState(&cParams, /* forCCtx */ 0); + ZSTD_useRowMatchFinderMode_e const useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(ZSTD_urm_auto, &cParams); + /* enableDedicatedDictSearch == 1 ensures matchstate is not too small in case this CDict will be used for DDS + row hash */ + size_t const matchStateSize = ZSTD_sizeof_matchState(&cParams, useRowMatchFinder, /* enableDedicatedDictSearch */ 1, /* forCCtx */ 0); size_t const neededSize = ZSTD_cwksp_alloc_size(sizeof(ZSTD_CDict)) + (dictLoadMethod == ZSTD_dlm_byRef ? 0 : ZSTD_cwksp_alloc_size(ZSTD_cwksp_align(dictSize, sizeof(void*)))) @@ -4706,6 +4897,8 @@ const ZSTD_CDict* ZSTD_initStaticCDict( ZSTD_CCtxParams_init(¶ms, 0); params.cParams = cParams; + params.useRowMatchFinder = useRowMatchFinder; + cdict->useRowMatchFinder = useRowMatchFinder; if (ZSTD_isError( ZSTD_initCDict_internal(cdict, dict, dictSize, @@ -5242,6 +5435,8 @@ static size_t ZSTD_CCtx_init_compressStream2(ZSTD_CCtx* cctx, params.splitBlocks = 1; } + params.useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(params.useRowMatchFinder, ¶ms.cParams); + #ifdef ZSTD_MULTITHREAD if ((cctx->pledgedSrcSizePlusOne-1) <= ZSTDMT_JOBSIZE_MIN) { params.nbWorkers = 0; /* do not invoke multi-threading when src size is too small */ @@ -6051,6 +6246,7 @@ static ZSTD_compressionParameters ZSTD_getCParams_internal(int compressionLevel, else row = compressionLevel; { ZSTD_compressionParameters cp = ZSTD_defaultCParameters[tableID][row]; + DEBUGLOG(5, "ZSTD_getCParams_internal selected tableID: %u row: %u strat: %u", tableID, row, (U32)cp.strategy); /* acceleration factor */ if (compressionLevel < 0) { int const clampedCompressionLevel = MAX(ZSTD_minCLevel(), compressionLevel); diff --git a/lib/compress/zstd_compress_internal.h b/lib/compress/zstd_compress_internal.h index 447115478..3c488c61c 100644 --- a/lib/compress/zstd_compress_internal.h +++ b/lib/compress/zstd_compress_internal.h @@ -196,6 +196,9 @@ typedef struct { } ZSTD_window_t; typedef struct ZSTD_matchState_t ZSTD_matchState_t; + +#define ZSTD_ROW_HASH_CACHE_SIZE 8 /* Size of prefetching hash cache for row-based matchfinder */ + struct ZSTD_matchState_t { ZSTD_window_t window; /* State for window round buffer management */ U32 loadedDictEnd; /* index of end of dictionary, within context's referential. @@ -207,9 +210,15 @@ struct ZSTD_matchState_t { */ U32 nextToUpdate; /* index from which to continue table update */ U32 hashLog3; /* dispatch table for matches of len==3 : larger == faster, more memory */ + + U32 rowHashLog; /* For row-based matchfinder: Hashlog based on nb of rows in the hashTable.*/ + U16* tagTable; /* For row-based matchFinder: A row-based table containing the hashes and head index. */ + U32 hashCache[ZSTD_ROW_HASH_CACHE_SIZE]; /* For row-based matchFinder: a cache of hashes to improve speed */ + U32* hashTable; U32* hashTable3; U32* chainTable; + int dedicatedDictSearch; /* Indicates whether this matchState is using the * dedicated dictionary search structure. */ @@ -305,6 +314,9 @@ struct ZSTD_CCtx_params_s { /* Block splitting */ int splitBlocks; + /* Param for deciding whether to use row-based matchfinder */ + ZSTD_useRowMatchFinderMode_e useRowMatchFinder; + /* Internal use, for createCCtxParams() and freeCCtxParams() only */ ZSTD_customMem customMem; }; /* typedef'd to ZSTD_CCtx_params within "zstd.h" */ @@ -420,7 +432,7 @@ typedef enum { typedef size_t (*ZSTD_blockCompressor) ( ZSTD_matchState_t* bs, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], void const* src, size_t srcSize); -ZSTD_blockCompressor ZSTD_selectBlockCompressor(ZSTD_strategy strat, ZSTD_dictMode_e dictMode); +ZSTD_blockCompressor ZSTD_selectBlockCompressor(ZSTD_strategy strat, ZSTD_useRowMatchFinderMode_e rowMatchfinderMode, ZSTD_dictMode_e dictMode); MEM_STATIC U32 ZSTD_LLcode(U32 litLength) diff --git a/lib/compress/zstd_lazy.c b/lib/compress/zstd_lazy.c index 75745a79b..5cc242880 100644 --- a/lib/compress/zstd_lazy.c +++ b/lib/compress/zstd_lazy.c @@ -438,43 +438,9 @@ static size_t ZSTD_BtFindBestMatch_extDict_selectMLS ( } } - - -/* ********************************* -* Hash Chain +/*********************************** +* Dedicated dict search ***********************************/ -#define NEXT_IN_CHAIN(d, mask) chainTable[(d) & (mask)] - -/* Update chains up to ip (excluded) - Assumption : always within prefix (i.e. not within extDict) */ -FORCE_INLINE_TEMPLATE U32 ZSTD_insertAndFindFirstIndex_internal( - ZSTD_matchState_t* ms, - const ZSTD_compressionParameters* const cParams, - const BYTE* ip, U32 const mls) -{ - U32* const hashTable = ms->hashTable; - const U32 hashLog = cParams->hashLog; - U32* const chainTable = ms->chainTable; - const U32 chainMask = (1 << cParams->chainLog) - 1; - const BYTE* const base = ms->window.base; - const U32 target = (U32)(ip - base); - U32 idx = ms->nextToUpdate; - - while(idx < target) { /* catch up */ - size_t const h = ZSTD_hashPtr(base+idx, hashLog, mls); - NEXT_IN_CHAIN(idx, chainMask) = hashTable[h]; - hashTable[h] = idx; - idx++; - } - - ms->nextToUpdate = target; - return hashTable[ZSTD_hashPtr(ip, hashLog, mls)]; -} - -U32 ZSTD_insertAndFindFirstIndex(ZSTD_matchState_t* ms, const BYTE* ip) { - const ZSTD_compressionParameters* const cParams = &ms->cParams; - return ZSTD_insertAndFindFirstIndex_internal(ms, cParams, ip, ms->cParams.minMatch); -} void ZSTD_dedicatedDictSearch_lazy_loadDictionary(ZSTD_matchState_t* ms, const BYTE* const ip) { @@ -500,7 +466,6 @@ void ZSTD_dedicatedDictSearch_lazy_loadDictionary(ZSTD_matchState_t* ms, const B U32* const tmpChainTable = hashTable + ((size_t)1 << hashLog); U32 const tmpChainSize = ((1 << ZSTD_LAZY_DDSS_BUCKET_LOG) - 1) << hashLog; U32 const tmpMinChain = tmpChainSize < target ? target - tmpChainSize : idx; - U32 hashIdx; assert(ms->cParams.chainLog <= 24); @@ -591,6 +556,139 @@ void ZSTD_dedicatedDictSearch_lazy_loadDictionary(ZSTD_matchState_t* ms, const B ms->nextToUpdate = target; } +/* Returns the longest match length found in the dedicated dict search structure. + * If none are longer than the argument ml, then ml will be returned. + */ +FORCE_INLINE_TEMPLATE +size_t ZSTD_dedicatedDictSearch_lazy_search(size_t* offsetPtr, size_t ml, U32 nbAttempts, + const ZSTD_matchState_t* const dms, + const BYTE* const ip, const BYTE* const iLimit, + const BYTE* const prefixStart, const U32 curr, + const U32 dictLimit, const size_t ddsIdx) { + const U32 ddsLowestIndex = dms->window.dictLimit; + const BYTE* const ddsBase = dms->window.base; + const BYTE* const ddsEnd = dms->window.nextSrc; + const U32 ddsSize = (U32)(ddsEnd - ddsBase); + const U32 ddsIndexDelta = dictLimit - ddsSize; + const U32 bucketSize = (1 << ZSTD_LAZY_DDSS_BUCKET_LOG); + const U32 bucketLimit = nbAttempts < bucketSize - 1 ? nbAttempts : bucketSize - 1; + U32 ddsAttempt; + U32 matchIndex; + + for (ddsAttempt = 0; ddsAttempt < bucketSize - 1; ddsAttempt++) { + PREFETCH_L1(ddsBase + dms->hashTable[ddsIdx + ddsAttempt]); + } + + { + U32 const chainPackedPointer = dms->hashTable[ddsIdx + bucketSize - 1]; + U32 const chainIndex = chainPackedPointer >> 8; + + PREFETCH_L1(&dms->chainTable[chainIndex]); + } + + for (ddsAttempt = 0; ddsAttempt < bucketLimit; ddsAttempt++) { + size_t currentMl=0; + const BYTE* match; + matchIndex = dms->hashTable[ddsIdx + ddsAttempt]; + match = ddsBase + matchIndex; + + if (!matchIndex) { + return ml; + } + + /* guaranteed by table construction */ + (void)ddsLowestIndex; + assert(matchIndex >= ddsLowestIndex); + assert(match+4 <= ddsEnd); + if (MEM_read32(match) == MEM_read32(ip)) { + /* assumption : matchIndex <= dictLimit-4 (by table construction) */ + currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, ddsEnd, prefixStart) + 4; + } + + /* save best solution */ + if (currentMl > ml) { + ml = currentMl; + *offsetPtr = curr - (matchIndex + ddsIndexDelta) + ZSTD_REP_MOVE; + if (ip+currentMl == iLimit) { + /* best possible, avoids read overflow on next attempt */ + return ml; + } + } + } + + { + U32 const chainPackedPointer = dms->hashTable[ddsIdx + bucketSize - 1]; + U32 chainIndex = chainPackedPointer >> 8; + U32 const chainLength = chainPackedPointer & 0xFF; + U32 const chainAttempts = nbAttempts - ddsAttempt; + U32 const chainLimit = chainAttempts > chainLength ? chainLength : chainAttempts; + U32 chainAttempt; + + for (chainAttempt = 0 ; chainAttempt < chainLimit; chainAttempt++) { + PREFETCH_L1(ddsBase + dms->chainTable[chainIndex + chainAttempt]); + } + + for (chainAttempt = 0 ; chainAttempt < chainLimit; chainAttempt++, chainIndex++) { + size_t currentMl=0; + const BYTE* match; + matchIndex = dms->chainTable[chainIndex]; + match = ddsBase + matchIndex; + + /* guaranteed by table construction */ + assert(matchIndex >= ddsLowestIndex); + assert(match+4 <= ddsEnd); + if (MEM_read32(match) == MEM_read32(ip)) { + /* assumption : matchIndex <= dictLimit-4 (by table construction) */ + currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, ddsEnd, prefixStart) + 4; + } + + /* save best solution */ + if (currentMl > ml) { + ml = currentMl; + *offsetPtr = curr - (matchIndex + ddsIndexDelta) + ZSTD_REP_MOVE; + if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */ + } + } + } + return ml; +} + + +/* ********************************* +* Hash Chain +***********************************/ +#define NEXT_IN_CHAIN(d, mask) chainTable[(d) & (mask)] + +/* Update chains up to ip (excluded) + Assumption : always within prefix (i.e. not within extDict) */ +FORCE_INLINE_TEMPLATE U32 ZSTD_insertAndFindFirstIndex_internal( + ZSTD_matchState_t* ms, + const ZSTD_compressionParameters* const cParams, + const BYTE* ip, U32 const mls) +{ + U32* const hashTable = ms->hashTable; + const U32 hashLog = cParams->hashLog; + U32* const chainTable = ms->chainTable; + const U32 chainMask = (1 << cParams->chainLog) - 1; + const BYTE* const base = ms->window.base; + const U32 target = (U32)(ip - base); + U32 idx = ms->nextToUpdate; + + while(idx < target) { /* catch up */ + size_t const h = ZSTD_hashPtr(base+idx, hashLog, mls); + NEXT_IN_CHAIN(idx, chainMask) = hashTable[h]; + hashTable[h] = idx; + idx++; + } + + ms->nextToUpdate = target; + return hashTable[ZSTD_hashPtr(ip, hashLog, mls)]; +} + +U32 ZSTD_insertAndFindFirstIndex(ZSTD_matchState_t* ms, const BYTE* ip) { + const ZSTD_compressionParameters* const cParams = &ms->cParams; + return ZSTD_insertAndFindFirstIndex_internal(ms, cParams, ip, ms->cParams.minMatch); +} /* inlining is important to hardwire a hot branch (template emulation) */ FORCE_INLINE_TEMPLATE @@ -661,90 +759,8 @@ size_t ZSTD_HcFindBestMatch_generic ( } if (dictMode == ZSTD_dedicatedDictSearch) { - const U32 ddsLowestIndex = dms->window.dictLimit; - const BYTE* const ddsBase = dms->window.base; - const BYTE* const ddsEnd = dms->window.nextSrc; - const U32 ddsSize = (U32)(ddsEnd - ddsBase); - const U32 ddsIndexDelta = dictLimit - ddsSize; - const U32 bucketSize = (1 << ZSTD_LAZY_DDSS_BUCKET_LOG); - const U32 bucketLimit = nbAttempts < bucketSize - 1 ? nbAttempts : bucketSize - 1; - U32 ddsAttempt; - - for (ddsAttempt = 0; ddsAttempt < bucketSize - 1; ddsAttempt++) { - PREFETCH_L1(ddsBase + dms->hashTable[ddsIdx + ddsAttempt]); - } - - { - U32 const chainPackedPointer = dms->hashTable[ddsIdx + bucketSize - 1]; - U32 const chainIndex = chainPackedPointer >> 8; - - PREFETCH_L1(&dms->chainTable[chainIndex]); - } - - for (ddsAttempt = 0; ddsAttempt < bucketLimit; ddsAttempt++) { - size_t currentMl=0; - const BYTE* match; - matchIndex = dms->hashTable[ddsIdx + ddsAttempt]; - match = ddsBase + matchIndex; - - if (!matchIndex) { - return ml; - } - - /* guaranteed by table construction */ - (void)ddsLowestIndex; - assert(matchIndex >= ddsLowestIndex); - assert(match+4 <= ddsEnd); - if (MEM_read32(match) == MEM_read32(ip)) { - /* assumption : matchIndex <= dictLimit-4 (by table construction) */ - currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, ddsEnd, prefixStart) + 4; - } - - /* save best solution */ - if (currentMl > ml) { - ml = currentMl; - *offsetPtr = curr - (matchIndex + ddsIndexDelta) + ZSTD_REP_MOVE; - if (ip+currentMl == iLimit) { - /* best possible, avoids read overflow on next attempt */ - return ml; - } - } - } - - { - U32 const chainPackedPointer = dms->hashTable[ddsIdx + bucketSize - 1]; - U32 chainIndex = chainPackedPointer >> 8; - U32 const chainLength = chainPackedPointer & 0xFF; - U32 const chainAttempts = nbAttempts - ddsAttempt; - U32 const chainLimit = chainAttempts > chainLength ? chainLength : chainAttempts; - U32 chainAttempt; - - for (chainAttempt = 0 ; chainAttempt < chainLimit; chainAttempt++) { - PREFETCH_L1(ddsBase + dms->chainTable[chainIndex + chainAttempt]); - } - - for (chainAttempt = 0 ; chainAttempt < chainLimit; chainAttempt++, chainIndex++) { - size_t currentMl=0; - const BYTE* match; - matchIndex = dms->chainTable[chainIndex]; - match = ddsBase + matchIndex; - - /* guaranteed by table construction */ - assert(matchIndex >= ddsLowestIndex); - assert(match+4 <= ddsEnd); - if (MEM_read32(match) == MEM_read32(ip)) { - /* assumption : matchIndex <= dictLimit-4 (by table construction) */ - currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, ddsEnd, prefixStart) + 4; - } - - /* save best solution */ - if (currentMl > ml) { - ml = currentMl; - *offsetPtr = curr - (matchIndex + ddsIndexDelta) + ZSTD_REP_MOVE; - if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */ - } - } - } + ml = ZSTD_dedicatedDictSearch_lazy_search(offsetPtr, ml, nbAttempts, dms, + ip, iLimit, prefixStart, curr, dictLimit, ddsIdx); } else if (dictMode == ZSTD_dictMatchState) { const U32* const dmsChainTable = dms->chainTable; const U32 dmsChainSize = (1 << dms->cParams.chainLog); @@ -845,11 +861,657 @@ FORCE_INLINE_TEMPLATE size_t ZSTD_HcFindBestMatch_extDict_selectMLS ( } } +/* ********************************* +* (SIMD) Row-based matchfinder +***********************************/ +/* Constants for row-based hash */ +#define ZSTD_ROW_HASH_TAG_OFFSET 1 /* byte offset of hashes in the match state's tagTable from the beginning of a row */ +#define ZSTD_ROW_HASH_TAG_BITS 8 /* nb bits to use for the tag */ +#define ZSTD_ROW_HASH_TAG_MASK ((1u << ZSTD_ROW_HASH_TAG_BITS) - 1) + +#define ZSTD_ROW_HASH_CACHE_MASK (ZSTD_ROW_HASH_CACHE_SIZE - 1) + +typedef U32 ZSTD_VecMask; /* Clarifies when we are interacting with a U32 representing a mask of matches */ + +#if !defined(ZSTD_NO_INTRINSICS) && defined(__SSE2__) /* SIMD SSE version */ + +#include +typedef __m128i ZSTD_Vec128; + +/* Returns a 128-bit container with 128-bits from src */ +static ZSTD_Vec128 ZSTD_Vec128_read(const void* const src) { + return _mm_loadu_si128((ZSTD_Vec128 const*)src); +} + +/* Returns a ZSTD_Vec128 with the byte "val" packed 16 times */ +static ZSTD_Vec128 ZSTD_Vec128_set8(BYTE val) { + return _mm_set1_epi8((char)val); +} + +/* Do byte-by-byte comparison result of x and y. Then collapse 128-bit resultant mask + * into a 32-bit mask that is the MSB of each byte. + * */ +static ZSTD_VecMask ZSTD_Vec128_cmpMask8(ZSTD_Vec128 x, ZSTD_Vec128 y) { + return (ZSTD_VecMask)_mm_movemask_epi8(_mm_cmpeq_epi8(x, y)); +} + +typedef struct { + __m128i fst; + __m128i snd; +} ZSTD_Vec256; + +static ZSTD_Vec256 ZSTD_Vec256_read(const void* const ptr) { + ZSTD_Vec256 v; + v.fst = ZSTD_Vec128_read(ptr); + v.snd = ZSTD_Vec128_read((ZSTD_Vec128 const*)ptr + 1); + return v; +} + +static ZSTD_Vec256 ZSTD_Vec256_set8(BYTE val) { + ZSTD_Vec256 v; + v.fst = ZSTD_Vec128_set8(val); + v.snd = ZSTD_Vec128_set8(val); + return v; +} + +static ZSTD_VecMask ZSTD_Vec256_cmpMask8(ZSTD_Vec256 x, ZSTD_Vec256 y) { + ZSTD_VecMask fstMask; + ZSTD_VecMask sndMask; + fstMask = ZSTD_Vec128_cmpMask8(x.fst, y.fst); + sndMask = ZSTD_Vec128_cmpMask8(x.snd, y.snd); + return fstMask | (sndMask << 16); +} + +#elif !defined(ZSTD_NO_INTRINSICS) && defined(__ARM_NEON) /* SIMD ARM NEON Version */ + +#include +typedef uint8x16_t ZSTD_Vec128; + +static ZSTD_Vec128 ZSTD_Vec128_read(const void* const src) { + return vld1q_u8((const BYTE* const)src); +} + +static ZSTD_Vec128 ZSTD_Vec128_set8(BYTE val) { + return vdupq_n_u8(val); +} + +/* Mimics '_mm_movemask_epi8()' from SSE */ +static U32 ZSTD_vmovmaskq_u8(ZSTD_Vec128 val) { + /* Shift out everything but the MSB bits in each byte */ + uint16x8_t highBits = vreinterpretq_u16_u8(vshrq_n_u8(val, 7)); + /* Merge the even lanes together with vsra (right shift and add) */ + uint32x4_t paired16 = vreinterpretq_u32_u16(vsraq_n_u16(highBits, highBits, 7)); + uint64x2_t paired32 = vreinterpretq_u64_u32(vsraq_n_u32(paired16, paired16, 14)); + uint8x16_t paired64 = vreinterpretq_u8_u64(vsraq_n_u64(paired32, paired32, 28)); + /* Extract the low 8 bits from each lane, merge */ + return vgetq_lane_u8(paired64, 0) | ((U32)vgetq_lane_u8(paired64, 8) << 8); +} + +static ZSTD_VecMask ZSTD_Vec128_cmpMask8(ZSTD_Vec128 x, ZSTD_Vec128 y) { + return (ZSTD_VecMask)ZSTD_vmovmaskq_u8(vceqq_u8(x, y)); +} + +typedef struct { + uint8x16_t fst; + uint8x16_t snd; +} ZSTD_Vec256; + +static ZSTD_Vec256 ZSTD_Vec256_read(const void* const ptr) { + ZSTD_Vec256 v; + v.fst = ZSTD_Vec128_read(ptr); + v.snd = ZSTD_Vec128_read((ZSTD_Vec128 const*)ptr + 1); + return v; +} + +static ZSTD_Vec256 ZSTD_Vec256_set8(BYTE val) { + ZSTD_Vec256 v; + v.fst = ZSTD_Vec128_set8(val); + v.snd = ZSTD_Vec128_set8(val); + return v; +} + +static ZSTD_VecMask ZSTD_Vec256_cmpMask8(ZSTD_Vec256 x, ZSTD_Vec256 y) { + ZSTD_VecMask fstMask; + ZSTD_VecMask sndMask; + fstMask = ZSTD_Vec128_cmpMask8(x.fst, y.fst); + sndMask = ZSTD_Vec128_cmpMask8(x.snd, y.snd); + return fstMask | (sndMask << 16); +} + +#else /* Scalar fallback version */ + +#define VEC128_NB_SIZE_T (16 / sizeof(size_t)) +typedef struct { + size_t vec[VEC128_NB_SIZE_T]; +} ZSTD_Vec128; + +static ZSTD_Vec128 ZSTD_Vec128_read(const void* const src) { + ZSTD_Vec128 ret; + ZSTD_memcpy(ret.vec, src, VEC128_NB_SIZE_T*sizeof(size_t)); + return ret; +} + +static ZSTD_Vec128 ZSTD_Vec128_set8(BYTE val) { + ZSTD_Vec128 ret = { {0} }; + int startBit = sizeof(size_t) * 8 - 8; + for (;startBit >= 0; startBit -= 8) { + unsigned j = 0; + for (;j < VEC128_NB_SIZE_T; ++j) { + ret.vec[j] |= ((size_t)val << startBit); + } + } + return ret; +} + +/* Compare x to y, byte by byte, generating a "matches" bitfield */ +static ZSTD_VecMask ZSTD_Vec128_cmpMask8(ZSTD_Vec128 x, ZSTD_Vec128 y) { + ZSTD_VecMask res = 0; + unsigned i = 0; + unsigned l = 0; + for (; i < VEC128_NB_SIZE_T; ++i) { + const size_t cmp1 = x.vec[i]; + const size_t cmp2 = y.vec[i]; + unsigned j = 0; + for (; j < sizeof(size_t); ++j, ++l) { + if (((cmp1 >> j*8) & 0xFF) == ((cmp2 >> j*8) & 0xFF)) { + res |= ((U32)1 << (j+i*sizeof(size_t))); + } + } + } + return res; +} + +#define VEC256_NB_SIZE_T 2*VEC128_NB_SIZE_T +typedef struct { + size_t vec[VEC256_NB_SIZE_T]; +} ZSTD_Vec256; + +static ZSTD_Vec256 ZSTD_Vec256_read(const void* const src) { + ZSTD_Vec256 ret; + ZSTD_memcpy(ret.vec, src, VEC256_NB_SIZE_T*sizeof(size_t)); + return ret; +} + +static ZSTD_Vec256 ZSTD_Vec256_set8(BYTE val) { + ZSTD_Vec256 ret = { {0} }; + int startBit = sizeof(size_t) * 8 - 8; + for (;startBit >= 0; startBit -= 8) { + unsigned j = 0; + for (;j < VEC256_NB_SIZE_T; ++j) { + ret.vec[j] |= ((size_t)val << startBit); + } + } + return ret; +} + +/* Compare x to y, byte by byte, generating a "matches" bitfield */ +static ZSTD_VecMask ZSTD_Vec256_cmpMask8(ZSTD_Vec256 x, ZSTD_Vec256 y) { + ZSTD_VecMask res = 0; + unsigned i = 0; + unsigned l = 0; + for (; i < VEC256_NB_SIZE_T; ++i) { + const size_t cmp1 = x.vec[i]; + const size_t cmp2 = y.vec[i]; + unsigned j = 0; + for (; j < sizeof(size_t); ++j, ++l) { + if (((cmp1 >> j*8) & 0xFF) == ((cmp2 >> j*8) & 0xFF)) { + res |= ((U32)1 << (j+i*sizeof(size_t))); + } + } + } + return res; +} + +#endif /* !defined(ZSTD_NO_INTRINSICS) && defined(__SSE2__) */ + +/* ZSTD_VecMask_next(): + * Starting from the LSB, returns the idx of the next non-zero bit. + * Basically counting the nb of trailing zeroes. + */ +static U32 ZSTD_VecMask_next(ZSTD_VecMask val) { +# if defined(_MSC_VER) /* Visual */ + unsigned long r=0; + return _BitScanForward(&r, val) ? (U32)r : 0; +# elif defined(__GNUC__) && (__GNUC__ >= 3) + return (U32)__builtin_ctz(val); +# else + /* Software ctz version: http://graphics.stanford.edu/~seander/bithacks.html#ZerosOnRightMultLookup */ + static const U32 multiplyDeBruijnBitPosition[32] = + { + 0, 1, 28, 2, 29, 14, 24, 3, 30, 22, 20, 15, 25, 17, 4, 8, + 31, 27, 13, 23, 21, 19, 16, 7, 26, 12, 18, 6, 11, 5, 10, 9 + }; + return multiplyDeBruijnBitPosition[((U32)((v & -(int)v) * 0x077CB531U)) >> 27]; +# endif +} + +/* ZSTD_VecMask_rotateRight(): + * Rotates a bitfield to the right by "rotation" bits. + * If the rotation is greater than totalBits, the returned mask is 0. + */ +FORCE_INLINE_TEMPLATE ZSTD_VecMask +ZSTD_VecMask_rotateRight(ZSTD_VecMask mask, U32 const rotation, U32 const totalBits) { + if (rotation == 0) + return mask; + switch (totalBits) { + default: + assert(0); + case 16: + return (mask >> rotation) | (U16)(mask << (16 - rotation)); + case 32: + return (mask >> rotation) | (U32)(mask << (32 - rotation)); + } +} + +/* ZSTD_row_nextIndex(): + * Returns the next index to insert at within a tagTable row, and updates the "head" + * value to reflect the update. Essentially cycles backwards from [0, {entries per row}) + */ +FORCE_INLINE_TEMPLATE U32 ZSTD_row_nextIndex(BYTE* const tagRow, U32 const rowMask) { + U32 const next = (*tagRow - 1) & rowMask; + *tagRow = (BYTE)next; + return next; +} + +/* ZSTD_isAligned(): + * Checks that a pointer is aligned to "align" bytes which must be a power of 2. + */ +MEM_STATIC int ZSTD_isAligned(void const* ptr, size_t align) { + assert((align & (align - 1)) == 0); + return (((size_t)ptr) & (align - 1)) == 0; +} + +/* ZSTD_row_prefetch(): + * Performs prefetching for the hashTable and tagTable at a given row. + */ +FORCE_INLINE_TEMPLATE void ZSTD_row_prefetch(U32 const* hashTable, U16 const* tagTable, U32 const relRow, U32 const rowLog) { + PREFETCH_L1(hashTable + relRow); + if (rowLog == 5) { + PREFETCH_L1(hashTable + relRow + 16); + } + PREFETCH_L1(tagTable + relRow); + assert(rowLog == 4 || rowLog == 5); + assert(ZSTD_isAligned(hashTable + relRow, 64)); /* prefetched hash row always 64-byte aligned */ + assert(ZSTD_isAligned(tagTable + relRow, (size_t)1 << rowLog)); /* prefetched tagRow sits on a multiple of 32 or 64 bytes */ +} + +/* ZSTD_row_fillHashCache(): + * Fill up the hash cache starting at idx, prefetching ZSTD_ROW_HASH_CACHE_SIZE entries. + */ +static void ZSTD_row_fillHashCache(ZSTD_matchState_t* ms, const BYTE* base, + U32 const rowLog, U32 const mls, + U32 idx, const BYTE* const iend) +{ + U32 const* const hashTable = ms->hashTable; + U16 const* const tagTable = ms->tagTable; + U32 const hashLog = ms->rowHashLog; + U32 const maxElemsToPrefetch = (base + idx) >= iend ? 0 : (U32)(iend - (base + idx)); + U32 const lim = idx + MIN(ZSTD_ROW_HASH_CACHE_SIZE, maxElemsToPrefetch); + + for (; idx < lim; ++idx) { + U32 const hash = (U32)ZSTD_hashPtr(base + idx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls); + U32 const row = (hash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog; + ZSTD_row_prefetch(hashTable, tagTable, row, rowLog); + ms->hashCache[idx & ZSTD_ROW_HASH_CACHE_MASK] = hash; + } + + DEBUGLOG(6, "ZSTD_row_fillHashCache(): [%u %u %u %u %u %u %u %u]", ms->hashCache[0], ms->hashCache[1], + ms->hashCache[2], ms->hashCache[3], ms->hashCache[4], + ms->hashCache[5], ms->hashCache[6], ms->hashCache[7]); +} + +/* ZSTD_row_nextCachedHash(): + * Returns the hash of base + idx, and replaces the hash in the hash cache with the byte at + * base + idx + ZSTD_ROW_HASH_CACHE_SIZE. Also prefetches the appropriate rows from hashTable and tagTable. + */ +FORCE_INLINE_TEMPLATE U32 ZSTD_row_nextCachedHash(U32* cache, U32 const* hashTable, + U16 const* tagTable, BYTE const* base, + U32 idx, U32 const hashLog, + U32 const rowLog, U32 const mls) +{ + U32 const newHash = (U32)ZSTD_hashPtr(base+idx+ZSTD_ROW_HASH_CACHE_SIZE, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls); + U32 const row = (newHash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog; + ZSTD_row_prefetch(hashTable, tagTable, row, rowLog); + { U32 const hash = cache[idx & ZSTD_ROW_HASH_CACHE_MASK]; + cache[idx & ZSTD_ROW_HASH_CACHE_MASK] = newHash; + return hash; + } +} + +/* ZSTD_row_update_internal(): + * Inserts the byte at ip into the appropriate position in the hash table. + * Determines the relative row, and the position within the {16, 32} entry row to insert at. + */ +FORCE_INLINE_TEMPLATE void ZSTD_row_update_internal(ZSTD_matchState_t* ms, const BYTE* ip, + U32 const mls, U32 const rowLog, + U32 const rowMask, U32 const useCache) +{ + U32* const hashTable = ms->hashTable; + U16* const tagTable = ms->tagTable; + U32 const hashLog = ms->rowHashLog; + const BYTE* const base = ms->window.base; + const U32 target = (U32)(ip - base); + U32 idx = ms->nextToUpdate; + + DEBUGLOG(6, "ZSTD_row_update_internal(): nextToUpdate=%u, current=%u", idx, target); + for (; idx < target; ++idx) { + U32 const hash = useCache ? ZSTD_row_nextCachedHash(ms->hashCache, hashTable, tagTable, base, idx, hashLog, rowLog, mls) + : (U32)ZSTD_hashPtr(base + idx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls); + U32 const relRow = (hash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog; + U32* const row = hashTable + relRow; + BYTE* tagRow = (BYTE*)(tagTable + relRow); /* Though tagTable is laid out as a table of U16, each tag is only 1 byte. + Explicit cast allows us to get exact desired position within each row */ + U32 const pos = ZSTD_row_nextIndex(tagRow, rowMask); + + assert(hash == ZSTD_hashPtr(base + idx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls)); + ((BYTE*)tagRow)[pos + ZSTD_ROW_HASH_TAG_OFFSET] = hash & ZSTD_ROW_HASH_TAG_MASK; + row[pos] = idx; + } + ms->nextToUpdate = target; +} + +/* ZSTD_row_update(): + * External wrapper for ZSTD_row_update_internal(). Used for filling the hashtable during dictionary + * processing. + */ +void ZSTD_row_update(ZSTD_matchState_t* const ms, const BYTE* ip) { + const U32 rowLog = ms->cParams.searchLog < 5 ? 4 : 5; + const U32 rowMask = (1u << rowLog) - 1; + const U32 mls = MIN(ms->cParams.minMatch, 6 /* mls caps out at 6 */); + + DEBUGLOG(5, "ZSTD_row_update(), rowLog=%u", rowLog); + ZSTD_row_update_internal(ms, ip, mls, rowLog, rowMask, 0 /* dont use cache */); +} + +/* Returns a ZSTD_VecMask (U32) that has the nth bit set to 1 if the newly-computed "tag" matches + * the hash at the nth position in a row of the tagTable. + */ +FORCE_INLINE_TEMPLATE +ZSTD_VecMask ZSTD_row_getMatchMask(const BYTE* const tagRow, const BYTE tag, const U32 head, const U32 rowEntries) { + ZSTD_VecMask matches = 0; + if (rowEntries == 16) { + ZSTD_Vec128 hashes = ZSTD_Vec128_read(tagRow + ZSTD_ROW_HASH_TAG_OFFSET); + ZSTD_Vec128 expandedTags = ZSTD_Vec128_set8(tag); + matches = ZSTD_Vec128_cmpMask8(hashes, expandedTags); + } else if (rowEntries == 32) { + ZSTD_Vec256 hashes = ZSTD_Vec256_read(tagRow + ZSTD_ROW_HASH_TAG_OFFSET); + ZSTD_Vec256 expandedTags = ZSTD_Vec256_set8(tag); + matches = ZSTD_Vec256_cmpMask8(hashes, expandedTags); + } else { + assert(0); + } + /* Each row is a circular buffer beginning at the value of "head". So we must rotate the "matches" bitfield + to match up with the actual layout of the entries within the hashTable */ + return ZSTD_VecMask_rotateRight(matches, head, rowEntries); +} + +/* The high-level approach of the SIMD row based match finder is as follows: + * - Figure out where to insert the new entry: + * - Generate a hash from a byte along with an additional 1-byte "short hash". The additional byte is our "tag" + * - The hashTable is effectively split into groups or "rows" of 16 or 32 entries of U32, and the hash determines + * which row to insert into. + * - Determine the correct position within the row to insert the entry into. Each row of 16 or 32 can + * be considered as a circular buffer with a "head" index that resides in the tagTable. + * - Also insert the "tag" into the equivalent row and position in the tagTable. + * - Note: The tagTable has 17 or 33 1-byte entries per row, due to 16 or 32 tags, and 1 "head" entry. + * The 17 or 33 entry rows are spaced out to occur every 32 or 64 bytes, respectively, + * for alignment/performance reasons, leaving some bytes unused. + * - Use SIMD to efficiently compare the tags in the tagTable to the 1-byte "short hash" and + * generate a bitfield that we can cycle through to check the collisions in the hash table. + * - Pick the longest match. + */ +FORCE_INLINE_TEMPLATE +size_t ZSTD_RowFindBestMatch_generic ( + ZSTD_matchState_t* ms, + const BYTE* const ip, const BYTE* const iLimit, + size_t* offsetPtr, + const U32 mls, const ZSTD_dictMode_e dictMode, + const U32 rowLog) +{ + U32* const hashTable = ms->hashTable; + U16* const tagTable = ms->tagTable; + U32* const hashCache = ms->hashCache; + const U32 hashLog = ms->rowHashLog; + const ZSTD_compressionParameters* const cParams = &ms->cParams; + const BYTE* const base = ms->window.base; + const BYTE* const dictBase = ms->window.dictBase; + const U32 dictLimit = ms->window.dictLimit; + const BYTE* const prefixStart = base + dictLimit; + const BYTE* const dictEnd = dictBase + dictLimit; + const U32 curr = (U32)(ip-base); + const U32 maxDistance = 1U << cParams->windowLog; + const U32 lowestValid = ms->window.lowLimit; + const U32 withinMaxDistance = (curr - lowestValid > maxDistance) ? curr - maxDistance : lowestValid; + const U32 isDictionary = (ms->loadedDictEnd != 0); + const U32 lowLimit = isDictionary ? lowestValid : withinMaxDistance; + const U32 rowEntries = (1U << rowLog); + const U32 rowMask = rowEntries - 1; + const U32 cappedSearchLog = MIN(cParams->searchLog, rowLog); /* nb of searches is capped at nb entries per row */ + U32 nbAttempts = 1U << cappedSearchLog; + size_t ml=4-1; + + /* DMS/DDS variables that may be referenced laster */ + const ZSTD_matchState_t* const dms = ms->dictMatchState; + size_t ddsIdx; + U32 ddsExtraAttempts; /* cctx hash tables are limited in searches, but allow extra searches into DDS */ + U32 dmsTag; + U32* dmsRow; + BYTE* dmsTagRow; + + if (dictMode == ZSTD_dedicatedDictSearch) { + const U32 ddsHashLog = dms->cParams.hashLog - ZSTD_LAZY_DDSS_BUCKET_LOG; + { /* Prefetch DDS hashtable entry */ + ddsIdx = ZSTD_hashPtr(ip, ddsHashLog, mls) << ZSTD_LAZY_DDSS_BUCKET_LOG; + PREFETCH_L1(&dms->hashTable[ddsIdx]); + } + ddsExtraAttempts = cParams->searchLog > rowLog ? 1U << (cParams->searchLog - 5) : 0; + } + + if (dictMode == ZSTD_dictMatchState) { + /* Prefetch DMS rows */ + U32* const dmsHashTable = dms->hashTable; + U16* const dmsTagTable = dms->tagTable; + U32 const dmsHash = (U32)ZSTD_hashPtr(ip, dms->rowHashLog + ZSTD_ROW_HASH_TAG_BITS, mls); + U32 const dmsRelRow = (dmsHash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog; + dmsTag = dmsHash & ZSTD_ROW_HASH_TAG_MASK; + dmsTagRow = (BYTE*)(dmsTagTable + dmsRelRow); + dmsRow = dmsHashTable + dmsRelRow; + ZSTD_row_prefetch(dmsHashTable, dmsTagTable, dmsRelRow, rowLog); + } + + /* Update the hashTable and tagTable up to (but not including) ip */ + ZSTD_row_update_internal(ms, ip, mls, rowLog, rowMask, 1 /* useCache */); + { /* Get the hash for ip, compute the appropriate row */ + U32 const hash = ZSTD_row_nextCachedHash(hashCache, hashTable, tagTable, base, curr, hashLog, rowLog, mls); + U32 const relRow = (hash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog; + U32 const tag = hash & ZSTD_ROW_HASH_TAG_MASK; + U32* const row = hashTable + relRow; + BYTE* tagRow = (BYTE*)(tagTable + relRow); + U32 const head = *tagRow & rowMask; + U32 matchBuffer[32 /* maximum nb entries per row */]; + size_t numMatches = 0; + size_t currMatch = 0; + ZSTD_VecMask matches = ZSTD_row_getMatchMask(tagRow, (BYTE)tag, head, rowEntries); + + /* Cycle through the matches and prefetch */ + for (; (matches > 0) && (nbAttempts > 0); --nbAttempts, matches &= (matches - 1)) { + U32 const matchPos = (head + ZSTD_VecMask_next(matches)) & rowMask; + U32 const matchIndex = row[matchPos]; + assert(numMatches < rowEntries); + if (matchIndex < lowLimit) + break; + if ((dictMode != ZSTD_extDict) || matchIndex >= dictLimit) { + PREFETCH_L1(base + matchIndex); + } else { + PREFETCH_L1(dictBase + matchIndex); + } + matchBuffer[numMatches++] = matchIndex; + } + + /* Speed opt: insert current byte into hashtable too. This allows us to avoid one iteration of the loop + in ZSTD_row_update_internal() at the next search. */ + { + U32 const pos = ZSTD_row_nextIndex(tagRow, rowMask); + tagRow[pos + ZSTD_ROW_HASH_TAG_OFFSET] = (BYTE)tag; + row[pos] = ms->nextToUpdate++; + } + + /* Return the longest match */ + for (; currMatch < numMatches; ++currMatch) { + U32 const matchIndex = matchBuffer[currMatch]; + size_t currentMl=0; + assert(matchIndex < curr); + assert(matchIndex >= lowLimit); + + if ((dictMode != ZSTD_extDict) || matchIndex >= dictLimit) { + const BYTE* const match = base + matchIndex; + assert(matchIndex >= dictLimit); /* ensures this is true if dictMode != ZSTD_extDict */ + if (match[ml] == ip[ml]) /* potentially better */ + currentMl = ZSTD_count(ip, match, iLimit); + } else { + const BYTE* const match = dictBase + matchIndex; + assert(match+4 <= dictEnd); + if (MEM_read32(match) == MEM_read32(ip)) /* assumption : matchIndex <= dictLimit-4 (by table construction) */ + currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, dictEnd, prefixStart) + 4; + } + + /* Save best solution */ + if (currentMl > ml) { + ml = currentMl; + *offsetPtr = curr - matchIndex + ZSTD_REP_MOVE; + if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */ + } + } + } + + if (dictMode == ZSTD_dedicatedDictSearch) { + ml = ZSTD_dedicatedDictSearch_lazy_search(offsetPtr, ml, nbAttempts + ddsExtraAttempts, dms, + ip, iLimit, prefixStart, curr, dictLimit, ddsIdx); + } else if (dictMode == ZSTD_dictMatchState) { + /* TODO: Measure and potentially add prefetching to DMS */ + const U32 dmsLowestIndex = dms->window.dictLimit; + const BYTE* const dmsBase = dms->window.base; + const BYTE* const dmsEnd = dms->window.nextSrc; + const U32 dmsSize = (U32)(dmsEnd - dmsBase); + const U32 dmsIndexDelta = dictLimit - dmsSize; + + { /* Get the hash for ip, compute the appropriate row */ + U32 const head = *dmsTagRow & rowMask; + U32 matchBuffer[32 /* maximum nb row entries */]; + size_t numMatches = 0; + size_t currMatch = 0; + ZSTD_VecMask matches = ZSTD_row_getMatchMask(dmsTagRow, (BYTE)dmsTag, head, rowEntries); + + for (; (matches > 0) && (nbAttempts > 0); --nbAttempts, matches &= (matches - 1)) { + U32 const matchPos = (head + ZSTD_VecMask_next(matches)) & rowMask; + U32 const matchIndex = dmsRow[matchPos]; + if (matchIndex < dmsLowestIndex) + break; + PREFETCH_L1(dmsBase + matchIndex); + matchBuffer[numMatches++] = matchIndex; + } + + /* Return the longest match */ + for (; currMatch < numMatches; ++currMatch) { + U32 const matchIndex = matchBuffer[currMatch]; + size_t currentMl=0; + assert(matchIndex >= dmsLowestIndex); + assert(matchIndex < curr); + + { const BYTE* const match = dmsBase + matchIndex; + assert(match+4 <= dmsEnd); + if (MEM_read32(match) == MEM_read32(ip)) + currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, dmsEnd, prefixStart) + 4; + } + + if (currentMl > ml) { + ml = currentMl; + *offsetPtr = curr - (matchIndex + dmsIndexDelta) + ZSTD_REP_MOVE; + if (ip+currentMl == iLimit) break; + } + } + } + } + return ml; +} + +/* Inlining is important to hardwire a hot branch (template emulation) */ +FORCE_INLINE_TEMPLATE size_t ZSTD_RowFindBestMatch_selectMLS ( + ZSTD_matchState_t* ms, + const BYTE* ip, const BYTE* const iLimit, + const ZSTD_dictMode_e dictMode, size_t* offsetPtr, const U32 rowLog) +{ + switch(ms->cParams.minMatch) + { + default : /* includes case 3 */ + case 4 : return ZSTD_RowFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 4, dictMode, rowLog); + case 5 : return ZSTD_RowFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 5, dictMode, rowLog); + case 7 : + case 6 : return ZSTD_RowFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 6, dictMode, rowLog); + } +} + +FORCE_INLINE_TEMPLATE size_t ZSTD_RowFindBestMatch_selectRowLog ( + ZSTD_matchState_t* ms, + const BYTE* ip, const BYTE* const iLimit, + size_t* offsetPtr) +{ + const U32 cappedSearchLog = MIN(ms->cParams.searchLog, 5); + switch(cappedSearchLog) + { + default : + case 4 : return ZSTD_RowFindBestMatch_selectMLS(ms, ip, iLimit, ZSTD_noDict, offsetPtr, 4); + case 5 : return ZSTD_RowFindBestMatch_selectMLS(ms, ip, iLimit, ZSTD_noDict, offsetPtr, 5); + } +} + +FORCE_INLINE_TEMPLATE size_t ZSTD_RowFindBestMatch_dictMatchState_selectRowLog( + ZSTD_matchState_t* ms, + const BYTE* ip, const BYTE* const iLimit, + size_t* offsetPtr) +{ + const U32 cappedSearchLog = MIN(ms->cParams.searchLog, 5); + switch(cappedSearchLog) + { + default : + case 4 : return ZSTD_RowFindBestMatch_selectMLS(ms, ip, iLimit, ZSTD_dictMatchState, offsetPtr, 4); + case 5 : return ZSTD_RowFindBestMatch_selectMLS(ms, ip, iLimit, ZSTD_dictMatchState, offsetPtr, 5); + } +} + +FORCE_INLINE_TEMPLATE size_t ZSTD_RowFindBestMatch_dedicatedDictSearch_selectRowLog( + ZSTD_matchState_t* ms, + const BYTE* ip, const BYTE* const iLimit, + size_t* offsetPtr) +{ + const U32 cappedSearchLog = MIN(ms->cParams.searchLog, 5); + switch(cappedSearchLog) + { + default : + case 4 : return ZSTD_RowFindBestMatch_selectMLS(ms, ip, iLimit, ZSTD_dedicatedDictSearch, offsetPtr, 4); + case 5 : return ZSTD_RowFindBestMatch_selectMLS(ms, ip, iLimit, ZSTD_dedicatedDictSearch, offsetPtr, 5); + } +} + +FORCE_INLINE_TEMPLATE size_t ZSTD_RowFindBestMatch_extDict_selectRowLog ( + ZSTD_matchState_t* ms, + const BYTE* ip, const BYTE* const iLimit, + size_t* offsetPtr) +{ + const U32 cappedSearchLog = MIN(ms->cParams.searchLog, 5); + switch(cappedSearchLog) + { + default : + case 4 : return ZSTD_RowFindBestMatch_selectMLS(ms, ip, iLimit, ZSTD_extDict, offsetPtr, 4); + case 5 : return ZSTD_RowFindBestMatch_selectMLS(ms, ip, iLimit, ZSTD_extDict, offsetPtr, 5); + } +} + /* ******************************* * Common parser - lazy strategy *********************************/ -typedef enum { search_hashChain, search_binaryTree } searchMethod_e; +typedef enum { search_hashChain=0, search_binaryTree=1, search_rowHash=2 } searchMethod_e; FORCE_INLINE_TEMPLATE size_t ZSTD_compressBlock_lazy_generic( @@ -863,10 +1525,11 @@ ZSTD_compressBlock_lazy_generic( const BYTE* ip = istart; const BYTE* anchor = istart; const BYTE* const iend = istart + srcSize; - const BYTE* const ilimit = iend - 8; + const BYTE* const ilimit = searchMethod == search_rowHash ? iend - 16 : iend - 8; const BYTE* const base = ms->window.base; const U32 prefixLowestIndex = ms->window.dictLimit; const BYTE* const prefixLowest = base + prefixLowestIndex; + const U32 rowLog = ms->cParams.searchLog < 5 ? 4 : 5; typedef size_t (*searchMax_f)( ZSTD_matchState_t* ms, @@ -878,26 +1541,30 @@ ZSTD_compressBlock_lazy_generic( * that should never occur (extDict modes go to the other implementation * below and there is no DDSS for binary tree search yet). */ - const searchMax_f searchFuncs[4][2] = { + const searchMax_f searchFuncs[4][3] = { { ZSTD_HcFindBestMatch_selectMLS, - ZSTD_BtFindBestMatch_selectMLS + ZSTD_BtFindBestMatch_selectMLS, + ZSTD_RowFindBestMatch_selectRowLog }, { + NULL, NULL, NULL }, { ZSTD_HcFindBestMatch_dictMatchState_selectMLS, - ZSTD_BtFindBestMatch_dictMatchState_selectMLS + ZSTD_BtFindBestMatch_dictMatchState_selectMLS, + ZSTD_RowFindBestMatch_dictMatchState_selectRowLog }, { ZSTD_HcFindBestMatch_dedicatedDictSearch_selectMLS, - NULL + NULL, + ZSTD_RowFindBestMatch_dedicatedDictSearch_selectRowLog } }; - searchMax_f const searchMax = searchFuncs[dictMode][searchMethod == search_binaryTree]; + searchMax_f const searchMax = searchFuncs[dictMode][(int)searchMethod]; U32 offset_1 = rep[0], offset_2 = rep[1], savedOffset=0; const int isDMS = dictMode == ZSTD_dictMatchState; @@ -915,9 +1582,7 @@ ZSTD_compressBlock_lazy_generic( assert(searchMax != NULL); - DEBUGLOG(5, "ZSTD_compressBlock_lazy_generic (dictMode=%u)", (U32)dictMode); - - /* init */ + DEBUGLOG(5, "ZSTD_compressBlock_lazy_generic (dictMode=%u) (searchFunc=%u)", (U32)dictMode, (U32)searchMethod); ip += (dictAndPrefixLength == 0); if (dictMode == ZSTD_noDict) { U32 const curr = (U32)(ip - base); @@ -933,6 +1598,12 @@ ZSTD_compressBlock_lazy_generic( assert(offset_2 <= dictAndPrefixLength); } + if (searchMethod == search_rowHash) { + ZSTD_row_fillHashCache(ms, base, rowLog, + MIN(ms->cParams.minMatch, 6 /* mls caps out at 6 */), + ms->nextToUpdate, ilimit); + } + /* Match Loop */ #if defined(__GNUC__) && defined(__x86_64__) /* I've measured random a 5% speed loss on levels 5 & 6 (greedy) when the @@ -1198,6 +1869,70 @@ size_t ZSTD_compressBlock_greedy_dedicatedDictSearch( return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0, ZSTD_dedicatedDictSearch); } +/* Row-based matchfinder */ +size_t ZSTD_compressBlock_lazy2_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize) +{ + return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2, ZSTD_noDict); +} + +size_t ZSTD_compressBlock_lazy_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize) +{ + return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1, ZSTD_noDict); +} + +size_t ZSTD_compressBlock_greedy_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize) +{ + return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0, ZSTD_noDict); +} + +size_t ZSTD_compressBlock_lazy2_dictMatchState_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize) +{ + return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2, ZSTD_dictMatchState); +} + +size_t ZSTD_compressBlock_lazy_dictMatchState_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize) +{ + return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1, ZSTD_dictMatchState); +} + +size_t ZSTD_compressBlock_greedy_dictMatchState_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize) +{ + return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0, ZSTD_dictMatchState); +} + + +size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize) +{ + return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2, ZSTD_dedicatedDictSearch); +} + +size_t ZSTD_compressBlock_lazy_dedicatedDictSearch_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize) +{ + return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1, ZSTD_dedicatedDictSearch); +} + +size_t ZSTD_compressBlock_greedy_dedicatedDictSearch_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize) +{ + return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0, ZSTD_dedicatedDictSearch); +} FORCE_INLINE_TEMPLATE size_t ZSTD_compressBlock_lazy_extDict_generic( @@ -1210,7 +1945,7 @@ size_t ZSTD_compressBlock_lazy_extDict_generic( const BYTE* ip = istart; const BYTE* anchor = istart; const BYTE* const iend = istart + srcSize; - const BYTE* const ilimit = iend - 8; + const BYTE* const ilimit = searchMethod == search_rowHash ? iend - 16 : iend - 8; const BYTE* const base = ms->window.base; const U32 dictLimit = ms->window.dictLimit; const BYTE* const prefixStart = base + dictLimit; @@ -1218,18 +1953,28 @@ size_t ZSTD_compressBlock_lazy_extDict_generic( const BYTE* const dictEnd = dictBase + dictLimit; const BYTE* const dictStart = dictBase + ms->window.lowLimit; const U32 windowLog = ms->cParams.windowLog; + const U32 rowLog = ms->cParams.searchLog < 5 ? 4 : 5; typedef size_t (*searchMax_f)( ZSTD_matchState_t* ms, const BYTE* ip, const BYTE* iLimit, size_t* offsetPtr); - searchMax_f searchMax = searchMethod==search_binaryTree ? ZSTD_BtFindBestMatch_extDict_selectMLS : ZSTD_HcFindBestMatch_extDict_selectMLS; - + const searchMax_f searchFuncs[3] = { + ZSTD_HcFindBestMatch_extDict_selectMLS, + ZSTD_BtFindBestMatch_extDict_selectMLS, + ZSTD_RowFindBestMatch_extDict_selectRowLog + }; + searchMax_f searchMax = searchFuncs[(int)searchMethod]; U32 offset_1 = rep[0], offset_2 = rep[1]; - DEBUGLOG(5, "ZSTD_compressBlock_lazy_extDict_generic"); + DEBUGLOG(5, "ZSTD_compressBlock_lazy_extDict_generic (searchFunc=%u)", (U32)searchMethod); /* init */ ip += (ip == prefixStart); + if (searchMethod == search_rowHash) { + ZSTD_row_fillHashCache(ms, base, rowLog, + MIN(ms->cParams.minMatch, 6 /* mls caps out at 6 */), + ms->nextToUpdate, ilimit); + } /* Match Loop */ #if defined(__GNUC__) && defined(__x86_64__) @@ -1410,3 +2155,26 @@ size_t ZSTD_compressBlock_btlazy2_extDict( { return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_binaryTree, 2); } + +size_t ZSTD_compressBlock_greedy_extDict_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize) +{ + return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0); +} + +size_t ZSTD_compressBlock_lazy_extDict_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize) + +{ + return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1); +} + +size_t ZSTD_compressBlock_lazy2_extDict_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize) + +{ + return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2); +} diff --git a/lib/compress/zstd_lazy.h b/lib/compress/zstd_lazy.h index b75f7e8ae..150f7b390 100644 --- a/lib/compress/zstd_lazy.h +++ b/lib/compress/zstd_lazy.h @@ -26,6 +26,7 @@ extern "C" { #define ZSTD_LAZY_DDSS_BUCKET_LOG 2 U32 ZSTD_insertAndFindFirstIndex(ZSTD_matchState_t* ms, const BYTE* ip); +void ZSTD_row_update(ZSTD_matchState_t* const ms, const BYTE* ip); void ZSTD_dedicatedDictSearch_lazy_loadDictionary(ZSTD_matchState_t* ms, const BYTE* const ip); @@ -43,6 +44,15 @@ size_t ZSTD_compressBlock_lazy( size_t ZSTD_compressBlock_greedy( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], void const* src, size_t srcSize); +size_t ZSTD_compressBlock_lazy2_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize); +size_t ZSTD_compressBlock_lazy_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize); +size_t ZSTD_compressBlock_greedy_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize); size_t ZSTD_compressBlock_btlazy2_dictMatchState( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], @@ -56,6 +66,15 @@ size_t ZSTD_compressBlock_lazy_dictMatchState( size_t ZSTD_compressBlock_greedy_dictMatchState( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], void const* src, size_t srcSize); +size_t ZSTD_compressBlock_lazy2_dictMatchState_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize); +size_t ZSTD_compressBlock_lazy_dictMatchState_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize); +size_t ZSTD_compressBlock_greedy_dictMatchState_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize); size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], @@ -66,6 +85,15 @@ size_t ZSTD_compressBlock_lazy_dedicatedDictSearch( size_t ZSTD_compressBlock_greedy_dedicatedDictSearch( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], void const* src, size_t srcSize); +size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize); +size_t ZSTD_compressBlock_lazy_dedicatedDictSearch_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize); +size_t ZSTD_compressBlock_greedy_dedicatedDictSearch_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize); size_t ZSTD_compressBlock_greedy_extDict( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], @@ -76,9 +104,19 @@ size_t ZSTD_compressBlock_lazy_extDict( size_t ZSTD_compressBlock_lazy2_extDict( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], void const* src, size_t srcSize); +size_t ZSTD_compressBlock_greedy_extDict_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize); +size_t ZSTD_compressBlock_lazy_extDict_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize); +size_t ZSTD_compressBlock_lazy2_extDict_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize); size_t ZSTD_compressBlock_btlazy2_extDict( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], void const* src, size_t srcSize); + #if defined (__cplusplus) } diff --git a/lib/compress/zstd_ldm.c b/lib/compress/zstd_ldm.c index 66dbc2875..7dea97aa8 100644 --- a/lib/compress/zstd_ldm.c +++ b/lib/compress/zstd_ldm.c @@ -622,12 +622,13 @@ void ZSTD_ldm_skipRawSeqStoreBytes(rawSeqStore_t* rawSeqStore, size_t nbBytes) { size_t ZSTD_ldm_blockCompress(rawSeqStore_t* rawSeqStore, ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + ZSTD_useRowMatchFinderMode_e useRowMatchFinder, void const* src, size_t srcSize) { const ZSTD_compressionParameters* const cParams = &ms->cParams; unsigned const minMatch = cParams->minMatch; ZSTD_blockCompressor const blockCompressor = - ZSTD_selectBlockCompressor(cParams->strategy, ZSTD_matchState_dictMode(ms)); + ZSTD_selectBlockCompressor(cParams->strategy, useRowMatchFinder, ZSTD_matchState_dictMode(ms)); /* Input bounds */ BYTE const* const istart = (BYTE const*)src; BYTE const* const iend = istart + srcSize; diff --git a/lib/compress/zstd_ldm.h b/lib/compress/zstd_ldm.h index 7e69cc9b4..393466fa9 100644 --- a/lib/compress/zstd_ldm.h +++ b/lib/compress/zstd_ldm.h @@ -66,6 +66,7 @@ size_t ZSTD_ldm_generateSequences( */ size_t ZSTD_ldm_blockCompress(rawSeqStore_t* rawSeqStore, ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + ZSTD_useRowMatchFinderMode_e useRowMatchFinder, void const* src, size_t srcSize); /** diff --git a/lib/zstd.h b/lib/zstd.h index 3a4a755e6..b95959a34 100644 --- a/lib/zstd.h +++ b/lib/zstd.h @@ -420,6 +420,7 @@ typedef enum { * ZSTD_c_blockDelimiters * ZSTD_c_validateSequences * ZSTD_c_splitBlocks + * ZSTD_c_useRowMatchFinder * Because they are not stable, it's necessary to define ZSTD_STATIC_LINKING_ONLY to access them. * note : never ever use experimentalParam? names directly; * also, the enums values themselves are unstable and can still change. @@ -436,7 +437,8 @@ typedef enum { ZSTD_c_experimentalParam10=1007, ZSTD_c_experimentalParam11=1008, ZSTD_c_experimentalParam12=1009, - ZSTD_c_experimentalParam13=1010 + ZSTD_c_experimentalParam13=1010, + ZSTD_c_experimentalParam14=1011 } ZSTD_cParameter; typedef struct { @@ -1272,6 +1274,11 @@ typedef enum { ZSTD_lcm_uncompressed = 2 /**< Always emit uncompressed literals. */ } ZSTD_literalCompressionMode_e; +typedef enum { + ZSTD_urm_auto = 0, /* Automatically determine whether or not we use row matchfinder */ + ZSTD_urm_disableRowMatchFinder = 1, /* Never use row matchfinder */ + ZSTD_urm_enableRowMatchFinder = 2 /* Always use row matchfinder when applicable */ +} ZSTD_useRowMatchFinderMode_e; /*************************************** * Frame size functions @@ -1843,6 +1850,19 @@ ZSTDLIB_API size_t ZSTD_CCtx_refPrefix_advanced(ZSTD_CCtx* cctx, const void* pre */ #define ZSTD_c_splitBlocks ZSTD_c_experimentalParam13 +/* ZSTD_c_useRowMatchFinder + * Default is ZSTD_urm_auto. + * Controlled with ZSTD_useRowMatchFinderMode_e enum. + * + * By default, in ZSTD_urm_auto, when finalizing the compression parameters, the library + * will decide at runtime whether to use the row-based matchfinder based on support for SIMD + * instructions as well as the windowLog. + * + * Set to ZSTD_urm_disableRowMatchFinder to never use row-based matchfinder. + * Set to ZSTD_urm_enableRowMatchFinder to force usage of row-based matchfinder. + */ +#define ZSTD_c_useRowMatchFinder ZSTD_c_experimentalParam14 + /*! ZSTD_CCtx_getParameter() : * Get the requested compression parameter value, selected by enum ZSTD_cParameter, * and store it into int* value. diff --git a/programs/benchzstd.c b/programs/benchzstd.c index cd7902fbe..c40f0a24c 100644 --- a/programs/benchzstd.c +++ b/programs/benchzstd.c @@ -137,7 +137,8 @@ BMK_advancedParams_t BMK_initAdvancedParams(void) { 0, /* ldmHashLog */ 0, /* ldmBuckSizeLog */ 0, /* ldmHashRateLog */ - ZSTD_lcm_auto /* literalCompressionMode */ + ZSTD_lcm_auto, /* literalCompressionMode */ + 0 /* useRowMatchFinder */ }; return res; } @@ -175,6 +176,7 @@ BMK_initCCtx(ZSTD_CCtx* ctx, CHECK_Z(ZSTD_CCtx_setParameter(ctx, ZSTD_c_nbWorkers, adv->nbWorkers)); } CHECK_Z(ZSTD_CCtx_setParameter(ctx, ZSTD_c_compressionLevel, cLevel)); + CHECK_Z(ZSTD_CCtx_setParameter(ctx, ZSTD_c_useRowMatchFinder, adv->useRowMatchFinder)); CHECK_Z(ZSTD_CCtx_setParameter(ctx, ZSTD_c_enableLongDistanceMatching, adv->ldmFlag)); CHECK_Z(ZSTD_CCtx_setParameter(ctx, ZSTD_c_ldmMinMatch, adv->ldmMinMatch)); CHECK_Z(ZSTD_CCtx_setParameter(ctx, ZSTD_c_ldmHashLog, adv->ldmHashLog)); diff --git a/programs/benchzstd.h b/programs/benchzstd.h index bdeecae7a..9b40dcc29 100644 --- a/programs/benchzstd.h +++ b/programs/benchzstd.h @@ -117,6 +117,7 @@ typedef struct { int ldmBucketSizeLog; int ldmHashRateLog; ZSTD_literalCompressionMode_e literalCompressionMode; + int useRowMatchFinder; /* use row-based matchfinder if possible */ } BMK_advancedParams_t; /* returns default parameters used by nonAdvanced functions */ diff --git a/programs/fileio.c b/programs/fileio.c index 66b86f710..fab7918d7 100644 --- a/programs/fileio.c +++ b/programs/fileio.c @@ -298,6 +298,7 @@ struct FIO_prefs_s { int blockSize; int overlapLog; U32 adaptiveMode; + U32 useRowMatchFinder; int rsyncable; int minAdaptLevel; int maxAdaptLevel; @@ -468,6 +469,10 @@ void FIO_setAdaptiveMode(FIO_prefs_t* const prefs, unsigned adapt) { prefs->adaptiveMode = adapt; } +void FIO_setUseRowMatchFinder(FIO_prefs_t* const prefs, int useRowMatchFinder) { + prefs->useRowMatchFinder = useRowMatchFinder; +} + void FIO_setRsyncable(FIO_prefs_t* const prefs, int rsyncable) { if ((rsyncable>0) && (prefs->nbWorkers==0)) EXM_THROW(1, "Rsyncable mode is not compatible with single thread mode \n"); @@ -986,6 +991,7 @@ static cRess_t FIO_createCResources(FIO_prefs_t* const prefs, if (prefs->ldmHashRateLog != FIO_LDM_PARAM_NOTSET) { CHECK( ZSTD_CCtx_setParameter(ress.cctx, ZSTD_c_ldmHashRateLog, prefs->ldmHashRateLog) ); } + CHECK( ZSTD_CCtx_setParameter(ress.cctx, ZSTD_c_useRowMatchFinder, prefs->useRowMatchFinder)); /* compression parameters */ CHECK( ZSTD_CCtx_setParameter(ress.cctx, ZSTD_c_windowLog, (int)comprParams.windowLog) ); CHECK( ZSTD_CCtx_setParameter(ress.cctx, ZSTD_c_chainLog, (int)comprParams.chainLog) ); diff --git a/programs/fileio.h b/programs/fileio.h index 252df2f26..1f5db4b52 100644 --- a/programs/fileio.h +++ b/programs/fileio.h @@ -77,6 +77,7 @@ void FIO_overwriteMode(FIO_prefs_t* const prefs); void FIO_setAdaptiveMode(FIO_prefs_t* const prefs, unsigned adapt); void FIO_setAdaptMin(FIO_prefs_t* const prefs, int minCLevel); void FIO_setAdaptMax(FIO_prefs_t* const prefs, int maxCLevel); +void FIO_setUseRowMatchFinder(FIO_prefs_t* const prefs, int useRowMatchFinder); void FIO_setBlockSize(FIO_prefs_t* const prefs, int blockSize); void FIO_setChecksumFlag(FIO_prefs_t* const prefs, int checksumFlag); void FIO_setDictIDFlag(FIO_prefs_t* const prefs, int dictIDFlag); diff --git a/programs/zstdcli.c b/programs/zstdcli.c index cda2d9578..d9d2c701e 100644 --- a/programs/zstdcli.c +++ b/programs/zstdcli.c @@ -205,6 +205,7 @@ static void usage_advanced(const char* programName) DISPLAYOUT( "--long[=#]: enable long distance matching with given window log (default: %u) \n", g_defaultMaxWindowLog); DISPLAYOUT( "--fast[=#]: switch to very fast compression levels (default: %u) \n", 1); DISPLAYOUT( "--adapt : dynamically adapt compression level to I/O conditions \n"); + DISPLAYOUT( "--[no-]row-match-finder : force enable/disable usage of fast row-based matchfinder for greedy, lazy, and lazy2 strategies \n"); # ifdef ZSTD_MULTITHREAD DISPLAYOUT( " -T# : spawns # compression threads (default: 1, 0==# cores) \n"); DISPLAYOUT( " -B# : select size of each job (default: 0==automatic) \n"); @@ -730,6 +731,7 @@ int main(int const argCount, const char* argv[]) main_pause = 0, nbWorkers = 0, adapt = 0, + useRowMatchFinder = 0, adaptMin = MINCLEVEL, adaptMax = MAXCLEVEL, rsyncable = 0, @@ -857,6 +859,8 @@ int main(int const argCount, const char* argv[]) if (!strcmp(argument, "--content-size")) { contentSize = 1; continue; } if (!strcmp(argument, "--no-content-size")) { contentSize = 0; continue; } if (!strcmp(argument, "--adapt")) { adapt = 1; continue; } + if (!strcmp(argument, "--no-row-match-finder")) { useRowMatchFinder = 1; continue; } + if (!strcmp(argument, "--row-match-finder")) { useRowMatchFinder = 2; continue; } if (longCommandWArg(&argument, "--adapt=")) { adapt = 1; if (!parseAdaptParameters(argument, &adaptMin, &adaptMax)) { badusage(programName); CLEAN_RETURN(1); } continue; } if (!strcmp(argument, "--single-thread")) { nbWorkers = 0; singleThread = 1; continue; } if (!strcmp(argument, "--format=zstd")) { suffix = ZSTD_EXTENSION; FIO_setCompressionType(prefs, FIO_zstdCompression); continue; } @@ -1196,6 +1200,7 @@ int main(int const argCount, const char* argv[]) benchParams.ldmFlag = ldmFlag; benchParams.ldmMinMatch = (int)g_ldmMinMatch; benchParams.ldmHashLog = (int)g_ldmHashLog; + benchParams.useRowMatchFinder = useRowMatchFinder; if (g_ldmBucketSizeLog != LDM_PARAM_DEFAULT) { benchParams.ldmBucketSizeLog = (int)g_ldmBucketSizeLog; } @@ -1348,6 +1353,7 @@ int main(int const argCount, const char* argv[]) if (g_ldmBucketSizeLog != LDM_PARAM_DEFAULT) FIO_setLdmBucketSizeLog(prefs, (int)g_ldmBucketSizeLog); if (g_ldmHashRateLog != LDM_PARAM_DEFAULT) FIO_setLdmHashRateLog(prefs, (int)g_ldmHashRateLog); FIO_setAdaptiveMode(prefs, (unsigned)adapt); + FIO_setUseRowMatchFinder(prefs, useRowMatchFinder); FIO_setAdaptMin(prefs, adaptMin); FIO_setAdaptMax(prefs, adaptMax); FIO_setRsyncable(prefs, rsyncable); @@ -1387,7 +1393,7 @@ int main(int const argCount, const char* argv[]) else operationResult = FIO_compressMultipleFilenames(fCtx, prefs, filenames->fileNames, outMirroredDirName, outDirName, outFileName, suffix, dictFileName, cLevel, compressionParams); #else - (void)contentSize; (void)suffix; (void)adapt; (void)rsyncable; (void)ultra; (void)cLevel; (void)ldmFlag; (void)literalCompressionMode; (void)targetCBlockSize; (void)streamSrcSize; (void)srcSizeHint; (void)ZSTD_strategyMap; /* not used when ZSTD_NOCOMPRESS set */ + (void)contentSize; (void)suffix; (void)adapt; (void)rsyncable; (void)ultra; (void)cLevel; (void)ldmFlag; (void)literalCompressionMode; (void)targetCBlockSize; (void)streamSrcSize; (void)srcSizeHint; (void)ZSTD_strategyMap; (void)useRowMatchFinder; /* not used when ZSTD_NOCOMPRESS set */ DISPLAY("Compression not supported \n"); #endif } else { /* decompression or test */ diff --git a/tests/fuzz/.gitignore b/tests/fuzz/.gitignore index 9bd280c08..8ef3a3efd 100644 --- a/tests/fuzz/.gitignore +++ b/tests/fuzz/.gitignore @@ -19,6 +19,7 @@ sequence_compression_api fuzz-*.log rt_lib_* d_lib_* +crash-* # misc trace diff --git a/tests/fuzz/zstd_helpers.c b/tests/fuzz/zstd_helpers.c index 0607a0745..fcbe3361d 100644 --- a/tests/fuzz/zstd_helpers.c +++ b/tests/fuzz/zstd_helpers.c @@ -91,6 +91,7 @@ void FUZZ_setRandomParameters(ZSTD_CCtx *cctx, size_t srcSize, FUZZ_dataProducer /* Set misc parameters */ setRand(cctx, ZSTD_c_nbWorkers, 0, 2, producer); setRand(cctx, ZSTD_c_rsyncable, 0, 1, producer); + setRand(cctx, ZSTD_c_useRowMatchFinder, 0, 2, producer); setRand(cctx, ZSTD_c_forceMaxWindow, 0, 1, producer); setRand(cctx, ZSTD_c_literalCompressionMode, 0, 2, producer); setRand(cctx, ZSTD_c_forceAttachDict, 0, 2, producer); diff --git a/tests/fuzzer.c b/tests/fuzzer.c index d58e6a568..1b1042393 100644 --- a/tests/fuzzer.c +++ b/tests/fuzzer.c @@ -1757,13 +1757,14 @@ static int basicUnitTests(U32 const seed, double compressibility) size_t const contentSize = 9 KB; const void* const dict = (const char*)CNBuffer; const void* const contentStart = (const char*)dict + flatdictSize; + /* These upper bounds are generally within a few bytes of the compressed size */ size_t const target_nodict_cSize[22+1] = { 3840, 3770, 3870, 3830, 3770, 3770, 3770, 3770, 3750, 3750, 3742, 3670, 3670, 3660, 3660, 3660, 3660, 3660, 3660, 3660, 3660, 3660, 3660 }; size_t const target_wdict_cSize[22+1] = { 2830, 2890, 2890, 2820, 2940, - 2950, 2950, 2921, 2900, 2891, + 2950, 2950, 2925, 2900, 2891, 2910, 2910, 2910, 2770, 2760, 2750, 2750, 2750, 2750, 2750, 2750, 2750, 2750 }; @@ -1800,6 +1801,22 @@ static int basicUnitTests(U32 const seed, double compressibility) DISPLAYLEVEL(4, "level %i with dictionary : max expected %u >= reached %u \n", l, (unsigned)target_wdict_cSize[l], (unsigned)wdict_cSize); } + /* Dict compression with DMS */ + for ( l=1 ; l <= maxLevel; l++) { + size_t wdict_cSize; + CHECK_Z( ZSTD_CCtx_loadDictionary(ctxOrig, dict, flatdictSize) ); + CHECK_Z( ZSTD_CCtx_setParameter(ctxOrig, ZSTD_c_compressionLevel, l) ); + CHECK_Z( ZSTD_CCtx_setParameter(ctxOrig, ZSTD_c_enableDedicatedDictSearch, 0) ); + CHECK_Z( ZSTD_CCtx_setParameter(ctxOrig, ZSTD_c_forceAttachDict, ZSTD_dictForceAttach) ); + wdict_cSize = ZSTD_compress2(ctxOrig, compressedBuffer, compressedBufferSize, contentStart, contentSize); + if (wdict_cSize > target_wdict_cSize[l]) { + DISPLAYLEVEL(1, "error : compression with dictionary and compress2 at level %i worse than expected (%u > %u) \n", + l, (unsigned)wdict_cSize, (unsigned)target_wdict_cSize[l]); + goto _output_error; + } + DISPLAYLEVEL(4, "level %i with dictionary and compress2 : max expected %u >= reached %u \n", + l, (unsigned)target_wdict_cSize[l], (unsigned)wdict_cSize); + } DISPLAYLEVEL(4, "compression efficiency tests OK \n"); }