Merge pull request #743 from facebook/fullbench

compress_generic() automatic optimization opportunities
This commit is contained in:
Yann Collet
2017-07-03 21:26:38 -07:00
committed by GitHub
9 changed files with 329 additions and 165 deletions
+7 -14
View File
@@ -72,6 +72,7 @@
#if defined(ZSTD_DEBUG) && (ZSTD_DEBUG>=2)
# include <stdio.h>
/* recommended values for ZSTD_DEBUG display levels :
* 1 : no display, enables assert() only
* 2 : reserved for currently active debugging path
* 3 : events once per object lifetime (CCtx, CDict)
* 4 : events once per frame
@@ -308,20 +309,6 @@ MEM_STATIC U32 ZSTD_highbit32(U32 val)
void ZSTD_invalidateRepCodes(ZSTD_CCtx* cctx);
typedef enum { ZSTDb_not_buffered, ZSTDb_buffered } ZSTD_buffered_policy_e;
#if 0
/*! ZSTD_compressBegin_internal() :
* innermost initialization function. Private use only.
* expects params to be valid.
* must receive dict, or cdict, or none, but not both.
* @return : 0, or an error code */
size_t ZSTD_compressBegin_internal(ZSTD_CCtx* cctx,
const void* dict, size_t dictSize,
const ZSTD_CDict* cdict,
ZSTD_parameters params, U64 pledgedSrcSize,
ZSTD_buffered_policy_e zbuff);
#endif
/*! ZSTD_initCStream_internal() :
* Private use only. Init streaming operation.
* expects params to be valid.
@@ -332,6 +319,12 @@ size_t ZSTD_initCStream_internal(ZSTD_CStream* zcs,
const ZSTD_CDict* cdict,
ZSTD_parameters params, unsigned long long pledgedSrcSize);
/*! ZSTD_compressStream_generic() :
* Private use only. To be called from zstdmt_compress.c in single-thread mode. */
size_t ZSTD_compressStream_generic(ZSTD_CStream* zcs,
ZSTD_outBuffer* output,
ZSTD_inBuffer* input,
ZSTD_EndDirective const flushMode);
/*! ZSTD_getParamsFromCDict() :
* as the name implies */
+32 -12
View File
@@ -627,6 +627,7 @@ static size_t ZSTD_continueCCtx(ZSTD_CCtx* cctx, ZSTD_parameters params, U64 ple
}
typedef enum { ZSTDcrp_continue, ZSTDcrp_noMemset } ZSTD_compResetPolicy_e;
typedef enum { ZSTDb_not_buffered, ZSTDb_buffered } ZSTD_buffered_policy_e;
/*! ZSTD_resetCCtx_internal() :
note : `params` are assumed fully validated at this stage */
@@ -2954,6 +2955,8 @@ static size_t ZSTD_compressContinue_internal (ZSTD_CCtx* cctx,
const BYTE* const ip = (const BYTE*) src;
size_t fhSize = 0;
DEBUGLOG(5, "ZSTD_compressContinue_internal");
DEBUGLOG(5, "stage: %u", cctx->stage);
if (cctx->stage==ZSTDcs_created) return ERROR(stage_wrong); /* missing init (ZSTD_compressBegin) */
if (frame && (cctx->stage==ZSTDcs_init)) {
@@ -3211,9 +3214,9 @@ static size_t ZSTD_compressBegin_internal(ZSTD_CCtx* cctx,
ZSTD_parameters params, U64 pledgedSrcSize,
ZSTD_buffered_policy_e zbuff)
{
DEBUGLOG(5, "ZSTD_compressBegin_internal");
DEBUGLOG(5, "dict ? %s", dict ? "dict" : cdict ? "cdict" : "none");
DEBUGLOG(5, "dictMode : %u", (U32)dictMode);
DEBUGLOG(4, "ZSTD_compressBegin_internal");
DEBUGLOG(4, "dict ? %s", dict ? "dict" : (cdict ? "cdict" : "none"));
DEBUGLOG(4, "dictMode : %u", (U32)dictMode);
/* params are supposed to be fully validated at this point */
assert(!ZSTD_isError(ZSTD_checkCParams(params.cParams)));
assert(!((dict) && (cdict))); /* either dict or cdict, not both */
@@ -3633,7 +3636,7 @@ static size_t ZSTD_resetCStream_internal(ZSTD_CStream* zcs,
const ZSTD_CDict* cdict,
ZSTD_parameters params, unsigned long long pledgedSrcSize)
{
DEBUGLOG(5, "ZSTD_resetCStream_internal");
DEBUGLOG(4, "ZSTD_resetCStream_internal");
/* params are supposed to be fully validated at this point */
assert(!ZSTD_isError(ZSTD_checkCParams(params.cParams)));
assert(!((dict) && (cdict))); /* either dict or cdict, not both */
@@ -3766,10 +3769,13 @@ MEM_STATIC size_t ZSTD_limitCopy(void* dst, size_t dstCapacity,
return length;
}
static size_t ZSTD_compressStream_generic(ZSTD_CStream* zcs,
ZSTD_outBuffer* output,
ZSTD_inBuffer* input,
ZSTD_EndDirective const flushMode)
/** ZSTD_compressStream_generic():
* internal function for all *compressStream*() variants and *compress_generic()
* @return : hint size for next input */
size_t ZSTD_compressStream_generic(ZSTD_CStream* zcs,
ZSTD_outBuffer* output,
ZSTD_inBuffer* input,
ZSTD_EndDirective const flushMode)
{
const char* const istart = (const char*)input->src;
const char* const iend = istart + input->size;
@@ -3780,7 +3786,7 @@ static size_t ZSTD_compressStream_generic(ZSTD_CStream* zcs,
U32 someMoreWork = 1;
/* check expectations */
DEBUGLOG(5, "ZSTD_compressStream_generic");
DEBUGLOG(5, "ZSTD_compressStream_generic, order %u", (U32)flushMode);
assert(zcs->inBuff != NULL);
assert(zcs->inBuffSize>0);
assert(zcs->outBuff!= NULL);
@@ -3796,7 +3802,21 @@ static size_t ZSTD_compressStream_generic(ZSTD_CStream* zcs,
return ERROR(init_missing);
case zcss_load:
/* complete inBuffer */
if ( (flushMode == ZSTD_e_end)
&& ((size_t)(oend-op) >= ZSTD_compressBound(iend-ip)) /* enough dstCapacity */
&& (zcs->inBuffPos == 0) ) {
/* shortcut to compression pass directly into output buffer */
size_t const cSize = ZSTD_compressEnd(zcs,
op, oend-op, ip, iend-ip);
DEBUGLOG(4, "ZSTD_compressEnd : %u", (U32)cSize);
if (ZSTD_isError(cSize)) return cSize;
ip = iend;
op += cSize;
zcs->frameEnded = 1;
ZSTD_startNewCompression(zcs);
someMoreWork = 0; break;
}
/* complete loading into inBuffer */
{ size_t const toLoad = zcs->inBuffTarget - zcs->inBuffPos;
size_t const loaded = ZSTD_limitCopy(
zcs->inBuff + zcs->inBuffPos, toLoad,
@@ -3922,8 +3942,8 @@ size_t ZSTD_compress_generic (ZSTD_CCtx* cctx,
if (input->pos > input->size) return ERROR(GENERIC);
assert(cctx!=NULL);
/* transparent initialization stage */
if (cctx->streamStage == zcss_init) {
/* transparent reset */
const void* const prefix = cctx->prefix;
size_t const prefixSize = cctx->prefixSize;
ZSTD_parameters params = cctx->requestedParams;
@@ -3944,6 +3964,7 @@ size_t ZSTD_compress_generic (ZSTD_CCtx* cctx,
CHECK_F( ZSTD_resetCStream_internal(cctx, prefix, prefixSize, cctx->dictMode, cctx->cdict, params, cctx->pledgedSrcSizePlusOne-1) );
} }
/* compression stage */
#ifdef ZSTD_MULTITHREAD
if (cctx->nbThreads > 1) {
size_t const flushMin = ZSTDMT_compressStream_generic(cctx->mtctx, output, input, endOp);
@@ -3956,7 +3977,6 @@ size_t ZSTD_compress_generic (ZSTD_CCtx* cctx,
}
#endif
DEBUGLOG(5, "calling ZSTD_compressStream_generic(%i,...)", endOp);
CHECK_F( ZSTD_compressStream_generic(cctx, output, input, endOp) );
DEBUGLOG(5, "completed ZSTD_compress_generic");
return cctx->outBuffContentSize - cctx->outBuffFlushedSize; /* remaining to flush */
+148 -86
View File
@@ -50,7 +50,7 @@ static unsigned long long GetCurrentClockTimeMicroseconds(void)
return ((((unsigned long long)newTicks)*(1000000))/_ticksPerSecond); }
}
#define MUTEX_WAIT_TIME_DLEVEL 5
#define MUTEX_WAIT_TIME_DLEVEL 6
#define PTHREAD_MUTEX_LOCK(mutex) { \
if (ZSTD_DEBUG>=MUTEX_WAIT_TIME_DLEVEL) { \
unsigned long long const beforeTime = GetCurrentClockTimeMicroseconds(); \
@@ -261,7 +261,7 @@ void ZSTDMT_compressChunk(void* jobDescription)
ZSTDMT_jobDescription* const job = (ZSTDMT_jobDescription*)jobDescription;
const void* const src = (const char*)job->srcStart + job->dictSize;
buffer_t const dstBuff = job->dstBuff;
DEBUGLOG(4, "job (first:%u) (last:%u) : dictSize %u, srcSize %u",
DEBUGLOG(5, "job (first:%u) (last:%u) : dictSize %u, srcSize %u",
job->firstChunk, job->lastChunk, (U32)job->dictSize, (U32)job->srcSize);
if (job->cdict) { /* should only happen for first segment */
size_t const initError = ZSTD_compressBegin_usingCDict_advanced(job->cctx, job->cdict, job->params.fParams, job->fullFrameSize);
@@ -285,7 +285,7 @@ void ZSTDMT_compressChunk(void* jobDescription)
job->cSize = (job->lastChunk) ?
ZSTD_compressEnd (job->cctx, dstBuff.start, dstBuff.size, src, job->srcSize) :
ZSTD_compressContinue(job->cctx, dstBuff.start, dstBuff.size, src, job->srcSize);
DEBUGLOG(4, "compressed %u bytes into %u bytes (first:%u) (last:%u)",
DEBUGLOG(5, "compressed %u bytes into %u bytes (first:%u) (last:%u)",
(unsigned)job->srcSize, (unsigned)job->cSize, job->firstChunk, job->lastChunk);
DEBUGLOG(5, "dstBuff.size : %u ; => %s", (U32)dstBuff.size, ZSTD_getErrorName(job->cSize));
@@ -331,14 +331,20 @@ struct ZSTDMT_CCtx_s {
const ZSTD_CDict* cdict;
};
static ZSTDMT_jobDescription* ZSTDMT_allocJobsTable(U32* nbJobsPtr, ZSTD_customMem cMem)
{
U32 const nbJobsLog2 = ZSTD_highbit32(*nbJobsPtr) + 1;
U32 const nbJobs = 1 << nbJobsLog2;
*nbJobsPtr = nbJobs;
return (ZSTDMT_jobDescription*) ZSTD_calloc(
nbJobs * sizeof(ZSTDMT_jobDescription), cMem);
}
ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced(unsigned nbThreads, ZSTD_customMem cMem)
{
ZSTDMT_CCtx* mtctx;
U32 const minNbJobs = nbThreads + 2;
U32 const nbJobsLog2 = ZSTD_highbit32(minNbJobs) + 1;
U32 const nbJobs = 1 << nbJobsLog2;
DEBUGLOG(5, "nbThreads: %u ; minNbJobs: %u ; nbJobsLog2: %u ; nbJobs: %u",
nbThreads, minNbJobs, nbJobsLog2, nbJobs);
U32 nbJobs = nbThreads + 2;
DEBUGLOG(3, "ZSTDMT_createCCtx_advanced");
if ((nbThreads < 1) | (nbThreads > ZSTDMT_NBTHREADS_MAX)) return NULL;
if ((cMem.customAlloc!=NULL) ^ (cMem.customFree!=NULL))
@@ -349,13 +355,12 @@ ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced(unsigned nbThreads, ZSTD_customMem cMem)
if (!mtctx) return NULL;
mtctx->cMem = cMem;
mtctx->nbThreads = nbThreads;
mtctx->jobIDMask = nbJobs - 1;
mtctx->allJobsCompleted = 1;
mtctx->sectionSize = 0;
mtctx->overlapRLog = 3;
mtctx->factory = POOL_create(nbThreads, 1);
mtctx->jobs = (ZSTDMT_jobDescription*)ZSTD_calloc(
nbJobs * sizeof(*mtctx->jobs), cMem);
mtctx->jobs = ZSTDMT_allocJobsTable(&nbJobs, cMem);
mtctx->jobIDMask = nbJobs - 1;
mtctx->buffPool = ZSTDMT_createBufferPool(nbThreads, cMem);
mtctx->cctxPool = ZSTDMT_createCCtxPool(nbThreads, cMem);
if (!mtctx->factory | !mtctx->jobs | !mtctx->buffPool | !mtctx->cctxPool) {
@@ -374,10 +379,11 @@ ZSTDMT_CCtx* ZSTDMT_createCCtx(unsigned nbThreads)
}
/* ZSTDMT_releaseAllJobResources() :
* Ensure all workers are killed first. */
* note : ensure all workers are killed first ! */
static void ZSTDMT_releaseAllJobResources(ZSTDMT_CCtx* mtctx)
{
unsigned jobID;
DEBUGLOG(4, "ZSTDMT_releaseAllJobResources");
for (jobID=0; jobID <= mtctx->jobIDMask; jobID++) {
ZSTDMT_releaseBuffer(mtctx->buffPool, mtctx->jobs[jobID].dstBuff);
mtctx->jobs[jobID].dstBuff = g_nullBuffer;
@@ -439,31 +445,50 @@ size_t ZSTDMT_setMTCtxParameter(ZSTDMT_CCtx* mtctx, ZSDTMT_parameter parameter,
/* ===== Multi-threaded compression ===== */
/* ------------------------------------------ */
size_t ZSTDMT_compressCCtx(ZSTDMT_CCtx* mtctx,
static unsigned computeNbChunks(size_t srcSize, unsigned windowLog, unsigned nbThreads) {
size_t const chunkSizeTarget = (size_t)1 << (windowLog + 2);
size_t const chunkMaxSize = chunkSizeTarget << 2;
size_t const passSizeMax = chunkMaxSize * nbThreads;
unsigned const multiplier = (unsigned)(srcSize / passSizeMax) + 1;
unsigned const nbChunksLarge = multiplier * nbThreads;
unsigned const nbChunksMax = (unsigned)(srcSize / chunkSizeTarget) + 1;
unsigned const nbChunksSmall = MIN(nbChunksMax, nbThreads);
return (multiplier>1) ? nbChunksLarge : nbChunksSmall;
}
size_t ZSTDMT_compress_advanced(ZSTDMT_CCtx* mtctx,
void* dst, size_t dstCapacity,
const void* src, size_t srcSize,
int compressionLevel)
const ZSTD_CDict* cdict,
ZSTD_parameters const params,
unsigned overlapRLog)
{
ZSTD_parameters params = ZSTD_getParams(compressionLevel, srcSize, 0);
U32 const overlapLog = (compressionLevel >= ZSTD_maxCLevel()) ? 0 : 3;
size_t const overlapSize = (size_t)1 << (params.cParams.windowLog - overlapLog);
size_t const chunkTargetSize = (size_t)1 << (params.cParams.windowLog + 2);
unsigned const nbChunksMax = (unsigned)(srcSize / chunkTargetSize) + 1;
unsigned nbChunks = MIN(nbChunksMax, mtctx->nbThreads);
size_t const overlapSize = (overlapRLog>=9) ? 0 : (size_t)1 << (params.cParams.windowLog - overlapRLog);
unsigned nbChunks = computeNbChunks(srcSize, params.cParams.windowLog, mtctx->nbThreads);
size_t const proposedChunkSize = (srcSize + (nbChunks-1)) / nbChunks;
size_t const avgChunkSize = ((proposedChunkSize & 0x1FFFF) < 0xFFFF) ? proposedChunkSize + 0xFFFF : proposedChunkSize; /* avoid too small last block */
size_t remainingSrcSize = srcSize;
size_t const avgChunkSize = ((proposedChunkSize & 0x1FFFF) < 0x7FFF) ? proposedChunkSize + 0xFFFF : proposedChunkSize; /* avoid too small last block */
const char* const srcStart = (const char*)src;
size_t remainingSrcSize = srcSize;
unsigned const compressWithinDst = (dstCapacity >= ZSTD_compressBound(srcSize)) ? nbChunks : (unsigned)(dstCapacity / ZSTD_compressBound(avgChunkSize)); /* presumes avgChunkSize >= 256 KB, which should be the case */
size_t frameStartPos = 0, dstBufferPos = 0;
DEBUGLOG(4, "windowLog : %2u => chunkTargetSize : %u bytes ", params.cParams.windowLog, (U32)chunkTargetSize);
DEBUGLOG(4, "nbChunks : %2u (chunkSize : %u bytes) ", nbChunks, (U32)avgChunkSize);
params.fParams.contentSizeFlag = 1;
assert(avgChunkSize >= 256 KB); /* required for ZSTD_compressBound(A) + ZSTD_compressBound(B) <= ZSTD_compressBound(A+B) */
if (nbChunks==1) { /* fallback to single-thread mode */
ZSTD_CCtx* const cctx = mtctx->cctxPool->cctx[0];
return ZSTD_compressCCtx(cctx, dst, dstCapacity, src, srcSize, compressionLevel);
if (cdict) return ZSTD_compress_usingCDict_advanced(cctx, dst, dstCapacity, src, srcSize, cdict, params.fParams);
return ZSTD_compress_advanced(cctx, dst, dstCapacity, src, srcSize, NULL, 0, params);
}
if (nbChunks > mtctx->jobIDMask+1) { /* enlarge job table */
U32 nbJobs = nbChunks;
ZSTD_free(mtctx->jobs, mtctx->cMem);
mtctx->jobIDMask = 0;
mtctx->jobs = ZSTDMT_allocJobsTable(&nbJobs, mtctx->cMem);
if (mtctx->jobs==NULL) return ERROR(memory_allocation);
mtctx->jobIDMask = nbJobs - 1;
}
{ unsigned u;
@@ -478,15 +503,18 @@ size_t ZSTDMT_compressCCtx(ZSTDMT_CCtx* mtctx,
if ((cctx==NULL) || (dstBuffer.start==NULL)) {
mtctx->jobs[u].cSize = ERROR(memory_allocation); /* job result */
mtctx->jobs[u].jobCompleted = 1;
nbChunks = u+1;
nbChunks = u+1; /* only wait and free u jobs, instead of initially expected nbChunks ones */
break; /* let's wait for previous jobs to complete, but don't start new ones */
}
mtctx->jobs[u].srcStart = srcStart + frameStartPos - dictSize;
mtctx->jobs[u].dictSize = dictSize;
mtctx->jobs[u].srcSize = chunkSize;
mtctx->jobs[u].cdict = mtctx->nextJobID==0 ? cdict : NULL;
mtctx->jobs[u].fullFrameSize = srcSize;
mtctx->jobs[u].params = params;
/* do not calculate checksum within sections, but write it in header for first section */
if (u!=0) mtctx->jobs[u].params.fParams.checksumFlag = 0;
mtctx->jobs[u].dstBuff = dstBuffer;
mtctx->jobs[u].cctx = cctx;
mtctx->jobs[u].firstChunk = (u==0);
@@ -495,16 +523,16 @@ size_t ZSTDMT_compressCCtx(ZSTDMT_CCtx* mtctx,
mtctx->jobs[u].jobCompleted_mutex = &mtctx->jobCompleted_mutex;
mtctx->jobs[u].jobCompleted_cond = &mtctx->jobCompleted_cond;
DEBUGLOG(4, "posting job %u (%u bytes)", u, (U32)chunkSize);
DEBUG_PRINTHEX(5, mtctx->jobs[u].srcStart, 12);
DEBUGLOG(5, "posting job %u (%u bytes)", u, (U32)chunkSize);
DEBUG_PRINTHEX(6, mtctx->jobs[u].srcStart, 12);
POOL_add(mtctx->factory, ZSTDMT_compressChunk, &mtctx->jobs[u]);
frameStartPos += chunkSize;
dstBufferPos += dstBufferCapacity;
remainingSrcSize -= chunkSize;
} }
/* note : since nbChunks <= nbThreads, all jobs should be running immediately in parallel */
/* collect result */
{ unsigned chunkID;
size_t error = 0, dstPos = 0;
for (chunkID=0; chunkID<nbChunks; chunkID++) {
@@ -523,11 +551,13 @@ size_t ZSTDMT_compressCCtx(ZSTDMT_CCtx* mtctx,
{ size_t const cSize = mtctx->jobs[chunkID].cSize;
if (ZSTD_isError(cSize)) error = cSize;
if ((!error) && (dstPos + cSize > dstCapacity)) error = ERROR(dstSize_tooSmall);
if (chunkID) { /* note : chunk 0 is already written directly into dst */
if (chunkID) { /* note : chunk 0 is written directly at dst, which is correct position */
if (!error)
memmove((char*)dst + dstPos, mtctx->jobs[chunkID].dstBuff.start, cSize); /* may overlap if chunk decompressed within dst */
if (chunkID >= compressWithinDst) /* otherwise, it decompresses within dst */
memmove((char*)dst + dstPos, mtctx->jobs[chunkID].dstBuff.start, cSize); /* may overlap when chunk compressed within dst */
if (chunkID >= compressWithinDst) { /* chunk compressed into its own buffer, which must be released */
DEBUGLOG(5, "releasing buffer %u>=%u", chunkID, compressWithinDst);
ZSTDMT_releaseBuffer(mtctx->buffPool, mtctx->jobs[chunkID].dstBuff);
}
mtctx->jobs[chunkID].dstBuff = g_nullBuffer;
}
dstPos += cSize ;
@@ -536,7 +566,18 @@ size_t ZSTDMT_compressCCtx(ZSTDMT_CCtx* mtctx,
if (!error) DEBUGLOG(4, "compressed size : %u ", (U32)dstPos);
return error ? error : dstPos;
}
}
size_t ZSTDMT_compressCCtx(ZSTDMT_CCtx* mtctx,
void* dst, size_t dstCapacity,
const void* src, size_t srcSize,
int compressionLevel)
{
U32 const overlapRLog = (compressionLevel >= ZSTD_maxCLevel()) ? 0 : 3;
ZSTD_parameters params = ZSTD_getParams(compressionLevel, srcSize, 0);
params.fParams.contentSizeFlag = 1;
return ZSTDMT_compress_advanced(mtctx, dst, dstCapacity, src, srcSize, NULL, params, overlapRLog);
}
@@ -546,6 +587,7 @@ size_t ZSTDMT_compressCCtx(ZSTDMT_CCtx* mtctx,
static void ZSTDMT_waitForAllJobsCompleted(ZSTDMT_CCtx* zcs)
{
DEBUGLOG(4, "ZSTDMT_waitForAllJobsCompleted");
while (zcs->doneJobID < zcs->nextJobID) {
unsigned const jobID = zcs->doneJobID & zcs->jobIDMask;
PTHREAD_MUTEX_LOCK(&zcs->jobCompleted_mutex);
@@ -559,17 +601,19 @@ static void ZSTDMT_waitForAllJobsCompleted(ZSTDMT_CCtx* zcs)
}
/** ZSTDMT_initCStream_internal() :
* internal usage only */
size_t ZSTDMT_initCStream_internal(ZSTDMT_CCtx* zcs,
const void* dict, size_t dictSize, const ZSTD_CDict* cdict,
ZSTD_parameters params, unsigned long long pledgedSrcSize)
{
DEBUGLOG(5, "ZSTDMT_initCStream_internal");
DEBUGLOG(4, "ZSTDMT_initCStream_internal");
/* params are supposed to be fully validated at this point */
assert(!ZSTD_isError(ZSTD_checkCParams(params.cParams)));
assert(!((dict) && (cdict))); /* either dict or cdict, not both */
if (zcs->nbThreads==1) {
DEBUGLOG(5, "single thread mode");
DEBUGLOG(4, "single thread mode");
return ZSTD_initCStream_internal(zcs->cctxPool->cctx[0],
dict, dictSize, cdict,
params, pledgedSrcSize);
@@ -584,6 +628,7 @@ size_t ZSTDMT_initCStream_internal(ZSTDMT_CCtx* zcs,
zcs->params = params;
zcs->frameContentSize = pledgedSrcSize;
if (dict) {
DEBUGLOG(4,"cdictLocal: %08X", (U32)(size_t)zcs->cdictLocal);
ZSTD_freeCDict(zcs->cdictLocal);
zcs->cdictLocal = ZSTD_createCDict_advanced(dict, dictSize,
0 /* byRef */, ZSTD_dm_auto, /* note : a loadPrefix becomes an internal CDict */
@@ -591,18 +636,19 @@ size_t ZSTDMT_initCStream_internal(ZSTDMT_CCtx* zcs,
zcs->cdict = zcs->cdictLocal;
if (zcs->cdictLocal == NULL) return ERROR(memory_allocation);
} else {
DEBUGLOG(4,"cdictLocal: %08X", (U32)(size_t)zcs->cdictLocal);
ZSTD_freeCDict(zcs->cdictLocal);
zcs->cdictLocal = NULL;
zcs->cdict = cdict;
}
zcs->targetDictSize = (zcs->overlapRLog>=9) ? 0 : (size_t)1 << (zcs->params.cParams.windowLog - zcs->overlapRLog);
DEBUGLOG(5, "overlapRLog : %u ", zcs->overlapRLog);
DEBUGLOG(5, "overlap Size : %u KB", (U32)(zcs->targetDictSize>>10));
DEBUGLOG(4, "overlapRLog : %u ", zcs->overlapRLog);
DEBUGLOG(4, "overlap Size : %u KB", (U32)(zcs->targetDictSize>>10));
zcs->targetSectionSize = zcs->sectionSize ? zcs->sectionSize : (size_t)1 << (zcs->params.cParams.windowLog + 2);
zcs->targetSectionSize = MAX(ZSTDMT_SECTION_SIZE_MIN, zcs->targetSectionSize);
zcs->targetSectionSize = MAX(zcs->targetDictSize, zcs->targetSectionSize);
DEBUGLOG(5, "Section Size : %u KB", (U32)(zcs->targetSectionSize>>10));
DEBUGLOG(4, "Section Size : %u KB", (U32)(zcs->targetSectionSize>>10));
zcs->marginSize = zcs->targetSectionSize >> 2;
zcs->inBuffSize = zcs->targetDictSize + zcs->targetSectionSize + zcs->marginSize;
zcs->inBuff.buffer = ZSTDMT_getBuffer(zcs->buffPool, zcs->inBuffSize);
@@ -797,32 +843,80 @@ static size_t ZSTDMT_flushNextJob(ZSTDMT_CCtx* zcs, ZSTD_outBuffer* output, unsi
} }
size_t ZSTDMT_compressStream(ZSTDMT_CCtx* zcs, ZSTD_outBuffer* output, ZSTD_inBuffer* input)
/** ZSTDMT_compressStream_generic() :
* internal use only
* assumption : output and input are valid (pos <= size)
* @return : minimum amount of data remaining to flush, 0 if none */
size_t ZSTDMT_compressStream_generic(ZSTDMT_CCtx* mtctx,
ZSTD_outBuffer* output,
ZSTD_inBuffer* input,
ZSTD_EndDirective endOp)
{
size_t const newJobThreshold = zcs->dictSize + zcs->targetSectionSize + zcs->marginSize;
if (zcs->frameEnded) {
/* current frame being ended. Only flush is allowed. Or start new job with init */
DEBUGLOG(5, "ZSTDMT_compressStream: zcs::frameEnded==1");
size_t const newJobThreshold = mtctx->dictSize + mtctx->targetSectionSize + mtctx->marginSize;
assert(output->pos <= output->size);
assert(input->pos <= input->size);
if ((mtctx->frameEnded) && (endOp==ZSTD_e_continue)) {
/* current frame being ended. Only flush/end are allowed. Or start new job with init */
return ERROR(stage_wrong);
}
if (zcs->nbThreads==1) {
return ZSTD_compressStream(zcs->cctxPool->cctx[0], output, input);
if (mtctx->nbThreads==1) {
return ZSTD_compressStream_generic(mtctx->cctxPool->cctx[0], output, input, endOp);
}
/* single-pass shortcut (note : this is blocking-mode) */
if ( (mtctx->nextJobID==0) /* just started */
&& (mtctx->inBuff.filled==0) /* nothing buffered yet */
&& (endOp==ZSTD_e_end) /* end order, immediately at beginning */
&& (output->size - output->pos >= ZSTD_compressBound(input->size - input->pos)) ) { /* enough room */
size_t const cSize = ZSTDMT_compress_advanced(mtctx,
(char*)output->dst + output->pos, output->size - output->pos,
(const char*)input->src + input->pos, input->size - input->pos,
mtctx->cdict, mtctx->params, mtctx->overlapRLog);
if (ZSTD_isError(cSize)) return cSize;
input->pos = input->size;
output->pos += cSize;
ZSTDMT_releaseBuffer(mtctx->buffPool, mtctx->inBuff.buffer); /* was allocated in initStream */
mtctx->allJobsCompleted = 1;
mtctx->frameEnded = 1;
return 0;
}
/* fill input buffer */
{ size_t const toLoad = MIN(input->size - input->pos, zcs->inBuffSize - zcs->inBuff.filled);
memcpy((char*)zcs->inBuff.buffer.start + zcs->inBuff.filled, (const char*)input->src + input->pos, toLoad);
if (input->src) { /* support NULL input */
size_t const toLoad = MIN(input->size - input->pos, mtctx->inBuffSize - mtctx->inBuff.filled);
memcpy((char*)mtctx->inBuff.buffer.start + mtctx->inBuff.filled, (const char*)input->src + input->pos, toLoad);
input->pos += toLoad;
zcs->inBuff.filled += toLoad;
mtctx->inBuff.filled += toLoad;
}
if ( (zcs->inBuff.filled >= newJobThreshold) /* filled enough : let's compress */
&& (zcs->nextJobID <= zcs->doneJobID + zcs->jobIDMask) ) { /* avoid overwriting job round buffer */
CHECK_F( ZSTDMT_createCompressionJob(zcs, zcs->targetSectionSize, 0 /* endFrame */) );
if ( (mtctx->inBuff.filled >= newJobThreshold) /* filled enough : let's compress */
&& (mtctx->nextJobID <= mtctx->doneJobID + mtctx->jobIDMask) ) { /* avoid overwriting job round buffer */
CHECK_F( ZSTDMT_createCompressionJob(mtctx, mtctx->targetSectionSize, 0 /* endFrame */) );
}
/* check for data to flush */
CHECK_F( ZSTDMT_flushNextJob(zcs, output, (zcs->inBuff.filled == zcs->inBuffSize) /* blockToFlush */) ); /* block if it wasn't possible to create new job due to saturation */
/* check for potential compressed data ready to be flushed */
CHECK_F( ZSTDMT_flushNextJob(mtctx, output, (mtctx->inBuff.filled == mtctx->inBuffSize) /* blockToFlush */) ); /* block if it wasn't possible to create new job due to saturation */
if (input->pos < input->size) /* input not consumed : do not flush yet */
endOp = ZSTD_e_continue;
switch(endOp)
{
case ZSTD_e_flush:
return ZSTDMT_flushStream(mtctx, output);
case ZSTD_e_end:
return ZSTDMT_endStream(mtctx, output);
case ZSTD_e_continue:
return 1;
default:
return ERROR(GENERIC); /* invalid endDirective */
}
}
size_t ZSTDMT_compressStream(ZSTDMT_CCtx* zcs, ZSTD_outBuffer* output, ZSTD_inBuffer* input)
{
CHECK_F( ZSTDMT_compressStream_generic(zcs, output, input, ZSTD_e_continue) );
/* recommended next input size : fill current input buffer */
return zcs->inBuffSize - zcs->inBuff.filled; /* note : could be zero when input buffer is fully filled and no more availability to create new job */
@@ -833,21 +927,13 @@ static size_t ZSTDMT_flushStream_internal(ZSTDMT_CCtx* zcs, ZSTD_outBuffer* outp
{
size_t const srcSize = zcs->inBuff.filled - zcs->dictSize;
if (srcSize)
DEBUGLOG(5, "ZSTDMT_flushStream_internal : %u bytes left to compress",
(U32)srcSize);
if ( ((srcSize > 0) || (endFrame && !zcs->frameEnded))
&& (zcs->nextJobID <= zcs->doneJobID + zcs->jobIDMask) ) {
DEBUGLOG(5, "create new job with %u bytes to compress", (U32)srcSize);
DEBUGLOG(5, "end order : %u", endFrame);
CHECK_F( ZSTDMT_createCompressionJob(zcs, srcSize, endFrame) );
DEBUGLOG(5, "resulting zcs->frameEnded : %u", zcs->frameEnded);
}
/* check if there is any data available to flush */
DEBUGLOG(5, "zcs->doneJobID : %u ; zcs->nextJobID : %u",
zcs->doneJobID, zcs->nextJobID);
return ZSTDMT_flushNextJob(zcs, output, 1 /*blockToFlush */);
return ZSTDMT_flushNextJob(zcs, output, 1 /* blockToFlush */);
}
@@ -861,32 +947,8 @@ size_t ZSTDMT_flushStream(ZSTDMT_CCtx* zcs, ZSTD_outBuffer* output)
size_t ZSTDMT_endStream(ZSTDMT_CCtx* zcs, ZSTD_outBuffer* output)
{
DEBUGLOG(5, "ZSTDMT_endStream");
DEBUGLOG(4, "ZSTDMT_endStream");
if (zcs->nbThreads==1)
return ZSTD_endStream(zcs->cctxPool->cctx[0], output);
return ZSTDMT_flushStream_internal(zcs, output, 1 /* endFrame */);
}
size_t ZSTDMT_compressStream_generic(ZSTDMT_CCtx* mtctx,
ZSTD_outBuffer* output,
ZSTD_inBuffer* input,
ZSTD_EndDirective endOp)
{
DEBUGLOG(5, "in: pos:%u / size:%u ; endOp=%u",
(U32)input->pos, (U32)input->size, (U32)endOp);
if (input->pos < input->size) /* exclude final flushes */
CHECK_F(ZSTDMT_compressStream(mtctx, output, input));
if (input->pos < input->size) endOp = ZSTD_e_continue;
switch(endOp)
{
case ZSTD_e_flush:
return ZSTDMT_flushStream(mtctx, output);
case ZSTD_e_end:
DEBUGLOG(5, "endOp:%u; calling ZSTDMT_endStream", (U32)endOp);
return ZSTDMT_endStream(mtctx, output);
case ZSTD_e_continue:
return 1;
default:
return ERROR(GENERIC); /* invalid endDirective */
}
}
+8
View File
@@ -44,6 +44,7 @@ ZSTDLIB_API size_t ZSTDMT_compressCCtx(ZSTDMT_CCtx* mtctx,
int compressionLevel);
/* === Streaming functions === */
ZSTDLIB_API size_t ZSTDMT_initCStream(ZSTDMT_CCtx* mtctx, int compressionLevel);
@@ -61,6 +62,13 @@ ZSTDLIB_API size_t ZSTDMT_endStream(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output);
# define ZSTDMT_SECTION_SIZE_MIN (1U << 20) /* 1 MB - Minimum size of each compression job */
#endif
ZSTDLIB_API size_t ZSTDMT_compress_advanced(ZSTDMT_CCtx* mtctx,
void* dst, size_t dstCapacity,
const void* src, size_t srcSize,
const ZSTD_CDict* cdict,
ZSTD_parameters const params,
unsigned overlapRLog);
ZSTDLIB_API size_t ZSTDMT_initCStream_advanced(ZSTDMT_CCtx* mtctx,
const void* dict, size_t dictSize, /* dict can be released after init, a local copy is preserved within zcs */
ZSTD_parameters params,
+1
View File
@@ -198,6 +198,7 @@ ZSTD_DCtx* ZSTD_createDCtx_advanced(ZSTD_customMem customMem)
if (!dctx) return NULL;
dctx->customMem = customMem;
dctx->legacyContext = NULL;
dctx->previousLegacyVersion = 0;
ZSTD_initDCtx_internal(dctx);
return dctx;
}