Changed nbThreads for nbWorkers

This makes it easier to explain that nbWorkers=0 --> single-threaded mode,
while nbWorkers=1 --> asynchronous mode (one mode thread on top of the "main" caller thread).
No need for an additional asynchronous mode flag.
nbWorkers>=2 works the same as nbThreads>=2 previously.
This commit is contained in:
Yann Collet
2018-02-01 19:29:30 -08:00
parent 4b6a94f0cc
commit 209df52ba2
15 changed files with 165 additions and 183 deletions
+22 -34
View File
@@ -281,13 +281,12 @@ size_t ZSTD_CCtx_setParameter(ZSTD_CCtx* cctx, ZSTD_cParameter param, unsigned v
* default : 0 when using a CDict, 1 when using a Prefix */
return ZSTD_CCtxParam_setParameter(&cctx->requestedParams, param, value);
case ZSTD_p_nbThreads:
if ((value > 1) && cctx->staticSize) {
case ZSTD_p_nbWorkers:
if ((value>0) && cctx->staticSize) {
return ERROR(parameter_unsupported); /* MT not compatible with static alloc */
}
return ZSTD_CCtxParam_setParameter(&cctx->requestedParams, param, value);
case ZSTD_p_nonBlockingMode:
case ZSTD_p_jobSize:
case ZSTD_p_overlapSizeLog:
return ZSTD_CCtxParam_setParameter(&cctx->requestedParams, param, value);
@@ -403,21 +402,12 @@ size_t ZSTD_CCtxParam_setParameter(
CCtxParams->forceWindow = (value > 0);
return CCtxParams->forceWindow;
case ZSTD_p_nbThreads :
if (value == 0) return CCtxParams->nbThreads;
case ZSTD_p_nbWorkers :
#ifndef ZSTD_MULTITHREAD
if (value > 1) return ERROR(parameter_unsupported);
return 1;
if (value > 0) return ERROR(parameter_unsupported);
return 0;
#else
return ZSTDMT_CCtxParam_setNbThreads(CCtxParams, value);
#endif
case ZSTD_p_nonBlockingMode :
#ifndef ZSTD_MULTITHREAD
return ERROR(parameter_unsupported);
#else
CCtxParams->nonBlockingMode = (value>0);
return CCtxParams->nonBlockingMode;
return ZSTDMT_CCtxParam_setNbWorkers(CCtxParams, value);
#endif
case ZSTD_p_jobSize :
@@ -489,7 +479,7 @@ size_t ZSTD_CCtx_setParametersUsingCCtxParams(
cctx->requestedParams = *params;
#ifdef ZSTD_MULTITHREAD
if (cctx->mtctx)
ZSTDMT_MTCtx_setParametersUsingCCtxParams(cctx->mtctx, params);
ZSTDMT_MTCtx_setParametersUsingCCtxParams_whileCompressing(cctx->mtctx, params);
#endif
return 0;
@@ -687,7 +677,7 @@ static size_t ZSTD_sizeof_matchState(ZSTD_compressionParameters const* cParams,
size_t ZSTD_estimateCCtxSize_usingCCtxParams(const ZSTD_CCtx_params* params)
{
/* Estimate CCtx size is supported for single-threaded compression only. */
if (params->nbThreads > 1) { return ERROR(GENERIC); }
if (params->nbWorkers > 0) { return ERROR(GENERIC); }
{ ZSTD_compressionParameters const cParams =
ZSTD_getCParamsFromCCtxParams(*params, 0, 0);
size_t const blockSize = MIN(ZSTD_BLOCKSIZE_MAX, (size_t)1 << cParams.windowLog);
@@ -736,7 +726,7 @@ size_t ZSTD_estimateCCtxSize(int compressionLevel)
size_t ZSTD_estimateCStreamSize_usingCCtxParams(const ZSTD_CCtx_params* params)
{
if (params->nbThreads > 1) { return ERROR(GENERIC); }
if (params->nbWorkers > 0) { return ERROR(GENERIC); }
{ size_t const CCtxSize = ZSTD_estimateCCtxSize_usingCCtxParams(params);
size_t const blockSize = MIN(ZSTD_BLOCKSIZE_MAX, (size_t)1 << params->cParams.windowLog);
size_t const inBuffSize = ((size_t)1 << params->cParams.windowLog) + blockSize;
@@ -775,7 +765,7 @@ size_t ZSTD_estimateCStreamSize(int compressionLevel) {
ZSTD_frameProgression ZSTD_getFrameProgression(const ZSTD_CCtx* cctx)
{
#ifdef ZSTD_MULTITHREAD
if ((cctx->appliedParams.nbThreads > 1) || (cctx->appliedParams.nonBlockingMode)) {
if (cctx->appliedParams.nbWorkers > 0) {
return ZSTDMT_getFrameProgression(cctx->mtctx);
}
#endif
@@ -3166,28 +3156,26 @@ size_t ZSTD_compress_generic (ZSTD_CCtx* cctx,
#ifdef ZSTD_MULTITHREAD
if ((cctx->pledgedSrcSizePlusOne-1) <= ZSTDMT_JOBSIZE_MIN) {
params.nbThreads = 1; /* do not invoke multi-threading when src size is too small */
params.nonBlockingMode = 0;
params.nbWorkers = 0; /* do not invoke multi-threading when src size is too small */
}
if ((params.nbThreads > 1) | (params.nonBlockingMode == 1)) {
if (cctx->mtctx == NULL || (params.nbThreads != ZSTDMT_getNbThreads(cctx->mtctx))) {
DEBUGLOG(4, "ZSTD_compress_generic: creating new mtctx for nbThreads=%u",
params.nbThreads);
if (params.nbWorkers > 0) {
if (cctx->mtctx == NULL || (params.nbWorkers != ZSTDMT_getNbWorkers(cctx->mtctx))) {
DEBUGLOG(4, "ZSTD_compress_generic: creating new mtctx for nbWorkers=%u",
params.nbWorkers);
if (cctx->mtctx != NULL)
DEBUGLOG(4, "ZSTD_compress_generic: previous nbThreads was %u",
ZSTDMT_getNbThreads(cctx->mtctx));
DEBUGLOG(4, "ZSTD_compress_generic: previous nbWorkers was %u",
ZSTDMT_getNbWorkers(cctx->mtctx));
ZSTDMT_freeCCtx(cctx->mtctx);
cctx->mtctx = ZSTDMT_createCCtx_advanced(params.nbThreads, cctx->customMem);
cctx->mtctx = ZSTDMT_createCCtx_advanced(params.nbWorkers, cctx->customMem);
if (cctx->mtctx == NULL) return ERROR(memory_allocation);
}
DEBUGLOG(4, "call ZSTDMT_initCStream_internal as nbThreads=%u", params.nbThreads);
DEBUGLOG(4, "call ZSTDMT_initCStream_internal as nbWorkers=%u", params.nbWorkers);
CHECK_F( ZSTDMT_initCStream_internal(
cctx->mtctx,
prefixDict.dict, prefixDict.dictSize, ZSTD_dm_rawContent,
cctx->cdict, params, cctx->pledgedSrcSizePlusOne-1) );
cctx->streamStage = zcss_load;
cctx->appliedParams.nbThreads = params.nbThreads;
cctx->appliedParams.nonBlockingMode = params.nonBlockingMode;
cctx->appliedParams.nbWorkers = params.nbWorkers;
} else
#endif
{ CHECK_F( ZSTD_resetCStream_internal(
@@ -3195,12 +3183,12 @@ size_t ZSTD_compress_generic (ZSTD_CCtx* cctx,
prefixDict.dictMode, cctx->cdict, params,
cctx->pledgedSrcSizePlusOne-1) );
assert(cctx->streamStage == zcss_load);
assert(cctx->appliedParams.nbThreads <= 1);
assert(cctx->appliedParams.nbWorkers == 0);
} }
/* compression stage */
#ifdef ZSTD_MULTITHREAD
if ((cctx->appliedParams.nbThreads > 1) | (cctx->appliedParams.nonBlockingMode==1)) {
if (cctx->appliedParams.nbWorkers > 0) {
size_t const flushMin = ZSTDMT_compressStream_generic(cctx->mtctx, output, input, endOp);
if ( ZSTD_isError(flushMin)
|| (endOp == ZSTD_e_end && flushMin == 0) ) { /* compression completed */
+2 -3
View File
@@ -145,12 +145,11 @@ struct ZSTD_CCtx_params_s {
ZSTD_frameParameters fParams;
int compressionLevel;
U32 forceWindow; /* force back-references to respect limit of
int forceWindow; /* force back-references to respect limit of
* 1<<wLog, even for dictionary */
/* Multithreading: used to pass parameters to mtctx */
U32 nbThreads;
int nonBlockingMode; /* will trigger ZSTDMT even with nbThreads==1 */
unsigned nbWorkers;
unsigned jobSize;
unsigned overlapSizeLog;
+57 -54
View File
@@ -10,7 +10,7 @@
/* ====== Tuning parameters ====== */
#define ZSTDMT_NBTHREADS_MAX 200
#define ZSTDMT_NBWORKERS_MAX 200
#define ZSTDMT_JOBSIZE_MAX (MEM_32bits() ? (512 MB) : (2 GB)) /* note : limited by `jobSize` type, which is `unsigned` */
#define ZSTDMT_OVERLAPLOG_DEFAULT 6
@@ -97,9 +97,9 @@ typedef struct ZSTDMT_bufferPool_s {
buffer_t bTable[1]; /* variable size */
} ZSTDMT_bufferPool;
static ZSTDMT_bufferPool* ZSTDMT_createBufferPool(unsigned nbThreads, ZSTD_customMem cMem)
static ZSTDMT_bufferPool* ZSTDMT_createBufferPool(unsigned nbWorkers, ZSTD_customMem cMem)
{
unsigned const maxNbBuffers = 2*nbThreads + 3;
unsigned const maxNbBuffers = 2*nbWorkers + 3;
ZSTDMT_bufferPool* const bufPool = (ZSTDMT_bufferPool*)ZSTD_calloc(
sizeof(ZSTDMT_bufferPool) + (maxNbBuffers-1) * sizeof(buffer_t), cMem);
if (bufPool==NULL) return NULL;
@@ -236,23 +236,24 @@ static void ZSTDMT_freeCCtxPool(ZSTDMT_CCtxPool* pool)
}
/* ZSTDMT_createCCtxPool() :
* implies nbThreads >= 1 , checked by caller ZSTDMT_createCCtx() */
static ZSTDMT_CCtxPool* ZSTDMT_createCCtxPool(unsigned nbThreads,
* implies nbWorkers >= 1 , checked by caller ZSTDMT_createCCtx() */
static ZSTDMT_CCtxPool* ZSTDMT_createCCtxPool(unsigned nbWorkers,
ZSTD_customMem cMem)
{
ZSTDMT_CCtxPool* const cctxPool = (ZSTDMT_CCtxPool*) ZSTD_calloc(
sizeof(ZSTDMT_CCtxPool) + (nbThreads-1)*sizeof(ZSTD_CCtx*), cMem);
sizeof(ZSTDMT_CCtxPool) + (nbWorkers-1)*sizeof(ZSTD_CCtx*), cMem);
assert(nbWorkers > 0);
if (!cctxPool) return NULL;
if (ZSTD_pthread_mutex_init(&cctxPool->poolMutex, NULL)) {
ZSTD_free(cctxPool, cMem);
return NULL;
}
cctxPool->cMem = cMem;
cctxPool->totalCCtx = nbThreads;
cctxPool->totalCCtx = nbWorkers;
cctxPool->availCCtx = 1; /* at least one cctx for single-thread mode */
cctxPool->cctx[0] = ZSTD_createCCtx_advanced(cMem);
if (!cctxPool->cctx[0]) { ZSTDMT_freeCCtxPool(cctxPool); return NULL; }
DEBUGLOG(3, "cctxPool created, with %u threads", nbThreads);
DEBUGLOG(3, "cctxPool created, with %u workers", nbWorkers);
return cctxPool;
}
@@ -260,15 +261,16 @@ static ZSTDMT_CCtxPool* ZSTDMT_createCCtxPool(unsigned nbThreads,
static size_t ZSTDMT_sizeof_CCtxPool(ZSTDMT_CCtxPool* cctxPool)
{
ZSTD_pthread_mutex_lock(&cctxPool->poolMutex);
{ unsigned const nbThreads = cctxPool->totalCCtx;
{ unsigned const nbWorkers = cctxPool->totalCCtx;
size_t const poolSize = sizeof(*cctxPool)
+ (nbThreads-1)*sizeof(ZSTD_CCtx*);
+ (nbWorkers-1) * sizeof(ZSTD_CCtx*);
unsigned u;
size_t totalCCtxSize = 0;
for (u=0; u<nbThreads; u++) {
for (u=0; u<nbWorkers; u++) {
totalCCtxSize += ZSTD_sizeof_CCtx(cctxPool->cctx[u]);
}
ZSTD_pthread_mutex_unlock(&cctxPool->poolMutex);
assert(nbWorkers > 0);
return poolSize + totalCCtxSize;
}
}
@@ -295,8 +297,8 @@ static void ZSTDMT_releaseCCtx(ZSTDMT_CCtxPool* pool, ZSTD_CCtx* cctx)
if (pool->availCCtx < pool->totalCCtx)
pool->cctx[pool->availCCtx++] = cctx;
else {
/* pool overflow : should not happen, since totalCCtx==nbThreads */
DEBUGLOG(5, "CCtx pool overflow : free cctx");
/* pool overflow : should not happen, since totalCCtx==nbWorkers */
DEBUGLOG(4, "CCtx pool overflow : free cctx");
ZSTD_freeCCtx(cctx);
}
ZSTD_pthread_mutex_unlock(&pool->poolMutex);
@@ -502,52 +504,52 @@ static ZSTDMT_jobDescription* ZSTDMT_createJobsTable(U32* nbJobsPtr, ZSTD_custom
return jobTable;
}
/* ZSTDMT_CCtxParam_setNbThreads():
/* ZSTDMT_CCtxParam_setNbWorkers():
* Internal use only */
size_t ZSTDMT_CCtxParam_setNbThreads(ZSTD_CCtx_params* params, unsigned nbThreads)
size_t ZSTDMT_CCtxParam_setNbWorkers(ZSTD_CCtx_params* params, unsigned nbWorkers)
{
if (nbThreads > ZSTDMT_NBTHREADS_MAX) nbThreads = ZSTDMT_NBTHREADS_MAX;
if (nbThreads < 1) nbThreads = 1;
params->nbThreads = nbThreads;
if (nbWorkers > ZSTDMT_NBWORKERS_MAX) nbWorkers = ZSTDMT_NBWORKERS_MAX;
if (nbWorkers < 1) nbWorkers = 1;
params->nbWorkers = nbWorkers;
params->overlapSizeLog = ZSTDMT_OVERLAPLOG_DEFAULT;
params->jobSize = 0;
return nbThreads;
return nbWorkers;
}
ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced(unsigned nbThreads, ZSTD_customMem cMem)
ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced(unsigned nbWorkers, ZSTD_customMem cMem)
{
ZSTDMT_CCtx* mtctx;
U32 nbJobs = nbThreads + 2;
DEBUGLOG(3, "ZSTDMT_createCCtx_advanced (nbThreads = %u)", nbThreads);
U32 nbJobs = nbWorkers + 2;
DEBUGLOG(3, "ZSTDMT_createCCtx_advanced (nbWorkers = %u)", nbWorkers);
if (nbThreads < 1) return NULL;
nbThreads = MIN(nbThreads , ZSTDMT_NBTHREADS_MAX);
if (nbWorkers < 1) return NULL;
nbWorkers = MIN(nbWorkers , ZSTDMT_NBWORKERS_MAX);
if ((cMem.customAlloc!=NULL) ^ (cMem.customFree!=NULL))
/* invalid custom allocator */
return NULL;
mtctx = (ZSTDMT_CCtx*) ZSTD_calloc(sizeof(ZSTDMT_CCtx), cMem);
if (!mtctx) return NULL;
ZSTDMT_CCtxParam_setNbThreads(&mtctx->params, nbThreads);
ZSTDMT_CCtxParam_setNbWorkers(&mtctx->params, nbWorkers);
mtctx->cMem = cMem;
mtctx->allJobsCompleted = 1;
mtctx->factory = POOL_create_advanced(nbThreads, 0, cMem);
mtctx->factory = POOL_create_advanced(nbWorkers, 0, cMem);
mtctx->jobs = ZSTDMT_createJobsTable(&nbJobs, cMem);
assert(nbJobs > 0); assert((nbJobs & (nbJobs - 1)) == 0); /* ensure nbJobs is a power of 2 */
mtctx->jobIDMask = nbJobs - 1;
mtctx->bufPool = ZSTDMT_createBufferPool(nbThreads, cMem);
mtctx->cctxPool = ZSTDMT_createCCtxPool(nbThreads, cMem);
mtctx->bufPool = ZSTDMT_createBufferPool(nbWorkers, cMem);
mtctx->cctxPool = ZSTDMT_createCCtxPool(nbWorkers, cMem);
if (!mtctx->factory | !mtctx->jobs | !mtctx->bufPool | !mtctx->cctxPool) {
ZSTDMT_freeCCtx(mtctx);
return NULL;
}
DEBUGLOG(3, "mt_cctx created, for %u threads", nbThreads);
DEBUGLOG(3, "mt_cctx created, for %u threads", nbWorkers);
return mtctx;
}
ZSTDMT_CCtx* ZSTDMT_createCCtx(unsigned nbThreads)
ZSTDMT_CCtx* ZSTDMT_createCCtx(unsigned nbWorkers)
{
return ZSTDMT_createCCtx_advanced(nbThreads, ZSTD_defaultCMem);
return ZSTDMT_createCCtx_advanced(nbWorkers, ZSTD_defaultCMem);
}
@@ -649,8 +651,8 @@ size_t ZSTDMT_setMTCtxParameter(ZSTDMT_CCtx* mtctx, ZSTDMT_parameter parameter,
}
}
/* Sets parameters relevant to the compression job, initializing others to
* default values. Notably, nbThreads should probably be zero. */
/* Sets parameters relevant to the compression job,
* initializing others to default values. */
static ZSTD_CCtx_params ZSTDMT_initJobCCtxParams(ZSTD_CCtx_params const params)
{
ZSTD_CCtx_params jobParams;
@@ -664,7 +666,7 @@ static ZSTD_CCtx_params ZSTDMT_initJobCCtxParams(ZSTD_CCtx_params const params)
return jobParams;
}
/*! ZSTDMT_MTCtx_setParametersUsingCCtxParams() :
/*! ZSTDMT_MTCtx_setParametersUsingCCtxParams_whileCompressing() :
* Apply a ZSTD_CCtx_params to the compression context.
* This entry point is accessed while compression is ongoing,
* new parameters will be applied to next compression job.
@@ -675,21 +677,23 @@ static ZSTD_CCtx_params ZSTDMT_initJobCCtxParams(ZSTD_CCtx_params const params)
* - job size
* - overlap size
*/
void ZSTDMT_MTCtx_setParametersUsingCCtxParams(ZSTDMT_CCtx* mtctx, const ZSTD_CCtx_params* params)
void ZSTDMT_MTCtx_setParametersUsingCCtxParams_whileCompressing(ZSTDMT_CCtx* mtctx, const ZSTD_CCtx_params* params)
{
U32 const wlog = mtctx->params.cParams.windowLog;
U32 const nbWorkers = mtctx->params.nbWorkers;
mtctx->params = *params;
mtctx->params.cParams.windowLog = wlog; /* Do not modify windowLog ! */
mtctx->params.cParams.windowLog = wlog; /* Do not modify windowLog ! Frame must keep same wlog during the whole process ! */
mtctx->params.nbWorkers = nbWorkers; /* Do not modify nbWorkers, it must remain synchronized with CCtx Pool ! */
/* note : other parameters not updated are simply not used beyond initialization */
}
/* ZSTDMT_getNbThreads():
/* ZSTDMT_getNbWorkers():
* @return nb threads currently active in mtctx.
* mtctx must be valid */
unsigned ZSTDMT_getNbThreads(const ZSTDMT_CCtx* mtctx)
unsigned ZSTDMT_getNbWorkers(const ZSTDMT_CCtx* mtctx)
{
assert(mtctx != NULL);
return mtctx->params.nbThreads;
return mtctx->params.nbWorkers;
}
/* ZSTDMT_getFrameProgression():
@@ -728,15 +732,15 @@ ZSTD_frameProgression ZSTDMT_getFrameProgression(ZSTDMT_CCtx* mtctx)
/* ===== Multi-threaded compression ===== */
/* ------------------------------------------ */
static unsigned ZSTDMT_computeNbJobs(size_t srcSize, unsigned windowLog, unsigned nbThreads) {
assert(nbThreads>0);
static unsigned ZSTDMT_computeNbJobs(size_t srcSize, unsigned windowLog, unsigned nbWorkers) {
assert(nbWorkers>0);
{ size_t const jobSizeTarget = (size_t)1 << (windowLog + 2);
size_t const jobMaxSize = jobSizeTarget << 2;
size_t const passSizeMax = jobMaxSize * nbThreads;
size_t const passSizeMax = jobMaxSize * nbWorkers;
unsigned const multiplier = (unsigned)(srcSize / passSizeMax) + 1;
unsigned const nbJobsLarge = multiplier * nbThreads;
unsigned const nbJobsLarge = multiplier * nbWorkers;
unsigned const nbJobsMax = (unsigned)(srcSize / jobSizeTarget) + 1;
unsigned const nbJobsSmall = MIN(nbJobsMax, nbThreads);
unsigned const nbJobsSmall = MIN(nbJobsMax, nbWorkers);
return (multiplier>1) ? nbJobsLarge : nbJobsSmall;
} }
@@ -753,7 +757,7 @@ static size_t ZSTDMT_compress_advanced_internal(
ZSTD_CCtx_params const jobParams = ZSTDMT_initJobCCtxParams(params);
unsigned const overlapRLog = (params.overlapSizeLog>9) ? 0 : 9-params.overlapSizeLog;
size_t const overlapSize = (overlapRLog>=9) ? 0 : (size_t)1 << (params.cParams.windowLog - overlapRLog);
unsigned const nbJobs = ZSTDMT_computeNbJobs(srcSize, params.cParams.windowLog, params.nbThreads);
unsigned const nbJobs = ZSTDMT_computeNbJobs(srcSize, params.cParams.windowLog, params.nbWorkers);
size_t const proposedJobSize = (srcSize + (nbJobs-1)) / nbJobs;
size_t const avgJobSize = (((proposedJobSize-1) & 0x1FFFF) < 0x7FFF) ? proposedJobSize + 0xFFFF : proposedJobSize; /* avoid too small last block */
const char* const srcStart = (const char*)src;
@@ -761,13 +765,13 @@ static size_t ZSTDMT_compress_advanced_internal(
unsigned const compressWithinDst = (dstCapacity >= ZSTD_compressBound(srcSize)) ? nbJobs : (unsigned)(dstCapacity / ZSTD_compressBound(avgJobSize)); /* presumes avgJobSize >= 256 KB, which should be the case */
size_t frameStartPos = 0, dstBufferPos = 0;
XXH64_state_t xxh64;
assert(jobParams.nbThreads == 0);
assert(mtctx->cctxPool->totalCCtx == params.nbThreads);
assert(jobParams.nbWorkers == 0);
assert(mtctx->cctxPool->totalCCtx == params.nbWorkers);
DEBUGLOG(4, "ZSTDMT_compress_advanced_internal: nbJobs=%2u (rawSize=%u bytes; fixedSize=%u) ",
nbJobs, (U32)proposedJobSize, (U32)avgJobSize);
if ((nbJobs==1) | (params.nbThreads<=1)) { /* fallback to single-thread mode : this is a blocking invocation anyway */
if ((nbJobs==1) | (params.nbWorkers<=1)) { /* fallback to single-thread mode : this is a blocking invocation anyway */
ZSTD_CCtx* const cctx = mtctx->cctxPool->cctx[0];
if (cdict) return ZSTD_compress_usingCDict_advanced(cctx, dst, dstCapacity, src, srcSize, cdict, jobParams.fParams);
return ZSTD_compress_advanced_internal(cctx, dst, dstCapacity, src, srcSize, NULL, 0, jobParams);
@@ -911,11 +915,12 @@ size_t ZSTDMT_initCStream_internal(
const ZSTD_CDict* cdict, ZSTD_CCtx_params params,
unsigned long long pledgedSrcSize)
{
DEBUGLOG(4, "ZSTDMT_initCStream_internal (pledgedSrcSize=%u)", (U32)pledgedSrcSize);
DEBUGLOG(2, "ZSTDMT_initCStream_internal (pledgedSrcSize=%u, nbWorkers=%u, cctxPool=%u)",
(U32)pledgedSrcSize, params.nbWorkers, mtctx->cctxPool->totalCCtx);
/* 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 */
assert(mtctx->cctxPool->totalCCtx == params.nbThreads);
assert(mtctx->cctxPool->totalCCtx == params.nbWorkers);
mtctx->singleBlockingThread = (pledgedSrcSize <= ZSTDMT_JOBSIZE_MIN); /* do not trigger multi-threading when srcSize is too small */
if (params.jobSize == 0) {
if (params.cParams.windowLog >= 29)
@@ -928,12 +933,12 @@ size_t ZSTDMT_initCStream_internal(
if (mtctx->singleBlockingThread) {
ZSTD_CCtx_params const singleThreadParams = ZSTDMT_initJobCCtxParams(params);
DEBUGLOG(4, "ZSTDMT_initCStream_internal: switch to single blocking thread mode");
assert(singleThreadParams.nbThreads == 0);
assert(singleThreadParams.nbWorkers == 0);
return ZSTD_initCStream_internal(mtctx->cctxPool->cctx[0],
dict, dictSize, cdict,
singleThreadParams, pledgedSrcSize);
}
DEBUGLOG(4, "ZSTDMT_initCStream_internal: %u threads", params.nbThreads);
DEBUGLOG(4, "ZSTDMT_initCStream_internal: %u workers", params.nbWorkers);
if (mtctx->allJobsCompleted == 0) { /* previous compression not correctly finished */
ZSTDMT_waitForAllJobsCompleted(mtctx);
@@ -1012,8 +1017,6 @@ size_t ZSTDMT_initCStream_usingCDict(ZSTDMT_CCtx* mtctx,
size_t ZSTDMT_resetCStream(ZSTDMT_CCtx* mtctx, unsigned long long pledgedSrcSize)
{
if (!pledgedSrcSize) pledgedSrcSize = ZSTD_CONTENTSIZE_UNKNOWN;
if (mtctx->params.nbThreads==1)
return ZSTD_resetCStream(mtctx->cctxPool->cctx[0], pledgedSrcSize);
return ZSTDMT_initCStream_internal(mtctx, NULL, 0, ZSTD_dm_auto, 0, mtctx->params,
pledgedSrcSize);
}
+9 -9
View File
@@ -30,8 +30,8 @@
/* === Memory management === */
typedef struct ZSTDMT_CCtx_s ZSTDMT_CCtx;
ZSTDLIB_API ZSTDMT_CCtx* ZSTDMT_createCCtx(unsigned nbThreads);
ZSTDLIB_API ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced(unsigned nbThreads,
ZSTDLIB_API ZSTDMT_CCtx* ZSTDMT_createCCtx(unsigned nbWorkers);
ZSTDLIB_API ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced(unsigned nbWorkers,
ZSTD_customMem cMem);
ZSTDLIB_API size_t ZSTDMT_freeCCtx(ZSTDMT_CCtx* mtctx);
@@ -116,12 +116,12 @@ ZSTDLIB_API size_t ZSTDMT_compressStream_generic(ZSTDMT_CCtx* mtctx,
size_t ZSTDMT_CCtxParam_setMTCtxParameter(ZSTD_CCtx_params* params, ZSTDMT_parameter parameter, unsigned value);
/* ZSTDMT_CCtxParam_setNbThreads()
* Set nbThreads, and clamp it.
/* ZSTDMT_CCtxParam_setNbWorkers()
* Set nbWorkers, and clamp it.
* Also reset jobSize and overlapLog */
size_t ZSTDMT_CCtxParam_setNbThreads(ZSTD_CCtx_params* params, unsigned nbThreads);
size_t ZSTDMT_CCtxParam_setNbWorkers(ZSTD_CCtx_params* params, unsigned nbWorkers);
/*! ZSTDMT_MTCtx_setParametersUsingCCtxParams() :
/*! ZSTDMT_MTCtx_setParametersUsingCCtxParams_whileCompressing() :
* Apply a ZSTD_CCtx_params to the compression context.
* This works even during compression, and will be applied to next compression job.
* However, the following parameters will NOT be updated after compression has been started :
@@ -131,12 +131,12 @@ size_t ZSTDMT_CCtxParam_setNbThreads(ZSTD_CCtx_params* params, unsigned nbThread
* - job size
* - overlap size
*/
void ZSTDMT_MTCtx_setParametersUsingCCtxParams(ZSTDMT_CCtx* mtctx, const ZSTD_CCtx_params* params);
void ZSTDMT_MTCtx_setParametersUsingCCtxParams_whileCompressing(ZSTDMT_CCtx* mtctx, const ZSTD_CCtx_params* params);
/* ZSTDMT_getNbThreads():
/* ZSTDMT_getNbWorkers():
* @return nb threads currently active in mtctx.
* mtctx must be valid */
unsigned ZSTDMT_getNbThreads(const ZSTDMT_CCtx* mtctx);
unsigned ZSTDMT_getNbWorkers(const ZSTDMT_CCtx* mtctx);
/* ZSTDMT_getFrameProgression():
* tells how much data has been consumed (input) and produced (output) for current frame.
+13 -17
View File
@@ -506,7 +506,7 @@ ZSTDLIB_API size_t ZSTD_sizeof_DDict(const ZSTD_DDict* ddict);
* It will also consider src size to be arbitrarily "large", which is worst case.
* If srcSize is known to always be small, ZSTD_estimateCCtxSize_usingCParams() can provide a tighter estimation.
* ZSTD_estimateCCtxSize_usingCParams() can be used in tandem with ZSTD_getCParams() to create cParams from compressionLevel.
* ZSTD_estimateCCtxSize_usingCCtxParams() can be used in tandem with ZSTD_CCtxParam_setParameter(). Only single-threaded compression is supported. This function will return an error code if ZSTD_p_nbThreads is > 1.
* ZSTD_estimateCCtxSize_usingCCtxParams() can be used in tandem with ZSTD_CCtxParam_setParameter(). Only single-threaded compression is supported. This function will return an error code if ZSTD_p_nbWorkers is >= 1.
* Note : CCtx size estimation is only correct for single-threaded compression. */
ZSTDLIB_API size_t ZSTD_estimateCCtxSize(int compressionLevel);
ZSTDLIB_API size_t ZSTD_estimateCCtxSize_usingCParams(ZSTD_compressionParameters cParams);
@@ -518,7 +518,7 @@ ZSTDLIB_API size_t ZSTD_estimateDCtxSize(void);
* It will also consider src size to be arbitrarily "large", which is worst case.
* If srcSize is known to always be small, ZSTD_estimateCStreamSize_usingCParams() can provide a tighter estimation.
* ZSTD_estimateCStreamSize_usingCParams() can be used in tandem with ZSTD_getCParams() to create cParams from compressionLevel.
* ZSTD_estimateCStreamSize_usingCCtxParams() can be used in tandem with ZSTD_CCtxParam_setParameter(). Only single-threaded compression is supported. This function will return an error code if ZSTD_p_nbThreads is set to a value > 1.
* ZSTD_estimateCStreamSize_usingCCtxParams() can be used in tandem with ZSTD_CCtxParam_setParameter(). Only single-threaded compression is supported. This function will return an error code if ZSTD_p_nbWorkers is >= 1.
* Note : CStream size estimation is only correct for single-threaded compression.
* ZSTD_DStream memory budget depends on window Size.
* This information can be passed manually, using ZSTD_estimateDStreamSize,
@@ -992,18 +992,13 @@ typedef enum {
/* multi-threading parameters */
/* These parameters are only useful if multi-threading is enabled (ZSTD_MULTITHREAD).
* They return an error otherwise. */
ZSTD_p_nbThreads=400, /* Select how many threads a compression job can spawn (default:1)
* More threads improve speed, but also increase memory usage.
* Can only receive a value > 1 if ZSTD_MULTITHREAD is enabled.
* Special: value 0 means "do not change nbThreads" */
ZSTD_p_nonBlockingMode, /* Single thread mode is by default "blocking" :
* it finishes its job as much as possible, and only then gives back control to caller.
* In contrast, multi-thread is by default "non-blocking" :
* it takes some input, flush some output if available, and immediately gives back control to caller.
* Compression work is performed in parallel, within worker threads.
* (note : a strong exception to this rule is when first job is called with ZSTD_e_end : it becomes blocking)
* Setting this parameter to 1 will enforce non-blocking mode even when only 1 thread is selected.
* It allows the caller to do other tasks while the worker thread compresses in parallel. */
ZSTD_p_nbWorkers=400, /* Select how many threads will be spawned to compress in parallel.
* When nbWorkers >= 1, triggers asynchronous mode :
* ZSTD_compress_generic() consumes some input, flush some output if possible, and immediately gives back control to caller,
* while compression work is performed in parallel, within worker threads.
* (note : a strong exception to this rule is when first invocation sets ZSTD_e_end : it becomes a blocking call).
* More workers improve speed, but also increase memory usage.
* Default value is `0`, aka "single-threaded mode" : no worker is spawned, compression is performed inside Caller's thread, all invocations are blocking */
ZSTD_p_jobSize, /* Size of a compression job. This value is only enforced in streaming (non-blocking) mode.
* Each compression job is completed in parallel, so indirectly controls the nb of active threads.
* 0 means default, which is dynamically determined based on compression parameters.
@@ -1231,9 +1226,10 @@ ZSTDLIB_API size_t ZSTD_CCtxParam_setParameter(ZSTD_CCtx_params* params, ZSTD_cP
/*! ZSTD_CCtx_setParametersUsingCCtxParams() :
* Apply a set of ZSTD_CCtx_params to the compression context.
* This must be done before the dictionary is loaded.
* The pledgedSrcSize is treated as unknown.
* Multithreading parameters are applied only if nbThreads > 1.
* This can be done even after compression is started,
* if nbWorkers==0, this will have no impact until a new compression is started.
* if nbWorkers>=1, new parameters will be picked up at next job,
* with a few restrictions (windowLog, pledgedSrcSize, nbWorkers, jobSize, and overlapLog are not updated).
*/
ZSTDLIB_API size_t ZSTD_CCtx_setParametersUsingCCtxParams(
ZSTD_CCtx* cctx, const ZSTD_CCtx_params* params);