zstd-ruby 1.2.0.0 → 1.3.0.0

data/ext/zstdruby/libzstd/compress/zstd_opt.h

@@ -43,6 +43,7 @@ MEM_STATIC void ZSTD_rescaleFreqs(seqStore_t* ssPtr, const BYTE* src, size_t src
     if (ssPtr->litLengthSum == 0) {
         if (srcSize <= 1024) ssPtr->staticPrices = 1;
 
+        assert(ssPtr->litFreq!=NULL);
         for (u=0; u<=MaxLit; u++)
             ssPtr->litFreq[u] = 0;
         for (u=0; u<srcSize; u++)
@@ -201,6 +202,20 @@ MEM_STATIC void ZSTD_updatePrice(seqStore_t* seqStorePtr, U32 litLength, const B
     }
 
 
+/* function safe only for comparisons */
+MEM_STATIC U32 ZSTD_readMINMATCH(const void* memPtr, U32 length)
+{
+    switch (length)
+    {
+    default :
+    case 4 : return MEM_read32(memPtr);
+    case 3 : if (MEM_isLittleEndian())
+                return MEM_read32(memPtr)<<8;
+             else
+                return MEM_read32(memPtr)>>8;
+    }
+}
+
 
 /* Update hashTable3 up to ip (excluded)
    Assumption : always within prefix (i.e. not within extDict) */
@@ -234,12 +249,12 @@ static U32 ZSTD_insertBtAndGetAllMatches (
 {
     const BYTE* const base = zc->base;
     const U32 current = (U32)(ip-base);
-    const U32 hashLog = zc->params.cParams.hashLog;
+    const U32 hashLog = zc->appliedParams.cParams.hashLog;
     const size_t h  = ZSTD_hashPtr(ip, hashLog, mls);
     U32* const hashTable = zc->hashTable;
     U32 matchIndex  = hashTable[h];
     U32* const bt   = zc->chainTable;
-    const U32 btLog = zc->params.cParams.chainLog - 1;
+    const U32 btLog = zc->appliedParams.cParams.chainLog - 1;
     const U32 btMask= (1U << btLog) - 1;
     size_t commonLengthSmaller=0, commonLengthLarger=0;
     const BYTE* const dictBase = zc->dictBase;
@@ -267,7 +282,7 @@ static U32 ZSTD_insertBtAndGetAllMatches (
                 if (match[bestLength] == ip[bestLength]) currentMl = ZSTD_count(ip, match, iLimit);
             } else {
                 match = dictBase + matchIndex3;
-                if (MEM_readMINMATCH(match, MINMATCH) == MEM_readMINMATCH(ip, MINMATCH))    /* assumption : matchIndex3 <= dictLimit-4 (by table construction) */
+                if (ZSTD_readMINMATCH(match, MINMATCH) == ZSTD_readMINMATCH(ip, MINMATCH))    /* assumption : matchIndex3 <= dictLimit-4 (by table construction) */
                     currentMl = ZSTD_count_2segments(ip+MINMATCH, match+MINMATCH, iLimit, dictEnd, prefixStart) + MINMATCH;
             }
 
@@ -410,10 +425,10 @@ void ZSTD_compressBlock_opt_generic(ZSTD_CCtx* ctx,
     const BYTE* const base = ctx->base;
     const BYTE* const prefixStart = base + ctx->dictLimit;
 
-    const U32 maxSearches = 1U << ctx->params.cParams.searchLog;
-    const U32 sufficient_len = ctx->params.cParams.targetLength;
-    const U32 mls = ctx->params.cParams.searchLength;
-    const U32 minMatch = (ctx->params.cParams.searchLength == 3) ? 3 : 4;
+    const U32 maxSearches = 1U << ctx->appliedParams.cParams.searchLog;
+    const U32 sufficient_len = ctx->appliedParams.cParams.targetLength;
+    const U32 mls = ctx->appliedParams.cParams.searchLength;
+    const U32 minMatch = (ctx->appliedParams.cParams.searchLength == 3) ? 3 : 4;
 
     ZSTD_optimal_t* opt = seqStorePtr->priceTable;
     ZSTD_match_t* matches = seqStorePtr->matchTable;
@@ -439,7 +454,7 @@ void ZSTD_compressBlock_opt_generic(ZSTD_CCtx* ctx,
             for (i=(ip == anchor); i<last_i; i++) {
                 const S32 repCur = (i==ZSTD_REP_MOVE_OPT) ? (rep[0] - 1) : rep[i];
                 if ( (repCur > 0) && (repCur < (S32)(ip-prefixStart))
-                    && (MEM_readMINMATCH(ip, minMatch) == MEM_readMINMATCH(ip - repCur, minMatch))) {
+                    && (ZSTD_readMINMATCH(ip, minMatch) == ZSTD_readMINMATCH(ip - repCur, minMatch))) {
                     mlen = (U32)ZSTD_count(ip+minMatch, ip+minMatch-repCur, iend) + minMatch;
                     if (mlen > sufficient_len || mlen >= ZSTD_OPT_NUM) {
                         best_mlen = mlen; best_off = i; cur = 0; last_pos = 1;
@@ -524,7 +539,7 @@ void ZSTD_compressBlock_opt_generic(ZSTD_CCtx* ctx,
                 for (i=(opt[cur].mlen != 1); i<last_i; i++) {  /* check rep */
                     const S32 repCur = (i==ZSTD_REP_MOVE_OPT) ? (opt[cur].rep[0] - 1) : opt[cur].rep[i];
                     if ( (repCur > 0) && (repCur < (S32)(inr-prefixStart))
-                       && (MEM_readMINMATCH(inr, minMatch) == MEM_readMINMATCH(inr - repCur, minMatch))) {
+                       && (ZSTD_readMINMATCH(inr, minMatch) == ZSTD_readMINMATCH(inr - repCur, minMatch))) {
                        mlen = (U32)ZSTD_count(inr+minMatch, inr+minMatch - repCur, iend) + minMatch;
 
                        if (mlen > sufficient_len || cur + mlen >= ZSTD_OPT_NUM) {
@@ -663,10 +678,10 @@ void ZSTD_compressBlock_opt_extDict_generic(ZSTD_CCtx* ctx,
     const BYTE* const dictBase = ctx->dictBase;
     const BYTE* const dictEnd  = dictBase + dictLimit;
 
-    const U32 maxSearches = 1U << ctx->params.cParams.searchLog;
-    const U32 sufficient_len = ctx->params.cParams.targetLength;
-    const U32 mls = ctx->params.cParams.searchLength;
-    const U32 minMatch = (ctx->params.cParams.searchLength == 3) ? 3 : 4;
+    const U32 maxSearches = 1U << ctx->appliedParams.cParams.searchLog;
+    const U32 sufficient_len = ctx->appliedParams.cParams.targetLength;
+    const U32 mls = ctx->appliedParams.cParams.searchLength;
+    const U32 minMatch = (ctx->appliedParams.cParams.searchLength == 3) ? 3 : 4;
 
     ZSTD_optimal_t* opt = seqStorePtr->priceTable;
     ZSTD_match_t* matches = seqStorePtr->matchTable;
@@ -698,7 +713,7 @@ void ZSTD_compressBlock_opt_extDict_generic(ZSTD_CCtx* ctx,
                 const BYTE* const repMatch = repBase + repIndex;
                 if ( (repCur > 0 && repCur <= (S32)current)
                    && (((U32)((dictLimit-1) - repIndex) >= 3) & (repIndex>lowestIndex))  /* intentional overflow */
-                   && (MEM_readMINMATCH(ip, minMatch) == MEM_readMINMATCH(repMatch, minMatch)) ) {
+                   && (ZSTD_readMINMATCH(ip, minMatch) == ZSTD_readMINMATCH(repMatch, minMatch)) ) {
                     /* repcode detected we should take it */
                     const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
                     mlen = (U32)ZSTD_count_2segments(ip+minMatch, repMatch+minMatch, iend, repEnd, prefixStart) + minMatch;
@@ -794,7 +809,7 @@ void ZSTD_compressBlock_opt_extDict_generic(ZSTD_CCtx* ctx,
                     const BYTE* const repMatch = repBase + repIndex;
                     if ( (repCur > 0 && repCur <= (S32)(current+cur))
                       && (((U32)((dictLimit-1) - repIndex) >= 3) & (repIndex>lowestIndex))  /* intentional overflow */
-                      && (MEM_readMINMATCH(inr, minMatch) == MEM_readMINMATCH(repMatch, minMatch)) ) {
+                      && (ZSTD_readMINMATCH(inr, minMatch) == ZSTD_readMINMATCH(repMatch, minMatch)) ) {
                         /* repcode detected */
                         const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
                         mlen = (U32)ZSTD_count_2segments(inr+minMatch, repMatch+minMatch, iend, repEnd, prefixStart) + minMatch;

data/ext/zstdruby/libzstd/compress/zstdmt_compress.c

@@ -14,34 +14,31 @@
 
 /* ======   Compiler specifics   ====== */
 #if defined(_MSC_VER)
-#  pragma warning(disable : 4204)        /* disable: C4204: non-constant aggregate initializer */
+#  pragma warning(disable : 4204)   /* disable: C4204: non-constant aggregate initializer */
 #endif
 
 
 /* ======   Dependencies   ====== */
-#include <stdlib.h>   /* malloc */
-#include <string.h>   /* memcpy */
-#include "pool.h"     /* threadpool */
-#include "threading.h"  /* mutex */
-#include "zstd_internal.h"   /* MIN, ERROR, ZSTD_*, ZSTD_highbit32 */
+#include <string.h>      /* memcpy, memset */
+#include "pool.h"        /* threadpool */
+#include "threading.h"   /* mutex */
+#include "zstd_internal.h"  /* MIN, ERROR, ZSTD_*, ZSTD_highbit32 */
 #include "zstdmt_compress.h"
 
 
 /* ======   Debug   ====== */
-#if 0
+#if defined(ZSTD_DEBUG) && (ZSTD_DEBUG>=2)
 
 #  include <stdio.h>
 #  include <unistd.h>
 #  include <sys/times.h>
-   static unsigned g_debugLevel = 5;
-#  define DEBUGLOGRAW(l, ...) if (l<=g_debugLevel) { fprintf(stderr, __VA_ARGS__); }
-#  define DEBUGLOG(l, ...) if (l<=g_debugLevel) { fprintf(stderr, __FILE__ ": "); fprintf(stderr, __VA_ARGS__); fprintf(stderr, " \n"); }
+#  define DEBUGLOGRAW(l, ...) if (l<=ZSTD_DEBUG) { fprintf(stderr, __VA_ARGS__); }
 
-#  define DEBUG_PRINTHEX(l,p,n) { \
-    unsigned debug_u;                   \
-    for (debug_u=0; debug_u<(n); debug_u++)           \
+#  define DEBUG_PRINTHEX(l,p,n) {            \
+    unsigned debug_u;                        \
+    for (debug_u=0; debug_u<(n); debug_u++)  \
         DEBUGLOGRAW(l, "%02X ", ((const unsigned char*)(p))[debug_u]); \
-    DEBUGLOGRAW(l, " \n");       \
+    DEBUGLOGRAW(l, " \n");                   \
 }
 
 static unsigned long long GetCurrentClockTimeMicroseconds(void)
@@ -53,22 +50,22 @@ static unsigned long long GetCurrentClockTimeMicroseconds(void)
      return ((((unsigned long long)newTicks)*(1000000))/_ticksPerSecond); }
 }
 
-#define MUTEX_WAIT_TIME_DLEVEL 5
-#define PTHREAD_MUTEX_LOCK(mutex) \
-if (g_debugLevel>=MUTEX_WAIT_TIME_DLEVEL) { \
-    unsigned long long const beforeTime = GetCurrentClockTimeMicroseconds(); \
-    pthread_mutex_lock(mutex); \
-    {   unsigned long long const afterTime = GetCurrentClockTimeMicroseconds(); \
-        unsigned long long const elapsedTime = (afterTime-beforeTime); \
-        if (elapsedTime > 1000) {  /* or whatever threshold you like; I'm using 1 millisecond here */ \
-            DEBUGLOG(MUTEX_WAIT_TIME_DLEVEL, "Thread took %llu microseconds to acquire mutex %s \n", \
-               elapsedTime, #mutex); \
-    }   } \
-} else pthread_mutex_lock(mutex);
+#define MUTEX_WAIT_TIME_DLEVEL 6
+#define PTHREAD_MUTEX_LOCK(mutex) {               \
+    if (ZSTD_DEBUG>=MUTEX_WAIT_TIME_DLEVEL) {   \
+        unsigned long long const beforeTime = GetCurrentClockTimeMicroseconds(); \
+        pthread_mutex_lock(mutex);                \
+        {   unsigned long long const afterTime = GetCurrentClockTimeMicroseconds(); \
+            unsigned long long const elapsedTime = (afterTime-beforeTime); \
+            if (elapsedTime > 1000) {  /* or whatever threshold you like; I'm using 1 millisecond here */ \
+                DEBUGLOG(MUTEX_WAIT_TIME_DLEVEL, "Thread took %llu microseconds to acquire mutex %s \n", \
+                   elapsedTime, #mutex);          \
+        }   }                                     \
+    } else pthread_mutex_lock(mutex);             \
+}
 
 #else
 
-#  define DEBUGLOG(l, ...)      {}    /* disabled */
 #  define PTHREAD_MUTEX_LOCK(m) pthread_mutex_lock(m)
 #  define DEBUG_PRINTHEX(l,p,n) {}
 
@@ -87,16 +84,19 @@ static const buffer_t g_nullBuffer = { NULL, 0 };
 typedef struct ZSTDMT_bufferPool_s {
     unsigned totalBuffers;
     unsigned nbBuffers;
+    ZSTD_customMem cMem;
     buffer_t bTable[1];   /* variable size */
 } ZSTDMT_bufferPool;
 
-static ZSTDMT_bufferPool* ZSTDMT_createBufferPool(unsigned nbThreads)
+static ZSTDMT_bufferPool* ZSTDMT_createBufferPool(unsigned nbThreads, ZSTD_customMem cMem)
 {
     unsigned const maxNbBuffers = 2*nbThreads + 2;
-    ZSTDMT_bufferPool* const bufPool = (ZSTDMT_bufferPool*)calloc(1, sizeof(ZSTDMT_bufferPool) + (maxNbBuffers-1) * sizeof(buffer_t));
+    ZSTDMT_bufferPool* const bufPool = (ZSTDMT_bufferPool*)ZSTD_calloc(
+        sizeof(ZSTDMT_bufferPool) + (maxNbBuffers-1) * sizeof(buffer_t), cMem);
     if (bufPool==NULL) return NULL;
     bufPool->totalBuffers = maxNbBuffers;
     bufPool->nbBuffers = 0;
+    bufPool->cMem = cMem;
     return bufPool;
 }
 
@@ -105,23 +105,39 @@ static void ZSTDMT_freeBufferPool(ZSTDMT_bufferPool* bufPool)
     unsigned u;
     if (!bufPool) return;   /* compatibility with free on NULL */
     for (u=0; u<bufPool->totalBuffers; u++)
-        free(bufPool->bTable[u].start);
-    free(bufPool);
+        ZSTD_free(bufPool->bTable[u].start, bufPool->cMem);
+    ZSTD_free(bufPool, bufPool->cMem);
+}
+
+/* only works at initialization, not during compression */
+static size_t ZSTDMT_sizeof_bufferPool(ZSTDMT_bufferPool* bufPool)
+{
+    size_t const poolSize = sizeof(*bufPool)
+                            + (bufPool->totalBuffers - 1) * sizeof(buffer_t);
+    unsigned u;
+    size_t totalBufferSize = 0;
+    for (u=0; u<bufPool->totalBuffers; u++)
+        totalBufferSize += bufPool->bTable[u].size;
+
+    return poolSize + totalBufferSize;
 }
 
-/* assumption : invocation from main thread only ! */
+/** ZSTDMT_getBuffer() :
+ *  assumption : invocation from main thread only ! */
 static buffer_t ZSTDMT_getBuffer(ZSTDMT_bufferPool* pool, size_t bSize)
 {
     if (pool->nbBuffers) {   /* try to use an existing buffer */
         buffer_t const buf = pool->bTable[--(pool->nbBuffers)];
         size_t const availBufferSize = buf.size;
-        if ((availBufferSize >= bSize) & (availBufferSize <= 10*bSize))   /* large enough, but not too much */
+        if ((availBufferSize >= bSize) & (availBufferSize <= 10*bSize))
+            /* large enough, but not too much */
             return buf;
-        free(buf.start);   /* size conditions not respected : scratch this buffer and create a new one */
+        /* size conditions not respected : scratch this buffer, create new one */
+        ZSTD_free(buf.start, pool->cMem);
     }
     /* create new buffer */
     {   buffer_t buffer;
-        void* const start = malloc(bSize);
+        void* const start = ZSTD_malloc(bSize, pool->cMem);
         if (start==NULL) bSize = 0;
         buffer.start = start;   /* note : start can be NULL if malloc fails ! */
         buffer.size = bSize;
@@ -138,7 +154,7 @@ static void ZSTDMT_releaseBuffer(ZSTDMT_bufferPool* pool, buffer_t buf)
         return;
     }
     /* Reached bufferPool capacity (should not happen) */
-    free(buf.start);
+    ZSTD_free(buf.start, pool->cMem);
 }
 
 
@@ -147,6 +163,7 @@ static void ZSTDMT_releaseBuffer(ZSTDMT_bufferPool* pool, buffer_t buf)
 typedef struct {
     unsigned totalCCtx;
     unsigned availCCtx;
+    ZSTD_customMem cMem;
     ZSTD_CCtx* cctx[1];   /* variable size */
 } ZSTDMT_CCtxPool;
 
@@ -158,23 +175,40 @@ static void ZSTDMT_freeCCtxPool(ZSTDMT_CCtxPool* pool)
     unsigned u;
     for (u=0; u<pool->totalCCtx; u++)
         ZSTD_freeCCtx(pool->cctx[u]);  /* note : compatible with free on NULL */
-    free(pool);
+    ZSTD_free(pool, pool->cMem);
 }
 
 /* ZSTDMT_createCCtxPool() :
  * implies nbThreads >= 1 , checked by caller ZSTDMT_createCCtx() */
-static ZSTDMT_CCtxPool* ZSTDMT_createCCtxPool(unsigned nbThreads)
+static ZSTDMT_CCtxPool* ZSTDMT_createCCtxPool(unsigned nbThreads,
+                                              ZSTD_customMem cMem)
 {
-    ZSTDMT_CCtxPool* const cctxPool = (ZSTDMT_CCtxPool*) calloc(1, sizeof(ZSTDMT_CCtxPool) + (nbThreads-1)*sizeof(ZSTD_CCtx*));
+    ZSTDMT_CCtxPool* const cctxPool = (ZSTDMT_CCtxPool*) ZSTD_calloc(
+        sizeof(ZSTDMT_CCtxPool) + (nbThreads-1)*sizeof(ZSTD_CCtx*), cMem);
     if (!cctxPool) return NULL;
+    cctxPool->cMem = cMem;
     cctxPool->totalCCtx = nbThreads;
     cctxPool->availCCtx = 1;   /* at least one cctx for single-thread mode */
-    cctxPool->cctx[0] = ZSTD_createCCtx();
+    cctxPool->cctx[0] = ZSTD_createCCtx_advanced(cMem);
     if (!cctxPool->cctx[0]) { ZSTDMT_freeCCtxPool(cctxPool); return NULL; }
-    DEBUGLOG(1, "cctxPool created, with %u threads", nbThreads);
+    DEBUGLOG(3, "cctxPool created, with %u threads", nbThreads);
     return cctxPool;
 }
 
+/* only works during initialization phase, not during compression */
+static size_t ZSTDMT_sizeof_CCtxPool(ZSTDMT_CCtxPool* cctxPool)
+{
+    unsigned const nbThreads = cctxPool->totalCCtx;
+    size_t const poolSize = sizeof(*cctxPool)
+                            + (nbThreads-1)*sizeof(ZSTD_CCtx*);
+    unsigned u;
+    size_t totalCCtxSize = 0;
+    for (u=0; u<nbThreads; u++)
+        totalCCtxSize += ZSTD_sizeof_CCtx(cctxPool->cctx[u]);
+
+    return poolSize + totalCCtxSize;
+}
+
 static ZSTD_CCtx* ZSTDMT_getCCtx(ZSTDMT_CCtxPool* pool)
 {
     if (pool->availCCtx) {
@@ -218,7 +252,7 @@ typedef struct {
     pthread_mutex_t* jobCompleted_mutex;
     pthread_cond_t* jobCompleted_cond;
     ZSTD_parameters params;
-    ZSTD_CDict* cdict;
+    const ZSTD_CDict* cdict;
     unsigned long long fullFrameSize;
 } ZSTDMT_jobDescription;
 
@@ -228,11 +262,11 @@ 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(3, "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);
-        if (job->cdict) DEBUGLOG(3, "using CDict ");
+        DEBUGLOG(5, "using CDict");
         if (ZSTD_isError(initError)) { job->cSize = initError; goto _endJob; }
     } else {  /* srcStart points at reloaded section */
         if (!job->firstChunk) job->params.fParams.contentSizeFlag = 0;  /* ensure no srcSize control */
@@ -247,12 +281,12 @@ void ZSTDMT_compressChunk(void* jobDescription)
         ZSTD_invalidateRepCodes(job->cctx);
     }
 
-    DEBUGLOG(4, "Compressing : ");
+    DEBUGLOG(5, "Compressing : ");
     DEBUG_PRINTHEX(4, job->srcStart, 12);
     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(3, "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));
 
@@ -271,6 +305,7 @@ _endJob:
 
 struct ZSTDMT_CCtx_s {
     POOL_ctx* factory;
+    ZSTDMT_jobDescription* jobs;
     ZSTDMT_bufferPool* buffPool;
     ZSTDMT_CCtxPool* cctxPool;
     pthread_mutex_t jobCompleted_mutex;
@@ -292,50 +327,64 @@ struct ZSTDMT_CCtx_s {
     unsigned overlapRLog;
     unsigned long long frameContentSize;
     size_t sectionSize;
-    ZSTD_CDict* cdict;
-    ZSTD_CStream* cstream;
-    ZSTDMT_jobDescription jobs[1];   /* variable size (must lies at the end) */
+    ZSTD_customMem cMem;
+    ZSTD_CDict* cdictLocal;
+    const ZSTD_CDict* cdict;
 };
 
-ZSTDMT_CCtx *ZSTDMT_createCCtx(unsigned nbThreads)
+static ZSTDMT_jobDescription* ZSTDMT_allocJobsTable(U32* nbJobsPtr, ZSTD_customMem cMem)
 {
-    ZSTDMT_CCtx* cctx;
-    U32 const minNbJobs = nbThreads + 2;
-    U32 const nbJobsLog2 = ZSTD_highbit32(minNbJobs) + 1;
+    U32 const nbJobsLog2 = ZSTD_highbit32(*nbJobsPtr) + 1;
     U32 const nbJobs = 1 << nbJobsLog2;
-    DEBUGLOG(5, "nbThreads : %u  ; minNbJobs : %u ;  nbJobsLog2 : %u ;  nbJobs : %u  \n",
-            nbThreads, minNbJobs, nbJobsLog2, nbJobs);
+    *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 nbJobs = nbThreads + 2;
+    DEBUGLOG(3, "ZSTDMT_createCCtx_advanced");
+
     if ((nbThreads < 1) | (nbThreads > ZSTDMT_NBTHREADS_MAX)) return NULL;
-    cctx = (ZSTDMT_CCtx*) calloc(1, sizeof(ZSTDMT_CCtx) + nbJobs*sizeof(ZSTDMT_jobDescription));
-    if (!cctx) return NULL;
-    cctx->nbThreads = nbThreads;
-    cctx->jobIDMask = nbJobs - 1;
-    cctx->allJobsCompleted = 1;
-    cctx->sectionSize = 0;
-    cctx->overlapRLog = 3;
-    cctx->factory = POOL_create(nbThreads, 1);
-    cctx->buffPool = ZSTDMT_createBufferPool(nbThreads);
-    cctx->cctxPool = ZSTDMT_createCCtxPool(nbThreads);
-    if (!cctx->factory | !cctx->buffPool | !cctx->cctxPool) {  /* one object was not created */
-        ZSTDMT_freeCCtx(cctx);
+    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;
+    mtctx->cMem = cMem;
+    mtctx->nbThreads = nbThreads;
+    mtctx->allJobsCompleted = 1;
+    mtctx->sectionSize = 0;
+    mtctx->overlapRLog = 3;
+    mtctx->factory = POOL_create(nbThreads, 1);
+    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) {
+        ZSTDMT_freeCCtx(mtctx);
         return NULL;
     }
-    if (nbThreads==1) {
-        cctx->cstream = ZSTD_createCStream();
-        if (!cctx->cstream) {
-            ZSTDMT_freeCCtx(cctx); return NULL;
-    }   }
-    pthread_mutex_init(&cctx->jobCompleted_mutex, NULL);   /* Todo : check init function return */
-    pthread_cond_init(&cctx->jobCompleted_cond, NULL);
-    DEBUGLOG(4, "mt_cctx created, for %u threads \n", nbThreads);
-    return cctx;
+    pthread_mutex_init(&mtctx->jobCompleted_mutex, NULL);   /* Todo : check init function return */
+    pthread_cond_init(&mtctx->jobCompleted_cond, NULL);
+    DEBUGLOG(3, "mt_cctx created, for %u threads", nbThreads);
+    return mtctx;
+}
+
+ZSTDMT_CCtx* ZSTDMT_createCCtx(unsigned nbThreads)
+{
+    return ZSTDMT_createCCtx_advanced(nbThreads, ZSTD_defaultCMem);
 }
 
 /* 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(3, "ZSTDMT_releaseAllJobResources");
     for (jobID=0; jobID <= mtctx->jobIDMask; jobID++) {
         ZSTDMT_releaseBuffer(mtctx->buffPool, mtctx->jobs[jobID].dstBuff);
         mtctx->jobs[jobID].dstBuff = g_nullBuffer;
@@ -356,15 +405,26 @@ size_t ZSTDMT_freeCCtx(ZSTDMT_CCtx* mtctx)
     POOL_free(mtctx->factory);
     if (!mtctx->allJobsCompleted) ZSTDMT_releaseAllJobResources(mtctx); /* stop workers first */
     ZSTDMT_freeBufferPool(mtctx->buffPool);  /* release job resources into pools first */
+    ZSTD_free(mtctx->jobs, mtctx->cMem);
     ZSTDMT_freeCCtxPool(mtctx->cctxPool);
-    ZSTD_freeCDict(mtctx->cdict);
-    ZSTD_freeCStream(mtctx->cstream);
+    ZSTD_freeCDict(mtctx->cdictLocal);
     pthread_mutex_destroy(&mtctx->jobCompleted_mutex);
     pthread_cond_destroy(&mtctx->jobCompleted_cond);
-    free(mtctx);
+    ZSTD_free(mtctx, mtctx->cMem);
     return 0;
 }
 
+size_t ZSTDMT_sizeof_CCtx(ZSTDMT_CCtx* mtctx)
+{
+    if (mtctx == NULL) return 0;   /* supports sizeof NULL */
+    return sizeof(*mtctx)
+        + POOL_sizeof(mtctx->factory)
+        + ZSTDMT_sizeof_bufferPool(mtctx->buffPool)
+        + (mtctx->jobIDMask+1) * sizeof(ZSTDMT_jobDescription)
+        + ZSTDMT_sizeof_CCtxPool(mtctx->cctxPool)
+        + ZSTD_sizeof_CDict(mtctx->cdictLocal);
+}
+
 size_t ZSTDMT_setMTCtxParameter(ZSTDMT_CCtx* mtctx, ZSDTMT_parameter parameter, unsigned value)
 {
     switch(parameter)
@@ -373,7 +433,7 @@ size_t ZSTDMT_setMTCtxParameter(ZSTDMT_CCtx* mtctx, ZSDTMT_parameter parameter,
         mtctx->sectionSize = value;
         return 0;
     case ZSTDMT_p_overlapSectionLog :
-    DEBUGLOG(4, "ZSTDMT_p_overlapSectionLog : %u", value);
+        DEBUGLOG(5, "ZSTDMT_p_overlapSectionLog : %u", value);
         mtctx->overlapRLog = (value >= 9) ? 0 : 9 - value;
         return 0;
     default :
@@ -386,31 +446,49 @@ 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(3, "windowLog : %2u => chunkTargetSize : %u bytes  ", params.cParams.windowLog, (U32)chunkTargetSize);
-    DEBUGLOG(2, "nbChunks  : %2u   (chunkSize : %u bytes)   ", nbChunks, (U32)avgChunkSize);
-    params.fParams.contentSizeFlag = 1;
-
+    DEBUGLOG(4, "nbChunks  : %2u   (chunkSize : %u bytes)   ", nbChunks, (U32)avgChunkSize);
     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);
+    }
+    assert(avgChunkSize >= 256 KB);  /* condition for ZSTD_compressBound(A) + ZSTD_compressBound(B) <= ZSTD_compressBound(A+B), which is useful to avoid allocating extra buffers */
+
+    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;
@@ -425,15 +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);
@@ -442,27 +523,27 @@ size_t ZSTDMT_compressCCtx(ZSTDMT_CCtx* mtctx,
             mtctx->jobs[u].jobCompleted_mutex = &mtctx->jobCompleted_mutex;
             mtctx->jobs[u].jobCompleted_cond = &mtctx->jobCompleted_cond;
 
-            DEBUGLOG(3, "posting job %u   (%u bytes)", u, (U32)chunkSize);
-            DEBUG_PRINTHEX(3, 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++) {
-            DEBUGLOG(3, "waiting for chunk %u ", chunkID);
+            DEBUGLOG(5, "waiting for chunk %u ", chunkID);
             PTHREAD_MUTEX_LOCK(&mtctx->jobCompleted_mutex);
             while (mtctx->jobs[chunkID].jobCompleted==0) {
-                DEBUGLOG(4, "waiting for jobCompleted signal from chunk %u", chunkID);
+                DEBUGLOG(5, "waiting for jobCompleted signal from chunk %u", chunkID);
                 pthread_cond_wait(&mtctx->jobCompleted_cond, &mtctx->jobCompleted_mutex);
             }
             pthread_mutex_unlock(&mtctx->jobCompleted_mutex);
-            DEBUGLOG(3, "ready to write chunk %u ", chunkID);
+            DEBUGLOG(5, "ready to write chunk %u ", chunkID);
 
             ZSTDMT_releaseCCtx(mtctx->cctxPool, mtctx->jobs[chunkID].cctx);
             mtctx->jobs[chunkID].cctx = NULL;
@@ -470,20 +551,33 @@ 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 ;
             }
         }
-        if (!error) DEBUGLOG(3, "compressed size : %u  ", (U32)dstPos);
+        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);
 }
 
 
@@ -491,12 +585,14 @@ size_t ZSTDMT_compressCCtx(ZSTDMT_CCtx* mtctx,
 /* =======      Streaming API     ======= */
 /* ====================================== */
 
-static void ZSTDMT_waitForAllJobsCompleted(ZSTDMT_CCtx* zcs) {
+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);
         while (zcs->jobs[jobID].jobCompleted==0) {
-            DEBUGLOG(4, "waiting for jobCompleted signal from chunk %u", zcs->doneJobID);   /* we want to block when waiting for data to flush */
+            DEBUGLOG(5, "waiting for jobCompleted signal from chunk %u", zcs->doneJobID);   /* we want to block when waiting for data to flush */
             pthread_cond_wait(&zcs->jobCompleted_cond, &zcs->jobCompleted_mutex);
         }
         pthread_mutex_unlock(&zcs->jobCompleted_mutex);
@@ -505,33 +601,54 @@ static void ZSTDMT_waitForAllJobsCompleted(ZSTDMT_CCtx* zcs) {
 }
 
 
-static size_t ZSTDMT_initCStream_internal(ZSTDMT_CCtx* zcs,
-                                    const void* dict, size_t dictSize, unsigned updateDict,
-                                    ZSTD_parameters params, unsigned long long pledgedSrcSize)
+/** 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)
 {
-    ZSTD_customMem const cmem = { NULL, NULL, NULL };
-    DEBUGLOG(3, "Started new compression, with windowLog : %u", params.cParams.windowLog);
-    if (zcs->nbThreads==1) return ZSTD_initCStream_advanced(zcs->cstream, dict, dictSize, params, pledgedSrcSize);
-    if (zcs->allJobsCompleted == 0) {   /* previous job not correctly finished */
+    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(4, "single thread mode");
+        return ZSTD_initCStream_internal(zcs->cctxPool->cctx[0],
+                                dict, dictSize, cdict,
+                                params, pledgedSrcSize);
+    }
+
+    if (zcs->allJobsCompleted == 0) {   /* previous compression not correctly finished */
         ZSTDMT_waitForAllJobsCompleted(zcs);
         ZSTDMT_releaseAllJobResources(zcs);
         zcs->allJobsCompleted = 1;
     }
+
     zcs->params = params;
-    if (updateDict) {
-        ZSTD_freeCDict(zcs->cdict); zcs->cdict = NULL;
-        if (dict && dictSize) {
-            zcs->cdict = ZSTD_createCDict_advanced(dict, dictSize, 0, params.cParams, cmem);
-            if (zcs->cdict == NULL) return ERROR(memory_allocation);
-    }   }
     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 */
+                                                    params.cParams, zcs->cMem);
+        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(4, "overlapRLog : %u ", zcs->overlapRLog);
-    DEBUGLOG(3, "overlap Size : %u KB", (U32)(zcs->targetDictSize>>10));
+    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(3, "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);
@@ -546,24 +663,39 @@ static size_t ZSTDMT_initCStream_internal(ZSTDMT_CCtx* zcs,
     return 0;
 }
 
-size_t ZSTDMT_initCStream_advanced(ZSTDMT_CCtx* zcs,
+size_t ZSTDMT_initCStream_advanced(ZSTDMT_CCtx* mtctx,
                                 const void* dict, size_t dictSize,
                                 ZSTD_parameters params, unsigned long long pledgedSrcSize)
 {
-    return ZSTDMT_initCStream_internal(zcs, dict, dictSize, 1, params, pledgedSrcSize);
+    DEBUGLOG(5, "ZSTDMT_initCStream_advanced");
+    return ZSTDMT_initCStream_internal(mtctx, dict, dictSize, NULL, params, pledgedSrcSize);
 }
 
+size_t ZSTDMT_initCStream_usingCDict(ZSTDMT_CCtx* mtctx,
+                               const ZSTD_CDict* cdict,
+                                     ZSTD_frameParameters fParams,
+                                     unsigned long long pledgedSrcSize)
+{
+    ZSTD_parameters params = ZSTD_getParamsFromCDict(cdict);
+    if (cdict==NULL) return ERROR(dictionary_wrong);   /* method incompatible with NULL cdict */
+    params.fParams = fParams;
+    return ZSTDMT_initCStream_internal(mtctx, NULL, 0 /*dictSize*/, cdict,
+                                        params, pledgedSrcSize);
+}
+
+
 /* ZSTDMT_resetCStream() :
  * pledgedSrcSize is optional and can be zero == unknown */
 size_t ZSTDMT_resetCStream(ZSTDMT_CCtx* zcs, unsigned long long pledgedSrcSize)
 {
-    if (zcs->nbThreads==1) return ZSTD_resetCStream(zcs->cstream, pledgedSrcSize);
+    if (zcs->nbThreads==1)
+        return ZSTD_resetCStream(zcs->cctxPool->cctx[0], pledgedSrcSize);
     return ZSTDMT_initCStream_internal(zcs, NULL, 0, 0, zcs->params, pledgedSrcSize);
 }
 
 size_t ZSTDMT_initCStream(ZSTDMT_CCtx* zcs, int compressionLevel) {
     ZSTD_parameters const params = ZSTD_getParams(compressionLevel, 0, 0);
-    return ZSTDMT_initCStream_internal(zcs, NULL, 0, 1, params, 0);
+    return ZSTDMT_initCStream_internal(zcs, NULL, 0, NULL, params, 0);
 }
 
 
@@ -582,13 +714,16 @@ static size_t ZSTDMT_createCompressionJob(ZSTDMT_CCtx* zcs, size_t srcSize, unsi
         return ERROR(memory_allocation);
     }
 
-    DEBUGLOG(4, "preparing job %u to compress %u bytes with %u preload ", zcs->nextJobID, (U32)srcSize, (U32)zcs->dictSize);
+    DEBUGLOG(4, "preparing job %u to compress %u bytes with %u preload ",
+                zcs->nextJobID, (U32)srcSize, (U32)zcs->dictSize);
     zcs->jobs[jobID].src = zcs->inBuff.buffer;
     zcs->jobs[jobID].srcStart = zcs->inBuff.buffer.start;
     zcs->jobs[jobID].srcSize = srcSize;
-    zcs->jobs[jobID].dictSize = zcs->dictSize;   /* note : zcs->inBuff.filled is presumed >= srcSize + dictSize */
+    zcs->jobs[jobID].dictSize = zcs->dictSize;
+    assert(zcs->inBuff.filled >= srcSize + zcs->dictSize);
     zcs->jobs[jobID].params = zcs->params;
-    if (zcs->nextJobID) zcs->jobs[jobID].params.fParams.checksumFlag = 0;  /* do not calculate checksum within sections, just keep it in header for first section */
+    /* do not calculate checksum within sections, but write it in header for first section */
+    if (zcs->nextJobID) zcs->jobs[jobID].params.fParams.checksumFlag = 0;
     zcs->jobs[jobID].cdict = zcs->nextJobID==0 ? zcs->cdict : NULL;
     zcs->jobs[jobID].fullFrameSize = zcs->frameContentSize;
     zcs->jobs[jobID].dstBuff = dstBuffer;
@@ -603,6 +738,7 @@ static size_t ZSTDMT_createCompressionJob(ZSTDMT_CCtx* zcs, size_t srcSize, unsi
     /* get a new buffer for next input */
     if (!endFrame) {
         size_t const newDictSize = MIN(srcSize + zcs->dictSize, zcs->targetDictSize);
+        DEBUGLOG(5, "ZSTDMT_createCompressionJob::endFrame = %u", endFrame);
         zcs->inBuff.buffer = ZSTDMT_getBuffer(zcs->buffPool, zcs->inBuffSize);
         if (zcs->inBuff.buffer.start == NULL) {   /* not enough memory to allocate next input buffer */
             zcs->jobs[jobID].jobCompleted = 1;
@@ -611,22 +747,33 @@ static size_t ZSTDMT_createCompressionJob(ZSTDMT_CCtx* zcs, size_t srcSize, unsi
             ZSTDMT_releaseAllJobResources(zcs);
             return ERROR(memory_allocation);
         }
-        DEBUGLOG(5, "inBuff filled to %u", (U32)zcs->inBuff.filled);
+        DEBUGLOG(5, "inBuff currently filled to %u", (U32)zcs->inBuff.filled);
         zcs->inBuff.filled -= srcSize + zcs->dictSize - newDictSize;
-        DEBUGLOG(5, "new job : filled to %u, with %u dict and %u src", (U32)zcs->inBuff.filled, (U32)newDictSize, (U32)(zcs->inBuff.filled - newDictSize));
-        memmove(zcs->inBuff.buffer.start, (const char*)zcs->jobs[jobID].srcStart + zcs->dictSize + srcSize - newDictSize, zcs->inBuff.filled);
+        DEBUGLOG(5, "new job : inBuff filled to %u, with %u dict and %u src",
+                    (U32)zcs->inBuff.filled, (U32)newDictSize,
+                    (U32)(zcs->inBuff.filled - newDictSize));
+        memmove(zcs->inBuff.buffer.start,
+            (const char*)zcs->jobs[jobID].srcStart + zcs->dictSize + srcSize - newDictSize,
+            zcs->inBuff.filled);
         DEBUGLOG(5, "new inBuff pre-filled");
         zcs->dictSize = newDictSize;
-    } else {
+    } else {   /* if (endFrame==1) */
+        DEBUGLOG(5, "ZSTDMT_createCompressionJob::endFrame = %u", endFrame);
         zcs->inBuff.buffer = g_nullBuffer;
         zcs->inBuff.filled = 0;
         zcs->dictSize = 0;
         zcs->frameEnded = 1;
         if (zcs->nextJobID == 0)
-            zcs->params.fParams.checksumFlag = 0;   /* single chunk : checksum is calculated directly within worker thread */
+            /* single chunk exception : checksum is calculated directly within worker thread */
+            zcs->params.fParams.checksumFlag = 0;
     }
 
-    DEBUGLOG(3, "posting job %u : %u bytes  (end:%u) (note : doneJob = %u=>%u)", zcs->nextJobID, (U32)zcs->jobs[jobID].srcSize, zcs->jobs[jobID].lastChunk, zcs->doneJobID, zcs->doneJobID & zcs->jobIDMask);
+    DEBUGLOG(4, "posting job %u : %u bytes  (end:%u) (note : doneJob = %u=>%u)",
+                zcs->nextJobID,
+                (U32)zcs->jobs[jobID].srcSize,
+                zcs->jobs[jobID].lastChunk,
+                zcs->doneJobID,
+                zcs->doneJobID & zcs->jobIDMask);
     POOL_add(zcs->factory, ZSTDMT_compressChunk, &zcs->jobs[jobID]);   /* this call is blocking when thread worker pool is exhausted */
     zcs->nextJobID++;
     return 0;
@@ -664,7 +811,7 @@ static size_t ZSTDMT_flushNextJob(ZSTDMT_CCtx* zcs, ZSTD_outBuffer* output, unsi
                 XXH64_update(&zcs->xxhState, (const char*)job.srcStart + job.dictSize, job.srcSize);
                 if (zcs->frameEnded && (zcs->doneJobID+1 == zcs->nextJobID)) {  /* write checksum at end of last section */
                     U32 const checksum = (U32)XXH64_digest(&zcs->xxhState);
-                    DEBUGLOG(4, "writing checksum : %08X \n", checksum);
+                    DEBUGLOG(5, "writing checksum : %08X \n", checksum);
                     MEM_writeLE32((char*)job.dstBuff.start + job.cSize, checksum);
                     job.cSize += 4;
                     zcs->jobs[wJobID].cSize += 4;
@@ -675,7 +822,7 @@ static size_t ZSTDMT_flushNextJob(ZSTDMT_CCtx* zcs, ZSTD_outBuffer* output, unsi
             zcs->jobs[wJobID].jobScanned = 1;
         }
         {   size_t const toWrite = MIN(job.cSize - job.dstFlushed, output->size - output->pos);
-            DEBUGLOG(4, "Flushing %u bytes from job %u ", (U32)toWrite, zcs->doneJobID);
+            DEBUGLOG(5, "Flushing %u bytes from job %u ", (U32)toWrite, zcs->doneJobID);
             memcpy((char*)output->dst + output->pos, (const char*)job.dstBuff.start + job.dstFlushed, toWrite);
             output->pos += toWrite;
             job.dstFlushed += toWrite;
@@ -696,26 +843,81 @@ 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) return ERROR(stage_wrong);   /* current frame being ended. Only flush is allowed. Restart with init */
-    if (zcs->nbThreads==1) return ZSTD_compressStream(zcs->cstream, output, input);
+    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 frame with init */
+        return ERROR(stage_wrong);
+    }
+    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 */
+      && (endOp==ZSTD_e_end)        /* end order */
+      && (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, input->src, toLoad);
+    if ((input->src) && (mtctx->inBuff.buffer.start)) {   /* support NULL input */
+        size_t const toLoad = MIN(input->size - input->pos, mtctx->inBuffSize - mtctx->inBuff.filled);
+        DEBUGLOG(2, "inBuff:%08X;  inBuffSize=%u;  ToCopy=%u", (U32)(size_t)mtctx->inBuff.buffer.start, (U32)mtctx->inBuffSize, (U32)toLoad);
+        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) );
+    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)) ); /* 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 */
@@ -726,26 +928,28 @@ static size_t ZSTDMT_flushStream_internal(ZSTDMT_CCtx* zcs, ZSTD_outBuffer* outp
 {
     size_t const srcSize = zcs->inBuff.filled - zcs->dictSize;
 
-    if (srcSize) DEBUGLOG(4, "flushing : %u bytes left to compress", (U32)srcSize);
     if ( ((srcSize > 0) || (endFrame && !zcs->frameEnded))
        && (zcs->nextJobID <= zcs->doneJobID + zcs->jobIDMask) ) {
         CHECK_F( ZSTDMT_createCompressionJob(zcs, srcSize, endFrame) );
     }
 
     /* 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);
+    return ZSTDMT_flushNextJob(zcs, output, 1 /* blockToFlush */);
 }
 
 
 size_t ZSTDMT_flushStream(ZSTDMT_CCtx* zcs, ZSTD_outBuffer* output)
 {
-    if (zcs->nbThreads==1) return ZSTD_flushStream(zcs->cstream, output);
-    return ZSTDMT_flushStream_internal(zcs, output, 0);
+    DEBUGLOG(5, "ZSTDMT_flushStream");
+    if (zcs->nbThreads==1)
+        return ZSTD_flushStream(zcs->cctxPool->cctx[0], output);
+    return ZSTDMT_flushStream_internal(zcs, output, 0 /* endFrame */);
 }
 
 size_t ZSTDMT_endStream(ZSTDMT_CCtx* zcs, ZSTD_outBuffer* output)
 {
-    if (zcs->nbThreads==1) return ZSTD_endStream(zcs->cstream, output);
-    return ZSTDMT_flushStream_internal(zcs, output, 1);
+    DEBUGLOG(4, "ZSTDMT_endStream");
+    if (zcs->nbThreads==1)
+        return ZSTD_endStream(zcs->cctxPool->cctx[0], output);
+    return ZSTDMT_flushStream_internal(zcs, output, 1 /* endFrame */);
 }