krak/zstd
1
0
Fork 0
mirror of https://github.com/facebook/zstd.git synced 2025-10-31 16:47:48 +02:00
Code Issues Releases 1 Activity

added ability to split input files for dictionary training

Browse Source 
using command -B#
This is the same behavior as benchmark module,
which can also split input into arbitrary size blocks, using -B#.
This commit is contained in:
Yann Collet
2017-09-14 16:45:10 -07:00
parent 77c137b3ae
commit 086b9597d9
5 changed files with 116 additions and 62 deletions
Download Patch File Download Diff File Expand all files Collapse all files

29
doc/zstd_manual.html
View File

@@ -845,6 +845,35 @@ size_t ZSTD_decodingBufferSize_min(unsigned long long windowSize, unsigned long
</b>/* advanced parameters - may not remain available after API update */<b>
ZSTD_p_forceMaxWindow=1100, </b>/* Force back-reference distances to remain < windowSize,<b>
* even when referencing into Dictionary content (default:0) */
ZSTD_p_enableLongDistanceMatching=1200, </b>/* Enable long distance matching.<b>
* This parameter is designed to improve the compression
* ratio for large inputs with long distance matches.
* This increases the memory usage as well as window size.
* Note: setting this parameter sets all the LDM parameters
* as well as ZSTD_p_windowLog. It should be set after
* ZSTD_p_compressionLevel and before ZSTD_p_windowLog and
* other LDM parameters. Setting the compression level
* after this parameter overrides the window log, though LDM
* will remain enabled until explicitly disabled. */
ZSTD_p_ldmHashLog, </b>/* Size of the table for long distance matching, as a power of 2.<b>
* Larger values increase memory usage and compression ratio, but decrease
* compression speed.
* Must be clamped between ZSTD_HASHLOG_MIN and ZSTD_HASHLOG_MAX
* (default: 20). */
ZSTD_p_ldmMinMatch, </b>/* Minimum size of searched matches for long distance matcher.<b>
* Larger/too small values usually decrease compression ratio.
* Must be clamped between ZSTD_LDM_MINMATCH_MIN
* and ZSTD_LDM_MINMATCH_MAX (default: 64). */
ZSTD_p_ldmBucketSizeLog, </b>/* Log size of each bucket in the LDM hash table for collision resolution.<b>
* Larger values usually improve collision resolution but may decrease
* compression speed.
* The maximum value is ZSTD_LDM_BUCKETSIZELOG_MAX (default: 3). */
ZSTD_p_ldmHashEveryLog, </b>/* Frequency of inserting/looking up entries in the LDM hash table.<b>
* The default is MAX(0, (windowLog - ldmHashLog)) to
* optimize hash table usage.
* Larger values improve compression speed. Deviating far from the
* default value will likely result in a decrease in compression ratio.
* Must be clamped between 0 and ZSTD_WINDOWLOG_MAX - ZSTD_HASHLOG_MIN. */
} ZSTD_cParameter;
</b></pre><BR>

139
programs/dibio.c
View File

@@ -53,13 +53,12 @@ static const size_t g_maxMemory = (sizeof(size_t) == 4) ? (2 GB - 64 MB) : ((siz
* Console display
***************************************/
#define DISPLAY(...) fprintf(stderr, __VA_ARGS__)
#define DISPLAYLEVEL(l, ...) if (g_displayLevel>=l) { DISPLAY(__VA_ARGS__); }
static int g_displayLevel = 0; /* 0 : no display; 1: errors; 2: default; 4: full information */
#define DISPLAYLEVEL(l, ...) if (displayLevel>=l) { DISPLAY(__VA_ARGS__); }
#define DISPLAYUPDATE(l, ...) if (g_displayLevel>=l) { \
if ((DIB_clockSpan(g_time) > refreshRate) || (g_displayLevel>=4)) \
#define DISPLAYUPDATE(l, ...) if (displayLevel>=l) { \
if ((DIB_clockSpan(g_time) > refreshRate) || (displayLevel>=4)) \
{ g_time = clock(); DISPLAY(__VA_ARGS__); \
if (g_displayLevel>=4) fflush(stderr); } }
if (displayLevel>=4) fflush(stderr); } }
static const clock_t refreshRate = CLOCKS_PER_SEC * 2 / 10;
static clock_t g_time = 0;
@@ -76,9 +75,9 @@ static clock_t DIB_clockSpan(clock_t nPrevious) { return clock() - nPrevious; }
#define EXM_THROW(error, ...) \
{ \
DEBUGOUTPUT("Error defined at %s, line %i : \n", __FILE__, __LINE__); \
DISPLAYLEVEL(1, "Error %i : ", error); \
DISPLAYLEVEL(1, __VA_ARGS__); \
DISPLAYLEVEL(1, "\n"); \
DISPLAY("Error %i : ", error); \
DISPLAY(__VA_ARGS__); \
DISPLAY("\n"); \
exit(error); \
}
@@ -102,30 +101,42 @@ const char* DiB_getErrorName(size_t errorCode) { return ERR_getErrorName(errorCo
* @return : nb of files effectively loaded into `buffer`
* *bufferSizePtr is modified, it provides the amount data loaded within buffer */
static unsigned DiB_loadFiles(void* buffer, size_t* bufferSizePtr,
size_t* fileSizes,
const char** fileNamesTable, unsigned nbFiles)
size_t* chunkSizes,
const char** fileNamesTable, unsigned nbFiles, size_t targetChunkSize,
unsigned displayLevel)
{
char* const buff = (char*)buffer;
size_t pos = 0;
unsigned n;
unsigned nbLoadedChunks = 0, fileIndex;
for (n=0; n<nbFiles; n++) {
const char* const fileName = fileNamesTable[n];
for (fileIndex=0; fileIndex<nbFiles; fileIndex++) {
const char* const fileName = fileNamesTable[fileIndex];
unsigned long long const fs64 = UTIL_getFileSize(fileName);
size_t const fileSize = (size_t) MIN(fs64, SAMPLESIZE_MAX);
if (fileSize > *bufferSizePtr-pos) break;
{ FILE* const f = fopen(fileName, "rb");
if (f==NULL) EXM_THROW(10, "zstd: dictBuilder: %s %s ", fileName, strerror(errno));
DISPLAYUPDATE(2, "Loading %s... \r", fileName);
{ size_t const readSize = fread(buff+pos, 1, fileSize, f);
if (readSize != fileSize) EXM_THROW(11, "Pb reading %s", fileName);
pos += readSize; }
fileSizes[n] = fileSize;
fclose(f);
} }
unsigned long long remainingToLoad = fs64;
U32 const nbChunks = targetChunkSize ? (U32)((fs64 + (targetChunkSize-1)) / targetChunkSize) : 1;
U64 const chunkSize = targetChunkSize ? MIN(targetChunkSize, fs64) : fs64;
size_t const maxChunkSize = MIN(chunkSize, SAMPLESIZE_MAX);
U32 cnb;
FILE* const f = fopen(fileName, "rb");
if (f==NULL) EXM_THROW(10, "zstd: dictBuilder: %s %s ", fileName, strerror(errno));
DISPLAYUPDATE(2, "Loading %s... \r", fileName);
for (cnb=0; cnb<nbChunks; cnb++) {
size_t const toLoad = MIN(maxChunkSize, remainingToLoad);
if (toLoad > *bufferSizePtr-pos) break;
{ size_t const readSize = fread(buff+pos, 1, toLoad, f);
if (readSize != toLoad) EXM_THROW(11, "Pb reading %s", fileName);
pos += readSize;
chunkSizes[nbLoadedChunks++] = toLoad;
remainingToLoad -= targetChunkSize;
if (toLoad < targetChunkSize) {
fseek(f, (targetChunkSize - toLoad), SEEK_CUR);
} } }
fclose(f);
}
DISPLAYLEVEL(2, "\r%79s\r", "");
*bufferSizePtr = pos;
return n;
DISPLAYLEVEL(4, "loaded : %u KB \n", (U32)(pos >> 10))
return nbLoadedChunks;
}
#define DiB_rotl32(x,r) ((x << r) | (x >> (32 - r)))
@@ -207,18 +218,28 @@ static void DiB_saveDict(const char* dictFileName,
}
static int g_tooLargeSamples = 0;
static U64 DiB_totalCappedFileSize(const char** fileNamesTable, unsigned nbFiles)
typedef struct {
U64 totalSizeToLoad;
unsigned oneSampleTooLarge;
unsigned nbChunks;
} fileStats;
static fileStats DiB_fileStats(const char** fileNamesTable, unsigned nbFiles, size_t chunkSize, unsigned displayLevel)
{
U64 total = 0;
fileStats fs;
unsigned n;
memset(&fs, 0, sizeof(fs));
for (n=0; n<nbFiles; n++) {
U64 const fileSize = UTIL_getFileSize(fileNamesTable[n]);
U64 const cappedFileSize = MIN(fileSize, SAMPLESIZE_MAX);
total += cappedFileSize;
g_tooLargeSamples |= (fileSize > 2*SAMPLESIZE_MAX);
U32 const nbChunks = (U32)(chunkSize ? (fileSize + (chunkSize-1)) / chunkSize : 1);
U64 const chunkToLoad = chunkSize ? MIN(chunkSize, fileSize) : fileSize;
size_t const cappedChunkSize = MIN(chunkToLoad, SAMPLESIZE_MAX);
fs.totalSizeToLoad += cappedChunkSize * nbChunks;
fs.oneSampleTooLarge |= (chunkSize > 2*SAMPLESIZE_MAX);
fs.nbChunks += nbChunks;
}
return total;
DISPLAYLEVEL(4, "Preparing to load : %u KB \n", (U32)(fs.totalSizeToLoad >> 10));
return fs;
}
@@ -235,63 +256,65 @@ size_t ZDICT_trainFromBuffer_unsafe_legacy(void* dictBuffer, size_t dictBufferCa
int DiB_trainFromFiles(const char* dictFileName, unsigned maxDictSize,
const char** fileNamesTable, unsigned nbFiles,
const char** fileNamesTable, unsigned nbFiles, size_t chunkSize,
ZDICT_legacy_params_t *params, ZDICT_cover_params_t *coverParams,
int optimizeCover)
{
unsigned displayLevel = params ? params->zParams.notificationLevel :
coverParams ? coverParams->zParams.notificationLevel :
0; /* should never happen */
void* const dictBuffer = malloc(maxDictSize);
size_t* const fileSizes = (size_t*)malloc(nbFiles * sizeof(size_t));
unsigned long long const totalSizeToLoad = DiB_totalCappedFileSize(fileNamesTable, nbFiles);
fileStats const fs = DiB_fileStats(fileNamesTable, nbFiles, chunkSize, displayLevel);
size_t* const chunkSizes = (size_t*)malloc(fs.nbChunks * sizeof(size_t));
size_t const memMult = params ? MEMMULT : COVER_MEMMULT;
size_t const maxMem = DiB_findMaxMem(totalSizeToLoad * memMult) / memMult;
size_t benchedSize = (size_t) MIN ((unsigned long long)maxMem, totalSizeToLoad);
void* const srcBuffer = malloc(benchedSize+NOISELENGTH);
size_t const maxMem = DiB_findMaxMem(fs.totalSizeToLoad * memMult) / memMult;
size_t loadedSize = (size_t) MIN ((unsigned long long)maxMem, fs.totalSizeToLoad);
void* const srcBuffer = malloc(loadedSize+NOISELENGTH);
int result = 0;
/* Checks */
if (params) g_displayLevel = params->zParams.notificationLevel;
else if (coverParams) g_displayLevel = coverParams->zParams.notificationLevel;
else EXM_THROW(13, "Neither dictionary algorithm selected"); /* should not happen */
if ((!fileSizes) || (!srcBuffer) || (!dictBuffer))
if ((!chunkSizes) || (!srcBuffer) || (!dictBuffer))
EXM_THROW(12, "not enough memory for DiB_trainFiles"); /* should not happen */
if (g_tooLargeSamples) {
DISPLAYLEVEL(2, "! Warning : some samples are very large \n");
DISPLAYLEVEL(2, "! Note that dictionary is only useful for small files or beginning of large files. \n");
DISPLAYLEVEL(2, "! As a consequence, only the first %u bytes of each file are loaded \n", SAMPLESIZE_MAX);
if (fs.oneSampleTooLarge) {
DISPLAYLEVEL(2, "! Warning : some sample(s) are very large \n");
DISPLAYLEVEL(2, "! Note that dictionary is only useful for small samples. \n");
DISPLAYLEVEL(2, "! As a consequence, only the first %u bytes of each sample are loaded \n", SAMPLESIZE_MAX);
}
if ((nbFiles < 5) || (totalSizeToLoad < 9 * (unsigned long long)maxDictSize)) {
if (fs.nbChunks < 5) {
DISPLAYLEVEL(2, "! Warning : nb of samples too low for proper processing ! \n");
DISPLAYLEVEL(2, "! Please provide _one file per sample_. \n");
DISPLAYLEVEL(2, "! Do not concatenate samples together into a single file, \n");
DISPLAYLEVEL(2, "! as dictBuilder will be unable to find the beginning of each sample, \n");
DISPLAYLEVEL(2, "! resulting in poor dictionary quality. \n");
EXM_THROW(14, "nb of samples too low"); /* we now clearly forbid this case */
}
if (fs.totalSizeToLoad < (unsigned long long)(8 * maxDictSize)) {
DISPLAYLEVEL(2, "! Warning : data size of samples too small for target dictionary size \n");
DISPLAYLEVEL(2, "! Samples should be about 100x larger than target dictionary size \n");
}
/* init */
if (benchedSize < totalSizeToLoad)
DISPLAYLEVEL(1, "Not enough memory; training on %u MB only...\n", (unsigned)(benchedSize >> 20));
if (loadedSize < fs.totalSizeToLoad)
DISPLAYLEVEL(1, "Not enough memory; training on %u MB only...\n", (unsigned)(loadedSize >> 20));
/* Load input buffer */
DISPLAYLEVEL(3, "Shuffling input files\n");
DiB_shuffle(fileNamesTable, nbFiles);
nbFiles = DiB_loadFiles(srcBuffer, &benchedSize, fileSizes, fileNamesTable, nbFiles);
nbFiles = DiB_loadFiles(srcBuffer, &loadedSize, chunkSizes, fileNamesTable, nbFiles, chunkSize, displayLevel);
{ size_t dictSize;
if (params) {
DiB_fillNoise((char*)srcBuffer + benchedSize, NOISELENGTH); /* guard band, for end of buffer condition */
DiB_fillNoise((char*)srcBuffer + loadedSize, NOISELENGTH); /* guard band, for end of buffer condition */
dictSize = ZDICT_trainFromBuffer_unsafe_legacy(dictBuffer, maxDictSize,
srcBuffer, fileSizes, nbFiles,
srcBuffer, chunkSizes, fs.nbChunks,
*params);
} else if (optimizeCover) {
dictSize = ZDICT_optimizeTrainFromBuffer_cover(dictBuffer, maxDictSize,
srcBuffer, fileSizes, nbFiles,
srcBuffer, chunkSizes, fs.nbChunks,
coverParams);
if (!ZDICT_isError(dictSize)) {
DISPLAYLEVEL(2, "k=%u\nd=%u\nsteps=%u\n", coverParams->k, coverParams->d, coverParams->steps);
}
} else {
dictSize = ZDICT_trainFromBuffer_cover(dictBuffer, maxDictSize, srcBuffer,
fileSizes, nbFiles, *coverParams);
chunkSizes, fs.nbChunks, *coverParams);
}
if (ZDICT_isError(dictSize)) {
DISPLAYLEVEL(1, "dictionary training failed : %s \n", ZDICT_getErrorName(dictSize)); /* should not happen */
@@ -306,7 +329,7 @@ int DiB_trainFromFiles(const char* dictFileName, unsigned maxDictSize,
/* clean up */
_cleanup:
free(srcBuffer);
free(chunkSizes);
free(dictBuffer);
free(fileSizes);
return result;
}

2
programs/dibio.h
View File

@@ -32,7 +32,7 @@
@return : 0 == ok. Any other : error.
*/
int DiB_trainFromFiles(const char* dictFileName, unsigned maxDictSize,
const char** fileNamesTable, unsigned nbFiles,
const char** fileNamesTable, unsigned nbFiles, size_t chunkSize,
ZDICT_legacy_params_t *params, ZDICT_cover_params_t *coverParams,
int optimizeCover);

4
programs/zstd.1.md
View File

@@ -184,6 +184,8 @@ Typical gains range from 10% (at 64KB) to x5 better (at <1KB).
Dictionary saved into `file` (default name: dictionary).
* `--maxdict=#`:
Limit dictionary to specified size (default: 112640).
* `-B#`:
Split input files in blocks of size # (default: no split)
* `--dictID=#`:
A dictionary ID is a locally unique ID that a decoder can use to verify it is
using the right dictionary.
@@ -373,7 +375,7 @@ The list of available _options_:
default value will likely result in a decrease in compression ratio.
The default value is `wlog - ldmhlog`.
### -B#:
Select the size of each compression job.
This parameter is available only when multi-threading is enabled.

4
programs/zstdcli.c
View File

@@ -759,13 +759,13 @@ int main(int argCount, const char* argv[])
int const optimize = !coverParams.k || !coverParams.d;
coverParams.nbThreads = nbThreads;
coverParams.zParams = zParams;
operationResult = DiB_trainFromFiles(outFileName, maxDictSize, filenameTable, filenameIdx, NULL, &coverParams, optimize);
operationResult = DiB_trainFromFiles(outFileName, maxDictSize, filenameTable, filenameIdx, blockSize, NULL, &coverParams, optimize);
} else {
ZDICT_legacy_params_t dictParams;
memset(&dictParams, 0, sizeof(dictParams));
dictParams.selectivityLevel = dictSelect;
dictParams.zParams = zParams;
operationResult = DiB_trainFromFiles(outFileName, maxDictSize, filenameTable, filenameIdx, &dictParams, NULL, 0);
operationResult = DiB_trainFromFiles(outFileName, maxDictSize, filenameTable, filenameIdx, blockSize, &dictParams, NULL, 0);
}
#endif
goto _end;