ddidderr/zstd-rs
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Merge fastCover into DictBuilder (#1274)

Browse Source 
* Minor fix

* Run non-optimize FASTCOVER 5 times in benchmark

* Merge fastCover into dictBuilder

* Fix mixed declaration issue

* Add fastcover to symbol.c

* Add fastCover.c and cover.h to build

* Change fastCover.c to fastcover.c

* Update benchmark to run FASTCOVER in dictBuilder

* Undo spliting fastcover_param into cover_param and f

* Remove convert param functions

* Assign f to parameter

* Add zdict.h to Makefile in lib

* Add cover.h to BUCK

* Cast 1 to U64 before shifting

* Remove trimming of zero freq head and tail in selectSegment and rebenchmark

* Remove f as a separate parameter of tryParam

* Read 8 bytes when d is 6

* Add trimming off zero frequency head and tail

* Use best functions from COVER and remove trimming part(which leads to worse compression ratio after previous bugs were fixed)

* Add finalize= argument to FASTCOVER to specify percentage of training samples passed to ZDICT_finalizeDictionary

* Change nbDmer to always read 8 bytes even when d=6

* Add skip=# argument to allow skipping dmers in computeFrequency in FASTCOVER

* Update comments and benchmarking result

* Change default method of ZDICT_trainFromBuffer to ZDICT_optimizeTrainFromBuffer_fastCover

* Add dictType enum and fix bug about passing zParam when converting to coverParam

* Combine finalize and skip into a single parameter

* Update acceleration parameters and benchmark on 3 sample sets

* Change default splitPoint of FASTCOVER to 0.75 and benchmark first 3 sample sets

* Initialize variables outside of for loop in benchmark.c

* Update benchmark result for hg-manifest

* Remove cover.h from install-includes

* Add explanation of f

* Set default compression level for trainFromBuffer to 3

* Add assertion of fastCoverParams in DiB_trainFromFiles

* Add checkTotalCompressedSize function + some minor fixes

* Add test for multithreading fastCovr

* Initialize segmentFreqs in every FASTCOVER_selectSegment and move mutex_unnlock to end of COVER_best_finish

* Free segmentFreqs

* Initialize segmentFreqs before calling FASTCOVER_buildDictionary instead of in FASTCOVER_selectSegment

* Add FASTCOVER_MEMMULT

* Minor fix

* Update benchmarking result
This commit is contained in:
Jennifer Liu
2018-08-23 12:06:20 -07:00
committed by Nick Terrell
parent 1af27a7ed7
commit 9d6ed9def3
26 changed files with 1999 additions and 246 deletions
Download Patch File Download Diff File Expand all files Collapse all files
programs/README.md
+1
View File
@@ -151,6 +151,7 @@ Advanced arguments :
Dictionary builder :
--train ## : create a dictionary from a training set of files
--train-cover[=k=#,d=#,steps=#,split=#] : use the cover algorithm with optional args
--train-fastcover[=k=#,d=#,f=#,steps=#,split=#,accel=#] : use the fastcover algorithm with optional args
--train-legacy[=s=#] : use the legacy algorithm with selectivity (default: 9)
-o file : `file` is dictionary name (default: dictionary)
--maxdict=# : limit dictionary to specified size (default: 112640)
programs/dibio.c
+36 -14
View File
@@ -44,6 +44,7 @@
#define SAMPLESIZE_MAX (128 KB)
#define MEMMULT 11 /* rough estimation : memory cost to analyze 1 byte of sample */
#define COVER_MEMMULT 9 /* rough estimation : memory cost to analyze 1 byte of sample */
#define FASTCOVER_MEMMULT 1 /* rough estimation : memory cost to analyze 1 byte of sample */
static const size_t g_maxMemory = (sizeof(size_t) == 4) ? (2 GB - 64 MB) : ((size_t)(512 MB) << sizeof(size_t));
#define NOISELENGTH 32
@@ -271,16 +272,19 @@ size_t ZDICT_trainFromBuffer_unsafe_legacy(void* dictBuffer, size_t dictBufferCa
int DiB_trainFromFiles(const char* dictFileName, unsigned maxDictSize,
const char** fileNamesTable, unsigned nbFiles, size_t chunkSize,
ZDICT_legacy_params_t *params, ZDICT_cover_params_t *coverParams,
int optimizeCover)
ZDICT_legacy_params_t* params, ZDICT_cover_params_t* coverParams,
ZDICT_fastCover_params_t* fastCoverParams, int optimize)
{
unsigned const displayLevel = params ? params->zParams.notificationLevel :
coverParams ? coverParams->zParams.notificationLevel :
fastCoverParams ? fastCoverParams->zParams.notificationLevel :
0; /* should never happen */
void* const dictBuffer = malloc(maxDictSize);
fileStats const fs = DiB_fileStats(fileNamesTable, nbFiles, chunkSize, displayLevel);
size_t* const sampleSizes = (size_t*)malloc(fs.nbSamples * sizeof(size_t));
size_t const memMult = params ? MEMMULT : COVER_MEMMULT;
size_t const memMult = params ? MEMMULT :
coverParams ? COVER_MEMMULT:
FASTCOVER_MEMMULT;
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);
@@ -312,6 +316,7 @@ int DiB_trainFromFiles(const char* dictFileName, unsigned maxDictSize,
/* Load input buffer */
DISPLAYLEVEL(3, "Shuffling input files\n");
DiB_shuffle(fileNamesTable, nbFiles);
DiB_loadFiles(srcBuffer, &loadedSize, sampleSizes, fs.nbSamples, fileNamesTable, nbFiles, chunkSize, displayLevel);
{ size_t dictSize;
@@ -320,19 +325,36 @@ int DiB_trainFromFiles(const char* dictFileName, unsigned maxDictSize,
dictSize = ZDICT_trainFromBuffer_unsafe_legacy(dictBuffer, maxDictSize,
srcBuffer, sampleSizes, fs.nbSamples,
*params);
} else if (optimizeCover) {
assert(coverParams != NULL);
dictSize = ZDICT_optimizeTrainFromBuffer_cover(dictBuffer, maxDictSize,
srcBuffer, sampleSizes, fs.nbSamples,
coverParams);
if (!ZDICT_isError(dictSize)) {
unsigned splitPercentage = (unsigned)(coverParams->splitPoint * 100);
DISPLAYLEVEL(2, "k=%u\nd=%u\nsteps=%u\nsplit=%u\n", coverParams->k, coverParams->d, coverParams->steps, splitPercentage);
} else if (coverParams) {
if (optimize) {
dictSize = ZDICT_optimizeTrainFromBuffer_cover(dictBuffer, maxDictSize,
srcBuffer, sampleSizes, fs.nbSamples,
coverParams);
if (!ZDICT_isError(dictSize)) {
unsigned splitPercentage = (unsigned)(coverParams->splitPoint * 100);
DISPLAYLEVEL(2, "k=%u\nd=%u\nsteps=%u\nsplit=%u\n", coverParams->k, coverParams->d,
coverParams->steps, splitPercentage);
}
} else {
dictSize = ZDICT_trainFromBuffer_cover(dictBuffer, maxDictSize, srcBuffer,
sampleSizes, fs.nbSamples, *coverParams);
}
} else {
assert(coverParams != NULL);
dictSize = ZDICT_trainFromBuffer_cover(dictBuffer, maxDictSize, srcBuffer,
sampleSizes, fs.nbSamples, *coverParams);
assert(fastCoverParams != NULL);
if (optimize) {
dictSize = ZDICT_optimizeTrainFromBuffer_fastCover(dictBuffer, maxDictSize,
srcBuffer, sampleSizes, fs.nbSamples,
fastCoverParams);
if (!ZDICT_isError(dictSize)) {
unsigned splitPercentage = (unsigned)(fastCoverParams->splitPoint * 100);
DISPLAYLEVEL(2, "k=%u\nd=%u\nf=%u\nsteps=%u\nsplit=%u\naccel=%u\n", fastCoverParams->k,
fastCoverParams->d, fastCoverParams->f, fastCoverParams->steps, splitPercentage,
fastCoverParams->accel);
}
} else {
dictSize = ZDICT_trainFromBuffer_fastCover(dictBuffer, maxDictSize, srcBuffer,
sampleSizes, fs.nbSamples, *fastCoverParams);
}
}
if (ZDICT_isError(dictSize)) {
DISPLAYLEVEL(1, "dictionary training failed : %s \n", ZDICT_getErrorName(dictSize)); /* should not happen */
programs/dibio.h
+1 -1
View File
@@ -34,6 +34,6 @@
int DiB_trainFromFiles(const char* dictFileName, unsigned maxDictSize,
const char** fileNamesTable, unsigned nbFiles, size_t chunkSize,
ZDICT_legacy_params_t* params, ZDICT_cover_params_t* coverParams,
int optimizeCover);
ZDICT_fastCover_params_t* fastCoverParams, int optimize);
#endif
programs/zstd.1
+20 -1
View File
@@ -194,7 +194,7 @@ All arguments after \fB\-\-\fR are treated as files
Use FILEs as training set to create a dictionary\. The training set should contain a lot of small files (> 100), and weight typically 100x the target dictionary size (for example, 10 MB for a 100 KB dictionary)\.
.
.IP
Supports multithreading if \fBzstd\fR is compiled with threading support\. Additional parameters can be specified with \fB\-\-train\-cover\fR\. The legacy dictionary builder can be accessed with \fB\-\-train\-legacy\fR\. Equivalent to \fB\-\-train\-cover=d=8,steps=4\fR\.
Supports multithreading if \fBzstd\fR is compiled with threading support\. Additional parameters can be specified with \fB\-\-train\-cover\fR\ or \fB\-\-train\-fastcover\fR\. The legacy dictionary builder can be accessed with \fB\-\-train\-legacy\fR\. Equivalent to \fB\-\-train\-fastCover=d=8,steps=4\fR\.
.
.TP
\fB\-o file\fR
@@ -240,6 +240,25 @@ Examples:
\fBzstd \-\-train\-cover=k=50 FILEs\fR
.
.TP
\fB\-\-train\-fastcover[=k#,d=#,f=#,steps=#,split=#,accel=#]\fR
Same as cover but with extra parameters \fIf\fR and \fIaccel\fR and different default value of split
.
.IP
If \fIsplit\fR is not specified, then it tries \fIsplit\fR = 75. If \fIf\fR is not specified, then it tries \fIf\fR = 18. Requires that 0 < \fIf\fR < 32. If \fIaccel\fR is not specified, then it tries \fIaccel\fR = 1. Requires that 0 < \fIaccel\fR <= 10. Requires that \fId\fR = 6 or \fId\fR = 8.
.
.IP
\fIf\fR is log of size of array that keeps track of frequency of subsegments of size \fId\fR. The subsegment is hashed to an index in the range [0,2^\fIf\fR - 1]. It is possible that 2 different subsegments are hashed to the same index, and they are considered as the same subsegment when computing frequency. Using a higher \fIf\fR reduces collision but takes longer.
.
.IP
Examples:
.
.IP
\fBzstd \-\-train\-fastcover FILEs\fR
.
.IP
\fBzstd \-\-train\-fastcover=d=8,f=15,accel=2 FILEs\fR
.
.TP
\fB\-\-train\-legacy[=selectivity=#]\fR
Use legacy dictionary builder algorithm with the given dictionary \fIselectivity\fR (default: 9)\. The smaller the \fIselectivity\fR value, the denser the dictionary, improving its efficiency but reducing its possible maximum size\. \fB\-\-train\-legacy=s=#\fR is also accepted\.
.
programs/zstd.1.md
+20
View File
@@ -254,6 +254,26 @@ Compression of small files similar to the sample set will be greatly improved.
`zstd --train-cover=k=50,split=60 FILEs`
* `--train-fastcover[=k#,d=#,f=#,steps=#,split=#,accel=#]`:
Same as cover but with extra parameters _f_ and _accel_ and different default value of split
If _split_ is not specified, then it tries _split_ = 75.
If _f_ is not specified, then it tries _f_ = 18.
Requires that 0 < _f_ < 32.
If _accel_ is not specified, then it tries _accel_ = 1.
Requires that 0 < _accel_ <= 10.
Requires that _d_ = 6 or _d_ = 8.
_f_ is log of size of array that keeps track of frequency of subsegments of size _d_.
The subsegment is hashed to an index in the range [0,2^_f_ - 1].
It is possible that 2 different subsegments are hashed to the same index, and they are considered as the same subsegment when computing frequency.
Using a higher _f_ reduces collision but takes longer.
Examples:
`zstd --train-fastcover FILEs`
`zstd --train-fastcover=d=8,f=15,accel=2 FILEs`
* `--train-legacy[=selectivity=#]`:
Use legacy dictionary builder algorithm with the given dictionary
_selectivity_ (default: 9).
programs/zstdcli.c
+68 -8
View File
@@ -84,7 +84,10 @@ static U32 g_ldmMinMatch = 0;
static U32 g_ldmHashEveryLog = LDM_PARAM_DEFAULT;
static U32 g_ldmBucketSizeLog = LDM_PARAM_DEFAULT;
#define DEFAULT_SPLITPOINT 1.0
#define DEFAULT_ACCEL 1
typedef enum { cover, fastCover, legacy } dictType;
/*-************************************
* Display Macros
@@ -172,6 +175,7 @@ static int usage_advanced(const char* programName)
DISPLAY( "Dictionary builder : \n");
DISPLAY( "--train ## : create a dictionary from a training set of files \n");
DISPLAY( "--train-cover[=k=#,d=#,steps=#,split=#] : use the cover algorithm with optional args\n");
DISPLAY( "--train-fastcover[=k=#,d=#,f=#,steps=#,split=#,accel=#] : use the fast cover algorithm with optional args\n");
DISPLAY( "--train-legacy[=s=#] : use the legacy algorithm with selectivity (default: %u)\n", g_defaultSelectivityLevel);
DISPLAY( " -o file : `file` is dictionary name (default: %s) \n", g_defaultDictName);
DISPLAY( "--maxdict=# : limit dictionary to specified size (default: %u) \n", g_defaultMaxDictSize);
@@ -295,6 +299,33 @@ static unsigned parseCoverParameters(const char* stringPtr, ZDICT_cover_params_t
return 1;
}
/**
* parseFastCoverParameters() :
* reads fastcover parameters from *stringPtr (e.g. "--train-fastcover=k=48,d=8,f=20,steps=32,accel=2") into *params
* @return 1 means that fastcover parameters were correct
* @return 0 in case of malformed parameters
*/
static unsigned parseFastCoverParameters(const char* stringPtr, ZDICT_fastCover_params_t* params)
{
memset(params, 0, sizeof(*params));
for (; ;) {
if (longCommandWArg(&stringPtr, "k=")) { params->k = readU32FromChar(&stringPtr); if (stringPtr[0]==',') { stringPtr++; continue; } else break; }
if (longCommandWArg(&stringPtr, "d=")) { params->d = readU32FromChar(&stringPtr); if (stringPtr[0]==',') { stringPtr++; continue; } else break; }
if (longCommandWArg(&stringPtr, "f=")) { params->f = readU32FromChar(&stringPtr); if (stringPtr[0]==',') { stringPtr++; continue; } else break; }
if (longCommandWArg(&stringPtr, "steps=")) { params->steps = readU32FromChar(&stringPtr); if (stringPtr[0]==',') { stringPtr++; continue; } else break; }
if (longCommandWArg(&stringPtr, "accel=")) { params->accel = readU32FromChar(&stringPtr); if (stringPtr[0]==',') { stringPtr++; continue; } else break; }
if (longCommandWArg(&stringPtr, "split=")) {
unsigned splitPercentage = readU32FromChar(&stringPtr);
params->splitPoint = (double)splitPercentage / 100.0;
if (stringPtr[0]==',') { stringPtr++; continue; } else break;
}
return 0;
}
if (stringPtr[0] != 0) return 0;
DISPLAYLEVEL(4, "cover: k=%u\nd=%u\nf=%u\nsteps=%u\nsplit=%u\naccel=%u\n", params->k, params->d, params->f, params->steps, (unsigned)(params->splitPoint * 100), params->accel);
return 1;
}
/**
* parseLegacyParameters() :
* reads legacy dictioanry builter parameters from *stringPtr (e.g. "--train-legacy=selectivity=8") into *selectivity
@@ -316,7 +347,19 @@ static ZDICT_cover_params_t defaultCoverParams(void)
memset(&params, 0, sizeof(params));
params.d = 8;
params.steps = 4;
params.splitPoint = DEFAULT_SPLITPOINT;
params.splitPoint = 1.0;
return params;
}
static ZDICT_fastCover_params_t defaultFastCoverParams(void)
{
ZDICT_fastCover_params_t params;
memset(&params, 0, sizeof(params));
params.d = 8;
params.f = 18;
params.steps = 4;
params.splitPoint = 0.75; /* different from default splitPoint of cover */
params.accel = DEFAULT_ACCEL;
return params;
}
#endif
@@ -431,7 +474,8 @@ int main(int argCount, const char* argv[])
#endif
#ifndef ZSTD_NODICT
ZDICT_cover_params_t coverParams = defaultCoverParams();
int cover = 1;
ZDICT_fastCover_params_t fastCoverParams = defaultFastCoverParams();
dictType dict = fastCover;
#endif
#ifndef ZSTD_NOBENCH
BMK_advancedParams_t benchParams = BMK_initAdvancedParams();
@@ -530,18 +574,29 @@ int main(int argCount, const char* argv[])
operation = zom_train;
if (outFileName == NULL)
outFileName = g_defaultDictName;
cover = 1;
dict = cover;
/* Allow optional arguments following an = */
if (*argument == 0) { memset(&coverParams, 0, sizeof(coverParams)); }
else if (*argument++ != '=') { CLEAN_RETURN(badusage(programName)); }
else if (!parseCoverParameters(argument, &coverParams)) { CLEAN_RETURN(badusage(programName)); }
continue;
}
if (longCommandWArg(&argument, "--train-fastcover")) {
operation = zom_train;
if (outFileName == NULL)
outFileName = g_defaultDictName;
dict = fastCover;
/* Allow optional arguments following an = */
if (*argument == 0) { memset(&fastCoverParams, 0, sizeof(fastCoverParams)); }
else if (*argument++ != '=') { CLEAN_RETURN(badusage(programName)); }
else if (!parseFastCoverParameters(argument, &fastCoverParams)) { CLEAN_RETURN(badusage(programName)); }
continue;
}
if (longCommandWArg(&argument, "--train-legacy")) {
operation = zom_train;
if (outFileName == NULL)
outFileName = g_defaultDictName;
cover = 0;
dict = legacy;
/* Allow optional arguments following an = */
if (*argument == 0) { continue; }
else if (*argument++ != '=') { CLEAN_RETURN(badusage(programName)); }
@@ -881,17 +936,22 @@ int main(int argCount, const char* argv[])
zParams.compressionLevel = dictCLevel;
zParams.notificationLevel = g_displayLevel;
zParams.dictID = dictID;
if (cover) {
if (dict == cover) {
int const optimize = !coverParams.k || !coverParams.d;
coverParams.nbThreads = nbWorkers;
coverParams.zParams = zParams;
operationResult = DiB_trainFromFiles(outFileName, maxDictSize, filenameTable, filenameIdx, blockSize, NULL, &coverParams, optimize);
operationResult = DiB_trainFromFiles(outFileName, maxDictSize, filenameTable, filenameIdx, blockSize, NULL, &coverParams, NULL, optimize);
} else if (dict == fastCover) {
int const optimize = !fastCoverParams.k || !fastCoverParams.d;
fastCoverParams.nbThreads = nbWorkers;
fastCoverParams.zParams = zParams;
operationResult = DiB_trainFromFiles(outFileName, maxDictSize, filenameTable, filenameIdx, blockSize, NULL, NULL, &fastCoverParams, 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, blockSize, &dictParams, NULL, 0);
operationResult = DiB_trainFromFiles(outFileName, maxDictSize, filenameTable, filenameIdx, blockSize, &dictParams, NULL, NULL, 0);
}
#endif
goto _end;