[CLI] Switch dictionary builder on CLI to cover
Nick Terrell
parent
020b960e13
commit
f376d47c11
|
|
@ -940,9 +940,9 @@ ZDICTLIB_API size_t COVER_optimizeTrainFromBuffer(void *dictBuffer,
|
|||
const unsigned nbThreads = parameters->nbThreads;
|
||||
const unsigned kMinD = parameters->d == 0 ? 6 : parameters->d;
|
||||
const unsigned kMaxD = parameters->d == 0 ? 8 : parameters->d;
|
||||
const unsigned kMinK = parameters->k == 0 ? 40 + kMaxD : parameters->k;
|
||||
const unsigned kMaxK = parameters->k == 0 ? 2048 : parameters->k;
|
||||
const unsigned kSteps = parameters->steps == 0 ? 32 : parameters->steps;
|
||||
const unsigned kMinK = parameters->k == 0 ? 50 : parameters->k;
|
||||
const unsigned kMaxK = parameters->k == 0 ? 2000 : parameters->k;
|
||||
const unsigned kSteps = parameters->steps == 0 ? 40 : parameters->steps;
|
||||
const unsigned kStepSize = MAX((kMaxK - kMinK) / kSteps, 1);
|
||||
const unsigned kIterations =
|
||||
(1 + (kMaxD - kMinD) / 2) * (1 + (kMaxK - kMinK) / kStepSize);
|
||||
|
|
|
|||
|
|
@ -168,49 +168,57 @@ All arguments after \fB\-\-\fR are treated as files
|
|||
.
|
||||
.TP
|
||||
\fB\-\-train FILEs\fR
|
||||
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)\.
|
||||
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\.
|
||||
.
|
||||
.TP
|
||||
\fB\-o file\fR
|
||||
dictionary saved into \fBfile\fR (default name: dictionary)
|
||||
Dictionary saved into \fBfile\fR (default name: dictionary)\.
|
||||
.
|
||||
.TP
|
||||
\fB\-\-maxdict=#\fR
|
||||
limit dictionary to specified size (default : (112640)
|
||||
Limit dictionary to specified size (default: 112640)\.
|
||||
.
|
||||
.TP
|
||||
\fB\-\-dictID=#\fR
|
||||
A dictionary ID is a locally unique ID that a decoder can use to verify it is using the right dictionary\. By default, zstd will create a 4\-bytes random number ID\. It\'s possible to give a precise number instead\. Short numbers have an advantage : an ID < 256 will only need 1 byte in the compressed frame header, and an ID < 65536 will only need 2 bytes\. This compares favorably to 4 bytes default\. However, it\'s up to the dictionary manager to not assign twice the same ID to 2 different dictionaries\.
|
||||
.
|
||||
.TP
|
||||
\fB\-s#\fR
|
||||
dictionary selectivity level (default: 9) the smaller the value, the denser the dictionary, improving its efficiency but reducing its possible maximum size\.
|
||||
\fB\-\-train\-cover[=k#,d=#,steps=#]\fR
|
||||
Select parameters for the default dictionary builder algorithm named cover\. If \fId\fR is not specified, then it tries \fId\fR = 6 and \fId\fR = 8\. If \fIk\fR is not specified, then it tries \fIsteps\fR values in the range [50, 2000]\. If \fIsteps\fR is not specified, then the default value of 40 is used\. Requires that \fId\fR <= \fIk\fR\.
|
||||
.
|
||||
.IP
|
||||
Selects segments of size \fIk\fR with highest score to put in the dictionary\. The score of a segment is computed by the sum of the frequencies of all the subsegments of size \fId\fR\. Generally \fId\fR should be in the range [6, 8], occasionally up to 16, but the algorithm will run faster with d <= \fI8\fR\. Good values for \fIk\fR vary widely based on the input data, but a safe range is [2 * \fId\fR, 2000]\. Supports multithreading if \fBzstd\fR is compiled with threading support\.
|
||||
.
|
||||
.IP
|
||||
Examples:
|
||||
.
|
||||
.IP
|
||||
\fBzstd \-\-train\-cover FILEs\fR
|
||||
.
|
||||
.IP
|
||||
\fBzstd \-\-train\-cover=k=50,d=8 FILEs\fR
|
||||
.
|
||||
.IP
|
||||
\fBzstd \-\-train\-cover=d=8,steps=500 FILEs\fR
|
||||
.
|
||||
.IP
|
||||
\fBzstd \-\-train\-cover=k=50 FILEs\fR
|
||||
.
|
||||
.TP
|
||||
\fB\-\-cover=k#,d=#\fR
|
||||
Use alternate dictionary builder algorithm named cover with parameters \fIk\fR and \fId\fR with \fId\fR <= \fIk\fR\. Selects segments of size \fIk\fR with the highest score to put in the dictionary\. The score of a segment is computed by the sum of the frequencies of all the subsegments of of size \fId\fR\. Generally \fId\fR should be in the range [6, 8], but no more than 24\. When \fId\fR <= 8, the dictionary builder will run significantly faster\. Good values for \fIk\fR vary widely based on the input data, but a safe range is [32, 2048]\.
|
||||
.
|
||||
.br
|
||||
Example: \fB\-\-train \-\-cover=k=64,d=8 FILEs\fR\.
|
||||
.
|
||||
.TP
|
||||
\fB\-\-optimize\-cover[=steps=#,k=#,d=#]\fR
|
||||
If \fIsteps\fR is not specified, the default value of 32 is used\. If \fIk\fR is not specified, the \fIk\fR values in [48, 2048] are checked for each value of \fId\fR\. If \fId\fR is not specified, the values checked are [6, 8]\.
|
||||
\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\.
|
||||
.
|
||||
.IP
|
||||
Runs the cover dictionary builder for each parameter set and saves the optimal parameters and dictionary\. Prints optimal parameters and writes optimal dictionary into output file\. Supports multithreading if \fBzstd\fR is compiled with threading support\.
|
||||
Examples:
|
||||
.
|
||||
.IP
|
||||
The parameter \fIk\fR is more sensitive than \fId\fR, and is faster to optimize over\.
|
||||
\fBzstd \-\-train\-legacy FILEs\fR
|
||||
.
|
||||
.IP
|
||||
Examples :
|
||||
.
|
||||
.IP
|
||||
\fBzstd \-\-train \-\-optimize\-cover FILEs\fR
|
||||
.
|
||||
.IP
|
||||
\fBzstd \-\-train \-\-optimize\-cover=d=8,steps=512 FILEs\fR
|
||||
\fBzstd \-\-train\-legacy=selectivity=8 FILEs\fR
|
||||
.
|
||||
.SH "BENCHMARK"
|
||||
.
|
||||
|
|
|
|||
|
|
@ -158,14 +158,19 @@ It will improve compression ratio of small files.
|
|||
Typical gains range from 10% (at 64KB) to x5 better (at <1KB).
|
||||
|
||||
* `--train FILEs`:
|
||||
use FILEs as training set to create a dictionary.
|
||||
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).
|
||||
|
||||
Supports multithreading if `zstd` is compiled with threading support.
|
||||
Additional parameters can be specified with `--train-cover`.
|
||||
The legacy dictionary builder can be accessed with `--train-legacy`.
|
||||
Equivalent to `--train-cover=d=8,steps=4`.
|
||||
* `-o file`:
|
||||
dictionary saved into `file` (default name: dictionary)
|
||||
Dictionary saved into `file` (default name: dictionary).
|
||||
* `--maxdict=#`:
|
||||
limit dictionary to specified size (default : (112640)
|
||||
Limit dictionary to specified size (default: 112640).
|
||||
* `--dictID=#`:
|
||||
A dictionary ID is a locally unique ID that a decoder can use to verify it is
|
||||
using the right dictionary.
|
||||
|
|
@ -176,40 +181,44 @@ Typical gains range from 10% (at 64KB) to x5 better (at <1KB).
|
|||
This compares favorably to 4 bytes default.
|
||||
However, it's up to the dictionary manager to not assign twice the same ID to
|
||||
2 different dictionaries.
|
||||
* `-s#`:
|
||||
dictionary selectivity level (default: 9)
|
||||
the smaller the value, the denser the dictionary,
|
||||
improving its efficiency but reducing its possible maximum size.
|
||||
* `--cover=k#,d=#`:
|
||||
Use alternate dictionary builder algorithm named cover with parameters
|
||||
_k_ and _d_ with _d_ <= _k_.
|
||||
Selects segments of size _k_ with the highest score to put in the dictionary.
|
||||
* `--train-cover[=k#,d=#,steps=#]`:
|
||||
Select parameters for the default dictionary builder algorithm named cover.
|
||||
If _d_ is not specified, then it tries _d_ = 6 and _d_ = 8.
|
||||
If _k_ is not specified, then it tries _steps_ values in the range [50, 2000].
|
||||
If _steps_ is not specified, then the default value of 40 is used.
|
||||
Requires that _d_ <= _k_.
|
||||
|
||||
Selects segments of size _k_ with highest score to put in the dictionary.
|
||||
The score of a segment is computed by the sum of the frequencies of all the
|
||||
subsegments of of size _d_.
|
||||
Generally _d_ should be in the range [6, 8], but no more than 24.
|
||||
When _d_ <= 8, the dictionary builder will run significantly faster.
|
||||
Good values for _k_ vary widely based on the input data,
|
||||
but a safe range is [32, 2048].<br />
|
||||
Example: `--train --cover=k=64,d=8 FILEs`.
|
||||
|
||||
* `--optimize-cover[=steps=#,k=#,d=#]`:
|
||||
If _steps_ is not specified, the default value of 32 is used.
|
||||
If _k_ is not specified, the _k_ values in [48, 2048] are checked for each
|
||||
value of _d_.
|
||||
If _d_ is not specified, the values checked are [6, 8].
|
||||
|
||||
Runs the cover dictionary builder for each parameter set
|
||||
and saves the optimal parameters and dictionary.
|
||||
Prints optimal parameters and writes optimal dictionary into output file.
|
||||
subsegments of size _d_.
|
||||
Generally _d_ should be in the range [6, 8], occasionally up to 16, but the
|
||||
algorithm will run faster with d <= _8_.
|
||||
Good values for _k_ vary widely based on the input data, but a safe range is
|
||||
[2 * _d_, 2000].
|
||||
Supports multithreading if `zstd` is compiled with threading support.
|
||||
|
||||
The parameter _k_ is more sensitive than _d_, and is faster to optimize over.
|
||||
Examples:
|
||||
|
||||
Examples :
|
||||
`zstd --train-cover FILEs`
|
||||
|
||||
`zstd --train --optimize-cover FILEs`
|
||||
`zstd --train-cover=k=50,d=8 FILEs`
|
||||
|
||||
`zstd --train --optimize-cover=d=8,steps=512 FILEs`
|
||||
`zstd --train-cover=d=8,steps=500 FILEs`
|
||||
|
||||
`zstd --train-cover=k=50 FILEs`
|
||||
|
||||
* `--train-legacy[=selectivity=#]`:
|
||||
Use legacy dictionary builder algorithm with the given dictionary
|
||||
_selectivity_ (default: 9).
|
||||
The smaller the _selectivity_ value, the denser the dictionary,
|
||||
improving its efficiency but reducing its possible maximum size.
|
||||
`--train-legacy=s=#` is also accepted.
|
||||
|
||||
Examples:
|
||||
|
||||
`zstd --train-legacy FILEs`
|
||||
|
||||
`zstd --train-legacy=selectivity=8 FILEs`
|
||||
|
||||
|
||||
BENCHMARK
|
||||
|
|
|
|||
|
|
@ -153,11 +153,10 @@ static int usage_advanced(const char* programName)
|
|||
DISPLAY( "\n");
|
||||
DISPLAY( "Dictionary builder :\n");
|
||||
DISPLAY( "--train ## : create a dictionary from a training set of files \n");
|
||||
DISPLAY( "--cover=k=#,d=# : use the cover algorithm with parameters k and d \n");
|
||||
DISPLAY( "--optimize-cover[=steps=#,k=#,d=#] : optimize cover parameters with optional parameters\n");
|
||||
DISPLAY( "--train-cover[=k=#,d=#,steps=#] : use the 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);
|
||||
DISPLAY( " -s# : dictionary selectivity level (default: %u)\n", g_defaultSelectivityLevel);
|
||||
DISPLAY( "--dictID=# : force dictionary ID to specified value (default: random)\n");
|
||||
#endif
|
||||
#ifndef ZSTD_NOBENCH
|
||||
|
|
@ -241,11 +240,11 @@ static unsigned longCommandWArg(const char** stringPtr, const char* longCommand)
|
|||
#ifndef ZSTD_NODICT
|
||||
/**
|
||||
* parseCoverParameters() :
|
||||
* reads cover parameters from *stringPtr (e.g. "--cover=smoothing=100,kmin=48,kstep=4,kmax=64,d=8") into *params
|
||||
* reads cover parameters from *stringPtr (e.g. "--train-cover=k=48,d=8,steps=32") into *params
|
||||
* @return 1 means that cover parameters were correct
|
||||
* @return 0 in case of malformed parameters
|
||||
*/
|
||||
static unsigned parseCoverParameters(const char* stringPtr, COVER_params_t *params)
|
||||
static unsigned parseCoverParameters(const char* stringPtr, COVER_params_t* params)
|
||||
{
|
||||
memset(params, 0, sizeof(*params));
|
||||
for (; ;) {
|
||||
|
|
@ -255,9 +254,33 @@ static unsigned parseCoverParameters(const char* stringPtr, COVER_params_t *para
|
|||
return 0;
|
||||
}
|
||||
if (stringPtr[0] != 0) return 0;
|
||||
DISPLAYLEVEL(4, "k=%u\nd=%u\nsteps=%u\n", params->k, params->d, params->steps);
|
||||
DISPLAYLEVEL(4, "cover: k=%u\nd=%u\nsteps=%u\n", params->k, params->d, params->steps);
|
||||
return 1;
|
||||
}
|
||||
|
||||
/**
|
||||
* parseLegacyParameters() :
|
||||
* reads legacy dictioanry builter parameters from *stringPtr (e.g. "--train-legacy=selectivity=8") into *selectivity
|
||||
* @return 1 means that legacy dictionary builder parameters were correct
|
||||
* @return 0 in case of malformed parameters
|
||||
*/
|
||||
static unsigned parseLegacyParameters(const char* stringPtr, unsigned* selectivity)
|
||||
{
|
||||
if (!longCommandWArg(&stringPtr, "s=") && !longCommandWArg(&stringPtr, "selectivity=")) { return 0; }
|
||||
*selectivity = readU32FromChar(&stringPtr);
|
||||
if (stringPtr[0] != 0) return 0;
|
||||
DISPLAYLEVEL(4, "legacy: selectivity=%u\n", *selectivity);
|
||||
return 1;
|
||||
}
|
||||
|
||||
static COVER_params_t defaultCoverParams(void)
|
||||
{
|
||||
COVER_params_t params;
|
||||
memset(¶ms, 0, sizeof(params));
|
||||
params.d = 8;
|
||||
params.steps = 4;
|
||||
return params;
|
||||
}
|
||||
#endif
|
||||
|
||||
|
||||
|
|
@ -331,8 +354,8 @@ int main(int argCount, const char* argv[])
|
|||
unsigned fileNamesNb;
|
||||
#endif
|
||||
#ifndef ZSTD_NODICT
|
||||
COVER_params_t coverParams;
|
||||
int cover = 0;
|
||||
COVER_params_t coverParams = defaultCoverParams();
|
||||
int cover = 1;
|
||||
#endif
|
||||
|
||||
/* init */
|
||||
|
|
@ -413,18 +436,26 @@ int main(int argCount, const char* argv[])
|
|||
|
||||
/* long commands with arguments */
|
||||
#ifndef ZSTD_NODICT
|
||||
if (longCommandWArg(&argument, "--cover=")) {
|
||||
cover=1; if (!parseCoverParameters(argument, &coverParams)) CLEAN_RETURN(badusage(programName));
|
||||
continue;
|
||||
}
|
||||
if (longCommandWArg(&argument, "--optimize-cover")) {
|
||||
cover=2;
|
||||
if (longCommandWArg(&argument, "--train-cover")) {
|
||||
operation = zom_train;
|
||||
outFileName = g_defaultDictName;
|
||||
cover = 1;
|
||||
/* 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-legacy")) {
|
||||
operation = zom_train;
|
||||
outFileName = g_defaultDictName;
|
||||
cover = 0;
|
||||
/* Allow optional arguments following an = */
|
||||
if (*argument == 0) { continue; }
|
||||
else if (*argument++ != '=') { CLEAN_RETURN(badusage(programName)); }
|
||||
else if (!parseLegacyParameters(argument, &dictSelect)) { CLEAN_RETURN(badusage(programName)); }
|
||||
continue;
|
||||
}
|
||||
#endif
|
||||
if (longCommandWArg(&argument, "--threads=")) { nbThreads = readU32FromChar(&argument); continue; }
|
||||
if (longCommandWArg(&argument, "--memlimit=")) { memLimit = readU32FromChar(&argument); continue; }
|
||||
|
|
@ -659,11 +690,12 @@ int main(int argCount, const char* argv[])
|
|||
if (operation==zom_train) {
|
||||
#ifndef ZSTD_NODICT
|
||||
if (cover) {
|
||||
int const optimize = !coverParams.k || !coverParams.d;
|
||||
coverParams.nbThreads = nbThreads;
|
||||
coverParams.compressionLevel = dictCLevel;
|
||||
coverParams.notificationLevel = g_displayLevel;
|
||||
coverParams.dictID = dictID;
|
||||
operationResult = DiB_trainFromFiles(outFileName, maxDictSize, filenameTable, filenameIdx, NULL, &coverParams, cover - 1);
|
||||
operationResult = DiB_trainFromFiles(outFileName, maxDictSize, filenameTable, filenameIdx, NULL, &coverParams, optimize);
|
||||
} else {
|
||||
ZDICT_params_t dictParams;
|
||||
memset(&dictParams, 0, sizeof(dictParams));
|
||||
|
|
|
|||
|
|
@ -314,9 +314,9 @@ esac
|
|||
rm -rf dirTestDict
|
||||
$ECHO "- dictionary builder on bogus input"
|
||||
$ECHO "Hello World" > tmp
|
||||
$ZSTD --train -q tmp && die "Dictionary training should fail : not enough input source"
|
||||
$ZSTD --train-legacy -q tmp && die "Dictionary training should fail : not enough input source"
|
||||
./datagen -P0 -g10M > tmp
|
||||
$ZSTD --train -q tmp && die "Dictionary training should fail : source is pure noise"
|
||||
$ZSTD --train-legacy -q tmp && die "Dictionary training should fail : source is pure noise"
|
||||
rm tmp*
|
||||
|
||||
|
||||
|
|
@ -325,19 +325,39 @@ $ECHO "\n**** cover dictionary tests **** "
|
|||
TESTFILE=../programs/zstdcli.c
|
||||
./datagen > tmpDict
|
||||
$ECHO "- Create first dictionary"
|
||||
$ZSTD --train --cover=k=46,d=8 *.c ../programs/*.c -o tmpDict
|
||||
$ZSTD --train-cover=k=46,d=8 *.c ../programs/*.c -o tmpDict
|
||||
cp $TESTFILE tmp
|
||||
$ZSTD -f tmp -D tmpDict
|
||||
$ZSTD -d tmp.zst -D tmpDict -fo result
|
||||
$DIFF $TESTFILE result
|
||||
$ECHO "- Create second (different) dictionary"
|
||||
$ZSTD --train --cover=k=56,d=8 *.c ../programs/*.c ../programs/*.h -o tmpDictC
|
||||
$ZSTD --train-cover=k=56,d=8 *.c ../programs/*.c ../programs/*.h -o tmpDictC
|
||||
$ZSTD -d tmp.zst -D tmpDictC -fo result && die "wrong dictionary not detected!"
|
||||
$ECHO "- Create dictionary with short dictID"
|
||||
$ZSTD --train --cover=k=46,d=8 *.c ../programs/*.c --dictID=1 -o tmpDict1
|
||||
$ZSTD --train-cover=k=46,d=8 *.c ../programs/*.c --dictID=1 -o tmpDict1
|
||||
cmp tmpDict tmpDict1 && die "dictionaries should have different ID !"
|
||||
$ECHO "- Create dictionary with size limit"
|
||||
$ZSTD --train --optimize-cover=steps=8 *.c ../programs/*.c -o tmpDict2 --maxdict=4K
|
||||
$ZSTD --train-cover=steps=8 *.c ../programs/*.c -o tmpDict2 --maxdict=4K
|
||||
rm tmp*
|
||||
|
||||
$ECHO "\n**** legacy dictionary tests **** "
|
||||
|
||||
TESTFILE=../programs/zstdcli.c
|
||||
./datagen > tmpDict
|
||||
$ECHO "- Create first dictionary"
|
||||
$ZSTD --train-legacy=selectivity=8 *.c ../programs/*.c -o tmpDict
|
||||
cp $TESTFILE tmp
|
||||
$ZSTD -f tmp -D tmpDict
|
||||
$ZSTD -d tmp.zst -D tmpDict -fo result
|
||||
$DIFF $TESTFILE result
|
||||
$ECHO "- Create second (different) dictionary"
|
||||
$ZSTD --train-legacy=s=5 *.c ../programs/*.c ../programs/*.h -o tmpDictC
|
||||
$ZSTD -d tmp.zst -D tmpDictC -fo result && die "wrong dictionary not detected!"
|
||||
$ECHO "- Create dictionary with short dictID"
|
||||
$ZSTD --train-legacy -s5 *.c ../programs/*.c --dictID=1 -o tmpDict1
|
||||
cmp tmpDict tmpDict1 && die "dictionaries should have different ID !"
|
||||
$ECHO "- Create dictionary with size limit"
|
||||
$ZSTD --train-legacy -s9 *.c ../programs/*.c -o tmpDict2 --maxdict=4K
|
||||
rm tmp*
|
||||
|
||||
|
||||
|
|
|
|||
Loading…
Reference in New Issue