Merge fastCover into DictBuilder (#1274)
* 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
committed by
Nick Terrell
Nick Terrell
parent
1af27a7ed7
commit
9d6ed9def3
@@ -69,6 +69,7 @@ cxx_library(
|
||||
]),
|
||||
headers=subdir_glob([
|
||||
('dictBuilder', 'divsufsort.h'),
|
||||
('dictBuilder', 'cover.h'),
|
||||
]),
|
||||
srcs=glob(['dictBuilder/*.c']),
|
||||
deps=[':common'],
|
||||
|
||||
+68
-79
@@ -29,6 +29,7 @@
|
||||
#include "mem.h" /* read */
|
||||
#include "pool.h"
|
||||
#include "threading.h"
|
||||
#include "cover.h"
|
||||
#include "zstd_internal.h" /* includes zstd.h */
|
||||
#ifndef ZDICT_STATIC_LINKING_ONLY
|
||||
#define ZDICT_STATIC_LINKING_ONLY
|
||||
@@ -185,7 +186,7 @@ static void COVER_map_remove(COVER_map_t *map, U32 key) {
|
||||
}
|
||||
|
||||
/**
|
||||
* Destroyes a map that is inited with COVER_map_init().
|
||||
* Destroys a map that is inited with COVER_map_init().
|
||||
*/
|
||||
static void COVER_map_destroy(COVER_map_t *map) {
|
||||
if (map->data) {
|
||||
@@ -223,7 +224,7 @@ static COVER_ctx_t *g_ctx = NULL;
|
||||
/**
|
||||
* Returns the sum of the sample sizes.
|
||||
*/
|
||||
static size_t COVER_sum(const size_t *samplesSizes, unsigned nbSamples) {
|
||||
size_t COVER_sum(const size_t *samplesSizes, unsigned nbSamples) {
|
||||
size_t sum = 0;
|
||||
unsigned i;
|
||||
for (i = 0; i < nbSamples; ++i) {
|
||||
@@ -380,14 +381,6 @@ static void COVER_group(COVER_ctx_t *ctx, const void *group,
|
||||
ctx->suffix[dmerId] = freq;
|
||||
}
|
||||
|
||||
/**
|
||||
* A segment is a range in the source as well as the score of the segment.
|
||||
*/
|
||||
typedef struct {
|
||||
U32 begin;
|
||||
U32 end;
|
||||
U32 score;
|
||||
} COVER_segment_t;
|
||||
|
||||
/**
|
||||
* Selects the best segment in an epoch.
|
||||
@@ -691,7 +684,7 @@ ZDICTLIB_API size_t ZDICT_trainFromBuffer_cover(
|
||||
BYTE* const dict = (BYTE*)dictBuffer;
|
||||
COVER_ctx_t ctx;
|
||||
COVER_map_t activeDmers;
|
||||
parameters.splitPoint = 1.0;
|
||||
parameters.splitPoint = parameters.splitPoint <= 0.0 ? DEFAULT_SPLITPOINT : parameters.splitPoint;
|
||||
/* Initialize global data */
|
||||
g_displayLevel = parameters.zParams.notificationLevel;
|
||||
/* Checks */
|
||||
@@ -737,28 +730,65 @@ ZDICTLIB_API size_t ZDICT_trainFromBuffer_cover(
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* COVER_best_t is used for two purposes:
|
||||
* 1. Synchronizing threads.
|
||||
* 2. Saving the best parameters and dictionary.
|
||||
*
|
||||
* All of the methods except COVER_best_init() are thread safe if zstd is
|
||||
* compiled with multithreaded support.
|
||||
*/
|
||||
typedef struct COVER_best_s {
|
||||
ZSTD_pthread_mutex_t mutex;
|
||||
ZSTD_pthread_cond_t cond;
|
||||
size_t liveJobs;
|
||||
void *dict;
|
||||
size_t dictSize;
|
||||
ZDICT_cover_params_t parameters;
|
||||
size_t compressedSize;
|
||||
} COVER_best_t;
|
||||
|
||||
|
||||
size_t COVER_checkTotalCompressedSize(const ZDICT_cover_params_t parameters,
|
||||
const size_t *samplesSizes, const BYTE *samples,
|
||||
size_t *offsets,
|
||||
size_t nbTrainSamples, size_t nbSamples,
|
||||
BYTE *const dict, size_t dictBufferCapacity) {
|
||||
size_t totalCompressedSize = ERROR(GENERIC);
|
||||
/* Pointers */
|
||||
ZSTD_CCtx *cctx;
|
||||
ZSTD_CDict *cdict;
|
||||
void *dst;
|
||||
/* Local variables */
|
||||
size_t dstCapacity;
|
||||
size_t i;
|
||||
/* Allocate dst with enough space to compress the maximum sized sample */
|
||||
{
|
||||
size_t maxSampleSize = 0;
|
||||
i = parameters.splitPoint < 1.0 ? nbTrainSamples : 0;
|
||||
for (; i < nbSamples; ++i) {
|
||||
maxSampleSize = MAX(samplesSizes[i], maxSampleSize);
|
||||
}
|
||||
dstCapacity = ZSTD_compressBound(maxSampleSize);
|
||||
dst = malloc(dstCapacity);
|
||||
}
|
||||
/* Create the cctx and cdict */
|
||||
cctx = ZSTD_createCCtx();
|
||||
cdict = ZSTD_createCDict(dict, dictBufferCapacity,
|
||||
parameters.zParams.compressionLevel);
|
||||
if (!dst || !cctx || !cdict) {
|
||||
goto _compressCleanup;
|
||||
}
|
||||
/* Compress each sample and sum their sizes (or error) */
|
||||
totalCompressedSize = dictBufferCapacity;
|
||||
i = parameters.splitPoint < 1.0 ? nbTrainSamples : 0;
|
||||
for (; i < nbSamples; ++i) {
|
||||
const size_t size = ZSTD_compress_usingCDict(
|
||||
cctx, dst, dstCapacity, samples + offsets[i],
|
||||
samplesSizes[i], cdict);
|
||||
if (ZSTD_isError(size)) {
|
||||
totalCompressedSize = ERROR(GENERIC);
|
||||
goto _compressCleanup;
|
||||
}
|
||||
totalCompressedSize += size;
|
||||
}
|
||||
_compressCleanup:
|
||||
ZSTD_freeCCtx(cctx);
|
||||
ZSTD_freeCDict(cdict);
|
||||
if (dst) {
|
||||
free(dst);
|
||||
}
|
||||
return totalCompressedSize;
|
||||
}
|
||||
|
||||
|
||||
/**
|
||||
* Initialize the `COVER_best_t`.
|
||||
*/
|
||||
static void COVER_best_init(COVER_best_t *best) {
|
||||
void COVER_best_init(COVER_best_t *best) {
|
||||
if (best==NULL) return; /* compatible with init on NULL */
|
||||
(void)ZSTD_pthread_mutex_init(&best->mutex, NULL);
|
||||
(void)ZSTD_pthread_cond_init(&best->cond, NULL);
|
||||
@@ -772,7 +802,7 @@ static void COVER_best_init(COVER_best_t *best) {
|
||||
/**
|
||||
* Wait until liveJobs == 0.
|
||||
*/
|
||||
static void COVER_best_wait(COVER_best_t *best) {
|
||||
void COVER_best_wait(COVER_best_t *best) {
|
||||
if (!best) {
|
||||
return;
|
||||
}
|
||||
@@ -786,7 +816,7 @@ static void COVER_best_wait(COVER_best_t *best) {
|
||||
/**
|
||||
* Call COVER_best_wait() and then destroy the COVER_best_t.
|
||||
*/
|
||||
static void COVER_best_destroy(COVER_best_t *best) {
|
||||
void COVER_best_destroy(COVER_best_t *best) {
|
||||
if (!best) {
|
||||
return;
|
||||
}
|
||||
@@ -802,7 +832,7 @@ static void COVER_best_destroy(COVER_best_t *best) {
|
||||
* Called when a thread is about to be launched.
|
||||
* Increments liveJobs.
|
||||
*/
|
||||
static void COVER_best_start(COVER_best_t *best) {
|
||||
void COVER_best_start(COVER_best_t *best) {
|
||||
if (!best) {
|
||||
return;
|
||||
}
|
||||
@@ -816,7 +846,7 @@ static void COVER_best_start(COVER_best_t *best) {
|
||||
* Decrements liveJobs and signals any waiting threads if liveJobs == 0.
|
||||
* If this dictionary is the best so far save it and its parameters.
|
||||
*/
|
||||
static void COVER_best_finish(COVER_best_t *best, size_t compressedSize,
|
||||
void COVER_best_finish(COVER_best_t *best, size_t compressedSize,
|
||||
ZDICT_cover_params_t parameters, void *dict,
|
||||
size_t dictSize) {
|
||||
if (!best) {
|
||||
@@ -847,10 +877,10 @@ static void COVER_best_finish(COVER_best_t *best, size_t compressedSize,
|
||||
best->parameters = parameters;
|
||||
best->compressedSize = compressedSize;
|
||||
}
|
||||
ZSTD_pthread_mutex_unlock(&best->mutex);
|
||||
if (liveJobs == 0) {
|
||||
ZSTD_pthread_cond_broadcast(&best->cond);
|
||||
}
|
||||
ZSTD_pthread_mutex_unlock(&best->mutex);
|
||||
}
|
||||
}
|
||||
|
||||
@@ -904,51 +934,10 @@ static void COVER_tryParameters(void *opaque) {
|
||||
}
|
||||
}
|
||||
/* Check total compressed size */
|
||||
{
|
||||
/* Pointers */
|
||||
ZSTD_CCtx *cctx;
|
||||
ZSTD_CDict *cdict;
|
||||
void *dst;
|
||||
/* Local variables */
|
||||
size_t dstCapacity;
|
||||
size_t i;
|
||||
/* Allocate dst with enough space to compress the maximum sized sample */
|
||||
{
|
||||
size_t maxSampleSize = 0;
|
||||
i = parameters.splitPoint < 1.0 ? ctx->nbTrainSamples : 0;
|
||||
for (; i < ctx->nbSamples; ++i) {
|
||||
maxSampleSize = MAX(ctx->samplesSizes[i], maxSampleSize);
|
||||
}
|
||||
dstCapacity = ZSTD_compressBound(maxSampleSize);
|
||||
dst = malloc(dstCapacity);
|
||||
}
|
||||
/* Create the cctx and cdict */
|
||||
cctx = ZSTD_createCCtx();
|
||||
cdict = ZSTD_createCDict(dict, dictBufferCapacity,
|
||||
parameters.zParams.compressionLevel);
|
||||
if (!dst || !cctx || !cdict) {
|
||||
goto _compressCleanup;
|
||||
}
|
||||
/* Compress each sample and sum their sizes (or error) */
|
||||
totalCompressedSize = dictBufferCapacity;
|
||||
i = parameters.splitPoint < 1.0 ? ctx->nbTrainSamples : 0;
|
||||
for (; i < ctx->nbSamples; ++i) {
|
||||
const size_t size = ZSTD_compress_usingCDict(
|
||||
cctx, dst, dstCapacity, ctx->samples + ctx->offsets[i],
|
||||
ctx->samplesSizes[i], cdict);
|
||||
if (ZSTD_isError(size)) {
|
||||
totalCompressedSize = ERROR(GENERIC);
|
||||
goto _compressCleanup;
|
||||
}
|
||||
totalCompressedSize += size;
|
||||
}
|
||||
_compressCleanup:
|
||||
ZSTD_freeCCtx(cctx);
|
||||
ZSTD_freeCDict(cdict);
|
||||
if (dst) {
|
||||
free(dst);
|
||||
}
|
||||
}
|
||||
totalCompressedSize = COVER_checkTotalCompressedSize(parameters, ctx->samplesSizes,
|
||||
ctx->samples, ctx->offsets,
|
||||
ctx->nbTrainSamples, ctx->nbSamples,
|
||||
dict, dictBufferCapacity);
|
||||
|
||||
_cleanup:
|
||||
COVER_best_finish(data->best, totalCompressedSize, parameters, dict,
|
||||
|
||||
@@ -0,0 +1,83 @@
|
||||
#include <stdio.h> /* fprintf */
|
||||
#include <stdlib.h> /* malloc, free, qsort */
|
||||
#include <string.h> /* memset */
|
||||
#include <time.h> /* clock */
|
||||
#include "mem.h" /* read */
|
||||
#include "pool.h"
|
||||
#include "threading.h"
|
||||
#include "zstd_internal.h" /* includes zstd.h */
|
||||
#ifndef ZDICT_STATIC_LINKING_ONLY
|
||||
#define ZDICT_STATIC_LINKING_ONLY
|
||||
#endif
|
||||
#include "zdict.h"
|
||||
|
||||
/**
|
||||
* COVER_best_t is used for two purposes:
|
||||
* 1. Synchronizing threads.
|
||||
* 2. Saving the best parameters and dictionary.
|
||||
*
|
||||
* All of the methods except COVER_best_init() are thread safe if zstd is
|
||||
* compiled with multithreaded support.
|
||||
*/
|
||||
typedef struct COVER_best_s {
|
||||
ZSTD_pthread_mutex_t mutex;
|
||||
ZSTD_pthread_cond_t cond;
|
||||
size_t liveJobs;
|
||||
void *dict;
|
||||
size_t dictSize;
|
||||
ZDICT_cover_params_t parameters;
|
||||
size_t compressedSize;
|
||||
} COVER_best_t;
|
||||
|
||||
/**
|
||||
* A segment is a range in the source as well as the score of the segment.
|
||||
*/
|
||||
typedef struct {
|
||||
U32 begin;
|
||||
U32 end;
|
||||
U32 score;
|
||||
} COVER_segment_t;
|
||||
|
||||
/**
|
||||
* Checks total compressed size of a dictionary
|
||||
*/
|
||||
size_t COVER_checkTotalCompressedSize(const ZDICT_cover_params_t parameters,
|
||||
const size_t *samplesSizes, const BYTE *samples,
|
||||
size_t *offsets,
|
||||
size_t nbTrainSamples, size_t nbSamples,
|
||||
BYTE *const dict, size_t dictBufferCapacity);
|
||||
|
||||
/**
|
||||
* Returns the sum of the sample sizes.
|
||||
*/
|
||||
size_t COVER_sum(const size_t *samplesSizes, unsigned nbSamples) ;
|
||||
|
||||
/**
|
||||
* Initialize the `COVER_best_t`.
|
||||
*/
|
||||
void COVER_best_init(COVER_best_t *best);
|
||||
|
||||
/**
|
||||
* Wait until liveJobs == 0.
|
||||
*/
|
||||
void COVER_best_wait(COVER_best_t *best);
|
||||
|
||||
/**
|
||||
* Call COVER_best_wait() and then destroy the COVER_best_t.
|
||||
*/
|
||||
void COVER_best_destroy(COVER_best_t *best);
|
||||
|
||||
/**
|
||||
* Called when a thread is about to be launched.
|
||||
* Increments liveJobs.
|
||||
*/
|
||||
void COVER_best_start(COVER_best_t *best);
|
||||
|
||||
/**
|
||||
* Called when a thread finishes executing, both on error or success.
|
||||
* Decrements liveJobs and signals any waiting threads if liveJobs == 0.
|
||||
* If this dictionary is the best so far save it and its parameters.
|
||||
*/
|
||||
void COVER_best_finish(COVER_best_t *best, size_t compressedSize,
|
||||
ZDICT_cover_params_t parameters, void *dict,
|
||||
size_t dictSize);
|
||||
@@ -0,0 +1,701 @@
|
||||
/*-*************************************
|
||||
* Dependencies
|
||||
***************************************/
|
||||
#include <stdio.h> /* fprintf */
|
||||
#include <stdlib.h> /* malloc, free, qsort */
|
||||
#include <string.h> /* memset */
|
||||
#include <time.h> /* clock */
|
||||
|
||||
#include "mem.h" /* read */
|
||||
#include "pool.h"
|
||||
#include "threading.h"
|
||||
#include "cover.h"
|
||||
#include "zstd_internal.h" /* includes zstd.h */
|
||||
#ifndef ZDICT_STATIC_LINKING_ONLY
|
||||
#define ZDICT_STATIC_LINKING_ONLY
|
||||
#endif
|
||||
#include "zdict.h"
|
||||
|
||||
|
||||
/*-*************************************
|
||||
* Constants
|
||||
***************************************/
|
||||
#define FASTCOVER_MAX_SAMPLES_SIZE (sizeof(size_t) == 8 ? ((U32)-1) : ((U32)1 GB))
|
||||
#define FASTCOVER_MAX_F 31
|
||||
#define FASTCOVER_MAX_ACCEL 10
|
||||
#define DEFAULT_SPLITPOINT 0.75
|
||||
#define DEFAULT_F 18
|
||||
#define DEFAULT_ACCEL 1
|
||||
|
||||
|
||||
/*-*************************************
|
||||
* Console display
|
||||
***************************************/
|
||||
static int g_displayLevel = 2;
|
||||
#define DISPLAY(...) \
|
||||
{ \
|
||||
fprintf(stderr, __VA_ARGS__); \
|
||||
fflush(stderr); \
|
||||
}
|
||||
#define LOCALDISPLAYLEVEL(displayLevel, l, ...) \
|
||||
if (displayLevel >= l) { \
|
||||
DISPLAY(__VA_ARGS__); \
|
||||
} /* 0 : no display; 1: errors; 2: default; 3: details; 4: debug */
|
||||
#define DISPLAYLEVEL(l, ...) LOCALDISPLAYLEVEL(g_displayLevel, l, __VA_ARGS__)
|
||||
|
||||
#define LOCALDISPLAYUPDATE(displayLevel, l, ...) \
|
||||
if (displayLevel >= l) { \
|
||||
if ((clock() - g_time > refreshRate) || (displayLevel >= 4)) { \
|
||||
g_time = clock(); \
|
||||
DISPLAY(__VA_ARGS__); \
|
||||
} \
|
||||
}
|
||||
#define DISPLAYUPDATE(l, ...) LOCALDISPLAYUPDATE(g_displayLevel, l, __VA_ARGS__)
|
||||
static const clock_t refreshRate = CLOCKS_PER_SEC * 15 / 100;
|
||||
static clock_t g_time = 0;
|
||||
|
||||
|
||||
/*-*************************************
|
||||
* Hash Functions
|
||||
***************************************/
|
||||
static const U64 prime6bytes = 227718039650203ULL;
|
||||
static size_t ZSTD_hash6(U64 u, U32 h) { return (size_t)(((u << (64-48)) * prime6bytes) >> (64-h)) ; }
|
||||
static size_t ZSTD_hash6Ptr(const void* p, U32 h) { return ZSTD_hash6(MEM_readLE64(p), h); }
|
||||
|
||||
static const U64 prime8bytes = 0xCF1BBCDCB7A56463ULL;
|
||||
static size_t ZSTD_hash8(U64 u, U32 h) { return (size_t)(((u) * prime8bytes) >> (64-h)) ; }
|
||||
static size_t ZSTD_hash8Ptr(const void* p, U32 h) { return ZSTD_hash8(MEM_readLE64(p), h); }
|
||||
|
||||
|
||||
/**
|
||||
* Hash the d-byte value pointed to by p and mod 2^f
|
||||
*/
|
||||
static size_t FASTCOVER_hashPtrToIndex(const void* p, U32 h, unsigned d) {
|
||||
if (d == 6) {
|
||||
return ZSTD_hash6Ptr(p, h) & ((1 << h) - 1);
|
||||
}
|
||||
return ZSTD_hash8Ptr(p, h) & ((1 << h) - 1);
|
||||
}
|
||||
|
||||
|
||||
/*-*************************************
|
||||
* Acceleration
|
||||
***************************************/
|
||||
typedef struct {
|
||||
unsigned finalize; /* Percentage of training samples used for ZDICT_finalizeDictionary */
|
||||
unsigned skip; /* Number of dmer skipped between each dmer counted in computeFrequency */
|
||||
} FASTCOVER_accel_t;
|
||||
|
||||
|
||||
static const FASTCOVER_accel_t FASTCOVER_defaultAccelParameters[FASTCOVER_MAX_ACCEL+1] = {
|
||||
{ 100, 0 }, /* accel = 0, should not happen because accel = 0 defaults to accel = 1 */
|
||||
{ 100, 0 }, /* accel = 1 */
|
||||
{ 50, 1 }, /* accel = 2 */
|
||||
{ 34, 2 }, /* accel = 3 */
|
||||
{ 25, 3 }, /* accel = 4 */
|
||||
{ 20, 4 }, /* accel = 5 */
|
||||
{ 17, 5 }, /* accel = 6 */
|
||||
{ 14, 6 }, /* accel = 7 */
|
||||
{ 13, 7 }, /* accel = 8 */
|
||||
{ 11, 8 }, /* accel = 9 */
|
||||
{ 10, 9 }, /* accel = 10 */
|
||||
};
|
||||
|
||||
|
||||
/*-*************************************
|
||||
* Context
|
||||
***************************************/
|
||||
typedef struct {
|
||||
const BYTE *samples;
|
||||
size_t *offsets;
|
||||
const size_t *samplesSizes;
|
||||
size_t nbSamples;
|
||||
size_t nbTrainSamples;
|
||||
size_t nbTestSamples;
|
||||
size_t nbDmers;
|
||||
U32 *freqs;
|
||||
unsigned d;
|
||||
unsigned f;
|
||||
FASTCOVER_accel_t accelParams;
|
||||
} FASTCOVER_ctx_t;
|
||||
|
||||
|
||||
/*-*************************************
|
||||
* Helper functions
|
||||
***************************************/
|
||||
/**
|
||||
* Selects the best segment in an epoch.
|
||||
* Segments of are scored according to the function:
|
||||
*
|
||||
* Let F(d) be the frequency of all dmers with hash value d.
|
||||
* Let S_i be hash value of the dmer at position i of segment S which has length k.
|
||||
*
|
||||
* Score(S) = F(S_1) + F(S_2) + ... + F(S_{k-d+1})
|
||||
*
|
||||
* Once the dmer with hash value d is in the dictionay we set F(d) = 0.
|
||||
*/
|
||||
static COVER_segment_t FASTCOVER_selectSegment(const FASTCOVER_ctx_t *ctx,
|
||||
U32 *freqs, U32 begin, U32 end,
|
||||
ZDICT_cover_params_t parameters,
|
||||
U16* segmentFreqs) {
|
||||
/* Constants */
|
||||
const U32 k = parameters.k;
|
||||
const U32 d = parameters.d;
|
||||
const U32 f = ctx->f;
|
||||
const U32 dmersInK = k - d + 1;
|
||||
|
||||
/* Try each segment (activeSegment) and save the best (bestSegment) */
|
||||
COVER_segment_t bestSegment = {0, 0, 0};
|
||||
COVER_segment_t activeSegment;
|
||||
|
||||
/* Reset the activeDmers in the segment */
|
||||
/* The activeSegment starts at the beginning of the epoch. */
|
||||
activeSegment.begin = begin;
|
||||
activeSegment.end = begin;
|
||||
activeSegment.score = 0;
|
||||
|
||||
/* Slide the activeSegment through the whole epoch.
|
||||
* Save the best segment in bestSegment.
|
||||
*/
|
||||
while (activeSegment.end < end) {
|
||||
/* Get hash value of current dmer */
|
||||
const size_t index = FASTCOVER_hashPtrToIndex(ctx->samples + activeSegment.end, f, d);
|
||||
|
||||
/* Add frequency of this index to score if this is the first occurence of index in active segment */
|
||||
if (segmentFreqs[index] == 0) {
|
||||
activeSegment.score += freqs[index];
|
||||
}
|
||||
/* Increment end of segment and segmentFreqs*/
|
||||
activeSegment.end += 1;
|
||||
segmentFreqs[index] += 1;
|
||||
/* If the window is now too large, drop the first position */
|
||||
if (activeSegment.end - activeSegment.begin == dmersInK + 1) {
|
||||
/* Get hash value of the dmer to be eliminated from active segment */
|
||||
const size_t delIndex = FASTCOVER_hashPtrToIndex(ctx->samples + activeSegment.begin, f, d);
|
||||
segmentFreqs[delIndex] -= 1;
|
||||
/* Subtract frequency of this index from score if this is the last occurrence of this index in active segment */
|
||||
if (segmentFreqs[delIndex] == 0) {
|
||||
activeSegment.score -= freqs[delIndex];
|
||||
}
|
||||
/* Increment start of segment */
|
||||
activeSegment.begin += 1;
|
||||
}
|
||||
|
||||
/* If this segment is the best so far save it */
|
||||
if (activeSegment.score > bestSegment.score) {
|
||||
bestSegment = activeSegment;
|
||||
}
|
||||
}
|
||||
|
||||
/* Zero out rest of segmentFreqs array */
|
||||
while (activeSegment.begin < end) {
|
||||
const size_t delIndex = FASTCOVER_hashPtrToIndex(ctx->samples + activeSegment.begin, f, d);
|
||||
segmentFreqs[delIndex] -= 1;
|
||||
activeSegment.begin += 1;
|
||||
}
|
||||
|
||||
{
|
||||
/* Zero the frequency of hash value of each dmer covered by the chosen segment. */
|
||||
U32 pos;
|
||||
for (pos = bestSegment.begin; pos != bestSegment.end; ++pos) {
|
||||
const size_t i = FASTCOVER_hashPtrToIndex(ctx->samples + pos, f, d);
|
||||
freqs[i] = 0;
|
||||
}
|
||||
}
|
||||
|
||||
return bestSegment;
|
||||
}
|
||||
|
||||
|
||||
static int FASTCOVER_checkParameters(ZDICT_cover_params_t parameters,
|
||||
size_t maxDictSize, unsigned f,
|
||||
unsigned accel) {
|
||||
/* k, d, and f are required parameters */
|
||||
if (parameters.d == 0 || parameters.k == 0) {
|
||||
return 0;
|
||||
}
|
||||
/* d has to be 6 or 8 */
|
||||
if (parameters.d != 6 && parameters.d != 8) {
|
||||
return 0;
|
||||
}
|
||||
/* k <= maxDictSize */
|
||||
if (parameters.k > maxDictSize) {
|
||||
return 0;
|
||||
}
|
||||
/* d <= k */
|
||||
if (parameters.d > parameters.k) {
|
||||
return 0;
|
||||
}
|
||||
/* 0 < f <= FASTCOVER_MAX_F*/
|
||||
if (f > FASTCOVER_MAX_F || f == 0) {
|
||||
return 0;
|
||||
}
|
||||
/* 0 < splitPoint <= 1 */
|
||||
if (parameters.splitPoint <= 0 || parameters.splitPoint > 1) {
|
||||
return 0;
|
||||
}
|
||||
/* 0 < accel <= 10 */
|
||||
if (accel > 10 || accel == 0) {
|
||||
return 0;
|
||||
}
|
||||
return 1;
|
||||
}
|
||||
|
||||
|
||||
/**
|
||||
* Clean up a context initialized with `FASTCOVER_ctx_init()`.
|
||||
*/
|
||||
static void FASTCOVER_ctx_destroy(FASTCOVER_ctx_t *ctx) {
|
||||
if (!ctx) {
|
||||
return;
|
||||
}
|
||||
|
||||
free(ctx->freqs);
|
||||
ctx->freqs = NULL;
|
||||
|
||||
free(ctx->offsets);
|
||||
ctx->offsets = NULL;
|
||||
}
|
||||
|
||||
|
||||
/**
|
||||
* Calculate for frequency of hash value of each dmer in ctx->samples
|
||||
*/
|
||||
static void FASTCOVER_computeFrequency(U32 *freqs, FASTCOVER_ctx_t *ctx){
|
||||
const unsigned f = ctx->f;
|
||||
const unsigned d = ctx->d;
|
||||
const unsigned skip = ctx->accelParams.skip;
|
||||
const unsigned readLength = MAX(d, 8);
|
||||
size_t start; /* start of current dmer */
|
||||
size_t i;
|
||||
for (i = 0; i < ctx->nbTrainSamples; i++) {
|
||||
size_t currSampleStart = ctx->offsets[i];
|
||||
size_t currSampleEnd = ctx->offsets[i+1];
|
||||
start = currSampleStart;
|
||||
while (start + readLength <= currSampleEnd) {
|
||||
const size_t dmerIndex = FASTCOVER_hashPtrToIndex(ctx->samples + start, f, d);
|
||||
freqs[dmerIndex]++;
|
||||
start = start + skip + 1;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
/**
|
||||
* Prepare a context for dictionary building.
|
||||
* The context is only dependent on the parameter `d` and can used multiple
|
||||
* times.
|
||||
* Returns 1 on success or zero on error.
|
||||
* The context must be destroyed with `FASTCOVER_ctx_destroy()`.
|
||||
*/
|
||||
static int FASTCOVER_ctx_init(FASTCOVER_ctx_t *ctx, const void *samplesBuffer,
|
||||
const size_t *samplesSizes, unsigned nbSamples,
|
||||
unsigned d, double splitPoint, unsigned f,
|
||||
FASTCOVER_accel_t accelParams) {
|
||||
const BYTE *const samples = (const BYTE *)samplesBuffer;
|
||||
const size_t totalSamplesSize = COVER_sum(samplesSizes, nbSamples);
|
||||
/* Split samples into testing and training sets */
|
||||
const unsigned nbTrainSamples = splitPoint < 1.0 ? (unsigned)((double)nbSamples * splitPoint) : nbSamples;
|
||||
const unsigned nbTestSamples = splitPoint < 1.0 ? nbSamples - nbTrainSamples : nbSamples;
|
||||
const size_t trainingSamplesSize = splitPoint < 1.0 ? COVER_sum(samplesSizes, nbTrainSamples) : totalSamplesSize;
|
||||
const size_t testSamplesSize = splitPoint < 1.0 ? COVER_sum(samplesSizes + nbTrainSamples, nbTestSamples) : totalSamplesSize;
|
||||
/* Checks */
|
||||
if (totalSamplesSize < MAX(d, sizeof(U64)) ||
|
||||
totalSamplesSize >= (size_t)FASTCOVER_MAX_SAMPLES_SIZE) {
|
||||
DISPLAYLEVEL(1, "Total samples size is too large (%u MB), maximum size is %u MB\n",
|
||||
(U32)(totalSamplesSize >> 20), (FASTCOVER_MAX_SAMPLES_SIZE >> 20));
|
||||
return 0;
|
||||
}
|
||||
/* Check if there are at least 5 training samples */
|
||||
if (nbTrainSamples < 5) {
|
||||
DISPLAYLEVEL(1, "Total number of training samples is %u and is invalid\n", nbTrainSamples);
|
||||
return 0;
|
||||
}
|
||||
/* Check if there's testing sample */
|
||||
if (nbTestSamples < 1) {
|
||||
DISPLAYLEVEL(1, "Total number of testing samples is %u and is invalid.\n", nbTestSamples);
|
||||
return 0;
|
||||
}
|
||||
/* Zero the context */
|
||||
memset(ctx, 0, sizeof(*ctx));
|
||||
DISPLAYLEVEL(2, "Training on %u samples of total size %u\n", nbTrainSamples,
|
||||
(U32)trainingSamplesSize);
|
||||
DISPLAYLEVEL(2, "Testing on %u samples of total size %u\n", nbTestSamples,
|
||||
(U32)testSamplesSize);
|
||||
|
||||
ctx->samples = samples;
|
||||
ctx->samplesSizes = samplesSizes;
|
||||
ctx->nbSamples = nbSamples;
|
||||
ctx->nbTrainSamples = nbTrainSamples;
|
||||
ctx->nbTestSamples = nbTestSamples;
|
||||
ctx->nbDmers = trainingSamplesSize - MAX(d, sizeof(U64)) + 1;
|
||||
ctx->d = d;
|
||||
ctx->f = f;
|
||||
ctx->accelParams = accelParams;
|
||||
|
||||
/* The offsets of each file */
|
||||
ctx->offsets = (size_t *)malloc((nbSamples + 1) * sizeof(size_t));
|
||||
if (!ctx->offsets) {
|
||||
DISPLAYLEVEL(1, "Failed to allocate scratch buffers\n");
|
||||
FASTCOVER_ctx_destroy(ctx);
|
||||
return 0;
|
||||
}
|
||||
|
||||
/* Fill offsets from the samplesSizes */
|
||||
{
|
||||
U32 i;
|
||||
ctx->offsets[0] = 0;
|
||||
for (i = 1; i <= nbSamples; ++i) {
|
||||
ctx->offsets[i] = ctx->offsets[i - 1] + samplesSizes[i - 1];
|
||||
}
|
||||
}
|
||||
|
||||
/* Initialize frequency array of size 2^f */
|
||||
ctx->freqs = (U32 *)calloc(((U64)1 << f), sizeof(U32));
|
||||
|
||||
DISPLAYLEVEL(2, "Computing frequencies\n");
|
||||
FASTCOVER_computeFrequency(ctx->freqs, ctx);
|
||||
|
||||
return 1;
|
||||
}
|
||||
|
||||
|
||||
/**
|
||||
* Given the prepared context build the dictionary.
|
||||
*/
|
||||
static size_t FASTCOVER_buildDictionary(const FASTCOVER_ctx_t *ctx, U32 *freqs,
|
||||
void *dictBuffer, size_t dictBufferCapacity,
|
||||
ZDICT_cover_params_t parameters, U16* segmentFreqs){
|
||||
BYTE *const dict = (BYTE *)dictBuffer;
|
||||
size_t tail = dictBufferCapacity;
|
||||
/* Divide the data up into epochs of equal size.
|
||||
* We will select at least one segment from each epoch.
|
||||
*/
|
||||
const U32 epochs = MAX(1, (U32)(dictBufferCapacity / parameters.k));
|
||||
const U32 epochSize = (U32)(ctx->nbDmers / epochs);
|
||||
size_t epoch;
|
||||
DISPLAYLEVEL(2, "Breaking content into %u epochs of size %u\n", epochs,
|
||||
epochSize);
|
||||
/* Loop through the epochs until there are no more segments or the dictionary
|
||||
* is full.
|
||||
*/
|
||||
for (epoch = 0; tail > 0; epoch = (epoch + 1) % epochs) {
|
||||
const U32 epochBegin = (U32)(epoch * epochSize);
|
||||
const U32 epochEnd = epochBegin + epochSize;
|
||||
size_t segmentSize;
|
||||
/* Select a segment */
|
||||
COVER_segment_t segment = FASTCOVER_selectSegment(
|
||||
ctx, freqs, epochBegin, epochEnd, parameters, segmentFreqs);
|
||||
|
||||
/* If the segment covers no dmers, then we are out of content */
|
||||
if (segment.score == 0) {
|
||||
break;
|
||||
}
|
||||
|
||||
/* Trim the segment if necessary and if it is too small then we are done */
|
||||
segmentSize = MIN(segment.end - segment.begin + parameters.d - 1, tail);
|
||||
if (segmentSize < parameters.d) {
|
||||
break;
|
||||
}
|
||||
|
||||
/* We fill the dictionary from the back to allow the best segments to be
|
||||
* referenced with the smallest offsets.
|
||||
*/
|
||||
tail -= segmentSize;
|
||||
memcpy(dict + tail, ctx->samples + segment.begin, segmentSize);
|
||||
DISPLAYUPDATE(
|
||||
2, "\r%u%% ",
|
||||
(U32)(((dictBufferCapacity - tail) * 100) / dictBufferCapacity));
|
||||
}
|
||||
DISPLAYLEVEL(2, "\r%79s\r", "");
|
||||
return tail;
|
||||
}
|
||||
|
||||
|
||||
/**
|
||||
* Parameters for FASTCOVER_tryParameters().
|
||||
*/
|
||||
typedef struct FASTCOVER_tryParameters_data_s {
|
||||
const FASTCOVER_ctx_t *ctx;
|
||||
COVER_best_t *best;
|
||||
size_t dictBufferCapacity;
|
||||
ZDICT_cover_params_t parameters;
|
||||
} FASTCOVER_tryParameters_data_t;
|
||||
|
||||
|
||||
/**
|
||||
* Tries a set of parameters and updates the COVER_best_t with the results.
|
||||
* This function is thread safe if zstd is compiled with multithreaded support.
|
||||
* It takes its parameters as an *OWNING* opaque pointer to support threading.
|
||||
*/
|
||||
static void FASTCOVER_tryParameters(void *opaque) {
|
||||
/* Save parameters as local variables */
|
||||
FASTCOVER_tryParameters_data_t *const data = (FASTCOVER_tryParameters_data_t *)opaque;
|
||||
const FASTCOVER_ctx_t *const ctx = data->ctx;
|
||||
const ZDICT_cover_params_t parameters = data->parameters;
|
||||
size_t dictBufferCapacity = data->dictBufferCapacity;
|
||||
size_t totalCompressedSize = ERROR(GENERIC);
|
||||
/* Initialize array to keep track of frequency of dmer within activeSegment */
|
||||
U16* segmentFreqs = (U16 *)calloc(((U64)1 << ctx->f), sizeof(U16));
|
||||
/* Allocate space for hash table, dict, and freqs */
|
||||
BYTE *const dict = (BYTE * const)malloc(dictBufferCapacity);
|
||||
U32 *freqs = (U32*) malloc(((U64)1 << ctx->f) * sizeof(U32));
|
||||
if (!segmentFreqs || !dict || !freqs) {
|
||||
DISPLAYLEVEL(1, "Failed to allocate buffers: out of memory\n");
|
||||
goto _cleanup;
|
||||
}
|
||||
/* Copy the frequencies because we need to modify them */
|
||||
memcpy(freqs, ctx->freqs, ((U64)1 << ctx->f) * sizeof(U32));
|
||||
/* Build the dictionary */
|
||||
{
|
||||
const size_t tail = FASTCOVER_buildDictionary(ctx, freqs, dict, dictBufferCapacity,
|
||||
parameters, segmentFreqs);
|
||||
const unsigned nbFinalizeSamples = (unsigned)(ctx->nbTrainSamples * ctx->accelParams.finalize / 100);
|
||||
dictBufferCapacity = ZDICT_finalizeDictionary(
|
||||
dict, dictBufferCapacity, dict + tail, dictBufferCapacity - tail,
|
||||
ctx->samples, ctx->samplesSizes, nbFinalizeSamples, parameters.zParams);
|
||||
if (ZDICT_isError(dictBufferCapacity)) {
|
||||
DISPLAYLEVEL(1, "Failed to finalize dictionary\n");
|
||||
goto _cleanup;
|
||||
}
|
||||
}
|
||||
/* Check total compressed size */
|
||||
totalCompressedSize = COVER_checkTotalCompressedSize(parameters, ctx->samplesSizes,
|
||||
ctx->samples, ctx->offsets,
|
||||
ctx->nbTrainSamples, ctx->nbSamples,
|
||||
dict, dictBufferCapacity);
|
||||
_cleanup:
|
||||
COVER_best_finish(data->best, totalCompressedSize, parameters, dict,
|
||||
dictBufferCapacity);
|
||||
free(data);
|
||||
free(segmentFreqs);
|
||||
free(dict);
|
||||
free(freqs);
|
||||
}
|
||||
|
||||
|
||||
|
||||
static void FASTCOVER_convertToCoverParams(ZDICT_fastCover_params_t fastCoverParams,
|
||||
ZDICT_cover_params_t *coverParams) {
|
||||
coverParams->k = fastCoverParams.k;
|
||||
coverParams->d = fastCoverParams.d;
|
||||
coverParams->steps = fastCoverParams.steps;
|
||||
coverParams->nbThreads = fastCoverParams.nbThreads;
|
||||
coverParams->splitPoint = fastCoverParams.splitPoint;
|
||||
coverParams->zParams = fastCoverParams.zParams;
|
||||
}
|
||||
|
||||
|
||||
static void FASTCOVER_convertToFastCoverParams(ZDICT_cover_params_t coverParams,
|
||||
ZDICT_fastCover_params_t *fastCoverParams,
|
||||
unsigned f, unsigned accel) {
|
||||
fastCoverParams->k = coverParams.k;
|
||||
fastCoverParams->d = coverParams.d;
|
||||
fastCoverParams->steps = coverParams.steps;
|
||||
fastCoverParams->nbThreads = coverParams.nbThreads;
|
||||
fastCoverParams->splitPoint = coverParams.splitPoint;
|
||||
fastCoverParams->f = f;
|
||||
fastCoverParams->accel = accel;
|
||||
fastCoverParams->zParams = coverParams.zParams;
|
||||
}
|
||||
|
||||
|
||||
ZDICTLIB_API size_t ZDICT_trainFromBuffer_fastCover(
|
||||
void *dictBuffer, size_t dictBufferCapacity, const void *samplesBuffer,
|
||||
const size_t *samplesSizes, unsigned nbSamples, ZDICT_fastCover_params_t parameters) {
|
||||
BYTE* const dict = (BYTE*)dictBuffer;
|
||||
FASTCOVER_ctx_t ctx;
|
||||
ZDICT_cover_params_t coverParams;
|
||||
FASTCOVER_accel_t accelParams;
|
||||
/* Initialize global data */
|
||||
g_displayLevel = parameters.zParams.notificationLevel;
|
||||
/* Assign splitPoint and f if not provided */
|
||||
parameters.splitPoint = parameters.splitPoint <= 0.0 ? DEFAULT_SPLITPOINT : parameters.splitPoint;
|
||||
parameters.f = parameters.f == 0 ? DEFAULT_F : parameters.f;
|
||||
parameters.accel = parameters.accel == 0 ? DEFAULT_ACCEL : parameters.accel;
|
||||
/* Convert to cover parameter */
|
||||
memset(&coverParams, 0 , sizeof(coverParams));
|
||||
FASTCOVER_convertToCoverParams(parameters, &coverParams);
|
||||
/* Checks */
|
||||
if (!FASTCOVER_checkParameters(coverParams, dictBufferCapacity, parameters.f,
|
||||
parameters.accel)) {
|
||||
DISPLAYLEVEL(1, "FASTCOVER parameters incorrect\n");
|
||||
return ERROR(GENERIC);
|
||||
}
|
||||
if (nbSamples == 0) {
|
||||
DISPLAYLEVEL(1, "FASTCOVER must have at least one input file\n");
|
||||
return ERROR(GENERIC);
|
||||
}
|
||||
if (dictBufferCapacity < ZDICT_DICTSIZE_MIN) {
|
||||
DISPLAYLEVEL(1, "dictBufferCapacity must be at least %u\n",
|
||||
ZDICT_DICTSIZE_MIN);
|
||||
return ERROR(dstSize_tooSmall);
|
||||
}
|
||||
/* Assign corresponding FASTCOVER_accel_t to accelParams*/
|
||||
accelParams = FASTCOVER_defaultAccelParameters[parameters.accel];
|
||||
/* Initialize context */
|
||||
if (!FASTCOVER_ctx_init(&ctx, samplesBuffer, samplesSizes, nbSamples,
|
||||
coverParams.d, parameters.splitPoint, parameters.f,
|
||||
accelParams)) {
|
||||
DISPLAYLEVEL(1, "Failed to initialize context\n");
|
||||
return ERROR(GENERIC);
|
||||
}
|
||||
/* Build the dictionary */
|
||||
DISPLAYLEVEL(2, "Building dictionary\n");
|
||||
{
|
||||
/* Initialize array to keep track of frequency of dmer within activeSegment */
|
||||
U16* segmentFreqs = (U16 *)calloc(((U64)1 << parameters.f), sizeof(U16));
|
||||
const size_t tail = FASTCOVER_buildDictionary(&ctx, ctx.freqs, dictBuffer,
|
||||
dictBufferCapacity, coverParams, segmentFreqs);
|
||||
const unsigned nbFinalizeSamples = (unsigned)(ctx.nbTrainSamples * ctx.accelParams.finalize / 100);
|
||||
const size_t dictionarySize = ZDICT_finalizeDictionary(
|
||||
dict, dictBufferCapacity, dict + tail, dictBufferCapacity - tail,
|
||||
samplesBuffer, samplesSizes, nbFinalizeSamples, coverParams.zParams);
|
||||
if (!ZSTD_isError(dictionarySize)) {
|
||||
DISPLAYLEVEL(2, "Constructed dictionary of size %u\n",
|
||||
(U32)dictionarySize);
|
||||
}
|
||||
FASTCOVER_ctx_destroy(&ctx);
|
||||
free(segmentFreqs);
|
||||
return dictionarySize;
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
ZDICTLIB_API size_t ZDICT_optimizeTrainFromBuffer_fastCover(
|
||||
void *dictBuffer, size_t dictBufferCapacity, const void *samplesBuffer,
|
||||
const size_t *samplesSizes, unsigned nbSamples,
|
||||
ZDICT_fastCover_params_t *parameters) {
|
||||
ZDICT_cover_params_t coverParams;
|
||||
FASTCOVER_accel_t accelParams;
|
||||
/* constants */
|
||||
const unsigned nbThreads = parameters->nbThreads;
|
||||
const double splitPoint =
|
||||
parameters->splitPoint <= 0.0 ? DEFAULT_SPLITPOINT : parameters->splitPoint;
|
||||
const unsigned kMinD = parameters->d == 0 ? 6 : parameters->d;
|
||||
const unsigned kMaxD = parameters->d == 0 ? 8 : parameters->d;
|
||||
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);
|
||||
const unsigned f = parameters->f == 0 ? DEFAULT_F : parameters->f;
|
||||
const unsigned accel = parameters->accel == 0 ? DEFAULT_ACCEL : parameters->accel;
|
||||
/* Local variables */
|
||||
const int displayLevel = parameters->zParams.notificationLevel;
|
||||
unsigned iteration = 1;
|
||||
unsigned d;
|
||||
unsigned k;
|
||||
COVER_best_t best;
|
||||
POOL_ctx *pool = NULL;
|
||||
/* Checks */
|
||||
if (splitPoint <= 0 || splitPoint > 1) {
|
||||
LOCALDISPLAYLEVEL(displayLevel, 1, "Incorrect splitPoint\n");
|
||||
return ERROR(GENERIC);
|
||||
}
|
||||
if (accel == 0 || accel > FASTCOVER_MAX_ACCEL) {
|
||||
LOCALDISPLAYLEVEL(displayLevel, 1, "Incorrect accel\n");
|
||||
return ERROR(GENERIC);
|
||||
}
|
||||
if (kMinK < kMaxD || kMaxK < kMinK) {
|
||||
LOCALDISPLAYLEVEL(displayLevel, 1, "Incorrect k\n");
|
||||
return ERROR(GENERIC);
|
||||
}
|
||||
if (nbSamples == 0) {
|
||||
LOCALDISPLAYLEVEL(displayLevel, 1, "FASTCOVER must have at least one input file\n");
|
||||
return ERROR(GENERIC);
|
||||
}
|
||||
if (dictBufferCapacity < ZDICT_DICTSIZE_MIN) {
|
||||
LOCALDISPLAYLEVEL(displayLevel, 1, "dictBufferCapacity must be at least %u\n",
|
||||
ZDICT_DICTSIZE_MIN);
|
||||
return ERROR(dstSize_tooSmall);
|
||||
}
|
||||
if (nbThreads > 1) {
|
||||
pool = POOL_create(nbThreads, 1);
|
||||
if (!pool) {
|
||||
return ERROR(memory_allocation);
|
||||
}
|
||||
}
|
||||
/* Initialization */
|
||||
COVER_best_init(&best);
|
||||
memset(&coverParams, 0 , sizeof(coverParams));
|
||||
FASTCOVER_convertToCoverParams(*parameters, &coverParams);
|
||||
accelParams = FASTCOVER_defaultAccelParameters[accel];
|
||||
/* Turn down global display level to clean up display at level 2 and below */
|
||||
g_displayLevel = displayLevel == 0 ? 0 : displayLevel - 1;
|
||||
/* Loop through d first because each new value needs a new context */
|
||||
LOCALDISPLAYLEVEL(displayLevel, 2, "Trying %u different sets of parameters\n",
|
||||
kIterations);
|
||||
for (d = kMinD; d <= kMaxD; d += 2) {
|
||||
/* Initialize the context for this value of d */
|
||||
FASTCOVER_ctx_t ctx;
|
||||
LOCALDISPLAYLEVEL(displayLevel, 3, "d=%u\n", d);
|
||||
if (!FASTCOVER_ctx_init(&ctx, samplesBuffer, samplesSizes, nbSamples, d, splitPoint, f, accelParams)) {
|
||||
LOCALDISPLAYLEVEL(displayLevel, 1, "Failed to initialize context\n");
|
||||
COVER_best_destroy(&best);
|
||||
POOL_free(pool);
|
||||
return ERROR(GENERIC);
|
||||
}
|
||||
/* Loop through k reusing the same context */
|
||||
for (k = kMinK; k <= kMaxK; k += kStepSize) {
|
||||
/* Prepare the arguments */
|
||||
FASTCOVER_tryParameters_data_t *data = (FASTCOVER_tryParameters_data_t *)malloc(
|
||||
sizeof(FASTCOVER_tryParameters_data_t));
|
||||
LOCALDISPLAYLEVEL(displayLevel, 3, "k=%u\n", k);
|
||||
if (!data) {
|
||||
LOCALDISPLAYLEVEL(displayLevel, 1, "Failed to allocate parameters\n");
|
||||
COVER_best_destroy(&best);
|
||||
FASTCOVER_ctx_destroy(&ctx);
|
||||
POOL_free(pool);
|
||||
return ERROR(GENERIC);
|
||||
}
|
||||
data->ctx = &ctx;
|
||||
data->best = &best;
|
||||
data->dictBufferCapacity = dictBufferCapacity;
|
||||
data->parameters = coverParams;
|
||||
data->parameters.k = k;
|
||||
data->parameters.d = d;
|
||||
data->parameters.splitPoint = splitPoint;
|
||||
data->parameters.steps = kSteps;
|
||||
data->parameters.zParams.notificationLevel = g_displayLevel;
|
||||
/* Check the parameters */
|
||||
if (!FASTCOVER_checkParameters(data->parameters, dictBufferCapacity,
|
||||
data->ctx->f, accel)) {
|
||||
DISPLAYLEVEL(1, "FASTCOVER parameters incorrect\n");
|
||||
free(data);
|
||||
continue;
|
||||
}
|
||||
/* Call the function and pass ownership of data to it */
|
||||
COVER_best_start(&best);
|
||||
if (pool) {
|
||||
POOL_add(pool, &FASTCOVER_tryParameters, data);
|
||||
} else {
|
||||
FASTCOVER_tryParameters(data);
|
||||
}
|
||||
/* Print status */
|
||||
LOCALDISPLAYUPDATE(displayLevel, 2, "\r%u%% ",
|
||||
(U32)((iteration * 100) / kIterations));
|
||||
++iteration;
|
||||
}
|
||||
COVER_best_wait(&best);
|
||||
FASTCOVER_ctx_destroy(&ctx);
|
||||
}
|
||||
LOCALDISPLAYLEVEL(displayLevel, 2, "\r%79s\r", "");
|
||||
/* Fill the output buffer and parameters with output of the best parameters */
|
||||
{
|
||||
const size_t dictSize = best.dictSize;
|
||||
if (ZSTD_isError(best.compressedSize)) {
|
||||
const size_t compressedSize = best.compressedSize;
|
||||
COVER_best_destroy(&best);
|
||||
POOL_free(pool);
|
||||
return compressedSize;
|
||||
}
|
||||
FASTCOVER_convertToFastCoverParams(best.parameters, parameters, f, accel);
|
||||
memcpy(dictBuffer, best.dict, dictSize);
|
||||
COVER_best_destroy(&best);
|
||||
POOL_free(pool);
|
||||
return dictSize;
|
||||
}
|
||||
|
||||
}
|
||||
@@ -863,8 +863,8 @@ _cleanup:
|
||||
|
||||
size_t ZDICT_finalizeDictionary(void* dictBuffer, size_t dictBufferCapacity,
|
||||
const void* customDictContent, size_t dictContentSize,
|
||||
const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples,
|
||||
ZDICT_params_t params)
|
||||
const void* samplesBuffer, const size_t* samplesSizes,
|
||||
unsigned nbSamples, ZDICT_params_t params)
|
||||
{
|
||||
size_t hSize;
|
||||
#define HBUFFSIZE 256 /* should prove large enough for all entropy headers */
|
||||
@@ -1078,17 +1078,17 @@ size_t ZDICT_trainFromBuffer_legacy(void* dictBuffer, size_t dictBufferCapacity,
|
||||
size_t ZDICT_trainFromBuffer(void* dictBuffer, size_t dictBufferCapacity,
|
||||
const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples)
|
||||
{
|
||||
ZDICT_cover_params_t params;
|
||||
ZDICT_fastCover_params_t params;
|
||||
DEBUGLOG(3, "ZDICT_trainFromBuffer");
|
||||
memset(¶ms, 0, sizeof(params));
|
||||
params.d = 8;
|
||||
params.steps = 4;
|
||||
/* Default to level 6 since no compression level information is available */
|
||||
params.zParams.compressionLevel = 6;
|
||||
params.zParams.compressionLevel = 3;
|
||||
#if defined(DEBUGLEVEL) && (DEBUGLEVEL>=1)
|
||||
params.zParams.notificationLevel = DEBUGLEVEL;
|
||||
#endif
|
||||
return ZDICT_optimizeTrainFromBuffer_cover(dictBuffer, dictBufferCapacity,
|
||||
return ZDICT_optimizeTrainFromBuffer_fastCover(dictBuffer, dictBufferCapacity,
|
||||
samplesBuffer, samplesSizes, nbSamples,
|
||||
¶ms);
|
||||
}
|
||||
|
||||
+56
-3
@@ -39,7 +39,8 @@ extern "C" {
|
||||
|
||||
/*! ZDICT_trainFromBuffer():
|
||||
* Train a dictionary from an array of samples.
|
||||
* Redirect towards ZDICT_optimizeTrainFromBuffer_cover() single-threaded, with d=8 and steps=4.
|
||||
* Redirect towards ZDICT_optimizeTrainFromBuffer_fastCover() single-threaded, with d=8, steps=4,
|
||||
* f=18, and accel=1.
|
||||
* Samples must be stored concatenated in a single flat buffer `samplesBuffer`,
|
||||
* supplied with an array of sizes `samplesSizes`, providing the size of each sample, in order.
|
||||
* The resulting dictionary will be saved into `dictBuffer`.
|
||||
@@ -90,6 +91,16 @@ typedef struct {
|
||||
ZDICT_params_t zParams;
|
||||
} ZDICT_cover_params_t;
|
||||
|
||||
typedef struct {
|
||||
unsigned k; /* Segment size : constraint: 0 < k : Reasonable range [16, 2048+] */
|
||||
unsigned d; /* dmer size : constraint: 0 < d <= k : Reasonable range [6, 16] */
|
||||
unsigned f; /* log of size of frequency array : constraint: 0 < f <= 31 : 1 means default(18)*/
|
||||
unsigned steps; /* Number of steps : Only used for optimization : 0 means default (32) : Higher means more parameters checked */
|
||||
unsigned nbThreads; /* Number of threads : constraint: 0 < nbThreads : 1 means single-threaded : Only used for optimization : Ignored if ZSTD_MULTITHREAD is not defined */
|
||||
double splitPoint; /* Percentage of samples used for training: the first nbSamples * splitPoint samples will be used to training, the last nbSamples * (1 - splitPoint) samples will be used for testing, 0 means default (0.75), 1.0 when all samples are used for both training and testing */
|
||||
unsigned accel; /* Acceleration level: constraint: 0 < accel <= 10, higher means faster and less accurate, 0 means default(1) */
|
||||
ZDICT_params_t zParams;
|
||||
} ZDICT_fastCover_params_t;
|
||||
|
||||
/*! ZDICT_trainFromBuffer_cover():
|
||||
* Train a dictionary from an array of samples using the COVER algorithm.
|
||||
@@ -116,9 +127,9 @@ ZDICTLIB_API size_t ZDICT_trainFromBuffer_cover(
|
||||
* dictionary constructed with those parameters is stored in `dictBuffer`.
|
||||
*
|
||||
* All of the parameters d, k, steps are optional.
|
||||
* If d is non-zero then we don't check multiple values of d, otherwise we check d = {6, 8, 10, 12, 14, 16}.
|
||||
* If d is non-zero then we don't check multiple values of d, otherwise we check d = {6, 8}.
|
||||
* if steps is zero it defaults to its default value.
|
||||
* If k is non-zero then we don't check multiple values of k, otherwise we check steps values in [16, 2048].
|
||||
* If k is non-zero then we don't check multiple values of k, otherwise we check steps values in [50, 2000].
|
||||
*
|
||||
* @return: size of dictionary stored into `dictBuffer` (<= `dictBufferCapacity`)
|
||||
* or an error code, which can be tested with ZDICT_isError().
|
||||
@@ -130,6 +141,48 @@ ZDICTLIB_API size_t ZDICT_optimizeTrainFromBuffer_cover(
|
||||
const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples,
|
||||
ZDICT_cover_params_t* parameters);
|
||||
|
||||
/*! ZDICT_trainFromBuffer_fastCover():
|
||||
* Train a dictionary from an array of samples using a modified version of COVER algorithm.
|
||||
* Samples must be stored concatenated in a single flat buffer `samplesBuffer`,
|
||||
* supplied with an array of sizes `samplesSizes`, providing the size of each sample, in order.
|
||||
* d and k are required.
|
||||
* All other parameters are optional, will use default values if not provided
|
||||
* The resulting dictionary will be saved into `dictBuffer`.
|
||||
* @return: size of dictionary stored into `dictBuffer` (<= `dictBufferCapacity`)
|
||||
* or an error code, which can be tested with ZDICT_isError().
|
||||
* Note: ZDICT_trainFromBuffer_fastCover() requires about 1 bytes of memory for each input byte and additionally another 6 * 2^f bytes of memory .
|
||||
* Tips: In general, a reasonable dictionary has a size of ~ 100 KB.
|
||||
* It's possible to select smaller or larger size, just by specifying `dictBufferCapacity`.
|
||||
* In general, it's recommended to provide a few thousands samples, though this can vary a lot.
|
||||
* It's recommended that total size of all samples be about ~x100 times the target size of dictionary.
|
||||
*/
|
||||
ZDICTLIB_API size_t ZDICT_trainFromBuffer_fastCover(void *dictBuffer,
|
||||
size_t dictBufferCapacity, const void *samplesBuffer,
|
||||
const size_t *samplesSizes, unsigned nbSamples,
|
||||
ZDICT_fastCover_params_t parameters);
|
||||
|
||||
/*! ZDICT_optimizeTrainFromBuffer_fastCover():
|
||||
* The same requirements as above hold for all the parameters except `parameters`.
|
||||
* This function tries many parameter combinations (specifically, k and d combinations)
|
||||
* and picks the best parameters. `*parameters` is filled with the best parameters found,
|
||||
* dictionary constructed with those parameters is stored in `dictBuffer`.
|
||||
* All of the parameters d, k, steps, f, and accel are optional.
|
||||
* If d is non-zero then we don't check multiple values of d, otherwise we check d = {6, 8}.
|
||||
* if steps is zero it defaults to its default value.
|
||||
* If k is non-zero then we don't check multiple values of k, otherwise we check steps values in [50, 2000].
|
||||
* If f is zero, default value of 18 is used.
|
||||
* If accel is zero, default value of 1 is used.
|
||||
*
|
||||
* @return: size of dictionary stored into `dictBuffer` (<= `dictBufferCapacity`)
|
||||
* or an error code, which can be tested with ZDICT_isError().
|
||||
* On success `*parameters` contains the parameters selected.
|
||||
* Note: ZDICT_optimizeTrainFromBuffer_fastCover() requires about 1 byte of memory for each input byte and additionally another 6 * 2^f bytes of memory for each thread.
|
||||
*/
|
||||
ZDICTLIB_API size_t ZDICT_optimizeTrainFromBuffer_fastCover(void* dictBuffer,
|
||||
size_t dictBufferCapacity, const void* samplesBuffer,
|
||||
const size_t* samplesSizes, unsigned nbSamples,
|
||||
ZDICT_fastCover_params_t* parameters);
|
||||
|
||||
/*! ZDICT_finalizeDictionary():
|
||||
* Given a custom content as a basis for dictionary, and a set of samples,
|
||||
* finalize dictionary by adding headers and statistics.
|
||||
|
||||
Reference in New Issue
Block a user