/*
* Copyright (c) Meta Platforms, Inc. and affiliates.
* All rights reserved.
*
* This source code is licensed under both the BSD-style license (found in the
* LICENSE file in the root directory of this source tree) and the GPLv2 (found
* in the COPYING file in the root directory of this source tree).
* You may select, at your option, one of the above-listed licenses.
*/
/*-*************************************
* Constants
***************************************/
/**
* There are 32bit indexes used to ref samples, so limit samples size to 4GB
* on 64bit builds.
* For 32bit builds we choose 1 GB.
* Most 32bit platforms have 2GB user-mode addressable space and we allocate a large
* contiguous buffer, so 1GB is already a high limit.
*/
#define FASTCOVER_MAX_SAMPLES_SIZE (sizeof(size_t) == 8 ? ((unsigned)-1) : ((unsigned)1 GB))
#define FASTCOVER_MAX_F 31
#define FASTCOVER_MAX_ACCEL 10
#define FASTCOVER_DEFAULT_SPLITPOINT 0.75
#define DEFAULT_F 20
#define DEFAULT_ACCEL 1
/*-*************************************
* Hash Functions
***************************************/
/**
* Hash the d-byte value pointed to by p and mod 2^f into the frequency vector
*/
static size_t FASTCOVER_hashPtrToIndex(const void* p, U32 f, unsigned d) {
if (d == 6) {
return ZSTD_hash6Ptr(p, f);
}
return ZSTD_hash8Ptr(p, f);
}
/*-*************************************
* 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;
/*-*************************************
* 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 dictionary 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 idx = FASTCOVER_hashPtrToIndex(ctx->samples + activeSegment.end, f, d);
/* Add frequency of this index to score if this is the first occurrence of index in active segment */
if (segmentFreqs[idx] == 0) {
activeSegment.score += freqs[idx];
}
/* Increment end of segment and segmentFreqs*/
activeSegment.end += 1;
segmentFreqs[idx] += 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, const 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 i;
assert(ctx->nbTrainSamples >= 5);
assert(ctx->nbTrainSamples <= ctx->nbSamples);
for (i = 0; i < ctx->nbTrainSamples; i++) {
size_t start = ctx->offsets[i]; /* start of current dmer */
size_t const currSampleEnd = ctx->offsets[i+1];
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 be used multiple
* times.
* Returns 0 on success or error code on error.
* The context must be destroyed with `FASTCOVER_ctx_destroy()`.
*/
static size_t
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",
(unsigned)(totalSamplesSize >> 20), (FASTCOVER_MAX_SAMPLES_SIZE >> 20));
return ERROR(srcSize_wrong);
}
/* 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 ERROR(srcSize_wrong);
}
/* Check if there's testing sample */
if (nbTestSamples < 1) {
DISPLAYLEVEL(1, "Total number of testing samples is %u and is invalid.\n", nbTestSamples);
return ERROR(srcSize_wrong);
}
/* Zero the context */
memset(ctx, 0, sizeof(*ctx));
DISPLAYLEVEL(2, "Training on %u samples of total size %u\n", nbTrainSamples,
(unsigned)trainingSamplesSize);
DISPLAYLEVEL(2, "Testing on %u samples of total size %u\n", nbTestSamples,
(unsigned)testSamplesSize);
/**
* 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 into epochs. We will select one segment from each epoch. */
const COVER_epoch_info_t epochs = COVER_computeEpochs(
(U32)dictBufferCapacity, (U32)ctx->nbDmers, parameters.k, 1);
const size_t maxZeroScoreRun = 10;
size_t zeroScoreRun = 0;
size_t epoch;
DISPLAYLEVEL(2, "Breaking content into %u epochs of size %u\n",
(U32)epochs.num, (U32)epochs.size);
/* Loop through the epochs until there are no more segments or the dictionary
* is full.
*/
for (epoch = 0; tail > 0; epoch = (epoch + 1) % epochs.num) {
const U32 epochBegin = (U32)(epoch * epochs.size);
const U32 epochEnd = epochBegin + epochs.size;
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.
* There may be new content in other epochs, for continue for some time.
*/
if (segment.score == 0) {
if (++zeroScoreRun >= maxZeroScoreRun) {
break;
}
continue;
}
zeroScoreRun = 0;
/* 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%% ",
(unsigned)(((dictBufferCapacity - tail) * 100) / dictBufferCapacity));
}
DISPLAYLEVEL(2, "\r%79s\r", "");
return tail;
}
/**
* 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*)malloc(dictBufferCapacity);
COVER_dictSelection_t selection = COVER_dictSelectionError(ERROR(GENERIC));
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);