Limit train samples (#2809)
* Limit training samples size to 2GB
* simplified DISPLAYLEVEL() macro to use global vqriable instead of local.
* refactored training samples loading
* fixed compiler warning
* addressed comments from the pull request
* addressed @terrelln comments
* missed some fixes
* fixed type mismatch
* Fixed bug passing estimated number of samples rather insted of the loaded number of samples.
Changed unit conversion not to use bit-shifts.
* fixed a declaration after code
* fixed type conversion compile errors
* fixed more type castting
* fixed more type mismatching
* changed sizes type to size_t
* move type casting
* more type cast fixes
diff --git a/programs/dibio.c b/programs/dibio.c
index d6c9f6d..49fa211 100644
--- a/programs/dibio.c
+++ b/programs/dibio.c
@@ -49,6 +49,7 @@
static const size_t g_maxMemory = (sizeof(size_t) == 4) ? (2 GB - 64 MB) : ((size_t)(512 MB) << sizeof(size_t));
#define NOISELENGTH 32
+#define MAX_SAMPLES_SIZE (2 GB) /* training dataset limited to 2GB */
/*-*************************************
@@ -88,6 +89,15 @@
#undef MIN
#define MIN(a,b) ((a) < (b) ? (a) : (b))
+/**
+ Returns the size of a file.
+ If error returns -1.
+*/
+static S64 DiB_getFileSize (const char * fileName)
+{
+ U64 const fileSize = UTIL_getFileSize(fileName);
+ return (fileSize == UTIL_FILESIZE_UNKNOWN) ? -1 : (S64)fileSize;
+}
/* ********************************************************
* File related operations
@@ -101,47 +111,67 @@
* *bufferSizePtr is modified, it provides the amount data loaded within buffer.
* sampleSizes is filled with the size of each sample.
*/
-static unsigned DiB_loadFiles(void* buffer, size_t* bufferSizePtr,
- size_t* sampleSizes, unsigned sstSize,
- const char** fileNamesTable, unsigned nbFiles, size_t targetChunkSize,
- unsigned displayLevel)
+static int DiB_loadFiles(
+ void* buffer, size_t* bufferSizePtr,
+ size_t* sampleSizes, int sstSize,
+ const char** fileNamesTable, int nbFiles,
+ size_t targetChunkSize, int displayLevel )
{
char* const buff = (char*)buffer;
- size_t pos = 0;
- unsigned nbLoadedChunks = 0, fileIndex;
+ size_t totalDataLoaded = 0;
+ int nbSamplesLoaded = 0;
+ int fileIndex = 0;
+ FILE * f = NULL;
- for (fileIndex=0; fileIndex<nbFiles; fileIndex++) {
- const char* const fileName = fileNamesTable[fileIndex];
- unsigned long long const fs64 = UTIL_getFileSize(fileName);
- unsigned long long remainingToLoad = (fs64 == UTIL_FILESIZE_UNKNOWN) ? 0 : fs64;
- U32 const nbChunks = targetChunkSize ? (U32)((fs64 + (targetChunkSize-1)) / targetChunkSize) : 1;
- U64 const chunkSize = targetChunkSize ? MIN(targetChunkSize, fs64) : fs64;
- size_t const maxChunkSize = (size_t)MIN(chunkSize, SAMPLESIZE_MAX);
- U32 cnb;
- FILE* const f = fopen(fileName, "rb");
- if (f==NULL) EXM_THROW(10, "zstd: dictBuilder: %s %s ", fileName, strerror(errno));
- DISPLAYUPDATE(2, "Loading %s... \r", fileName);
- for (cnb=0; cnb<nbChunks; cnb++) {
- size_t const toLoad = (size_t)MIN(maxChunkSize, remainingToLoad);
- if (toLoad > *bufferSizePtr-pos) break;
- { size_t const readSize = fread(buff+pos, 1, toLoad, f);
- if (readSize != toLoad) EXM_THROW(11, "Pb reading %s", fileName);
- pos += readSize;
- sampleSizes[nbLoadedChunks++] = toLoad;
- remainingToLoad -= targetChunkSize;
- if (nbLoadedChunks == sstSize) { /* no more space left in sampleSizes table */
- fileIndex = nbFiles; /* stop there */
+ assert(targetChunkSize <= SAMPLESIZE_MAX);
+
+ while ( nbSamplesLoaded < sstSize && fileIndex < nbFiles ) {
+ size_t fileDataLoaded;
+ S64 const fileSize = DiB_getFileSize(fileNamesTable[fileIndex]);
+ if (fileSize <= 0) /* skip if zero-size or file error */
+ continue;
+
+ f = fopen( fileNamesTable[fileIndex], "rb");
+ if (f == NULL)
+ EXM_THROW(10, "zstd: dictBuilder: %s %s ", fileNamesTable[fileIndex], strerror(errno));
+ DISPLAYUPDATE(2, "Loading %s... \r", fileNamesTable[fileIndex]);
+
+ /* Load the first chunk of data from the file */
+ fileDataLoaded = targetChunkSize > 0 ?
+ (size_t)MIN(fileSize, (S64)targetChunkSize) :
+ (size_t)MIN(fileSize, SAMPLESIZE_MAX );
+ if (totalDataLoaded + fileDataLoaded > *bufferSizePtr)
+ break;
+ if (fread( buff+totalDataLoaded, 1, fileDataLoaded, f ) != fileDataLoaded)
+ EXM_THROW(11, "Pb reading %s", fileNamesTable[fileIndex]);
+ sampleSizes[nbSamplesLoaded++] = fileDataLoaded;
+ totalDataLoaded += fileDataLoaded;
+
+ /* If file-chunking is enabled, load the rest of the file as more samples */
+ if (targetChunkSize > 0) {
+ while( (S64)fileDataLoaded < fileSize && nbSamplesLoaded < sstSize ) {
+ size_t const chunkSize = MIN((size_t)(fileSize-fileDataLoaded), targetChunkSize);
+ if (totalDataLoaded + chunkSize > *bufferSizePtr) /* buffer is full */
break;
- }
- if (toLoad < targetChunkSize) {
- fseek(f, (long)(targetChunkSize - toLoad), SEEK_CUR);
- } } }
- fclose(f);
+
+ if (fread( buff+totalDataLoaded, 1, chunkSize, f ) != chunkSize)
+ EXM_THROW(11, "Pb reading %s", fileNamesTable[fileIndex]);
+ sampleSizes[nbSamplesLoaded++] = chunkSize;
+ totalDataLoaded += chunkSize;
+ fileDataLoaded += chunkSize;
+ }
+ }
+ fileIndex += 1;
+ fclose(f); f = NULL;
}
+ if (f != NULL)
+ fclose(f);
+
DISPLAYLEVEL(2, "\r%79s\r", "");
- *bufferSizePtr = pos;
- DISPLAYLEVEL(4, "loaded : %u KB \n", (unsigned)(pos >> 10))
- return nbLoadedChunks;
+ DISPLAYLEVEL(4, "Loaded %d KB total training data, %d nb samples \n",
+ (int)(totalDataLoaded / (1 KB)), nbSamplesLoaded );
+ *bufferSizePtr = totalDataLoaded;
+ return nbSamplesLoaded;
}
#define DiB_rotl32(x,r) ((x << r) | (x >> (32 - r)))
@@ -223,11 +253,10 @@
if (n!=0) EXM_THROW(5, "%s : flush error", dictFileName) }
}
-
typedef struct {
- U64 totalSizeToLoad;
- unsigned oneSampleTooLarge;
- unsigned nbSamples;
+ S64 totalSizeToLoad;
+ int nbSamples;
+ int oneSampleTooLarge;
} fileStats;
/*! DiB_fileStats() :
@@ -235,46 +264,87 @@
* provides the amount of data to be loaded and the resulting nb of samples.
* This is useful primarily for allocation purpose => sample buffer, and sample sizes table.
*/
-static fileStats DiB_fileStats(const char** fileNamesTable, unsigned nbFiles, size_t chunkSize, unsigned displayLevel)
+static fileStats DiB_fileStats(const char** fileNamesTable, int nbFiles, size_t chunkSize, int displayLevel)
{
fileStats fs;
- unsigned n;
+ int n;
memset(&fs, 0, sizeof(fs));
+
+ // We assume that if chunking is requsted, the chunk size is < SAMPLESIZE_MAX
+ assert( chunkSize <= SAMPLESIZE_MAX );
+
for (n=0; n<nbFiles; n++) {
- U64 const fileSize = UTIL_getFileSize(fileNamesTable[n]);
- U64 const srcSize = (fileSize == UTIL_FILESIZE_UNKNOWN) ? 0 : fileSize;
- U32 const nbSamples = (U32)(chunkSize ? (srcSize + (chunkSize-1)) / chunkSize : 1);
- U64 const chunkToLoad = chunkSize ? MIN(chunkSize, srcSize) : srcSize;
- size_t const cappedChunkSize = (size_t)MIN(chunkToLoad, SAMPLESIZE_MAX);
- fs.totalSizeToLoad += cappedChunkSize * nbSamples;
- fs.oneSampleTooLarge |= (chunkSize > 2*SAMPLESIZE_MAX);
- fs.nbSamples += nbSamples;
+ S64 const fileSize = DiB_getFileSize(fileNamesTable[n]);
+ // TODO: is there a minimum sample size? What if the file is 1-byte?
+ if (fileSize == 0) {
+ DISPLAYLEVEL(3, "Sample file '%s' has zero size, skipping...\n", fileNamesTable[n]);
+ continue;
+ }
+
+ /* the case where we are breaking up files in sample chunks */
+ if (chunkSize > 0)
+ {
+ // TODO: is there a minimum sample size? Can we have a 1-byte sample?
+ fs.nbSamples += (int)((fileSize + chunkSize-1) / chunkSize);
+ fs.totalSizeToLoad += fileSize;
+ }
+ else {
+ /* the case where one file is one sample */
+ if (fileSize > SAMPLESIZE_MAX) {
+ /* flag excessively large sample files */
+ fs.oneSampleTooLarge |= (fileSize > 2*SAMPLESIZE_MAX);
+
+ /* Limit to the first SAMPLESIZE_MAX (128kB) of the file */
+ DISPLAYLEVEL(3, "Sample file '%s' is too large, limiting to %d KB",
+ fileNamesTable[n], SAMPLESIZE_MAX / (1 KB));
+ }
+ fs.nbSamples += 1;
+ fs.totalSizeToLoad += MIN(fileSize, SAMPLESIZE_MAX);
+ }
}
- DISPLAYLEVEL(4, "Preparing to load : %u KB \n", (unsigned)(fs.totalSizeToLoad >> 10));
+ DISPLAYLEVEL(4, "Found training data %d files, %d KB, %d samples\n", nbFiles, (int)(fs.totalSizeToLoad / (1 KB)), fs.nbSamples);
return fs;
}
-
-int DiB_trainFromFiles(const char* dictFileName, unsigned maxDictSize,
- const char** fileNamesTable, unsigned nbFiles, size_t chunkSize,
+int DiB_trainFromFiles(const char* dictFileName, size_t maxDictSize,
+ const char** fileNamesTable, int nbFiles, size_t chunkSize,
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 */
+ fileStats fs;
+ size_t* sampleSizes; /* vector of sample sizes. Each sample can be up to SAMPLESIZE_MAX */
+ int nbSamplesLoaded; /* nb of samples effectively loaded in srcBuffer */
+ size_t loadedSize; /* total data loaded in srcBuffer for all samples */
+ void* srcBuffer /* contiguous buffer with training data/samples */;
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 :
- 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);
int result = 0;
+ int const displayLevel = params ? params->zParams.notificationLevel :
+ coverParams ? coverParams->zParams.notificationLevel :
+ fastCoverParams ? fastCoverParams->zParams.notificationLevel : 0;
+
+ /* Shuffle input files before we start assessing how much sample datA to load.
+ The purpose of the shuffle is to pick random samples when the sample
+ set is larger than what we can load in memory. */
+ DISPLAYLEVEL(3, "Shuffling input files\n");
+ DiB_shuffle(fileNamesTable, nbFiles);
+
+ /* Figure out how much sample data to load with how many samples */
+ fs = DiB_fileStats(fileNamesTable, nbFiles, chunkSize, displayLevel);
+
+ {
+ int const memMult = params ? MEMMULT :
+ coverParams ? COVER_MEMMULT:
+ FASTCOVER_MEMMULT;
+ size_t const maxMem = DiB_findMaxMem(fs.totalSizeToLoad * memMult) / memMult;
+ /* Limit the size of the training data to the free memory */
+ /* Limit the size of the training data to 2GB */
+ /* TODO: there is oportunity to stop DiB_fileStats() early when the data limit is reached */
+ loadedSize = (size_t)MIN( MIN((S64)maxMem, fs.totalSizeToLoad), MAX_SAMPLES_SIZE );
+ srcBuffer = malloc(loadedSize+NOISELENGTH);
+ sampleSizes = (size_t*)malloc(fs.nbSamples * sizeof(size_t));
+ }
+
/* Checks */
if ((!sampleSizes) || (!srcBuffer) || (!dictBuffer))
EXM_THROW(12, "not enough memory for DiB_trainFiles"); /* should not happen */
@@ -289,31 +359,32 @@
DISPLAYLEVEL(2, "! Alternatively, split files into fixed-size blocks representative of samples, with -B# \n");
EXM_THROW(14, "nb of samples too low"); /* we now clearly forbid this case */
}
- if (fs.totalSizeToLoad < (unsigned long long)maxDictSize * 8) {
+ if (fs.totalSizeToLoad < (S64)maxDictSize * 8) {
DISPLAYLEVEL(2, "! Warning : data size of samples too small for target dictionary size \n");
DISPLAYLEVEL(2, "! Samples should be about 100x larger than target dictionary size \n");
}
/* init */
- if (loadedSize < fs.totalSizeToLoad)
- DISPLAYLEVEL(1, "Not enough memory; training on %u MB only...\n", (unsigned)(loadedSize >> 20));
+ if ((S64)loadedSize < fs.totalSizeToLoad)
+ DISPLAYLEVEL(1, "Training samples set too large (%u MB); training on %u MB only...\n",
+ (unsigned)(fs.totalSizeToLoad / (1 MB)),
+ (unsigned)(loadedSize / (1 MB)));
/* Load input buffer */
- DISPLAYLEVEL(3, "Shuffling input files\n");
- DiB_shuffle(fileNamesTable, nbFiles);
-
- DiB_loadFiles(srcBuffer, &loadedSize, sampleSizes, fs.nbSamples, fileNamesTable, nbFiles, chunkSize, displayLevel);
+ nbSamplesLoaded = DiB_loadFiles(
+ srcBuffer, &loadedSize, sampleSizes, fs.nbSamples, fileNamesTable,
+ nbFiles, chunkSize, displayLevel);
{ size_t dictSize;
if (params) {
DiB_fillNoise((char*)srcBuffer + loadedSize, NOISELENGTH); /* guard band, for end of buffer condition */
dictSize = ZDICT_trainFromBuffer_legacy(dictBuffer, maxDictSize,
- srcBuffer, sampleSizes, fs.nbSamples,
+ srcBuffer, sampleSizes, nbSamplesLoaded,
*params);
} else if (coverParams) {
if (optimize) {
dictSize = ZDICT_optimizeTrainFromBuffer_cover(dictBuffer, maxDictSize,
- srcBuffer, sampleSizes, fs.nbSamples,
+ srcBuffer, sampleSizes, nbSamplesLoaded,
coverParams);
if (!ZDICT_isError(dictSize)) {
unsigned splitPercentage = (unsigned)(coverParams->splitPoint * 100);
@@ -322,13 +393,13 @@
}
} else {
dictSize = ZDICT_trainFromBuffer_cover(dictBuffer, maxDictSize, srcBuffer,
- sampleSizes, fs.nbSamples, *coverParams);
+ sampleSizes, nbSamplesLoaded, *coverParams);
}
} else {
assert(fastCoverParams != NULL);
if (optimize) {
dictSize = ZDICT_optimizeTrainFromBuffer_fastCover(dictBuffer, maxDictSize,
- srcBuffer, sampleSizes, fs.nbSamples,
+ srcBuffer, sampleSizes, nbSamplesLoaded,
fastCoverParams);
if (!ZDICT_isError(dictSize)) {
unsigned splitPercentage = (unsigned)(fastCoverParams->splitPoint * 100);
@@ -338,7 +409,7 @@
}
} else {
dictSize = ZDICT_trainFromBuffer_fastCover(dictBuffer, maxDictSize, srcBuffer,
- sampleSizes, fs.nbSamples, *fastCoverParams);
+ sampleSizes, nbSamplesLoaded, *fastCoverParams);
}
}
if (ZDICT_isError(dictSize)) {