1 /*
2 bench.c - Demo program to benchmark open-source compression algorithms
3 Copyright (C) Yann Collet 2012-2020
4
5 GPL v2 License
6
7 This program is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 2 of the License, or
10 (at your option) any later version.
11
12 This program is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
16
17 You should have received a copy of the GNU General Public License along
18 with this program; if not, write to the Free Software Foundation, Inc.,
19 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
20
21 You can contact the author at :
22 - LZ4 homepage : http://www.lz4.org
23 - LZ4 source repository : https://github.com/lz4/lz4
24 */
25
26
27 /*-************************************
28 * Compiler options
29 **************************************/
30 #ifdef _MSC_VER /* Visual Studio */
31 # pragma warning(disable : 4127) /* disable: C4127: conditional expression is constant */
32 #endif
33
34
35 /* *************************************
36 * Includes
37 ***************************************/
38 #include "platform.h" /* Compiler options */
39 #include "util.h" /* UTIL_GetFileSize, UTIL_sleep */
40 #include <stdlib.h> /* malloc, free */
41 #include <string.h> /* memset */
42 #include <stdio.h> /* fprintf, fopen, ftello */
43 #include <time.h> /* clock_t, clock, CLOCKS_PER_SEC */
44 #include <assert.h> /* assert */
45
46 #include "lorem.h" /* LOREM_genBuffer */
47 #include "xxhash.h"
48 #include "bench.h"
49 #include "timefn.h"
50
51 #define LZ4_STATIC_LINKING_ONLY
52 #include "lz4.h"
53 #define LZ4_HC_STATIC_LINKING_ONLY
54 #include "lz4hc.h"
55 #include "lz4frame.h" /* LZ4F_decompress */
56
57
58 /* *************************************
59 * Constants
60 ***************************************/
61 #ifndef LZ4_GIT_COMMIT_STRING
62 # define LZ4_GIT_COMMIT_STRING ""
63 #else
64 # define LZ4_GIT_COMMIT_STRING LZ4_EXPAND_AND_QUOTE(LZ4_GIT_COMMIT)
65 #endif
66
67 #define NBSECONDS 3
68 #define TIMELOOP_MICROSEC 1*1000000ULL /* 1 second */
69 #define TIMELOOP_NANOSEC 1*1000000000ULL /* 1 second */
70 #define ACTIVEPERIOD_NANOSEC 70*1000000000ULL /* 70 seconds */
71 #define COOLPERIOD_SEC 10
72 #define DECOMP_MULT 1 /* test decompression DECOMP_MULT times longer than compression */
73
74 #define KB *(1 <<10)
75 #define MB *(1 <<20)
76 #define GB *(1U<<30)
77
78 #define LZ4_MAX_DICT_SIZE (64 KB)
79
80 static const size_t maxMemory = (sizeof(size_t)==4) ? (2 GB - 64 MB) : (size_t)(1ULL << ((sizeof(size_t)*8)-31));
81
82
83 /* *************************************
84 * console display
85 ***************************************/
86 #define DISPLAYOUT(...) fprintf(stdout, __VA_ARGS__)
87 #define OUTLEVEL(l, ...) if (g_displayLevel>=(l)) { DISPLAYOUT(__VA_ARGS__); }
88 #define DISPLAY(...) fprintf(stderr, __VA_ARGS__)
89 #define DISPLAYLEVEL(l, ...) if (g_displayLevel>=(l)) { DISPLAY(__VA_ARGS__); }
90 static U32 g_displayLevel = 2; /* 0 : no display; 1: errors; 2 : + result + interaction + warnings; 3 : + progression; 4 : + information */
91
92 #define DISPLAYUPDATE(l, ...) if (g_displayLevel>=l) { \
93 if ((clock() - g_time > refreshRate) || (g_displayLevel>=4)) \
94 { g_time = clock(); DISPLAY(__VA_ARGS__); \
95 if (g_displayLevel>=4) fflush(stdout); } }
96 static const clock_t refreshRate = CLOCKS_PER_SEC * 15 / 100;
97 static clock_t g_time = 0;
98
99
100 /* *************************************
101 * DEBUG and error conditions
102 ***************************************/
103 #ifndef DEBUG
104 # define DEBUG 0
105 #endif
106 #define DEBUGOUTPUT(...) if (DEBUG) DISPLAY(__VA_ARGS__);
107 #define END_PROCESS(error, ...) \
108 do { \
109 DEBUGOUTPUT("Error defined at %s, line %i : \n", __FILE__, __LINE__); \
110 DISPLAYLEVEL(1, "Error %i : ", error); \
111 DISPLAYLEVEL(1, __VA_ARGS__); \
112 DISPLAYLEVEL(1, "\n"); \
113 exit(error); \
114 } while (0)
115
116 #define LZ4_isError(errcode) (errcode==0)
117
118
119 /* *************************************
120 * Benchmark Parameters
121 ***************************************/
122 static U32 g_nbSeconds = NBSECONDS;
123 static size_t g_blockSize = 0;
124 int g_additionalParam = 0;
125 int g_benchSeparately = 0;
126 int g_decodeOnly = 0;
127 unsigned g_skipChecksums = 0;
128
BMK_setNotificationLevel(unsigned level)129 void BMK_setNotificationLevel(unsigned level) { g_displayLevel=level; }
130
BMK_setAdditionalParam(int additionalParam)131 void BMK_setAdditionalParam(int additionalParam) { g_additionalParam=additionalParam; }
132
BMK_setNbSeconds(unsigned nbSeconds)133 void BMK_setNbSeconds(unsigned nbSeconds)
134 {
135 g_nbSeconds = nbSeconds;
136 DISPLAYLEVEL(3, "- test >= %u seconds per compression / decompression -\n", g_nbSeconds);
137 }
138
BMK_setBlockSize(size_t blockSize)139 void BMK_setBlockSize(size_t blockSize) { g_blockSize = blockSize; }
140
BMK_setBenchSeparately(int separate)141 void BMK_setBenchSeparately(int separate) { g_benchSeparately = (separate!=0); }
142
BMK_setDecodeOnlyMode(int set)143 void BMK_setDecodeOnlyMode(int set) { g_decodeOnly = (set!=0); }
144
BMK_skipChecksums(int skip)145 void BMK_skipChecksums(int skip) { g_skipChecksums = (skip!=0); }
146
147
148 /* *************************************
149 * Compression state management
150 ***************************************/
151
152 struct compressionParameters
153 {
154 int cLevel;
155 const char* dictBuf;
156 int dictSize;
157
158 LZ4_stream_t* LZ4_stream;
159 LZ4_stream_t* LZ4_dictStream;
160 LZ4_streamHC_t* LZ4_streamHC;
161 LZ4_streamHC_t* LZ4_dictStreamHC;
162
163 void (*initFunction)(
164 struct compressionParameters* pThis);
165 void (*resetFunction)(
166 const struct compressionParameters* pThis);
167 int (*blockFunction)(
168 const struct compressionParameters* pThis,
169 const char* src, char* dst, int srcSize, int dstSize);
170 void (*cleanupFunction)(
171 const struct compressionParameters* pThis);
172 };
173
174 static void
LZ4_compressInitNoStream(struct compressionParameters * pThis)175 LZ4_compressInitNoStream(struct compressionParameters* pThis)
176 {
177 pThis->LZ4_stream = NULL;
178 pThis->LZ4_dictStream = NULL;
179 pThis->LZ4_streamHC = NULL;
180 pThis->LZ4_dictStreamHC = NULL;
181 }
182
183 static void
LZ4_compressInitStream(struct compressionParameters * pThis)184 LZ4_compressInitStream(struct compressionParameters* pThis)
185 {
186 pThis->LZ4_stream = LZ4_createStream();
187 pThis->LZ4_dictStream = LZ4_createStream();
188 pThis->LZ4_streamHC = NULL;
189 pThis->LZ4_dictStreamHC = NULL;
190 LZ4_loadDictSlow(pThis->LZ4_dictStream, pThis->dictBuf, pThis->dictSize);
191 }
192
193 static void
LZ4_compressInitStreamHC(struct compressionParameters * pThis)194 LZ4_compressInitStreamHC(struct compressionParameters* pThis)
195 {
196 pThis->LZ4_stream = NULL;
197 pThis->LZ4_dictStream = NULL;
198 pThis->LZ4_streamHC = LZ4_createStreamHC();
199 pThis->LZ4_dictStreamHC = LZ4_createStreamHC();
200 LZ4_resetStreamHC_fast(pThis->LZ4_dictStreamHC, pThis->cLevel);
201 LZ4_loadDictHC(pThis->LZ4_dictStreamHC, pThis->dictBuf, pThis->dictSize);
202 }
203
204 static void
LZ4_compressResetNoStream(const struct compressionParameters * cparams)205 LZ4_compressResetNoStream(const struct compressionParameters* cparams)
206 {
207 (void)cparams;
208 }
209
210 static void
LZ4_compressResetStream(const struct compressionParameters * cparams)211 LZ4_compressResetStream(const struct compressionParameters* cparams)
212 {
213 LZ4_resetStream_fast(cparams->LZ4_stream);
214 LZ4_attach_dictionary(cparams->LZ4_stream, cparams->LZ4_dictStream);
215 }
216
217 static void
LZ4_compressResetStreamHC(const struct compressionParameters * cparams)218 LZ4_compressResetStreamHC(const struct compressionParameters* cparams)
219 {
220 LZ4_resetStreamHC_fast(cparams->LZ4_streamHC, cparams->cLevel);
221 LZ4_attach_HC_dictionary(cparams->LZ4_streamHC, cparams->LZ4_dictStreamHC);
222 }
223
224 static int
LZ4_compressBlockNoStream(const struct compressionParameters * cparams,const char * src,char * dst,int srcSize,int dstSize)225 LZ4_compressBlockNoStream(const struct compressionParameters* cparams,
226 const char* src, char* dst,
227 int srcSize, int dstSize)
228 {
229 int const acceleration = (cparams->cLevel < 0) ? -cparams->cLevel + 1 : 1;
230 return LZ4_compress_fast(src, dst, srcSize, dstSize, acceleration);
231 }
232
233 static int
LZ4_compressBlockNoStreamHC(const struct compressionParameters * cparams,const char * src,char * dst,int srcSize,int dstSize)234 LZ4_compressBlockNoStreamHC(const struct compressionParameters* cparams,
235 const char* src, char* dst,
236 int srcSize, int dstSize)
237 {
238 return LZ4_compress_HC(src, dst, srcSize, dstSize, cparams->cLevel);
239 }
240
241 static int
LZ4_compressBlockStream(const struct compressionParameters * cparams,const char * src,char * dst,int srcSize,int dstSize)242 LZ4_compressBlockStream(const struct compressionParameters* cparams,
243 const char* src, char* dst,
244 int srcSize, int dstSize)
245 {
246 int const acceleration = (cparams->cLevel < 0) ? -cparams->cLevel + 1 : 1;
247 LZ4_compressResetStream(cparams);
248 return LZ4_compress_fast_continue(cparams->LZ4_stream, src, dst, srcSize, dstSize, acceleration);
249 }
250
251 static int
LZ4_compressBlockStreamHC(const struct compressionParameters * cparams,const char * src,char * dst,int srcSize,int dstSize)252 LZ4_compressBlockStreamHC(const struct compressionParameters* cparams,
253 const char* src, char* dst,
254 int srcSize, int dstSize)
255 {
256 LZ4_compressResetStreamHC(cparams);
257 return LZ4_compress_HC_continue(cparams->LZ4_streamHC, src, dst, srcSize, dstSize);
258 }
259
260 static void
LZ4_compressCleanupNoStream(const struct compressionParameters * cparams)261 LZ4_compressCleanupNoStream(const struct compressionParameters* cparams)
262 {
263 (void)cparams;
264 }
265
266 static void
LZ4_compressCleanupStream(const struct compressionParameters * cparams)267 LZ4_compressCleanupStream(const struct compressionParameters* cparams)
268 {
269 LZ4_freeStream(cparams->LZ4_stream);
270 LZ4_freeStream(cparams->LZ4_dictStream);
271 }
272
273 static void
LZ4_compressCleanupStreamHC(const struct compressionParameters * cparams)274 LZ4_compressCleanupStreamHC(const struct compressionParameters* cparams)
275 {
276 LZ4_freeStreamHC(cparams->LZ4_streamHC);
277 LZ4_freeStreamHC(cparams->LZ4_dictStreamHC);
278 }
279
280 static void
LZ4_buildCompressionParameters(struct compressionParameters * pParams,int cLevel,const char * dictBuf,int dictSize)281 LZ4_buildCompressionParameters(struct compressionParameters* pParams,
282 int cLevel,
283 const char* dictBuf, int dictSize)
284 {
285 pParams->cLevel = cLevel;
286 pParams->dictBuf = dictBuf;
287 pParams->dictSize = dictSize;
288
289 if (dictSize) {
290 if (cLevel < LZ4HC_CLEVEL_MIN) {
291 pParams->initFunction = LZ4_compressInitStream;
292 pParams->resetFunction = LZ4_compressResetStream;
293 pParams->blockFunction = LZ4_compressBlockStream;
294 pParams->cleanupFunction = LZ4_compressCleanupStream;
295 } else {
296 pParams->initFunction = LZ4_compressInitStreamHC;
297 pParams->resetFunction = LZ4_compressResetStreamHC;
298 pParams->blockFunction = LZ4_compressBlockStreamHC;
299 pParams->cleanupFunction = LZ4_compressCleanupStreamHC;
300 }
301 } else {
302 pParams->initFunction = LZ4_compressInitNoStream;
303 pParams->resetFunction = LZ4_compressResetNoStream;
304 pParams->cleanupFunction = LZ4_compressCleanupNoStream;
305
306 if (cLevel < LZ4HC_CLEVEL_MIN) {
307 pParams->blockFunction = LZ4_compressBlockNoStream;
308 } else {
309 pParams->blockFunction = LZ4_compressBlockNoStreamHC;
310 }
311 }
312 }
313
314
315 typedef int (*DecFunction_f)(const char* src, char* dst,
316 int srcSize, int dstCapacity,
317 const char* dictStart, int dictSize);
318
319 static LZ4F_dctx* g_dctx = NULL;
320
321 static int
LZ4F_decompress_binding(const char * src,char * dst,int srcSize,int dstCapacity,const char * dictStart,int dictSize)322 LZ4F_decompress_binding(const char* src, char* dst,
323 int srcSize, int dstCapacity,
324 const char* dictStart, int dictSize)
325 {
326 size_t dstSize = (size_t)dstCapacity;
327 size_t readSize = (size_t)srcSize;
328 LZ4F_decompressOptions_t dOpt = { 1, 0, 0, 0 };
329 size_t decStatus;
330 dOpt.skipChecksums = g_skipChecksums;
331 decStatus = LZ4F_decompress(g_dctx,
332 dst, &dstSize,
333 src, &readSize,
334 &dOpt);
335 if ( (decStatus == 0) /* decompression successful */
336 && ((int)readSize==srcSize) /* consume all input */ )
337 return (int)dstSize;
338 /* else, error */
339 return -1;
340 (void)dictStart; (void)dictSize; /* not compatible with dictionary yet */
341 }
342
343
344 /* ********************************************************
345 * Bench functions
346 **********************************************************/
347 typedef struct {
348 const char* srcPtr;
349 size_t srcSize;
350 char* cPtr;
351 size_t cRoom;
352 size_t cSize;
353 char* resPtr;
354 size_t resSize;
355 } blockParam_t;
356
357 #define MIN(a,b) ((a)<(b) ? (a) : (b))
358 #define MAX(a,b) ((a)>(b) ? (a) : (b))
359
BMK_benchMem(const void * srcBuffer,size_t srcSize,const char * displayName,int cLevel,const size_t * fileSizes,U32 nbFiles,const char * dictBuf,int dictSize)360 static int BMK_benchMem(const void* srcBuffer, size_t srcSize,
361 const char* displayName, int cLevel,
362 const size_t* fileSizes, U32 nbFiles,
363 const char* dictBuf, int dictSize)
364 {
365 size_t const blockSize = (g_blockSize>=32 && !g_decodeOnly ? g_blockSize : srcSize) + (!srcSize) /* avoid div by 0 */ ;
366 U32 const maxNbBlocks = (U32)((srcSize + (blockSize-1)) / blockSize) + nbFiles;
367 blockParam_t* const blockTable = (blockParam_t*) malloc(maxNbBlocks * sizeof(blockParam_t));
368 size_t const maxCompressedSize = (size_t)LZ4_compressBound((int)srcSize) + (maxNbBlocks * 1024); /* add some room for safety */
369 void* const compressedBuffer = malloc(maxCompressedSize);
370 size_t const decMultiplier = g_decodeOnly ? 255 : 1;
371 size_t const maxInSize = (size_t)LZ4_MAX_INPUT_SIZE / decMultiplier;
372 size_t const maxDecSize = srcSize < maxInSize ? srcSize * decMultiplier : LZ4_MAX_INPUT_SIZE;
373 void* const resultBuffer = malloc(maxDecSize);
374 int benchError = 0;
375 U32 nbBlocks;
376 struct compressionParameters compP;
377
378 /* checks */
379 if (!compressedBuffer || !resultBuffer || !blockTable)
380 END_PROCESS(31, "allocation error : not enough memory");
381
382 if (strlen(displayName)>17) displayName += strlen(displayName)-17; /* can only display 17 characters */
383
384 /* init */
385 LZ4_buildCompressionParameters(&compP, cLevel, dictBuf, dictSize);
386 compP.initFunction(&compP);
387 if (g_dctx==NULL) {
388 LZ4F_createDecompressionContext(&g_dctx, LZ4F_VERSION);
389 if (g_dctx==NULL)
390 END_PROCESS(1, "allocation error - decompression state");
391 }
392
393 /* Init blockTable data */
394 { const char* srcPtr = (const char*)srcBuffer;
395 char* cPtr = (char*)compressedBuffer;
396 char* resPtr = (char*)resultBuffer;
397 U32 fileNb;
398 for (nbBlocks=0, fileNb=0; fileNb<nbFiles; fileNb++) {
399 size_t remaining = fileSizes[fileNb];
400 U32 const nbBlocksforThisFile = (U32)((remaining + (blockSize-1)) / blockSize);
401 U32 const blockEnd = nbBlocks + nbBlocksforThisFile;
402 for ( ; nbBlocks<blockEnd; nbBlocks++) {
403 size_t const thisBlockSize = MIN(remaining, blockSize);
404 size_t const resMaxSize = thisBlockSize * decMultiplier;
405 size_t const resCapa = (thisBlockSize < maxInSize) ? resMaxSize : LZ4_MAX_INPUT_SIZE;
406 blockTable[nbBlocks].srcPtr = srcPtr;
407 blockTable[nbBlocks].cPtr = cPtr;
408 blockTable[nbBlocks].resPtr = resPtr;
409 blockTable[nbBlocks].srcSize = thisBlockSize;
410 blockTable[nbBlocks].cRoom = (size_t)LZ4_compressBound((int)thisBlockSize);
411 srcPtr += thisBlockSize;
412 cPtr += blockTable[nbBlocks].cRoom;
413 resPtr += resCapa;
414 remaining -= thisBlockSize;
415 } } }
416
417 /* warming up memory */
418 memset(compressedBuffer, ' ', maxCompressedSize);
419
420 /* decode-only mode : copy input to @compressedBuffer */
421 if (g_decodeOnly) {
422 U32 blockNb;
423 for (blockNb=0; blockNb < nbBlocks; blockNb++) {
424 memcpy(blockTable[blockNb].cPtr, blockTable[blockNb].srcPtr, blockTable[blockNb].srcSize);
425 blockTable[blockNb].cSize = blockTable[blockNb].srcSize;
426 } }
427
428 /* Bench */
429 { U64 fastestC = (U64)(-1LL), fastestD = (U64)(-1LL);
430 U64 const crcOrig = XXH64(srcBuffer, srcSize, 0);
431 TIME_t coolTime = TIME_getTime();
432 U64 const maxTime = (g_nbSeconds * TIMELOOP_NANOSEC) + 100;
433 U32 nbCompressionLoops = (U32)((5 MB) / (srcSize+1)) + 1; /* conservative initial compression speed estimate */
434 U32 nbDecodeLoops = (U32)((200 MB) / (srcSize+1)) + 1; /* conservative initial decode speed estimate */
435 Duration_ns totalCTime=0, totalDTime=0;
436 U32 cCompleted=(g_decodeOnly==1), dCompleted=0;
437 # define NB_MARKS 4
438 const char* const marks[NB_MARKS] = { " |", " /", " =", "\\" };
439 U32 markNb = 0;
440 size_t cSize = srcSize;
441 size_t totalRSize = srcSize;
442 double ratio = 0.;
443
444 DISPLAYLEVEL(2, "\r%79s\r", "");
445 if (g_nbSeconds==0) { nbCompressionLoops = 1; nbDecodeLoops = 1; }
446 while (!cCompleted || !dCompleted) {
447 /* overheat protection */
448 if (TIME_clockSpan_ns(coolTime) > ACTIVEPERIOD_NANOSEC) {
449 DISPLAYLEVEL(2, "\rcooling down ... \r");
450 UTIL_sleep(COOLPERIOD_SEC);
451 coolTime = TIME_getTime();
452 }
453
454 /* Compression */
455 DISPLAYLEVEL(2, "%2s-%-17.17s :%10u ->\r", marks[markNb], displayName, (U32)totalRSize);
456 if (!cCompleted) {
457 memset(compressedBuffer, 0xE5, maxCompressedSize); /* warm up and erase compressed buffer */
458 { U32 blockNb; for (blockNb=0; blockNb<nbBlocks; blockNb++) blockTable[blockNb].cSize = 0; }
459 }
460
461 UTIL_sleepMilli(1); /* give processor time to other processes */
462 TIME_waitForNextTick();
463
464 if (!cCompleted) { /* still some time to do compression tests */
465 TIME_t const timeStart = TIME_getTime();
466 U32 nbLoops;
467 for (nbLoops=0; nbLoops < nbCompressionLoops; nbLoops++) {
468 U32 blockNb;
469 compP.resetFunction(&compP);
470 for (blockNb=0; blockNb<nbBlocks; blockNb++) {
471 size_t const rSize = (size_t)compP.blockFunction(
472 &compP,
473 blockTable[blockNb].srcPtr, blockTable[blockNb].cPtr,
474 (int)blockTable[blockNb].srcSize, (int)blockTable[blockNb].cRoom);
475 if (LZ4_isError(rSize)) {
476 DISPLAY("LZ4 compression failed on block %u \n", blockNb);
477 benchError =1 ;
478 }
479 blockTable[blockNb].cSize = rSize;
480 } }
481 { Duration_ns const duration_ns = TIME_clockSpan_ns(timeStart);
482 if (duration_ns > 0) {
483 if (duration_ns < fastestC * nbCompressionLoops)
484 fastestC = duration_ns / nbCompressionLoops;
485 assert(fastestC > 0);
486 nbCompressionLoops = (U32)(TIMELOOP_NANOSEC / fastestC) + 1; /* aim for ~1sec */
487 } else {
488 assert(nbCompressionLoops < 40000000); /* avoid overflow */
489 nbCompressionLoops *= 100;
490 }
491 totalCTime += duration_ns;
492 cCompleted = totalCTime>maxTime;
493 }
494
495 cSize = 0;
496 { U32 blockNb; for (blockNb=0; blockNb<nbBlocks; blockNb++) cSize += blockTable[blockNb].cSize; }
497 cSize += !cSize; /* avoid div by 0 */
498 ratio = (double)totalRSize / (double)cSize;
499 markNb = (markNb+1) % NB_MARKS;
500 OUTLEVEL(2, "%2s-%-17.17s :%10u ->%10u (%5.3f),%6.1f MB/s\r",
501 marks[markNb], displayName,
502 (U32)totalRSize, (U32)cSize, ratio,
503 ((double)totalRSize / (double)fastestC) * 1000 );
504 fflush(NULL);
505 }
506 (void)fastestD; (void)crcOrig; /* unused when decompression disabled */
507 #if 1
508 /* Decompression */
509 if (!dCompleted) memset(resultBuffer, 0xD6, srcSize); /* warm result buffer */
510
511 UTIL_sleepMilli(5); /* give processor time to other processes */
512 TIME_waitForNextTick();
513
514 if (!dCompleted) {
515 const DecFunction_f decFunction = g_decodeOnly ?
516 LZ4F_decompress_binding : LZ4_decompress_safe_usingDict;
517 const char* const decString = g_decodeOnly ?
518 "LZ4F_decompress" : "LZ4_decompress_safe_usingDict";
519 TIME_t const timeStart = TIME_getTime();
520 U32 nbLoops;
521
522 for (nbLoops=0; nbLoops < nbDecodeLoops; nbLoops++) {
523 U32 blockNb;
524 for (blockNb=0; blockNb<nbBlocks; blockNb++) {
525 size_t const inMaxSize = (size_t)INT_MAX / decMultiplier;
526 size_t const resCapa = (blockTable[blockNb].srcSize < inMaxSize) ?
527 blockTable[blockNb].srcSize * decMultiplier :
528 INT_MAX;
529 int const regenSize = decFunction(
530 blockTable[blockNb].cPtr, blockTable[blockNb].resPtr,
531 (int)blockTable[blockNb].cSize, (int)resCapa,
532 dictBuf, dictSize);
533 if (regenSize < 0) {
534 DISPLAY("%s() failed on block %u of size %u \n",
535 decString, blockNb, (unsigned)blockTable[blockNb].srcSize);
536 if (g_decodeOnly)
537 DISPLAY("Is input using LZ4 Frame format ? \n");
538 benchError = 1;
539 break;
540 }
541 blockTable[blockNb].resSize = (size_t)regenSize;
542 } }
543 { Duration_ns const duration_ns = TIME_clockSpan_ns(timeStart);
544 if (duration_ns > 0) {
545 if (duration_ns < fastestD * nbDecodeLoops)
546 fastestD = duration_ns / nbDecodeLoops;
547 assert(fastestD > 0);
548 nbDecodeLoops = (U32)(TIMELOOP_NANOSEC / fastestD) + 1; /* aim for ~1sec */
549 } else {
550 assert(nbDecodeLoops < 40000000); /* avoid overflow */
551 nbDecodeLoops *= 100;
552 }
553 totalDTime += duration_ns;
554 dCompleted = totalDTime > (DECOMP_MULT*maxTime);
555 } }
556
557 if (g_decodeOnly) {
558 unsigned u;
559 totalRSize = 0;
560 for (u=0; u<nbBlocks; u++) totalRSize += blockTable[u].resSize;
561 }
562 markNb = (markNb+1) % NB_MARKS;
563 ratio = (double)totalRSize / (double)cSize;
564 OUTLEVEL(2, "%2s-%-17.17s :%10u ->%10u (%5.3f),%6.1f MB/s, %6.1f MB/s\r",
565 marks[markNb], displayName,
566 (U32)totalRSize, (U32)cSize, ratio,
567 ((double)totalRSize / (double)fastestC) * 1000,
568 ((double)totalRSize / (double)fastestD) * 1000);
569 fflush(NULL);
570
571 /* CRC Checking (not possible in decode-only mode)*/
572 if (!g_decodeOnly) {
573 U64 const crcCheck = XXH64(resultBuffer, srcSize, 0);
574 if (crcOrig!=crcCheck) {
575 size_t u;
576 DISPLAY("\n!!! WARNING !!! %17s : Invalid Checksum : %x != %x \n", displayName, (unsigned)crcOrig, (unsigned)crcCheck);
577 benchError = 1;
578 for (u=0; u<srcSize; u++) {
579 if (((const BYTE*)srcBuffer)[u] != ((const BYTE*)resultBuffer)[u]) {
580 U32 segNb, bNb, pos;
581 size_t bacc = 0;
582 DISPLAY("Decoding error at pos %u ", (U32)u);
583 for (segNb = 0; segNb < nbBlocks; segNb++) {
584 if (bacc + blockTable[segNb].srcSize > u) break;
585 bacc += blockTable[segNb].srcSize;
586 }
587 pos = (U32)(u - bacc);
588 bNb = pos / (128 KB);
589 DISPLAY("(block %u, sub %u, pos %u) \n", segNb, bNb, pos);
590 break;
591 }
592 if (u==srcSize-1) { /* should never happen */
593 DISPLAY("no difference detected\n");
594 } }
595 break;
596 } } /* CRC Checking */
597 #endif
598 } /* for (testNb = 1; testNb <= (g_nbSeconds + !g_nbSeconds); testNb++) */
599
600 OUTLEVEL(2, "%2i#\n", cLevel);
601
602 /* quiet mode */
603 if (g_displayLevel == 1) {
604 double const cSpeed = ((double)srcSize / (double)fastestC) * 1000;
605 double const dSpeed = ((double)srcSize / (double)fastestD) * 1000;
606 DISPLAYOUT("-%-3i%11i (%5.3f) %6.2f MB/s %6.1f MB/s %s ", cLevel, (int)cSize, ratio, cSpeed, dSpeed, displayName);
607 if (g_additionalParam)
608 DISPLAYOUT("(param=%d)", g_additionalParam);
609 DISPLAYOUT("\n");
610 }
611 } /* Bench */
612
613 /* clean up */
614 compP.cleanupFunction(&compP);
615 free(blockTable);
616 free(compressedBuffer);
617 free(resultBuffer);
618 return benchError;
619 }
620
621
BMK_findMaxMem(U64 requiredMem)622 static size_t BMK_findMaxMem(U64 requiredMem)
623 {
624 size_t step = 64 MB;
625 BYTE* testmem=NULL;
626
627 requiredMem = (((requiredMem >> 26) + 1) << 26);
628 requiredMem += 2*step;
629 if (requiredMem > maxMemory) requiredMem = maxMemory;
630
631 while (!testmem) {
632 if (requiredMem > step) requiredMem -= step;
633 else requiredMem >>= 1;
634 testmem = (BYTE*) malloc ((size_t)requiredMem);
635 }
636 free (testmem);
637
638 /* keep some space available */
639 if (requiredMem > step) requiredMem -= step;
640 else requiredMem >>= 1;
641
642 return (size_t)requiredMem;
643 }
644
645
BMK_benchCLevel(void * srcBuffer,size_t benchedSize,const char * displayName,int cLevel,int cLevelLast,const size_t * fileSizes,unsigned nbFiles,const char * dictBuf,int dictSize)646 static int BMK_benchCLevel(void* srcBuffer, size_t benchedSize,
647 const char* displayName, int cLevel, int cLevelLast,
648 const size_t* fileSizes, unsigned nbFiles,
649 const char* dictBuf, int dictSize)
650 {
651 int l;
652 int benchError = 0;
653 const char* pch = strrchr(displayName, '\\'); /* Windows */
654 if (!pch) pch = strrchr(displayName, '/'); /* Linux */
655 if (pch) displayName = pch+1;
656
657 SET_REALTIME_PRIORITY;
658
659 if (g_displayLevel == 1 && !g_additionalParam)
660 DISPLAY("bench %s %s: input %u bytes, %u seconds, %u KB blocks\n", LZ4_VERSION_STRING, LZ4_GIT_COMMIT_STRING, (U32)benchedSize, g_nbSeconds, (U32)(g_blockSize>>10));
661
662 if (cLevelLast < cLevel) cLevelLast = cLevel;
663
664 for (l=cLevel; l <= cLevelLast; l++) {
665 benchError |= BMK_benchMem(
666 srcBuffer, benchedSize,
667 displayName, l,
668 fileSizes, nbFiles,
669 dictBuf, dictSize);
670 }
671 return benchError;
672 }
673
674
675 /*! BMK_loadFiles() :
676 Loads `buffer` with content of files listed within `fileNamesTable`.
677 At most, fills `buffer` entirely */
BMK_loadFiles(void * buffer,size_t bufferSize,size_t * fileSizes,const char ** fileNamesTable,unsigned nbFiles)678 static void BMK_loadFiles(void* buffer, size_t bufferSize,
679 size_t* fileSizes,
680 const char** fileNamesTable, unsigned nbFiles)
681 {
682 size_t pos = 0, totalSize = 0;
683 unsigned n;
684 for (n=0; n<nbFiles; n++) {
685 FILE* f;
686 U64 fileSize = UTIL_getFileSize(fileNamesTable[n]);
687 if (UTIL_isDirectory(fileNamesTable[n])) {
688 DISPLAYLEVEL(2, "Ignoring %s directory... \n", fileNamesTable[n]);
689 fileSizes[n] = 0;
690 continue;
691 }
692 f = fopen(fileNamesTable[n], "rb");
693 if (f==NULL) END_PROCESS(10, "impossible to open file %s", fileNamesTable[n]);
694 DISPLAYUPDATE(2, "Loading %s... \r", fileNamesTable[n]);
695 if (fileSize > bufferSize-pos) { /* buffer too small - stop after this file */
696 fileSize = bufferSize-pos;
697 nbFiles=n;
698 }
699 { size_t const readSize = fread(((char*)buffer)+pos, 1, (size_t)fileSize, f);
700 if (readSize != (size_t)fileSize) END_PROCESS(11, "could not read %s", fileNamesTable[n]);
701 pos += readSize; }
702 fileSizes[n] = (size_t)fileSize;
703 totalSize += (size_t)fileSize;
704 fclose(f);
705 }
706
707 if (totalSize == 0) END_PROCESS(12, "no data to bench");
708 }
709
BMK_benchFileTable(const char ** fileNamesTable,unsigned nbFiles,int cLevel,int cLevelLast,const char * dictBuf,int dictSize)710 static int BMK_benchFileTable(const char** fileNamesTable, unsigned nbFiles,
711 int cLevel, int cLevelLast,
712 const char* dictBuf, int dictSize)
713 {
714 void* srcBuffer;
715 size_t benchedSize;
716 int benchError = 0;
717 size_t* fileSizes = (size_t*)malloc(nbFiles * sizeof(size_t));
718 U64 const totalSizeToLoad = UTIL_getTotalFileSize(fileNamesTable, nbFiles);
719 char mfName[20] = {0};
720
721 if (!fileSizes) END_PROCESS(12, "not enough memory for fileSizes");
722
723 /* Memory allocation & restrictions */
724 benchedSize = BMK_findMaxMem(totalSizeToLoad * 3) / 3;
725 if (benchedSize==0) END_PROCESS(12, "not enough memory");
726 if ((U64)benchedSize > totalSizeToLoad) benchedSize = (size_t)totalSizeToLoad;
727 if (benchedSize > LZ4_MAX_INPUT_SIZE) {
728 benchedSize = LZ4_MAX_INPUT_SIZE;
729 DISPLAY("File(s) bigger than LZ4's max input size; testing %u MB only...\n", (U32)(benchedSize >> 20));
730 } else {
731 if (benchedSize < totalSizeToLoad)
732 DISPLAY("Not enough memory; testing %u MB only...\n", (U32)(benchedSize >> 20));
733 }
734 srcBuffer = malloc(benchedSize + !benchedSize); /* avoid alloc of zero */
735 if (!srcBuffer) END_PROCESS(12, "not enough memory");
736
737 /* Load input buffer */
738 BMK_loadFiles(srcBuffer, benchedSize, fileSizes, fileNamesTable, nbFiles);
739
740 /* Bench */
741 snprintf (mfName, sizeof(mfName), " %u files", nbFiles);
742 { const char* displayName = (nbFiles > 1) ? mfName : fileNamesTable[0];
743 benchError = BMK_benchCLevel(srcBuffer, benchedSize,
744 displayName, cLevel, cLevelLast,
745 fileSizes, nbFiles,
746 dictBuf, dictSize);
747 }
748
749 /* clean up */
750 free(srcBuffer);
751 free(fileSizes);
752 return benchError;
753 }
754
755
BMK_syntheticTest(int cLevel,int cLevelLast,const char * dictBuf,int dictSize)756 static int BMK_syntheticTest(int cLevel, int cLevelLast,
757 const char* dictBuf, int dictSize)
758 {
759 int benchError = 0;
760 size_t const benchedSize = 10000000;
761 void* const srcBuffer = malloc(benchedSize);
762
763 /* Memory allocation */
764 if (!srcBuffer) END_PROCESS(21, "not enough memory");
765
766 /* Fill input buffer */
767 LOREM_genBuffer(srcBuffer, benchedSize, 0);
768
769 /* Bench */
770 benchError = BMK_benchCLevel(srcBuffer, benchedSize,
771 "Lorem ipsum",
772 cLevel, cLevelLast,
773 &benchedSize,
774 1,
775 dictBuf, dictSize);
776
777 /* clean up */
778 free(srcBuffer);
779
780 return benchError;
781 }
782
783
784 static int
BMK_benchFilesSeparately(const char ** fileNamesTable,unsigned nbFiles,int cLevel,int cLevelLast,const char * dictBuf,int dictSize)785 BMK_benchFilesSeparately(const char** fileNamesTable, unsigned nbFiles,
786 int cLevel, int cLevelLast,
787 const char* dictBuf, int dictSize)
788 {
789 int benchError = 0;
790 unsigned fileNb;
791 if (cLevel > LZ4HC_CLEVEL_MAX) cLevel = LZ4HC_CLEVEL_MAX;
792 if (cLevelLast > LZ4HC_CLEVEL_MAX) cLevelLast = LZ4HC_CLEVEL_MAX;
793 if (cLevelLast < cLevel) cLevelLast = cLevel;
794
795 for (fileNb=0; fileNb<nbFiles; fileNb++)
796 benchError |= BMK_benchFileTable(fileNamesTable+fileNb, 1, cLevel, cLevelLast, dictBuf, dictSize);
797
798 return benchError;
799 }
800
801
BMK_benchFiles(const char ** fileNamesTable,unsigned nbFiles,int cLevel,int cLevelLast,const char * dictFileName)802 int BMK_benchFiles(const char** fileNamesTable, unsigned nbFiles,
803 int cLevel, int cLevelLast,
804 const char* dictFileName)
805 {
806 int benchError = 0;
807 char* dictBuf = NULL;
808 size_t dictSize = 0;
809
810 if (cLevel > LZ4HC_CLEVEL_MAX) cLevel = LZ4HC_CLEVEL_MAX;
811 if (g_decodeOnly) {
812 DISPLAYLEVEL(2, "Benchmark Decompression of LZ4 Frame ");
813 if (g_skipChecksums) {
814 DISPLAYLEVEL(2, "_without_ checksum even when present \n");
815 } else {
816 DISPLAYLEVEL(2, "+ Checksum when present \n");
817 }
818 cLevelLast = cLevel;
819 }
820 if (cLevelLast > LZ4HC_CLEVEL_MAX) cLevelLast = LZ4HC_CLEVEL_MAX;
821 if (cLevelLast < cLevel) cLevelLast = cLevel;
822 if (cLevelLast > cLevel)
823 DISPLAYLEVEL(2, "Benchmarking levels from %d to %d\n", cLevel, cLevelLast);
824
825 if (dictFileName) {
826 FILE* dictFile = NULL;
827 U64 const dictFileSize = UTIL_getFileSize(dictFileName);
828 if (!dictFileSize)
829 END_PROCESS(25, "Dictionary error : could not stat dictionary file");
830 if (g_decodeOnly)
831 END_PROCESS(26, "Error : LZ4 Frame decoder mode not compatible with dictionary yet");
832
833 dictFile = fopen(dictFileName, "rb");
834 if (!dictFile)
835 END_PROCESS(25, "Dictionary error : could not open dictionary file");
836
837 if (dictFileSize > LZ4_MAX_DICT_SIZE) {
838 dictSize = LZ4_MAX_DICT_SIZE;
839 if (UTIL_fseek(dictFile, (long)(dictFileSize - dictSize), SEEK_SET))
840 END_PROCESS(25, "Dictionary error : could not seek dictionary file");
841 } else {
842 dictSize = (size_t)dictFileSize;
843 }
844
845 dictBuf = (char*)malloc(dictSize);
846 if (!dictBuf) END_PROCESS(25, "Allocation error : not enough memory");
847
848 if (fread(dictBuf, 1, dictSize, dictFile) != dictSize)
849 END_PROCESS(25, "Dictionary error : could not read dictionary file");
850
851 fclose(dictFile);
852 }
853
854 if (nbFiles == 0) {
855 benchError = BMK_syntheticTest(cLevel, cLevelLast, dictBuf, (int)dictSize);
856 } else {
857 if (g_benchSeparately)
858 benchError = BMK_benchFilesSeparately(fileNamesTable, nbFiles, cLevel, cLevelLast, dictBuf, (int)dictSize);
859 else
860 benchError = BMK_benchFileTable(fileNamesTable, nbFiles, cLevel, cLevelLast, dictBuf, (int)dictSize);
861 }
862
863 free(dictBuf);
864 return benchError;
865 }
866