blob: 77d7c533a8e01ed6398a3d590f6b0cd232d7768b [file] [log] [blame]
Yann Collet4856a002015-01-24 01:58:16 +01001/*
Yann Collet99909862016-04-09 16:17:18 +02002 bench.c - open-source compression benchmark module
3 Copyright (C) Yann Collet 2012-2016
Yann Collet4856a002015-01-24 01:58:16 +01004
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 :
Yann Collet99909862016-04-09 16:17:18 +020022 - zstd homepage : http://www.zstd.net
Yann Collet4856a002015-01-24 01:58:16 +010023 - zstd source repository : https://github.com/Cyan4973/zstd
Yann Collet4856a002015-01-24 01:58:16 +010024*/
25
Yann Colletf3eca252015-10-22 15:31:46 +010026/* **************************************
Yann Collet4856a002015-01-24 01:58:16 +010027* Compiler Options
Yann Colletf3eca252015-10-22 15:31:46 +010028****************************************/
Yann Collet4856a002015-01-24 01:58:16 +010029/* Disable some Visual warning messages */
Yann Collet6c8b9252015-12-16 02:44:56 +010030#ifdef _MSC_VER
31# define _CRT_SECURE_NO_WARNINGS /* fopen */
32# pragma warning(disable : 4127) /* disable: C4127: conditional expression is constant */
33#endif
Yann Collet4856a002015-01-24 01:58:16 +010034
Yann Colletf3eca252015-10-22 15:31:46 +010035/* Unix Large Files support (>4GB) */
Yann Collet4856a002015-01-24 01:58:16 +010036#define _FILE_OFFSET_BITS 64
Yann Colletf3eca252015-10-22 15:31:46 +010037#if (defined(__sun__) && (!defined(__LP64__))) /* Sun Solaris 32-bits requires specific definitions */
Yann Collet4856a002015-01-24 01:58:16 +010038# define _LARGEFILE_SOURCE
Yann Colletf3eca252015-10-22 15:31:46 +010039#elif ! defined(__LP64__) /* No point defining Large file for 64 bit */
Yann Collet4856a002015-01-24 01:58:16 +010040# define _LARGEFILE64_SOURCE
41#endif
42
Yann Collet4856a002015-01-24 01:58:16 +010043
Yann Colletf3eca252015-10-22 15:31:46 +010044/* *************************************
Yann Collet4856a002015-01-24 01:58:16 +010045* Includes
Yann Colletf3eca252015-10-22 15:31:46 +010046***************************************/
Yann Collet99909862016-04-09 16:17:18 +020047#define _POSIX_C_SOURCE 199309L /* before <time.h> - needed for nanosleep() */
Yann Collet4856a002015-01-24 01:58:16 +010048#include <stdlib.h> /* malloc, free */
49#include <string.h> /* memset */
Yann Colleteeb8ba12015-10-22 16:55:40 +010050#include <stdio.h> /* fprintf, fopen, ftello64 */
51#include <sys/types.h> /* stat64 */
52#include <sys/stat.h> /* stat64 */
Yann Collet99909862016-04-09 16:17:18 +020053#include <time.h> /* clock_t, nanosleep, clock, CLOCKS_PER_SEC */
Yann Collet4856a002015-01-24 01:58:16 +010054
Yann Collet27d3dad2016-03-11 13:41:20 +010055/* sleep : posix - windows - others */
56#if !defined(_WIN32) && (defined(__unix__) || defined(__unix) || (defined(__APPLE__) && defined(__MACH__)))
57# include <unistd.h>
inikep472638c2016-03-23 12:28:28 +010058# include <sys/resource.h> /* setpriority */
Yann Collet27d3dad2016-03-11 13:41:20 +010059# define BMK_sleep(s) sleep(s)
inikep4c12f232016-03-29 14:52:13 +020060# define mili_sleep(mili) { struct timespec t; t.tv_sec=0; t.tv_nsec=mili*1000000ULL; nanosleep(&t, NULL); }
inikep19bd48f2016-04-04 12:10:00 +020061# define SET_HIGH_PRIORITY setpriority(PRIO_PROCESS, 0, -20)
Yann Collet27d3dad2016-03-11 13:41:20 +010062#elif defined(_WIN32)
63# include <windows.h>
64# define BMK_sleep(s) Sleep(1000*s)
inikep472638c2016-03-23 12:28:28 +010065# define mili_sleep(mili) Sleep(mili)
inikep19bd48f2016-04-04 12:10:00 +020066# define SET_HIGH_PRIORITY SetPriorityClass(GetCurrentProcess(), REALTIME_PRIORITY_CLASS)
Yann Collet4856a002015-01-24 01:58:16 +010067#else
Yann Collet27d3dad2016-03-11 13:41:20 +010068# define BMK_sleep(s) /* disabled */
inikep472638c2016-03-23 12:28:28 +010069# define mili_sleep(mili) /* disabled */
inikep19bd48f2016-04-04 12:10:00 +020070# define SET_HIGH_PRIORITY /* disabled */
Yann Collet4856a002015-01-24 01:58:16 +010071#endif
72
inikep1eeddde2016-04-08 16:55:17 +020073#if !defined(_WIN32)
inikep4c12f232016-03-29 14:52:13 +020074 typedef clock_t BMK_time_t;
inikep1c556a32016-03-30 10:59:48 +020075# define BMK_initTimer(ticksPerSecond) ticksPerSecond=0
inikep4c12f232016-03-29 14:52:13 +020076# define BMK_getTime(x) x = clock()
77# define BMK_getSpanTimeMicro(ticksPerSecond, clockStart, clockEnd) (1000000ULL * (clockEnd - clockStart) / CLOCKS_PER_SEC)
78# define BMK_getSpanTimeNano(ticksPerSecond, clockStart, clockEnd) (1000000000ULL * (clockEnd - clockStart) / CLOCKS_PER_SEC)
inikep1eeddde2016-04-08 16:55:17 +020079#else
inikep4c12f232016-03-29 14:52:13 +020080 typedef LARGE_INTEGER BMK_time_t;
inikep4c12f232016-03-29 14:52:13 +020081# define BMK_initTimer(x) if (!QueryPerformanceFrequency(&x)) { fprintf(stderr, "ERROR: QueryPerformance not present\n"); }
82# define BMK_getTime(x) QueryPerformanceCounter(&x)
83# define BMK_getSpanTimeMicro(ticksPerSecond, clockStart, clockEnd) (1000000ULL*(clockEnd.QuadPart - clockStart.QuadPart)/ticksPerSecond.QuadPart)
84# define BMK_getSpanTimeNano(ticksPerSecond, clockStart, clockEnd) (1000000000ULL*(clockEnd.QuadPart - clockStart.QuadPart)/ticksPerSecond.QuadPart)
inikep4c12f232016-03-29 14:52:13 +020085#endif
86
Yann Colletf3eca252015-10-22 15:31:46 +010087#include "mem.h"
Yann Collet31683c02015-12-18 01:26:48 +010088#include "zstd_static.h"
inikep69fcd7c2016-04-28 12:23:33 +020089#include "datagen.h" /* RDG_genBuffer */
Yann Collet4856a002015-01-24 01:58:16 +010090#include "xxhash.h"
inikep69fcd7c2016-04-28 12:23:33 +020091#include "util.h" /* UTIL_GetFileSize */
Yann Collet4856a002015-01-24 01:58:16 +010092
93
Yann Colletf3eca252015-10-22 15:31:46 +010094/* *************************************
Yann Collet4856a002015-01-24 01:58:16 +010095* Compiler specifics
Yann Colletf3eca252015-10-22 15:31:46 +010096***************************************/
Yann Collet99909862016-04-09 16:17:18 +020097#if defined(_MSC_VER)
98# define snprintf sprintf_s
99#elif defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L))
100 /* part of <stdio.h> */
101#else
102 extern int snprintf (char* s, size_t maxlen, const char* format, ...); /* not declared in <stdio.h> when C version < c99 */
Yann Colleted699e62015-12-16 02:37:24 +0100103#endif
104
Yann Collet4856a002015-01-24 01:58:16 +0100105
Yann Colletf3eca252015-10-22 15:31:46 +0100106/* *************************************
Yann Collet4856a002015-01-24 01:58:16 +0100107* Constants
Yann Colletf3eca252015-10-22 15:31:46 +0100108***************************************/
inikep44af12d2016-03-14 15:59:04 +0100109#ifndef ZSTD_VERSION
110# define ZSTD_VERSION ""
111#endif
112
Yann Collet27d3dad2016-03-11 13:41:20 +0100113#define NBLOOPS 3
Yann Collet699b14d2016-03-17 19:37:33 +0100114#define TIMELOOP_S 1
Yann Collet27d3dad2016-03-11 13:41:20 +0100115#define ACTIVEPERIOD_S 70
116#define COOLPERIOD_S 10
Yann Collet4856a002015-01-24 01:58:16 +0100117
118#define KB *(1 <<10)
119#define MB *(1 <<20)
120#define GB *(1U<<30)
121
Yann Colletd062f132015-12-01 01:31:17 +0100122static const size_t maxMemory = (sizeof(size_t)==4) ? (2 GB - 64 MB) : (size_t)(1ULL << ((sizeof(size_t)*8)-31));
Yann Collet4856a002015-01-24 01:58:16 +0100123
124static U32 g_compressibilityDefault = 50;
Yann Collet4856a002015-01-24 01:58:16 +0100125
126
Yann Colletf3eca252015-10-22 15:31:46 +0100127/* *************************************
Yann Colleted699e62015-12-16 02:37:24 +0100128* console display
Yann Colletf3eca252015-10-22 15:31:46 +0100129***************************************/
Yann Colleted699e62015-12-16 02:37:24 +0100130#define DISPLAY(...) fprintf(stderr, __VA_ARGS__)
131#define DISPLAYLEVEL(l, ...) if (g_displayLevel>=l) { DISPLAY(__VA_ARGS__); }
132static U32 g_displayLevel = 2; /* 0 : no display; 1: errors; 2 : + result + interaction + warnings; 3 : + progression; 4 : + information */
133
134
135/* *************************************
136* Exceptions
137***************************************/
138#ifndef DEBUG
139# define DEBUG 0
140#endif
141#define DEBUGOUTPUT(...) if (DEBUG) DISPLAY(__VA_ARGS__);
142#define EXM_THROW(error, ...) \
143{ \
144 DEBUGOUTPUT("Error defined at %s, line %i : \n", __FILE__, __LINE__); \
145 DISPLAYLEVEL(1, "Error %i : ", error); \
146 DISPLAYLEVEL(1, __VA_ARGS__); \
147 DISPLAYLEVEL(1, "\n"); \
148 exit(error); \
149}
Yann Collet4856a002015-01-24 01:58:16 +0100150
151
Yann Colletf3eca252015-10-22 15:31:46 +0100152/* *************************************
Yann Collet4856a002015-01-24 01:58:16 +0100153* Benchmark Parameters
Yann Colletf3eca252015-10-22 15:31:46 +0100154***************************************/
Yann Collet1d1ae402016-03-17 19:51:02 +0100155static U32 g_nbIterations = NBLOOPS;
Yann Collet1c00dc32015-10-21 08:22:25 +0100156static size_t g_blockSize = 0;
inikepeaba91a2016-03-23 20:30:26 +0100157int g_additionalParam = 0;
Yann Collet4856a002015-01-24 01:58:16 +0100158
inikep4e26bb62016-03-14 12:48:51 +0100159void BMK_setNotificationLevel(unsigned level) { g_displayLevel=level; }
160
inikepeaba91a2016-03-23 20:30:26 +0100161void BMK_setAdditionalParam(int additionalParam) { g_additionalParam=additionalParam; }
162
Yann Collet1d1ae402016-03-17 19:51:02 +0100163void BMK_SetNbIterations(unsigned nbLoops)
Yann Collet4856a002015-01-24 01:58:16 +0100164{
Yann Collet1d1ae402016-03-17 19:51:02 +0100165 g_nbIterations = nbLoops;
inikep2872b6f2016-03-22 14:38:34 +0100166 DISPLAYLEVEL(2, "- %i iterations -\n", g_nbIterations);
Yann Collet4856a002015-01-24 01:58:16 +0100167}
168
Yann Collet1c00dc32015-10-21 08:22:25 +0100169void BMK_SetBlockSize(size_t blockSize)
170{
171 g_blockSize = blockSize;
inikep4e26bb62016-03-14 12:48:51 +0100172 DISPLAYLEVEL(2, "using blocks of size %u KB \n", (U32)(blockSize>>10));
Yann Collet1c00dc32015-10-21 08:22:25 +0100173}
174
Yann Collet4856a002015-01-24 01:58:16 +0100175
Yann Colletf3eca252015-10-22 15:31:46 +0100176/* ********************************************************
Yann Collet4856a002015-01-24 01:58:16 +0100177* Private functions
Yann Colletf3eca252015-10-22 15:31:46 +0100178**********************************************************/
inikep4c12f232016-03-29 14:52:13 +0200179/* returns time span in microseconds */
180static U64 BMK_clockSpan( BMK_time_t clockStart, BMK_time_t ticksPerSecond )
Yann Collet4856a002015-01-24 01:58:16 +0100181{
inikep4c12f232016-03-29 14:52:13 +0200182 BMK_time_t clockEnd;
Yann Collet99909862016-04-09 16:17:18 +0200183
inikep4c12f232016-03-29 14:52:13 +0200184 (void)ticksPerSecond;
185 BMK_getTime(clockEnd);
186 return BMK_getSpanTimeMicro(ticksPerSecond, clockStart, clockEnd);
inikep06f793a2016-03-29 11:17:58 +0200187}
188
Yann Collet4856a002015-01-24 01:58:16 +0100189
Yann Colletf3eca252015-10-22 15:31:46 +0100190/* ********************************************************
Yann Collet4856a002015-01-24 01:58:16 +0100191* Bench functions
Yann Colletf3eca252015-10-22 15:31:46 +0100192**********************************************************/
Yann Collet1c00dc32015-10-21 08:22:25 +0100193typedef struct
Yann Collet4856a002015-01-24 01:58:16 +0100194{
Yann Colleted699e62015-12-16 02:37:24 +0100195 const char* srcPtr;
Yann Collet1c00dc32015-10-21 08:22:25 +0100196 size_t srcSize;
197 char* cPtr;
198 size_t cRoom;
199 size_t cSize;
200 char* resPtr;
201 size_t resSize;
202} blockParam_t;
203
inikep4e26bb62016-03-14 12:48:51 +0100204typedef struct
205{
inikepc034b732016-03-14 13:13:42 +0100206 double ratio;
inikep4e26bb62016-03-14 12:48:51 +0100207 size_t cSize;
inikepc034b732016-03-14 13:13:42 +0100208 double cSpeed;
209 double dSpeed;
inikep4e26bb62016-03-14 12:48:51 +0100210} benchResult_t;
211
Yann Collet99909862016-04-09 16:17:18 +0200212
Yann Colletbe2010e2015-10-31 12:57:14 +0100213#define MIN(a,b) ((a)<(b) ? (a) : (b))
Yann Collet2ce49232016-02-02 14:36:49 +0100214#define MAX(a,b) ((a)>(b) ? (a) : (b))
Yann Collet1c00dc32015-10-21 08:22:25 +0100215
Yann Colleted699e62015-12-16 02:37:24 +0100216static int BMK_benchMem(const void* srcBuffer, size_t srcSize,
inikep472638c2016-03-23 12:28:28 +0100217 const char* displayName, int cLevel,
Yann Collet31683c02015-12-18 01:26:48 +0100218 const size_t* fileSizes, U32 nbFiles,
inikep4e26bb62016-03-14 12:48:51 +0100219 const void* dictBuffer, size_t dictBufferSize, benchResult_t *result)
Yann Collet1c00dc32015-10-21 08:22:25 +0100220{
inikep38654982016-04-21 12:18:47 +0200221 size_t const blockSize = (g_blockSize>=32 ? g_blockSize : srcSize) + (!srcSize); /* avoid div by 0 */
Yann Colletd64f4352016-03-21 00:07:42 +0100222 U32 const maxNbBlocks = (U32) ((srcSize + (blockSize-1)) / blockSize) + nbFiles;
Yann Colleted699e62015-12-16 02:37:24 +0100223 blockParam_t* const blockTable = (blockParam_t*) malloc(maxNbBlocks * sizeof(blockParam_t));
Yann Colletb9151402016-03-26 17:18:11 +0100224 size_t const maxCompressedSize = ZSTD_compressBound(srcSize) + (maxNbBlocks * 1024); /* add some room for safety */
Yann Collet1c00dc32015-10-21 08:22:25 +0100225 void* const compressedBuffer = malloc(maxCompressedSize);
226 void* const resultBuffer = malloc(srcSize);
Yann Collet2630a5e2016-01-14 19:13:22 +0100227 ZSTD_CCtx* refCtx = ZSTD_createCCtx();
Yann Collet31683c02015-12-18 01:26:48 +0100228 ZSTD_CCtx* ctx = ZSTD_createCCtx();
Yann Collet7b51a292016-01-26 15:58:49 +0100229 ZSTD_DCtx* refDCtx = ZSTD_createDCtx();
Yann Collet31683c02015-12-18 01:26:48 +0100230 ZSTD_DCtx* dctx = ZSTD_createDCtx();
Yann Colletde406ee2016-03-20 15:46:10 +0100231 U32 nbBlocks;
inikep4c12f232016-03-29 14:52:13 +0200232 BMK_time_t ticksPerSecond;
Yann Colletde406ee2016-03-20 15:46:10 +0100233
234 /* checks */
Yann Collet7b51a292016-01-26 15:58:49 +0100235 if (!compressedBuffer || !resultBuffer || !blockTable || !refCtx || !ctx || !refDCtx || !dctx)
Yann Colleted699e62015-12-16 02:37:24 +0100236 EXM_THROW(31, "not enough memory");
Yann Collet4856a002015-01-24 01:58:16 +0100237
Yann Collet4856a002015-01-24 01:58:16 +0100238 /* init */
239 if (strlen(displayName)>17) displayName += strlen(displayName)-17; /* can only display 17 characters */
inikep4c12f232016-03-29 14:52:13 +0200240 BMK_initTimer(ticksPerSecond);
241
Yann Collet1c00dc32015-10-21 08:22:25 +0100242 /* Init blockTable data */
Yann Colletde406ee2016-03-20 15:46:10 +0100243 { const char* srcPtr = (const char*)srcBuffer;
Yann Collet1c00dc32015-10-21 08:22:25 +0100244 char* cPtr = (char*)compressedBuffer;
245 char* resPtr = (char*)resultBuffer;
Yann Collet699b14d2016-03-17 19:37:33 +0100246 U32 fileNb;
Yann Colletde406ee2016-03-20 15:46:10 +0100247 for (nbBlocks=0, fileNb=0; fileNb<nbFiles; fileNb++) {
Yann Collet70611352015-12-16 03:01:03 +0100248 size_t remaining = fileSizes[fileNb];
Yann Collet699b14d2016-03-17 19:37:33 +0100249 U32 const nbBlocksforThisFile = (U32)((remaining + (blockSize-1)) / blockSize);
250 U32 const blockEnd = nbBlocks + nbBlocksforThisFile;
Yann Colletfd416f12016-01-30 03:14:15 +0100251 for ( ; nbBlocks<blockEnd; nbBlocks++) {
Yann Colletde406ee2016-03-20 15:46:10 +0100252 size_t const thisBlockSize = MIN(remaining, blockSize);
Yann Colleted699e62015-12-16 02:37:24 +0100253 blockTable[nbBlocks].srcPtr = srcPtr;
254 blockTable[nbBlocks].cPtr = cPtr;
255 blockTable[nbBlocks].resPtr = resPtr;
256 blockTable[nbBlocks].srcSize = thisBlockSize;
257 blockTable[nbBlocks].cRoom = ZSTD_compressBound(thisBlockSize);
258 srcPtr += thisBlockSize;
259 cPtr += blockTable[nbBlocks].cRoom;
260 resPtr += thisBlockSize;
261 remaining -= thisBlockSize;
Yann Colletfd416f12016-01-30 03:14:15 +0100262 } } }
Yann Collet1c00dc32015-10-21 08:22:25 +0100263
Yann Collet4856a002015-01-24 01:58:16 +0100264 /* warmimg up memory */
Yann Colletd062f132015-12-01 01:31:17 +0100265 RDG_genBuffer(compressedBuffer, maxCompressedSize, 0.10, 0.50, 1);
Yann Collet4856a002015-01-24 01:58:16 +0100266
267 /* Bench */
inikep6d157f12016-04-15 16:54:11 +0200268 { U64 fastestC = (U64)(-1LL), fastestD = (U64)(-1LL);
Yann Colletb9151402016-03-26 17:18:11 +0100269 U64 const crcOrig = XXH64(srcBuffer, srcSize, 0);
Yann Collet4856a002015-01-24 01:58:16 +0100270 U64 crcCheck = 0;
inikep4c12f232016-03-29 14:52:13 +0200271 BMK_time_t coolTime;
Yann Colletde406ee2016-03-20 15:46:10 +0100272 U32 testNb;
inikep19bd48f2016-04-04 12:10:00 +0200273 size_t cSize = 0;
274 double ratio = 0.;
Yann Collet4856a002015-01-24 01:58:16 +0100275
inikep4c12f232016-03-29 14:52:13 +0200276 BMK_getTime(coolTime);
inikep4e26bb62016-03-14 12:48:51 +0100277 DISPLAYLEVEL(2, "\r%79s\r", "");
Yann Colletde406ee2016-03-20 15:46:10 +0100278 for (testNb = 1; testNb <= (g_nbIterations + !g_nbIterations); testNb++) {
inikep4c12f232016-03-29 14:52:13 +0200279 BMK_time_t clockStart, clockEnd;
inikepc5e1d292016-04-19 09:37:59 +0200280 U64 clockLoop = g_nbIterations ? TIMELOOP_S*1000000ULL : 1;
Yann Collet4856a002015-01-24 01:58:16 +0100281
Yann Collet27d3dad2016-03-11 13:41:20 +0100282 /* overheat protection */
inikep4c12f232016-03-29 14:52:13 +0200283 if (BMK_clockSpan(coolTime, ticksPerSecond) > ACTIVEPERIOD_S*1000000ULL) {
Yann Collet27d3dad2016-03-11 13:41:20 +0100284 DISPLAY("\rcooling down ... \r");
285 BMK_sleep(COOLPERIOD_S);
inikep4c12f232016-03-29 14:52:13 +0200286 BMK_getTime(coolTime);
Yann Collet27d3dad2016-03-11 13:41:20 +0100287 }
Yann Collet4856a002015-01-24 01:58:16 +0100288
289 /* Compression */
inikep2872b6f2016-03-22 14:38:34 +0100290 DISPLAYLEVEL(2, "%2i-%-17.17s :%10u ->\r", testNb, displayName, (U32)srcSize);
Yann Collet699b14d2016-03-17 19:37:33 +0100291 memset(compressedBuffer, 0xE5, maxCompressedSize); /* warm up and erase result buffer */
Yann Collet4856a002015-01-24 01:58:16 +0100292
inikep472638c2016-03-23 12:28:28 +0100293 mili_sleep(1); /* give processor time to other processes */
inikep4c12f232016-03-29 14:52:13 +0200294 BMK_getTime(clockStart);
295 do { BMK_getTime(clockEnd); }
296 while (BMK_getSpanTimeNano(ticksPerSecond, clockStart, clockEnd) == 0);
297 BMK_getTime(clockStart);
Yann Colletde406ee2016-03-20 15:46:10 +0100298
inikepc5e1d292016-04-19 09:37:59 +0200299 { U32 nbLoops = 0;
300 do {
Yann Colleta5b66e32016-03-26 01:48:27 +0100301 U32 blockNb;
inikep4b3c5ee2016-04-14 13:43:51 +0200302 { ZSTD_parameters params;
303 params.cParams = ZSTD_getCParams(cLevel, blockSize, dictBufferSize);
304 params.fParams.contentSizeFlag = 1;
305 ZSTD_adjustCParams(&params.cParams, blockSize, dictBufferSize);
306 { size_t const initResult = ZSTD_compressBegin_advanced(refCtx, dictBuffer, dictBufferSize, params, blockSize);
307 if (ZSTD_isError(initResult)) break;
308 } }
Yann Colleta5b66e32016-03-26 01:48:27 +0100309 for (blockNb=0; blockNb<nbBlocks; blockNb++) {
310 size_t const rSize = ZSTD_compress_usingPreparedCCtx(ctx, refCtx,
311 blockTable[blockNb].cPtr, blockTable[blockNb].cRoom,
312 blockTable[blockNb].srcPtr,blockTable[blockNb].srcSize);
313 if (ZSTD_isError(rSize)) EXM_THROW(1, "ZSTD_compress_usingPreparedCCtx() failed : %s", ZSTD_getErrorName(rSize));
314 blockTable[blockNb].cSize = rSize;
inikepc5e1d292016-04-19 09:37:59 +0200315 }
316 nbLoops++;
317 } while (BMK_clockSpan(clockStart, ticksPerSecond) < clockLoop);
inikep19bd48f2016-04-04 12:10:00 +0200318 { U64 const clockSpan = BMK_clockSpan(clockStart, ticksPerSecond);
inikep6d157f12016-04-15 16:54:11 +0200319 if (clockSpan < fastestC*nbLoops) fastestC = clockSpan / nbLoops;
Yann Colletde406ee2016-03-20 15:46:10 +0100320 } }
Yann Collet4856a002015-01-24 01:58:16 +0100321
Yann Collet1c00dc32015-10-21 08:22:25 +0100322 cSize = 0;
Yann Colletde406ee2016-03-20 15:46:10 +0100323 { U32 blockNb; for (blockNb=0; blockNb<nbBlocks; blockNb++) cSize += blockTable[blockNb].cSize; }
Yann Collet2acb5d32015-10-29 16:49:43 +0100324 ratio = (double)srcSize / (double)cSize;
inikep2872b6f2016-03-22 14:38:34 +0100325 DISPLAYLEVEL(2, "%2i-%-17.17s :%10u ->%10u (%5.3f),%6.1f MB/s\r",
Yann Colletde406ee2016-03-20 15:46:10 +0100326 testNb, displayName, (U32)srcSize, (U32)cSize, ratio,
inikep06f793a2016-03-29 11:17:58 +0200327 (double)srcSize / fastestC );
Yann Collet4856a002015-01-24 01:58:16 +0100328
Yann Colleta5b66e32016-03-26 01:48:27 +0100329 (void)fastestD; (void)crcOrig; /* unused when decompression disabled */
Yann Collete93d6ce2016-01-31 00:58:06 +0100330#if 1
Yann Collet4856a002015-01-24 01:58:16 +0100331 /* Decompression */
Yann Collet7b51a292016-01-26 15:58:49 +0100332 memset(resultBuffer, 0xD6, srcSize); /* warm result buffer */
Yann Collet4856a002015-01-24 01:58:16 +0100333
inikep472638c2016-03-23 12:28:28 +0100334 mili_sleep(1); /* give processor time to other processes */
inikep4c12f232016-03-29 14:52:13 +0200335 BMK_getTime(clockStart);
336 do { BMK_getTime(clockEnd); }
337 while (BMK_getSpanTimeNano(ticksPerSecond, clockStart, clockEnd) == 0);
338 BMK_getTime(clockStart);
Yann Collet7b51a292016-01-26 15:58:49 +0100339
inikepc5e1d292016-04-19 09:37:59 +0200340 { U32 nbLoops = 0;
341 do {
Yann Colleta5b66e32016-03-26 01:48:27 +0100342 U32 blockNb;
343 ZSTD_decompressBegin_usingDict(refDCtx, dictBuffer, dictBufferSize);
344 for (blockNb=0; blockNb<nbBlocks; blockNb++) {
345 size_t const regenSize = ZSTD_decompress_usingPreparedDCtx(dctx, refDCtx,
346 blockTable[blockNb].resPtr, blockTable[blockNb].srcSize,
347 blockTable[blockNb].cPtr, blockTable[blockNb].cSize);
348 if (ZSTD_isError(regenSize)) {
349 DISPLAY("ZSTD_decompress_usingPreparedDCtx() failed on block %u : %s \n",
350 blockNb, ZSTD_getErrorName(regenSize));
inikep19bd48f2016-04-04 12:10:00 +0200351 clockLoop = 0; /* force immediate test end */
Yann Colleta5b66e32016-03-26 01:48:27 +0100352 break;
353 }
354 blockTable[blockNb].resSize = regenSize;
inikepc5e1d292016-04-19 09:37:59 +0200355 }
356 nbLoops++;
357 } while (BMK_clockSpan(clockStart, ticksPerSecond) < clockLoop);
inikep19bd48f2016-04-04 12:10:00 +0200358 { U64 const clockSpan = BMK_clockSpan(clockStart, ticksPerSecond);
inikep6d157f12016-04-15 16:54:11 +0200359 if (clockSpan < fastestD*nbLoops) fastestD = clockSpan / nbLoops;
Yann Collet7b51a292016-01-26 15:58:49 +0100360 } }
361
inikep2872b6f2016-03-22 14:38:34 +0100362 DISPLAYLEVEL(2, "%2i-%-17.17s :%10u ->%10u (%5.3f),%6.1f MB/s ,%6.1f MB/s\r",
Yann Colletde406ee2016-03-20 15:46:10 +0100363 testNb, displayName, (U32)srcSize, (U32)cSize, ratio,
inikep06f793a2016-03-29 11:17:58 +0200364 (double)srcSize / fastestC,
365 (double)srcSize / fastestD );
Yann Collet4856a002015-01-24 01:58:16 +0100366
367 /* CRC Checking */
inikep19bd48f2016-04-04 12:10:00 +0200368 { crcCheck = XXH64(resultBuffer, srcSize, 0);
Yann Colleta5b66e32016-03-26 01:48:27 +0100369 if (crcOrig!=crcCheck) {
370 size_t u;
371 DISPLAY("!!! WARNING !!! %14s : Invalid Checksum : %x != %x \n", displayName, (unsigned)crcOrig, (unsigned)crcCheck);
372 for (u=0; u<srcSize; u++) {
373 if (((const BYTE*)srcBuffer)[u] != ((const BYTE*)resultBuffer)[u]) {
374 U32 segNb, bNb, pos;
375 size_t bacc = 0;
376 DISPLAY("Decoding error at pos %u ", (U32)u);
377 for (segNb = 0; segNb < nbBlocks; segNb++) {
378 if (bacc + blockTable[segNb].srcSize > u) break;
379 bacc += blockTable[segNb].srcSize;
380 }
381 pos = (U32)(u - bacc);
382 bNb = pos / (128 KB);
383 DISPLAY("(block %u, sub %u, pos %u) \n", segNb, bNb, pos);
384 break;
Yann Collet03a6dab2016-01-21 02:21:17 +0100385 }
Yann Colleta5b66e32016-03-26 01:48:27 +0100386 if (u==srcSize-1) { /* should never happen */
387 DISPLAY("no difference detected\n");
388 } }
389 break;
390 } } /* CRC Checking */
Yann Collete8c6bb12015-07-26 00:23:57 +0100391#endif
Yann Colletde406ee2016-03-20 15:46:10 +0100392 } /* for (testNb = 1; testNb <= (g_nbIterations + !g_nbIterations); testNb++) */
Yann Collet4856a002015-01-24 01:58:16 +0100393
inikep2872b6f2016-03-22 14:38:34 +0100394 if (crcOrig == crcCheck) {
inikep4e26bb62016-03-14 12:48:51 +0100395 result->ratio = ratio;
396 result->cSize = cSize;
Yann Collet99909862016-04-09 16:17:18 +0200397 result->cSpeed = (double)srcSize / fastestC;
inikep06f793a2016-03-29 11:17:58 +0200398 result->dSpeed = (double)srcSize / fastestD;
inikep4e26bb62016-03-14 12:48:51 +0100399 }
inikep2872b6f2016-03-22 14:38:34 +0100400 DISPLAYLEVEL(2, "%2i#\n", cLevel);
Yann Colletde406ee2016-03-20 15:46:10 +0100401 } /* Bench */
Yann Collet4856a002015-01-24 01:58:16 +0100402
Yann Colleted699e62015-12-16 02:37:24 +0100403 /* clean up */
Yann Collet4856a002015-01-24 01:58:16 +0100404 free(compressedBuffer);
405 free(resultBuffer);
Yann Collet2630a5e2016-01-14 19:13:22 +0100406 ZSTD_freeCCtx(refCtx);
Yann Collet31683c02015-12-18 01:26:48 +0100407 ZSTD_freeCCtx(ctx);
Yann Collet7b51a292016-01-26 15:58:49 +0100408 ZSTD_freeDCtx(refDCtx);
Yann Collet31683c02015-12-18 01:26:48 +0100409 ZSTD_freeDCtx(dctx);
Yann Collet4856a002015-01-24 01:58:16 +0100410 return 0;
411}
412
413
Yann Collet4856a002015-01-24 01:58:16 +0100414static size_t BMK_findMaxMem(U64 requiredMem)
415{
Yann Colletde406ee2016-03-20 15:46:10 +0100416 size_t const step = 64 MB;
Yann Collet4856a002015-01-24 01:58:16 +0100417 BYTE* testmem = NULL;
418
419 requiredMem = (((requiredMem >> 26) + 1) << 26);
Yann Colletde406ee2016-03-20 15:46:10 +0100420 requiredMem += step;
Yann Collet050efba2015-11-03 09:49:30 +0100421 if (requiredMem > maxMemory) requiredMem = maxMemory;
Yann Collet4856a002015-01-24 01:58:16 +0100422
Yann Colletde406ee2016-03-20 15:46:10 +0100423 do {
Yann Collet4856a002015-01-24 01:58:16 +0100424 testmem = (BYTE*)malloc((size_t)requiredMem);
Yann Colletde406ee2016-03-20 15:46:10 +0100425 requiredMem -= step;
426 } while (!testmem);
427
Yann Collet4856a002015-01-24 01:58:16 +0100428 free(testmem);
Yann Colletde406ee2016-03-20 15:46:10 +0100429 return (size_t)(requiredMem);
Yann Collet4856a002015-01-24 01:58:16 +0100430}
431
Yann Colleted699e62015-12-16 02:37:24 +0100432static void BMK_benchCLevel(void* srcBuffer, size_t benchedSize,
inikepeaba91a2016-03-23 20:30:26 +0100433 const char* displayName, int cLevel, int cLevelLast,
Yann Collet31683c02015-12-18 01:26:48 +0100434 const size_t* fileSizes, unsigned nbFiles,
435 const void* dictBuffer, size_t dictBufferSize)
Yann Collet4856a002015-01-24 01:58:16 +0100436{
inikep4e26bb62016-03-14 12:48:51 +0100437 benchResult_t result, total;
inikepe9554b72016-03-14 18:10:30 +0100438 int l;
inikep4e26bb62016-03-14 12:48:51 +0100439
inikepe9554b72016-03-14 18:10:30 +0100440 const char* pch = strrchr(displayName, '\\'); /* Windows */
441 if (!pch) pch = strrchr(displayName, '/'); /* Linux */
442 if (pch) displayName = pch+1;
inikep4e26bb62016-03-14 12:48:51 +0100443
inikepea4ee3e2016-04-25 13:09:06 +0200444 SET_HIGH_PRIORITY;
445
inikepe9554b72016-03-14 18:10:30 +0100446 memset(&result, 0, sizeof(result));
447 memset(&total, 0, sizeof(total));
inikep5fdd0b42016-03-14 19:51:11 +0100448
inikepeaba91a2016-03-23 20:30:26 +0100449 if (g_displayLevel == 1 && !g_additionalParam)
inikep2872b6f2016-03-22 14:38:34 +0100450 DISPLAY("bench %s: input %u bytes, %i iterations, %u KB blocks\n", ZSTD_VERSION, (U32)benchedSize, g_nbIterations, (U32)(g_blockSize>>10));
inikepe9554b72016-03-14 18:10:30 +0100451
452 if (cLevelLast < cLevel) cLevelLast = cLevel;
453
Yann Collet99909862016-04-09 16:17:18 +0200454 for (l=cLevel; l <= cLevelLast; l++) {
inikepe9554b72016-03-14 18:10:30 +0100455 BMK_benchMem(srcBuffer, benchedSize,
inikep472638c2016-03-23 12:28:28 +0100456 displayName, l,
inikepe9554b72016-03-14 18:10:30 +0100457 fileSizes, nbFiles,
458 dictBuffer, dictBufferSize, &result);
459 if (g_displayLevel == 1) {
inikepeaba91a2016-03-23 20:30:26 +0100460 if (g_additionalParam)
inikep38654982016-04-21 12:18:47 +0200461 DISPLAY("%-3i%11i (%5.3f) %6.2f MB/s %6.1f MB/s %s (param=%d)\n", -l, (int)result.cSize, result.ratio, result.cSpeed, result.dSpeed, displayName, g_additionalParam);
inikepd700a1a2016-03-15 12:18:44 +0100462 else
inikep38654982016-04-21 12:18:47 +0200463 DISPLAY("%-3i%11i (%5.3f) %6.2f MB/s %6.1f MB/s %s\n", -l, (int)result.cSize, result.ratio, result.cSpeed, result.dSpeed, displayName);
inikepe9554b72016-03-14 18:10:30 +0100464 total.cSize += result.cSize;
465 total.cSpeed += result.cSpeed;
466 total.dSpeed += result.dSpeed;
467 total.ratio += result.ratio;
Yann Collet99909862016-04-09 16:17:18 +0200468 } }
469 if (g_displayLevel == 1 && cLevelLast > cLevel) {
inikepe9554b72016-03-14 18:10:30 +0100470 total.cSize /= 1+cLevelLast-cLevel;
471 total.cSpeed /= 1+cLevelLast-cLevel;
472 total.dSpeed /= 1+cLevelLast-cLevel;
473 total.ratio /= 1+cLevelLast-cLevel;
inikep38654982016-04-21 12:18:47 +0200474 DISPLAY("avg%11i (%5.3f) %6.2f MB/s %6.1f MB/s %s\n", (int)total.cSize, total.ratio, total.cSpeed, total.dSpeed, displayName);
inikepe9554b72016-03-14 18:10:30 +0100475 }
Yann Colleted699e62015-12-16 02:37:24 +0100476}
477
478static U64 BMK_getTotalFileSize(const char** fileNamesTable, unsigned nbFiles)
479{
480 U64 total = 0;
481 unsigned n;
482 for (n=0; n<nbFiles; n++)
inikep69fcd7c2016-04-28 12:23:33 +0200483 total += UTIL_getFileSize(fileNamesTable[n]);
Yann Colleted699e62015-12-16 02:37:24 +0100484 return total;
485}
486
Yann Colleta5b66e32016-03-26 01:48:27 +0100487/*! BMK_loadFiles() :
488 Loads `buffer` with content of files listed within `fileNamesTable`.
489 At most, fills `buffer` entirely */
Yann Collet70611352015-12-16 03:01:03 +0100490static void BMK_loadFiles(void* buffer, size_t bufferSize,
491 size_t* fileSizes,
Yann Colleta5b66e32016-03-26 01:48:27 +0100492 const char** fileNamesTable, unsigned nbFiles)
Yann Colleted699e62015-12-16 02:37:24 +0100493{
inikepc0d5f4e2016-04-13 10:48:04 +0200494 size_t pos = 0, totalSize = 0;
inikepea4ee3e2016-04-25 13:09:06 +0200495 FILE* f;
Yann Collet699b14d2016-03-17 19:37:33 +0100496 unsigned n;
Yann Colletfd416f12016-01-30 03:14:15 +0100497 for (n=0; n<nbFiles; n++) {
inikep69fcd7c2016-04-28 12:23:33 +0200498 U64 fileSize = UTIL_getFileSize(fileNamesTable[n]);
499 if (UTIL_isDirectory(fileNamesTable[n])) {
inikepc0d5f4e2016-04-13 10:48:04 +0200500 DISPLAYLEVEL(2, "Ignoring %s directory... \n", fileNamesTable[n]);
501 continue;
502 }
inikepea4ee3e2016-04-25 13:09:06 +0200503 f = fopen(fileNamesTable[n], "rb");
Yann Colleted699e62015-12-16 02:37:24 +0100504 if (f==NULL) EXM_THROW(10, "impossible to open file %s", fileNamesTable[n]);
505 DISPLAYLEVEL(2, "Loading %s... \r", fileNamesTable[n]);
Yann Colleta5b66e32016-03-26 01:48:27 +0100506 if (fileSize > bufferSize-pos) fileSize = bufferSize-pos, nbFiles=n; /* buffer too small - stop after this file */
507 { size_t const readSize = fread(((char*)buffer)+pos, 1, (size_t)fileSize, f);
508 if (readSize != (size_t)fileSize) EXM_THROW(11, "could not read %s", fileNamesTable[n]);
509 pos += readSize; }
Yann Colleta52c98d2015-12-16 03:12:31 +0100510 fileSizes[n] = (size_t)fileSize;
inikepc0d5f4e2016-04-13 10:48:04 +0200511 totalSize += (size_t)fileSize;
Yann Colleted699e62015-12-16 02:37:24 +0100512 fclose(f);
513 }
inikepc0d5f4e2016-04-13 10:48:04 +0200514
515 if (totalSize == 0) EXM_THROW(12, "no data to bench");
Yann Colleted699e62015-12-16 02:37:24 +0100516}
517
Yann Collet31683c02015-12-18 01:26:48 +0100518static void BMK_benchFileTable(const char** fileNamesTable, unsigned nbFiles,
inikepeaba91a2016-03-23 20:30:26 +0100519 const char* dictFileName, int cLevel, int cLevelLast)
Yann Colleted699e62015-12-16 02:37:24 +0100520{
521 void* srcBuffer;
522 size_t benchedSize;
Yann Collet31683c02015-12-18 01:26:48 +0100523 void* dictBuffer = NULL;
524 size_t dictBufferSize = 0;
525 size_t* fileSizes = (size_t*)malloc(nbFiles * sizeof(size_t));
Yann Colleted699e62015-12-16 02:37:24 +0100526 U64 totalSizeToLoad = BMK_getTotalFileSize(fileNamesTable, nbFiles);
527 char mfName[20] = {0};
528 const char* displayName = NULL;
529
Yann Collet31683c02015-12-18 01:26:48 +0100530 if (!fileSizes) EXM_THROW(12, "not enough memory for fileSizes");
531
532 /* Load dictionary */
Yann Colletfd416f12016-01-30 03:14:15 +0100533 if (dictFileName != NULL) {
inikep69fcd7c2016-04-28 12:23:33 +0200534 U64 dictFileSize = UTIL_getFileSize(dictFileName);
Yann Collet31683c02015-12-18 01:26:48 +0100535 if (dictFileSize > 64 MB) EXM_THROW(10, "dictionary file %s too large", dictFileName);
536 dictBufferSize = (size_t)dictFileSize;
537 dictBuffer = malloc(dictBufferSize);
538 if (dictBuffer==NULL) EXM_THROW(11, "not enough memory for dictionary (%u bytes)", (U32)dictBufferSize);
539 BMK_loadFiles(dictBuffer, dictBufferSize, fileSizes, &dictFileName, 1);
540 }
541
Yann Colleted699e62015-12-16 02:37:24 +0100542 /* Memory allocation & restrictions */
543 benchedSize = BMK_findMaxMem(totalSizeToLoad * 3) / 3;
544 if ((U64)benchedSize > totalSizeToLoad) benchedSize = (size_t)totalSizeToLoad;
545 if (benchedSize < totalSizeToLoad)
546 DISPLAY("Not enough memory; testing %u MB only...\n", (U32)(benchedSize >> 20));
547 srcBuffer = malloc(benchedSize);
548 if (!srcBuffer) EXM_THROW(12, "not enough memory");
549
550 /* Load input buffer */
Yann Collet70611352015-12-16 03:01:03 +0100551 BMK_loadFiles(srcBuffer, benchedSize, fileSizes, fileNamesTable, nbFiles);
Yann Colleted699e62015-12-16 02:37:24 +0100552
553 /* Bench */
554 snprintf (mfName, sizeof(mfName), " %u files", nbFiles);
555 if (nbFiles > 1) displayName = mfName;
556 else displayName = fileNamesTable[0];
557
Yann Collet31683c02015-12-18 01:26:48 +0100558 BMK_benchCLevel(srcBuffer, benchedSize,
inikepeaba91a2016-03-23 20:30:26 +0100559 displayName, cLevel, cLevelLast,
Yann Collet31683c02015-12-18 01:26:48 +0100560 fileSizes, nbFiles,
561 dictBuffer, dictBufferSize);
Yann Collet4856a002015-01-24 01:58:16 +0100562
Yann Colleteeb8ba12015-10-22 16:55:40 +0100563 /* clean up */
Yann Collet4856a002015-01-24 01:58:16 +0100564 free(srcBuffer);
Yann Collet31683c02015-12-18 01:26:48 +0100565 free(dictBuffer);
Yann Collet70611352015-12-16 03:01:03 +0100566 free(fileSizes);
Yann Collet4856a002015-01-24 01:58:16 +0100567}
568
569
inikepeaba91a2016-03-23 20:30:26 +0100570static void BMK_syntheticTest(int cLevel, int cLevelLast, double compressibility)
Yann Collet4856a002015-01-24 01:58:16 +0100571{
Yann Colleted699e62015-12-16 02:37:24 +0100572 char name[20] = {0};
Yann Collet4856a002015-01-24 01:58:16 +0100573 size_t benchedSize = 10000000;
574 void* srcBuffer = malloc(benchedSize);
Yann Collet4856a002015-01-24 01:58:16 +0100575
Yann Collet4856a002015-01-24 01:58:16 +0100576 /* Memory allocation */
Yann Colleted699e62015-12-16 02:37:24 +0100577 if (!srcBuffer) EXM_THROW(21, "not enough memory");
Yann Collet4856a002015-01-24 01:58:16 +0100578
579 /* Fill input buffer */
Yann Colletd062f132015-12-01 01:31:17 +0100580 RDG_genBuffer(srcBuffer, benchedSize, compressibility, 0.0, 0);
Yann Collet4856a002015-01-24 01:58:16 +0100581
582 /* Bench */
Yann Colleted699e62015-12-16 02:37:24 +0100583 snprintf (name, sizeof(name), "Synthetic %2u%%", (unsigned)(compressibility*100));
inikepeaba91a2016-03-23 20:30:26 +0100584 BMK_benchCLevel(srcBuffer, benchedSize, name, cLevel, cLevelLast, &benchedSize, 1, NULL, 0);
Yann Collet4856a002015-01-24 01:58:16 +0100585
Yann Colleted699e62015-12-16 02:37:24 +0100586 /* clean up */
Yann Collet4856a002015-01-24 01:58:16 +0100587 free(srcBuffer);
Yann Collet4856a002015-01-24 01:58:16 +0100588}
589
590
Yann Collet31683c02015-12-18 01:26:48 +0100591int BMK_benchFiles(const char** fileNamesTable, unsigned nbFiles,
inikepeaba91a2016-03-23 20:30:26 +0100592 const char* dictFileName, int cLevel, int cLevelLast)
Yann Collet4856a002015-01-24 01:58:16 +0100593{
Yann Collet699b14d2016-03-17 19:37:33 +0100594 double const compressibility = (double)g_compressibilityDefault / 100;
Yann Collet4856a002015-01-24 01:58:16 +0100595
596 if (nbFiles == 0)
inikepeaba91a2016-03-23 20:30:26 +0100597 BMK_syntheticTest(cLevel, cLevelLast, compressibility);
Yann Collet4856a002015-01-24 01:58:16 +0100598 else
inikepeaba91a2016-03-23 20:30:26 +0100599 BMK_benchFileTable(fileNamesTable, nbFiles, dictFileName, cLevel, cLevelLast);
Yann Collet4856a002015-01-24 01:58:16 +0100600 return 0;
601}
602