summaryrefslogtreecommitdiff
path: root/programs/bench.c
diff options
context:
space:
mode:
Diffstat (limited to 'programs/bench.c')
-rw-r--r--programs/bench.c1240
1 files changed, 819 insertions, 421 deletions
diff --git a/programs/bench.c b/programs/bench.c
index 014a4fd41b1cf..326c1c1c56e59 100644
--- a/programs/bench.c
+++ b/programs/bench.c
@@ -41,6 +41,8 @@
#include "zstd.h"
#include "datagen.h" /* RDG_genBuffer */
#include "xxhash.h"
+#include "bench.h"
+#include "zstd_errors.h"
/* *************************************
@@ -61,25 +63,27 @@
#define MB *(1 <<20)
#define GB *(1U<<30)
-static const size_t maxMemory = (sizeof(size_t)==4) ? (2 GB - 64 MB) : (size_t)(1ULL << ((sizeof(size_t)*8)-31));
+#define BMK_RUNTEST_DEFAULT_MS 1000
-static U32 g_compressibilityDefault = 50;
+static const size_t maxMemory = (sizeof(size_t)==4) ?
+ /* 32-bit */ (2 GB - 64 MB) :
+ /* 64-bit */ (size_t)(1ULL << ((sizeof(size_t)*8)-31));
/* *************************************
* console display
***************************************/
#define DISPLAY(...) fprintf(stderr, __VA_ARGS__)
-#define DISPLAYLEVEL(l, ...) if (g_displayLevel>=l) { DISPLAY(__VA_ARGS__); }
-static int g_displayLevel = 2; /* 0 : no display; 1: errors; 2 : + result + interaction + warnings; 3 : + progression; 4 : + information */
+#define DISPLAYLEVEL(l, ...) if (displayLevel>=l) { DISPLAY(__VA_ARGS__); }
+/* 0 : no display; 1: errors; 2 : + result + interaction + warnings; 3 : + progression; 4 : + information */
static const U64 g_refreshRate = SEC_TO_MICRO / 6;
static UTIL_time_t g_displayClock = UTIL_TIME_INITIALIZER;
-#define DISPLAYUPDATE(l, ...) { if (g_displayLevel>=l) { \
- if ((UTIL_clockSpanMicro(g_displayClock) > g_refreshRate) || (g_displayLevel>=4)) \
+#define DISPLAYUPDATE(l, ...) { if (displayLevel>=l) { \
+ if ((UTIL_clockSpanMicro(g_displayClock) > g_refreshRate) || (displayLevel>=4)) \
{ g_displayClock = UTIL_getTime(); DISPLAY(__VA_ARGS__); \
- if (g_displayLevel>=4) fflush(stderr); } } }
+ if (displayLevel>=4) fflush(stderr); } } }
/* *************************************
@@ -89,467 +93,867 @@ static UTIL_time_t g_displayClock = UTIL_TIME_INITIALIZER;
# define DEBUG 0
#endif
#define DEBUGOUTPUT(...) { if (DEBUG) DISPLAY(__VA_ARGS__); }
-#define EXM_THROW(error, ...) { \
+
+#define EXM_THROW_INT(errorNum, ...) { \
DEBUGOUTPUT("%s: %i: \n", __FILE__, __LINE__); \
- DISPLAYLEVEL(1, "Error %i : ", error); \
+ DISPLAYLEVEL(1, "Error %i : ", errorNum); \
DISPLAYLEVEL(1, __VA_ARGS__); \
DISPLAYLEVEL(1, " \n"); \
- exit(error); \
+ return errorNum; \
}
+#define RETURN_ERROR(errorNum, retType, ...) { \
+ retType r; \
+ memset(&r, 0, sizeof(retType)); \
+ DEBUGOUTPUT("%s: %i: \n", __FILE__, __LINE__); \
+ DISPLAYLEVEL(1, "Error %i : ", errorNum); \
+ DISPLAYLEVEL(1, __VA_ARGS__); \
+ DISPLAYLEVEL(1, " \n"); \
+ r.tag = errorNum; \
+ return r; \
+}
+
+/* error without displaying */
+#define RETURN_QUIET_ERROR(errorNum, retType, ...) { \
+ retType r; \
+ memset(&r, 0, sizeof(retType)); \
+ DEBUGOUTPUT("%s: %i: \n", __FILE__, __LINE__); \
+ DEBUGOUTPUT("Error %i : ", errorNum); \
+ DEBUGOUTPUT(__VA_ARGS__); \
+ DEBUGOUTPUT(" \n"); \
+ r.tag = errorNum; \
+ return r; \
+}
/* *************************************
* Benchmark Parameters
***************************************/
-static int g_additionalParam = 0;
-static U32 g_decodeOnly = 0;
-void BMK_setNotificationLevel(unsigned level) { g_displayLevel=level; }
+BMK_advancedParams_t BMK_initAdvancedParams(void) {
+ BMK_advancedParams_t const res = {
+ BMK_both, /* mode */
+ BMK_TIMETEST_DEFAULT_S, /* nbSeconds */
+ 0, /* blockSize */
+ 0, /* nbWorkers */
+ 0, /* realTime */
+ 0, /* additionalParam */
+ 0, /* ldmFlag */
+ 0, /* ldmMinMatch */
+ 0, /* ldmHashLog */
+ 0, /* ldmBuckSizeLog */
+ 0 /* ldmHashEveryLog */
+ };
+ return res;
+}
+
+
+/* ********************************************************
+* Bench functions
+**********************************************************/
+typedef struct {
+ const void* srcPtr;
+ size_t srcSize;
+ void* cPtr;
+ size_t cRoom;
+ size_t cSize;
+ void* resPtr;
+ size_t resSize;
+} blockParam_t;
+
+#undef MIN
+#undef MAX
+#define MIN(a,b) ((a) < (b) ? (a) : (b))
+#define MAX(a,b) ((a) > (b) ? (a) : (b))
+
+static void BMK_initCCtx(ZSTD_CCtx* ctx,
+ const void* dictBuffer, size_t dictBufferSize, int cLevel,
+ const ZSTD_compressionParameters* comprParams, const BMK_advancedParams_t* adv) {
+ ZSTD_CCtx_reset(ctx);
+ ZSTD_CCtx_resetParameters(ctx);
+ if (adv->nbWorkers==1) {
+ ZSTD_CCtx_setParameter(ctx, ZSTD_p_nbWorkers, 0);
+ } else {
+ ZSTD_CCtx_setParameter(ctx, ZSTD_p_nbWorkers, adv->nbWorkers);
+ }
+ ZSTD_CCtx_setParameter(ctx, ZSTD_p_compressionLevel, cLevel);
+ ZSTD_CCtx_setParameter(ctx, ZSTD_p_enableLongDistanceMatching, adv->ldmFlag);
+ ZSTD_CCtx_setParameter(ctx, ZSTD_p_ldmMinMatch, adv->ldmMinMatch);
+ ZSTD_CCtx_setParameter(ctx, ZSTD_p_ldmHashLog, adv->ldmHashLog);
+ ZSTD_CCtx_setParameter(ctx, ZSTD_p_ldmBucketSizeLog, adv->ldmBucketSizeLog);
+ ZSTD_CCtx_setParameter(ctx, ZSTD_p_ldmHashEveryLog, adv->ldmHashEveryLog);
+ ZSTD_CCtx_setParameter(ctx, ZSTD_p_windowLog, comprParams->windowLog);
+ ZSTD_CCtx_setParameter(ctx, ZSTD_p_hashLog, comprParams->hashLog);
+ ZSTD_CCtx_setParameter(ctx, ZSTD_p_chainLog, comprParams->chainLog);
+ ZSTD_CCtx_setParameter(ctx, ZSTD_p_searchLog, comprParams->searchLog);
+ ZSTD_CCtx_setParameter(ctx, ZSTD_p_minMatch, comprParams->searchLength);
+ ZSTD_CCtx_setParameter(ctx, ZSTD_p_targetLength, comprParams->targetLength);
+ ZSTD_CCtx_setParameter(ctx, ZSTD_p_compressionStrategy, comprParams->strategy);
+ ZSTD_CCtx_loadDictionary(ctx, dictBuffer, dictBufferSize);
+}
+
+static void BMK_initDCtx(ZSTD_DCtx* dctx,
+ const void* dictBuffer, size_t dictBufferSize) {
+ ZSTD_DCtx_reset(dctx);
+ ZSTD_DCtx_loadDictionary(dctx, dictBuffer, dictBufferSize);
+}
+
-void BMK_setAdditionalParam(int additionalParam) { g_additionalParam=additionalParam; }
+typedef struct {
+ ZSTD_CCtx* cctx;
+ const void* dictBuffer;
+ size_t dictBufferSize;
+ int cLevel;
+ const ZSTD_compressionParameters* comprParams;
+ const BMK_advancedParams_t* adv;
+} BMK_initCCtxArgs;
+
+static size_t local_initCCtx(void* payload) {
+ BMK_initCCtxArgs* ag = (BMK_initCCtxArgs*)payload;
+ BMK_initCCtx(ag->cctx, ag->dictBuffer, ag->dictBufferSize, ag->cLevel, ag->comprParams, ag->adv);
+ return 0;
+}
-static U32 g_nbSeconds = BMK_TIMETEST_DEFAULT_S;
-void BMK_setNbSeconds(unsigned nbSeconds)
+typedef struct {
+ ZSTD_DCtx* dctx;
+ const void* dictBuffer;
+ size_t dictBufferSize;
+} BMK_initDCtxArgs;
+
+static size_t local_initDCtx(void* payload) {
+ BMK_initDCtxArgs* ag = (BMK_initDCtxArgs*)payload;
+ BMK_initDCtx(ag->dctx, ag->dictBuffer, ag->dictBufferSize);
+ return 0;
+}
+
+
+/* `addArgs` is the context */
+static size_t local_defaultCompress(
+ const void* srcBuffer, size_t srcSize,
+ void* dstBuffer, size_t dstSize,
+ void* addArgs)
{
- g_nbSeconds = nbSeconds;
- DISPLAYLEVEL(3, "- test >= %u seconds per compression / decompression - \n", g_nbSeconds);
+ size_t moreToFlush = 1;
+ ZSTD_CCtx* const cctx = (ZSTD_CCtx*)addArgs;
+ ZSTD_inBuffer in;
+ ZSTD_outBuffer out;
+ in.src = srcBuffer; in.size = srcSize; in.pos = 0;
+ out.dst = dstBuffer; out.size = dstSize; out.pos = 0;
+ while (moreToFlush) {
+ if(out.pos == out.size) {
+ return (size_t)-ZSTD_error_dstSize_tooSmall;
+ }
+ moreToFlush = ZSTD_compress_generic(cctx, &out, &in, ZSTD_e_end);
+ if (ZSTD_isError(moreToFlush)) {
+ return moreToFlush;
+ }
+ }
+ return out.pos;
}
-static size_t g_blockSize = 0;
-void BMK_setBlockSize(size_t blockSize)
+/* `addArgs` is the context */
+static size_t local_defaultDecompress(
+ const void* srcBuffer, size_t srcSize,
+ void* dstBuffer, size_t dstCapacity,
+ void* addArgs)
{
- g_blockSize = blockSize;
- if (g_blockSize) DISPLAYLEVEL(2, "using blocks of size %u KB \n", (U32)(blockSize>>10));
+ size_t moreToFlush = 1;
+ ZSTD_DCtx* const dctx = (ZSTD_DCtx*)addArgs;
+ ZSTD_inBuffer in;
+ ZSTD_outBuffer out;
+ in.src = srcBuffer; in.size = srcSize; in.pos = 0;
+ out.dst = dstBuffer; out.size = dstCapacity; out.pos = 0;
+ while (moreToFlush) {
+ if(out.pos == out.size) {
+ return (size_t)-ZSTD_error_dstSize_tooSmall;
+ }
+ moreToFlush = ZSTD_decompress_generic(dctx, &out, &in);
+ if (ZSTD_isError(moreToFlush)) {
+ return moreToFlush;
+ }
+ }
+ return out.pos;
+
}
-void BMK_setDecodeOnlyMode(unsigned decodeFlag) { g_decodeOnly = (decodeFlag>0); }
-static U32 g_nbWorkers = 0;
-void BMK_setNbWorkers(unsigned nbWorkers) {
-#ifndef ZSTD_MULTITHREAD
- if (nbWorkers > 0) DISPLAYLEVEL(2, "Note : multi-threading is disabled \n");
-#endif
- g_nbWorkers = nbWorkers;
+/*=== Benchmarking an arbitrary function ===*/
+
+int BMK_isSuccessful_runOutcome(BMK_runOutcome_t outcome)
+{
+ return outcome.tag == 0;
}
-static U32 g_realTime = 0;
-void BMK_setRealTime(unsigned priority) {
- g_realTime = (priority>0);
+/* warning : this function will stop program execution if outcome is invalid !
+ * check outcome validity first, using BMK_isValid_runResult() */
+BMK_runTime_t BMK_extract_runTime(BMK_runOutcome_t outcome)
+{
+ assert(outcome.tag == 0);
+ return outcome.internal_never_use_directly;
}
-static U32 g_separateFiles = 0;
-void BMK_setSeparateFiles(unsigned separate) {
- g_separateFiles = (separate>0);
+static BMK_runOutcome_t BMK_runOutcome_error(void)
+{
+ BMK_runOutcome_t b;
+ memset(&b, 0, sizeof(b));
+ b.tag = 1;
+ return b;
+}
+
+static BMK_runOutcome_t BMK_setValid_runTime(BMK_runTime_t runTime)
+{
+ BMK_runOutcome_t outcome;
+ outcome.tag = 0;
+ outcome.internal_never_use_directly = runTime;
+ return outcome;
+}
+
+
+/* initFn will be measured once, benchFn will be measured `nbLoops` times */
+/* initFn is optional, provide NULL if none */
+/* benchFn must return size_t field compliant with ZSTD_isError for error valuee */
+/* takes # of blocks and list of size & stuff for each. */
+/* can report result of benchFn for each block into blockResult. */
+/* blockResult is optional, provide NULL if this information is not required */
+/* note : time per loop could be zero if run time < timer resolution */
+BMK_runOutcome_t BMK_benchFunction(
+ BMK_benchFn_t benchFn, void* benchPayload,
+ BMK_initFn_t initFn, void* initPayload,
+ size_t blockCount,
+ const void* const * srcBlockBuffers, const size_t* srcBlockSizes,
+ void* const * dstBlockBuffers, const size_t* dstBlockCapacities,
+ size_t* blockResults,
+ unsigned nbLoops)
+{
+ size_t dstSize = 0;
+
+ if(!nbLoops) {
+ RETURN_QUIET_ERROR(2, BMK_runOutcome_t, "nbLoops must be nonzero ");
+ }
+
+ /* init */
+ { size_t i;
+ for(i = 0; i < blockCount; i++) {
+ memset(dstBlockBuffers[i], 0xE5, dstBlockCapacities[i]); /* warm up and erase result buffer */
+ }
+#if 0
+ /* based on testing these seem to lower accuracy of multiple calls of 1 nbLoops vs 1 call of multiple nbLoops
+ * (Makes former slower)
+ */
+ UTIL_sleepMilli(5); /* give processor time to other processes */
+ UTIL_waitForNextTick();
+#endif
+ }
+
+ /* benchmark */
+ { UTIL_time_t const clockStart = UTIL_getTime();
+ unsigned loopNb, blockNb;
+ if (initFn != NULL) initFn(initPayload);
+ for (loopNb = 0; loopNb < nbLoops; loopNb++) {
+ for (blockNb = 0; blockNb < blockCount; blockNb++) {
+ size_t const res = benchFn(srcBlockBuffers[blockNb], srcBlockSizes[blockNb],
+ dstBlockBuffers[blockNb], dstBlockCapacities[blockNb],
+ benchPayload);
+ if(ZSTD_isError(res)) {
+ RETURN_QUIET_ERROR(2, BMK_runOutcome_t,
+ "Function benchmark failed on block %u of size %u : %s",
+ blockNb, (U32)dstBlockCapacities[blockNb], ZSTD_getErrorName(res));
+ } else if (loopNb == 0) {
+ dstSize += res;
+ if (blockResults != NULL) blockResults[blockNb] = res;
+ } }
+ } /* for (loopNb = 0; loopNb < nbLoops; loopNb++) */
+
+ { U64 const totalTime = UTIL_clockSpanNano(clockStart);
+ BMK_runTime_t rt;
+ rt.nanoSecPerRun = totalTime / nbLoops;
+ rt.sumOfReturn = dstSize;
+ return BMK_setValid_runTime(rt);
+ } }
}
-static U32 g_ldmFlag = 0;
-void BMK_setLdmFlag(unsigned ldmFlag) {
- g_ldmFlag = ldmFlag;
+
+/* ==== Benchmarking any function, providing intermediate results ==== */
+
+struct BMK_timedFnState_s {
+ U64 timeSpent_ns;
+ U64 timeBudget_ns;
+ U64 runBudget_ns;
+ BMK_runTime_t fastestRun;
+ unsigned nbLoops;
+ UTIL_time_t coolTime;
+}; /* typedef'd to BMK_timedFnState_t within bench.h */
+
+BMK_timedFnState_t* BMK_createTimedFnState(unsigned total_ms, unsigned run_ms)
+{
+ BMK_timedFnState_t* const r = (BMK_timedFnState_t*)malloc(sizeof(*r));
+ if (r == NULL) return NULL; /* malloc() error */
+ BMK_resetTimedFnState(r, total_ms, run_ms);
+ return r;
}
-static U32 g_ldmMinMatch = 0;
-void BMK_setLdmMinMatch(unsigned ldmMinMatch) {
- g_ldmMinMatch = ldmMinMatch;
+void BMK_freeTimedFnState(BMK_timedFnState_t* state) {
+ free(state);
}
-static U32 g_ldmHashLog = 0;
-void BMK_setLdmHashLog(unsigned ldmHashLog) {
- g_ldmHashLog = ldmHashLog;
+void BMK_resetTimedFnState(BMK_timedFnState_t* timedFnState, unsigned total_ms, unsigned run_ms)
+{
+ if (!total_ms) total_ms = 1 ;
+ if (!run_ms) run_ms = 1;
+ if (run_ms > total_ms) run_ms = total_ms;
+ timedFnState->timeSpent_ns = 0;
+ timedFnState->timeBudget_ns = (U64)total_ms * TIMELOOP_NANOSEC / 1000;
+ timedFnState->runBudget_ns = (U64)run_ms * TIMELOOP_NANOSEC / 1000;
+ timedFnState->fastestRun.nanoSecPerRun = (U64)(-1LL);
+ timedFnState->fastestRun.sumOfReturn = (size_t)(-1LL);
+ timedFnState->nbLoops = 1;
+ timedFnState->coolTime = UTIL_getTime();
}
-#define BMK_LDM_PARAM_NOTSET 9999
-static U32 g_ldmBucketSizeLog = BMK_LDM_PARAM_NOTSET;
-void BMK_setLdmBucketSizeLog(unsigned ldmBucketSizeLog) {
- g_ldmBucketSizeLog = ldmBucketSizeLog;
+/* Tells if nb of seconds set in timedFnState for all runs is spent.
+ * note : this function will return 1 if BMK_benchFunctionTimed() has actually errored. */
+int BMK_isCompleted_TimedFn(const BMK_timedFnState_t* timedFnState)
+{
+ return (timedFnState->timeSpent_ns >= timedFnState->timeBudget_ns);
}
-static U32 g_ldmHashEveryLog = BMK_LDM_PARAM_NOTSET;
-void BMK_setLdmHashEveryLog(unsigned ldmHashEveryLog) {
- g_ldmHashEveryLog = ldmHashEveryLog;
+
+#define MINUSABLETIME (TIMELOOP_NANOSEC / 2) /* 0.5 seconds */
+
+BMK_runOutcome_t BMK_benchTimedFn(
+ BMK_timedFnState_t* cont,
+ BMK_benchFn_t benchFn, void* benchPayload,
+ BMK_initFn_t initFn, void* initPayload,
+ size_t blockCount,
+ const void* const* srcBlockBuffers, const size_t* srcBlockSizes,
+ void * const * dstBlockBuffers, const size_t * dstBlockCapacities,
+ size_t* blockResults)
+{
+ U64 const runBudget_ns = cont->runBudget_ns;
+ U64 const runTimeMin_ns = runBudget_ns / 2;
+ int completed = 0;
+ BMK_runTime_t bestRunTime = cont->fastestRun;
+
+ while (!completed) {
+ BMK_runOutcome_t runResult;
+
+ /* Overheat protection */
+ if (UTIL_clockSpanMicro(cont->coolTime) > ACTIVEPERIOD_MICROSEC) {
+ DEBUGOUTPUT("\rcooling down ... \r");
+ UTIL_sleep(COOLPERIOD_SEC);
+ cont->coolTime = UTIL_getTime();
+ }
+
+ /* reinitialize capacity */
+ runResult = BMK_benchFunction(benchFn, benchPayload,
+ initFn, initPayload,
+ blockCount,
+ srcBlockBuffers, srcBlockSizes,
+ dstBlockBuffers, dstBlockCapacities,
+ blockResults,
+ cont->nbLoops);
+
+ if(!BMK_isSuccessful_runOutcome(runResult)) { /* error : move out */
+ return BMK_runOutcome_error();
+ }
+
+ { BMK_runTime_t const newRunTime = BMK_extract_runTime(runResult);
+ U64 const loopDuration_ns = newRunTime.nanoSecPerRun * cont->nbLoops;
+
+ cont->timeSpent_ns += loopDuration_ns;
+
+ /* estimate nbLoops for next run to last approximately 1 second */
+ if (loopDuration_ns > (runBudget_ns / 50)) {
+ U64 const fastestRun_ns = MIN(bestRunTime.nanoSecPerRun, newRunTime.nanoSecPerRun);
+ cont->nbLoops = (U32)(runBudget_ns / fastestRun_ns) + 1;
+ } else {
+ /* previous run was too short : blindly increase workload by x multiplier */
+ const unsigned multiplier = 10;
+ assert(cont->nbLoops < ((unsigned)-1) / multiplier); /* avoid overflow */
+ cont->nbLoops *= multiplier;
+ }
+
+ if(loopDuration_ns < runTimeMin_ns) {
+ /* don't report results for which benchmark run time was too small : increased risks of rounding errors */
+ assert(completed == 0);
+ continue;
+ } else {
+ if(newRunTime.nanoSecPerRun < bestRunTime.nanoSecPerRun) {
+ bestRunTime = newRunTime;
+ }
+ completed = 1;
+ }
+ }
+ } /* while (!completed) */
+
+ return BMK_setValid_runTime(bestRunTime);
}
-/* ********************************************************
-* Bench functions
-**********************************************************/
-typedef struct {
- const void* srcPtr;
- size_t srcSize;
- void* cPtr;
- size_t cRoom;
- size_t cSize;
- void* resPtr;
- size_t resSize;
-} blockParam_t;
+/* ================================================================= */
+/* Benchmark Zstandard, mem-to-mem scenarios */
+/* ================================================================= */
+int BMK_isSuccessful_benchOutcome(BMK_benchOutcome_t outcome)
+{
+ return outcome.tag == 0;
+}
+BMK_benchResult_t BMK_extract_benchResult(BMK_benchOutcome_t outcome)
+{
+ assert(outcome.tag == 0);
+ return outcome.internal_never_use_directly;
+}
-#undef MIN
-#undef MAX
-#define MIN(a,b) ((a) < (b) ? (a) : (b))
-#define MAX(a,b) ((a) > (b) ? (a) : (b))
+static BMK_benchOutcome_t BMK_benchOutcome_error(void)
+{
+ BMK_benchOutcome_t b;
+ memset(&b, 0, sizeof(b));
+ b.tag = 1;
+ return b;
+}
-static int BMK_benchMem(const void* srcBuffer, size_t srcSize,
- const char* displayName, int cLevel,
- const size_t* fileSizes, U32 nbFiles,
- const void* dictBuffer, size_t dictBufferSize,
- const ZSTD_compressionParameters* const comprParams)
+static BMK_benchOutcome_t BMK_benchOutcome_setValidResult(BMK_benchResult_t result)
{
- size_t const blockSize = ((g_blockSize>=32 && !g_decodeOnly) ? g_blockSize : srcSize) + (!srcSize) /* avoid div by 0 */ ;
- U32 const maxNbBlocks = (U32) ((srcSize + (blockSize-1)) / blockSize) + nbFiles;
- blockParam_t* const blockTable = (blockParam_t*) malloc(maxNbBlocks * sizeof(blockParam_t));
- size_t const maxCompressedSize = ZSTD_compressBound(srcSize) + (maxNbBlocks * 1024); /* add some room for safety */
- void* const compressedBuffer = malloc(maxCompressedSize);
- void* resultBuffer = malloc(srcSize);
- ZSTD_CCtx* const ctx = ZSTD_createCCtx();
- ZSTD_DCtx* const dctx = ZSTD_createDCtx();
+ BMK_benchOutcome_t b;
+ b.tag = 0;
+ b.internal_never_use_directly = result;
+ return b;
+}
+
+
+/* benchMem with no allocation */
+static BMK_benchOutcome_t BMK_benchMemAdvancedNoAlloc(
+ const void** srcPtrs, size_t* srcSizes,
+ void** cPtrs, size_t* cCapacities, size_t* cSizes,
+ void** resPtrs, size_t* resSizes,
+ void** resultBufferPtr, void* compressedBuffer,
+ size_t maxCompressedSize,
+ BMK_timedFnState_t* timeStateCompress,
+ BMK_timedFnState_t* timeStateDecompress,
+
+ const void* srcBuffer, size_t srcSize,
+ const size_t* fileSizes, unsigned nbFiles,
+ const int cLevel, const ZSTD_compressionParameters* comprParams,
+ const void* dictBuffer, size_t dictBufferSize,
+ ZSTD_CCtx* cctx, ZSTD_DCtx* dctx,
+ int displayLevel, const char* displayName,
+ const BMK_advancedParams_t* adv)
+{
+ size_t const blockSize = ((adv->blockSize>=32 && (adv->mode != BMK_decodeOnly)) ? adv->blockSize : srcSize) + (!srcSize); /* avoid div by 0 */
+ BMK_benchResult_t benchResult;
size_t const loadedCompressedSize = srcSize;
size_t cSize = 0;
double ratio = 0.;
U32 nbBlocks;
- /* checks */
- if (!compressedBuffer || !resultBuffer || !blockTable || !ctx || !dctx)
- EXM_THROW(31, "allocation error : not enough memory");
+ assert(cctx != NULL); assert(dctx != NULL);
/* init */
- if (strlen(displayName)>17) displayName += strlen(displayName)-17; /* display last 17 characters */
- if (g_nbWorkers==1) g_nbWorkers=0; /* prefer synchronous mode */
-
- if (g_decodeOnly) { /* benchmark only decompression : source must be already compressed */
+ memset(&benchResult, 0, sizeof(benchResult));
+ if (strlen(displayName)>17) displayName += strlen(displayName) - 17; /* display last 17 characters */
+ if (adv->mode == BMK_decodeOnly) { /* benchmark only decompression : source must be already compressed */
const char* srcPtr = (const char*)srcBuffer;
U64 totalDSize64 = 0;
U32 fileNb;
for (fileNb=0; fileNb<nbFiles; fileNb++) {
U64 const fSize64 = ZSTD_findDecompressedSize(srcPtr, fileSizes[fileNb]);
- if (fSize64==0) EXM_THROW(32, "Impossible to determine original size ");
+ if (fSize64==0) RETURN_ERROR(32, BMK_benchOutcome_t, "Impossible to determine original size ");
totalDSize64 += fSize64;
srcPtr += fileSizes[fileNb];
}
{ size_t const decodedSize = (size_t)totalDSize64;
- if (totalDSize64 > decodedSize) EXM_THROW(32, "original size is too large"); /* size_t overflow */
- free(resultBuffer);
- resultBuffer = malloc(decodedSize);
- if (!resultBuffer) EXM_THROW(33, "not enough memory");
+ assert((U64)decodedSize == totalDSize64); /* check overflow */
+ free(*resultBufferPtr);
+ *resultBufferPtr = malloc(decodedSize);
+ if (!(*resultBufferPtr)) {
+ RETURN_ERROR(33, BMK_benchOutcome_t, "not enough memory");
+ }
+ if (totalDSize64 > decodedSize) { /* size_t overflow */
+ free(*resultBufferPtr);
+ RETURN_ERROR(32, BMK_benchOutcome_t, "original size is too large");
+ }
cSize = srcSize;
srcSize = decodedSize;
ratio = (double)srcSize / (double)cSize;
- } }
+ }
+ }
- /* Init blockTable data */
+ /* Init data blocks */
{ const char* srcPtr = (const char*)srcBuffer;
char* cPtr = (char*)compressedBuffer;
- char* resPtr = (char*)resultBuffer;
+ char* resPtr = (char*)(*resultBufferPtr);
U32 fileNb;
for (nbBlocks=0, fileNb=0; fileNb<nbFiles; fileNb++) {
size_t remaining = fileSizes[fileNb];
- U32 const nbBlocksforThisFile = g_decodeOnly ? 1 : (U32)((remaining + (blockSize-1)) / blockSize);
+ U32 const nbBlocksforThisFile = (adv->mode == BMK_decodeOnly) ? 1 : (U32)((remaining + (blockSize-1)) / blockSize);
U32 const blockEnd = nbBlocks + nbBlocksforThisFile;
for ( ; nbBlocks<blockEnd; nbBlocks++) {
size_t const thisBlockSize = MIN(remaining, blockSize);
- blockTable[nbBlocks].srcPtr = (const void*)srcPtr;
- blockTable[nbBlocks].srcSize = thisBlockSize;
- blockTable[nbBlocks].cPtr = (void*)cPtr;
- blockTable[nbBlocks].cRoom = g_decodeOnly ? thisBlockSize : ZSTD_compressBound(thisBlockSize);
- blockTable[nbBlocks].cSize = blockTable[nbBlocks].cRoom;
- blockTable[nbBlocks].resPtr = (void*)resPtr;
- blockTable[nbBlocks].resSize = g_decodeOnly ? (size_t) ZSTD_findDecompressedSize(srcPtr, thisBlockSize) : thisBlockSize;
+ srcPtrs[nbBlocks] = srcPtr;
+ srcSizes[nbBlocks] = thisBlockSize;
+ cPtrs[nbBlocks] = cPtr;
+ cCapacities[nbBlocks] = (adv->mode == BMK_decodeOnly) ? thisBlockSize : ZSTD_compressBound(thisBlockSize);
+ resPtrs[nbBlocks] = resPtr;
+ resSizes[nbBlocks] = (adv->mode == BMK_decodeOnly) ? (size_t) ZSTD_findDecompressedSize(srcPtr, thisBlockSize) : thisBlockSize;
srcPtr += thisBlockSize;
- cPtr += blockTable[nbBlocks].cRoom;
+ cPtr += cCapacities[nbBlocks];
resPtr += thisBlockSize;
remaining -= thisBlockSize;
- } } }
+ }
+ }
+ }
- /* warmimg up memory */
- if (g_decodeOnly) {
+ /* warmimg up `compressedBuffer` */
+ if (adv->mode == BMK_decodeOnly) {
memcpy(compressedBuffer, srcBuffer, loadedCompressedSize);
} else {
RDG_genBuffer(compressedBuffer, maxCompressedSize, 0.10, 0.50, 1);
}
/* Bench */
- { U64 fastestC = (U64)(-1LL), fastestD = (U64)(-1LL);
- U64 const crcOrig = g_decodeOnly ? 0 : XXH64(srcBuffer, srcSize, 0);
- UTIL_time_t coolTime;
- U64 const maxTime = (g_nbSeconds * TIMELOOP_NANOSEC) + 1;
- U32 nbDecodeLoops = (U32)((100 MB) / (srcSize+1)) + 1; /* initial conservative speed estimate */
- U32 nbCompressionLoops = (U32)((2 MB) / (srcSize+1)) + 1; /* initial conservative speed estimate */
- U64 totalCTime=0, totalDTime=0;
- U32 cCompleted=g_decodeOnly, dCompleted=0;
+ { U64 const crcOrig = (adv->mode == BMK_decodeOnly) ? 0 : XXH64(srcBuffer, srcSize, 0);
# define NB_MARKS 4
- const char* const marks[NB_MARKS] = { " |", " /", " =", "\\" };
+ const char* marks[NB_MARKS] = { " |", " /", " =", " \\" };
U32 markNb = 0;
+ int compressionCompleted = (adv->mode == BMK_decodeOnly);
+ int decompressionCompleted = (adv->mode == BMK_compressOnly);
+ BMK_initCCtxArgs cctxprep;
+ BMK_initDCtxArgs dctxprep;
+ cctxprep.cctx = cctx;
+ cctxprep.dictBuffer = dictBuffer;
+ cctxprep.dictBufferSize = dictBufferSize;
+ cctxprep.cLevel = cLevel;
+ cctxprep.comprParams = comprParams;
+ cctxprep.adv = adv;
+ dctxprep.dctx = dctx;
+ dctxprep.dictBuffer = dictBuffer;
+ dctxprep.dictBufferSize = dictBufferSize;
+
+ DISPLAYLEVEL(2, "\r%70s\r", ""); /* blank line */
+ DISPLAYLEVEL(2, "%2s-%-17.17s :%10u ->\r", marks[markNb], displayName, (U32)srcSize);
+
+ while (!(compressionCompleted && decompressionCompleted)) {
+
+ if (!compressionCompleted) {
+ BMK_runOutcome_t const cOutcome =
+ BMK_benchTimedFn( timeStateCompress,
+ &local_defaultCompress, cctx,
+ &local_initCCtx, &cctxprep,
+ nbBlocks,
+ srcPtrs, srcSizes,
+ cPtrs, cCapacities,
+ cSizes);
+
+ if (!BMK_isSuccessful_runOutcome(cOutcome)) {
+ return BMK_benchOutcome_error();
+ }
- coolTime = UTIL_getTime();
- DISPLAYLEVEL(2, "\r%79s\r", "");
- while (!cCompleted || !dCompleted) {
-
- /* overheat protection */
- if (UTIL_clockSpanMicro(coolTime) > ACTIVEPERIOD_MICROSEC) {
- DISPLAYLEVEL(2, "\rcooling down ... \r");
- UTIL_sleep(COOLPERIOD_SEC);
- coolTime = UTIL_getTime();
- }
-
- if (!g_decodeOnly) {
- /* Compression */
- DISPLAYLEVEL(2, "%2s-%-17.17s :%10u ->\r", marks[markNb], displayName, (U32)srcSize);
- if (!cCompleted) memset(compressedBuffer, 0xE5, maxCompressedSize); /* warm up and erase result buffer */
-
- UTIL_sleepMilli(5); /* give processor time to other processes */
- UTIL_waitForNextTick();
-
- if (!cCompleted) { /* still some time to do compression tests */
- U32 nbLoops = 0;
- UTIL_time_t const clockStart = UTIL_getTime();
- ZSTD_CCtx_setParameter(ctx, ZSTD_p_nbWorkers, g_nbWorkers);
- ZSTD_CCtx_setParameter(ctx, ZSTD_p_compressionLevel, cLevel);
- ZSTD_CCtx_setParameter(ctx, ZSTD_p_enableLongDistanceMatching, g_ldmFlag);
- ZSTD_CCtx_setParameter(ctx, ZSTD_p_ldmMinMatch, g_ldmMinMatch);
- ZSTD_CCtx_setParameter(ctx, ZSTD_p_ldmHashLog, g_ldmHashLog);
- if (g_ldmBucketSizeLog != BMK_LDM_PARAM_NOTSET) {
- ZSTD_CCtx_setParameter(ctx, ZSTD_p_ldmBucketSizeLog, g_ldmBucketSizeLog);
- }
- if (g_ldmHashEveryLog != BMK_LDM_PARAM_NOTSET) {
- ZSTD_CCtx_setParameter(ctx, ZSTD_p_ldmHashEveryLog, g_ldmHashEveryLog);
- }
- ZSTD_CCtx_setParameter(ctx, ZSTD_p_windowLog, comprParams->windowLog);
- ZSTD_CCtx_setParameter(ctx, ZSTD_p_hashLog, comprParams->hashLog);
- ZSTD_CCtx_setParameter(ctx, ZSTD_p_chainLog, comprParams->chainLog);
- ZSTD_CCtx_setParameter(ctx, ZSTD_p_searchLog, comprParams->searchLog);
- ZSTD_CCtx_setParameter(ctx, ZSTD_p_minMatch, comprParams->searchLength);
- ZSTD_CCtx_setParameter(ctx, ZSTD_p_targetLength, comprParams->targetLength);
- ZSTD_CCtx_setParameter(ctx, ZSTD_p_compressionStrategy, comprParams->strategy);
- ZSTD_CCtx_loadDictionary(ctx, dictBuffer, dictBufferSize);
-
- if (!g_nbSeconds) nbCompressionLoops=1;
- for (nbLoops=0; nbLoops<nbCompressionLoops; nbLoops++) {
- U32 blockNb;
- for (blockNb=0; blockNb<nbBlocks; blockNb++) {
-#if 0 /* direct compression function, for occasional comparison */
- ZSTD_parameters const params = ZSTD_getParams(cLevel, blockTable[blockNb].srcSize, dictBufferSize);
- blockTable[blockNb].cSize = ZSTD_compress_advanced(ctx,
- blockTable[blockNb].cPtr, blockTable[blockNb].cRoom,
- blockTable[blockNb].srcPtr, blockTable[blockNb].srcSize,
- dictBuffer, dictBufferSize,
- params);
-#else
- size_t moreToFlush = 1;
- ZSTD_outBuffer out;
- ZSTD_inBuffer in;
- in.src = blockTable[blockNb].srcPtr;
- in.size = blockTable[blockNb].srcSize;
- in.pos = 0;
- out.dst = blockTable[blockNb].cPtr;
- out.size = blockTable[blockNb].cRoom;
- out.pos = 0;
- while (moreToFlush) {
- moreToFlush = ZSTD_compress_generic(ctx,
- &out, &in, ZSTD_e_end);
- if (ZSTD_isError(moreToFlush))
- EXM_THROW(1, "ZSTD_compress_generic() error : %s",
- ZSTD_getErrorName(moreToFlush));
- }
- blockTable[blockNb].cSize = out.pos;
-#endif
- } }
- { U64 const loopDuration = UTIL_clockSpanNano(clockStart);
- if (loopDuration > 0) {
- if (loopDuration < fastestC * nbCompressionLoops)
- fastestC = loopDuration / nbCompressionLoops;
- nbCompressionLoops = (U32)(TIMELOOP_NANOSEC / fastestC) + 1;
- } else {
- assert(nbCompressionLoops < 40000000); /* avoid overflow */
- nbCompressionLoops *= 100;
- }
- totalCTime += loopDuration;
- cCompleted = (totalCTime >= maxTime); /* end compression tests */
+ { BMK_runTime_t const cResult = BMK_extract_runTime(cOutcome);
+ cSize = cResult.sumOfReturn;
+ ratio = (double)srcSize / cSize;
+ { BMK_benchResult_t newResult;
+ newResult.cSpeed = ((U64)srcSize * TIMELOOP_NANOSEC / cResult.nanoSecPerRun);
+ benchResult.cSize = cSize;
+ if (newResult.cSpeed > benchResult.cSpeed)
+ benchResult.cSpeed = newResult.cSpeed;
} }
- cSize = 0;
- { U32 blockNb; for (blockNb=0; blockNb<nbBlocks; blockNb++) cSize += blockTable[blockNb].cSize; }
- ratio = (double)srcSize / (double)cSize;
- markNb = (markNb+1) % NB_MARKS;
{ int const ratioAccuracy = (ratio < 10.) ? 3 : 2;
- double const compressionSpeed = ((double)srcSize / fastestC) * 1000;
- int const cSpeedAccuracy = (compressionSpeed < 10.) ? 2 : 1;
+ markNb = (markNb+1) % NB_MARKS;
DISPLAYLEVEL(2, "%2s-%-17.17s :%10u ->%10u (%5.*f),%6.*f MB/s\r",
- marks[markNb], displayName, (U32)srcSize, (U32)cSize,
+ marks[markNb], displayName,
+ (U32)srcSize, (U32)cSize,
ratioAccuracy, ratio,
- cSpeedAccuracy, compressionSpeed );
+ benchResult.cSpeed < (10 MB) ? 2 : 1, (double)benchResult.cSpeed / MB_UNIT);
}
- } /* if (!g_decodeOnly) */
+ compressionCompleted = BMK_isCompleted_TimedFn(timeStateCompress);
+ }
-#if 0 /* disable decompression test */
- dCompleted=1;
- (void)totalDTime; (void)fastestD; (void)crcOrig; /* unused when decompression disabled */
-#else
- /* Decompression */
- if (!dCompleted) memset(resultBuffer, 0xD6, srcSize); /* warm result buffer */
-
- UTIL_sleepMilli(5); /* give processor time to other processes */
- UTIL_waitForNextTick();
-
- if (!dCompleted) {
- U32 nbLoops = 0;
- ZSTD_DDict* const ddict = ZSTD_createDDict(dictBuffer, dictBufferSize);
- UTIL_time_t const clockStart = UTIL_getTime();
- if (!ddict) EXM_THROW(2, "ZSTD_createDDict() allocation failure");
- if (!g_nbSeconds) nbDecodeLoops = 1;
- for (nbLoops=0; nbLoops < nbDecodeLoops; nbLoops++) {
- U32 blockNb;
- for (blockNb=0; blockNb<nbBlocks; blockNb++) {
- size_t const regenSize = ZSTD_decompress_usingDDict(dctx,
- blockTable[blockNb].resPtr, blockTable[blockNb].resSize,
- blockTable[blockNb].cPtr, blockTable[blockNb].cSize,
- ddict);
- if (ZSTD_isError(regenSize)) {
- EXM_THROW(2, "ZSTD_decompress_usingDDict() failed on block %u of size %u : %s \n",
- blockNb, (U32)blockTable[blockNb].cSize, ZSTD_getErrorName(regenSize));
- }
- blockTable[blockNb].resSize = regenSize;
- } }
- ZSTD_freeDDict(ddict);
- { U64 const loopDuration = UTIL_clockSpanNano(clockStart);
- if (loopDuration > 0) {
- if (loopDuration < fastestD * nbDecodeLoops)
- fastestD = loopDuration / nbDecodeLoops;
- nbDecodeLoops = (U32)(TIMELOOP_NANOSEC / fastestD) + 1;
- } else {
- assert(nbDecodeLoops < 40000000); /* avoid overflow */
- nbDecodeLoops *= 100;
- }
- totalDTime += loopDuration;
- dCompleted = (totalDTime >= maxTime);
- } }
+ if(!decompressionCompleted) {
+ BMK_runOutcome_t const dOutcome =
+ BMK_benchTimedFn(timeStateDecompress,
+ &local_defaultDecompress, dctx,
+ &local_initDCtx, &dctxprep,
+ nbBlocks,
+ (const void *const *)cPtrs, cSizes,
+ resPtrs, resSizes,
+ NULL);
+
+ if(!BMK_isSuccessful_runOutcome(dOutcome)) {
+ return BMK_benchOutcome_error();
+ }
- markNb = (markNb+1) % NB_MARKS;
- { int const ratioAccuracy = (ratio < 10.) ? 3 : 2;
- double const compressionSpeed = ((double)srcSize / fastestC) * 1000;
- int const cSpeedAccuracy = (compressionSpeed < 10.) ? 2 : 1;
- double const decompressionSpeed = ((double)srcSize / fastestD) * 1000;
- DISPLAYLEVEL(2, "%2s-%-17.17s :%10u ->%10u (%5.*f),%6.*f MB/s ,%6.1f MB/s \r",
- marks[markNb], displayName, (U32)srcSize, (U32)cSize,
- ratioAccuracy, ratio,
- cSpeedAccuracy, compressionSpeed,
- decompressionSpeed);
- }
+ { BMK_runTime_t const dResult = BMK_extract_runTime(dOutcome);
+ U64 const newDSpeed = (srcSize * TIMELOOP_NANOSEC / dResult.nanoSecPerRun);
+ if (newDSpeed > benchResult.dSpeed)
+ benchResult.dSpeed = newDSpeed;
+ }
- /* CRC Checking */
- { U64 const crcCheck = XXH64(resultBuffer, srcSize, 0);
- if (!g_decodeOnly && (crcOrig!=crcCheck)) {
- size_t u;
- DISPLAY("!!! WARNING !!! %14s : Invalid Checksum : %x != %x \n", displayName, (unsigned)crcOrig, (unsigned)crcCheck);
- for (u=0; u<srcSize; u++) {
- if (((const BYTE*)srcBuffer)[u] != ((const BYTE*)resultBuffer)[u]) {
- U32 segNb, bNb, pos;
- size_t bacc = 0;
- DISPLAY("Decoding error at pos %u ", (U32)u);
- for (segNb = 0; segNb < nbBlocks; segNb++) {
- if (bacc + blockTable[segNb].srcSize > u) break;
- bacc += blockTable[segNb].srcSize;
- }
- pos = (U32)(u - bacc);
- bNb = pos / (128 KB);
- DISPLAY("(sample %u, block %u, pos %u) \n", segNb, bNb, pos);
- if (u>5) {
- int n;
- DISPLAY("origin: ");
- for (n=-5; n<0; n++) DISPLAY("%02X ", ((const BYTE*)srcBuffer)[u+n]);
- DISPLAY(" :%02X: ", ((const BYTE*)srcBuffer)[u]);
- for (n=1; n<3; n++) DISPLAY("%02X ", ((const BYTE*)srcBuffer)[u+n]);
- DISPLAY(" \n");
- DISPLAY("decode: ");
- for (n=-5; n<0; n++) DISPLAY("%02X ", ((const BYTE*)resultBuffer)[u+n]);
- DISPLAY(" :%02X: ", ((const BYTE*)resultBuffer)[u]);
- for (n=1; n<3; n++) DISPLAY("%02X ", ((const BYTE*)resultBuffer)[u+n]);
- DISPLAY(" \n");
- }
- break;
+ { int const ratioAccuracy = (ratio < 10.) ? 3 : 2;
+ markNb = (markNb+1) % NB_MARKS;
+ DISPLAYLEVEL(2, "%2s-%-17.17s :%10u ->%10u (%5.*f),%6.*f MB/s ,%6.1f MB/s \r",
+ marks[markNb], displayName,
+ (U32)srcSize, (U32)benchResult.cSize,
+ ratioAccuracy, ratio,
+ benchResult.cSpeed < (10 MB) ? 2 : 1, (double)benchResult.cSpeed / MB_UNIT,
+ (double)benchResult.dSpeed / MB_UNIT);
+ }
+ decompressionCompleted = BMK_isCompleted_TimedFn(timeStateDecompress);
+ }
+ } /* while (!(compressionCompleted && decompressionCompleted)) */
+
+ /* CRC Checking */
+ { const BYTE* resultBuffer = (const BYTE*)(*resultBufferPtr);
+ U64 const crcCheck = XXH64(resultBuffer, srcSize, 0);
+ if ((adv->mode == BMK_both) && (crcOrig!=crcCheck)) {
+ size_t u;
+ DISPLAY("!!! WARNING !!! %14s : Invalid Checksum : %x != %x \n", displayName, (unsigned)crcOrig, (unsigned)crcCheck);
+ for (u=0; u<srcSize; u++) {
+ if (((const BYTE*)srcBuffer)[u] != resultBuffer[u]) {
+ U32 segNb, bNb, pos;
+ size_t bacc = 0;
+ DISPLAY("Decoding error at pos %u ", (U32)u);
+ for (segNb = 0; segNb < nbBlocks; segNb++) {
+ if (bacc + srcSizes[segNb] > u) break;
+ bacc += srcSizes[segNb];
}
- if (u==srcSize-1) { /* should never happen */
- DISPLAY("no difference detected\n");
- } }
- break;
- } } /* CRC Checking */
-#endif
- } /* for (testNb = 1; testNb <= (g_nbSeconds + !g_nbSeconds); testNb++) */
-
- if (g_displayLevel == 1) { /* hidden display mode -q, used by python speed benchmark */
- double const cSpeed = ((double)srcSize / fastestC) * 1000;
- double const dSpeed = ((double)srcSize / fastestD) * 1000;
- if (g_additionalParam)
- DISPLAY("-%-3i%11i (%5.3f) %6.2f MB/s %6.1f MB/s %s (param=%d)\n", cLevel, (int)cSize, ratio, cSpeed, dSpeed, displayName, g_additionalParam);
- else
+ pos = (U32)(u - bacc);
+ bNb = pos / (128 KB);
+ DISPLAY("(sample %u, block %u, pos %u) \n", segNb, bNb, pos);
+ if (u>5) {
+ int n;
+ DISPLAY("origin: ");
+ for (n=-5; n<0; n++) DISPLAY("%02X ", ((const BYTE*)srcBuffer)[u+n]);
+ DISPLAY(" :%02X: ", ((const BYTE*)srcBuffer)[u]);
+ for (n=1; n<3; n++) DISPLAY("%02X ", ((const BYTE*)srcBuffer)[u+n]);
+ DISPLAY(" \n");
+ DISPLAY("decode: ");
+ for (n=-5; n<0; n++) DISPLAY("%02X ", resultBuffer[u+n]);
+ DISPLAY(" :%02X: ", resultBuffer[u]);
+ for (n=1; n<3; n++) DISPLAY("%02X ", resultBuffer[u+n]);
+ DISPLAY(" \n");
+ }
+ break;
+ }
+ if (u==srcSize-1) { /* should never happen */
+ DISPLAY("no difference detected\n");
+ }
+ }
+ }
+ } /* CRC Checking */
+
+ if (displayLevel == 1) { /* hidden display mode -q, used by python speed benchmark */
+ double const cSpeed = (double)benchResult.cSpeed / MB_UNIT;
+ double const dSpeed = (double)benchResult.dSpeed / MB_UNIT;
+ if (adv->additionalParam) {
+ DISPLAY("-%-3i%11i (%5.3f) %6.2f MB/s %6.1f MB/s %s (param=%d)\n", cLevel, (int)cSize, ratio, cSpeed, dSpeed, displayName, adv->additionalParam);
+ } else {
DISPLAY("-%-3i%11i (%5.3f) %6.2f MB/s %6.1f MB/s %s\n", cLevel, (int)cSize, ratio, cSpeed, dSpeed, displayName);
+ }
}
+
DISPLAYLEVEL(2, "%2i#\n", cLevel);
} /* Bench */
- /* clean up */
- free(blockTable);
- free(compressedBuffer);
- free(resultBuffer);
- ZSTD_freeCCtx(ctx);
- ZSTD_freeDCtx(dctx);
- return 0;
+ benchResult.cMem = (1ULL << (comprParams->windowLog)) + ZSTD_sizeof_CCtx(cctx);
+ return BMK_benchOutcome_setValidResult(benchResult);
}
+BMK_benchOutcome_t BMK_benchMemAdvanced(const void* srcBuffer, size_t srcSize,
+ void* dstBuffer, size_t dstCapacity,
+ const size_t* fileSizes, unsigned nbFiles,
+ int cLevel, const ZSTD_compressionParameters* comprParams,
+ const void* dictBuffer, size_t dictBufferSize,
+ int displayLevel, const char* displayName, const BMK_advancedParams_t* adv)
-static size_t BMK_findMaxMem(U64 requiredMem)
{
- size_t const step = 64 MB;
- BYTE* testmem = NULL;
+ int const dstParamsError = !dstBuffer ^ !dstCapacity; /* must be both NULL or none */
- requiredMem = (((requiredMem >> 26) + 1) << 26);
- requiredMem += step;
- if (requiredMem > maxMemory) requiredMem = maxMemory;
+ size_t const blockSize = ((adv->blockSize>=32 && (adv->mode != BMK_decodeOnly)) ? adv->blockSize : srcSize) + (!srcSize) /* avoid div by 0 */ ;
+ U32 const maxNbBlocks = (U32) ((srcSize + (blockSize-1)) / blockSize) + nbFiles;
- do {
- testmem = (BYTE*)malloc((size_t)requiredMem);
- requiredMem -= step;
- } while (!testmem);
+ /* these are the blockTable parameters, just split up */
+ const void ** const srcPtrs = (const void**)malloc(maxNbBlocks * sizeof(void*));
+ size_t* const srcSizes = (size_t*)malloc(maxNbBlocks * sizeof(size_t));
- free(testmem);
- return (size_t)(requiredMem);
+
+ void ** const cPtrs = (void**)malloc(maxNbBlocks * sizeof(void*));
+ size_t* const cSizes = (size_t*)malloc(maxNbBlocks * sizeof(size_t));
+ size_t* const cCapacities = (size_t*)malloc(maxNbBlocks * sizeof(size_t));
+
+ void ** const resPtrs = (void**)malloc(maxNbBlocks * sizeof(void*));
+ size_t* const resSizes = (size_t*)malloc(maxNbBlocks * sizeof(size_t));
+
+ BMK_timedFnState_t* timeStateCompress = BMK_createTimedFnState(adv->nbSeconds * 1000, BMK_RUNTEST_DEFAULT_MS);
+ BMK_timedFnState_t* timeStateDecompress = BMK_createTimedFnState(adv->nbSeconds * 1000, BMK_RUNTEST_DEFAULT_MS);
+
+ ZSTD_CCtx* const cctx = ZSTD_createCCtx();
+ ZSTD_DCtx* const dctx = ZSTD_createDCtx();
+
+ const size_t maxCompressedSize = dstCapacity ? dstCapacity : ZSTD_compressBound(srcSize) + (maxNbBlocks * 1024);
+
+ void* const internalDstBuffer = dstBuffer ? NULL : malloc(maxCompressedSize);
+ void* const compressedBuffer = dstBuffer ? dstBuffer : internalDstBuffer;
+
+ BMK_benchOutcome_t outcome = BMK_benchOutcome_error(); /* error by default */
+
+ void* resultBuffer = srcSize ? malloc(srcSize) : NULL;
+
+ int allocationincomplete = !srcPtrs || !srcSizes || !cPtrs ||
+ !cSizes || !cCapacities || !resPtrs || !resSizes ||
+ !timeStateCompress || !timeStateDecompress ||
+ !cctx || !dctx ||
+ !compressedBuffer || !resultBuffer;
+
+
+ if (!allocationincomplete && !dstParamsError) {
+ outcome = BMK_benchMemAdvancedNoAlloc(srcPtrs, srcSizes,
+ cPtrs, cCapacities, cSizes,
+ resPtrs, resSizes,
+ &resultBuffer,
+ compressedBuffer, maxCompressedSize,
+ timeStateCompress, timeStateDecompress,
+ srcBuffer, srcSize,
+ fileSizes, nbFiles,
+ cLevel, comprParams,
+ dictBuffer, dictBufferSize,
+ cctx, dctx,
+ displayLevel, displayName, adv);
+ }
+
+ /* clean up */
+ BMK_freeTimedFnState(timeStateCompress);
+ BMK_freeTimedFnState(timeStateDecompress);
+
+ ZSTD_freeCCtx(cctx);
+ ZSTD_freeDCtx(dctx);
+
+ free(internalDstBuffer);
+ free(resultBuffer);
+
+ free((void*)srcPtrs);
+ free(srcSizes);
+ free(cPtrs);
+ free(cSizes);
+ free(cCapacities);
+ free(resPtrs);
+ free(resSizes);
+
+ if(allocationincomplete) {
+ RETURN_ERROR(31, BMK_benchOutcome_t, "allocation error : not enough memory");
+ }
+
+ if(dstParamsError) {
+ RETURN_ERROR(32, BMK_benchOutcome_t, "Dst parameters not coherent");
+ }
+ return outcome;
+}
+
+BMK_benchOutcome_t BMK_benchMem(const void* srcBuffer, size_t srcSize,
+ const size_t* fileSizes, unsigned nbFiles,
+ int cLevel, const ZSTD_compressionParameters* comprParams,
+ const void* dictBuffer, size_t dictBufferSize,
+ int displayLevel, const char* displayName) {
+
+ BMK_advancedParams_t const adv = BMK_initAdvancedParams();
+ return BMK_benchMemAdvanced(srcBuffer, srcSize,
+ NULL, 0,
+ fileSizes, nbFiles,
+ cLevel, comprParams,
+ dictBuffer, dictBufferSize,
+ displayLevel, displayName, &adv);
}
-static void BMK_benchCLevel(const void* srcBuffer, size_t benchedSize,
- const char* displayName, int cLevel, int cLevelLast,
+static BMK_benchOutcome_t BMK_benchCLevel(const void* srcBuffer, size_t benchedSize,
const size_t* fileSizes, unsigned nbFiles,
+ int cLevel, const ZSTD_compressionParameters* comprParams,
const void* dictBuffer, size_t dictBufferSize,
- const ZSTD_compressionParameters* const compressionParams)
+ int displayLevel, const char* displayName,
+ BMK_advancedParams_t const * const adv)
{
- int l;
-
const char* pch = strrchr(displayName, '\\'); /* Windows */
- if (!pch) pch = strrchr(displayName, '/'); /* Linux */
+ if (!pch) pch = strrchr(displayName, '/'); /* Linux */
if (pch) displayName = pch+1;
- if (g_realTime) {
+ if (adv->realTime) {
DISPLAYLEVEL(2, "Note : switching to real-time priority \n");
SET_REALTIME_PRIORITY;
}
- if (g_displayLevel == 1 && !g_additionalParam)
- DISPLAY("bench %s %s: input %u bytes, %u seconds, %u KB blocks\n", ZSTD_VERSION_STRING, ZSTD_GIT_COMMIT_STRING, (U32)benchedSize, g_nbSeconds, (U32)(g_blockSize>>10));
+ if (displayLevel == 1 && !adv->additionalParam) /* --quiet mode */
+ DISPLAY("bench %s %s: input %u bytes, %u seconds, %u KB blocks\n",
+ ZSTD_VERSION_STRING, ZSTD_GIT_COMMIT_STRING,
+ (U32)benchedSize, adv->nbSeconds, (U32)(adv->blockSize>>10));
+
+ return BMK_benchMemAdvanced(srcBuffer, benchedSize,
+ NULL, 0,
+ fileSizes, nbFiles,
+ cLevel, comprParams,
+ dictBuffer, dictBufferSize,
+ displayLevel, displayName, adv);
+}
- for (l=cLevel; l <= cLevelLast; l++) {
- if (l==0) continue; /* skip level 0 */
- BMK_benchMem(srcBuffer, benchedSize,
- displayName, l,
- fileSizes, nbFiles,
- dictBuffer, dictBufferSize, compressionParams);
+BMK_benchOutcome_t BMK_syntheticTest(int cLevel, double compressibility,
+ const ZSTD_compressionParameters* compressionParams,
+ int displayLevel, const BMK_advancedParams_t* adv)
+{
+ char name[20] = {0};
+ size_t const benchedSize = 10000000;
+ void* srcBuffer;
+ BMK_benchOutcome_t res;
+
+ if (cLevel > ZSTD_maxCLevel()) {
+ RETURN_ERROR(15, BMK_benchOutcome_t, "Invalid Compression Level");
}
+
+ /* Memory allocation */
+ srcBuffer = malloc(benchedSize);
+ if (!srcBuffer) RETURN_ERROR(21, BMK_benchOutcome_t, "not enough memory");
+
+ /* Fill input buffer */
+ RDG_genBuffer(srcBuffer, benchedSize, compressibility, 0.0, 0);
+
+ /* Bench */
+ snprintf (name, sizeof(name), "Synthetic %2u%%", (unsigned)(compressibility*100));
+ res = BMK_benchCLevel(srcBuffer, benchedSize,
+ &benchedSize /* ? */, 1 /* ? */,
+ cLevel, compressionParams,
+ NULL, 0, /* dictionary */
+ displayLevel, name, adv);
+
+ /* clean up */
+ free(srcBuffer);
+
+ return res;
}
+
+static size_t BMK_findMaxMem(U64 requiredMem)
+{
+ size_t const step = 64 MB;
+ BYTE* testmem = NULL;
+
+ requiredMem = (((requiredMem >> 26) + 1) << 26);
+ requiredMem += step;
+ if (requiredMem > maxMemory) requiredMem = maxMemory;
+
+ do {
+ testmem = (BYTE*)malloc((size_t)requiredMem);
+ requiredMem -= step;
+ } while (!testmem && requiredMem > 0);
+
+ free(testmem);
+ return (size_t)(requiredMem);
+}
+
/*! BMK_loadFiles() :
* Loads `buffer` with content of files listed within `fileNamesTable`.
* At most, fills `buffer` entirely. */
-static void BMK_loadFiles(void* buffer, size_t bufferSize,
- size_t* fileSizes,
- const char* const * const fileNamesTable, unsigned nbFiles)
+static int BMK_loadFiles(void* buffer, size_t bufferSize,
+ size_t* fileSizes,
+ const char* const * fileNamesTable, unsigned nbFiles,
+ int displayLevel)
{
size_t pos = 0, totalSize = 0;
unsigned n;
@@ -567,45 +971,69 @@ static void BMK_loadFiles(void* buffer, size_t bufferSize,
continue;
}
f = fopen(fileNamesTable[n], "rb");
- if (f==NULL) EXM_THROW(10, "impossible to open file %s", fileNamesTable[n]);
+ if (f==NULL) EXM_THROW_INT(10, "impossible to open file %s", fileNamesTable[n]);
DISPLAYUPDATE(2, "Loading %s... \r", fileNamesTable[n]);
if (fileSize > bufferSize-pos) fileSize = bufferSize-pos, nbFiles=n; /* buffer too small - stop after this file */
- { size_t const readSize = fread(((char*)buffer)+pos, 1, (size_t)fileSize, f);
- if (readSize != (size_t)fileSize) EXM_THROW(11, "could not read %s", fileNamesTable[n]);
- pos += readSize; }
+ { size_t const readSize = fread(((char*)buffer)+pos, 1, (size_t)fileSize, f);
+ if (readSize != (size_t)fileSize) EXM_THROW_INT(11, "could not read %s", fileNamesTable[n]);
+ pos += readSize;
+ }
fileSizes[n] = (size_t)fileSize;
totalSize += (size_t)fileSize;
fclose(f);
}
- if (totalSize == 0) EXM_THROW(12, "no data to bench");
+ if (totalSize == 0) EXM_THROW_INT(12, "no data to bench");
+ return 0;
}
-static void BMK_benchFileTable(const char* const * const fileNamesTable, unsigned const nbFiles,
- const char* const dictFileName,
- int const cLevel, int const cLevelLast,
- const ZSTD_compressionParameters* const compressionParams)
+BMK_benchOutcome_t BMK_benchFilesAdvanced(
+ const char* const * fileNamesTable, unsigned nbFiles,
+ const char* dictFileName, int cLevel,
+ const ZSTD_compressionParameters* compressionParams,
+ int displayLevel, const BMK_advancedParams_t* adv)
{
- void* srcBuffer;
+ void* srcBuffer = NULL;
size_t benchedSize;
void* dictBuffer = NULL;
size_t dictBufferSize = 0;
- size_t* const fileSizes = (size_t*)malloc(nbFiles * sizeof(size_t));
+ size_t* fileSizes = NULL;
+ BMK_benchOutcome_t res;
U64 const totalSizeToLoad = UTIL_getTotalFileSize(fileNamesTable, nbFiles);
- if (!fileSizes) EXM_THROW(12, "not enough memory for fileSizes");
+ if (!nbFiles) {
+ RETURN_ERROR(14, BMK_benchOutcome_t, "No Files to Benchmark");
+ }
+
+ if (cLevel > ZSTD_maxCLevel()) {
+ RETURN_ERROR(15, BMK_benchOutcome_t, "Invalid Compression Level");
+ }
+
+ fileSizes = (size_t*)calloc(nbFiles, sizeof(size_t));
+ if (!fileSizes) RETURN_ERROR(12, BMK_benchOutcome_t, "not enough memory for fileSizes");
/* Load dictionary */
if (dictFileName != NULL) {
U64 const dictFileSize = UTIL_getFileSize(dictFileName);
- if (dictFileSize > 64 MB)
- EXM_THROW(10, "dictionary file %s too large", dictFileName);
+ if (dictFileSize > 64 MB) {
+ free(fileSizes);
+ RETURN_ERROR(10, BMK_benchOutcome_t, "dictionary file %s too large", dictFileName);
+ }
dictBufferSize = (size_t)dictFileSize;
dictBuffer = malloc(dictBufferSize);
- if (dictBuffer==NULL)
- EXM_THROW(11, "not enough memory for dictionary (%u bytes)",
+ if (dictBuffer==NULL) {
+ free(fileSizes);
+ RETURN_ERROR(11, BMK_benchOutcome_t, "not enough memory for dictionary (%u bytes)",
(U32)dictBufferSize);
- BMK_loadFiles(dictBuffer, dictBufferSize, fileSizes, &dictFileName, 1);
+ }
+
+ { int const errorCode = BMK_loadFiles(dictBuffer, dictBufferSize,
+ fileSizes, &dictFileName /*?*/,
+ 1 /*?*/, displayLevel);
+ if (errorCode) {
+ res = BMK_benchOutcome_error();
+ goto _cleanUp;
+ } }
}
/* Memory allocation & restrictions */
@@ -613,79 +1041,49 @@ static void BMK_benchFileTable(const char* const * const fileNamesTable, unsigne
if ((U64)benchedSize > totalSizeToLoad) benchedSize = (size_t)totalSizeToLoad;
if (benchedSize < totalSizeToLoad)
DISPLAY("Not enough memory; testing %u MB only...\n", (U32)(benchedSize >> 20));
- srcBuffer = malloc(benchedSize);
- if (!srcBuffer) EXM_THROW(12, "not enough memory");
+
+ srcBuffer = benchedSize ? malloc(benchedSize) : NULL;
+ if (!srcBuffer) {
+ free(dictBuffer);
+ free(fileSizes);
+ RETURN_ERROR(12, BMK_benchOutcome_t, "not enough memory");
+ }
/* Load input buffer */
- BMK_loadFiles(srcBuffer, benchedSize, fileSizes, fileNamesTable, nbFiles);
+ { int const errorCode = BMK_loadFiles(srcBuffer, benchedSize,
+ fileSizes, fileNamesTable, nbFiles,
+ displayLevel);
+ if (errorCode) {
+ res = BMK_benchOutcome_error();
+ goto _cleanUp;
+ } }
/* Bench */
- if (g_separateFiles) {
- const BYTE* srcPtr = (const BYTE*)srcBuffer;
- U32 fileNb;
- for (fileNb=0; fileNb<nbFiles; fileNb++) {
- size_t const fileSize = fileSizes[fileNb];
- BMK_benchCLevel(srcPtr, fileSize,
- fileNamesTable[fileNb], cLevel, cLevelLast,
- fileSizes+fileNb, 1,
- dictBuffer, dictBufferSize, compressionParams);
- srcPtr += fileSize;
- }
- } else {
- char mfName[20] = {0};
+ { char mfName[20] = {0};
snprintf (mfName, sizeof(mfName), " %u files", nbFiles);
{ const char* const displayName = (nbFiles > 1) ? mfName : fileNamesTable[0];
- BMK_benchCLevel(srcBuffer, benchedSize,
- displayName, cLevel, cLevelLast,
- fileSizes, nbFiles,
- dictBuffer, dictBufferSize, compressionParams);
+ res = BMK_benchCLevel(srcBuffer, benchedSize,
+ fileSizes, nbFiles,
+ cLevel, compressionParams,
+ dictBuffer, dictBufferSize,
+ displayLevel, displayName,
+ adv);
} }
- /* clean up */
+_cleanUp:
free(srcBuffer);
free(dictBuffer);
free(fileSizes);
+ return res;
}
-static void BMK_syntheticTest(int cLevel, int cLevelLast, double compressibility,
- const ZSTD_compressionParameters* compressionParams)
+BMK_benchOutcome_t BMK_benchFiles(
+ const char* const * fileNamesTable, unsigned nbFiles,
+ const char* dictFileName,
+ int cLevel, const ZSTD_compressionParameters* compressionParams,
+ int displayLevel)
{
- char name[20] = {0};
- size_t benchedSize = 10000000;
- void* const srcBuffer = malloc(benchedSize);
-
- /* Memory allocation */
- if (!srcBuffer) EXM_THROW(21, "not enough memory");
-
- /* Fill input buffer */
- RDG_genBuffer(srcBuffer, benchedSize, compressibility, 0.0, 0);
-
- /* Bench */
- snprintf (name, sizeof(name), "Synthetic %2u%%", (unsigned)(compressibility*100));
- BMK_benchCLevel(srcBuffer, benchedSize, name, cLevel, cLevelLast, &benchedSize, 1, NULL, 0, compressionParams);
-
- /* clean up */
- free(srcBuffer);
-}
-
-
-int BMK_benchFiles(const char** fileNamesTable, unsigned nbFiles,
- const char* dictFileName,
- int cLevel, int cLevelLast,
- const ZSTD_compressionParameters* compressionParams)
-{
- double const compressibility = (double)g_compressibilityDefault / 100;
-
- if (cLevel > ZSTD_maxCLevel()) cLevel = ZSTD_maxCLevel();
- if (cLevelLast > ZSTD_maxCLevel()) cLevelLast = ZSTD_maxCLevel();
- if (cLevelLast < cLevel) cLevelLast = cLevel;
- if (cLevelLast > cLevel)
- DISPLAYLEVEL(2, "Benchmarking levels from %d to %d\n", cLevel, cLevelLast);
-
- if (nbFiles == 0)
- BMK_syntheticTest(cLevel, cLevelLast, compressibility, compressionParams);
- else
- BMK_benchFileTable(fileNamesTable, nbFiles, dictFileName, cLevel, cLevelLast, compressionParams);
- return 0;
+ BMK_advancedParams_t const adv = BMK_initAdvancedParams();
+ return BMK_benchFilesAdvanced(fileNamesTable, nbFiles, dictFileName, cLevel, compressionParams, displayLevel, &adv);
}
generated by cgit v1.2.3 (git 2.25.1) at 2025-08-02 00:51:52 +0000