/* * Copyright (c) 2016-present, Yann Collet, Facebook, Inc. * All rights reserved. * * This source code is licensed under both the BSD-style license (found in the * LICENSE file in the root directory of this source tree) and the GPLv2 (found * in the COPYING file in the root directory of this source tree). * You may select, at your option, one of the above-listed licenses. */ /* ************************************* * Includes ***************************************/ #include "platform.h" /* Large Files support */ #include "util.h" /* UTIL_getFileSize, UTIL_sleep */ #include /* malloc, free */ #include /* memset */ #include /* fprintf, fopen */ #undef NDEBUG /* assert must not be disabled */ #include /* assert */ #include "mem.h" #include "benchfn.h" /* ************************************* * Constants ***************************************/ #define TIMELOOP_MICROSEC (1*1000000ULL) /* 1 second */ #define TIMELOOP_NANOSEC (1*1000000000ULL) /* 1 second */ #define ACTIVEPERIOD_MICROSEC (70*TIMELOOP_MICROSEC) /* 70 seconds */ #define COOLPERIOD_SEC 10 #define KB *(1 <<10) #define MB *(1 <<20) #define GB *(1U<<30) /* ************************************* * Errors ***************************************/ #ifndef DEBUG # define DEBUG 0 #endif #define DISPLAY(...) fprintf(stderr, __VA_ARGS__) #define DEBUGOUTPUT(...) { if (DEBUG) DISPLAY(__VA_ARGS__); } /* error without displaying */ #define RETURN_QUIET_ERROR(retValue, ...) { \ DEBUGOUTPUT("%s: %i: \n", __FILE__, __LINE__); \ DEBUGOUTPUT("Error : "); \ DEBUGOUTPUT(__VA_ARGS__); \ DEBUGOUTPUT(" \n"); \ return retValue; \ } /* ************************************* * Benchmarking an arbitrary function ***************************************/ int BMK_isSuccessful_runOutcome(BMK_runOutcome_t outcome) { return outcome.error_tag_never_ever_use_directly == 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.error_tag_never_ever_use_directly == 0); return outcome.internal_never_ever_use_directly; } size_t BMK_extract_errorResult(BMK_runOutcome_t outcome) { assert(outcome.error_tag_never_ever_use_directly != 0); return outcome.error_result_never_ever_use_directly; } static BMK_runOutcome_t BMK_runOutcome_error(size_t errorResult) { BMK_runOutcome_t b; memset(&b, 0, sizeof(b)); b.error_tag_never_ever_use_directly = 1; b.error_result_never_ever_use_directly = errorResult; return b; } static BMK_runOutcome_t BMK_setValid_runTime(BMK_runTime_t runTime) { BMK_runOutcome_t outcome; outcome.error_tag_never_ever_use_directly = 0; outcome.internal_never_ever_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 a size_t value that errorFn can interpret */ /* 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 can be reported as zero if run time < timer resolution */ BMK_runOutcome_t BMK_benchFunction(BMK_benchParams_t p, unsigned nbLoops) { size_t dstSize = 0; nbLoops += !nbLoops; /* minimum nbLoops is 1 */ /* init */ { size_t i; for(i = 0; i < p.blockCount; i++) { memset(p.dstBuffers[i], 0xE5, p.dstCapacities[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 (p.initFn != NULL) p.initFn(p.initPayload); for (loopNb = 0; loopNb < nbLoops; loopNb++) { for (blockNb = 0; blockNb < p.blockCount; blockNb++) { size_t const res = p.benchFn(p.srcBuffers[blockNb], p.srcSizes[blockNb], p.dstBuffers[blockNb], p.dstCapacities[blockNb], p.benchPayload); if (loopNb == 0) { if (p.blockResults != NULL) p.blockResults[blockNb] = res; if ((p.errorFn != NULL) && (p.errorFn(res))) { RETURN_QUIET_ERROR(BMK_runOutcome_error(res), "Function benchmark failed on block %u (of size %u) with error %i", blockNb, (U32)p.srcSizes[blockNb], (int)res); } dstSize += 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); } } } /* ==== 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; } void BMK_freeTimedFnState(BMK_timedFnState_t* state) { free(state); } 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(); } /* 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); } #undef MIN #define MIN(a,b) ( (a) < (b) ? (a) : (b) ) #define MINUSABLETIME (TIMELOOP_NANOSEC / 2) /* 0.5 seconds */ BMK_runOutcome_t BMK_benchTimedFn(BMK_timedFnState_t* cont, BMK_benchParams_t p) { 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(p, cont->nbLoops); if(!BMK_isSuccessful_runOutcome(runResult)) { /* error : move out */ return runResult; } { 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); }