/*
* Copyright (c) Meta Platforms, Inc. and affiliates.
* 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.
*/
/* Abort execution if a condition is not met */
#define CONTROL(c) { if (!(c)) { DEBUGOUTPUT("error: %s \n", #c); abort(); } }
/* *************************************
* 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)
{
CONTROL(outcome.error_tag_never_ever_use_directly == 0);
return outcome.internal_never_ever_use_directly;
}
/* 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)
{
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 */
} }
/* benchmark */
{ size_t dstSize = 0;
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, (unsigned)p.srcSizes[blockNb], (int)res);
}
dstSize += res;
} }
} /* for (loopNb = 0; loopNb < nbLoops; loopNb++) */
BMK_timedFnState_t*
BMK_initStatic_timedFnState(void* buffer, size_t size, unsigned total_ms, unsigned run_ms)
{
typedef char check_size[ 2 * (sizeof(BMK_timedFnState_shell) >= sizeof(struct BMK_timedFnState_s)) - 1]; /* static assert : a compilation failure indicates that BMK_timedFnState_shell is not large enough */
typedef struct { check_size c; BMK_timedFnState_t tfs; } tfs_align; /* force tfs to be aligned at its next best position */
size_t const tfs_alignment = offsetof(tfs_align, tfs); /* provides the minimal alignment restriction for BMK_timedFnState_t */
BMK_timedFnState_t* const r = (BMK_timedFnState_t*)buffer;
if (buffer == NULL) return NULL;
if (size < sizeof(struct BMK_timedFnState_s)) return NULL;
if ((size_t)buffer % tfs_alignment) return NULL; /* buffer must be properly aligned */
BMK_resetTimedFnState(r, total_ms, run_ms);
return r;
}
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 = (PTime)total_ms * TIMELOOP_NANOSEC / 1000;
timedFnState->runBudget_ns = (PTime)run_ms * TIMELOOP_NANOSEC / 1000;
timedFnState->fastestRun.nanoSecPerRun = (double)TIMELOOP_NANOSEC * 2000000000; /* hopefully large enough : must be larger than any potential measurement */
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);
}
cont->timeSpent_ns += (unsigned long long)loopDuration_ns;
/* estimate nbLoops for next run to last approximately 1 second */
if (loopDuration_ns > ((double)runBudget_ns / 50)) {
double const fastestRun_ns = MIN(bestRunTime.nanoSecPerRun, newRunTime.nanoSecPerRun);
cont->nbLoops = (unsigned)((double)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 < (double)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) */