void SamplerBase::getGPUCycles(FlowID fid, float ratio) {
char str[255];
sprintf(str,"P(%u):clockTicks", getsFid());
GStats *gref = 0;
gref = GStats::getRef(str);
I(gref);
uint64_t cticks = gref->getSamples(); // clockTicks
if (doPower)
gpuEstimatedCycles = cticks*(ratio);
else
gpuEstimatedCycles = cticks;
gpuSampledRatio = ratio;
}
void PowerModel::updateActivity(uint64_t timeinterval, FlowID fid)
/* updateActivity {{{1 */
{
// Update stats that are passed to McPat
timeInterval = timeinterval;
PowerStats *pwrstat = 0;
uint32_t zcnt = 0;
for(size_t i = 0; i < stats->size(); i++) {
pwrstat = (*stats)[i];
(*activity)[i] = pwrstat->getActivity();
if (!(*activity)[i])
zcnt++;
}
dumpActivity();
#ifdef DEBUG
static int pcnt = 0;
if ((2*zcnt > stats->size())) {
if (pcnt++ < 20)
printf("WARNING: Too many blocks have zero activity. Check the eSESC to McPAT mapping.\n");
}
#endif
// Update internal performance counters
GStats *gref = 0;
char str[128];
for (size_t i=0; i<ncores; i++) {
sprintf(str,"S(%lu):globalClock_Timing", i);
gref = GStats::getRef(str);
I(gref);
uint64_t cticks = gref->getSamples(); // clockTicks
if (cticks - clockPrev[i] == 0) {
I(0);
clockInterval[i] = 0;
} else {
clockInterval[i] = cticks - clockPrev[i]; // + 1000;
}
clockPrev[i] = cticks;
}
// Just use the total stats, not average
avgTimingIPC = static_cast<float>(1.0)/static_cast<float>(getEmul(fid)->getSampler()->getMeaCPI());
ipc = avgTimingIPC;
}
void EmuSampler::calcCPI()
/* calculates cpi for the last EmuTiming mode */
{
if (mode != EmuTiming)
return;
if (terminated)
return;
I(calcCPIJustCalled == false);
I(!stopJustCalled);
calcCPIJustCalled = true;
//get instruction count
uint64_t timingInst = iusage[EmuTiming]->getSamples();
uint64_t instCount = timingInst - instPrev[sFid];
char str[255];
sprintf(str, "P(%d):clockTicks", sFid); //FIXME: needs mapping
GStats *gref = 0;
gref = GStats::getRef(str);
I(gref);
//fprintf(stderr, "\n\nCalcCPI: %s->getSamples returns %lld (previously : %lld)\n\n",str,gref->getSamples(),clockPrev[sFid]);
uint64_t cticks = gref->getSamples();
uint64_t clockInterval = cticks - clockPrev[sFid] + 1;
if(instCount<100 || clockInterval <100) // To avoid too frequent sampling errors
return;
instPrev[sFid] = timingInst;
clockPrev[sFid] = cticks;
//Get freezed cycles due to thermal throttling
sprintf(str, "P(%d):nFreeze", sFid); //FIXME: needs mapping
gref = 0;
gref = GStats::getRef(str);
I(gref);
uint64_t fticks = gref->getSamples();
//fprintf(stderr,"\n\nCalcCPI: %s->getSamples returns %lld\n\n",str,gref->getSamples());
uint64_t fticksInterval = fticks - fticksPrev[sFid] + 1;
fticksPrev[sFid] = fticks;
uint64_t uInstCount = nCommitted->getSamples() - prevnCommitted;
prevnCommitted = nCommitted->getSamples();
uint64_t adjustedClock = clockInterval - fticksInterval;
float newCPI = static_cast<float>(adjustedClock)/static_cast<float>(instCount);
// newuCPI should be used, but as the sampler, we can only scale Inst, not uInst.
//float newCPI = static_cast<float>(clockInterval)/static_cast<float>(uInstCount);
I(newCPI>0);
/*
//printf("InstCnt:%ld, prevInst:%ld, clock:%ld, ipc:%f\n", instCount, instPrev, clockInterval, static_cast<float>(instCount)/static_cast<float>(clockInterval));
if (newCPI <= 0){
fprintf(stderr,"newCPI = %f, clockInterval = %lld, adjustedClock = %lld, uInstCount = %lld",newCPI,clockInterval,adjustedClock, uInstCount);
}
*/
float newipc = static_cast<float>(instCount)/static_cast<float>(adjustedClock);
float newuipc = static_cast<float>(uInstCount)/static_cast<float>(adjustedClock);
uipc->sample(100*newuipc,true);
ipc->sample(100*newipc,true);
//I(newCPI<=4);
if (newCPI > 5) {
//newCPI = 5.0;
//return;
}
//meaCPI = newCPI;
meaCPI = newCPI;
meauCPI = 1.0/newuipc;
}