void
TestLocalAllocations( int nthread ) {
ASSERT(nthread > 0, "nthread must be positive");
#define NITERATIONS 1000
Body myBody(nthread, NITERATIONS);
tbb::combinable<int> myCombinable;
myBody.locals = &myCombinable;
NativeParallelFor( nthread, myBody );
int mySum = 0;
int mySlots = 0;
CombineEachHelperCnt<int> myCountCombine(mySum, mySlots);
myCombinable.combine_each(myCountCombine);
ASSERT(nthread == mySlots, "Incorrect number of slots");
ASSERT(mySum == (nthread - 1) * nthread / 2, "Incorrect values in result");
}
void
RunOneCriticalSectionTest(int nThreads, int csWorkRatio, bool test_throw) {
tbb::task_scheduler_init init(tbb::task_scheduler_init::deferred);
tbb::enumerable_thread_specific<double> test_locals;
int myCount = 0;
BusyBody myBody(nThreads, csWorkRatio, test_locals, myCount, test_throw);
BusyBodyScoped myScopedBody(nThreads, csWorkRatio, test_locals, myCount, test_throw);
init.initialize(nThreads);
tbb::tick_count t0;
{
t0 = tbb::tick_count::now();
myCount = 0;
NativeParallelFor(nThreads, myBody);
ASSERT(myCount == (MAX_WORK - (MAX_WORK % nThreads)), NULL);
REMARK("%d threads, work ratio %d per cent, time %g", nThreads, csWorkRatio, (tbb::tick_count::now() - t0).seconds());
if (nThreads > 1) {
double etsSum = 0;
double etsMax = 0;
double etsMin = 0;
double etsSigmaSq = 0;
double etsSigma = 0;
for(tbb::enumerable_thread_specific<double>::const_iterator ci = test_locals.begin(); ci != test_locals.end(); ci++) {
etsSum += *ci;
if(etsMax==0.0) {
etsMin = *ci;
}
else {
if(etsMin > *ci) etsMin = *ci;
}
if(etsMax < *ci) etsMax = *ci;
}
double etsAvg = etsSum / (double)nThreads;
for(tbb::enumerable_thread_specific<double>::const_iterator ci = test_locals.begin(); ci != test_locals.end(); ci++) {
etsSigma = etsAvg - *ci;
etsSigmaSq += etsSigma * etsSigma;
}
// an attempt to gauge the "fairness" of the scheduling of the threads. We figure
// the standard deviation, and compare it with the maximum deviation from the
// average time. If the difference is 0 that means all threads finished in the same
// amount of time. If non-zero, the difference is divided by the time, and the
// negative log is taken. If > 2, then the difference is on the order of 0.01*t
// where T is the average time. We aritrarily define this as "fair."
etsSigma = sqrt(etsSigmaSq/double(nThreads));
etsMax -= etsAvg; // max - a == delta1
etsMin = etsAvg - etsMin; // a - min == delta2
if(etsMax < etsMin) etsMax = etsMin;
etsMax -= etsSigma;
// ASSERT(etsMax >= 0, NULL); // shouldn't the maximum difference from the mean be > the stddev?
etsMax = (etsMax > 0.0) ? etsMax : 0.0; // possible rounding error
double fairness = etsMax / etsAvg;
if(fairness == 0.0) {
fairness = 100.0;
}
else fairness = - log10(fairness);
if(fairness > 2.0 ) {
REMARK(" Fair (%g)\n", fairness);
}
else {
REMARK(" Unfair (%g)\n", fairness);
}
}
myCount = 0;
NativeParallelFor(nThreads, myScopedBody);
ASSERT(myCount == (MAX_WORK - (MAX_WORK % nThreads)), NULL);
}
init.terminate();
}