int main() {
int data[] = {1,2,3,4,5,6,7,8,9,10};
int m = reduce(max, data, 10);
int s = reduce(sum, data, 10);
printf("max : %i; sum: %i\n", m, s);
int pm = parallel_reduce(max, data, 10);
int ps = parallel_reduce(sum, data, 10);
printf("parallel max : %i; parallel sum: %i\n", pm, ps);
return 0;
}
void dot( const ConstVectorType & X ,
const Finalize & finalize )
{
typedef DotSingle< ConstVectorType > functor ;
parallel_reduce( X.dimension_0() , functor( X ) , finalize );
}
void dot( const ConstVectorType & X ,
const Finalize & finalize )
{
typedef DotSingle< ConstVectorType > functor ;
parallel_reduce( X.extent(0) , functor( X ) , finalize );
}
double parallelPearsonCorrelationCoefficient(double *a, double *b) {
double meanA = parallelMean(a);
double meanB = parallelMean(b);
double standardDeviationA = parallelStandardDeviation(a,&meanA);
double standardDeviationB = parallelStandardDeviation(b,&meanB);
// TOP of fraction
double value = parallel_reduce(blocked_range<size_t>(0, vec_size, 10000), double(0),
[=](blocked_range<size_t> &r, double sum) -> double {
for(size_t i=r.begin();i!=r.end();++i){
sum += ((a[i]-meanA) * (b[i]-meanB));
}
return sum;
},
[=](double a, double b){
return a+b;
});
value *= 1.0/vec_size;
// BOTTOM of fraction
value /= (standardDeviationA*standardDeviationB);
return value;
}
void CowichanTBB::life(BoolMatrix input, BoolMatrix output) {
GameOfLife game(input, output, nr, nc);
for (index_t i = 0; i < LIFE_ITERATIONS; ++i) {
// update CA simulation
parallel_reduce(Range2D(0, nr, 0, nc), game, auto_partitioner());
// check if there are alive cells
if (!game.isAlive()) {
no_cells_alive();
}
// swap arrays (ping-pong approach)
game.swap();
}
// final result is in input - copy to output
if (LIFE_ITERATIONS % 2 == 0) {
for (index_t r = 0; r < nr; r++) {
for (index_t c = 0; c < nc; c++) {
MATRIX_RECT(output, r, c) = MATRIX_RECT(input, r, c);
}
}
}
}
void
SPHSolver::CalculateDensity(
FluidParticle* particle
)
{
// -- Compute density over a kernel function of neighbors
#ifdef USE_TBB
float density = parallel_reduce(
blocked_range<FluidParticle**>( particle->neighbors, particle->neighbors + particle->neighborsCount ),
0.f,
[&](const blocked_range<FluidParticle**>& range, float init)->float {
for( FluidParticle** neighbor = range.begin(); neighbor != range.end(); ++neighbor)
{
init += KernelPoly6(glm::distance((*neighbor)->Position(), particle->Position()), m_kernelRadius);
}
return init;
},
[]( float x, float y )->float {
return x + y;
}
);
#else
float density = 0.0f;
for (size_t i = 0; i < particle->neighborsCount; ++i) {
FluidParticle* neighbor = particle->neighbors[i];
float tempDensity = KernelPoly6(glm::distance(neighbor->Position(), particle->Position()), m_kernelRadius);
density += tempDensity;
}
#endif;
density = FluidParticle::mass * density;
particle->SetDensity(density);
}
void testFakeAttributeRead()
{
task_scheduler_init scheduler( 100 );
TestSceneCache task( "fake" );
parallel_reduce( blocked_range<size_t>( 0, 100 ), task );
BOOST_CHECK( task.errors() == 100000 );
}
inline static
value_type apply( const size_t n ,
const scalar_vector & x )
{
value_type result = 0 ;
Dot op ; op.x = x ;
parallel_reduce( n , op , result );
return result ;
}
void InitThreadPool() {
boost::uint32_t systemCores = Threading::GetAvailableCoresMask();
boost::uint32_t mainAffinity = systemCores;
#ifndef UNIT_TEST
mainAffinity &= configHandler->GetUnsigned("SetCoreAffinity");
#endif
boost::uint32_t ompAvailCores = systemCores & ~mainAffinity;
{
#ifndef UNIT_TEST
int workerCount = std::min(ThreadPool::GetMaxThreads() - 1, configHandler->GetUnsigned("WorkerThreadCount"));
ThreadPool::SetThreadSpinTime(configHandler->GetUnsigned("WorkerThreadSpinTime"));
#else
int workerCount = -1;
#endif
const int numCores = ThreadPool::GetMaxThreads();
// For latency reasons our worker threads yield rarely and so eat a lot cputime with idleing.
// So it's better we always leave 1 core free for our other threads, drivers & OS
if (workerCount < 0) {
if (numCores == 2) {
workerCount = numCores;
} else if (numCores < 6) {
workerCount = numCores - 1;
} else {
workerCount = numCores / 2;
}
}
if (workerCount > numCores) {
LOG_L(L_WARNING, "Set ThreadPool workers to %i, but there are just %i cores!", workerCount, numCores);
workerCount = numCores;
}
ThreadPool::SetThreadCount(workerCount);
}
// set affinity of worker threads
boost::uint32_t ompCores = 0;
ompCores = parallel_reduce([&]() -> boost::uint32_t {
const int i = ThreadPool::GetThreadNum();
// 0 is the source thread, skip
if (i == 0)
return 0;
boost::uint32_t ompCore = GetCpuCoreForWorkerThread(i - 1, ompAvailCores, mainAffinity);
//boost::uint32_t ompCore = ompAvailCores;
Threading::SetAffinity(ompCore);
return ompCore;
}, [](boost::uint32_t a, boost::unique_future<boost::uint32_t>& b) -> boost::uint32_t { return a | b.get(); });
// affinity of mainthread
boost::uint32_t nonOmpCores = ~ompCores;
if (mainAffinity == 0) mainAffinity = systemCores;
Threading::SetAffinityHelper("Main", mainAffinity & nonOmpCores);
}
void
SPHSolver::CalculatePressureForceField(
FluidParticle* particle
)
{
// -- Compute pressure gradient
glm::vec3 pressureGrad(0.0f);
float particleDensitySquared = particle->Density() * particle->Density();
#ifdef USE_TBB
pressureGrad = parallel_reduce(
blocked_range<FluidParticle**>( particle->neighbors, particle->neighbors + particle->neighborsCount ),
glm::vec3(0.0),
[&](const blocked_range<FluidParticle**>& range, glm::vec3 init)->glm::vec3 {
for( FluidParticle** neighbor = range.begin(); neighbor != range.end(); ++neighbor )
{
glm::vec3 r = particle->Position() - (*neighbor)->Position();
float x = glm::distance((*neighbor)->Position(), particle->Position());
glm::vec3 kernelGrad = GradKernelSpiky(r, x, m_kernelRadius);
float neighborDensitySquared = (*neighbor)->Density() * (*neighbor)->Density();
float tempPressureForce =
(
(particle->Pressure() / particleDensitySquared) +
((*neighbor)->Pressure() / neighborDensitySquared)
);
init += tempPressureForce * kernelGrad;
}
return init;
},
[]( glm::vec3 x, glm::vec3 y )->glm::vec3 {
return x + y;
}
);
#else
for (size_t i = 0; i < particle->neighborsCount; ++i) {
FluidParticle* neighbor = particle->neighbors[i];
glm::vec3 r = particle->Position() - neighbor->Position();
float x = glm::distance(neighbor->Position(), particle->Position());
glm::vec3 kernelGrad = GradKernelSpiky(r, x, m_kernelRadius);
float neighborDensitySquared = neighbor->Density() * neighbor->Density();
float tempPressureForce =
(
particle->Pressure() / particleDensitySquared +
neighbor->Pressure() / neighborDensitySquared
);
pressureGrad += tempPressureForce * kernelGrad;
}
#endif
pressureGrad = -pressureGrad * FluidParticle::mass * FluidParticle::mass;
particle->SetPressureForce(pressureGrad);
}
double parallelStandardDeviation(double *a, double *meanValue) {
// reduce by pointing to a place in memory for i in a[] then remove the meanValue and then square
// emnumerate the above for all a[] then squareroot and divide by array size
// This is the Lambda syntax,
return sqrt(parallel_reduce(blocked_range<double*>(a, a+vec_size),0.0,
[&](const blocked_range<double*>& r, double sum)->double {
for(double* i=r.begin(); i!=r.end(); ++i)
sum += pow(*i-*meanValue,2.0);
return sum;
}, [](double x, double y)->double {return x+y;}
)/vec_size);
}
void buildGraph(std::vector<std::pair<datadim_t, mdMap_t>> data, std::vector<std::pair<discretedim_t, discretedim_t>> dataDiscrete, std::shared_ptr<gc::Graph> graph, data_t threshold) {
std::cout << "Build initial graph: " << std::flush;
long edgeCount = 0;
data_t xMax = std::numeric_limits<data_t>::lowest();
for (std::size_t i = 0; i < data.size(); i++) {
std::list<std::size_t> refs;
// reverse search existing vertices
TBBSearchHelper helper1(i, graph);
parallel_reduce(tbb::blocked_range<std::size_t>(0, i), helper1);
refs.splice(refs.end(), helper1.refs);
// do not add self reference (=i)
// test edges too non exisiting vertices
TBBEdgeHelper helper2(data[i].first, data[i].second, dataDiscrete[i].first, dataDiscrete[i].second, &data, &dataDiscrete, threshold);
parallel_reduce(tbb::blocked_range<std::size_t>(i + 1, data.size()), helper2);
xMax = std::max(xMax, helper2.xMax);
refs.splice(refs.end(), helper2.refs);
// store
graph->add(refs);
edgeCount += refs.size();
// report progress
if (i % 100 == 0) {
std::cout << i << std::flush;
} else if (i % 10 == 0) {
std::cout << "." << std::flush;
}
}
std::cout << "done (" << (edgeCount / 2) << " edges, max="<< xMax << ")" << std::endl;
}
bool operator() ()
{
bool passed = true;
printf("%s::%s ... ",TOSTRING(isa),name);
fflush(stdout);
const size_t M = 10;
for (size_t N=10; N<10000000; N*=2.1f)
{
/* sequentially calculate sum of squares */
size_t sum0 = 0;
for (size_t i=0; i<N; i++) {
sum0 += i*i;
}
/* parallel calculation of sum of squares */
double t0 = getSeconds();
for (size_t m=0; m<M; m++)
{
size_t sum1 = parallel_reduce( size_t(0), size_t(N), size_t(1024), size_t(0), [&](const range<size_t>& r) -> size_t
{
size_t s = 0;
for (size_t i=r.begin(); i<r.end(); i++)
s += i*i;
return s;
},
[](const size_t v0, const size_t v1) {
return v0+v1;
});
passed = sum0 == sum1;
}
double t1 = getSeconds();
printf("%zu/%3.2fM ",N,1E-6*double(N*M)/(t1-t0));
}
/* output if test passed or not */
if (passed) printf("[passed]\n");
else printf("[failed]\n");
return passed;
}
void InitThreadPool() {
boost::uint32_t systemCores = Threading::GetAvailableCoresMask();
boost::uint32_t mainAffinity = systemCores;
boost::uint32_t ompAvailCores = systemCores & ~mainAffinity;
#ifndef UNIT_TEST
mainAffinity = systemCores & configHandler->GetUnsigned("SetCoreAffinity");
#endif
{
int workerCount = -1;
#ifndef UNIT_TEST
workerCount = configHandler->GetUnsigned("WorkerThreadCount");
ThreadPool::SetThreadSpinTime(configHandler->GetUnsigned("WorkerThreadSpinTime"));
#endif
// For latency reasons our worker threads yield rarely and so eat a lot cputime with idleing.
// So it's better we always leave 1 core free for our other threads, drivers & OS
if (workerCount < 0) workerCount = ThreadPool::GetMaxThreads() - 1;
//if (workerCount > ThreadPool::GetMaxThreads()) LOG_L(L_WARNING, "");
ThreadPool::SetThreadCount(workerCount);
}
// set affinity of worker threads
boost::uint32_t ompCores = 0;
ompCores = parallel_reduce([&]() -> boost::uint32_t {
const int i = ThreadPool::GetThreadNum();
if (i != 0) {
// 0 is the source thread, skip
boost::uint32_t ompCore = GetCpuCoreForWorkerThread(i - 1, ompAvailCores, mainAffinity);
Threading::SetAffinity(ompCore);
return ompCore;
}
return 0;
}, [](boost::uint32_t a, boost::unique_future<boost::uint32_t>& b) -> boost::uint32_t { return a | b.get(); });
// affinity of mainthread
boost::uint32_t nonOmpCores = ~ompCores;
if (mainAffinity == 0) mainAffinity = systemCores;
Threading::SetAffinityHelper("Main", mainAffinity & nonOmpCores);
}
void CowichanTBB::norm(PointVector pointsIn, PointVector pointsOut) {
MinMaxReducer minmax(pointsIn);
// find min/max coordinates
parallel_reduce(Range(0, n), minmax, auto_partitioner());
Point minPoint = minmax.getMinimum();
Point maxPoint = minmax.getMaximum();
// compute scaling factors
real xfactor = (real)((maxPoint.x == minPoint.x) ?
0.0 : 1.0 / (maxPoint.x - minPoint.x));
real yfactor = (real)((maxPoint.y == minPoint.y) ?
0.0 : 1.0 / (maxPoint.y - minPoint.y));
Normalizer normalizer(pointsIn, pointsOut, minPoint.x, minPoint.y, xfactor,
yfactor);
// normalize the vector
parallel_for(Range(0, n), normalizer, auto_partitioner());
}
typename XVector::scalar_type V_Dot( const XVector & x, const YVector & y)
{
V_DotFunctor<XVector,YVector> f(x,y);
return parallel_reduce(x.dimension_0(),f);
}
void BVH4BuilderTwoLevel::build(size_t threadIndex, size_t threadCount)
{
/* delete some objects */
size_t N = scene->size();
if (N < objects.size()) {
parallel_for(N, objects.size(), [&] (const range<size_t>& r) {
for (size_t i=r.begin(); i<r.end(); i++) {
delete builders[i]; builders[i] = nullptr;
delete objects[i]; objects[i] = nullptr;
}
});
}
/* reset memory allocator */
bvh->alloc.reset();
/* skip build for empty scene */
const size_t numPrimitives = scene->getNumPrimitives<TriangleMesh,1>();
if (numPrimitives == 0) {
prims.resize(0);
bvh->set(BVH4::emptyNode,empty,0);
return;
}
double t0 = bvh->preBuild(TOSTRING(isa) "::BVH4BuilderTwoLevel");
#if PROFILE
profile(2,20,numPrimitives,[&] (ProfileTimer& timer)
{
#endif
/* resize object array if scene got larger */
if (objects.size() < N) objects.resize(N);
if (builders.size() < N) builders.resize(N);
if (refs.size() < N) refs.resize(N);
nextRef = 0;
/* create of acceleration structures */
parallel_for(size_t(0), N, [&] (const range<size_t>& r)
{
for (size_t objectID=r.begin(); objectID<r.end(); objectID++)
{
TriangleMesh* mesh = scene->getTriangleMeshSafe(objectID);
/* verify meshes got deleted properly */
if (mesh == nullptr || mesh->numTimeSteps != 1) {
assert(objectID < objects.size () && objects[objectID] == nullptr);
assert(objectID < builders.size() && builders[objectID] == nullptr);
continue;
}
/* create BVH and builder for new meshes */
if (objects[objectID] == nullptr)
createTriangleMeshAccel(mesh,(AccelData*&)objects[objectID],builders[objectID]);
}
});
/* parallel build of acceleration structures */
parallel_for(size_t(0), N, [&] (const range<size_t>& r)
{
for (size_t objectID=r.begin(); objectID<r.end(); objectID++)
{
/* ignore if no triangle mesh or not enabled */
TriangleMesh* mesh = scene->getTriangleMeshSafe(objectID);
if (mesh == nullptr || !mesh->isEnabled() || mesh->numTimeSteps != 1)
continue;
BVH4* object = objects [objectID]; assert(object);
Builder* builder = builders[objectID]; assert(builder);
/* build object if it got modified */
#if !PROFILE
if (mesh->isModified())
#endif
builder->build(0,0);
/* create build primitive */
if (!object->bounds.empty())
refs[nextRef++] = BVH4BuilderTwoLevel::BuildRef(object->bounds,object->root);
}
});
/* fast path for single geometry scenes */
if (nextRef == 1) {
bvh->set(refs[0].node,refs[0].bounds(),numPrimitives);
return;
}
/* open all large nodes */
refs.resize(nextRef);
open_sequential(numPrimitives);
/* fast path for small geometries */
if (refs.size() == 1) {
bvh->set(refs[0].node,refs[0].bounds(),numPrimitives);
return;
}
/* compute PrimRefs */
prims.resize(refs.size());
const PrimInfo pinfo = parallel_reduce(size_t(0), refs.size(), size_t(1024), PrimInfo(empty), [&] (const range<size_t>& r) -> PrimInfo
{
PrimInfo pinfo(empty);
for (size_t i=r.begin(); i<r.end(); i++) {
pinfo.add(refs[i].bounds());
prims[i] = PrimRef(refs[i].bounds(),(size_t)refs[i].node);
}
return pinfo;
}, [] (const PrimInfo& a, const PrimInfo& b) { return PrimInfo::merge(a,b); });
/* skip if all objects where empty */
if (pinfo.size() == 0)
bvh->set(BVH4::emptyNode,empty,0);
/* otherwise build toplevel hierarchy */
else
{
BVH4::NodeRef root;
BVHBuilderBinnedSAH::build<BVH4::NodeRef>
(root,
[&] { return bvh->alloc.threadLocal2(); },
[&] (const isa::BVHBuilderBinnedSAH::BuildRecord& current, BVHBuilderBinnedSAH::BuildRecord* children, const size_t N, FastAllocator::ThreadLocal2* alloc) -> int
{
BVH4::Node* node = (BVH4::Node*) alloc->alloc0.malloc(sizeof(BVH4::Node)); node->clear();
for (size_t i=0; i<N; i++) {
node->set(i,children[i].pinfo.geomBounds);
children[i].parent = (size_t*)&node->child(i);
}
*current.parent = bvh->encodeNode(node);
return 0;
},
[&] (const BVHBuilderBinnedSAH::BuildRecord& current, FastAllocator::ThreadLocal2* alloc) -> int
{
assert(current.prims.size() == 1);
*current.parent = (BVH4::NodeRef) prims[current.prims.begin()].ID();
return 1;
},
[&] (size_t dn) { bvh->scene->progressMonitor(0); },
prims.data(),pinfo,BVH4::N,BVH4::maxBuildDepthLeaf,4,1,1,1.0f,1.0f);
bvh->set(root,pinfo.geomBounds,numPrimitives);
}
#if PROFILE
});
#endif
bvh->alloc.cleanup();
bvh->postBuild(t0);
}
double ParallelTBBSclMlt(const double* A, const double* B, const int len){
ScalarMultiplicator mul(A, B);
parallel_reduce(blocked_range<int>(0, len), mul);
return mul.Result();
}
double parallelMean(double *a) {
return parallel_reduce(blocked_range<double*>(a,a+vec_size),0.0,
[](const blocked_range<double*>& r, double value)->double {
return accumulate(r.begin(),r.end(),value);
},plus<double>())/vec_size;
}
size_type apply() const
{
size_type count = 0u;
parallel_reduce(m_bitset.m_blocks.dimension_0(), *this, count);
return count;
}
BBox<dim> find_bounding_box(Reals coords) {
CHECK(coords.size() % dim == 0);
return parallel_reduce(coords.size() / dim, BBoxFunctor<dim>(coords));
}
double ParallelTBBSum(const double* src, const int len){
VectorSummator sum(src);
parallel_reduce(blocked_range<int>(0, len), sum);
return sum.Result();
}