portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-04-09 21:50:45 +01:00
Code Releases Activity

Hadrons: contractor performances

Browse Source
This commit is contained in:
Antonin Portelli 2018-11-07 19:59:11 +00:00
parent 0c6e581336
commit 32376f0437
2 changed files with 53 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
Hadrons/A2AMatrix.hpp
View File

@ -196,6 +196,14 @@ public:
} }
} }
template <typename C, typename MatLeft, typename MatRight>
static inline double accTrMulFlops(C &acc, const MatLeft &a, const MatRight &b)
{
double n = a.rows()*a.cols();
return 8.*n;
}
// mul(res, a, b): res = a*b // mul(res, a, b): res = a*b
#ifdef USE_MKL #ifdef USE_MKL
template <template <class, int...> class Mat, int... Opts> template <template <class, int...> class Mat, int... Opts>
@ -246,7 +254,13 @@ public:
res = a*b; res = a*b;
} }
#endif #endif
template <typename Mat>
static inline double mulFlops(Mat &res, const Mat &a, const Mat &b)
{
double nr = a.rows(), nc = a.cols();
return nr*nr*(6.*nc + 2.*(nc - 1.));
}
private: private:
template <typename C, typename MatLeft, typename MatRight> template <typename C, typename MatLeft, typename MatRight>
static inline void makeDotRowPt(C * &aPt, unsigned int &aInc, C * &bPt, static inline void makeDotRowPt(C * &aPt, unsigned int &aInc, C * &bPt,

40
Hadrons/Utilities/Contractor.cc
View File

@ -142,6 +142,13 @@ std::set<unsigned int> parseTimeRange(const std::string str, const unsigned int
return tSet; return tSet;
} }
void printPerf(const double flops, const double fusec, const double bytes, const double busec)
{
std::cout << std::setw(10) << flops/fusec/1.0e3 << " GFlop/s "
<< std::setw(10) << bytes/busec*1.0e6/1024/1024/1024 << " GB/s"
<< std::endl;
}
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
// parse command line // parse command line
@ -204,6 +211,7 @@ int main(int argc, char* argv[])
std::vector<ComplexD> corr(par.global.nt); std::vector<ComplexD> corr(par.global.nt);
std::vector<A2AMatrixTr<ComplexD>> lastTerm(par.global.nt); std::vector<A2AMatrixTr<ComplexD>> lastTerm(par.global.nt);
A2AMatrix<ComplexD> prod, buf, tmp; A2AMatrix<ComplexD> prod, buf, tmp;
double fusec, busec, flops, bytes;
std::cout << "======== Product tr("; std::cout << "======== Product tr(";
for (unsigned int g = 0; g < term.size(); ++g) for (unsigned int g = 0; g < term.size(); ++g)
@ -242,17 +250,47 @@ int main(int argc, char* argv[])
for (auto &dt: translations) for (auto &dt: translations)
{ {
std::cout << "-- position " << t << ", translation " << dt << std::endl; std::cout << "-- position " << t << ", translation " << dt << std::endl;
if (term.size() > 2)
{
std::cout << "* matrix products" << std::endl;
}
flops = 0.;
bytes = 0.;
fusec = 0.;
busec = 0.;
prod = a2aMat.at(term[0])[TIME_MOD(t[0] + dt)]; prod = a2aMat.at(term[0])[TIME_MOD(t[0] + dt)];
for (unsigned int i = 1; i < term.size() - 1; ++i) for (unsigned int i = 1; i < term.size() - 1; ++i)
{ {
const A2AMatrix<ComplexD> &ref = a2aMat.at(term[i])[TIME_MOD(t[i] + dt)]; const A2AMatrix<ComplexD> &ref = a2aMat.at(term[i])[TIME_MOD(t[i] + dt)];
fusec -= usecond();
busec -= usecond();
A2AContraction::mul(tmp, prod, ref); A2AContraction::mul(tmp, prod, ref);
prod = tmp; fusec += usecond();
flops += A2AContraction::mulFlops(tmp, prod, ref);
prod = tmp;
busec += usecond();
bytes += 3.*tmp.rows()*tmp.cols()*sizeof(ComplexD);
} }
if (term.size() > 2)
{
printPerf(flops, fusec, bytes, busec);
}
std::cout << "* traces" << std::endl;
flops = 0.;
bytes = 0.;
fusec = 0.;
busec = 0.;
for (unsigned int tLast = 0; tLast < par.global.nt; ++tLast) for (unsigned int tLast = 0; tLast < par.global.nt; ++tLast)
{ {
fusec -= usecond();
busec -= usecond();
A2AContraction::accTrMul(corr[tLast], prod, lastTerm[tLast]); A2AContraction::accTrMul(corr[tLast], prod, lastTerm[tLast]);
fusec += usecond();
busec += usecond();
flops += A2AContraction::accTrMulFlops(corr[tLast], prod, lastTerm[tLast]);
bytes += 2.*prod.rows()*prod.cols()*sizeof(ComplexD);
} }
printPerf(flops, fusec, bytes, busec);
} }
for (unsigned int tLast = 0; tLast < par.global.nt; ++tLast) for (unsigned int tLast = 0; tLast < par.global.nt; ++tLast)
{ {