portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-10 07:55:35 +00:00
Code Releases Activity

Performance metrics for the Scalar Action force term

Browse Source
This commit is contained in:
Guido Cossu 2017-11-14 10:01:48 +00:00
parent e0cae833da
commit 1f1d77b01a
3 changed files with 53 additions and 15 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
lib/qcd/action/scalar/ScalarImpl.h
View File

@ -168,7 +168,11 @@ class ScalarImplTypes {
static inline void update_field(Field &P, Field &U, double ep)
{
#ifndef USE_FFT_ACCELERATION
double t0=usecond();
U += P * ep;
double t1=usecond();
double total_time = (t1-t0)/1e6;
std::cout << GridLogIntegrator << "Total time for updating field (s) : " << total_time << std::endl;
#else
// FFT transform P(x) -> P(p)
// divide by (M^2+p^2) M external parameter (how to pass?)

48
lib/qcd/action/scalar/ScalarInteractionAction.h
View File

@ -54,8 +54,8 @@ private:
SimpleCompressor<vobj> compressor;
int npoint = 2 * Ndim;
std::vector<int> directions; // = {0,1,2,3,0,1,2,3}; // forcing 4 dimensions
std::vector<int> displacements; // = {1,1,1,1, -1,-1,-1,-1};
std::vector<int> directions; //
std::vector<int> displacements; //
public:
ScalarInteractionAction(RealD ms, RealD l, RealD gval) : mass_square(ms), lambda(l), g(gval), displacements(2 * Ndim, 0), directions(2 * Ndim, 0)
@ -129,34 +129,50 @@ public:
virtual void deriv(const Field &p, Field &force)
{
double t0 = usecond();
assert(p._grid->Nd() == Ndim);
force = (2. * Ndim + mass_square) * p - 2. * lambda * p * p * p;
double interm_t = usecond();
// move this outside
static Stencil phiStencil(p._grid, npoint, 0, directions, displacements);
phiStencil.HaloExchange(p, compressor);
phiStencil.HaloExchange(p, compressor);
double halo_t = usecond();
int chunk = 128;
//for (int mu = 0; mu < QCD::Nd; mu++) force -= Cshift(p, mu, -1) + Cshift(p, mu, 1);
// inverting the order of the loops slows down the code(! g++ 7)
// cannot try to reduce the number of force writes by factor npoint...
// use cache blocking
for (int point = 0; point < npoint; point++)
{
parallel_for(int i = 0; i < p._grid->oSites(); i++)
#pragma omp parallel
{
const vobj *temp;
vobj temp2;
int permute_type;
StencilEntry *SE;
const vobj *temp;
#pragma omp for schedule(static, chunk)
for (int i = 0; i < p._grid->oSites(); i++)
{
SE = phiStencil.GetEntry(permute_type, point, i);
// prefetch next p?
if (SE->_is_local)
{
temp = &p._odata[SE->_offset];
if (SE->_permute)
{
vobj temp2;
permute(temp2, *temp, permute_type);
force._odata[i] -= temp2;
}
else
{
force._odata[i] -= *temp;
force._odata[i] -= *temp; // slow part. Dominated by this read/write (BW)
}
}
else
@ -164,8 +180,26 @@ public:
force._odata[i] -= phiStencil.CommBuf()[SE->_offset];
}
}
}
}
force *= N / g;
double t1 = usecond();
double total_time = (t1 - t0) / 1e6;
double interm_time = (interm_t - t0) / 1e6;
double halo_time = (halo_t - interm_t) / 1e6;
double stencil_time = (t1 - halo_t) / 1e6;
std::cout << GridLogIntegrator << "Total time for force computation (s) : " << total_time << std::endl;
std::cout << GridLogIntegrator << "Intermediate time for force computation (s): " << interm_time << std::endl;
std::cout << GridLogIntegrator << "Halo time in force computation (s) : " << halo_time << std::endl;
std::cout << GridLogIntegrator << "Stencil time in force computation (s) : " << stencil_time << std::endl;
double flops = p._grid->gSites() * (14 * N * N * N + 18 * N * N + 2);
double flops_no_stencil = p._grid->gSites() * (14 * N * N * N + 6 * N * N + 2);
double Gflops = flops / (total_time * 1e9);
double Gflops_no_stencil = flops_no_stencil / (interm_time * 1e9);
std::cout << GridLogIntegrator << "Flops: " << flops << " - Gflop/s : " << Gflops << std::endl;
std::cout << GridLogIntegrator << "Flops NS: " << flops_no_stencil << " - Gflop/s NS: " << Gflops_no_stencil << std::endl;
}
};

2
lib/qcd/hmc/GenericHMCrunner.h
View File

@ -211,7 +211,7 @@ typedef HMCWrapperTemplate<ScalarAdjImplR, MinimumNorm2, ScalarMatrixFields>
ScalarAdjGenericHMCRunner;
template <int Colours>
using ScalarNxNAdjGenericHMCRunner = HMCWrapperTemplate < ScalarNxNAdjImplR<Colours>, MinimumNorm2, ScalarNxNMatrixFields<Colours> >;
using ScalarNxNAdjGenericHMCRunner = HMCWrapperTemplate < ScalarNxNAdjImplR<Colours>, ForceGradient, ScalarNxNMatrixFields<Colours> >;
} // namespace QCD
} // namespace Grid