portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-17 15:27:06 +01:00
Code Releases Activity

Synchronize blocking infrastructure with GPT

Browse Source
This commit is contained in:
Christoph Lehner
2020-05-06 08:42:28 -04:00
parent 6b64727161
commit e9b295f967
4 changed files with 216 additions and 282 deletions
Download Patch File Download Diff File Expand all files Collapse all files

205
Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h
View File

@ -37,211 +37,6 @@ Author: Christoph Lehner <clehner@bnl.gov>
NAMESPACE_BEGIN(Grid);
////////////////////////////////////////////////////////
// Move following 100 LOC to lattice/Lattice_basis.h
////////////////////////////////////////////////////////
template<class Field>
void basisOrthogonalize(std::vector<Field> &basis,Field &w,int k)
{
// If assume basis[j] are already orthonormal,
// can take all inner products in parallel saving 2x bandwidth
// Save 3x bandwidth on the second line of loop.
// perhaps 2.5x speed up.
// 2x overall in Multigrid Lanczos
for(int j=0; j<k; ++j){
auto ip = innerProduct(basis[j],w);
w = w - ip*basis[j];
}
}
template<class Field>
void basisRotate(std::vector<Field> &basis,Eigen::MatrixXd& Qt,int j0, int j1, int k0,int k1,int Nm)
{
typedef decltype(basis[0].View()) View;
auto tmp_v = basis[0].View();
Vector<View> basis_v(basis.size(),tmp_v);
typedef typename Field::vector_object vobj;
GridBase* grid = basis[0].Grid();
for(int k=0;k<basis.size();k++){
basis_v[k] = basis[k].View();
}
#if 0
std::vector < vobj , commAllocator<vobj> > Bt(thread_max() * Nm); // Thread private
thread_region
{
vobj* B = Bt.data() + Nm * thread_num();
thread_for_in_region(ss, grid->oSites(),{
for(int j=j0; j<j1; ++j) B[j]=0.;
for(int j=j0; j<j1; ++j){
for(int k=k0; k<k1; ++k){
B[j] +=Qt(j,k) * basis_v[k][ss];
}
}
for(int j=j0; j<j1; ++j){
basis_v[j][ss] = B[j];
}
});
}
#else
int nrot = j1-j0;
uint64_t oSites =grid->oSites();
uint64_t siteBlock=(grid->oSites()+nrot-1)/nrot; // Maximum 1 additional vector overhead
// printf("BasisRotate %d %d nrot %d siteBlock %d\n",j0,j1,nrot,siteBlock);
Vector <vobj> Bt(siteBlock * nrot);
auto Bp=&Bt[0];
// GPU readable copy of Eigen matrix
Vector<double> Qt_jv(Nm*Nm);
double *Qt_p = & Qt_jv[0];
for(int k=0;k<Nm;++k){
for(int j=0;j<Nm;++j){
Qt_p[j*Nm+k]=Qt(j,k);
}
}
// Block the loop to keep storage footprint down
vobj zz=Zero();
for(uint64_t s=0;s<oSites;s+=siteBlock){
// remaining work in this block
int ssites=MIN(siteBlock,oSites-s);
// zero out the accumulators
accelerator_for(ss,siteBlock*nrot,vobj::Nsimd(),{
auto z=coalescedRead(zz);
coalescedWrite(Bp[ss],z);
});
accelerator_for(sj,ssites*nrot,vobj::Nsimd(),{
int j =sj%nrot;
int jj =j0+j;
int ss =sj/nrot;
int sss=ss+s;
for(int k=k0; k<k1; ++k){
auto tmp = coalescedRead(Bp[ss*nrot+j]);
coalescedWrite(Bp[ss*nrot+j],tmp+ Qt_p[jj*Nm+k] * coalescedRead(basis_v[k][sss]));
}
});
accelerator_for(sj,ssites*nrot,vobj::Nsimd(),{
int j =sj%nrot;
int jj =j0+j;
int ss =sj/nrot;
int sss=ss+s;
coalescedWrite(basis_v[jj][sss],coalescedRead(Bp[ss*nrot+j]));
});
}
#endif
}
// Extract a single rotated vector
template<class Field>
void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,int j, int k0,int k1,int Nm)
{
typedef decltype(basis[0].View()) View;
typedef typename Field::vector_object vobj;
GridBase* grid = basis[0].Grid();
result.Checkerboard() = basis[0].Checkerboard();
auto result_v=result.View();
Vector<View> basis_v(basis.size(),result_v);
for(int k=0;k<basis.size();k++){
basis_v[k] = basis[k].View();
}
vobj zz=Zero();
Vector<double> Qt_jv(Nm);
double * Qt_j = & Qt_jv[0];
for(int k=0;k<Nm;++k) Qt_j[k]=Qt(j,k);
accelerator_for(ss, grid->oSites(),vobj::Nsimd(),{
auto B=coalescedRead(zz);
for(int k=k0; k<k1; ++k){
B +=Qt_j[k] * coalescedRead(basis_v[k][ss]);
}
coalescedWrite(result_v[ss], B);
});
}
template<class Field>
void basisReorderInPlace(std::vector<Field> &_v,std::vector<RealD>& sort_vals, std::vector<int>& idx)
{
int vlen = idx.size();
assert(vlen>=1);
assert(vlen<=sort_vals.size());
assert(vlen<=_v.size());
for (size_t i=0;i<vlen;i++) {
if (idx[i] != i) {
//////////////////////////////////////
// idx[i] is a table of desired sources giving a permutation.
// Swap v[i] with v[idx[i]].
// Find j>i for which _vnew[j] = _vold[i],
// track the move idx[j] => idx[i]
// track the move idx[i] => i
//////////////////////////////////////
size_t j;
for (j=i;j<idx.size();j++)
if (idx[j]==i)
break;
assert(idx[i] > i); assert(j!=idx.size()); assert(idx[j]==i);
swap(_v[i],_v[idx[i]]); // should use vector move constructor, no data copy
std::swap(sort_vals[i],sort_vals[idx[i]]);
idx[j] = idx[i];
idx[i] = i;
}
}
}
inline std::vector<int> basisSortGetIndex(std::vector<RealD>& sort_vals)
{
std::vector<int> idx(sort_vals.size());
std::iota(idx.begin(), idx.end(), 0);
// sort indexes based on comparing values in v
std::sort(idx.begin(), idx.end(), [&sort_vals](int i1, int i2) {
return ::fabs(sort_vals[i1]) < ::fabs(sort_vals[i2]);
});
return idx;
}
template<class Field>
void basisSortInPlace(std::vector<Field> & _v,std::vector<RealD>& sort_vals, bool reverse)
{
std::vector<int> idx = basisSortGetIndex(sort_vals);
if (reverse)
std::reverse(idx.begin(), idx.end());
basisReorderInPlace(_v,sort_vals,idx);
}
// PAB: faster to compute the inner products first then fuse loops.
// If performance critical can improve.
template<class Field>
void basisDeflate(const std::vector<Field> &_v,const std::vector<RealD>& eval,const Field& src_orig,Field& result) {
result = Zero();
assert(_v.size()==eval.size());
int N = (int)_v.size();
for (int i=0;i<N;i++) {
Field& tmp = _v[i];
axpy(result,TensorRemove(innerProduct(tmp,src_orig)) / eval[i],tmp,result);
}
}
/////////////////////////////////////////////////////////////
// Implicitly restarted lanczos
/////////////////////////////////////////////////////////////