1
0
mirror of https://github.com/paboyle/Grid.git synced 2025-04-04 19:25:56 +01:00

Partial dirichlet support

This commit is contained in:
Peter Boyle 2022-11-15 16:23:41 -05:00
parent 5fe480d81c
commit 0db4f1803f

View File

@ -3,26 +3,108 @@
NAMESPACE_BEGIN(Grid);
template<class vobj>
accelerator_inline void exchangeSIMT(vobj &mp0,vobj &mp1,const vobj &vp0,const vobj &vp1,Integer type)
{
typedef decltype(coalescedRead(mp0)) sobj;
unsigned int Nsimd = vobj::Nsimd();
unsigned int mask = Nsimd >> (type + 1);
int lane = acceleratorSIMTlane(Nsimd);
int j0 = lane &(~mask); // inner coor zero
int j1 = lane |(mask) ; // inner coor one
const vobj *vpa = &vp0;
const vobj *vpb = &vp1;
const vobj *vp = (lane&mask) ? (vpb) : (vpa);
auto sa = coalescedRead(vp[0],j0);
auto sb = coalescedRead(vp[0],j1);
coalescedWrite(mp0,sa);
coalescedWrite(mp1,sb);
}
class SimpleStencilParams{
public:
Coordinate dirichlet;
int partialDirichlet;
SimpleStencilParams() { partialDirichlet = 0; };
};
template<class vobj>
class SimpleCompressor {
// Compressors will inherit buffer management policies
// Standard comms buffer management
class FaceGatherSimple
{
public:
static int PartialCompressionFactor(GridBase *grid) {return 1;};
// Decompress is after merge so ok
template<class vobj,class cobj,class compressor>
static void Gather_plane_simple (commVector<std::pair<int,int> >& table,
const Lattice<vobj> &rhs,
cobj *buffer,
compressor &compress,
int off,int so,int partial)
{
int num=table.size();
std::pair<int,int> *table_v = & table[0];
auto rhs_v = rhs.View(AcceleratorRead);
accelerator_forNB( i,num, vobj::Nsimd(), {
compress.Compress(buffer[off+table_v[i].first],rhs_v[so+table_v[i].second]);
});
rhs_v.ViewClose();
}
template<class vobj,class cobj,class compressor>
static void Gather_plane_exchange(commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
Vector<cobj *> pointers,int dimension,int plane,int cbmask,
compressor &compress,int type,int partial)
{
assert( (table.size()&0x1)==0);
int num=table.size()/2;
int so = plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane
auto rhs_v = rhs.View(AcceleratorRead);
auto p0=&pointers[0][0];
auto p1=&pointers[1][0];
auto tp=&table[0];
auto rhs_p = &rhs_v[0];
accelerator_forNB(j, num, vobj::Nsimd(), {
compress.CompressExchange(p0[j],p1[j],
rhs_p[so+tp[2*j ].second],
rhs_p[so+tp[2*j+1].second],
type);
});
rhs_v.ViewClose();
}
template<class decompressor,class Decompression>
static void DecompressFace(decompressor decompress,Decompression &dd)
{
auto kp = dd.kernel_p;
auto mp = dd.mpi_p;
accelerator_forNB(o,dd.buffer_size,1,{
decompress.Decompress(kp[o],mp[o]);
});
}
template<class decompressor,class Merger>
static void MergeFace(decompressor decompress,Merger &mm)
{
auto mp = &mm.mpointer[0];
auto vp0= &mm.vpointers[0][0];
auto vp1= &mm.vpointers[1][0];
auto type= mm.type;
accelerator_forNB(o,mm.buffer_size/2,vobj::Nsimd(),{
decompress.Exchange(mp[2*o],mp[2*o+1],vp0[o],vp1[o],type);
});
}
};
////////////////////////////////////
// Wilson compressor will add alternate policies for Dirichlet
// and possibly partial Dirichlet for DWF
////////////////////////////////////
/*
class FaceGatherDirichlet
{
// If it's dirichlet we don't assemble comms buffers
//
// Rely on zeroes in gauge field to drive the correct result
// NAN propgagation: field will locally wrap, so fermion should NOT contain NAN and just permute
template<class vobj,class cobj,class compressor>
static void Gather_plane_simple (commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,cobj *buffer,compressor &compress, int off,int so){};
template<class vobj,class cobj,class compressor>
static void Gather_plane_exchange(commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
Vector<cobj *> pointers,int dimension,int plane,int cbmask,
compressor &compress,int type) {}
template<class decompressor,class Merger>
static void Merge(decompressor decompress,Merge &mm) { }
template<class decompressor,class Decompression>
static void Decompress(decompressor decompress,Decompression &dd) {}
};
*/
template<class vobj,class FaceGather>
class SimpleCompressorGather : public FaceGather {
public:
void Point(int) {};
accelerator_inline int CommDatumSize(void) const { return sizeof(vobj); }
@ -30,20 +112,19 @@ public:
accelerator_inline void Compress(vobj &buf,const vobj &in) const {
coalescedWrite(buf,coalescedRead(in));
}
accelerator_inline void Exchange(vobj *mp,vobj *vp0,vobj *vp1,Integer type,Integer o) const {
accelerator_inline void Exchange(vobj &mp0,vobj &mp1,vobj &vp0,vobj &vp1,Integer type) const {
#ifdef GRID_SIMT
exchangeSIMT(mp[2*o],mp[2*o+1],vp0[o],vp1[o],type);
exchangeSIMT(mp0,mp1,vp0,vp1,type);
#else
exchange(mp[2*o],mp[2*o+1],vp0[o],vp1[o],type);
exchange(mp0,mp1,vp0,vp1,type);
#endif
}
accelerator_inline void Decompress(vobj *out,vobj *in, int o) const { assert(0); }
accelerator_inline void CompressExchange(vobj *out0,vobj *out1,const vobj *in,
int j,int k, int m,int type) const {
accelerator_inline void Decompress(vobj &out,vobj &in) const { };
accelerator_inline void CompressExchange(vobj &out0,vobj &out1,const vobj &in0,const vobj &in1,int type) const {
#ifdef GRID_SIMT
exchangeSIMT(out0[j],out1[j],in[k],in[m],type);
exchangeSIMT(out0,out1,in0,in1,type);
#else
exchange(out0[j],out1[j],in[k],in[m],type);
exchange(out0,out1,in0,in1,type);
#endif
}
// For cshift. Cshift should drop compressor coupling altogether
@ -52,11 +133,18 @@ public:
return arg;
}
};
class SimpleStencilParams{
public:
Coordinate dirichlet;
SimpleStencilParams() {};
};
// Standard compressor never needs dirichlet.
//
// Get away with a local period wrap and rely on dirac operator to use a zero gauge link as it is faster
//
// Compressors that inherit Dirichlet and Non-dirichlet behaviour.
//
// Currently run-time behaviour through StencilParameters paramaters, p.dirichlet
// combined with the FaceGatherSimple behaviour
template <class vobj> using SimpleCompressor = SimpleCompressorGather<vobj,FaceGatherSimple>;
//template <class vobj> using SimpleCompressorDirichlet = SimpleCompressorGather<vobj,FaceGatherDirichlet>;
NAMESPACE_END(Grid);