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

No half prec comms; coalesced access on GPU

Browse Source
This commit is contained in:
Peter Boyle 2021-09-14 05:04:56 +01:00
parent 0fc662bb24
commit 7440cde92f
1 changed files with 51 additions and 47 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

98
Grid/qcd/action/fermion/WilsonCompressor.h
View File

@ -68,11 +68,12 @@ public:
/*****************************************************/
/* Compress includes precision change if mpi data is not same */
/*****************************************************/
template<class _SiteHalfSpinor, class _SiteSpinor>
accelerator_inline void Compress(_SiteHalfSpinor *buf,Integer o,const _SiteSpinor &in) const {
_SiteHalfSpinor tmp;
projector::Proj(tmp,in,mu,dag);
vstream(buf[o],tmp);
accelerator_inline void Compress(SiteHalfSpinor &buf,const SiteSpinor &in) const {
typedef decltype(coalescedRead(buf)) sobj;
sobj sp;
auto sin = coalescedRead(in);
projector::Proj(sp,sin,mu,dag);
coalescedWrite(buf,sp);
}
/*****************************************************/
@ -82,13 +83,18 @@ public:
const SiteHalfSpinor * __restrict__ vp0,
const SiteHalfSpinor * __restrict__ vp1,
Integer type,Integer o) const {
#ifdef GRID_SIMT
exchangeSIMT(mp[2*o],mp[2*o+1],vp0[o],vp1[o],type);
#else
SiteHalfSpinor tmp1;
SiteHalfSpinor tmp2;
exchange(tmp1,tmp2,vp0[o],vp1[o],type);
vstream(mp[2*o ],tmp1);
vstream(mp[2*o+1],tmp2);
#endif
}
/*****************************************************/
/* Have a decompression step if mpi data is not same */
/*****************************************************/
@ -105,6 +111,28 @@ public:
const SiteSpinor * __restrict__ in,
Integer j,Integer k, Integer m,Integer type) const
{
#ifdef GRID_SIMT
typedef SiteSpinor vobj;
typedef SiteHalfSpinor hvobj;
typedef decltype(coalescedRead(*in)) sobj;
typedef decltype(coalescedRead(*out0)) hsobj;
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 *vp0 = &in[k];
const vobj *vp1 = &in[m];
const vobj *vp = (lane&mask) ? vp1:vp0;
auto sa = coalescedRead(*vp,j0);
auto sb = coalescedRead(*vp,j1);
hsobj psa, psb;
projector::Proj(psa,sa,mu,dag);
projector::Proj(psb,sb,mu,dag);
coalescedWrite(out0[j],psa);
coalescedWrite(out1[j],psb);
#else
SiteHalfSpinor temp1, temp2;
SiteHalfSpinor temp3, temp4;
projector::Proj(temp1,in[k],mu,dag);
@ -112,6 +140,7 @@ public:
exchange(temp3,temp4,temp1,temp2,type);
vstream(out0[j],temp3);
vstream(out1[j],temp4);
#endif
}
/*****************************************************/
@ -121,6 +150,7 @@ public:
};
#if 0
template<class _HCspinor,class _Hspinor,class _Spinor, class projector>
class WilsonCompressorTemplate< _HCspinor, _Hspinor, _Spinor, projector,
typename std::enable_if<!std::is_same<_HCspinor,_Hspinor>::value>::type >
@ -149,13 +179,23 @@ public:
/*****************************************************/
/* Compress includes precision change if mpi data is not same */
/*****************************************************/
template<class _SiteHalfSpinor, class _SiteSpinor>
accelerator_inline void Compress(_SiteHalfSpinor *buf,Integer o,const _SiteSpinor &in) const {
_SiteHalfSpinor hsp;
accelerator_inline void Compress(SiteHalfSpinor &buf,const SiteSpinor &in) const {
SiteHalfSpinor hsp;
SiteHalfCommSpinor *hbuf = (SiteHalfCommSpinor *)buf;
projector::Proj(hsp,in,mu,dag);
precisionChange((vComplexLow *)&hbuf[o],(vComplexHigh *)&hsp,Nw);
}
accelerator_inline void Compress(SiteHalfSpinor &buf,const SiteSpinor &in) const {
#ifdef GRID_SIMT
typedef decltype(coalescedRead(buf)) sobj;
sobj sp;
auto sin = coalescedRead(in);
projector::Proj(sp,sin,mu,dag);
coalescedWrite(buf,sp);
#else
projector::Proj(buf,in,mu,dag);
#endif
}
/*****************************************************/
/* Exchange includes precision change if mpi data is not same */
@ -203,6 +243,7 @@ public:
accelerator_inline bool DecompressionStep(void) const { return true; }
};
#endif
#define DECLARE_PROJ(Projector,Compressor,spProj) \
class Projector { \
@ -253,33 +294,8 @@ public:
typedef typename Base::View_type View_type;
typedef typename Base::StencilVector StencilVector;
double timer0;
double timer1;
double timer2;
double timer3;
double timer4;
double timer5;
double timer6;
uint64_t callsi;
void ZeroCountersi(void)
{
timer0=0;
timer1=0;
timer2=0;
timer3=0;
timer4=0;
timer5=0;
timer6=0;
callsi=0;
}
void Reporti(int calls)
{
if ( timer0 ) std::cout << GridLogMessage << " timer0 (HaloGatherOpt) " <<timer0/calls <<std::endl;
if ( timer1 ) std::cout << GridLogMessage << " timer1 (Communicate) " <<timer1/calls <<std::endl;
if ( timer2 ) std::cout << GridLogMessage << " timer2 (CommsMerge ) " <<timer2/calls <<std::endl;
if ( timer3 ) std::cout << GridLogMessage << " timer3 (commsMergeShm) " <<timer3/calls <<std::endl;
if ( timer4 ) std::cout << GridLogMessage << " timer4 " <<timer4 <<std::endl;
}
void ZeroCountersi(void) { }
void Reporti(int calls) { }
std::vector<int> surface_list;
@ -321,26 +337,18 @@ public:
{
std::vector<std::vector<CommsRequest_t> > reqs;
this->HaloExchangeOptGather(source,compress);
double t1=usecond();
// Asynchronous MPI calls multidirectional, Isend etc...
// Non-overlapped directions within a thread. Asynchronous calls except MPI3, threaded up to comm threads ways.
this->Communicate();
double t2=usecond(); timer1 += t2-t1;
this->CommsMerge(compress);
double t3=usecond(); timer2 += t3-t2;
this->CommsMergeSHM(compress);
double t4=usecond(); timer3 += t4-t3;
}
template <class compressor>
void HaloExchangeOptGather(const Lattice<vobj> &source,compressor &compress)
{
this->Prepare();
double t0=usecond();
this->HaloGatherOpt(source,compress);
double t1=usecond();
timer0 += t1-t0;
callsi++;
}
template <class compressor>
@ -352,12 +360,9 @@ public:
typedef typename compressor::SiteHalfSpinor SiteHalfSpinor;
typedef typename compressor::SiteHalfCommSpinor SiteHalfCommSpinor;
this->mpi3synctime_g-=usecond();
this->_grid->StencilBarrier();
this->mpi3synctime_g+=usecond();
assert(source.Grid()==this->_grid);
this->halogtime-=usecond();
this->u_comm_offset=0;
@ -393,7 +398,6 @@ public:
}
this->face_table_computed=1;
assert(this->u_comm_offset==this->_unified_buffer_size);
this->halogtime+=usecond();
accelerator_barrier();
}