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

Move looping over "s" and "site" into kernels for GPU optimisatoin

Browse Source
This commit is contained in:
paboyle 2018-06-27 21:29:43 +01:00
parent 31f65beac8
commit 3e947527cb
1 changed files with 60 additions and 24 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

84
lib/qcd/action/fermion/WilsonKernels.h
View File

@ -51,11 +51,66 @@ public:
INHERIT_IMPL_TYPES(Impl); INHERIT_IMPL_TYPES(Impl);
public: public:
static void Dhop(int Opt,StencilImpl &st, DoubledGaugeField &U, SiteHalfSpinor * buf,
int Ls, int Nsite, const FermionField &in, FermionField &out,
int interior=1,int exterior=1)
{
auto U_v = U.View();
auto in_v = in.View();
auto out_v = out.View();
auto st_v = st.View();
if ( (Opt == WilsonKernelsStatic::OptGpu) && interior && exterior ) {
const uint64_t nsimd = Simd::Nsimd();
const uint64_t NN = Nsite*Ls*Simd::Nsimd();
accelerator_loopN( sss, NN, {
uint64_t cur = sss;
/* uint64_t lane = cur % nsimd; */ cur = cur / nsimd;
uint64_t sF = cur; cur = cur / Ls;
uint64_t sU = cur;
WilsonKernels<Impl>::GpuDhopSite(st_v,U_v,buf,sF,sU,in_v,out_v);
});
} else {
accelerator_loop( ss, U_v, {
int sU = ss;
int sF = Ls * sU;
DhopSite(Opt,st_v,U_v,st.CommBuf(),sF,sU,Ls,1,in_v,out_v);
});
}
}
static void DhopDag(int Opt,StencilImpl &st, DoubledGaugeField &U, SiteHalfSpinor * buf,
int Ls, int Nsite, const FermionField &in, FermionField &out,
int interior=1,int exterior=1)
{
auto U_v = U.View();
auto in_v = in.View();
auto out_v = out.View();
auto st_v = st.View();
if ( (Opt == WilsonKernelsStatic::OptGpu) && interior && exterior ) {
const uint64_t nsimd = Simd::Nsimd();
const uint64_t NN = Nsite*Ls*Simd::Nsimd();
accelerator_loopN( sss, NN, {
uint64_t cur = sss;
/* uint64_t lane = cur % nsimd; */ cur = cur / nsimd;
uint64_t sF = cur; cur = cur / Ls;
uint64_t sU = cur;
WilsonKernels<Impl>::GpuDhopSiteDag(st_v,U_v,buf,sF,sU,in_v,out_v);
});
} else {
accelerator_loop( ss, U_v, {
int sU = ss;
int sF = Ls * sU;
DhopSiteDag(Opt,st,U_v,st.CommBuf(),sF,sU,Ls,1,in_v,out_v);
});
}
}
template <bool EnableBool = true> static accelerator template <bool EnableBool = true> static accelerator
typename std::enable_if<Impl::Dimension == 3 && Nc == 3 &&EnableBool, void>::type typename std::enable_if<Impl::Dimension == 3 && Nc == 3 &&EnableBool, void>::type
DhopSite(int Opt,StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor * buf, DhopSite(int Opt,StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
int sF, int sU, int Ls, int Nsite, const FermionFieldView &in, FermionFieldView &out,int interior=1,int exterior=1) int sF, int sU, int Ls, int Nsite,
const FermionFieldView &in, FermionFieldView &out,int interior=1,int exterior=1)
{ {
//bgq_l1p_optimisation(1); //bgq_l1p_optimisation(1);
switch(Opt) { switch(Opt) {
@ -68,15 +123,6 @@ public:
else assert(0); else assert(0);
break; break;
#endif #endif
case WilsonKernelsStatic::OptGpu:
for (int site = 0; site < Nsite; site++) {
if(interior&&exterior) WilsonKernels<Impl>::GpuDhopSite(st,U,buf,sF,Ls,sU,in,out);
else if (interior) for (int s = 0; s < Ls; s++) WilsonKernels<Impl>::GenericDhopSiteInt(st,U,buf,sF,sU,in,out);
else if (exterior) for (int s = 0; s < Ls; s++) WilsonKernels<Impl>::GenericDhopSiteExt(st,U,buf,sF,sU,in,out);
else assert(0);
sU++;
}
break;
case WilsonKernelsStatic::OptHandUnroll: case WilsonKernelsStatic::OptHandUnroll:
for (int site = 0; site < Nsite; site++) { for (int site = 0; site < Nsite; site++) {
for (int s = 0; s < Ls; s++) { for (int s = 0; s < Ls; s++) {
@ -88,6 +134,7 @@ public:
sU++; sU++;
} }
break; break;
case WilsonKernelsStatic::OptGpu:
case WilsonKernelsStatic::OptGeneric: case WilsonKernelsStatic::OptGeneric:
for (int site = 0; site < Nsite; site++) { for (int site = 0; site < Nsite; site++) {
for (int s = 0; s < Ls; s++) { for (int s = 0; s < Ls; s++) {
@ -138,18 +185,6 @@ public:
else assert(0); else assert(0);
break; break;
#endif #endif
case WilsonKernelsStatic::OptGpu:
for (int site = 0; site < Nsite; site++) {
for (int s = 0; s < Ls; s++) {
if(interior&&exterior) WilsonKernels<Impl>::GenericDhopSiteDag(st,U,buf,sF,sU,in,out);
else if (interior) WilsonKernels<Impl>::GenericDhopSiteDagInt(st,U,buf,sF,sU,in,out);
else if (exterior) WilsonKernels<Impl>::GenericDhopSiteDagExt(st,U,buf,sF,sU,in,out);
else assert(0);
sF++;
}
sU++;
}
break;
case WilsonKernelsStatic::OptHandUnroll: case WilsonKernelsStatic::OptHandUnroll:
for (int site = 0; site < Nsite; site++) { for (int site = 0; site < Nsite; site++) {
for (int s = 0; s < Ls; s++) { for (int s = 0; s < Ls; s++) {
@ -162,6 +197,7 @@ public:
sU++; sU++;
} }
break; break;
case WilsonKernelsStatic::OptGpu:
case WilsonKernelsStatic::OptGeneric: case WilsonKernelsStatic::OptGeneric:
for (int site = 0; site < Nsite; site++) { for (int site = 0; site < Nsite; site++) {
for (int s = 0; s < Ls; s++) { for (int s = 0; s < Ls; s++) {
@ -238,10 +274,10 @@ public:
private: private:
// Specialised variants // Specialised variants
static accelerator void GpuDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor * buf, static accelerator void GpuDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
int sF, int LLs, int sU, const FermionFieldView &in, FermionFieldView &out); int sF, int sU, const FermionFieldView &in, FermionFieldView &out);
static accelerator void GpuDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor * buf, static accelerator void GpuDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
int sF, int LLs, int sU, const FermionFieldView &in, FermionFieldView &out); int sF, int sU, const FermionFieldView &in, FermionFieldView &out);
static accelerator void GenericDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor * buf, static accelerator void GenericDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
int sF, int sU, const FermionFieldView &in, FermionFieldView &out); int sF, int sU, const FermionFieldView &in, FermionFieldView &out);