portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-13 04:37:05 +01:00
Code Releases Activity

Dslash testing for reproduce

Browse Source
This commit is contained in:
Peter Boyle
2024-11-11 23:11:11 +00:00
parent b91fc1b6b4
commit 8fe429346f
5 changed files with 324 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
Grid/qcd/action/fermion/WilsonFermion5D.h
View File

@ -119,6 +119,9 @@ public:
void DhopOE(const FermionField &in, FermionField &out,int dag);
void DhopEO(const FermionField &in, FermionField &out,int dag);
void DhopComms (const FermionField &in, FermionField &out);
void DhopCalc (const FermionField &in, FermionField &out,uint64_t *ids);
// add a DhopComm
// -- suboptimal interface will presently trigger multiple comms.
void DhopDir(const FermionField &in, FermionField &out,int dir,int disp);

4
Grid/qcd/action/fermion/WilsonKernels.h
View File

@ -57,6 +57,10 @@ public:
int Ls, int Nsite, const FermionField &in, FermionField &out,
int interior=1,int exterior=1) ;
static void DhopKernel(int Opt,StencilImpl &st, DoubledGaugeField &U, SiteHalfSpinor * buf,
int Ls, int Nsite, const FermionField &in, FermionField &out,
uint64_t *ids);
static void DhopDagKernel(int Opt,StencilImpl &st, DoubledGaugeField &U, SiteHalfSpinor * buf,
int Ls, int Nsite, const FermionField &in, FermionField &out,
int interior=1,int exterior=1) ;

23
Grid/qcd/action/fermion/implementation/WilsonFermion5DImplementation.h
View File

@ -438,6 +438,29 @@ void WilsonFermion5D<Impl>::DhopEO(const FermionField &in, FermionField &out,int
DhopInternal(StencilOdd,UmuEven,in,out,dag);
}
template<class Impl>
void WilsonFermion5D<Impl>::DhopComms(const FermionField &in, FermionField &out)
{
int dag =0 ;
conformable(in.Grid(),FermionGrid()); // verifies full grid
conformable(in.Grid(),out.Grid());
out.Checkerboard() = in.Checkerboard();
Compressor compressor(dag);
Stencil.HaloExchangeOpt(in,compressor);
}
template<class Impl>
void WilsonFermion5D<Impl>::DhopCalc(const FermionField &in, FermionField &out,uint64_t *ids)
{
conformable(in.Grid(),FermionGrid()); // verifies full grid
conformable(in.Grid(),out.Grid());
out.Checkerboard() = in.Checkerboard();
int LLs = in.Grid()->_rdimensions[0];
int Opt = WilsonKernelsStatic::Opt;
Kernels::DhopKernel(Opt,Stencil,Umu,Stencil.CommBuf(),LLs,Umu.oSites(),in,out,ids);
}
template<class Impl>
void WilsonFermion5D<Impl>::Dhop(const FermionField &in, FermionField &out,int dag)
{

56
Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h
View File

@ -411,6 +411,46 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
#undef LoopBody
}
#ifdef GRID_SYCL
extern "C" {
ulong SYCL_EXTERNAL __attribute__((overloadable)) intel_get_cycle_counter( void );
uint SYCL_EXTERNAL __attribute__((overloadable)) intel_get_active_channel_mask( void );
uint SYCL_EXTERNAL __attribute__((overloadable)) intel_get_grf_register( uint reg );
uint SYCL_EXTERNAL __attribute__((overloadable)) intel_get_flag_register( uint flag );
uint SYCL_EXTERNAL __attribute__((overloadable)) intel_get_control_register( uint reg );
uint SYCL_EXTERNAL __attribute__((overloadable)) intel_get_hw_thread_id( void );
uint SYCL_EXTERNAL __attribute__((overloadable)) intel_get_slice_id( void );
uint SYCL_EXTERNAL __attribute__((overloadable)) intel_get_subslice_id( void );
uint SYCL_EXTERNAL __attribute__((overloadable)) intel_get_eu_id( void );
uint SYCL_EXTERNAL __attribute__((overloadable)) intel_get_eu_thread_id( void );
void SYCL_EXTERNAL __attribute__((overloadable)) intel_eu_thread_pause( uint value );
}
#ifdef GRID_SIMT
#define MAKE_ID(A) (intel_get_eu_id()<<16)|(intel_get_slice_id()<<8)|(intel_get_subslice_id())
#else
#define MAKE_ID(A) (0)
#endif
#else
#define MAKE_ID(A) (0)
#endif
#define KERNEL_CALL_ID(A) \
const uint64_t NN = Nsite*Ls; \
accelerator_forNB( ss, NN, Simd::Nsimd(), { \
int sF = ss; \
int sU = ss/Ls; \
WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,in_v,out_v); \
const int Nsimd = SiteHalfSpinor::Nsimd(); \
const int lane=acceleratorSIMTlane(Nsimd); \
int idx=sF*Nsimd+lane; \
uint64_t id = MAKE_ID(); \
ids[idx]=id; \
}); \
accelerator_barrier();
#define KERNEL_CALLNB(A) \
const uint64_t NN = Nsite*Ls; \
@ -418,7 +458,7 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
int sF = ss; \
int sU = ss/Ls; \
WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,in_v,out_v); \
});
});
#define KERNEL_CALL(A) KERNEL_CALLNB(A); accelerator_barrier();
@ -451,6 +491,8 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,Ls,1,in_v,out_v); \
});}
template <class Impl>
void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st, DoubledGaugeField &U, SiteHalfSpinor * buf,
int Ls, int Nsite, const FermionField &in, FermionField &out,
@ -485,6 +527,18 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st, DoubledGaugeField
}
assert(0 && " Kernel optimisation case not covered ");
}
template <class Impl>
void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st, DoubledGaugeField &U, SiteHalfSpinor * buf,
int Ls, int Nsite, const FermionField &in, FermionField &out,
uint64_t *ids)
{
autoView(U_v , U,AcceleratorRead);
autoView(in_v , in,AcceleratorRead);
autoView(out_v,out,AcceleratorWrite);
autoView(st_v , st,AcceleratorRead);
KERNEL_CALL_ID(GenericDhopSite);
}
template <class Impl>
void WilsonKernels<Impl>::DhopDagKernel(int Opt,StencilImpl &st, DoubledGaugeField &U, SiteHalfSpinor * buf,
int Ls, int Nsite, const FermionField &in, FermionField &out,