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

GPU dslash updates

Browse Source
This commit is contained in:
paboyle 2018-06-27 22:32:21 +01:00
parent f8e880b445
commit 3a50afe7e7
1 changed files with 34 additions and 43 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

77
lib/qcd/action/fermion/WilsonKernelsGpu.cc
View File

@ -41,9 +41,6 @@ __host__ __device__ inline void synchronise(void)
return; return;
} }
#define GPU_DSLASH_COALESCE
#ifdef GPU_DSLASH_COALESCE
__host__ __device__ inline int get_my_lanes(int Nsimd) __host__ __device__ inline int get_my_lanes(int Nsimd)
{ {
#ifdef __CUDA_ARCH__ #ifdef __CUDA_ARCH__
@ -118,40 +115,42 @@ typename vobj::scalar_object extractLaneFloat2(int lane, const vobj & __restrict
return extracted; return extracted;
} }
#define GPU_COALESCED_STENCIL_LEG_PROJ(Dir,spProj) \
#define GPU_COALESCED_STENCIL_LEG_PROJ(Dir,spProj) \ if (SE->_is_local) { \
if (SE->_is_local) { \ int mask = Nsimd >> (ptype + 1); \
auto in_l = extractLaneGpu(lane,in[SE->_offset]); \ int plane= lane; \
spProj(chi,in_l); \ if (SE->_permute) plane = (lane ^ mask); \
} else { \ auto in_l = extractLaneGpu(plane,in[SE->_offset]); \
chi = extractLaneGpu(lane,buf[SE->_offset]); \ spProj(chi,in_l); \
} else { \
chi = extractLaneGpu(lane,buf[SE->_offset]); \
} }
template <class Impl> template <class Impl>
accelerator void WilsonKernels<Impl>::GpuDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, accelerator void WilsonKernels<Impl>::GpuDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,
SiteHalfSpinor *buf, int sF,int LLs, SiteHalfSpinor *buf, int sF,
int sU, const FermionFieldView &in, FermionFieldView &out) int sU, const FermionFieldView &in, FermionFieldView &out)
{ {
typename SiteHalfSpinor::scalar_object tmp;
typename SiteHalfSpinor::scalar_object chi; typename SiteHalfSpinor::scalar_object chi;
typename SiteHalfSpinor::scalar_object Uchi; typename SiteHalfSpinor::scalar_object Uchi;
typename SiteSpinor::scalar_object result; typename SiteSpinor::scalar_object result;
typedef typename SiteSpinor::scalar_type scalar_type; typedef typename SiteSpinor::scalar_type scalar_type;
typedef typename SiteSpinor::vector_type vector_type; typedef typename SiteSpinor::vector_type vector_type;
constexpr int Nsimd = sizeof(vector_type)/sizeof(scalar_type); constexpr int Nsimd = sizeof(vector_type)/sizeof(scalar_type);
uint64_t lane_offset= get_my_lane_offset(Nsimd); uint64_t lane_offset= get_my_lane_offset(Nsimd);
uint64_t lanes = get_my_lanes (Nsimd); uint64_t lanes = get_my_lanes(Nsimd);
StencilEntry *SE; StencilEntry *SE;
int ptype; int ptype;
for(int lane = lane_offset;lane<lane_offset+lanes;lane++){ for(int lane = lane_offset;lane<lane_offset+lanes;lane++){
for(int s=0;s<LLs;s++){
for(int mu=0;mu<2*Nd;mu++) { for(int mu=0;mu<2*Nd;mu++) {
SE = st.GetEntry(ptype, mu, sF); SE = st.GetEntry(ptype, mu, sF);
switch(mu){ switch(mu){
case Xp: case Xp:
GPU_COALESCED_STENCIL_LEG_PROJ(Xp,spProjXp); break; GPU_COALESCED_STENCIL_LEG_PROJ(Xp,spProjXp); break;
@ -171,14 +170,12 @@ accelerator void WilsonKernels<Impl>::GpuDhopSiteDag(StencilView &st, DoubledGau
default: default:
GPU_COALESCED_STENCIL_LEG_PROJ(Tm,spProjTm); break; GPU_COALESCED_STENCIL_LEG_PROJ(Tm,spProjTm); break;
} }
synchronise();
auto U_l = extractLaneGpu(lane,U[sU](mu)); Impl::multLinkGpu(lane,Uchi,U[sU],chi,mu);
Uchi()=U_l*chi();
switch(mu){ switch(mu){
case Xp: case Xp:
spReconXp(result, Uchi); break; spReconXp(result, Uchi); break;
case Yp: case Yp:
accumReconYp(result, Uchi); break; accumReconYp(result, Uchi); break;
case Zp: case Zp:
@ -195,19 +192,16 @@ accelerator void WilsonKernels<Impl>::GpuDhopSiteDag(StencilView &st, DoubledGau
default: default:
accumReconTm(result, Uchi); break; accumReconTm(result, Uchi); break;
} }
synchronise();
} }
insertLane (lane,out[sF],result); insertLaneFloat2 (lane,out[sF],result);
sF++; }
}} }
};
template <class Impl> template <class Impl>
accelerator void WilsonKernels<Impl>::GpuDhopSite(StencilView &st, DoubledGaugeFieldView &U, accelerator void WilsonKernels<Impl>::GpuDhopSite(StencilView &st, DoubledGaugeFieldView &U,
SiteHalfSpinor *buf, int sF,int LLs, SiteHalfSpinor *buf, int sF,
int sU, const FermionFieldView &in, FermionFieldView &out) int sU, const FermionFieldView &in, FermionFieldView &out)
{ {
typename SiteHalfSpinor::scalar_object tmp;
typename SiteHalfSpinor::scalar_object chi; typename SiteHalfSpinor::scalar_object chi;
typename SiteHalfSpinor::scalar_object Uchi; typename SiteHalfSpinor::scalar_object Uchi;
typename SiteSpinor::scalar_object result; typename SiteSpinor::scalar_object result;
@ -219,14 +213,11 @@ accelerator void WilsonKernels<Impl>::GpuDhopSite(StencilView &st, DoubledGaugeF
uint64_t lane_offset= get_my_lane_offset(Nsimd); uint64_t lane_offset= get_my_lane_offset(Nsimd);
uint64_t lanes = get_my_lanes(Nsimd); uint64_t lanes = get_my_lanes(Nsimd);
// printf("Evaluating site %d Nsimd %d : lanes %ld %ld - %ld\n",sF,Nsimd,lanes,lane_offset,lane_offset+lanes-1);
StencilEntry *SE; StencilEntry *SE;
int ptype; int ptype;
for(int lane = lane_offset;lane<lane_offset+lanes;lane++){ for(int lane = lane_offset;lane<lane_offset+lanes;lane++){
for(int s=0;s<LLs;s++){ #if 0
#if 1
int mu=0; int mu=0;
SE = st.GetEntry(ptype, mu, sF); SE = st.GetEntry(ptype, mu, sF);
GPU_COALESCED_STENCIL_LEG_PROJ(Xp,spProjXm); GPU_COALESCED_STENCIL_LEG_PROJ(Xp,spProjXm);
@ -294,11 +285,7 @@ accelerator void WilsonKernels<Impl>::GpuDhopSite(StencilView &st, DoubledGaugeF
GPU_COALESCED_STENCIL_LEG_PROJ(Tm,spProjTp); break; GPU_COALESCED_STENCIL_LEG_PROJ(Tm,spProjTp); break;
} }
auto U_l = extractLaneGpu(lane,U[sU](mu)); Impl::multLinkGpu(lane,Uchi,U[sU],chi,mu);
auto tmp = U_l * chi();
Uchi() = tmp;
switch(mu){ switch(mu){
case Xp: case Xp:
@ -322,8 +309,7 @@ accelerator void WilsonKernels<Impl>::GpuDhopSite(StencilView &st, DoubledGaugeF
} }
#endif #endif
insertLaneFloat2 (lane,out[sF],result); insertLaneFloat2 (lane,out[sF],result);
sF++; }
}}
}; };
// Template specialise Gparity to empty for now // Template specialise Gparity to empty for now
@ -332,23 +318,25 @@ accelerator void WilsonKernels<Impl>::GpuDhopSite(StencilView &st, DoubledGaugeF
accelerator void \ accelerator void \
WilsonKernels<A>::GpuDhopSite(StencilView &st, \ WilsonKernels<A>::GpuDhopSite(StencilView &st, \
DoubledGaugeFieldView &U, \ DoubledGaugeFieldView &U, \
SiteHalfSpinor *buf, int sF, int LLs, \ SiteHalfSpinor *buf, int sF, \
int sU, \ int sU, \
const FermionFieldView &in, \ const FermionFieldView &in, \
FermionFieldView &out) {}; \ FermionFieldView &out) { assert(0);}; \
template <> \ template <> \
accelerator void \ accelerator void \
WilsonKernels<A>::GpuDhopSiteDag(StencilView &st, \ WilsonKernels<A>::GpuDhopSiteDag(StencilView &st, \
DoubledGaugeFieldView &U, \ DoubledGaugeFieldView &U, \
SiteHalfSpinor *buf, int sF,int LLs, \ SiteHalfSpinor *buf, int sF, \
int sU, \ int sU, \
const FermionFieldView &in, \ const FermionFieldView &in, \
FermionFieldView &out) {}; FermionFieldView &out) { assert(0);};
GPU_EMPTY(GparityWilsonImplF); GPU_EMPTY(GparityWilsonImplF);
GPU_EMPTY(GparityWilsonImplFH); GPU_EMPTY(GparityWilsonImplFH);
GPU_EMPTY(GparityWilsonImplD); GPU_EMPTY(GparityWilsonImplD);
GPU_EMPTY(GparityWilsonImplDF); GPU_EMPTY(GparityWilsonImplDF);
/*
GPU_EMPTY(DomainWallVec5dImplF); GPU_EMPTY(DomainWallVec5dImplF);
GPU_EMPTY(DomainWallVec5dImplFH); GPU_EMPTY(DomainWallVec5dImplFH);
GPU_EMPTY(DomainWallVec5dImplD); GPU_EMPTY(DomainWallVec5dImplD);
@ -357,8 +345,11 @@ GPU_EMPTY(ZDomainWallVec5dImplF);
GPU_EMPTY(ZDomainWallVec5dImplFH); GPU_EMPTY(ZDomainWallVec5dImplFH);
GPU_EMPTY(ZDomainWallVec5dImplD); GPU_EMPTY(ZDomainWallVec5dImplD);
GPU_EMPTY(ZDomainWallVec5dImplDF); GPU_EMPTY(ZDomainWallVec5dImplDF);
*/
FermOpTemplateInstantiate(WilsonKernels); FermOpTemplateInstantiate(WilsonKernels);
AdjointFermOpTemplateInstantiate(WilsonKernels);
TwoIndexFermOpTemplateInstantiate(WilsonKernels);
NAMESPACE_END(Grid); NAMESPACE_END(Grid);