portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-19 16:27:05 +01:00
Code Releases Activity

Make view specify where and drive data motion - first cut.

Browse Source 
This is a compile tiime option --enable-unified=yes/no
This commit is contained in:
Peter Boyle
2020-05-21 16:13:16 -04:00
parent ebb60330c9
commit 7860a50f70
48 changed files with 688 additions and 718 deletions
Download Patch File Download Diff File Expand all files Collapse all files

50
Grid/qcd/utils/A2Autils.h
View File

@ -185,13 +185,13 @@ void A2Autils<FImpl>::MesonField(TensorType &mat,
for(int i=0;i<Lblock;i++){
auto lhs_v = lhs_wi[i].View();
auto lhs_v = lhs_wi[i].View(CpuRead);
auto left = conjugate(lhs_v[ss]);
for(int j=0;j<Rblock;j++){
SpinMatrix_v vv;
auto rhs_v = rhs_vj[j].View();
auto rhs_v = rhs_vj[j].View(CpuRead);
auto right = rhs_v[ss];
for(int s1=0;s1<Ns;s1++){
for(int s2=0;s2<Ns;s2++){
@ -204,7 +204,7 @@ void A2Autils<FImpl>::MesonField(TensorType &mat,
int base = Nmom*i+Nmom*Lblock*j+Nmom*Lblock*Rblock*r;
for ( int m=0;m<Nmom;m++){
int idx = m+base;
auto mom_v = mom[m].View();
auto mom_v = mom[m].View(CpuRead);
auto phase = mom_v[ss];
mac(&lvSum[idx],&vv,&phase);
}
@ -371,7 +371,7 @@ void A2Autils<FImpl>::PionFieldXX(Eigen::Tensor<ComplexD,3> &mat,
for(int i=0;i<Lblock;i++){
auto wi_v = wi[i].View();
auto wi_v = wi[i].View(CpuRead);
auto w = conjugate(wi_v[ss]);
if (g5) {
w()(2)(0) = - w()(2)(0);
@ -383,7 +383,7 @@ void A2Autils<FImpl>::PionFieldXX(Eigen::Tensor<ComplexD,3> &mat,
}
for(int j=0;j<Rblock;j++){
auto vj_v=vj[j].View();
auto vj_v=vj[j].View(CpuRead);
auto v = vj_v[ss];
auto vv = v()(0)(0);
@ -518,12 +518,12 @@ void A2Autils<FImpl>::PionFieldWVmom(Eigen::Tensor<ComplexD,4> &mat,
for(int i=0;i<Lblock;i++){
auto wi_v = wi[i].View();
auto wi_v = wi[i].View(CpuRead);
auto w = conjugate(wi_v[ss]);
for(int j=0;j<Rblock;j++){
auto vj_v = vj[j].View();
auto vj_v = vj[j].View(CpuRead);
auto v = vj_v[ss];
auto vv = w()(0)(0) * v()(0)(0)// Gamma5 Dirac basis explicitly written out
@ -544,7 +544,7 @@ void A2Autils<FImpl>::PionFieldWVmom(Eigen::Tensor<ComplexD,4> &mat,
int base = Nmom*i+Nmom*Lblock*j+Nmom*Lblock*Rblock*r;
for ( int m=0;m<Nmom;m++){
int idx = m+base;
auto mom_v = mom[m].View();
auto mom_v = mom[m].View(CpuRead);
auto phase = mom_v[ss];
mac(&lvSum[idx],&vv,&phase()()());
}
@ -730,13 +730,13 @@ void A2Autils<FImpl>::AslashField(TensorType &mat,
for(int i=0;i<Lblock;i++)
{
auto wi_v = lhs_wi[i].View();
auto wi_v = lhs_wi[i].View(CpuRead);
auto left = conjugate(wi_v[ss]);
for(int j=0;j<Rblock;j++)
{
SpinMatrix_v vv;
auto vj_v = rhs_vj[j].View();
auto vj_v = rhs_vj[j].View(CpuRead);
auto right = vj_v[ss];
for(int s1=0;s1<Ns;s1++)
@ -752,8 +752,8 @@ void A2Autils<FImpl>::AslashField(TensorType &mat,
for ( int m=0;m<Nem;m++)
{
auto emB0_v = emB0[m].View();
auto emB1_v = emB1[m].View();
auto emB0_v = emB0[m].View(CpuRead);
auto emB1_v = emB1[m].View(CpuRead);
int idx = m+base;
auto b0 = emB0_v[ss];
auto b1 = emB1_v[ss];
@ -1014,12 +1014,12 @@ A2Autils<FImpl>::ContractWWVV(std::vector<PropagatorField> &WWVV,
for(int d_o=0;d_o<N_d;d_o+=d_unroll){
for(int t=0;t<N_t;t++){
for(int s=0;s<N_s;s++){
auto vs_v = vs[s].View();
auto vs_v = vs[s].View(CpuRead);
auto tmp1 = vs_v[ss];
vobj tmp2 = Zero();
vobj tmp3 = Zero();
for(int d=d_o;d<MIN(d_o+d_unroll,N_d);d++){
auto vd_v = vd[d].View();
auto vd_v = vd[d].View(CpuRead);
Scalar_v coeff = WW_sd(t,s,d);
tmp3 = conjugate(vd_v[ss]);
mac(&tmp2, &coeff, &tmp3);
@ -1067,12 +1067,12 @@ A2Autils<FImpl>::ContractWWVV(std::vector<PropagatorField> &WWVV,
thread_for(ss,grid->oSites(),{
for(int d_o=0;d_o<N_d;d_o+=d_unroll){
for(int s=0;s<N_s;s++){
auto vs_v = vs[s].View();
auto vs_v = vs[s].View(CpuRead);
auto tmp1 = vs_v[ss];
vobj tmp2 = Zero();
vobj tmp3 = Zero();
for(int d=d_o;d<MIN(d_o+d_unroll,N_d);d++){
auto vd_v = vd[d].View();
auto vd_v = vd[d].View(CpuRead);
Scalar_v coeff = buf(s,d);
tmp3 = conjugate(vd_v[ss]);
mac(&tmp2, &coeff, &tmp3);
@ -1093,7 +1093,7 @@ inline void A2Autils<FImpl>::OuterProductWWVV(PropagatorField &WWVV,
const vobj &rhs,
const int Ns, const int ss)
{
auto WWVV_v = WWVV.View();
auto WWVV_v = WWVV.View(CpuWrite);
for (int s1 = 0; s1 < Ns; s1++){
for (int s2 = 0; s2 < Ns; s2++){
WWVV_v[ss]()(s1,s2)(0, 0) += lhs()(s1)(0) * rhs()(s2)(0);
@ -1122,10 +1122,10 @@ void A2Autils<FImpl>::ContractFourQuarkColourDiagonal(const PropagatorField &WWV
GridBase *grid = WWVV0.Grid();
auto WWVV0_v = WWVV0.View();
auto WWVV1_v = WWVV1.View();
auto O_trtr_v= O_trtr.View();
auto O_fig8_v= O_fig8.View();
auto WWVV0_v = WWVV0.View(CpuRead);
auto WWVV1_v = WWVV1.View(CpuRead);
auto O_trtr_v= O_trtr.View(CpuWrite);
auto O_fig8_v= O_fig8.View(CpuWrite);
thread_for(ss,grid->oSites(),{
typedef typename ComplexField::vector_object vobj;
@ -1166,10 +1166,10 @@ void A2Autils<FImpl>::ContractFourQuarkColourMix(const PropagatorField &WWVV0,
GridBase *grid = WWVV0.Grid();
auto WWVV0_v = WWVV0.View();
auto WWVV1_v = WWVV1.View();
auto O_trtr_v= O_trtr.View();
auto O_fig8_v= O_fig8.View();
auto WWVV0_v = WWVV0.View(CpuRead);
auto WWVV1_v = WWVV1.View(CpuRead);
auto O_trtr_v= O_trtr.View(CpuWrite);
auto O_fig8_v= O_fig8.View(CpuWrite);
thread_for(ss,grid->oSites(),{

26
Grid/qcd/utils/BaryonUtils.h
View File

@ -273,10 +273,10 @@ void BaryonUtils<FImpl>::ContractBaryons(const PropagatorField &q1_left,
for (int ie=0; ie < 6 ; ie++)
wick_contraction[ie] = (quarks_left[0] == quarks_right[epsilon[ie][0]] && quarks_left[1] == quarks_right[epsilon[ie][1]] && quarks_left[2] == quarks_right[epsilon[ie][2]]) ? 1 : 0;
auto vbaryon_corr= baryon_corr.View();
auto v1 = q1_left.View();
auto v2 = q2_left.View();
auto v3 = q3_left.View();
auto vbaryon_corr= baryon_corr.View(CpuWrite);
auto v1 = q1_left.View(CpuRead);
auto v2 = q2_left.View(CpuRead);
auto v3 = q3_left.View(CpuRead);
// accelerator_for(ss, grid->oSites(), grid->Nsimd(), {
thread_for(ss,grid->oSites(),{
@ -560,10 +560,10 @@ void BaryonUtils<FImpl>::Sigma_to_Nucleon_Eye(const PropagatorField &qq_loop,
{
GridBase *grid = qs_ti.Grid();
auto vcorr= stn_corr.View();
auto vq_loop = qq_loop.View();
auto vd_tf = qd_tf.View();
auto vs_ti = qs_ti.View();
auto vcorr= stn_corr.View(CpuWrite);
auto vq_loop = qq_loop.View(CpuRead);
auto vd_tf = qd_tf.View(CpuRead);
auto vs_ti = qs_ti.View(CpuRead);
// accelerator_for(ss, grid->oSites(), grid->Nsimd(), {
thread_for(ss,grid->oSites(),{
@ -597,11 +597,11 @@ void BaryonUtils<FImpl>::Sigma_to_Nucleon_NonEye(const PropagatorField &qq_ti,
{
GridBase *grid = qs_ti.Grid();
auto vcorr= stn_corr.View();
auto vq_ti = qq_ti.View();
auto vq_tf = qq_tf.View();
auto vd_tf = qd_tf.View();
auto vs_ti = qs_ti.View();
auto vcorr= stn_corr.View(CpuWrite);
auto vq_ti = qq_ti.View(CpuRead);
auto vq_tf = qq_tf.View(CpuRead);
auto vd_tf = qd_tf.View(CpuRead);
auto vs_ti = qs_ti.View(CpuRead);
// accelerator_for(ss, grid->oSites(), grid->Nsimd(), {
thread_for(ss,grid->oSites(),{

48
Grid/qcd/utils/LinalgUtils.h
View File

@ -47,8 +47,8 @@ void axpibg5x(Lattice<vobj> &z,const Lattice<vobj> &x,Coeff a,Coeff b)
GridBase *grid=x.Grid();
Gamma G5(Gamma::Algebra::Gamma5);
auto x_v = x.View();
auto z_v = z.View();
auto x_v = x.View(AcceleratorRead);
auto z_v = z.View(AcceleratorWrite);
accelerator_for( ss, x_v.size(),vobj::Nsimd(), {
auto tmp = a*x_v(ss) + G5*(b*timesI(x_v(ss)));
coalescedWrite(z_v[ss],tmp);
@ -63,9 +63,9 @@ void axpby_ssp(Lattice<vobj> &z, Coeff a,const Lattice<vobj> &x,Coeff b,const La
conformable(x,z);
GridBase *grid=x.Grid();
int Ls = grid->_rdimensions[0];
auto x_v = x.View();
auto y_v = y.View();
auto z_v = z.View();
auto x_v = x.View(AcceleratorRead);
auto y_v = y.View(AcceleratorRead);
auto z_v = z.View(AcceleratorWrite);
// FIXME -- need a new class of accelerator_loop to implement this
//
uint64_t nloop = grid->oSites()/Ls;
@ -85,9 +85,9 @@ void ag5xpby_ssp(Lattice<vobj> &z,Coeff a,const Lattice<vobj> &x,Coeff b,const L
GridBase *grid=x.Grid();
int Ls = grid->_rdimensions[0];
Gamma G5(Gamma::Algebra::Gamma5);
auto x_v = x.View();
auto y_v = y.View();
auto z_v = z.View();
auto x_v = x.View(AcceleratorRead);
auto y_v = y.View(AcceleratorRead);
auto z_v = z.View(AcceleratorWrite);
uint64_t nloop = grid->oSites()/Ls;
accelerator_for(sss,nloop,vobj::Nsimd(),{
uint64_t ss = sss*Ls;
@ -104,9 +104,9 @@ void axpbg5y_ssp(Lattice<vobj> &z,Coeff a,const Lattice<vobj> &x,Coeff b,const L
conformable(x,z);
GridBase *grid=x.Grid();
int Ls = grid->_rdimensions[0];
auto x_v = x.View();
auto y_v = y.View();
auto z_v = z.View();
auto x_v = x.View(AcceleratorRead);
auto y_v = y.View(AcceleratorRead);
auto z_v = z.View(AcceleratorWrite);
Gamma G5(Gamma::Algebra::Gamma5);
uint64_t nloop = grid->oSites()/Ls;
accelerator_for(sss,nloop,vobj::Nsimd(),{
@ -125,9 +125,9 @@ void ag5xpbg5y_ssp(Lattice<vobj> &z,Coeff a,const Lattice<vobj> &x,Coeff b,const
GridBase *grid=x.Grid();
int Ls = grid->_rdimensions[0];
auto x_v = x.View();
auto y_v = y.View();
auto z_v = z.View();
auto x_v = x.View(AcceleratorRead);
auto y_v = y.View(AcceleratorRead);
auto z_v = z.View(AcceleratorWrite);
Gamma G5(Gamma::Algebra::Gamma5);
uint64_t nloop = grid->oSites()/Ls;
accelerator_for(sss,nloop,vobj::Nsimd(),{
@ -147,9 +147,9 @@ void axpby_ssp_pminus(Lattice<vobj> &z,Coeff a,const Lattice<vobj> &x,Coeff b,co
GridBase *grid=x.Grid();
int Ls = grid->_rdimensions[0];
auto x_v = x.View();
auto y_v = y.View();
auto z_v = z.View();
auto x_v = x.View(AcceleratorRead);
auto y_v = y.View(AcceleratorRead);
auto z_v = z.View(AcceleratorWrite);
uint64_t nloop = grid->oSites()/Ls;
accelerator_for(sss,nloop,vobj::Nsimd(),{
uint64_t ss = sss*Ls;
@ -168,9 +168,9 @@ void axpby_ssp_pplus(Lattice<vobj> &z,Coeff a,const Lattice<vobj> &x,Coeff b,con
conformable(x,z);
GridBase *grid=x.Grid();
int Ls = grid->_rdimensions[0];
auto x_v = x.View();
auto y_v = y.View();
auto z_v = z.View();
auto x_v = x.View(AcceleratorRead);
auto y_v = y.View(AcceleratorRead);
auto z_v = z.View(AcceleratorWrite);
uint64_t nloop = grid->oSites()/Ls;
accelerator_for(sss,nloop,vobj::Nsimd(),{
uint64_t ss = sss*Ls;
@ -189,8 +189,8 @@ void G5R5(Lattice<vobj> &z,const Lattice<vobj> &x)
conformable(x,z);
int Ls = grid->_rdimensions[0];
Gamma G5(Gamma::Algebra::Gamma5);
auto x_v = x.View();
auto z_v = z.View();
auto x_v = x.View(AcceleratorRead);
auto z_v = z.View(AcceleratorWrite);
uint64_t nloop = grid->oSites()/Ls;
accelerator_for(sss,nloop,vobj::Nsimd(),{
uint64_t ss = sss*Ls;
@ -222,8 +222,8 @@ void G5C(Lattice<iVector<CComplex, nbasis>> &z, const Lattice<iVector<CComplex,
static_assert(nbasis % 2 == 0, "");
int nb = nbasis / 2;
auto z_v = z.View();
auto x_v = x.View();
auto z_v = z.View(AcceleratorWrite);
auto x_v = x.View(AcceleratorRead);
accelerator_for(ss,grid->oSites(),CComplex::Nsimd(),
{
for(int n = 0; n < nb; ++n) {

10
Grid/qcd/utils/SUn.h
View File

@ -222,9 +222,9 @@ public:
conformable(subgroup, Determinant);
int i0, i1;
su2SubGroupIndex(i0, i1, su2_index);
auto subgroup_v = subgroup.View();
auto source_v = source.View();
auto Determinant_v = Determinant.View();
auto subgroup_v = subgroup.View(CpuWrite);
auto source_v = source.View(CpuRead);
auto Determinant_v = Determinant.View(CpuWrite);
thread_for(ss, grid->oSites(), {
@ -257,8 +257,8 @@ public:
su2SubGroupIndex(i0, i1, su2_index);
dest = 1.0; // start out with identity
auto dest_v = dest.View();
auto subgroup_v = subgroup.View();
auto dest_v = dest.View(CpuWrite);
auto subgroup_v = subgroup.View(CpuRead);
thread_for(ss, grid->oSites(),
{
dest_v[ss]()()(i0, i0) = subgroup_v[ss]()()(0, 0);