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

Merge pull request #4 from paboyle/develop

Browse Source 
merge
This commit is contained in:
Christoph Lehner 2020-05-11 20:59:29 +02:00 committed by GitHub
commit b1c86900b2
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
16 changed files with 582 additions and 452 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
Grid/algorithms/LinearOperator.h
View File

@ -257,13 +257,11 @@ public:
virtual RealD Mpc (const Field &in, Field &out) {
Field tmp(in.Grid());
tmp.Checkerboard() = !in.Checkerboard();
//std::cout <<"grid pointers: in._grid="<< in._grid << " out._grid=" << out._grid << " _Mat.Grid=" << _Mat.Grid() << " _Mat.RedBlackGrid=" << _Mat.RedBlackGrid() << std::endl;
_Mat.Meooe(in,tmp);
_Mat.MooeeInv(tmp,out);
_Mat.Meooe(out,tmp);
//std::cout << "cb in " << in.Checkerboard() << " cb out " << out.Checkerboard() << std::endl;
_Mat.Mooee(in,out);
return axpy_norm(out,-1.0,tmp,out);
}
@ -366,6 +364,9 @@ public:
void OpDir(const Field& in, Field& out, int dir, int disp) {
assert(0);
}
void OpDirAll(const Field& in, std::vector<Field>& out){
assert(0);
};
};
template<class Matrix, class Field>

24
Grid/algorithms/approx/Chebyshev.h
View File

@ -234,10 +234,8 @@ public:
GridBase *grid=in.Grid();
// std::cout << "Chevyshef(): in.Grid()="<<in.Grid()<<std::endl;
//std::cout <<" Linop.Grid()="<<Linop.Grid()<<"Linop.RedBlackGrid()="<<Linop.RedBlackGrid()<<std::endl;
int vol=grid->gSites();
typedef typename Field::vector_type vector_type;
Field T0(grid); T0 = in;
Field T1(grid);
@ -260,12 +258,26 @@ public:
for(int n=2;n<order;n++){
Linop.HermOp(*Tn,y);
// y=xscale*y+mscale*(*Tn);
// *Tnp=2.0*y-(*Tnm);
// out=out+Coeffs[n]* (*Tnp);
#if 0
auto y_v = y.View();
auto Tn_v = Tn->View();
auto Tnp_v = Tnp->View();
auto Tnm_v = Tnm->View();
constexpr int Nsimd = vector_type::Nsimd();
accelerator_forNB(ss, in.Grid()->oSites(), Nsimd, {
coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss));
coalescedWrite(Tnp_v[ss],2.0*y_v(ss)-Tnm_v(ss));
});
if ( Coeffs[n] != 0.0) {
axpy(out,Coeffs[n],*Tnp,out);
}
#else
axpby(y,xscale,mscale,y,(*Tn));
axpby(*Tnp,2.0,-1.0,y,(*Tnm));
if ( Coeffs[n] != 0.0) {
axpy(out,Coeffs[n],*Tnp,out);
}
#endif
// Cycle pointers to avoid copies
Field *swizzle = Tnm;
Tnm =Tn;

81
Grid/allocator/AlignedAllocator.cc
View File

@ -6,21 +6,39 @@ NAMESPACE_BEGIN(Grid);
MemoryStats *MemoryProfiler::stats = nullptr;
bool MemoryProfiler::debug = false;
#ifdef GRID_NVCC
#define SMALL_LIMIT (0)
int PointerCache::NcacheSmall = PointerCache::NcacheSmallMax;
#ifdef GRID_CUDA
int PointerCache::Ncache = 32;
#else
#define SMALL_LIMIT (4096)
int PointerCache::Ncache = 8;
#endif
int PointerCache::Victim;
int PointerCache::VictimSmall;
PointerCache::PointerCacheEntry PointerCache::Entries[PointerCache::NcacheMax];
PointerCache::PointerCacheEntry PointerCache::EntriesSmall[PointerCache::NcacheSmallMax];
#ifdef POINTER_CACHE
int PointerCache::victim;
void PointerCache::Init(void)
{
char * str;
PointerCache::PointerCacheEntry PointerCache::Entries[PointerCache::Ncache];
str= getenv("GRID_ALLOC_NCACHE_LARGE");
if ( str ) Ncache = atoi(str);
if ( (Ncache<0) || (Ncache > NcacheMax)) Ncache = NcacheMax;
void *PointerCache::Insert(void *ptr,size_t bytes) {
if (bytes < SMALL_LIMIT ) return ptr;
str= getenv("GRID_ALLOC_NCACHE_SMALL");
if ( str ) NcacheSmall = atoi(str);
if ( (NcacheSmall<0) || (NcacheSmall > NcacheSmallMax)) NcacheSmall = NcacheSmallMax;
// printf("Aligned alloocator cache: large %d/%d small %d/%d\n",Ncache,NcacheMax,NcacheSmall,NcacheSmallMax);
}
void *PointerCache::Insert(void *ptr,size_t bytes)
{
if (bytes < GRID_ALLOC_SMALL_LIMIT )
return Insert(ptr,bytes,EntriesSmall,NcacheSmall,VictimSmall);
return Insert(ptr,bytes,Entries,Ncache,Victim);
}
void *PointerCache::Insert(void *ptr,size_t bytes,PointerCacheEntry *entries,int ncache,int &victim)
{
#ifdef GRID_OMP
assert(omp_in_parallel()==0);
#endif
@ -28,8 +46,8 @@ void *PointerCache::Insert(void *ptr,size_t bytes) {
void * ret = NULL;
int v = -1;
for(int e=0;e<Ncache;e++) {
if ( Entries[e].valid==0 ) {
for(int e=0;e<ncache;e++) {
if ( entries[e].valid==0 ) {
v=e;
break;
}
@ -37,40 +55,43 @@ void *PointerCache::Insert(void *ptr,size_t bytes) {
if ( v==-1 ) {
v=victim;
victim = (victim+1)%Ncache;
victim = (victim+1)%ncache;
}
if ( Entries[v].valid ) {
ret = Entries[v].address;
Entries[v].valid = 0;
Entries[v].address = NULL;
Entries[v].bytes = 0;
if ( entries[v].valid ) {
ret = entries[v].address;
entries[v].valid = 0;
entries[v].address = NULL;
entries[v].bytes = 0;
}
Entries[v].address=ptr;
Entries[v].bytes =bytes;
Entries[v].valid =1;
entries[v].address=ptr;
entries[v].bytes =bytes;
entries[v].valid =1;
return ret;
}
void *PointerCache::Lookup(size_t bytes) {
if (bytes < SMALL_LIMIT ) return NULL;
void *PointerCache::Lookup(size_t bytes)
{
if (bytes < GRID_ALLOC_SMALL_LIMIT )
return Lookup(bytes,EntriesSmall,NcacheSmall);
return Lookup(bytes,Entries,Ncache);
}
void *PointerCache::Lookup(size_t bytes,PointerCacheEntry *entries,int ncache)
{
#ifdef GRID_OMP
assert(omp_in_parallel()==0);
#endif
for(int e=0;e<Ncache;e++){
if ( Entries[e].valid && ( Entries[e].bytes == bytes ) ) {
Entries[e].valid = 0;
return Entries[e].address;
for(int e=0;e<ncache;e++){
if ( entries[e].valid && ( entries[e].bytes == bytes ) ) {
entries[e].valid = 0;
return entries[e].address;
}
}
return NULL;
}
#endif
void check_huge_pages(void *Buf,uint64_t BYTES)
{

25
Grid/allocator/AlignedAllocator.h
View File

@ -42,21 +42,21 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#define POINTER_CACHE
#define GRID_ALLOC_ALIGN (2*1024*1024)
#define GRID_ALLOC_SMALL_LIMIT (4096)
NAMESPACE_BEGIN(Grid);
// Move control to configure.ac and Config.h?
#ifdef POINTER_CACHE
class PointerCache {
private:
/*Pinning pages is costly*/
/*Could maintain separate large and small allocation caches*/
#ifdef GRID_NVCC
static const int Ncache=128;
#else
static const int Ncache=8;
#endif
static int victim;
/* Could make these configurable, perhaps up to a max size*/
static const int NcacheSmallMax=128;
static const int NcacheMax=16;
static int NcacheSmall;
static int Ncache;
typedef struct {
void *address;
@ -64,15 +64,18 @@ private:
int valid;
} PointerCacheEntry;
static PointerCacheEntry Entries[Ncache];
static PointerCacheEntry Entries[NcacheMax];
static int Victim;
static PointerCacheEntry EntriesSmall[NcacheSmallMax];
static int VictimSmall;
public:
static void Init(void);
static void *Insert(void *ptr,size_t bytes) ;
static void *Insert(void *ptr,size_t bytes,PointerCacheEntry *entries,int ncache,int &victim) ;
static void *Lookup(size_t bytes) ;
static void *Lookup(size_t bytes,PointerCacheEntry *entries,int ncache) ;
};
#endif
std::string sizeString(size_t bytes);

4
Grid/communicator/SharedMemory.cc
View File

@ -74,7 +74,9 @@ void *SharedMemory::ShmBufferMalloc(size_t bytes){
if (heap_bytes >= heap_size) {
std::cout<< " ShmBufferMalloc exceeded shared heap size -- try increasing with --shm <MB> flag" <<std::endl;
std::cout<< " Parameter specified in units of MB (megabytes) " <<std::endl;
std::cout<< " Current value is " << (heap_size/(1024*1024)) <<std::endl;
std::cout<< " Current alloc is " << (bytes/(1024*1024)) <<"MB"<<std::endl;
std::cout<< " Current bytes is " << (heap_bytes/(1024*1024)) <<"MB"<<std::endl;
std::cout<< " Current heap is " << (heap_size/(1024*1024)) <<"MB"<<std::endl;
assert(heap_bytes<heap_size);
}
//std::cerr << "ShmBufferMalloc "<<std::hex<< ptr<<" - "<<((uint64_t)ptr+bytes)<<std::dec<<std::endl;

2
Grid/lattice/Lattice_reality.h
View File

@ -40,6 +40,7 @@ NAMESPACE_BEGIN(Grid);
template<class vobj> inline Lattice<vobj> adj(const Lattice<vobj> &lhs){
Lattice<vobj> ret(lhs.Grid());
ret.Checkerboard()=lhs.Checkerboard();
auto lhs_v = lhs.View();
auto ret_v = ret.View();
accelerator_for( ss, lhs_v.size(), vobj::Nsimd(), {
@ -50,6 +51,7 @@ template<class vobj> inline Lattice<vobj> adj(const Lattice<vobj> &lhs){
template<class vobj> inline Lattice<vobj> conjugate(const Lattice<vobj> &lhs){
Lattice<vobj> ret(lhs.Grid());
ret.Checkerboard() = lhs.Checkerboard();
auto lhs_v = lhs.View();
auto ret_v = ret.View();
accelerator_for( ss, lhs_v.size(), vobj::Nsimd(), {

6
Grid/perfmon/Timer.h
View File

@ -110,15 +110,15 @@ public:
#endif
accumulator = std::chrono::duration_cast<GridUsecs>(start-start);
}
GridTime Elapsed(void) {
GridTime Elapsed(void) const {
assert(running == false);
return std::chrono::duration_cast<GridTime>( accumulator );
}
uint64_t useconds(void){
uint64_t useconds(void) const {
assert(running == false);
return (uint64_t) accumulator.count();
}
bool isRunning(void){
bool isRunning(void) const {
return running;
}
};

2
Grid/qcd/action/fermion/MobiusFermion.h
View File

@ -59,7 +59,7 @@ public:
{
RealD eps = 1.0;
std::cout<<GridLogMessage << "MobiusFermion (b="<<b<<",c="<<c<<") with Ls= "<<this->Ls<<" Tanh approx"<<std::endl;
// std::cout<<GridLogMessage << "MobiusFermion (b="<<b<<",c="<<c<<") with Ls= "<<this->Ls<<" Tanh approx"<<std::endl;
Approx::zolotarev_data *zdata = Approx::higham(eps,this->Ls);// eps is ignored for higham
assert(zdata->n==this->Ls);

3
Grid/qcd/action/fermion/implementation/CayleyFermion5DImplementation.h
View File

@ -779,9 +779,9 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
assert(mu>=0);
assert(mu<Nd);
int tshift = (mu == Nd-1) ? 1 : 0;
#if 0
int tshift = (mu == Nd-1) ? 1 : 0;
////////////////////////////////////////////////
// SHAMIR CASE
////////////////////////////////////////////////
@ -829,6 +829,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
#endif
#ifndef GRID_NVCC
int tshift = (mu == Nd-1) ? 1 : 0;
////////////////////////////////////////////////
// GENERAL CAYLEY CASE
////////////////////////////////////////////////

80
Grid/qcd/utils/BaryonUtils.h
View File

@ -159,6 +159,7 @@ const Complex BaryonUtils<FImpl>::epsilon_sgn[6] = {Complex(1),
Complex(-1),
Complex(-1)};
//This is the old version
template <class FImpl>
template <class mobj, class robj>
void BaryonUtils<FImpl>::baryon_site(const mobj &D1,
@ -180,6 +181,10 @@ void BaryonUtils<FImpl>::baryon_site(const mobj &D1,
auto pD1 = 0.5* (gD1a + (double)parity * gD1b);
auto gD3 = GammaB_right * D3;
auto D2g = D2 * GammaB_left;
auto pD1g = pD1 * GammaB_left;
auto gD3g = gD3 * GammaB_left;
for (int ie_left=0; ie_left < 6 ; ie_left++){
int a_left = epsilon[ie_left][0]; //a
int b_left = epsilon[ie_left][1]; //b
@ -188,58 +193,71 @@ void BaryonUtils<FImpl>::baryon_site(const mobj &D1,
int a_right = epsilon[ie_right][0]; //a'
int b_right = epsilon[ie_right][1]; //b'
int c_right = epsilon[ie_right][2]; //c'
Complex ee = epsilon_sgn[ie_left] * epsilon_sgn[ie_right];
//This is the \delta_{456}^{123} part
if (wick_contraction[0]){
auto D2g = D2 * GammaB_left;
for (int gamma_left=0; gamma_left<Ns; gamma_left++){
auto eepD1 = ee * pD1()(gamma_left,gamma_left)(c_right,c_left);
for (int alpha_right=0; alpha_right<Ns; alpha_right++){
for (int beta_left=0; beta_left<Ns; beta_left++){
for (int gamma_left=0; gamma_left<Ns; gamma_left++){
result()()() += epsilon_sgn[ie_left] * epsilon_sgn[ie_right] * pD1()(gamma_left,gamma_left)(c_right,c_left)*D2g()(alpha_right,beta_left)(a_right,a_left)*gD3()(alpha_right,beta_left)(b_right,b_left);
auto D2g_ab = D2g()(alpha_right,beta_left)(a_right,a_left);
auto gD3_ab = gD3()(alpha_right,beta_left)(b_right,b_left);
result()()() += eepD1*D2g_ab*gD3_ab;
}}}
}
//This is the \delta_{456}^{231} part
if (wick_contraction[1]){
auto pD1g = pD1 * GammaB_left;
for (int alpha_right=0; alpha_right<Ns; alpha_right++){
for (int beta_left=0; beta_left<Ns; beta_left++){
for (int gamma_left=0; gamma_left<Ns; gamma_left++){
result()()() += epsilon_sgn[ie_left] * epsilon_sgn[ie_right] * pD1g()(gamma_left,beta_left)(c_right,a_left)*D2()(alpha_right,beta_left)(a_right,b_left)*gD3()(alpha_right,gamma_left)(b_right,c_left);
for (int alpha_right=0; alpha_right<Ns; alpha_right++){
auto gD3_ag = gD3()(alpha_right,gamma_left)(b_right,c_left);
for (int beta_left=0; beta_left<Ns; beta_left++){
auto eepD1g_gb = ee * pD1g()(gamma_left,beta_left)(c_right,a_left);
auto D2_ab = D2()(alpha_right,beta_left)(a_right,b_left);
result()()() += eepD1g_gb*D2_ab*gD3_ag;
}}}
}
//This is the \delta_{456}^{312} part
if (wick_contraction[2]){
auto gD3g = gD3 * GammaB_left;
for (int alpha_right=0; alpha_right<Ns; alpha_right++){
for (int beta_left=0; beta_left<Ns; beta_left++){
for (int gamma_left=0; gamma_left<Ns; gamma_left++){
result()()() += epsilon_sgn[ie_left] * epsilon_sgn[ie_right] * pD1()(gamma_left,beta_left)(c_right,b_left)*D2()(alpha_right,gamma_left)(a_right,c_left)*gD3g()(alpha_right,beta_left)(b_right,a_left);
for (int alpha_right=0; alpha_right<Ns; alpha_right++){
auto D2_ag = D2()(alpha_right,gamma_left)(a_right,c_left);
for (int beta_left=0; beta_left<Ns; beta_left++){
auto eepD1_gb = ee * pD1()(gamma_left,beta_left)(c_right,b_left);
auto gD3g_ab = gD3g()(alpha_right,beta_left)(b_right,a_left);
result()()() += eepD1_gb*D2_ag*gD3g_ab;
}}}
}
//This is the \delta_{456}^{132} part
if (wick_contraction[3]){
auto gD3g = gD3 * GammaB_left;
for (int gamma_left=0; gamma_left<Ns; gamma_left++){
auto eepD1 = ee * pD1()(gamma_left,gamma_left)(c_right,c_left);
for (int alpha_right=0; alpha_right<Ns; alpha_right++){
for (int beta_left=0; beta_left<Ns; beta_left++){
for (int gamma_left=0; gamma_left<Ns; gamma_left++){
result()()() -= epsilon_sgn[ie_left] * epsilon_sgn[ie_right] * pD1()(gamma_left,gamma_left)(c_right,c_left)*D2()(alpha_right,beta_left)(a_right,b_left)*gD3g()(alpha_right,beta_left)(b_right,a_left);
auto D2_ab = D2()(alpha_right,beta_left)(a_right,b_left);
auto gD3g_ab = gD3g()(alpha_right,beta_left)(b_right,a_left);
result()()() -= eepD1*D2_ab*gD3g_ab;
}}}
}
//This is the \delta_{456}^{321} part
if (wick_contraction[4]){
auto D2g = D2 * GammaB_left;
for (int alpha_right=0; alpha_right<Ns; alpha_right++){
for (int beta_left=0; beta_left<Ns; beta_left++){
for (int gamma_left=0; gamma_left<Ns; gamma_left++){
result()()() -= epsilon_sgn[ie_left] * epsilon_sgn[ie_right] * pD1()(gamma_left,beta_left)(c_right,b_left)*D2g()(alpha_right,beta_left)(a_right,a_left)*gD3()(alpha_right,gamma_left)(b_right,c_left);
for (int alpha_right=0; alpha_right<Ns; alpha_right++){
auto gD3_ag = gD3()(alpha_right,gamma_left)(b_right,c_left);
for (int beta_left=0; beta_left<Ns; beta_left++){
auto eepD1_gb = ee * pD1()(gamma_left,beta_left)(c_right,b_left);
auto D2g_ab = D2g()(alpha_right,beta_left)(a_right,a_left);
result()()() -= eepD1_gb*D2g_ab*gD3_ag;
}}}
}
//This is the \delta_{456}^{213} part
if (wick_contraction[5]){
auto pD1g = pD1 * GammaB_left;
for (int alpha_right=0; alpha_right<Ns; alpha_right++){
for (int beta_left=0; beta_left<Ns; beta_left++){
for (int gamma_left=0; gamma_left<Ns; gamma_left++){
result()()() -= epsilon_sgn[ie_left] * epsilon_sgn[ie_right] * pD1g()(gamma_left,beta_left)(c_right,a_left)*D2()(alpha_right,gamma_left)(a_right,c_left)*gD3()(alpha_right,beta_left)(b_right,b_left);
for (int alpha_right=0; alpha_right<Ns; alpha_right++){
auto D2_ag = D2()(alpha_right,gamma_left)(a_right,c_left);
for (int beta_left=0; beta_left<Ns; beta_left++){
auto eepD1g_gb = ee * pD1g()(gamma_left,beta_left)(c_right,a_left);
auto gD3_ab = gD3()(alpha_right,beta_left)(b_right,b_left);
result()()() -= eepD1g_gb*D2_ag*gD3_ab;
}}}
}
}
@ -259,6 +277,10 @@ void BaryonUtils<FImpl>::ContractBaryons(const PropagatorField &q1_left,
const int parity,
ComplexField &baryon_corr)
{
assert(Ns==4 && "Baryon code only implemented for N_spin = 4");
assert(Nc==3 && "Baryon code only implemented for N_colour = 3");
std::cout << "Contraction <" << quarks_right[0] << quarks_right[1] << quarks_right[2] << "|" << quarks_left[0] << quarks_left[1] << quarks_left[2] << ">" << std::endl;
std::cout << "GammaA (left) " << (GammaA_left.g) << std::endl;
std::cout << "GammaB (left) " << (GammaB_left.g) << std::endl;
@ -305,6 +327,10 @@ void BaryonUtils<FImpl>::ContractBaryons_Sliced(const mobj &D1,
const int parity,
robj &result)
{
assert(Ns==4 && "Baryon code only implemented for N_spin = 4");
assert(Nc==3 && "Baryon code only implemented for N_colour = 3");
std::cout << "Contraction <" << quarks_right[0] << quarks_right[1] << quarks_right[2] << "|" << quarks_left[0] << quarks_left[1] << quarks_left[2] << ">" << std::endl;
std::cout << "GammaA (left) " << (GammaA_left.g) << std::endl;
std::cout << "GammaB (left) " << (GammaB_left.g) << std::endl;
@ -318,7 +344,7 @@ void BaryonUtils<FImpl>::ContractBaryons_Sliced(const mobj &D1,
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;
result=Zero();
baryon_site(D1,D2,D3,GammaA_left,GammaB_left,GammaA_right,GammaB_right,parity,wick_contraction,result);
baryon_site<decltype(D1),decltype(result)>(D1,D2,D3,GammaA_left,GammaB_left,GammaA_right,GammaB_right,parity,wick_contraction,result);
}
/***********************************************************************
@ -558,6 +584,10 @@ void BaryonUtils<FImpl>::Sigma_to_Nucleon_Eye(const PropagatorField &qq_loop,
const std::string op,
SpinMatrixField &stn_corr)
{
assert(Ns==4 && "Baryon code only implemented for N_spin = 4");
assert(Nc==3 && "Baryon code only implemented for N_colour = 3");
GridBase *grid = qs_ti.Grid();
auto vcorr= stn_corr.View();
@ -595,6 +625,10 @@ void BaryonUtils<FImpl>::Sigma_to_Nucleon_NonEye(const PropagatorField &qq_ti,
const std::string op,
SpinMatrixField &stn_corr)
{
assert(Ns==4 && "Baryon code only implemented for N_spin = 4");
assert(Nc==3 && "Baryon code only implemented for N_colour = 3");
GridBase *grid = qs_ti.Grid();
auto vcorr= stn_corr.View();

2
Grid/util/Init.cc
View File

@ -355,6 +355,8 @@ void Grid_init(int *argc,char ***argv)
//////////////////////////////////////////////////////////
GridGpuInit(); // Must come first to set device prior to MPI init
PointerCache::Init();
if( GridCmdOptionExists(*argv,*argv+*argc,"--shm") ){
int MB;
arg= GridCmdOptionPayload(*argv,*argv+*argc,"--shm");

1
Grid/util/Init.h
View File

@ -56,6 +56,7 @@ std::string GridCmdVectorIntToString(const VectorInt & vec);
void GridCmdOptionCSL(std::string str,std::vector<std::string> & vec);
template<class VectorInt>
void GridCmdOptionIntVector(std::string &str,VectorInt & vec);
void GridCmdOptionInt(std::string &str,int & val);
void GridParseLayout(char **argv,int argc,

538
benchmarks/Benchmark_ITT.cc
View File

@ -30,7 +30,6 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
using namespace Grid;
std::vector<int> L_list;
std::vector<int> Ls_list;
std::vector<double> mflop_list;
@ -76,7 +75,6 @@ struct controls {
int Opt;
int CommsOverlap;
Grid::CartesianCommunicator::CommunicatorPolicy_t CommsAsynch;
// int HugePages;
};
class Benchmark {
@ -119,14 +117,15 @@ public:
std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
comms_header();
for(int lat=4;lat<=maxlat;lat+=4){
for(int Ls=8;Ls<=8;Ls*=2){
for(int lat=16;lat<=maxlat;lat+=8){
// for(int Ls=8;Ls<=8;Ls*=2){
{ int Ls=12;
Coordinate latt_size ({lat*mpi_layout[0],
lat*mpi_layout[1],
lat*mpi_layout[2],
lat*mpi_layout[3]});
std::cout << GridLogMessage<< latt_size <<std::endl;
GridCartesian Grid(latt_size,simd_layout,mpi_layout);
RealD Nrank = Grid._Nprocessors;
RealD Nnode = Grid.NodeCount();
@ -184,9 +183,6 @@ public:
}
timestat.statistics(t_time);
// for(int i=0;i<t_time.size();i++){
// std::cout << i<<" "<<t_time[i]<<std::endl;
// }
dbytes=dbytes*ppn;
double xbytes = dbytes*0.5;
@ -200,8 +196,6 @@ public:
<< "\t\t"<<std::setw(7)<< bidibytes/timestat.mean<< " " << bidibytes*timestat.err/(timestat.mean*timestat.mean) << " "
<< bidibytes/timestat.max << " " << bidibytes/timestat.min << std::endl;
}
}
@ -227,14 +221,15 @@ public:
uint64_t NN;
uint64_t lmax=48;
uint64_t lmax=32;
#define NLOOP (100*lmax*lmax*lmax*lmax/lat/lat/lat/lat)
GridSerialRNG sRNG; sRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9}));
for(int lat=8;lat<=lmax;lat+=4){
for(int lat=8;lat<=lmax;lat+=8){
Coordinate latt_size ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
int64_t vol= latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
GridCartesian Grid(latt_size,simd_layout,mpi_layout);
// NP= Grid.RankCount();
@ -270,191 +265,8 @@ public:
}
};
#if 0
static double DWF5(int Ls,int L)
{
// RealD mass=0.1;
RealD M5 =1.8;
double mflops;
double mflops_best = 0;
double mflops_worst= 0;
std::vector<double> mflops_all;
///////////////////////////////////////////////////////
// Set/Get the layout & grid size
///////////////////////////////////////////////////////
int threads = GridThread::GetThreads();
Coordinate mpi = GridDefaultMpi(); assert(mpi.size()==4);
Coordinate local({L,L,L,L});
GridCartesian * TmpGrid = SpaceTimeGrid::makeFourDimGrid(Coordinate({64,64,64,64}),
GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
uint64_t NP = TmpGrid->RankCount();
uint64_t NN = TmpGrid->NodeCount();
NN_global=NN;
uint64_t SHM=NP/NN;
Coordinate internal;
if ( SHM == 1 ) internal = Coordinate({1,1,1,1});
else if ( SHM == 2 ) internal = Coordinate({2,1,1,1});
else if ( SHM == 4 ) internal = Coordinate({2,2,1,1});
else if ( SHM == 8 ) internal = Coordinate({2,2,2,1});
else assert(0);
Coordinate nodes({mpi[0]/internal[0],mpi[1]/internal[1],mpi[2]/internal[2],mpi[3]/internal[3]});
Coordinate latt4({local[0]*nodes[0],local[1]*nodes[1],local[2]*nodes[2],local[3]*nodes[3]});
///////// Welcome message ////////////
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
std::cout<<GridLogMessage << "Benchmark DWF Ls vec on "<<L<<"^4 local volume "<<std::endl;
std::cout<<GridLogMessage << "* Global volume : "<<GridCmdVectorIntToString(latt4)<<std::endl;
std::cout<<GridLogMessage << "* Ls : "<<Ls<<std::endl;
std::cout<<GridLogMessage << "* MPI ranks : "<<GridCmdVectorIntToString(mpi)<<std::endl;
std::cout<<GridLogMessage << "* Intranode : "<<GridCmdVectorIntToString(internal)<<std::endl;
std::cout<<GridLogMessage << "* nodes : "<<GridCmdVectorIntToString(nodes)<<std::endl;
std::cout<<GridLogMessage << "* Using "<<threads<<" threads"<<std::endl;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
///////// Lattice Init ////////////
GridCartesian * UGrid = SpaceTimeGrid::makeFourDimGrid(latt4, GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
GridCartesian * sUGrid = SpaceTimeGrid::makeFourDimDWFGrid(latt4,GridDefaultMpi());
GridRedBlackCartesian * sUrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(sUGrid);
GridCartesian * sFGrid = SpaceTimeGrid::makeFiveDimDWFGrid(Ls,UGrid);
GridRedBlackCartesian * sFrbGrid = SpaceTimeGrid::makeFiveDimDWFRedBlackGrid(Ls,UGrid);
///////// RNG Init ////////////
std::vector<int> seeds4({1,2,3,4});
std::vector<int> seeds5({5,6,7,8});
GridParallelRNG RNG4(UGrid); RNG4.SeedFixedIntegers(seeds4);
GridParallelRNG RNG5(sFGrid); RNG5.SeedFixedIntegers(seeds5);
std::cout << GridLogMessage << "Initialised RNGs" << std::endl;
///////// Source preparation ////////////
LatticeFermion src (sFGrid);
LatticeFermion tmp (sFGrid);
std::cout << GridLogMessage << "allocated src and tmp" << std::endl;
random(RNG5,src);
std::cout << GridLogMessage << "intialised random source" << std::endl;
RealD N2 = 1.0/::sqrt(norm2(src));
src = src*N2;
LatticeGaugeField Umu(UGrid); SU3::HotConfiguration(RNG4,Umu);
WilsonFermion5DR sDw(Umu,*sFGrid,*sFrbGrid,*sUGrid,*sUrbGrid,M5);
LatticeFermion src_e (sFrbGrid);
LatticeFermion src_o (sFrbGrid);
LatticeFermion r_e (sFrbGrid);
LatticeFermion r_o (sFrbGrid);
LatticeFermion r_eo (sFGrid);
LatticeFermion err (sFGrid);
{
pickCheckerboard(Even,src_e,src);
pickCheckerboard(Odd,src_o,src);
#if defined(AVX512)
const int num_cases = 6;
std::string fmt("A/S ; A/O ; U/S ; U/O ; G/S ; G/O ");
#else
const int num_cases = 4;
std::string fmt("U/S ; U/O ; G/S ; G/O ");
#endif
controls Cases [] = {
#ifdef AVX512
{ WilsonKernelsStatic::OptInlineAsm , WilsonKernelsStatic::CommsThenCompute ,CartesianCommunicator::CommunicatorPolicySequential },
{ WilsonKernelsStatic::OptInlineAsm , WilsonKernelsStatic::CommsAndCompute ,CartesianCommunicator::CommunicatorPolicySequential },
#endif
{ WilsonKernelsStatic::OptHandUnroll, WilsonKernelsStatic::CommsThenCompute ,CartesianCommunicator::CommunicatorPolicySequential },
{ WilsonKernelsStatic::OptHandUnroll, WilsonKernelsStatic::CommsAndCompute ,CartesianCommunicator::CommunicatorPolicySequential },
{ WilsonKernelsStatic::OptGeneric , WilsonKernelsStatic::CommsThenCompute ,CartesianCommunicator::CommunicatorPolicySequential },
{ WilsonKernelsStatic::OptGeneric , WilsonKernelsStatic::CommsAndCompute ,CartesianCommunicator::CommunicatorPolicySequential }
};
for(int c=0;c<num_cases;c++) {
WilsonKernelsStatic::Comms = Cases[c].CommsOverlap;
WilsonKernelsStatic::Opt = Cases[c].Opt;
CartesianCommunicator::SetCommunicatorPolicy(Cases[c].CommsAsynch);
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGeneric ) std::cout << GridLogMessage<< "* Using GENERIC Nc WilsonKernels" <<std::endl;
if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptHandUnroll) std::cout << GridLogMessage<< "* Using Nc=3 WilsonKernels" <<std::endl;
if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm ) std::cout << GridLogMessage<< "* Using Asm Nc=3 WilsonKernels" <<std::endl;
if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute ) std::cout << GridLogMessage<< "* Using Overlapped Comms/Compute" <<std::endl;
if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsThenCompute) std::cout << GridLogMessage<< "* Using sequential comms compute" <<std::endl;
if ( sizeof(Real)==4 ) std::cout << GridLogMessage<< "* SINGLE precision "<<std::endl;
if ( sizeof(Real)==8 ) std::cout << GridLogMessage<< "* DOUBLE precision "<<std::endl;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
int nwarm = 100;
uint64_t ncall = 1000;
double t0=usecond();
sFGrid->Barrier();
for(int i=0;i<nwarm;i++){
sDw.DhopEO(src_o,r_e,DaggerNo);
}
sFGrid->Barrier();
double t1=usecond();
sDw.ZeroCounters();
time_statistics timestat;
std::vector<double> t_time(ncall);
for(uint64_t i=0;i<ncall;i++){
t0=usecond();
sDw.DhopEO(src_o,r_e,DaggerNo);
t1=usecond();
t_time[i] = t1-t0;
}
sFGrid->Barrier();
double volume=Ls; for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
double flops=(1344.0*volume)/2;
double mf_hi, mf_lo, mf_err;
timestat.statistics(t_time);
mf_hi = flops/timestat.min;
mf_lo = flops/timestat.max;
mf_err= flops/timestat.min * timestat.err/timestat.mean;
mflops = flops/timestat.mean;
mflops_all.push_back(mflops);
if ( mflops_best == 0 ) mflops_best = mflops;
if ( mflops_worst== 0 ) mflops_worst= mflops;
if ( mflops>mflops_best ) mflops_best = mflops;
if ( mflops<mflops_worst) mflops_worst= mflops;
std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"sDeo mflop/s = "<< mflops << " ("<<mf_err<<") " << mf_lo<<"-"<<mf_hi <<std::endl;
std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"sDeo mflop/s per rank "<< mflops/NP<<std::endl;
std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"sDeo mflop/s per node "<< mflops/NN<<std::endl;
sDw.Report();
}
double robust = mflops_worst/mflops_best;;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
std::cout<<GridLogMessage << L<<"^4 x "<<Ls<< " sDeo Best mflop/s = "<< mflops_best << " ; " << mflops_best/NN<<" per node " <<std::endl;
std::cout<<GridLogMessage << L<<"^4 x "<<Ls<< " sDeo Worst mflop/s = "<< mflops_worst<< " ; " << mflops_worst/NN<<" per node " <<std::endl;
std::cout<<GridLogMessage <<std::setprecision(3)<< L<<"^4 x "<<Ls<< " Performance Robustness = "<< robust <<std::endl;
std::cout<<GridLogMessage <<fmt << std::endl;
std::cout<<GridLogMessage;
for(int i=0;i<mflops_all.size();i++){
std::cout<<mflops_all[i]/NN<<" ; " ;
}
std::cout<<std::endl;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
}
return mflops_best;
}
#endif
static double DWF(int Ls,int L, double & robust)
static double DWF(int Ls,int L)
{
RealD mass=0.1;
RealD M5 =1.8;
@ -471,37 +283,30 @@ public:
Coordinate mpi = GridDefaultMpi(); assert(mpi.size()==4);
Coordinate local({L,L,L,L});
GridCartesian * TmpGrid = SpaceTimeGrid::makeFourDimGrid(Coordinate({64,64,64,64}),
GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
GridCartesian * TmpGrid = SpaceTimeGrid::makeFourDimGrid(Coordinate({72,72,72,72}),
GridDefaultSimd(Nd,vComplex::Nsimd()),
GridDefaultMpi());
uint64_t NP = TmpGrid->RankCount();
uint64_t NN = TmpGrid->NodeCount();
NN_global=NN;
uint64_t SHM=NP/NN;
Coordinate internal;
if ( SHM == 1 ) internal = Coordinate({1,1,1,1});
else if ( SHM == 2 ) internal = Coordinate({2,1,1,1});
else if ( SHM == 4 ) internal = Coordinate({2,2,1,1});
else if ( SHM == 8 ) internal = Coordinate({2,2,2,1});
else assert(0);
Coordinate nodes({mpi[0]/internal[0],mpi[1]/internal[1],mpi[2]/internal[2],mpi[3]/internal[3]});
Coordinate latt4({local[0]*nodes[0],local[1]*nodes[1],local[2]*nodes[2],local[3]*nodes[3]});
Coordinate latt4({local[0]*mpi[0],local[1]*mpi[1],local[2]*mpi[2],local[3]*mpi[3]});
///////// Welcome message ////////////
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
std::cout<<GridLogMessage << "Benchmark DWF on "<<L<<"^4 local volume "<<std::endl;
std::cout<<GridLogMessage << "* Global volume : "<<GridCmdVectorIntToString(latt4)<<std::endl;
std::cout<<GridLogMessage << "* Ls : "<<Ls<<std::endl;
std::cout<<GridLogMessage << "* MPI ranks : "<<GridCmdVectorIntToString(mpi)<<std::endl;
std::cout<<GridLogMessage << "* Intranode : "<<GridCmdVectorIntToString(internal)<<std::endl;
std::cout<<GridLogMessage << "* nodes : "<<GridCmdVectorIntToString(nodes)<<std::endl;
std::cout<<GridLogMessage << "* ranks : "<<NP <<std::endl;
std::cout<<GridLogMessage << "* nodes : "<<NN <<std::endl;
std::cout<<GridLogMessage << "* ranks/node : "<<SHM <<std::endl;
std::cout<<GridLogMessage << "* ranks geom : "<<GridCmdVectorIntToString(mpi)<<std::endl;
std::cout<<GridLogMessage << "* Using "<<threads<<" threads"<<std::endl;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
///////// Lattice Init ////////////
GridCartesian * UGrid = SpaceTimeGrid::makeFourDimGrid(latt4, GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
GridCartesian * UGrid = SpaceTimeGrid::makeFourDimGrid(latt4, GridDefaultSimd(Nd,vComplexF::Nsimd()),GridDefaultMpi());
GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
GridCartesian * FGrid = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
@ -514,74 +319,31 @@ public:
GridParallelRNG RNG5(FGrid); RNG5.SeedFixedIntegers(seeds5);
std::cout << GridLogMessage << "Initialised RNGs" << std::endl;
typedef DomainWallFermionF Action;
typedef typename Action::FermionField Fermion;
typedef LatticeGaugeFieldF Gauge;
///////// Source preparation ////////////
LatticeFermion src (FGrid); random(RNG5,src);
LatticeFermion ref (FGrid);
LatticeFermion tmp (FGrid);
Gauge Umu(UGrid); SU3::HotConfiguration(RNG4,Umu);
Fermion src (FGrid); random(RNG5,src);
Fermion src_e (FrbGrid);
Fermion src_o (FrbGrid);
Fermion r_e (FrbGrid);
Fermion r_o (FrbGrid);
Fermion r_eo (FGrid);
Action Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
RealD N2 = 1.0/::sqrt(norm2(src));
std::cout<<GridLogMessage << "Normalising src "<< N2 <<std::endl;
src = src*N2;
LatticeGaugeField Umu(UGrid); SU3::HotConfiguration(RNG4,Umu);
DomainWallFermionR Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
////////////////////////////////////
// Naive wilson implementation
////////////////////////////////////
{
LatticeGaugeField Umu5d(FGrid);
std::vector<LatticeColourMatrix> U(4,FGrid);
auto Umu_v = Umu.View();
auto Umu5d_v = Umu5d.View();
for(int ss=0;ss<Umu.Grid()->oSites();ss++){
for(int s=0;s<Ls;s++){
Umu5d_v[Ls*ss+s] = Umu_v[ss];
}
}
ref = Zero();
for(int mu=0;mu<Nd;mu++){
U[mu] = PeekIndex<LorentzIndex>(Umu5d,mu);
}
for(int mu=0;mu<Nd;mu++){
tmp = U[mu]*Cshift(src,mu+1,1);
ref=ref + tmp - Gamma(Gmu[mu])*tmp;
tmp =adj(U[mu])*src;
tmp =Cshift(tmp,mu+1,-1);
ref=ref + tmp + Gamma(Gmu[mu])*tmp;
}
ref = -0.5*ref;
}
LatticeFermion src_e (FrbGrid);
LatticeFermion src_o (FrbGrid);
LatticeFermion r_e (FrbGrid);
LatticeFermion r_o (FrbGrid);
LatticeFermion r_eo (FGrid);
LatticeFermion err (FGrid);
{
pickCheckerboard(Even,src_e,src);
pickCheckerboard(Odd,src_o,src);
#if defined(AVX512)
const int num_cases = 6;
std::string fmt("A/S ; A/O ; U/S ; U/O ; G/S ; G/O ");
#else
const int num_cases = 4;
std::string fmt("U/S ; U/O ; G/S ; G/O ");
#endif
std::string fmt("G/S/C ; G/O/C ; G/S/S ; G/O/S ");
controls Cases [] = {
#ifdef AVX512
{ WilsonKernelsStatic::OptInlineAsm , WilsonKernelsStatic::CommsThenCompute ,CartesianCommunicator::CommunicatorPolicySequential },
{ WilsonKernelsStatic::OptInlineAsm , WilsonKernelsStatic::CommsAndCompute ,CartesianCommunicator::CommunicatorPolicySequential },
#endif
{ WilsonKernelsStatic::OptHandUnroll, WilsonKernelsStatic::CommsThenCompute ,CartesianCommunicator::CommunicatorPolicySequential },
{ WilsonKernelsStatic::OptHandUnroll, WilsonKernelsStatic::CommsAndCompute ,CartesianCommunicator::CommunicatorPolicySequential },
{ WilsonKernelsStatic::OptGeneric , WilsonKernelsStatic::CommsThenCompute ,CartesianCommunicator::CommunicatorPolicyConcurrent },
{ WilsonKernelsStatic::OptGeneric , WilsonKernelsStatic::CommsAndCompute ,CartesianCommunicator::CommunicatorPolicyConcurrent },
{ WilsonKernelsStatic::OptGeneric , WilsonKernelsStatic::CommsThenCompute ,CartesianCommunicator::CommunicatorPolicySequential },
{ WilsonKernelsStatic::OptGeneric , WilsonKernelsStatic::CommsAndCompute ,CartesianCommunicator::CommunicatorPolicySequential }
};
@ -594,15 +356,12 @@ public:
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGeneric ) std::cout << GridLogMessage<< "* Using GENERIC Nc WilsonKernels" <<std::endl;
if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptHandUnroll) std::cout << GridLogMessage<< "* Using Nc=3 WilsonKernels" <<std::endl;
if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm ) std::cout << GridLogMessage<< "* Using Asm Nc=3 WilsonKernels" <<std::endl;
if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute ) std::cout << GridLogMessage<< "* Using Overlapped Comms/Compute" <<std::endl;
if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsThenCompute) std::cout << GridLogMessage<< "* Using sequential comms compute" <<std::endl;
if ( sizeof(Real)==4 ) std::cout << GridLogMessage<< "* SINGLE precision "<<std::endl;
if ( sizeof(Real)==8 ) std::cout << GridLogMessage<< "* DOUBLE precision "<<std::endl;
if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsThenCompute) std::cout << GridLogMessage<< "* Using sequential Comms/Compute" <<std::endl;
std::cout << GridLogMessage<< "* SINGLE precision "<<std::endl;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
int nwarm = 200;
int nwarm = 10;
double t0=usecond();
FGrid->Barrier();
for(int i=0;i<nwarm;i++){
@ -610,9 +369,7 @@ public:
}
FGrid->Barrier();
double t1=usecond();
// uint64_t ncall = (uint64_t) 2.5*1000.0*1000.0*nwarm/(t1-t0);
// if (ncall < 500) ncall = 500;
uint64_t ncall = 1000;
uint64_t ncall = 50;
FGrid->Broadcast(0,&ncall,sizeof(ncall));
@ -649,24 +406,11 @@ public:
std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Deo mflop/s per rank "<< mflops/NP<<std::endl;
std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Deo mflop/s per node "<< mflops/NN<<std::endl;
Dw.Report();
Dw.DhopEO(src_o,r_e,DaggerNo);
Dw.DhopOE(src_e,r_o,DaggerNo);
setCheckerboard(r_eo,r_o);
setCheckerboard(r_eo,r_e);
err = r_eo-ref;
RealD absref = norm2(ref);
RealD abserr = norm2(err);
std::cout<<GridLogMessage << "norm diff "<< abserr << " / " << absref<<std::endl;
assert(abserr<1.0e-4);
}
robust = mflops_worst/mflops_best;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
std::cout<<GridLogMessage << L<<"^4 x "<<Ls<< " Deo Best mflop/s = "<< mflops_best << " ; " << mflops_best/NN<<" per node " <<std::endl;
std::cout<<GridLogMessage << L<<"^4 x "<<Ls<< " Deo Worst mflop/s = "<< mflops_worst<< " ; " << mflops_worst/NN<<" per node " <<std::endl;
std::cout<<GridLogMessage << std::fixed<<std::setprecision(3)<< L<<"^4 x "<<Ls<< " Performance Robustness = "<< robust <<std::endl;
std::cout<<GridLogMessage <<fmt << std::endl;
std::cout<<GridLogMessage ;
@ -680,8 +424,166 @@ public:
return mflops_best;
}
static double Staggered(int L)
{
double mflops;
double mflops_best = 0;
double mflops_worst= 0;
std::vector<double> mflops_all;
///////////////////////////////////////////////////////
// Set/Get the layout & grid size
///////////////////////////////////////////////////////
int threads = GridThread::GetThreads();
Coordinate mpi = GridDefaultMpi(); assert(mpi.size()==4);
Coordinate local({L,L,L,L});
GridCartesian * TmpGrid = SpaceTimeGrid::makeFourDimGrid(Coordinate({72,72,72,72}),
GridDefaultSimd(Nd,vComplex::Nsimd()),
GridDefaultMpi());
uint64_t NP = TmpGrid->RankCount();
uint64_t NN = TmpGrid->NodeCount();
NN_global=NN;
uint64_t SHM=NP/NN;
Coordinate latt4({local[0]*mpi[0],local[1]*mpi[1],local[2]*mpi[2],local[3]*mpi[3]});
///////// Welcome message ////////////
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
std::cout<<GridLogMessage << "Benchmark ImprovedStaggered on "<<L<<"^4 local volume "<<std::endl;
std::cout<<GridLogMessage << "* Global volume : "<<GridCmdVectorIntToString(latt4)<<std::endl;
std::cout<<GridLogMessage << "* ranks : "<<NP <<std::endl;
std::cout<<GridLogMessage << "* nodes : "<<NN <<std::endl;
std::cout<<GridLogMessage << "* ranks/node : "<<SHM <<std::endl;
std::cout<<GridLogMessage << "* ranks geom : "<<GridCmdVectorIntToString(mpi)<<std::endl;
std::cout<<GridLogMessage << "* Using "<<threads<<" threads"<<std::endl;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
///////// Lattice Init ////////////
GridCartesian * FGrid = SpaceTimeGrid::makeFourDimGrid(latt4, GridDefaultSimd(Nd,vComplexF::Nsimd()),GridDefaultMpi());
GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(FGrid);
///////// RNG Init ////////////
std::vector<int> seeds4({1,2,3,4});
GridParallelRNG RNG4(FGrid); RNG4.SeedFixedIntegers(seeds4);
std::cout << GridLogMessage << "Initialised RNGs" << std::endl;
RealD mass=0.1;
RealD c1=9.0/8.0;
RealD c2=-1.0/24.0;
RealD u0=1.0;
typedef ImprovedStaggeredFermionF Action;
typedef typename Action::FermionField Fermion;
typedef LatticeGaugeFieldF Gauge;
Gauge Umu(FGrid); SU3::HotConfiguration(RNG4,Umu);
typename Action::ImplParams params;
Action Ds(Umu,Umu,*FGrid,*FrbGrid,mass,c1,c2,u0,params);
///////// Source preparation ////////////
Fermion src (FGrid); random(RNG4,src);
Fermion src_e (FrbGrid);
Fermion src_o (FrbGrid);
Fermion r_e (FrbGrid);
Fermion r_o (FrbGrid);
Fermion r_eo (FGrid);
{
pickCheckerboard(Even,src_e,src);
pickCheckerboard(Odd,src_o,src);
const int num_cases = 4;
std::string fmt("G/S/C ; G/O/C ; G/S/S ; G/O/S ");
controls Cases [] = {
{ StaggeredKernelsStatic::OptGeneric , StaggeredKernelsStatic::CommsThenCompute ,CartesianCommunicator::CommunicatorPolicyConcurrent },
{ StaggeredKernelsStatic::OptGeneric , StaggeredKernelsStatic::CommsAndCompute ,CartesianCommunicator::CommunicatorPolicyConcurrent },
{ StaggeredKernelsStatic::OptGeneric , StaggeredKernelsStatic::CommsThenCompute ,CartesianCommunicator::CommunicatorPolicySequential },
{ StaggeredKernelsStatic::OptGeneric , StaggeredKernelsStatic::CommsAndCompute ,CartesianCommunicator::CommunicatorPolicySequential }
};
for(int c=0;c<num_cases;c++) {
StaggeredKernelsStatic::Comms = Cases[c].CommsOverlap;
StaggeredKernelsStatic::Opt = Cases[c].Opt;
CartesianCommunicator::SetCommunicatorPolicy(Cases[c].CommsAsynch);
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
if ( StaggeredKernelsStatic::Opt == StaggeredKernelsStatic::OptGeneric ) std::cout << GridLogMessage<< "* Using GENERIC Nc StaggeredKernels" <<std::endl;
if ( StaggeredKernelsStatic::Comms == StaggeredKernelsStatic::CommsAndCompute ) std::cout << GridLogMessage<< "* Using Overlapped Comms/Compute" <<std::endl;
if ( StaggeredKernelsStatic::Comms == StaggeredKernelsStatic::CommsThenCompute) std::cout << GridLogMessage<< "* Using sequential Comms/Compute" <<std::endl;
std::cout << GridLogMessage<< "* SINGLE precision "<<std::endl;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
int nwarm = 10;
double t0=usecond();
FGrid->Barrier();
for(int i=0;i<nwarm;i++){
Ds.DhopEO(src_o,r_e,DaggerNo);
}
FGrid->Barrier();
double t1=usecond();
uint64_t ncall = 500;
FGrid->Broadcast(0,&ncall,sizeof(ncall));
// std::cout << GridLogMessage << " Estimate " << ncall << " calls per second"<<std::endl;
Ds.ZeroCounters();
time_statistics timestat;
std::vector<double> t_time(ncall);
for(uint64_t i=0;i<ncall;i++){
t0=usecond();
Ds.DhopEO(src_o,r_e,DaggerNo);
t1=usecond();
t_time[i] = t1-t0;
}
FGrid->Barrier();
double volume=1; for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
double flops=(1146.0*volume)/2;
double mf_hi, mf_lo, mf_err;
timestat.statistics(t_time);
mf_hi = flops/timestat.min;
mf_lo = flops/timestat.max;
mf_err= flops/timestat.min * timestat.err/timestat.mean;
mflops = flops/timestat.mean;
mflops_all.push_back(mflops);
if ( mflops_best == 0 ) mflops_best = mflops;
if ( mflops_worst== 0 ) mflops_worst= mflops;
if ( mflops>mflops_best ) mflops_best = mflops;
if ( mflops<mflops_worst) mflops_worst= mflops;
std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Deo mflop/s = "<< mflops << " ("<<mf_err<<") " << mf_lo<<"-"<<mf_hi <<std::endl;
std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Deo mflop/s per rank "<< mflops/NP<<std::endl;
std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Deo mflop/s per node "<< mflops/NN<<std::endl;
}
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
std::cout<<GridLogMessage << L<<"^4 Deo Best mflop/s = "<< mflops_best << " ; " << mflops_best/NN<<" per node " <<std::endl;
std::cout<<GridLogMessage << L<<"^4 Deo Worst mflop/s = "<< mflops_worst<< " ; " << mflops_worst/NN<<" per node " <<std::endl;
std::cout<<GridLogMessage <<fmt << std::endl;
std::cout<<GridLogMessage ;
for(int i=0;i<mflops_all.size();i++){
std::cout<<mflops_all[i]/NN<<" ; " ;
}
std::cout<<std::endl;
}
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
return mflops_best;
}
};
int main (int argc, char ** argv)
{
Grid_init(&argc,&argv);
@ -696,62 +598,50 @@ int main (int argc, char ** argv)
int do_memory=1;
int do_comms =1;
int do_su3 =0;
int do_wilson=1;
int do_dwf =1;
if ( do_su3 ) {
// empty for now
}
#if 1
int sel=2;
Coordinate L_list({8,12,16,24});
#else
int sel=1;
Coordinate L_list({8,12});
#endif
std::vector<int> L_list({16,24,32});
int selm1=sel-1;
std::vector<double> robust_list;
std::vector<double> wilson;
std::vector<double> dwf4;
std::vector<double> dwf5;
std::vector<double> staggered;
if ( do_wilson ) {
int Ls=1;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
std::cout<<GridLogMessage << " Wilson dslash 4D vectorised" <<std::endl;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
for(int l=0;l<L_list.size();l++){
double robust;
wilson.push_back(Benchmark::DWF(Ls,L_list[l],robust));
}
wilson.push_back(Benchmark::DWF(Ls,L_list[l]));
}
int Ls=16;
if ( do_dwf ) {
Ls=12;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
std::cout<<GridLogMessage << " Domain wall dslash 4D vectorised" <<std::endl;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
for(int l=0;l<L_list.size();l++){
double robust;
double result = Benchmark::DWF(Ls,L_list[l],robust) ;
double result = Benchmark::DWF(Ls,L_list[l]) ;
dwf4.push_back(result);
robust_list.push_back(robust);
}
}
if ( do_dwf ) {
/*
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
std::cout<<GridLogMessage << " Improved Staggered dslash 4D vectorised" <<std::endl;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
for(int l=0;l<L_list.size();l++){
double result = Benchmark::Staggered(L_list[l]) ;
staggered.push_back(result);
}
*/
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
std::cout<<GridLogMessage << " Summary table Ls="<<Ls <<std::endl;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
std::cout<<GridLogMessage << "L \t\t Wilson \t DWF4 " <<std::endl;
std::cout<<GridLogMessage << "L \t\t Wilson \t\t DWF4 \t\tt Staggered" <<std::endl;
for(int l=0;l<L_list.size();l++){
std::cout<<GridLogMessage << L_list[l] <<" \t\t "<< wilson[l]<<" \t "<<dwf4[l] <<std::endl;
std::cout<<GridLogMessage << L_list[l] <<" \t\t "<< wilson[l]<<" \t\t "<<dwf4[l] <<std::endl;
}
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
}
int NN=NN_global;
if ( do_memory ) {
@ -768,7 +658,6 @@ int main (int argc, char ** argv)
Benchmark::Comms();
}
if ( do_dwf ) {
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
std::cout<<GridLogMessage << " Per Node Summary table Ls="<<Ls <<std::endl;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
@ -782,10 +671,7 @@ int main (int argc, char ** argv)
std::cout<<GridLogMessage << " Comparison point result: " << 0.5*(dwf4[sel]+dwf4[selm1])/NN << " Mflop/s per node"<<std::endl;
std::cout<<GridLogMessage << " Comparison point is 0.5*("<<dwf4[sel]/NN<<"+"<<dwf4[selm1]/NN << ") "<<std::endl;
std::cout<<std::setprecision(3);
std::cout<<GridLogMessage << " Comparison point robustness: " << robust_list[sel] <<std::endl;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
}
Grid_finalize();
}

176
benchmarks/Benchmark_schur.cc Normal file
View File

@ -0,0 +1,176 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./benchmarks/Benchmark_dwf.cc
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: paboyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/Grid.h>
using namespace std;
using namespace Grid;
Gamma::Algebra Gmu [] = {
Gamma::Algebra::GammaX,
Gamma::Algebra::GammaY,
Gamma::Algebra::GammaZ,
Gamma::Algebra::GammaT
};
void benchDw(std::vector<int> & L, int Ls);
int main (int argc, char ** argv)
{
Grid_init(&argc,&argv);
const int Ls=12;
std::vector< std::vector<int> > latts;
#if 1
latts.push_back(std::vector<int> ({24,24,24,24}) );
latts.push_back(std::vector<int> ({48,24,24,24}) );
latts.push_back(std::vector<int> ({96,24,24,24}) );
latts.push_back(std::vector<int> ({96,48,24,24}) );
// latts.push_back(std::vector<int> ({96,48,48,24}) );
// latts.push_back(std::vector<int> ({96,48,48,48}) );
#else
// latts.push_back(std::vector<int> ({96,48,48,48}) );
latts.push_back(std::vector<int> ({96,96,96,192}) );
#endif
std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
std::cout << GridLogMessage<< "* Kernel options --dslash-generic, --dslash-unroll, --dslash-asm" <<std::endl;
std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGeneric ) std::cout << GridLogMessage<< "* Using GENERIC Nc WilsonKernels" <<std::endl;
if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptHandUnroll) std::cout << GridLogMessage<< "* Using Nc=3 WilsonKernels" <<std::endl;
if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm ) std::cout << GridLogMessage<< "* Using Asm Nc=3 WilsonKernels" <<std::endl;
std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
int threads = GridThread::GetThreads();
std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;
std::cout<<GridLogMessage << "=========================================================================="<<std::endl;
std::cout<<GridLogMessage << "= Benchmarking DWF"<<std::endl;
std::cout<<GridLogMessage << "=========================================================================="<<std::endl;
std::cout<<GridLogMessage << "Volume \t\t\tProcs \t SchurDiagOne "<<std::endl;
std::cout<<GridLogMessage << "=========================================================================="<<std::endl;
for (int l=0;l<latts.size();l++){
std::vector<int> latt4 = latts[l];
std::cout << GridLogMessage <<"\t";
for(int d=0;d<Nd;d++){
std::cout<<latt4[d]<<"x";
}
std::cout <<Ls<<"\t" ;
benchDw (latt4,Ls);
}
std::cout<<GridLogMessage << "=========================================================================="<<std::endl;
Grid_finalize();
}
void benchDw(std::vector<int> & latt4, int Ls)
{
/////////////////////////////////////////////////////////////////////////////////////
// for Nc=3
/////////////////////////////////////////////////////////////////////////////////////
// Dw : Ls*24*(7+48)= Ls*1320
//
// M5D: Ls*(4*2*Nc mul + 4*2*Nc madd ) = 3*4*2*Nc*Ls = Ls*72
// Meo: Ls*24*(7+48) + Ls*72 = Ls*1392
//
// Mee: 3*Ns*2*Nc*Ls // Chroma 6*N5*Nc*Ns
//
// LeemInv : 2*2*Nc*madd*Ls
// LeeInv : 2*2*Nc*madd*Ls
// DeeInv : 4*2*Nc*mul *Ls
// UeeInv : 2*2*Nc*madd*Ls
// UeemInv : 2*2*Nc*madd*Ls = Nc*Ls*(8+8+8+8+8) = 40*Nc*Ls// Chroma (10*N5 - 8)*Nc*Ns ~ (40 N5 - 32)Nc flops
// QUDA counts as dense LsxLs real matrix x Ls x NcNsNreim => Nc*4*2 x Ls^2 FMA = 16Nc Ls^2 flops
// Mpc => 1452*cbvol*2*Ls flops //
// => (1344+Ls*48)*Ls*cbvol*2 flops QUDA = 1920 @Ls=12 and 2112 @Ls=16
/////////////////////////////////////////////////////////////////////////////////////
GridCartesian * UGrid = SpaceTimeGrid::makeFourDimGrid(latt4, GridDefaultSimd(Nd,vComplexF::Nsimd()),GridDefaultMpi());
GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
GridCartesian * FGrid = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
// long unsigned int single_site_flops = 8*Nc*(7+16*Nc)*Ls;
long unsigned int single_site_mpc_flops = 8*Nc*(7+16*Nc)*2*Ls + 40*Nc*2*Ls + 4*Nc*2*Ls;
long unsigned int single_site_quda_flops = 8*Nc*(7+16*Nc)*2*Ls + 16*Nc*Ls*Ls + 4*Nc*2*Ls;
std::vector<int> seeds4({1,2,3,4});
std::vector<int> seeds5({5,6,7,8});
ColourMatrixF cm = ComplexF(1.0,0.0);
int ncall=300;
RealD mass=0.1;
RealD M5 =1.8;
RealD NP = UGrid->_Nprocessors;
double volume=1; for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
LatticeGaugeFieldF Umu(UGrid); Umu=Zero();
MobiusFermionF Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5,1.5,0.5);
LatticeFermionF src_o (FrbGrid); src_o=1.0;
LatticeFermionF r_o (FrbGrid); r_o=Zero();
int order =151;
SchurDiagOneOperator<MobiusFermionF,LatticeFermionF> Mpc(Dw);
Chebyshev<LatticeFermionF> Cheby(0.0,60.0,order);
{
Mpc.Mpc(src_o,r_o);
Mpc.Mpc(src_o,r_o);
Mpc.Mpc(src_o,r_o);
double t0=usecond();
for(int i=0;i<ncall;i++){
Mpc.Mpc(src_o,r_o);
}
double t1=usecond();
double flops=(single_site_mpc_flops*volume*ncall); // Mpc has 1 - Moo^-1 Moe Mee^-1 Meo so CB cancels.
std::cout <<"\t"<<NP<< "\t"<<flops/(t1-t0);
flops=(single_site_quda_flops*volume*ncall);
std::cout <<"\t"<<flops/(t1-t0)<<"\t"<<(t1-t0)/1000./1000.<<" s\t";
// Cheby uses MpcDagMpc so 2x flops
for(int i=0;i<1;i++){
Cheby(Mpc,src_o,r_o);
t0=usecond();
Cheby(Mpc,src_o,r_o);
t1=usecond();
flops=(single_site_mpc_flops*volume*2*order);
std::cout <<"\t"<<flops/(t1-t0);
flops=(single_site_quda_flops*volume*2*order);
std::cout <<"\t"<<flops/(t1-t0) << "\t" << (t1-t0)/1000./1000. <<" s";
std::cout <<std::endl;
}
}
// Dw.Report();
}

19
benchmarks/Benchmark_staggered.cc
View File

@ -88,25 +88,6 @@ int main (int argc, char ** argv)
U[mu] = PeekIndex<LorentzIndex>(Umu,mu);
}
ref = Zero();
/*
{ // Naive wilson implementation
ref = Zero();
for(int mu=0;mu<Nd;mu++){
// ref = src + Gamma(Gamma::GammaX)* src ; // 1-gamma_x
tmp = U[mu]*Cshift(src,mu,1);
for(int i=0;i<ref._odata.size();i++){
ref[i]+= tmp[i] - Gamma(Gmu[mu])*tmp[i]; ;
}
tmp =adj(U[mu])*src;
tmp =Cshift(tmp,mu,-1);
for(int i=0;i<ref._odata.size();i++){
ref[i]+= tmp[i] + Gamma(Gmu[mu])*tmp[i]; ;
}
}
}
ref = -0.5*ref;
*/
RealD mass=0.1;
RealD c1=9.0/8.0;

14
configure.ac
View File

@ -274,12 +274,20 @@ case ${ac_gen_scalar} in
esac
##################### Compiler dependent choices
case ${CXX} in
#Strip any optional compiler arguments from nvcc call (eg -ccbin) for compiler comparison
CXXBASE=${CXX}
CXXTEST=${CXX}
if echo "${CXX}" | grep -q "nvcc"; then
CXXTEST="nvcc"
fi
case ${CXXTEST} in
nvcc)
# CXX="nvcc -keep -v -x cu "
# CXXLD="nvcc -v -link"
CXX="nvcc -x cu "
CXXLD="nvcc -link"
CXX="${CXXBASE} -x cu "
CXXLD="${CXXBASE} -link"
# CXXFLAGS="$CXXFLAGS -Xcompiler -fno-strict-aliasing -Xcompiler -Wno-unusable-partial-specialization --expt-extended-lambda --expt-relaxed-constexpr"
CXXFLAGS="$CXXFLAGS -Xcompiler -fno-strict-aliasing --expt-extended-lambda --expt-relaxed-constexpr"
if test $ac_openmp = yes; then