portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-09 23:45:36 +00:00
Code Releases Activity

Merge branch 'develop' into gparity_HMC_merge_develop

Browse Source
This commit is contained in:
Christopher Kelly 2022-02-22 14:25:27 -05:00
commit deac621c2c
99 changed files with 4060 additions and 656 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
Grid/algorithms/CoarsenedMatrix.h
View File

@ -358,7 +358,7 @@ public:
autoView( in_v , in, AcceleratorRead);
autoView( out_v , out, AcceleratorWrite);
autoView( Stencil_v , Stencil, AcceleratorRead);
auto& geom_v = geom;
int npoint = geom.npoint;
typedef LatticeView<Cobj> Aview;
Vector<Aview> AcceleratorViewContainer;
@ -380,7 +380,7 @@ public:
int ptype;
StencilEntry *SE;
for(int point=0;point<geom_v.npoint;point++){
for(int point=0;point<npoint;point++){
SE=Stencil_v.GetEntry(ptype,point,ss);
@ -424,7 +424,7 @@ public:
autoView( in_v , in, AcceleratorRead);
autoView( out_v , out, AcceleratorWrite);
autoView( Stencil_v , Stencil, AcceleratorRead);
auto& geom_v = geom;
int npoint = geom.npoint;
typedef LatticeView<Cobj> Aview;
Vector<Aview> AcceleratorViewContainer;
@ -454,7 +454,7 @@ public:
int ptype;
StencilEntry *SE;
for(int p=0;p<geom_v.npoint;p++){
for(int p=0;p<npoint;p++){
int point = points_p[p];
SE=Stencil_v.GetEntry(ptype,point,ss);

5
Grid/algorithms/LinearOperator.h
View File

@ -52,6 +52,7 @@ public:
virtual void AdjOp (const Field &in, Field &out) = 0; // Abstract base
virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2)=0;
virtual void HermOp(const Field &in, Field &out)=0;
virtual ~LinearOperatorBase(){};
};
@ -507,7 +508,7 @@ class SchurStaggeredOperator : public SchurOperatorBase<Field> {
virtual void MpcDag (const Field &in, Field &out){
Mpc(in,out);
}
virtual void MpcDagMpc(const Field &in, Field &out,RealD &ni,RealD &no) {
virtual void MpcDagMpc(const Field &in, Field &out) {
assert(0);// Never need with staggered
}
};
@ -585,6 +586,7 @@ class HermOpOperatorFunction : public OperatorFunction<Field> {
template<typename Field>
class PlainHermOp : public LinearFunction<Field> {
public:
using LinearFunction<Field>::operator();
LinearOperatorBase<Field> &_Linop;
PlainHermOp(LinearOperatorBase<Field>& linop) : _Linop(linop)
@ -598,6 +600,7 @@ public:
template<typename Field>
class FunctionHermOp : public LinearFunction<Field> {
public:
using LinearFunction<Field>::operator();
OperatorFunction<Field> & _poly;
LinearOperatorBase<Field> &_Linop;

10
Grid/algorithms/Preconditioner.h
View File

@ -30,13 +30,19 @@ Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
NAMESPACE_BEGIN(Grid);
template<class Field> class Preconditioner : public LinearFunction<Field> {
template<class Field> using Preconditioner = LinearFunction<Field> ;
/*
template<class Field> class Preconditioner : public LinearFunction<Field> {
using LinearFunction<Field>::operator();
virtual void operator()(const Field &src, Field & psi)=0;
};
*/
template<class Field> class TrivialPrecon : public Preconditioner<Field> {
public:
void operator()(const Field &src, Field & psi){
using Preconditioner<Field>::operator();
virtual void operator()(const Field &src, Field & psi){
psi = src;
}
TrivialPrecon(void){};

3
Grid/algorithms/SparseMatrix.h
View File

@ -48,6 +48,7 @@ public:
virtual void Mdiag (const Field &in, Field &out)=0;
virtual void Mdir (const Field &in, Field &out,int dir, int disp)=0;
virtual void MdirAll (const Field &in, std::vector<Field> &out)=0;
virtual ~SparseMatrixBase() {};
};
/////////////////////////////////////////////////////////////////////////////////////////////
@ -72,7 +73,7 @@ public:
virtual void MeooeDag (const Field &in, Field &out)=0;
virtual void MooeeDag (const Field &in, Field &out)=0;
virtual void MooeeInvDag (const Field &in, Field &out)=0;
virtual ~CheckerBoardedSparseMatrixBase() {};
};
NAMESPACE_END(Grid);

3
Grid/algorithms/iterative/BiCGSTABMixedPrec.h
View File

@ -36,7 +36,8 @@ NAMESPACE_BEGIN(Grid);
template<class FieldD, class FieldF, typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0, typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0>
class MixedPrecisionBiCGSTAB : public LinearFunction<FieldD>
{
public:
public:
using LinearFunction<FieldD>::operator();
RealD Tolerance;
RealD InnerTolerance; // Initial tolerance for inner CG. Defaults to Tolerance but can be changed
Integer MaxInnerIterations;

3
Grid/algorithms/iterative/ConjugateGradientMixedPrec.h
View File

@ -35,7 +35,8 @@ NAMESPACE_BEGIN(Grid);
typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,
typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0>
class MixedPrecisionConjugateGradient : public LinearFunction<FieldD> {
public:
public:
using LinearFunction<FieldD>::operator();
RealD Tolerance;
RealD InnerTolerance; //Initial tolerance for inner CG. Defaults to Tolerance but can be changed
Integer MaxInnerIterations;

19
Grid/algorithms/iterative/Deflation.h
View File

@ -33,16 +33,19 @@ namespace Grid {
template<class Field>
class ZeroGuesser: public LinearFunction<Field> {
public:
using LinearFunction<Field>::operator();
virtual void operator()(const Field &src, Field &guess) { guess = Zero(); };
};
template<class Field>
class DoNothingGuesser: public LinearFunction<Field> {
public:
using LinearFunction<Field>::operator();
virtual void operator()(const Field &src, Field &guess) { };
};
template<class Field>
class SourceGuesser: public LinearFunction<Field> {
public:
using LinearFunction<Field>::operator();
virtual void operator()(const Field &src, Field &guess) { guess = src; };
};
@ -54,15 +57,24 @@ class DeflatedGuesser: public LinearFunction<Field> {
private:
const std::vector<Field> &evec;
const std::vector<RealD> &eval;
const unsigned int N;
public:
using LinearFunction<Field>::operator();
DeflatedGuesser(const std::vector<Field> & _evec,const std::vector<RealD> & _eval) : evec(_evec), eval(_eval) {};
DeflatedGuesser(const std::vector<Field> & _evec,const std::vector<RealD> & _eval)
: DeflatedGuesser(_evec, _eval, _evec.size())
{}
DeflatedGuesser(const std::vector<Field> & _evec, const std::vector<RealD> & _eval, const unsigned int _N)
: evec(_evec), eval(_eval), N(_N)
{
assert(evec.size()==eval.size());
assert(N <= evec.size());
}
virtual void operator()(const Field &src,Field &guess) {
guess = Zero();
assert(evec.size()==eval.size());
auto N = evec.size();
for (int i=0;i<N;i++) {
const Field& tmp = evec[i];
axpy(guess,TensorRemove(innerProduct(tmp,src)) / eval[i],tmp,guess);
@ -79,6 +91,7 @@ private:
const std::vector<RealD> &eval_coarse;
public:
using LinearFunction<FineField>::operator();
LocalCoherenceDeflatedGuesser(const std::vector<FineField> &_subspace,
const std::vector<CoarseField> &_evec_coarse,
const std::vector<RealD> &_eval_coarse)

2
Grid/algorithms/iterative/LocalCoherenceLanczos.h
View File

@ -68,6 +68,7 @@ public:
template<class Fobj,class CComplex,int nbasis>
class ProjectedHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > {
public:
using LinearFunction<Lattice<iVector<CComplex,nbasis > > >::operator();
typedef iVector<CComplex,nbasis > CoarseSiteVector;
typedef Lattice<CoarseSiteVector> CoarseField;
typedef Lattice<CComplex> CoarseScalar; // used for inner products on fine field
@ -98,6 +99,7 @@ public:
template<class Fobj,class CComplex,int nbasis>
class ProjectedFunctionHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > {
public:
using LinearFunction<Lattice<iVector<CComplex,nbasis > > >::operator();
typedef iVector<CComplex,nbasis > CoarseSiteVector;
typedef Lattice<CoarseSiteVector> CoarseField;
typedef Lattice<CComplex> CoarseScalar; // used for inner products on fine field

2
Grid/algorithms/iterative/PrecGeneralisedConjugateResidual.h
View File

@ -43,7 +43,7 @@ NAMESPACE_BEGIN(Grid);
template<class Field>
class PrecGeneralisedConjugateResidual : public LinearFunction<Field> {
public:
using LinearFunction<Field>::operator();
RealD Tolerance;
Integer MaxIterations;
int verbose;

11
Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h
View File

@ -43,7 +43,7 @@ NAMESPACE_BEGIN(Grid);
template<class Field>
class PrecGeneralisedConjugateResidualNonHermitian : public LinearFunction<Field> {
public:
using LinearFunction<Field>::operator();
RealD Tolerance;
Integer MaxIterations;
int verbose;
@ -119,7 +119,8 @@ public:
RealD GCRnStep(const Field &src, Field &psi,RealD rsq){
RealD cp;
ComplexD a, b, zAz;
ComplexD a, b;
// ComplexD zAz;
RealD zAAz;
ComplexD rq;
@ -146,7 +147,7 @@ public:
//////////////////////////////////
MatTimer.Start();
Linop.Op(psi,Az);
zAz = innerProduct(Az,psi);
// zAz = innerProduct(Az,psi);
zAAz= norm2(Az);
MatTimer.Stop();
@ -170,7 +171,7 @@ public:
LinalgTimer.Start();
zAz = innerProduct(Az,psi);
// zAz = innerProduct(Az,psi);
zAAz= norm2(Az);
//p[0],q[0],qq[0]
@ -212,7 +213,7 @@ public:
MatTimer.Start();
Linop.Op(z,Az);
MatTimer.Stop();
zAz = innerProduct(Az,psi);
// zAz = innerProduct(Az,psi);
zAAz= norm2(Az);
LinalgTimer.Start();

90
Grid/allocator/MemoryManager.cc
View File

@ -9,14 +9,30 @@ NAMESPACE_BEGIN(Grid);
#define AccSmall (3)
#define Shared (4)
#define SharedSmall (5)
#undef GRID_MM_VERBOSE
uint64_t total_shared;
uint64_t total_device;
uint64_t total_host;;
void MemoryManager::PrintBytes(void)
{
std::cout << " MemoryManager : "<<total_shared<<" shared bytes "<<std::endl;
std::cout << " MemoryManager : "<<total_device<<" accelerator bytes "<<std::endl;
std::cout << " MemoryManager : "<<total_host <<" cpu bytes "<<std::endl;
std::cout << " MemoryManager : ------------------------------------ "<<std::endl;
std::cout << " MemoryManager : PrintBytes "<<std::endl;
std::cout << " MemoryManager : ------------------------------------ "<<std::endl;
std::cout << " MemoryManager : "<<(total_shared>>20)<<" shared Mbytes "<<std::endl;
std::cout << " MemoryManager : "<<(total_device>>20)<<" accelerator Mbytes "<<std::endl;
std::cout << " MemoryManager : "<<(total_host>>20) <<" cpu Mbytes "<<std::endl;
uint64_t cacheBytes;
cacheBytes = CacheBytes[Cpu];
std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" cpu cache Mbytes "<<std::endl;
cacheBytes = CacheBytes[Acc];
std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" acc cache Mbytes "<<std::endl;
cacheBytes = CacheBytes[Shared];
std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" shared cache Mbytes "<<std::endl;
#ifdef GRID_CUDA
cuda_mem();
#endif
}
//////////////////////////////////////////////////////////////////////
@ -24,86 +40,114 @@ void MemoryManager::PrintBytes(void)
//////////////////////////////////////////////////////////////////////
MemoryManager::AllocationCacheEntry MemoryManager::Entries[MemoryManager::NallocType][MemoryManager::NallocCacheMax];
int MemoryManager::Victim[MemoryManager::NallocType];
int MemoryManager::Ncache[MemoryManager::NallocType] = { 8, 32, 8, 32, 8, 32 };
int MemoryManager::Ncache[MemoryManager::NallocType] = { 2, 8, 2, 8, 2, 8 };
uint64_t MemoryManager::CacheBytes[MemoryManager::NallocType];
//////////////////////////////////////////////////////////////////////
// Actual allocation and deallocation utils
//////////////////////////////////////////////////////////////////////
void *MemoryManager::AcceleratorAllocate(size_t bytes)
{
total_device+=bytes;
void *ptr = (void *) Lookup(bytes,Acc);
if ( ptr == (void *) NULL ) {
ptr = (void *) acceleratorAllocDevice(bytes);
total_device+=bytes;
}
#ifdef GRID_MM_VERBOSE
std::cout <<"AcceleratorAllocate "<<std::endl;
PrintBytes();
#endif
return ptr;
}
void MemoryManager::AcceleratorFree (void *ptr,size_t bytes)
{
total_device-=bytes;
void *__freeme = Insert(ptr,bytes,Acc);
if ( __freeme ) {
acceleratorFreeDevice(__freeme);
total_device-=bytes;
// PrintBytes();
}
#ifdef GRID_MM_VERBOSE
std::cout <<"AcceleratorFree "<<std::endl;
PrintBytes();
#endif
}
void *MemoryManager::SharedAllocate(size_t bytes)
{
total_shared+=bytes;
void *ptr = (void *) Lookup(bytes,Shared);
if ( ptr == (void *) NULL ) {
ptr = (void *) acceleratorAllocShared(bytes);
total_shared+=bytes;
// std::cout <<"AcceleratorAllocate: allocated Shared pointer "<<std::hex<<ptr<<std::dec<<std::endl;
// PrintBytes();
}
#ifdef GRID_MM_VERBOSE
std::cout <<"SharedAllocate "<<std::endl;
PrintBytes();
#endif
return ptr;
}
void MemoryManager::SharedFree (void *ptr,size_t bytes)
{
total_shared-=bytes;
void *__freeme = Insert(ptr,bytes,Shared);
if ( __freeme ) {
acceleratorFreeShared(__freeme);
total_shared-=bytes;
// PrintBytes();
}
#ifdef GRID_MM_VERBOSE
std::cout <<"SharedFree "<<std::endl;
PrintBytes();
#endif
}
#ifdef GRID_UVM
void *MemoryManager::CpuAllocate(size_t bytes)
{
total_host+=bytes;
void *ptr = (void *) Lookup(bytes,Cpu);
if ( ptr == (void *) NULL ) {
ptr = (void *) acceleratorAllocShared(bytes);
total_host+=bytes;
}
#ifdef GRID_MM_VERBOSE
std::cout <<"CpuAllocate "<<std::endl;
PrintBytes();
#endif
return ptr;
}
void MemoryManager::CpuFree (void *_ptr,size_t bytes)
{
total_host-=bytes;
NotifyDeletion(_ptr);
void *__freeme = Insert(_ptr,bytes,Cpu);
if ( __freeme ) {
acceleratorFreeShared(__freeme);
total_host-=bytes;
}
#ifdef GRID_MM_VERBOSE
std::cout <<"CpuFree "<<std::endl;
PrintBytes();
#endif
}
#else
void *MemoryManager::CpuAllocate(size_t bytes)
{
total_host+=bytes;
void *ptr = (void *) Lookup(bytes,Cpu);
if ( ptr == (void *) NULL ) {
ptr = (void *) acceleratorAllocCpu(bytes);
total_host+=bytes;
}
#ifdef GRID_MM_VERBOSE
std::cout <<"CpuAllocate "<<std::endl;
PrintBytes();
#endif
return ptr;
}
void MemoryManager::CpuFree (void *_ptr,size_t bytes)
{
total_host-=bytes;
NotifyDeletion(_ptr);
void *__freeme = Insert(_ptr,bytes,Cpu);
if ( __freeme ) {
acceleratorFreeCpu(__freeme);
total_host-=bytes;
}
#ifdef GRID_MM_VERBOSE
std::cout <<"CpuFree "<<std::endl;
PrintBytes();
#endif
}
#endif
@ -115,7 +159,6 @@ void MemoryManager::Init(void)
char * str;
int Nc;
int NcS;
str= getenv("GRID_ALLOC_NCACHE_LARGE");
if ( str ) {
@ -181,13 +224,13 @@ void *MemoryManager::Insert(void *ptr,size_t bytes,int type)
#ifdef ALLOCATION_CACHE
bool small = (bytes < GRID_ALLOC_SMALL_LIMIT);
int cache = type + small;
return Insert(ptr,bytes,Entries[cache],Ncache[cache],Victim[cache]);
return Insert(ptr,bytes,Entries[cache],Ncache[cache],Victim[cache],CacheBytes[cache]);
#else
return ptr;
#endif
}
void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim)
void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim, uint64_t &cacheBytes)
{
assert(ncache>0);
#ifdef GRID_OMP
@ -211,6 +254,7 @@ void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries
if ( entries[v].valid ) {
ret = entries[v].address;
cacheBytes -= entries[v].bytes;
entries[v].valid = 0;
entries[v].address = NULL;
entries[v].bytes = 0;
@ -219,6 +263,7 @@ void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries
entries[v].address=ptr;
entries[v].bytes =bytes;
entries[v].valid =1;
cacheBytes += bytes;
return ret;
}
@ -228,13 +273,13 @@ void *MemoryManager::Lookup(size_t bytes,int type)
#ifdef ALLOCATION_CACHE
bool small = (bytes < GRID_ALLOC_SMALL_LIMIT);
int cache = type+small;
return Lookup(bytes,Entries[cache],Ncache[cache]);
return Lookup(bytes,Entries[cache],Ncache[cache],CacheBytes[cache]);
#else
return NULL;
#endif
}
void *MemoryManager::Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache)
void *MemoryManager::Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache,uint64_t & cacheBytes)
{
assert(ncache>0);
#ifdef GRID_OMP
@ -243,6 +288,7 @@ void *MemoryManager::Lookup(size_t bytes,AllocationCacheEntry *entries,int ncach
for(int e=0;e<ncache;e++){
if ( entries[e].valid && ( entries[e].bytes == bytes ) ) {
entries[e].valid = 0;
cacheBytes -= entries[e].bytes;
return entries[e].address;
}
}

6
Grid/allocator/MemoryManager.h
View File

@ -82,14 +82,15 @@ private:
static AllocationCacheEntry Entries[NallocType][NallocCacheMax];
static int Victim[NallocType];
static int Ncache[NallocType];
static uint64_t CacheBytes[NallocType];
/////////////////////////////////////////////////
// Free pool
/////////////////////////////////////////////////
static void *Insert(void *ptr,size_t bytes,int type) ;
static void *Lookup(size_t bytes,int type) ;
static void *Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim) ;
static void *Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache) ;
static void *Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim,uint64_t &cbytes) ;
static void *Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache,uint64_t &cbytes) ;
static void PrintBytes(void);
public:
@ -169,6 +170,7 @@ private:
public:
static void Print(void);
static void PrintState( void* CpuPtr);
static int isOpen (void* CpuPtr);
static void ViewClose(void* CpuPtr,ViewMode mode);
static void *ViewOpen (void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint);

29
Grid/allocator/MemoryManagerCache.cc
View File

@ -3,7 +3,7 @@
#warning "Using explicit device memory copies"
NAMESPACE_BEGIN(Grid);
//define dprintf(...) printf ( __VA_ARGS__ ); fflush(stdout);
//#define dprintf(...) printf ( __VA_ARGS__ ); fflush(stdout);
#define dprintf(...)
@ -429,6 +429,7 @@ void MemoryManager::NotifyDeletion(void *_ptr)
}
void MemoryManager::Print(void)
{
PrintBytes();
std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
std::cout << GridLogDebug << "Memory Manager " << std::endl;
std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
@ -473,6 +474,32 @@ int MemoryManager::isOpen (void* _CpuPtr)
}
}
void MemoryManager::PrintState(void* _CpuPtr)
{
uint64_t CpuPtr = (uint64_t)_CpuPtr;
if ( EntryPresent(CpuPtr) ){
auto AccCacheIterator = EntryLookup(CpuPtr);
auto & AccCache = AccCacheIterator->second;
std::string str;
if ( AccCache.state==Empty ) str = std::string("Empty");
if ( AccCache.state==CpuDirty ) str = std::string("CpuDirty");
if ( AccCache.state==AccDirty ) str = std::string("AccDirty");
if ( AccCache.state==Consistent)str = std::string("Consistent");
if ( AccCache.state==EvictNext) str = std::string("EvictNext");
std::cout << GridLogMessage << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
std::cout << GridLogMessage << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
<< "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
<< "\t" << AccCache.cpuLock
<< "\t" << AccCache.accLock
<< "\t" << AccCache.LRU_valid<<std::endl;
} else {
std::cout << GridLogMessage << "No Entry in AccCache table." << std::endl;
}
}
NAMESPACE_END(Grid);
#endif

4
Grid/allocator/MemoryManagerShared.cc
View File

@ -16,6 +16,10 @@ uint64_t MemoryManager::DeviceToHostXfer;
void MemoryManager::ViewClose(void* AccPtr,ViewMode mode){};
void *MemoryManager::ViewOpen(void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint){ return CpuPtr; };
int MemoryManager::isOpen (void* CpuPtr) { return 0;}
void MemoryManager::PrintState(void* CpuPtr)
{
std::cout << GridLogMessage << "Host<->Device memory movement not currently managed by Grid." << std::endl;
};
void MemoryManager::Print(void){};
void MemoryManager::NotifyDeletion(void *ptr){};

5
Grid/communicator/Communicator_mpi3.cc
View File

@ -388,8 +388,8 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
// TODO : make a OMP loop on CPU, call threaded bcopy
void *shm = (void *) this->ShmBufferTranslate(dest,recv);
assert(shm!=NULL);
// std::cout <<"acceleratorCopyDeviceToDeviceAsynch"<< std::endl;
acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
acceleratorCopySynchronise(); // MPI prob slower
}
if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
@ -400,6 +400,9 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
}
void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
{
// std::cout << "Copy Synchronised\n"<<std::endl;
acceleratorCopySynchronise();
int nreq=list.size();
if (nreq==0) return;

7
Grid/lattice/Lattice_base.h
View File

@ -88,6 +88,13 @@ public:
LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this),mode);
accessor.ViewClose();
}
// Helper function to print the state of this object in the AccCache
void PrintCacheState(void)
{
MemoryManager::PrintState(this->_odata);
}
/////////////////////////////////////////////////////////////////////////////////
// Return a view object that may be dereferenced in site loops.
// The view is trivially copy constructible and may be copied to an accelerator device

5
Grid/lattice/Lattice_reduction_gpu.h
View File

@ -42,7 +42,6 @@ void getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator
std::cout << GridLogDebug << "\twarpSize = " << warpSize << std::endl;
std::cout << GridLogDebug << "\tsharedMemPerBlock = " << sharedMemPerBlock << std::endl;
std::cout << GridLogDebug << "\tmaxThreadsPerBlock = " << maxThreadsPerBlock << std::endl;
std::cout << GridLogDebug << "\tmaxThreadsPerBlock = " << warpSize << std::endl;
std::cout << GridLogDebug << "\tmultiProcessorCount = " << multiProcessorCount << std::endl;
if (warpSize != WARP_SIZE) {
@ -52,6 +51,10 @@ void getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator
// let the number of threads in a block be a multiple of 2, starting from warpSize
threads = warpSize;
if ( threads*sizeofsobj > sharedMemPerBlock ) {
std::cout << GridLogError << "The object is too large for the shared memory." << std::endl;
exit(EXIT_FAILURE);
}
while( 2*threads*sizeofsobj < sharedMemPerBlock && 2*threads <= maxThreadsPerBlock ) threads *= 2;
// keep all the streaming multiprocessors busy
blocks = nextPow2(multiProcessorCount);

70
Grid/lattice/Lattice_transfer.h
View File

@ -85,6 +85,76 @@ template<class vobj> inline void setCheckerboard(Lattice<vobj> &full,const Latti
});
}
template<class vobj> inline void acceleratorPickCheckerboard(int cb,Lattice<vobj> &half,const Lattice<vobj> &full, int checker_dim_half=0)
{
half.Checkerboard() = cb;
autoView(half_v, half, AcceleratorWrite);
autoView(full_v, full, AcceleratorRead);
Coordinate rdim_full = full.Grid()->_rdimensions;
Coordinate rdim_half = half.Grid()->_rdimensions;
unsigned long ndim_half = half.Grid()->_ndimension;
Coordinate checker_dim_mask_half = half.Grid()->_checker_dim_mask;
Coordinate ostride_half = half.Grid()->_ostride;
accelerator_for(ss, full.Grid()->oSites(),full.Grid()->Nsimd(),{
Coordinate coor;
int cbos;
int linear=0;
Lexicographic::CoorFromIndex(coor,ss,rdim_full);
assert(coor.size()==ndim_half);
for(int d=0;d<ndim_half;d++){
if(checker_dim_mask_half[d]) linear += coor[d];
}
cbos = (linear&0x1);
if (cbos==cb) {
int ssh=0;
for(int d=0;d<ndim_half;d++) {
if (d == checker_dim_half) ssh += ostride_half[d] * ((coor[d] / 2) % rdim_half[d]);
else ssh += ostride_half[d] * (coor[d] % rdim_half[d]);
}
coalescedWrite(half_v[ssh],full_v(ss));
}
});
}
template<class vobj> inline void acceleratorSetCheckerboard(Lattice<vobj> &full,const Lattice<vobj> &half, int checker_dim_half=0)
{
int cb = half.Checkerboard();
autoView(half_v , half, AcceleratorRead);
autoView(full_v , full, AcceleratorWrite);
Coordinate rdim_full = full.Grid()->_rdimensions;
Coordinate rdim_half = half.Grid()->_rdimensions;
unsigned long ndim_half = half.Grid()->_ndimension;
Coordinate checker_dim_mask_half = half.Grid()->_checker_dim_mask;
Coordinate ostride_half = half.Grid()->_ostride;
accelerator_for(ss,full.Grid()->oSites(),full.Grid()->Nsimd(),{
Coordinate coor;
int cbos;
int linear=0;
Lexicographic::CoorFromIndex(coor,ss,rdim_full);
assert(coor.size()==ndim_half);
for(int d=0;d<ndim_half;d++){
if(checker_dim_mask_half[d]) linear += coor[d];
}
cbos = (linear&0x1);
if (cbos==cb) {
int ssh=0;
for(int d=0;d<ndim_half;d++){
if (d == checker_dim_half) ssh += ostride_half[d] * ((coor[d] / 2) % rdim_half[d]);
else ssh += ostride_half[d] * (coor[d] % rdim_half[d]);
}
coalescedWrite(full_v[ss],half_v(ssh));
}
});
}
////////////////////////////////////////////////////////////////////////////////////////////
// Flexible Type Conversion for internal promotion to double as well as graceful
// treatment of scalar-compatible types

2
Grid/parallelIO/IldgIO.h
View File

@ -576,6 +576,8 @@ class ScidacReader : public GridLimeReader {
std::string rec_name(ILDG_BINARY_DATA);
while ( limeReaderNextRecord(LimeR) == LIME_SUCCESS ) {
if ( !strncmp(limeReaderType(LimeR), rec_name.c_str(),strlen(rec_name.c_str()) ) ) {
// in principle should do the line below, but that breaks backard compatibility with old data
// skipPastObjectRecord(std::string(GRID_FIELD_NORM));
skipPastObjectRecord(std::string(SCIDAC_CHECKSUM));
return;
}

240
Grid/qcd/action/fermion/CompactWilsonCloverFermion.h Normal file
View File

@ -0,0 +1,240 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/CompactWilsonCloverFermion.h
Copyright (C) 2020 - 2022
Author: Daniel Richtmann <daniel.richtmann@gmail.com>
Author: Nils Meyer <nils.meyer@ur.de>
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 */
#pragma once
#include <Grid/qcd/action/fermion/WilsonCloverTypes.h>
#include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
NAMESPACE_BEGIN(Grid);
// see Grid/qcd/action/fermion/WilsonCloverFermion.h for description
//
// Modifications done here:
//
// Original: clover term = 12x12 matrix per site
//
// But: Only two diagonal 6x6 hermitian blocks are non-zero (also true for original, verified by running)
// Sufficient to store/transfer only the real parts of the diagonal and one triangular part
// 2 * (6 + 15 * 2) = 72 real or 36 complex words to be stored/transfered
//
// Here: Above but diagonal as complex numbers, i.e., need to store/transfer
// 2 * (6 * 2 + 15 * 2) = 84 real or 42 complex words
//
// Words per site and improvement compared to original (combined with the input and output spinors):
//
// - Original: 2*12 + 12*12 = 168 words -> 1.00 x less
// - Minimal: 2*12 + 36 = 60 words -> 2.80 x less
// - Here: 2*12 + 42 = 66 words -> 2.55 x less
//
// These improvements directly translate to wall-clock time
//
// Data layout:
//
// - diagonal and triangle part as separate lattice fields,
// this was faster than as 1 combined field on all tested machines
// - diagonal: as expected
// - triangle: store upper right triangle in row major order
// - graphical:
// 0 1 2 3 4
// 5 6 7 8
// 9 10 11 = upper right triangle indices
// 12 13
// 14
// 0
// 1
// 2
// 3 = diagonal indices
// 4
// 5
// 0
// 1 5
// 2 6 9 = lower left triangle indices
// 3 7 10 12
// 4 8 11 13 14
//
// Impact on total memory consumption:
// - Original: (2 * 1 + 8 * 1/2) 12x12 matrices = 6 12x12 matrices = 864 complex words per site
// - Here: (2 * 1 + 4 * 1/2) diagonal parts = 4 diagonal parts = 24 complex words per site
// + (2 * 1 + 4 * 1/2) triangle parts = 4 triangle parts = 60 complex words per site
// = 84 complex words per site
template<class Impl>
class CompactWilsonCloverFermion : public WilsonFermion<Impl>,
public WilsonCloverHelpers<Impl>,
public CompactWilsonCloverHelpers<Impl> {
/////////////////////////////////////////////
// Sizes
/////////////////////////////////////////////
public:
INHERIT_COMPACT_CLOVER_SIZES(Impl);
/////////////////////////////////////////////
// Type definitions
/////////////////////////////////////////////
public:
INHERIT_IMPL_TYPES(Impl);
INHERIT_CLOVER_TYPES(Impl);
INHERIT_COMPACT_CLOVER_TYPES(Impl);
typedef WilsonFermion<Impl> WilsonBase;
typedef WilsonCloverHelpers<Impl> Helpers;
typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
/////////////////////////////////////////////
// Constructors
/////////////////////////////////////////////
public:
CompactWilsonCloverFermion(GaugeField& _Umu,
GridCartesian& Fgrid,
GridRedBlackCartesian& Hgrid,
const RealD _mass,
const RealD _csw_r = 0.0,
const RealD _csw_t = 0.0,
const RealD _cF = 1.0,
const WilsonAnisotropyCoefficients& clover_anisotropy = WilsonAnisotropyCoefficients(),
const ImplParams& impl_p = ImplParams());
/////////////////////////////////////////////
// Member functions (implementing interface)
/////////////////////////////////////////////
public:
virtual void Instantiatable() {};
int ConstEE() override { return 0; };
int isTrivialEE() override { return 0; };
void Dhop(const FermionField& in, FermionField& out, int dag) override;
void DhopOE(const FermionField& in, FermionField& out, int dag) override;
void DhopEO(const FermionField& in, FermionField& out, int dag) override;
void DhopDir(const FermionField& in, FermionField& out, int dir, int disp) override;
void DhopDirAll(const FermionField& in, std::vector<FermionField>& out) /* override */;
void M(const FermionField& in, FermionField& out) override;
void Mdag(const FermionField& in, FermionField& out) override;
void Meooe(const FermionField& in, FermionField& out) override;
void MeooeDag(const FermionField& in, FermionField& out) override;
void Mooee(const FermionField& in, FermionField& out) override;
void MooeeDag(const FermionField& in, FermionField& out) override;
void MooeeInv(const FermionField& in, FermionField& out) override;
void MooeeInvDag(const FermionField& in, FermionField& out) override;
void Mdir(const FermionField& in, FermionField& out, int dir, int disp) override;
void MdirAll(const FermionField& in, std::vector<FermionField>& out) override;
void MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) override;
void MooDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) override;
void MeeDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) override;
/////////////////////////////////////////////
// Member functions (internals)
/////////////////////////////////////////////
void MooeeInternal(const FermionField& in,
FermionField& out,
const CloverDiagonalField& diagonal,
const CloverTriangleField& triangle);
/////////////////////////////////////////////
// Helpers
/////////////////////////////////////////////
void ImportGauge(const GaugeField& _Umu) override;
/////////////////////////////////////////////
// Helpers
/////////////////////////////////////////////
private:
template<class Field>
const MaskField* getCorrectMaskField(const Field &in) const {
if(in.Grid()->_isCheckerBoarded) {
if(in.Checkerboard() == Odd) {
return &this->BoundaryMaskOdd;
} else {
return &this->BoundaryMaskEven;
}
} else {
return &this->BoundaryMask;
}
}
template<class Field>
void ApplyBoundaryMask(Field& f) {
const MaskField* m = getCorrectMaskField(f); assert(m != nullptr);
assert(m != nullptr);
CompactHelpers::ApplyBoundaryMask(f, *m);
}
/////////////////////////////////////////////
// Member Data
/////////////////////////////////////////////
public:
RealD csw_r;
RealD csw_t;
RealD cF;
bool open_boundaries;
CloverDiagonalField Diagonal, DiagonalEven, DiagonalOdd;
CloverDiagonalField DiagonalInv, DiagonalInvEven, DiagonalInvOdd;
CloverTriangleField Triangle, TriangleEven, TriangleOdd;
CloverTriangleField TriangleInv, TriangleInvEven, TriangleInvOdd;
FermionField Tmp;
MaskField BoundaryMask, BoundaryMaskEven, BoundaryMaskOdd;
};
NAMESPACE_END(Grid);

18
Grid/qcd/action/fermion/Fermion.h
View File

@ -53,6 +53,7 @@ NAMESPACE_CHECK(Wilson);
#include <Grid/qcd/action/fermion/WilsonTMFermion.h> // 4d wilson like
NAMESPACE_CHECK(WilsonTM);
#include <Grid/qcd/action/fermion/WilsonCloverFermion.h> // 4d wilson clover fermions
#include <Grid/qcd/action/fermion/CompactWilsonCloverFermion.h> // 4d compact wilson clover fermions
NAMESPACE_CHECK(WilsonClover);
#include <Grid/qcd/action/fermion/WilsonFermion5D.h> // 5d base used by all 5d overlap types
NAMESPACE_CHECK(Wilson5D);
@ -153,6 +154,23 @@ typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplR> WilsonCloverTwoInd
typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplF> WilsonCloverTwoIndexAntiSymmetricFermionF;
typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplD> WilsonCloverTwoIndexAntiSymmetricFermionD;
// Compact Clover fermions
typedef CompactWilsonCloverFermion<WilsonImplR> CompactWilsonCloverFermionR;
typedef CompactWilsonCloverFermion<WilsonImplF> CompactWilsonCloverFermionF;
typedef CompactWilsonCloverFermion<WilsonImplD> CompactWilsonCloverFermionD;
typedef CompactWilsonCloverFermion<WilsonAdjImplR> CompactWilsonCloverAdjFermionR;
typedef CompactWilsonCloverFermion<WilsonAdjImplF> CompactWilsonCloverAdjFermionF;
typedef CompactWilsonCloverFermion<WilsonAdjImplD> CompactWilsonCloverAdjFermionD;
typedef CompactWilsonCloverFermion<WilsonTwoIndexSymmetricImplR> CompactWilsonCloverTwoIndexSymmetricFermionR;
typedef CompactWilsonCloverFermion<WilsonTwoIndexSymmetricImplF> CompactWilsonCloverTwoIndexSymmetricFermionF;
typedef CompactWilsonCloverFermion<WilsonTwoIndexSymmetricImplD> CompactWilsonCloverTwoIndexSymmetricFermionD;
typedef CompactWilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplR> CompactWilsonCloverTwoIndexAntiSymmetricFermionR;
typedef CompactWilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplF> CompactWilsonCloverTwoIndexAntiSymmetricFermionF;
typedef CompactWilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplD> CompactWilsonCloverTwoIndexAntiSymmetricFermionD;
// Domain Wall fermions
typedef DomainWallFermion<WilsonImplR> DomainWallFermionR;
typedef DomainWallFermion<WilsonImplF> DomainWallFermionF;

301
Grid/qcd/action/fermion/WilsonCloverFermion.h
View File

@ -4,10 +4,11 @@
Source file: ./lib/qcd/action/fermion/WilsonCloverFermion.h
Copyright (C) 2017
Copyright (C) 2017 - 2022
Author: Guido Cossu <guido.cossu@ed.ac.uk>
Author: David Preti <>
Author: Daniel Richtmann <daniel.richtmann@gmail.com>
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
@ -29,7 +30,8 @@
#pragma once
#include <Grid/Grid.h>
#include <Grid/qcd/action/fermion/WilsonCloverTypes.h>
#include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
NAMESPACE_BEGIN(Grid);
@ -50,18 +52,15 @@ NAMESPACE_BEGIN(Grid);
//////////////////////////////////////////////////////////////////
template <class Impl>
class WilsonCloverFermion : public WilsonFermion<Impl>
class WilsonCloverFermion : public WilsonFermion<Impl>,
public WilsonCloverHelpers<Impl>
{
public:
// Types definitions
INHERIT_IMPL_TYPES(Impl);
template <typename vtype>
using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
typedef iImplClover<Simd> SiteCloverType;
typedef Lattice<SiteCloverType> CloverFieldType;
INHERIT_CLOVER_TYPES(Impl);
public:
typedef WilsonFermion<Impl> WilsonBase;
typedef WilsonFermion<Impl> WilsonBase;
typedef WilsonCloverHelpers<Impl> Helpers;
virtual int ConstEE(void) { return 0; };
virtual void Instantiatable(void){};
@ -72,42 +71,7 @@ public:
const RealD _csw_r = 0.0,
const RealD _csw_t = 0.0,
const WilsonAnisotropyCoefficients &clover_anisotropy = WilsonAnisotropyCoefficients(),
const ImplParams &impl_p = ImplParams()) : WilsonFermion<Impl>(_Umu,
Fgrid,
Hgrid,
_mass, impl_p, clover_anisotropy),
CloverTerm(&Fgrid),
CloverTermInv(&Fgrid),
CloverTermEven(&Hgrid),
CloverTermOdd(&Hgrid),
CloverTermInvEven(&Hgrid),
CloverTermInvOdd(&Hgrid),
CloverTermDagEven(&Hgrid),
CloverTermDagOdd(&Hgrid),
CloverTermInvDagEven(&Hgrid),
CloverTermInvDagOdd(&Hgrid)
{
assert(Nd == 4); // require 4 dimensions
if (clover_anisotropy.isAnisotropic)
{
csw_r = _csw_r * 0.5 / clover_anisotropy.xi_0;
diag_mass = _mass + 1.0 + (Nd - 1) * (clover_anisotropy.nu / clover_anisotropy.xi_0);
}
else
{
csw_r = _csw_r * 0.5;
diag_mass = 4.0 + _mass;
}
csw_t = _csw_t * 0.5;
if (csw_r == 0)
std::cout << GridLogWarning << "Initializing WilsonCloverFermion with csw_r = 0" << std::endl;
if (csw_t == 0)
std::cout << GridLogWarning << "Initializing WilsonCloverFermion with csw_t = 0" << std::endl;
ImportGauge(_Umu);
}
const ImplParams &impl_p = ImplParams());
virtual void M(const FermionField &in, FermionField &out);
virtual void Mdag(const FermionField &in, FermionField &out);
@ -124,250 +88,21 @@ public:
void ImportGauge(const GaugeField &_Umu);
// Derivative parts unpreconditioned pseudofermions
void MDeriv(GaugeField &force, const FermionField &X, const FermionField &Y, int dag)
{
conformable(X.Grid(), Y.Grid());
conformable(X.Grid(), force.Grid());
GaugeLinkField force_mu(force.Grid()), lambda(force.Grid());
GaugeField clover_force(force.Grid());
PropagatorField Lambda(force.Grid());
void MDeriv(GaugeField &force, const FermionField &X, const FermionField &Y, int dag);
// Guido: Here we are hitting some performance issues:
// need to extract the components of the DoubledGaugeField
// for each call
// Possible solution
// Create a vector object to store them? (cons: wasting space)
std::vector<GaugeLinkField> U(Nd, this->Umu.Grid());
Impl::extractLinkField(U, this->Umu);
force = Zero();
// Derivative of the Wilson hopping term
this->DhopDeriv(force, X, Y, dag);
///////////////////////////////////////////////////////////
// Clover term derivative
///////////////////////////////////////////////////////////
Impl::outerProductImpl(Lambda, X, Y);
//std::cout << "Lambda:" << Lambda << std::endl;
Gamma::Algebra sigma[] = {
Gamma::Algebra::SigmaXY,
Gamma::Algebra::SigmaXZ,
Gamma::Algebra::SigmaXT,
Gamma::Algebra::MinusSigmaXY,
Gamma::Algebra::SigmaYZ,
Gamma::Algebra::SigmaYT,
Gamma::Algebra::MinusSigmaXZ,
Gamma::Algebra::MinusSigmaYZ,
Gamma::Algebra::SigmaZT,
Gamma::Algebra::MinusSigmaXT,
Gamma::Algebra::MinusSigmaYT,
Gamma::Algebra::MinusSigmaZT};
/*
sigma_{\mu \nu}=
| 0 sigma[0] sigma[1] sigma[2] |
| sigma[3] 0 sigma[4] sigma[5] |
| sigma[6] sigma[7] 0 sigma[8] |
| sigma[9] sigma[10] sigma[11] 0 |
*/
int count = 0;
clover_force = Zero();
for (int mu = 0; mu < 4; mu++)
{
force_mu = Zero();
for (int nu = 0; nu < 4; nu++)
{
if (mu == nu)
continue;
RealD factor;
if (nu == 4 || mu == 4)
{
factor = 2.0 * csw_t;
}
else
{
factor = 2.0 * csw_r;
}
PropagatorField Slambda = Gamma(sigma[count]) * Lambda; // sigma checked
Impl::TraceSpinImpl(lambda, Slambda); // traceSpin ok
force_mu -= factor*Cmunu(U, lambda, mu, nu); // checked
count++;
}
pokeLorentz(clover_force, U[mu] * force_mu, mu);
}
//clover_force *= csw;
force += clover_force;
}
// Computing C_{\mu \nu}(x) as in Eq.(B.39) in Zbigniew Sroczynski's PhD thesis
GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu)
{
conformable(lambda.Grid(), U[0].Grid());
GaugeLinkField out(lambda.Grid()), tmp(lambda.Grid());
// insertion in upper staple
// please check redundancy of shift operations
// C1+
tmp = lambda * U[nu];
out = Impl::ShiftStaple(Impl::CovShiftForward(tmp, nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu);
// C2+
tmp = U[mu] * Impl::ShiftStaple(adj(lambda), mu);
out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(tmp, mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu);
// C3+
tmp = U[nu] * Impl::ShiftStaple(adj(lambda), nu);
out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(tmp, nu))), mu);
// C4+
out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu) * lambda;
// insertion in lower staple
// C1-
out -= Impl::ShiftStaple(lambda, mu) * Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu);
// C2-
tmp = adj(lambda) * U[nu];
out -= Impl::ShiftStaple(Impl::CovShiftBackward(tmp, nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu);
// C3-
tmp = lambda * U[nu];
out -= Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, tmp)), mu);
// C4-
out -= Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu) * lambda;
return out;
}
protected:
public:
// here fixing the 4 dimensions, make it more general?
RealD csw_r; // Clover coefficient - spatial
RealD csw_t; // Clover coefficient - temporal
RealD diag_mass; // Mass term
CloverFieldType CloverTerm, CloverTermInv; // Clover term
CloverFieldType CloverTermEven, CloverTermOdd; // Clover term EO
CloverFieldType CloverTermInvEven, CloverTermInvOdd; // Clover term Inv EO
CloverFieldType CloverTermDagEven, CloverTermDagOdd; // Clover term Dag EO
CloverFieldType CloverTermInvDagEven, CloverTermInvDagOdd; // Clover term Inv Dag EO
public:
// eventually these can be compressed into 6x6 blocks instead of the 12x12
// using the DeGrand-Rossi basis for the gamma matrices
CloverFieldType fillCloverYZ(const GaugeLinkField &F)
{
CloverFieldType T(F.Grid());
T = Zero();
autoView(T_v,T,AcceleratorWrite);
autoView(F_v,F,AcceleratorRead);
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
{
T_v[i]()(0, 1) = timesMinusI(F_v[i]()());
T_v[i]()(1, 0) = timesMinusI(F_v[i]()());
T_v[i]()(2, 3) = timesMinusI(F_v[i]()());
T_v[i]()(3, 2) = timesMinusI(F_v[i]()());
});
return T;
}
CloverFieldType fillCloverXZ(const GaugeLinkField &F)
{
CloverFieldType T(F.Grid());
T = Zero();
autoView(T_v, T,AcceleratorWrite);
autoView(F_v, F,AcceleratorRead);
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
{
T_v[i]()(0, 1) = -F_v[i]()();
T_v[i]()(1, 0) = F_v[i]()();
T_v[i]()(2, 3) = -F_v[i]()();
T_v[i]()(3, 2) = F_v[i]()();
});
return T;
}
CloverFieldType fillCloverXY(const GaugeLinkField &F)
{
CloverFieldType T(F.Grid());
T = Zero();
autoView(T_v,T,AcceleratorWrite);
autoView(F_v,F,AcceleratorRead);
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
{
T_v[i]()(0, 0) = timesMinusI(F_v[i]()());
T_v[i]()(1, 1) = timesI(F_v[i]()());
T_v[i]()(2, 2) = timesMinusI(F_v[i]()());
T_v[i]()(3, 3) = timesI(F_v[i]()());
});
return T;
}
CloverFieldType fillCloverXT(const GaugeLinkField &F)
{
CloverFieldType T(F.Grid());
T = Zero();
autoView( T_v , T, AcceleratorWrite);
autoView( F_v , F, AcceleratorRead);
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
{
T_v[i]()(0, 1) = timesI(F_v[i]()());
T_v[i]()(1, 0) = timesI(F_v[i]()());
T_v[i]()(2, 3) = timesMinusI(F_v[i]()());
T_v[i]()(3, 2) = timesMinusI(F_v[i]()());
});
return T;
}
CloverFieldType fillCloverYT(const GaugeLinkField &F)
{
CloverFieldType T(F.Grid());
T = Zero();
autoView( T_v ,T,AcceleratorWrite);
autoView( F_v ,F,AcceleratorRead);
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
{
T_v[i]()(0, 1) = -(F_v[i]()());
T_v[i]()(1, 0) = (F_v[i]()());
T_v[i]()(2, 3) = (F_v[i]()());
T_v[i]()(3, 2) = -(F_v[i]()());
});
return T;
}
CloverFieldType fillCloverZT(const GaugeLinkField &F)
{
CloverFieldType T(F.Grid());
T = Zero();
autoView( T_v , T,AcceleratorWrite);
autoView( F_v , F,AcceleratorRead);
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
{
T_v[i]()(0, 0) = timesI(F_v[i]()());
T_v[i]()(1, 1) = timesMinusI(F_v[i]()());
T_v[i]()(2, 2) = timesMinusI(F_v[i]()());
T_v[i]()(3, 3) = timesI(F_v[i]()());
});
return T;
}
CloverField CloverTerm, CloverTermInv; // Clover term
CloverField CloverTermEven, CloverTermOdd; // Clover term EO
CloverField CloverTermInvEven, CloverTermInvOdd; // Clover term Inv EO
CloverField CloverTermDagEven, CloverTermDagOdd; // Clover term Dag EO
CloverField CloverTermInvDagEven, CloverTermInvDagOdd; // Clover term Inv Dag EO
};
NAMESPACE_END(Grid);

761
Grid/qcd/action/fermion/WilsonCloverHelpers.h Normal file
View File

@ -0,0 +1,761 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonCloverHelpers.h
Copyright (C) 2021 - 2022
Author: Daniel Richtmann <daniel.richtmann@gmail.com>
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 */
#pragma once
// Helper routines that implement common clover functionality
NAMESPACE_BEGIN(Grid);
template<class Impl> class WilsonCloverHelpers {
public:
INHERIT_IMPL_TYPES(Impl);
INHERIT_CLOVER_TYPES(Impl);
// Computing C_{\mu \nu}(x) as in Eq.(B.39) in Zbigniew Sroczynski's PhD thesis
static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu)
{
conformable(lambda.Grid(), U[0].Grid());
GaugeLinkField out(lambda.Grid()), tmp(lambda.Grid());
// insertion in upper staple
// please check redundancy of shift operations
// C1+
tmp = lambda * U[nu];
out = Impl::ShiftStaple(Impl::CovShiftForward(tmp, nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu);
// C2+
tmp = U[mu] * Impl::ShiftStaple(adj(lambda), mu);
out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(tmp, mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu);
// C3+
tmp = U[nu] * Impl::ShiftStaple(adj(lambda), nu);
out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(tmp, nu))), mu);
// C4+
out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu) * lambda;
// insertion in lower staple
// C1-
out -= Impl::ShiftStaple(lambda, mu) * Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu);
// C2-
tmp = adj(lambda) * U[nu];
out -= Impl::ShiftStaple(Impl::CovShiftBackward(tmp, nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu);
// C3-
tmp = lambda * U[nu];
out -= Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, tmp)), mu);
// C4-
out -= Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu) * lambda;
return out;
}
static CloverField fillCloverYZ(const GaugeLinkField &F)
{
CloverField T(F.Grid());
T = Zero();
autoView(T_v,T,AcceleratorWrite);
autoView(F_v,F,AcceleratorRead);
accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
{
coalescedWrite(T_v[i]()(0, 1), coalescedRead(timesMinusI(F_v[i]()())));
coalescedWrite(T_v[i]()(1, 0), coalescedRead(timesMinusI(F_v[i]()())));
coalescedWrite(T_v[i]()(2, 3), coalescedRead(timesMinusI(F_v[i]()())));
coalescedWrite(T_v[i]()(3, 2), coalescedRead(timesMinusI(F_v[i]()())));
});
return T;
}
static CloverField fillCloverXZ(const GaugeLinkField &F)
{
CloverField T(F.Grid());
T = Zero();
autoView(T_v, T,AcceleratorWrite);
autoView(F_v, F,AcceleratorRead);
accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
{
coalescedWrite(T_v[i]()(0, 1), coalescedRead(-F_v[i]()()));
coalescedWrite(T_v[i]()(1, 0), coalescedRead(F_v[i]()()));
coalescedWrite(T_v[i]()(2, 3), coalescedRead(-F_v[i]()()));
coalescedWrite(T_v[i]()(3, 2), coalescedRead(F_v[i]()()));
});
return T;
}
static CloverField fillCloverXY(const GaugeLinkField &F)
{
CloverField T(F.Grid());
T = Zero();
autoView(T_v,T,AcceleratorWrite);
autoView(F_v,F,AcceleratorRead);
accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
{
coalescedWrite(T_v[i]()(0, 0), coalescedRead(timesMinusI(F_v[i]()())));
coalescedWrite(T_v[i]()(1, 1), coalescedRead(timesI(F_v[i]()())));
coalescedWrite(T_v[i]()(2, 2), coalescedRead(timesMinusI(F_v[i]()())));
coalescedWrite(T_v[i]()(3, 3), coalescedRead(timesI(F_v[i]()())));
});
return T;
}
static CloverField fillCloverXT(const GaugeLinkField &F)
{
CloverField T(F.Grid());
T = Zero();
autoView( T_v , T, AcceleratorWrite);
autoView( F_v , F, AcceleratorRead);
accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
{
coalescedWrite(T_v[i]()(0, 1), coalescedRead(timesI(F_v[i]()())));
coalescedWrite(T_v[i]()(1, 0), coalescedRead(timesI(F_v[i]()())));
coalescedWrite(T_v[i]()(2, 3), coalescedRead(timesMinusI(F_v[i]()())));
coalescedWrite(T_v[i]()(3, 2), coalescedRead(timesMinusI(F_v[i]()())));
});
return T;
}
static CloverField fillCloverYT(const GaugeLinkField &F)
{
CloverField T(F.Grid());
T = Zero();
autoView( T_v ,T,AcceleratorWrite);
autoView( F_v ,F,AcceleratorRead);
accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
{
coalescedWrite(T_v[i]()(0, 1), coalescedRead(-(F_v[i]()())));
coalescedWrite(T_v[i]()(1, 0), coalescedRead((F_v[i]()())));
coalescedWrite(T_v[i]()(2, 3), coalescedRead((F_v[i]()())));
coalescedWrite(T_v[i]()(3, 2), coalescedRead(-(F_v[i]()())));
});
return T;
}
static CloverField fillCloverZT(const GaugeLinkField &F)
{
CloverField T(F.Grid());
T = Zero();
autoView( T_v , T,AcceleratorWrite);
autoView( F_v , F,AcceleratorRead);
accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
{
coalescedWrite(T_v[i]()(0, 0), coalescedRead(timesI(F_v[i]()())));
coalescedWrite(T_v[i]()(1, 1), coalescedRead(timesMinusI(F_v[i]()())));
coalescedWrite(T_v[i]()(2, 2), coalescedRead(timesMinusI(F_v[i]()())));
coalescedWrite(T_v[i]()(3, 3), coalescedRead(timesI(F_v[i]()())));
});
return T;
}
template<class _Spinor>
static accelerator_inline void multClover(_Spinor& phi, const SiteClover& C, const _Spinor& chi) {
auto CC = coalescedRead(C);
mult(&phi, &CC, &chi);
}
template<class _SpinorField>
inline void multCloverField(_SpinorField& out, const CloverField& C, const _SpinorField& phi) {
const int Nsimd = SiteSpinor::Nsimd();
autoView(out_v, out, AcceleratorWrite);
autoView(phi_v, phi, AcceleratorRead);
autoView(C_v, C, AcceleratorRead);
typedef decltype(coalescedRead(out_v[0])) calcSpinor;
accelerator_for(sss,out.Grid()->oSites(),Nsimd,{
calcSpinor tmp;
multClover(tmp,C_v[sss],phi_v(sss));
coalescedWrite(out_v[sss],tmp);
});
}
};
template<class Impl> class CompactWilsonCloverHelpers {
public:
INHERIT_COMPACT_CLOVER_SIZES(Impl);
INHERIT_IMPL_TYPES(Impl);
INHERIT_CLOVER_TYPES(Impl);
INHERIT_COMPACT_CLOVER_TYPES(Impl);
#if 0
static accelerator_inline typename SiteCloverTriangle::vector_type triangle_elem(const SiteCloverTriangle& triangle, int block, int i, int j) {
assert(i != j);
if(i < j) {
return triangle()(block)(triangle_index(i, j));
} else { // i > j
return conjugate(triangle()(block)(triangle_index(i, j)));
}
}
#else
template<typename vobj>
static accelerator_inline vobj triangle_elem(const iImplCloverTriangle<vobj>& triangle, int block, int i, int j) {
assert(i != j);
if(i < j) {
return triangle()(block)(triangle_index(i, j));
} else { // i > j
return conjugate(triangle()(block)(triangle_index(i, j)));
}
}
#endif
static accelerator_inline int triangle_index(int i, int j) {
if(i == j)
return 0;
else if(i < j)
return Nred * (Nred - 1) / 2 - (Nred - i) * (Nred - i - 1) / 2 + j - i - 1;
else // i > j
return Nred * (Nred - 1) / 2 - (Nred - j) * (Nred - j - 1) / 2 + i - j - 1;
}
static void MooeeKernel_gpu(int Nsite,
int Ls,
const FermionField& in,
FermionField& out,
const CloverDiagonalField& diagonal,
const CloverTriangleField& triangle) {
autoView(diagonal_v, diagonal, AcceleratorRead);
autoView(triangle_v, triangle, AcceleratorRead);
autoView(in_v, in, AcceleratorRead);
autoView(out_v, out, AcceleratorWrite);
typedef decltype(coalescedRead(out_v[0])) CalcSpinor;
const uint64_t NN = Nsite * Ls;
accelerator_for(ss, NN, Simd::Nsimd(), {
int sF = ss;
int sU = ss/Ls;
CalcSpinor res;
CalcSpinor in_t = in_v(sF);
auto diagonal_t = diagonal_v(sU);
auto triangle_t = triangle_v(sU);
for(int block=0; block<Nhs; block++) {
int s_start = block*Nhs;
for(int i=0; i<Nred; i++) {
int si = s_start + i/Nc, ci = i%Nc;
res()(si)(ci) = diagonal_t()(block)(i) * in_t()(si)(ci);
for(int j=0; j<Nred; j++) {
if (j == i) continue;
int sj = s_start + j/Nc, cj = j%Nc;
res()(si)(ci) = res()(si)(ci) + triangle_elem(triangle_t, block, i, j) * in_t()(sj)(cj);
};
};
};
coalescedWrite(out_v[sF], res);
});
}
static void MooeeKernel_cpu(int Nsite,
int Ls,
const FermionField& in,
FermionField& out,
const CloverDiagonalField& diagonal,
const CloverTriangleField& triangle) {
autoView(diagonal_v, diagonal, CpuRead);
autoView(triangle_v, triangle, CpuRead);
autoView(in_v, in, CpuRead);
autoView(out_v, out, CpuWrite);
typedef SiteSpinor CalcSpinor;
#if defined(A64FX) || defined(A64FXFIXEDSIZE)
#define PREFETCH_CLOVER(BASE) { \
uint64_t base; \
int pf_dist_L1 = 1; \
int pf_dist_L2 = -5; /* -> penalty -> disable */ \
\
if ((pf_dist_L1 >= 0) && (sU + pf_dist_L1 < Nsite)) { \
base = (uint64_t)&diag_t()(pf_dist_L1+BASE)(0); \
svprfd(svptrue_b64(), (int64_t*)(base + 0), SV_PLDL1STRM); \
svprfd(svptrue_b64(), (int64_t*)(base + 256), SV_PLDL1STRM); \
svprfd(svptrue_b64(), (int64_t*)(base + 512), SV_PLDL1STRM); \
svprfd(svptrue_b64(), (int64_t*)(base + 768), SV_PLDL1STRM); \
svprfd(svptrue_b64(), (int64_t*)(base + 1024), SV_PLDL1STRM); \
svprfd(svptrue_b64(), (int64_t*)(base + 1280), SV_PLDL1STRM); \
} \
\
if ((pf_dist_L2 >= 0) && (sU + pf_dist_L2 < Nsite)) { \
base = (uint64_t)&diag_t()(pf_dist_L2+BASE)(0); \
svprfd(svptrue_b64(), (int64_t*)(base + 0), SV_PLDL2STRM); \
svprfd(svptrue_b64(), (int64_t*)(base + 256), SV_PLDL2STRM); \
svprfd(svptrue_b64(), (int64_t*)(base + 512), SV_PLDL2STRM); \
svprfd(svptrue_b64(), (int64_t*)(base + 768), SV_PLDL2STRM); \
svprfd(svptrue_b64(), (int64_t*)(base + 1024), SV_PLDL2STRM); \
svprfd(svptrue_b64(), (int64_t*)(base + 1280), SV_PLDL2STRM); \
} \
}
// TODO: Implement/generalize this for other architectures
// I played around a bit on KNL (see below) but didn't bring anything
// #elif defined(AVX512)
// #define PREFETCH_CLOVER(BASE) { \
// uint64_t base; \
// int pf_dist_L1 = 1; \
// int pf_dist_L2 = +4; \
// \
// if ((pf_dist_L1 >= 0) && (sU + pf_dist_L1 < Nsite)) { \
// base = (uint64_t)&diag_t()(pf_dist_L1+BASE)(0); \
// _mm_prefetch((const char*)(base + 0), _MM_HINT_T0); \
// _mm_prefetch((const char*)(base + 64), _MM_HINT_T0); \
// _mm_prefetch((const char*)(base + 128), _MM_HINT_T0); \
// _mm_prefetch((const char*)(base + 192), _MM_HINT_T0); \
// _mm_prefetch((const char*)(base + 256), _MM_HINT_T0); \
// _mm_prefetch((const char*)(base + 320), _MM_HINT_T0); \
// } \
// \
// if ((pf_dist_L2 >= 0) && (sU + pf_dist_L2 < Nsite)) { \
// base = (uint64_t)&diag_t()(pf_dist_L2+BASE)(0); \
// _mm_prefetch((const char*)(base + 0), _MM_HINT_T1); \
// _mm_prefetch((const char*)(base + 64), _MM_HINT_T1); \
// _mm_prefetch((const char*)(base + 128), _MM_HINT_T1); \
// _mm_prefetch((const char*)(base + 192), _MM_HINT_T1); \
// _mm_prefetch((const char*)(base + 256), _MM_HINT_T1); \
// _mm_prefetch((const char*)(base + 320), _MM_HINT_T1); \
// } \
// }
#else
#define PREFETCH_CLOVER(BASE)
#endif
const uint64_t NN = Nsite * Ls;
thread_for(ss, NN, {
int sF = ss;
int sU = ss/Ls;
CalcSpinor res;
CalcSpinor in_t = in_v[sF];
auto diag_t = diagonal_v[sU]; // "diag" instead of "diagonal" here to make code below easier to read
auto triangle_t = triangle_v[sU];
// upper half
PREFETCH_CLOVER(0);
auto in_cc_0_0 = conjugate(in_t()(0)(0)); // Nils: reduces number
auto in_cc_0_1 = conjugate(in_t()(0)(1)); // of conjugates from
auto in_cc_0_2 = conjugate(in_t()(0)(2)); // 30 to 20
auto in_cc_1_0 = conjugate(in_t()(1)(0));
auto in_cc_1_1 = conjugate(in_t()(1)(1));
res()(0)(0) = diag_t()(0)( 0) * in_t()(0)(0)
+ triangle_t()(0)( 0) * in_t()(0)(1)
+ triangle_t()(0)( 1) * in_t()(0)(2)
+ triangle_t()(0)( 2) * in_t()(1)(0)
+ triangle_t()(0)( 3) * in_t()(1)(1)
+ triangle_t()(0)( 4) * in_t()(1)(2);
res()(0)(1) = triangle_t()(0)( 0) * in_cc_0_0;
res()(0)(1) = diag_t()(0)( 1) * in_t()(0)(1)
+ triangle_t()(0)( 5) * in_t()(0)(2)
+ triangle_t()(0)( 6) * in_t()(1)(0)
+ triangle_t()(0)( 7) * in_t()(1)(1)
+ triangle_t()(0)( 8) * in_t()(1)(2)
+ conjugate( res()(0)( 1));
res()(0)(2) = triangle_t()(0)( 1) * in_cc_0_0
+ triangle_t()(0)( 5) * in_cc_0_1;
res()(0)(2) = diag_t()(0)( 2) * in_t()(0)(2)
+ triangle_t()(0)( 9) * in_t()(1)(0)
+ triangle_t()(0)(10) * in_t()(1)(1)
+ triangle_t()(0)(11) * in_t()(1)(2)
+ conjugate( res()(0)( 2));
res()(1)(0) = triangle_t()(0)( 2) * in_cc_0_0
+ triangle_t()(0)( 6) * in_cc_0_1
+ triangle_t()(0)( 9) * in_cc_0_2;
res()(1)(0) = diag_t()(0)( 3) * in_t()(1)(0)
+ triangle_t()(0)(12) * in_t()(1)(1)
+ triangle_t()(0)(13) * in_t()(1)(2)
+ conjugate( res()(1)( 0));
res()(1)(1) = triangle_t()(0)( 3) * in_cc_0_0
+ triangle_t()(0)( 7) * in_cc_0_1
+ triangle_t()(0)(10) * in_cc_0_2
+ triangle_t()(0)(12) * in_cc_1_0;
res()(1)(1) = diag_t()(0)( 4) * in_t()(1)(1)
+ triangle_t()(0)(14) * in_t()(1)(2)
+ conjugate( res()(1)( 1));
res()(1)(2) = triangle_t()(0)( 4) * in_cc_0_0
+ triangle_t()(0)( 8) * in_cc_0_1
+ triangle_t()(0)(11) * in_cc_0_2
+ triangle_t()(0)(13) * in_cc_1_0
+ triangle_t()(0)(14) * in_cc_1_1;
res()(1)(2) = diag_t()(0)( 5) * in_t()(1)(2)
+ conjugate( res()(1)( 2));
vstream(out_v[sF]()(0)(0), res()(0)(0));
vstream(out_v[sF]()(0)(1), res()(0)(1));
vstream(out_v[sF]()(0)(2), res()(0)(2));
vstream(out_v[sF]()(1)(0), res()(1)(0));
vstream(out_v[sF]()(1)(1), res()(1)(1));
vstream(out_v[sF]()(1)(2), res()(1)(2));
// lower half
PREFETCH_CLOVER(1);
auto in_cc_2_0 = conjugate(in_t()(2)(0));
auto in_cc_2_1 = conjugate(in_t()(2)(1));
auto in_cc_2_2 = conjugate(in_t()(2)(2));
auto in_cc_3_0 = conjugate(in_t()(3)(0));
auto in_cc_3_1 = conjugate(in_t()(3)(1));
res()(2)(0) = diag_t()(1)( 0) * in_t()(2)(0)
+ triangle_t()(1)( 0) * in_t()(2)(1)
+ triangle_t()(1)( 1) * in_t()(2)(2)
+ triangle_t()(1)( 2) * in_t()(3)(0)
+ triangle_t()(1)( 3) * in_t()(3)(1)
+ triangle_t()(1)( 4) * in_t()(3)(2);
res()(2)(1) = triangle_t()(1)( 0) * in_cc_2_0;
res()(2)(1) = diag_t()(1)( 1) * in_t()(2)(1)
+ triangle_t()(1)( 5) * in_t()(2)(2)
+ triangle_t()(1)( 6) * in_t()(3)(0)
+ triangle_t()(1)( 7) * in_t()(3)(1)
+ triangle_t()(1)( 8) * in_t()(3)(2)
+ conjugate( res()(2)( 1));
res()(2)(2) = triangle_t()(1)( 1) * in_cc_2_0
+ triangle_t()(1)( 5) * in_cc_2_1;
res()(2)(2) = diag_t()(1)( 2) * in_t()(2)(2)
+ triangle_t()(1)( 9) * in_t()(3)(0)
+ triangle_t()(1)(10) * in_t()(3)(1)
+ triangle_t()(1)(11) * in_t()(3)(2)
+ conjugate( res()(2)( 2));
res()(3)(0) = triangle_t()(1)( 2) * in_cc_2_0
+ triangle_t()(1)( 6) * in_cc_2_1
+ triangle_t()(1)( 9) * in_cc_2_2;
res()(3)(0) = diag_t()(1)( 3) * in_t()(3)(0)
+ triangle_t()(1)(12) * in_t()(3)(1)
+ triangle_t()(1)(13) * in_t()(3)(2)
+ conjugate( res()(3)( 0));
res()(3)(1) = triangle_t()(1)( 3) * in_cc_2_0
+ triangle_t()(1)( 7) * in_cc_2_1
+ triangle_t()(1)(10) * in_cc_2_2
+ triangle_t()(1)(12) * in_cc_3_0;
res()(3)(1) = diag_t()(1)( 4) * in_t()(3)(1)
+ triangle_t()(1)(14) * in_t()(3)(2)
+ conjugate( res()(3)( 1));
res()(3)(2) = triangle_t()(1)( 4) * in_cc_2_0
+ triangle_t()(1)( 8) * in_cc_2_1
+ triangle_t()(1)(11) * in_cc_2_2
+ triangle_t()(1)(13) * in_cc_3_0
+ triangle_t()(1)(14) * in_cc_3_1;
res()(3)(2) = diag_t()(1)( 5) * in_t()(3)(2)
+ conjugate( res()(3)( 2));
vstream(out_v[sF]()(2)(0), res()(2)(0));
vstream(out_v[sF]()(2)(1), res()(2)(1));
vstream(out_v[sF]()(2)(2), res()(2)(2));
vstream(out_v[sF]()(3)(0), res()(3)(0));
vstream(out_v[sF]()(3)(1), res()(3)(1));
vstream(out_v[sF]()(3)(2), res()(3)(2));
});
}
static void MooeeKernel(int Nsite,
int Ls,
const FermionField& in,
FermionField& out,
const CloverDiagonalField& diagonal,
const CloverTriangleField& triangle) {
#if defined(GRID_CUDA) || defined(GRID_HIP)
MooeeKernel_gpu(Nsite, Ls, in, out, diagonal, triangle);
#else
MooeeKernel_cpu(Nsite, Ls, in, out, diagonal, triangle);
#endif
}
static void Invert(const CloverDiagonalField& diagonal,
const CloverTriangleField& triangle,
CloverDiagonalField& diagonalInv,
CloverTriangleField& triangleInv) {
conformable(diagonal, diagonalInv);
conformable(triangle, triangleInv);
conformable(diagonal, triangle);
diagonalInv.Checkerboard() = diagonal.Checkerboard();
triangleInv.Checkerboard() = triangle.Checkerboard();
GridBase* grid = diagonal.Grid();
long lsites = grid->lSites();
typedef typename SiteCloverDiagonal::scalar_object scalar_object_diagonal;
typedef typename SiteCloverTriangle::scalar_object scalar_object_triangle;
autoView(diagonal_v, diagonal, CpuRead);
autoView(triangle_v, triangle, CpuRead);
autoView(diagonalInv_v, diagonalInv, CpuWrite);
autoView(triangleInv_v, triangleInv, CpuWrite);
thread_for(site, lsites, { // NOTE: Not on GPU because of Eigen & (peek/poke)LocalSite
Eigen::MatrixXcd clover_inv_eigen = Eigen::MatrixXcd::Zero(Ns*Nc, Ns*Nc);
Eigen::MatrixXcd clover_eigen = Eigen::MatrixXcd::Zero(Ns*Nc, Ns*Nc);
scalar_object_diagonal diagonal_tmp = Zero();
scalar_object_diagonal diagonal_inv_tmp = Zero();
scalar_object_triangle triangle_tmp = Zero();
scalar_object_triangle triangle_inv_tmp = Zero();
Coordinate lcoor;
grid->LocalIndexToLocalCoor(site, lcoor);
peekLocalSite(diagonal_tmp, diagonal_v, lcoor);
peekLocalSite(triangle_tmp, triangle_v, lcoor);
// TODO: can we save time here by inverting the two 6x6 hermitian matrices separately?
for (long s_row=0;s_row<Ns;s_row++) {
for (long s_col=0;s_col<Ns;s_col++) {
if(abs(s_row - s_col) > 1 || s_row + s_col == 3) continue;
int block = s_row / Nhs;
int s_row_block = s_row % Nhs;
int s_col_block = s_col % Nhs;
for (long c_row=0;c_row<Nc;c_row++) {
for (long c_col=0;c_col<Nc;c_col++) {
int i = s_row_block * Nc + c_row;
int j = s_col_block * Nc + c_col;
if(i == j)
clover_eigen(s_row*Nc+c_row, s_col*Nc+c_col) = static_cast<ComplexD>(TensorRemove(diagonal_tmp()(block)(i)));
else
clover_eigen(s_row*Nc+c_row, s_col*Nc+c_col) = static_cast<ComplexD>(TensorRemove(triangle_elem(triangle_tmp, block, i, j)));
}
}
}
}
clover_inv_eigen = clover_eigen.inverse();
for (long s_row=0;s_row<Ns;s_row++) {
for (long s_col=0;s_col<Ns;s_col++) {
if(abs(s_row - s_col) > 1 || s_row + s_col == 3) continue;
int block = s_row / Nhs;
int s_row_block = s_row % Nhs;
int s_col_block = s_col % Nhs;
for (long c_row=0;c_row<Nc;c_row++) {
for (long c_col=0;c_col<Nc;c_col++) {
int i = s_row_block * Nc + c_row;
int j = s_col_block * Nc + c_col;
if(i == j)
diagonal_inv_tmp()(block)(i) = clover_inv_eigen(s_row*Nc+c_row, s_col*Nc+c_col);
else if(i < j)
triangle_inv_tmp()(block)(triangle_index(i, j)) = clover_inv_eigen(s_row*Nc+c_row, s_col*Nc+c_col);
else
continue;
}
}
}
}
pokeLocalSite(diagonal_inv_tmp, diagonalInv_v, lcoor);
pokeLocalSite(triangle_inv_tmp, triangleInv_v, lcoor);
});
}
static void ConvertLayout(const CloverField& full,
CloverDiagonalField& diagonal,
CloverTriangleField& triangle) {
conformable(full, diagonal);
conformable(full, triangle);
diagonal.Checkerboard() = full.Checkerboard();
triangle.Checkerboard() = full.Checkerboard();
autoView(full_v, full, AcceleratorRead);
autoView(diagonal_v, diagonal, AcceleratorWrite);
autoView(triangle_v, triangle, AcceleratorWrite);
// NOTE: this function cannot be 'private' since nvcc forbids this for kernels
accelerator_for(ss, full.Grid()->oSites(), 1, {
for(int s_row = 0; s_row < Ns; s_row++) {
for(int s_col = 0; s_col < Ns; s_col++) {
if(abs(s_row - s_col) > 1 || s_row + s_col == 3) continue;
int block = s_row / Nhs;
int s_row_block = s_row % Nhs;
int s_col_block = s_col % Nhs;
for(int c_row = 0; c_row < Nc; c_row++) {
for(int c_col = 0; c_col < Nc; c_col++) {
int i = s_row_block * Nc + c_row;
int j = s_col_block * Nc + c_col;
if(i == j)
diagonal_v[ss]()(block)(i) = full_v[ss]()(s_row, s_col)(c_row, c_col);
else if(i < j)
triangle_v[ss]()(block)(triangle_index(i, j)) = full_v[ss]()(s_row, s_col)(c_row, c_col);
else
continue;
}
}
}
}
});
}
static void ConvertLayout(const CloverDiagonalField& diagonal,
const CloverTriangleField& triangle,
CloverField& full) {
conformable(full, diagonal);
conformable(full, triangle);
full.Checkerboard() = diagonal.Checkerboard();
full = Zero();
autoView(diagonal_v, diagonal, AcceleratorRead);
autoView(triangle_v, triangle, AcceleratorRead);
autoView(full_v, full, AcceleratorWrite);
// NOTE: this function cannot be 'private' since nvcc forbids this for kernels
accelerator_for(ss, full.Grid()->oSites(), 1, {
for(int s_row = 0; s_row < Ns; s_row++) {
for(int s_col = 0; s_col < Ns; s_col++) {
if(abs(s_row - s_col) > 1 || s_row + s_col == 3) continue;
int block = s_row / Nhs;
int s_row_block = s_row % Nhs;
int s_col_block = s_col % Nhs;
for(int c_row = 0; c_row < Nc; c_row++) {
for(int c_col = 0; c_col < Nc; c_col++) {
int i = s_row_block * Nc + c_row;
int j = s_col_block * Nc + c_col;
if(i == j)
full_v[ss]()(s_row, s_col)(c_row, c_col) = diagonal_v[ss]()(block)(i);
else
full_v[ss]()(s_row, s_col)(c_row, c_col) = triangle_elem(triangle_v[ss], block, i, j);
}
}
}
}
});
}
static void ModifyBoundaries(CloverDiagonalField& diagonal, CloverTriangleField& triangle, RealD csw_t, RealD cF, RealD diag_mass) {
// Checks/grid
double t0 = usecond();
conformable(diagonal, triangle);
GridBase* grid = diagonal.Grid();
// Determine the boundary coordinates/sites
double t1 = usecond();
int t_dir = Nd - 1;
Lattice<iScalar<vInteger>> t_coor(grid);
LatticeCoordinate(t_coor, t_dir);
int T = grid->GlobalDimensions()[t_dir];
// Set off-diagonal parts at boundary to zero -- OK
double t2 = usecond();
CloverTriangleField zeroTriangle(grid);
zeroTriangle.Checkerboard() = triangle.Checkerboard();
zeroTriangle = Zero();
triangle = where(t_coor == 0, zeroTriangle, triangle);
triangle = where(t_coor == T-1, zeroTriangle, triangle);
// Set diagonal to unity (scaled correctly) -- OK
double t3 = usecond();
CloverDiagonalField tmp(grid);
tmp.Checkerboard() = diagonal.Checkerboard();
tmp = -1.0 * csw_t + diag_mass;
diagonal = where(t_coor == 0, tmp, diagonal);
diagonal = where(t_coor == T-1, tmp, diagonal);
// Correct values next to boundary
double t4 = usecond();
if(cF != 1.0) {
tmp = cF - 1.0;
tmp += diagonal;
diagonal = where(t_coor == 1, tmp, diagonal);
diagonal = where(t_coor == T-2, tmp, diagonal);
}
// Report timings
double t5 = usecond();
#if 0
std::cout << GridLogMessage << "CompactWilsonCloverHelpers::ModifyBoundaries timings:"
<< " checks = " << (t1 - t0) / 1e6
<< ", coordinate = " << (t2 - t1) / 1e6
<< ", off-diag zero = " << (t3 - t2) / 1e6
<< ", diagonal unity = " << (t4 - t3) / 1e6
<< ", near-boundary = " << (t5 - t4) / 1e6
<< ", total = " << (t5 - t0) / 1e6
<< std::endl;
#endif
}
template<class Field, class Mask>
static strong_inline void ApplyBoundaryMask(Field& f, const Mask& m) {
conformable(f, m);
auto grid = f.Grid();
const uint32_t Nsite = grid->oSites();
const uint32_t Nsimd = grid->Nsimd();
autoView(f_v, f, AcceleratorWrite);
autoView(m_v, m, AcceleratorRead);
// NOTE: this function cannot be 'private' since nvcc forbids this for kernels
accelerator_for(ss, Nsite, Nsimd, {
coalescedWrite(f_v[ss], m_v(ss) * f_v(ss));
});
}
template<class MaskField>
static void SetupMasks(MaskField& full, MaskField& even, MaskField& odd) {
assert(even.Grid()->_isCheckerBoarded && even.Checkerboard() == Even);
assert(odd.Grid()->_isCheckerBoarded && odd.Checkerboard() == Odd);
assert(!full.Grid()->_isCheckerBoarded);
GridBase* grid = full.Grid();
int t_dir = Nd-1;
Lattice<iScalar<vInteger>> t_coor(grid);
LatticeCoordinate(t_coor, t_dir);
int T = grid->GlobalDimensions()[t_dir];
MaskField zeroMask(grid); zeroMask = Zero();
full = 1.0;
full = where(t_coor == 0, zeroMask, full);
full = where(t_coor == T-1, zeroMask, full);
pickCheckerboard(Even, even, full);
pickCheckerboard(Odd, odd, full);
}
};
NAMESPACE_END(Grid);

92
Grid/qcd/action/fermion/WilsonCloverTypes.h Normal file
View File

@ -0,0 +1,92 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonCloverTypes.h
Copyright (C) 2021 - 2022
Author: Daniel Richtmann <daniel.richtmann@gmail.com>
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 */
#pragma once
NAMESPACE_BEGIN(Grid);
template<class Impl>
class WilsonCloverTypes {
public:
INHERIT_IMPL_TYPES(Impl);
template <typename vtype> using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
typedef iImplClover<Simd> SiteClover;
typedef Lattice<SiteClover> CloverField;
};
template<class Impl>
class CompactWilsonCloverTypes {
public:
INHERIT_IMPL_TYPES(Impl);
static_assert(Nd == 4 && Nc == 3 && Ns == 4 && Impl::Dimension == 3, "Wrong dimensions");
static constexpr int Nred = Nc * Nhs; // 6
static constexpr int Nblock = Nhs; // 2
static constexpr int Ndiagonal = Nred; // 6
static constexpr int Ntriangle = (Nred - 1) * Nc; // 15
template<typename vtype> using iImplCloverDiagonal = iScalar<iVector<iVector<vtype, Ndiagonal>, Nblock>>;
template<typename vtype> using iImplCloverTriangle = iScalar<iVector<iVector<vtype, Ntriangle>, Nblock>>;
typedef iImplCloverDiagonal<Simd> SiteCloverDiagonal;
typedef iImplCloverTriangle<Simd> SiteCloverTriangle;
typedef iSinglet<Simd> SiteMask;
typedef Lattice<SiteCloverDiagonal> CloverDiagonalField;
typedef Lattice<SiteCloverTriangle> CloverTriangleField;
typedef Lattice<SiteMask> MaskField;
};
#define INHERIT_CLOVER_TYPES(Impl) \
typedef typename WilsonCloverTypes<Impl>::SiteClover SiteClover; \
typedef typename WilsonCloverTypes<Impl>::CloverField CloverField;
#define INHERIT_COMPACT_CLOVER_TYPES(Impl) \
typedef typename CompactWilsonCloverTypes<Impl>::SiteCloverDiagonal SiteCloverDiagonal; \
typedef typename CompactWilsonCloverTypes<Impl>::SiteCloverTriangle SiteCloverTriangle; \
typedef typename CompactWilsonCloverTypes<Impl>::SiteMask SiteMask; \
typedef typename CompactWilsonCloverTypes<Impl>::CloverDiagonalField CloverDiagonalField; \
typedef typename CompactWilsonCloverTypes<Impl>::CloverTriangleField CloverTriangleField; \
typedef typename CompactWilsonCloverTypes<Impl>::MaskField MaskField; \
/* ugly duplication but needed inside functionality classes */ \
template<typename vtype> using iImplCloverDiagonal = \
iScalar<iVector<iVector<vtype, CompactWilsonCloverTypes<Impl>::Ndiagonal>, CompactWilsonCloverTypes<Impl>::Nblock>>; \
template<typename vtype> using iImplCloverTriangle = \
iScalar<iVector<iVector<vtype, CompactWilsonCloverTypes<Impl>::Ntriangle>, CompactWilsonCloverTypes<Impl>::Nblock>>;
#define INHERIT_COMPACT_CLOVER_SIZES(Impl) \
static constexpr int Nred = CompactWilsonCloverTypes<Impl>::Nred; \
static constexpr int Nblock = CompactWilsonCloverTypes<Impl>::Nblock; \
static constexpr int Ndiagonal = CompactWilsonCloverTypes<Impl>::Ndiagonal; \
static constexpr int Ntriangle = CompactWilsonCloverTypes<Impl>::Ntriangle;
NAMESPACE_END(Grid);

13
Grid/qcd/action/fermion/implementation/CayleyFermion5DImplementation.h
View File

@ -828,6 +828,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
#if (!defined(GRID_HIP))
int tshift = (mu == Nd-1) ? 1 : 0;
unsigned int LLt = GridDefaultLatt()[Tp];
////////////////////////////////////////////////
// GENERAL CAYLEY CASE
////////////////////////////////////////////////
@ -880,7 +881,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
}
std::vector<RealD> G_s(Ls,1.0);
RealD sign = 1; // sign flip for vector/tadpole
RealD sign = 1.0; // sign flip for vector/tadpole
if ( curr_type == Current::Axial ) {
for(int s=0;s<Ls/2;s++){
G_s[s] = -1.0;
@ -890,7 +891,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
auto b=this->_b;
auto c=this->_c;
if ( b == 1 && c == 0 ) {
sign = -1;
sign = -1.0;
}
else {
std::cerr << "Error: Tadpole implementation currently unavailable for non-Shamir actions." << std::endl;
@ -934,7 +935,13 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
tmp = Cshift(tmp,mu,-1);
Impl::multLinkField(Utmp,this->Umu,tmp,mu+Nd); // Adjoint link
tmp = -G_s[s]*( Utmp + gmu*Utmp );
tmp = where((lcoor>=tmin+tshift),tmp,zz); // Mask the time
// Mask the time
if (tmax == LLt - 1 && tshift == 1){ // quick fix to include timeslice 0 if tmax + tshift is over the last timeslice
unsigned int t0 = 0;
tmp = where(((lcoor==t0) || (lcoor>=tmin+tshift)),tmp,zz);
} else {
tmp = where((lcoor>=tmin+tshift),tmp,zz);
}
L_Q += where((lcoor<=tmax+tshift),tmp,zz); // Position of current complicated
InsertSlice(L_Q, q_out, s , 0);

363
Grid/qcd/action/fermion/implementation/CompactWilsonCloverFermionImplementation.h Normal file
View File

@ -0,0 +1,363 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/CompactWilsonCloverFermionImplementation.h
Copyright (C) 2017 - 2022
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Guido Cossu <guido.cossu@ed.ac.uk>
Author: Daniel Richtmann <daniel.richtmann@gmail.com>
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>
#include <Grid/qcd/spin/Dirac.h>
#include <Grid/qcd/action/fermion/CompactWilsonCloverFermion.h>
NAMESPACE_BEGIN(Grid);
template<class Impl>
CompactWilsonCloverFermion<Impl>::CompactWilsonCloverFermion(GaugeField& _Umu,
GridCartesian& Fgrid,
GridRedBlackCartesian& Hgrid,
const RealD _mass,
const RealD _csw_r,
const RealD _csw_t,
const RealD _cF,
const WilsonAnisotropyCoefficients& clover_anisotropy,
const ImplParams& impl_p)
: WilsonBase(_Umu, Fgrid, Hgrid, _mass, impl_p, clover_anisotropy)
, csw_r(_csw_r)
, csw_t(_csw_t)
, cF(_cF)
, open_boundaries(impl_p.boundary_phases[Nd-1] == 0.0)
, Diagonal(&Fgrid), Triangle(&Fgrid)
, DiagonalEven(&Hgrid), TriangleEven(&Hgrid)
, DiagonalOdd(&Hgrid), TriangleOdd(&Hgrid)
, DiagonalInv(&Fgrid), TriangleInv(&Fgrid)
, DiagonalInvEven(&Hgrid), TriangleInvEven(&Hgrid)
, DiagonalInvOdd(&Hgrid), TriangleInvOdd(&Hgrid)
, Tmp(&Fgrid)
, BoundaryMask(&Fgrid)
, BoundaryMaskEven(&Hgrid), BoundaryMaskOdd(&Hgrid)
{
csw_r *= 0.5;
csw_t *= 0.5;
if (clover_anisotropy.isAnisotropic)
csw_r /= clover_anisotropy.xi_0;
ImportGauge(_Umu);
if (open_boundaries)
CompactHelpers::SetupMasks(this->BoundaryMask, this->BoundaryMaskEven, this->BoundaryMaskOdd);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::Dhop(const FermionField& in, FermionField& out, int dag) {
WilsonBase::Dhop(in, out, dag);
if(open_boundaries) ApplyBoundaryMask(out);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::DhopOE(const FermionField& in, FermionField& out, int dag) {
WilsonBase::DhopOE(in, out, dag);
if(open_boundaries) ApplyBoundaryMask(out);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::DhopEO(const FermionField& in, FermionField& out, int dag) {
WilsonBase::DhopEO(in, out, dag);
if(open_boundaries) ApplyBoundaryMask(out);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::DhopDir(const FermionField& in, FermionField& out, int dir, int disp) {
WilsonBase::DhopDir(in, out, dir, disp);
if(this->open_boundaries) ApplyBoundaryMask(out);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::DhopDirAll(const FermionField& in, std::vector<FermionField>& out) {
WilsonBase::DhopDirAll(in, out);
if(this->open_boundaries) {
for(auto& o : out) ApplyBoundaryMask(o);
}
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::M(const FermionField& in, FermionField& out) {
out.Checkerboard() = in.Checkerboard();
WilsonBase::Dhop(in, out, DaggerNo); // call base to save applying bc
Mooee(in, Tmp);
axpy(out, 1.0, out, Tmp);
if(open_boundaries) ApplyBoundaryMask(out);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::Mdag(const FermionField& in, FermionField& out) {
out.Checkerboard() = in.Checkerboard();
WilsonBase::Dhop(in, out, DaggerYes); // call base to save applying bc
MooeeDag(in, Tmp);
axpy(out, 1.0, out, Tmp);
if(open_boundaries) ApplyBoundaryMask(out);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::Meooe(const FermionField& in, FermionField& out) {
WilsonBase::Meooe(in, out);
if(open_boundaries) ApplyBoundaryMask(out);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::MeooeDag(const FermionField& in, FermionField& out) {
WilsonBase::MeooeDag(in, out);
if(open_boundaries) ApplyBoundaryMask(out);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::Mooee(const FermionField& in, FermionField& out) {
if(in.Grid()->_isCheckerBoarded) {
if(in.Checkerboard() == Odd) {
MooeeInternal(in, out, DiagonalOdd, TriangleOdd);
} else {
MooeeInternal(in, out, DiagonalEven, TriangleEven);
}
} else {
MooeeInternal(in, out, Diagonal, Triangle);
}
if(open_boundaries) ApplyBoundaryMask(out);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::MooeeDag(const FermionField& in, FermionField& out) {
Mooee(in, out); // blocks are hermitian
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::MooeeInv(const FermionField& in, FermionField& out) {
if(in.Grid()->_isCheckerBoarded) {
if(in.Checkerboard() == Odd) {
MooeeInternal(in, out, DiagonalInvOdd, TriangleInvOdd);
} else {
MooeeInternal(in, out, DiagonalInvEven, TriangleInvEven);
}
} else {
MooeeInternal(in, out, DiagonalInv, TriangleInv);
}
if(open_boundaries) ApplyBoundaryMask(out);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::MooeeInvDag(const FermionField& in, FermionField& out) {
MooeeInv(in, out); // blocks are hermitian
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::Mdir(const FermionField& in, FermionField& out, int dir, int disp) {
DhopDir(in, out, dir, disp);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::MdirAll(const FermionField& in, std::vector<FermionField>& out) {
DhopDirAll(in, out);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) {
assert(!open_boundaries); // TODO check for changes required for open bc
// NOTE: code copied from original clover term
conformable(X.Grid(), Y.Grid());
conformable(X.Grid(), force.Grid());
GaugeLinkField force_mu(force.Grid()), lambda(force.Grid());
GaugeField clover_force(force.Grid());
PropagatorField Lambda(force.Grid());
// Guido: Here we are hitting some performance issues:
// need to extract the components of the DoubledGaugeField
// for each call
// Possible solution
// Create a vector object to store them? (cons: wasting space)
std::vector<GaugeLinkField> U(Nd, this->Umu.Grid());
Impl::extractLinkField(U, this->Umu);
force = Zero();
// Derivative of the Wilson hopping term
this->DhopDeriv(force, X, Y, dag);
///////////////////////////////////////////////////////////
// Clover term derivative
///////////////////////////////////////////////////////////
Impl::outerProductImpl(Lambda, X, Y);
//std::cout << "Lambda:" << Lambda << std::endl;
Gamma::Algebra sigma[] = {
Gamma::Algebra::SigmaXY,
Gamma::Algebra::SigmaXZ,
Gamma::Algebra::SigmaXT,
Gamma::Algebra::MinusSigmaXY,
Gamma::Algebra::SigmaYZ,
Gamma::Algebra::SigmaYT,
Gamma::Algebra::MinusSigmaXZ,
Gamma::Algebra::MinusSigmaYZ,
Gamma::Algebra::SigmaZT,
Gamma::Algebra::MinusSigmaXT,
Gamma::Algebra::MinusSigmaYT,
Gamma::Algebra::MinusSigmaZT};
/*
sigma_{\mu \nu}=
| 0 sigma[0] sigma[1] sigma[2] |
| sigma[3] 0 sigma[4] sigma[5] |
| sigma[6] sigma[7] 0 sigma[8] |
| sigma[9] sigma[10] sigma[11] 0 |
*/
int count = 0;
clover_force = Zero();
for (int mu = 0; mu < 4; mu++)
{
force_mu = Zero();
for (int nu = 0; nu < 4; nu++)
{
if (mu == nu)
continue;
RealD factor;
if (nu == 4 || mu == 4)
{
factor = 2.0 * csw_t;
}
else
{
factor = 2.0 * csw_r;
}
PropagatorField Slambda = Gamma(sigma[count]) * Lambda; // sigma checked
Impl::TraceSpinImpl(lambda, Slambda); // traceSpin ok
force_mu -= factor*Helpers::Cmunu(U, lambda, mu, nu); // checked
count++;
}
pokeLorentz(clover_force, U[mu] * force_mu, mu);
}
//clover_force *= csw;
force += clover_force;
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::MooDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
assert(0);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::MeeDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
assert(0);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::MooeeInternal(const FermionField& in,
FermionField& out,
const CloverDiagonalField& diagonal,
const CloverTriangleField& triangle) {
assert(in.Checkerboard() == Odd || in.Checkerboard() == Even);
out.Checkerboard() = in.Checkerboard();
conformable(in, out);
conformable(in, diagonal);
conformable(in, triangle);
CompactHelpers::MooeeKernel(diagonal.oSites(), 1, in, out, diagonal, triangle);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::ImportGauge(const GaugeField& _Umu) {
// NOTE: parts copied from original implementation
// Import gauge into base class
double t0 = usecond();
WilsonBase::ImportGauge(_Umu); // NOTE: called here and in wilson constructor -> performed twice, but can't avoid that
// Initialize temporary variables
double t1 = usecond();
conformable(_Umu.Grid(), this->GaugeGrid());
GridBase* grid = _Umu.Grid();
typename Impl::GaugeLinkField Bx(grid), By(grid), Bz(grid), Ex(grid), Ey(grid), Ez(grid);
CloverField TmpOriginal(grid);
// Compute the field strength terms mu>nu
double t2 = usecond();
WilsonLoops<Impl>::FieldStrength(Bx, _Umu, Zdir, Ydir);
WilsonLoops<Impl>::FieldStrength(By, _Umu, Zdir, Xdir);
WilsonLoops<Impl>::FieldStrength(Bz, _Umu, Ydir, Xdir);
WilsonLoops<Impl>::FieldStrength(Ex, _Umu, Tdir, Xdir);
WilsonLoops<Impl>::FieldStrength(Ey, _Umu, Tdir, Ydir);
WilsonLoops<Impl>::FieldStrength(Ez, _Umu, Tdir, Zdir);
// Compute the Clover Operator acting on Colour and Spin
// multiply here by the clover coefficients for the anisotropy
double t3 = usecond();
TmpOriginal = Helpers::fillCloverYZ(Bx) * csw_r;
TmpOriginal += Helpers::fillCloverXZ(By) * csw_r;
TmpOriginal += Helpers::fillCloverXY(Bz) * csw_r;
TmpOriginal += Helpers::fillCloverXT(Ex) * csw_t;
TmpOriginal += Helpers::fillCloverYT(Ey) * csw_t;
TmpOriginal += Helpers::fillCloverZT(Ez) * csw_t;
TmpOriginal += this->diag_mass;
// Convert the data layout of the clover term
double t4 = usecond();
CompactHelpers::ConvertLayout(TmpOriginal, Diagonal, Triangle);
// Possible modify the boundary values
double t5 = usecond();
if(open_boundaries) CompactHelpers::ModifyBoundaries(Diagonal, Triangle, csw_t, cF, this->diag_mass);
// Invert the clover term in the improved layout
double t6 = usecond();
CompactHelpers::Invert(Diagonal, Triangle, DiagonalInv, TriangleInv);
// Fill the remaining clover fields
double t7 = usecond();
pickCheckerboard(Even, DiagonalEven, Diagonal);
pickCheckerboard(Even, TriangleEven, Triangle);
pickCheckerboard(Odd, DiagonalOdd, Diagonal);
pickCheckerboard(Odd, TriangleOdd, Triangle);
pickCheckerboard(Even, DiagonalInvEven, DiagonalInv);
pickCheckerboard(Even, TriangleInvEven, TriangleInv);
pickCheckerboard(Odd, DiagonalInvOdd, DiagonalInv);
pickCheckerboard(Odd, TriangleInvOdd, TriangleInv);
// Report timings
double t8 = usecond();
#if 0
std::cout << GridLogMessage << "CompactWilsonCloverFermion::ImportGauge timings:"
<< " WilsonFermion::Importgauge = " << (t1 - t0) / 1e6
<< ", allocations = " << (t2 - t1) / 1e6
<< ", field strength = " << (t3 - t2) / 1e6
<< ", fill clover = " << (t4 - t3) / 1e6
<< ", convert = " << (t5 - t4) / 1e6
<< ", boundaries = " << (t6 - t5) / 1e6
<< ", inversions = " << (t7 - t6) / 1e6
<< ", pick cbs = " << (t8 - t7) / 1e6
<< ", total = " << (t8 - t0) / 1e6
<< std::endl;
#endif
}
NAMESPACE_END(Grid);

171
Grid/qcd/action/fermion/implementation/WilsonCloverFermionImplementation.h
View File

@ -2,12 +2,13 @@
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonCloverFermion.cc
Source file: ./lib/qcd/action/fermion/WilsonCloverFermionImplementation.h
Copyright (C) 2017
Copyright (C) 2017 - 2022
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Guido Cossu <guido.cossu@ed.ac.uk>
Author: Daniel Richtmann <daniel.richtmann@gmail.com>
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
@ -33,6 +34,45 @@
NAMESPACE_BEGIN(Grid);
template<class Impl>
WilsonCloverFermion<Impl>::WilsonCloverFermion(GaugeField& _Umu,
GridCartesian& Fgrid,
GridRedBlackCartesian& Hgrid,
const RealD _mass,
const RealD _csw_r,
const RealD _csw_t,
const WilsonAnisotropyCoefficients& clover_anisotropy,
const ImplParams& impl_p)
: WilsonFermion<Impl>(_Umu, Fgrid, Hgrid, _mass, impl_p, clover_anisotropy)
, CloverTerm(&Fgrid)
, CloverTermInv(&Fgrid)
, CloverTermEven(&Hgrid)
, CloverTermOdd(&Hgrid)
, CloverTermInvEven(&Hgrid)
, CloverTermInvOdd(&Hgrid)
, CloverTermDagEven(&Hgrid)
, CloverTermDagOdd(&Hgrid)
, CloverTermInvDagEven(&Hgrid)
, CloverTermInvDagOdd(&Hgrid) {
assert(Nd == 4); // require 4 dimensions
if(clover_anisotropy.isAnisotropic) {
csw_r = _csw_r * 0.5 / clover_anisotropy.xi_0;
diag_mass = _mass + 1.0 + (Nd - 1) * (clover_anisotropy.nu / clover_anisotropy.xi_0);
} else {
csw_r = _csw_r * 0.5;
diag_mass = 4.0 + _mass;
}
csw_t = _csw_t * 0.5;
if(csw_r == 0)
std::cout << GridLogWarning << "Initializing WilsonCloverFermion with csw_r = 0" << std::endl;
if(csw_t == 0)
std::cout << GridLogWarning << "Initializing WilsonCloverFermion with csw_t = 0" << std::endl;
ImportGauge(_Umu);
}
// *NOT* EO
template <class Impl>
void WilsonCloverFermion<Impl>::M(const FermionField &in, FermionField &out)
@ -67,10 +107,13 @@ void WilsonCloverFermion<Impl>::Mdag(const FermionField &in, FermionField &out)
template <class Impl>
void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
{
double t0 = usecond();
WilsonFermion<Impl>::ImportGauge(_Umu);
double t1 = usecond();
GridBase *grid = _Umu.Grid();
typename Impl::GaugeLinkField Bx(grid), By(grid), Bz(grid), Ex(grid), Ey(grid), Ez(grid);
double t2 = usecond();
// Compute the field strength terms mu>nu
WilsonLoops<Impl>::FieldStrength(Bx, _Umu, Zdir, Ydir);
WilsonLoops<Impl>::FieldStrength(By, _Umu, Zdir, Xdir);
@ -79,19 +122,22 @@ void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
WilsonLoops<Impl>::FieldStrength(Ey, _Umu, Tdir, Ydir);
WilsonLoops<Impl>::FieldStrength(Ez, _Umu, Tdir, Zdir);
double t3 = usecond();
// Compute the Clover Operator acting on Colour and Spin
// multiply here by the clover coefficients for the anisotropy
CloverTerm = fillCloverYZ(Bx) * csw_r;
CloverTerm += fillCloverXZ(By) * csw_r;
CloverTerm += fillCloverXY(Bz) * csw_r;
CloverTerm += fillCloverXT(Ex) * csw_t;
CloverTerm += fillCloverYT(Ey) * csw_t;
CloverTerm += fillCloverZT(Ez) * csw_t;
CloverTerm = Helpers::fillCloverYZ(Bx) * csw_r;
CloverTerm += Helpers::fillCloverXZ(By) * csw_r;
CloverTerm += Helpers::fillCloverXY(Bz) * csw_r;
CloverTerm += Helpers::fillCloverXT(Ex) * csw_t;
CloverTerm += Helpers::fillCloverYT(Ey) * csw_t;
CloverTerm += Helpers::fillCloverZT(Ez) * csw_t;
CloverTerm += diag_mass;
double t4 = usecond();
int lvol = _Umu.Grid()->lSites();
int DimRep = Impl::Dimension;
double t5 = usecond();
{
autoView(CTv,CloverTerm,CpuRead);
autoView(CTIv,CloverTermInv,CpuWrite);
@ -100,7 +146,7 @@ void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
grid->LocalIndexToLocalCoor(site, lcoor);
Eigen::MatrixXcd EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
Eigen::MatrixXcd EigenInvCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
typename SiteCloverType::scalar_object Qx = Zero(), Qxinv = Zero();
typename SiteClover::scalar_object Qx = Zero(), Qxinv = Zero();
peekLocalSite(Qx, CTv, lcoor);
//if (csw!=0){
for (int j = 0; j < Ns; j++)
@ -125,6 +171,7 @@ void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
});
}
double t6 = usecond();
// Separate the even and odd parts
pickCheckerboard(Even, CloverTermEven, CloverTerm);
pickCheckerboard(Odd, CloverTermOdd, CloverTerm);
@ -137,6 +184,20 @@ void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
pickCheckerboard(Even, CloverTermInvDagEven, adj(CloverTermInv));
pickCheckerboard(Odd, CloverTermInvDagOdd, adj(CloverTermInv));
double t7 = usecond();
#if 0
std::cout << GridLogMessage << "WilsonCloverFermion::ImportGauge timings:"
<< " WilsonFermion::Importgauge = " << (t1 - t0) / 1e6
<< ", allocations = " << (t2 - t1) / 1e6
<< ", field strength = " << (t3 - t2) / 1e6
<< ", fill clover = " << (t4 - t3) / 1e6
<< ", misc = " << (t5 - t4) / 1e6
<< ", inversions = " << (t6 - t5) / 1e6
<< ", pick cbs = " << (t7 - t6) / 1e6
<< ", total = " << (t7 - t0) / 1e6
<< std::endl;
#endif
}
template <class Impl>
@ -167,7 +228,7 @@ template <class Impl>
void WilsonCloverFermion<Impl>::MooeeInternal(const FermionField &in, FermionField &out, int dag, int inv)
{
out.Checkerboard() = in.Checkerboard();
CloverFieldType *Clover;
CloverField *Clover;
assert(in.Checkerboard() == Odd || in.Checkerboard() == Even);
if (dag)
@ -182,12 +243,12 @@ void WilsonCloverFermion<Impl>::MooeeInternal(const FermionField &in, FermionFie
{
Clover = (inv) ? &CloverTermInvDagEven : &CloverTermDagEven;
}
out = *Clover * in;
Helpers::multCloverField(out, *Clover, in);
}
else
{
Clover = (inv) ? &CloverTermInv : &CloverTerm;
out = adj(*Clover) * in;
Helpers::multCloverField(out, *Clover, in); // don't bother with adj, hermitian anyway
}
}
else
@ -205,18 +266,98 @@ void WilsonCloverFermion<Impl>::MooeeInternal(const FermionField &in, FermionFie
// std::cout << "Calling clover term Even" << std::endl;
Clover = (inv) ? &CloverTermInvEven : &CloverTermEven;
}
out = *Clover * in;
Helpers::multCloverField(out, *Clover, in);
// std::cout << GridLogMessage << "*Clover.Checkerboard() " << (*Clover).Checkerboard() << std::endl;
}
else
{
Clover = (inv) ? &CloverTermInv : &CloverTerm;
out = *Clover * in;
Helpers::multCloverField(out, *Clover, in);
}
}
} // MooeeInternal
// Derivative parts unpreconditioned pseudofermions
template <class Impl>
void WilsonCloverFermion<Impl>::MDeriv(GaugeField &force, const FermionField &X, const FermionField &Y, int dag)
{
conformable(X.Grid(), Y.Grid());
conformable(X.Grid(), force.Grid());
GaugeLinkField force_mu(force.Grid()), lambda(force.Grid());
GaugeField clover_force(force.Grid());
PropagatorField Lambda(force.Grid());
// Guido: Here we are hitting some performance issues:
// need to extract the components of the DoubledGaugeField
// for each call
// Possible solution
// Create a vector object to store them? (cons: wasting space)
std::vector<GaugeLinkField> U(Nd, this->Umu.Grid());
Impl::extractLinkField(U, this->Umu);
force = Zero();
// Derivative of the Wilson hopping term
this->DhopDeriv(force, X, Y, dag);
///////////////////////////////////////////////////////////
// Clover term derivative
///////////////////////////////////////////////////////////
Impl::outerProductImpl(Lambda, X, Y);
//std::cout << "Lambda:" << Lambda << std::endl;
Gamma::Algebra sigma[] = {
Gamma::Algebra::SigmaXY,
Gamma::Algebra::SigmaXZ,
Gamma::Algebra::SigmaXT,
Gamma::Algebra::MinusSigmaXY,
Gamma::Algebra::SigmaYZ,
Gamma::Algebra::SigmaYT,
Gamma::Algebra::MinusSigmaXZ,
Gamma::Algebra::MinusSigmaYZ,
Gamma::Algebra::SigmaZT,
Gamma::Algebra::MinusSigmaXT,
Gamma::Algebra::MinusSigmaYT,
Gamma::Algebra::MinusSigmaZT};
/*
sigma_{\mu \nu}=
| 0 sigma[0] sigma[1] sigma[2] |
| sigma[3] 0 sigma[4] sigma[5] |
| sigma[6] sigma[7] 0 sigma[8] |
| sigma[9] sigma[10] sigma[11] 0 |
*/
int count = 0;
clover_force = Zero();
for (int mu = 0; mu < 4; mu++)
{
force_mu = Zero();
for (int nu = 0; nu < 4; nu++)
{
if (mu == nu)
continue;
RealD factor;
if (nu == 4 || mu == 4)
{
factor = 2.0 * csw_t;
}
else
{
factor = 2.0 * csw_r;
}
PropagatorField Slambda = Gamma(sigma[count]) * Lambda; // sigma checked
Impl::TraceSpinImpl(lambda, Slambda); // traceSpin ok
force_mu -= factor*Helpers::Cmunu(U, lambda, mu, nu); // checked
count++;
}
pokeLorentz(clover_force, U[mu] * force_mu, mu);
}
//clover_force *= csw;
force += clover_force;
}
// Derivative parts
template <class Impl>

218
Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h
View File

@ -77,23 +77,23 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#define REGISTER
#ifdef GRID_SIMT
#define LOAD_CHIMU(ptype) \
#define LOAD_CHIMU(Ptype) \
{const SiteSpinor & ref (in[offset]); \
Chimu_00=coalescedReadPermute<ptype>(ref()(0)(0),perm,lane); \
Chimu_01=coalescedReadPermute<ptype>(ref()(0)(1),perm,lane); \
Chimu_02=coalescedReadPermute<ptype>(ref()(0)(2),perm,lane); \
Chimu_10=coalescedReadPermute<ptype>(ref()(1)(0),perm,lane); \
Chimu_11=coalescedReadPermute<ptype>(ref()(1)(1),perm,lane); \
Chimu_12=coalescedReadPermute<ptype>(ref()(1)(2),perm,lane); \
Chimu_20=coalescedReadPermute<ptype>(ref()(2)(0),perm,lane); \
Chimu_21=coalescedReadPermute<ptype>(ref()(2)(1),perm,lane); \
Chimu_22=coalescedReadPermute<ptype>(ref()(2)(2),perm,lane); \
Chimu_30=coalescedReadPermute<ptype>(ref()(3)(0),perm,lane); \
Chimu_31=coalescedReadPermute<ptype>(ref()(3)(1),perm,lane); \
Chimu_32=coalescedReadPermute<ptype>(ref()(3)(2),perm,lane); }
Chimu_00=coalescedReadPermute<Ptype>(ref()(0)(0),perm,lane); \
Chimu_01=coalescedReadPermute<Ptype>(ref()(0)(1),perm,lane); \
Chimu_02=coalescedReadPermute<Ptype>(ref()(0)(2),perm,lane); \
Chimu_10=coalescedReadPermute<Ptype>(ref()(1)(0),perm,lane); \
Chimu_11=coalescedReadPermute<Ptype>(ref()(1)(1),perm,lane); \
Chimu_12=coalescedReadPermute<Ptype>(ref()(1)(2),perm,lane); \
Chimu_20=coalescedReadPermute<Ptype>(ref()(2)(0),perm,lane); \
Chimu_21=coalescedReadPermute<Ptype>(ref()(2)(1),perm,lane); \
Chimu_22=coalescedReadPermute<Ptype>(ref()(2)(2),perm,lane); \
Chimu_30=coalescedReadPermute<Ptype>(ref()(3)(0),perm,lane); \
Chimu_31=coalescedReadPermute<Ptype>(ref()(3)(1),perm,lane); \
Chimu_32=coalescedReadPermute<Ptype>(ref()(3)(2),perm,lane); }
#define PERMUTE_DIR(dir) ;
#else
#define LOAD_CHIMU(ptype) \
#define LOAD_CHIMU(Ptype) \
{const SiteSpinor & ref (in[offset]); \
Chimu_00=ref()(0)(0);\
Chimu_01=ref()(0)(1);\
@ -109,12 +109,12 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
Chimu_32=ref()(3)(2);}
#define PERMUTE_DIR(dir) \
permute##dir(Chi_00,Chi_00); \
permute##dir(Chi_01,Chi_01);\
permute##dir(Chi_02,Chi_02);\
permute##dir(Chi_10,Chi_10); \
permute##dir(Chi_11,Chi_11);\
permute##dir(Chi_12,Chi_12);
permute##dir(Chi_00,Chi_00); \
permute##dir(Chi_01,Chi_01); \
permute##dir(Chi_02,Chi_02); \
permute##dir(Chi_10,Chi_10); \
permute##dir(Chi_11,Chi_11); \
permute##dir(Chi_12,Chi_12);
#endif
@ -371,88 +371,91 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
result_32-= UChi_12;
#define HAND_STENCIL_LEGB(PROJ,PERM,DIR,RECON) \
SE=st.GetEntry(ptype,DIR,ss); \
offset = SE->_offset; \
local = SE->_is_local; \
perm = SE->_permute; \
if ( local ) { \
LOAD_CHIMU(PERM); \
PROJ; \
if ( perm) { \
PERMUTE_DIR(PERM); \
} \
} else { \
LOAD_CHI; \
} \
acceleratorSynchronise(); \
MULT_2SPIN(DIR); \
RECON;
{int ptype; \
SE=st.GetEntry(ptype,DIR,ss); \
auto offset = SE->_offset; \
auto local = SE->_is_local; \
auto perm = SE->_permute; \
if ( local ) { \
LOAD_CHIMU(PERM); \
PROJ; \
if ( perm) { \
PERMUTE_DIR(PERM); \
} \
} else { \
LOAD_CHI; \
} \
acceleratorSynchronise(); \
MULT_2SPIN(DIR); \
RECON; }
#define HAND_STENCIL_LEG(PROJ,PERM,DIR,RECON) \
SE=&st_p[DIR+8*ss]; \
ptype=st_perm[DIR]; \
offset = SE->_offset; \
local = SE->_is_local; \
perm = SE->_permute; \
if ( local ) { \
LOAD_CHIMU(PERM); \
PROJ; \
if ( perm) { \
PERMUTE_DIR(PERM); \
} \
} else { \
LOAD_CHI; \
} \
acceleratorSynchronise(); \
MULT_2SPIN(DIR); \
RECON;
#define HAND_STENCIL_LEG(PROJ,PERM,DIR,RECON) \
{ SE=&st_p[DIR+8*ss]; \
auto ptype=st_perm[DIR]; \
auto offset = SE->_offset; \
auto local = SE->_is_local; \
auto perm = SE->_permute; \
if ( local ) { \
LOAD_CHIMU(PERM); \
PROJ; \
if ( perm) { \
PERMUTE_DIR(PERM); \
} \
} else { \
LOAD_CHI; \
} \
acceleratorSynchronise(); \
MULT_2SPIN(DIR); \
RECON; }
#define HAND_STENCIL_LEGA(PROJ,PERM,DIR,RECON) \
SE=&st_p[DIR+8*ss]; \
ptype=st_perm[DIR]; \
/*SE=st.GetEntry(ptype,DIR,ss);*/ \
offset = SE->_offset; \
perm = SE->_permute; \
LOAD_CHIMU(PERM); \
PROJ; \
MULT_2SPIN(DIR); \
RECON;
{ SE=&st_p[DIR+8*ss]; \
auto ptype=st_perm[DIR]; \
/*SE=st.GetEntry(ptype,DIR,ss);*/ \
auto offset = SE->_offset; \
auto perm = SE->_permute; \
LOAD_CHIMU(PERM); \
PROJ; \
MULT_2SPIN(DIR); \
RECON; }
#define HAND_STENCIL_LEG_INT(PROJ,PERM,DIR,RECON) \
SE=st.GetEntry(ptype,DIR,ss); \
offset = SE->_offset; \
local = SE->_is_local; \
perm = SE->_permute; \
if ( local ) { \
LOAD_CHIMU(PERM); \
PROJ; \
if ( perm) { \
PERMUTE_DIR(PERM); \
} \
} else if ( st.same_node[DIR] ) { \
LOAD_CHI; \
} \
acceleratorSynchronise(); \
if (local || st.same_node[DIR] ) { \
MULT_2SPIN(DIR); \
RECON; \
} \
acceleratorSynchronise();
{ int ptype; \
SE=st.GetEntry(ptype,DIR,ss); \
auto offset = SE->_offset; \
auto local = SE->_is_local; \
auto perm = SE->_permute; \
if ( local ) { \
LOAD_CHIMU(PERM); \
PROJ; \
if ( perm) { \
PERMUTE_DIR(PERM); \
} \
} else if ( st.same_node[DIR] ) { \
LOAD_CHI; \
} \
acceleratorSynchronise(); \
if (local || st.same_node[DIR] ) { \
MULT_2SPIN(DIR); \
RECON; \
} \
acceleratorSynchronise(); }
#define HAND_STENCIL_LEG_EXT(PROJ,PERM,DIR,RECON) \
SE=st.GetEntry(ptype,DIR,ss); \
offset = SE->_offset; \
if((!SE->_is_local)&&(!st.same_node[DIR]) ) { \
LOAD_CHI; \
MULT_2SPIN(DIR); \
RECON; \
nmu++; \
} \
acceleratorSynchronise();
{ int ptype; \
SE=st.GetEntry(ptype,DIR,ss); \
auto offset = SE->_offset; \
if((!SE->_is_local)&&(!st.same_node[DIR]) ) { \
LOAD_CHI; \
MULT_2SPIN(DIR); \
RECON; \
nmu++; \
} \
acceleratorSynchronise(); }
#define HAND_RESULT(ss) \
{ \
SiteSpinor & ref (out[ss]); \
#define HAND_RESULT(ss) \
{ \
SiteSpinor & ref (out[ss]); \
coalescedWrite(ref()(0)(0),result_00,lane); \
coalescedWrite(ref()(0)(1),result_01,lane); \
coalescedWrite(ref()(0)(2),result_02,lane); \
@ -563,7 +566,6 @@ WilsonKernels<Impl>::HandDhopSiteSycl(StencilVector st_perm,StencilEntry *st_p,
HAND_DECLARATIONS(Simt);
int offset,local,perm, ptype;
StencilEntry *SE;
HAND_STENCIL_LEG(XM_PROJ,3,Xp,XM_RECON);
HAND_STENCIL_LEG(YM_PROJ,2,Yp,YM_RECON_ACCUM);
@ -593,9 +595,7 @@ WilsonKernels<Impl>::HandDhopSite(StencilView &st, DoubledGaugeFieldView &U,Site
HAND_DECLARATIONS(Simt);
int offset,local,perm, ptype;
StencilEntry *SE;
HAND_STENCIL_LEG(XM_PROJ,3,Xp,XM_RECON);
HAND_STENCIL_LEG(YM_PROJ,2,Yp,YM_RECON_ACCUM);
HAND_STENCIL_LEG(ZM_PROJ,1,Zp,ZM_RECON_ACCUM);
@ -623,8 +623,6 @@ void WilsonKernels<Impl>::HandDhopSiteDag(StencilView &st,DoubledGaugeFieldView
HAND_DECLARATIONS(Simt);
StencilEntry *SE;
int offset,local,perm, ptype;
HAND_STENCIL_LEG(XP_PROJ,3,Xp,XP_RECON);
HAND_STENCIL_LEG(YP_PROJ,2,Yp,YP_RECON_ACCUM);
HAND_STENCIL_LEG(ZP_PROJ,1,Zp,ZP_RECON_ACCUM);
@ -640,8 +638,8 @@ template<class Impl> accelerator_inline void
WilsonKernels<Impl>::HandDhopSiteInt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
{
auto st_p = st._entries_p;
auto st_perm = st._permute_type;
// auto st_p = st._entries_p;
// auto st_perm = st._permute_type;
// T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
typedef typename Simd::scalar_type S;
typedef typename Simd::vector_type V;
@ -652,7 +650,6 @@ WilsonKernels<Impl>::HandDhopSiteInt(StencilView &st,DoubledGaugeFieldView &U,Si
HAND_DECLARATIONS(Simt);
int offset,local,perm, ptype;
StencilEntry *SE;
ZERO_RESULT;
HAND_STENCIL_LEG_INT(XM_PROJ,3,Xp,XM_RECON_ACCUM);
@ -670,8 +667,8 @@ template<class Impl> accelerator_inline
void WilsonKernels<Impl>::HandDhopSiteDagInt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
{
auto st_p = st._entries_p;
auto st_perm = st._permute_type;
// auto st_p = st._entries_p;
// auto st_perm = st._permute_type;
typedef typename Simd::scalar_type S;
typedef typename Simd::vector_type V;
typedef decltype( coalescedRead( in[0]()(0)(0) )) Simt;
@ -682,7 +679,6 @@ void WilsonKernels<Impl>::HandDhopSiteDagInt(StencilView &st,DoubledGaugeFieldVi
HAND_DECLARATIONS(Simt);
StencilEntry *SE;
int offset,local,perm, ptype;
ZERO_RESULT;
HAND_STENCIL_LEG_INT(XP_PROJ,3,Xp,XP_RECON_ACCUM);
HAND_STENCIL_LEG_INT(YP_PROJ,2,Yp,YP_RECON_ACCUM);
@ -699,8 +695,8 @@ template<class Impl> accelerator_inline void
WilsonKernels<Impl>::HandDhopSiteExt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
{
auto st_p = st._entries_p;
auto st_perm = st._permute_type;
// auto st_p = st._entries_p;
// auto st_perm = st._permute_type;
// T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
typedef typename Simd::scalar_type S;
typedef typename Simd::vector_type V;
@ -711,7 +707,7 @@ WilsonKernels<Impl>::HandDhopSiteExt(StencilView &st,DoubledGaugeFieldView &U,Si
HAND_DECLARATIONS(Simt);
int offset, ptype;
// int offset, ptype;
StencilEntry *SE;
int nmu=0;
ZERO_RESULT;
@ -730,8 +726,8 @@ template<class Impl> accelerator_inline
void WilsonKernels<Impl>::HandDhopSiteDagExt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
{
auto st_p = st._entries_p;
auto st_perm = st._permute_type;
// auto st_p = st._entries_p;
// auto st_perm = st._permute_type;
typedef typename Simd::scalar_type S;
typedef typename Simd::vector_type V;
typedef decltype( coalescedRead( in[0]()(0)(0) )) Simt;
@ -742,7 +738,7 @@ void WilsonKernels<Impl>::HandDhopSiteDagExt(StencilView &st,DoubledGaugeFieldVi
HAND_DECLARATIONS(Simt);
StencilEntry *SE;
int offset, ptype;
// int offset, ptype;
int nmu=0;
ZERO_RESULT;
HAND_STENCIL_LEG_EXT(XP_PROJ,3,Xp,XP_RECON_ACCUM);

41
Grid/qcd/action/fermion/instantiation/CompactWilsonCloverFermionInstantiation.cc.master Normal file
View File

@ -0,0 +1,41 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/ qcd/action/fermion/instantiation/CompactWilsonCloverFermionInstantiation.cc.master
Copyright (C) 2017 - 2022
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Guido Cossu <guido.cossu@ed.ac.uk>
Author: Daniel Richtmann <daniel.richtmann@gmail.com>
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>
#include <Grid/qcd/spin/Dirac.h>
#include <Grid/qcd/action/fermion/CompactWilsonCloverFermion.h>
#include <Grid/qcd/action/fermion/implementation/CompactWilsonCloverFermionImplementation.h>
NAMESPACE_BEGIN(Grid);
#include "impl.h"
template class CompactWilsonCloverFermion<IMPLEMENTATION>;
NAMESPACE_END(Grid);

1
Grid/qcd/action/fermion/instantiation/WilsonImplD/CompactWilsonCloverFermionInstantiationWilsonImplD.cc Symbolic link
View File

@ -0,0 +1 @@
../CompactWilsonCloverFermionInstantiation.cc.master

1
Grid/qcd/action/fermion/instantiation/WilsonImplF/CompactWilsonCloverFermionInstantiationWilsonImplF.cc Symbolic link
View File

@ -0,0 +1 @@
../CompactWilsonCloverFermionInstantiation.cc.master

2
Grid/qcd/action/fermion/instantiation/generate_instantiations.sh
View File

@ -40,7 +40,7 @@ EOF
done
CC_LIST="WilsonCloverFermionInstantiation WilsonFermionInstantiation WilsonKernelsInstantiation WilsonTMFermionInstantiation"
CC_LIST="WilsonCloverFermionInstantiation CompactWilsonCloverFermionInstantiation WilsonFermionInstantiation WilsonKernelsInstantiation WilsonTMFermionInstantiation"
for impl in $WILSON_IMPL_LIST
do

10
Grid/qcd/action/gauge/GaugeImplTypes.h
View File

@ -78,6 +78,8 @@ public:
typedef Lattice<SiteLink> LinkField;
typedef Lattice<SiteField> Field;
typedef SU<Nrepresentation> Group;
// Guido: we can probably separate the types from the HMC functions
// this will create 2 kind of implementations
// probably confusing the users
@ -118,7 +120,7 @@ public:
LinkField Pmu(P.Grid());
Pmu = Zero();
for (int mu = 0; mu < Nd; mu++) {
SU<Nrepresentation>::GaussianFundamentalLieAlgebraMatrix(pRNG, Pmu);
Group::GaussianFundamentalLieAlgebraMatrix(pRNG, Pmu);
RealD scale = ::sqrt(HMC_MOMENTUM_DENOMINATOR) ;
Pmu = Pmu*scale;
PokeIndex<LorentzIndex>(P, Pmu, mu);
@ -159,15 +161,15 @@ public:
}
static inline void HotConfiguration(GridParallelRNG &pRNG, Field &U) {
SU<Nc>::HotConfiguration(pRNG, U);
Group::HotConfiguration(pRNG, U);
}
static inline void TepidConfiguration(GridParallelRNG &pRNG, Field &U) {
SU<Nc>::TepidConfiguration(pRNG, U);
Group::TepidConfiguration(pRNG, U);
}
static inline void ColdConfiguration(GridParallelRNG &pRNG, Field &U) {
SU<Nc>::ColdConfiguration(pRNG, U);
Group::ColdConfiguration(pRNG, U);
}
};

90
Grid/qcd/hmc/UsingHMC.md
View File

@ -1,61 +1,63 @@
Using HMC in Grid version 0.5.1
# Using HMC in Grid
These are the instructions to use the Generalised HMC on Grid version 0.5.1.
Disclaimer: GRID is still under active development so any information here can be changed in future releases.
These are the instructions to use the Generalised HMC on Grid as of commit `749b802`.
Disclaimer: Grid is still under active development so any information here can be changed in future releases.
Command line options
===================
(relevant file GenericHMCrunner.h)
## Command line options
(relevant file `GenericHMCrunner.h`)
The initial configuration can be changed at the command line using
--StartType <your choice>
valid choices, one among these
HotStart, ColdStart, TepidStart, CheckpointStart
default: HotStart
`--StartingType STARTING_TYPE`, where `STARTING_TYPE` is one of
`HotStart`, `ColdStart`, `TepidStart`, and `CheckpointStart`.
Default: `--StartingType HotStart`
example
./My_hmc_exec --StartType HotStart
Example:
```
./My_hmc_exec --StartingType HotStart
```
The CheckpointStart option uses the prefix for the configurations and rng seed files defined in your executable and the initial configuration is specified by
--StartTrajectory <integer>
default: 0
The `CheckpointStart` option uses the prefix for the configurations and rng seed files defined in your executable and the initial configuration is specified by
`--StartingTrajectory STARTING_TRAJECTORY`, where `STARTING_TRAJECTORY` is an integer.
Default: `--StartingTrajectory 0`
The number of trajectories for a specific run are specified at command line by
--Trajectories <integer>
default: 1
`--Trajectories TRAJECTORIES`, where `TRAJECTORIES` is an integer.
Default: `--Trajectories 1`
The number of thermalization steps (i.e. steps when the Metropolis acceptance check is turned off) is specified by
--Thermalizations <integer>
default: 10
`--Thermalizations THERMALIZATIONS`, where `THERMALIZATIONS` is an integer.
Default: `--Thermalizations 10`
Any other parameter is defined in the source for the executable.
HMC controls
===========
## HMC controls
The lines
```
std::vector<int> SerSeed({1, 2, 3, 4, 5});
std::vector<int> ParSeed({6, 7, 8, 9, 10});
```
define the seeds for the serial and the parallel RNG.
The line
```
TheHMC.MDparameters.set(20, 1.0);// MDsteps, traj length
```
declares the number of molecular dynamics steps and the total trajectory length.
Actions
======
## Actions
Action names are defined in the file
lib/qcd/Actions.h
Action names are defined in the directory `Grid/qcd/action`.
Gauge actions list:
Gauge actions list (from `Grid/qcd/action/gauge/Gauge.h`):
```
WilsonGaugeActionR;
WilsonGaugeActionF;
WilsonGaugeActionD;
@ -68,8 +70,9 @@ IwasakiGaugeActionD;
SymanzikGaugeActionR;
SymanzikGaugeActionF;
SymanzikGaugeActionD;
```
```
ConjugateWilsonGaugeActionR;
ConjugateWilsonGaugeActionF;
ConjugateWilsonGaugeActionD;
@ -82,26 +85,23 @@ ConjugateIwasakiGaugeActionD;
ConjugateSymanzikGaugeActionR;
ConjugateSymanzikGaugeActionF;
ConjugateSymanzikGaugeActionD;
```
Each of these action accepts one single parameter at creation time (beta).
Example for creating a Symanzik action with beta=4.0
```
SymanzikGaugeActionR(4.0)
```
Scalar actions list (from `Grid/qcd/action/scalar/Scalar.h`):
```
ScalarActionR;
ScalarActionF;
ScalarActionD;
```
each of these action accept one single parameter at creation time (beta).
Example for creating a Symanzik action with beta=4.0
SymanzikGaugeActionR(4.0)
The suffixes R,F,D in the action names refer to the Real
(the precision is defined at compile time by the --enable-precision flag in the configure),
Float and Double, that force the precision of the action to be 32, 64 bit respectively.
The suffixes `R`, `F`, `D` in the action names refer to the `Real`
(the precision is defined at compile time by the `--enable-precision` flag in the configure),
`Float` and `Double`, that force the precision of the action to be 32, 64 bit respectively.

4
Grid/stencil/Stencil.h
View File

@ -322,8 +322,8 @@ public:
int simd_layout = _grid->_simd_layout[dimension];
int comm_dim = _grid->_processors[dimension] >1 ;
int recv_from_rank;
int xmit_to_rank;
// int recv_from_rank;
// int xmit_to_rank;
if ( ! comm_dim ) return 1;
if ( displacement == 0 ) return 1;

6
Grid/tensors/Tensor_traits.h
View File

@ -47,20 +47,20 @@ NAMESPACE_BEGIN(Grid);
class TypePair {
public:
T _internal[2];
TypePair<T>& operator=(const Grid::Zero& o) {
accelerator TypePair<T>& operator=(const Grid::Zero& o) {
_internal[0] = Zero();
_internal[1] = Zero();
return *this;
}
TypePair<T> operator+(const TypePair<T>& o) const {
accelerator TypePair<T> operator+(const TypePair<T>& o) const {
TypePair<T> r;
r._internal[0] = _internal[0] + o._internal[0];
r._internal[1] = _internal[1] + o._internal[1];
return r;
}
TypePair<T>& operator+=(const TypePair<T>& o) {
accelerator TypePair<T>& operator+=(const TypePair<T>& o) {
_internal[0] += o._internal[0];
_internal[1] += o._internal[1];
return *this;

33
Grid/threads/Accelerator.cc
View File

@ -74,31 +74,43 @@ void acceleratorInit(void)
// GPU_PROP(singleToDoublePrecisionPerfRatio);
}
}
MemoryManager::DeviceMaxBytes = (8*totalDeviceMem)/10; // Assume 80% ours
#undef GPU_PROP_FMT
#undef GPU_PROP
#ifdef GRID_DEFAULT_GPU
int device = 0;
// IBM Jsrun makes cuda Device numbering screwy and not match rank
if ( world_rank == 0 ) {
printf("AcceleratorCudaInit: using default device \n");
printf("AcceleratorCudaInit: assume user either uses a) IBM jsrun, or \n");
printf("AcceleratorCudaInit: assume user either uses\n");
printf("AcceleratorCudaInit: a) IBM jsrun, or \n");
printf("AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding \n");
printf("AcceleratorCudaInit: Configure options --enable-setdevice=no \n");
}
#else
int device = rank;
printf("AcceleratorCudaInit: rank %d setting device to node rank %d\n",world_rank,rank);
printf("AcceleratorCudaInit: Configure options --enable-setdevice=yes \n");
cudaSetDevice(rank);
#endif
cudaSetDevice(device);
cudaStreamCreate(&copyStream);
const int len=64;
char busid[len];
if( rank == world_rank ) {
cudaDeviceGetPCIBusId(busid, len, device);
printf("local rank %d device %d bus id: %s\n", rank, device, busid);
}
if ( world_rank == 0 ) printf("AcceleratorCudaInit: ================================================\n");
}
#endif
#ifdef GRID_HIP
hipDeviceProp_t *gpu_props;
hipStream_t copyStream;
void acceleratorInit(void)
{
int nDevices = 1;
@ -156,16 +168,25 @@ void acceleratorInit(void)
#ifdef GRID_DEFAULT_GPU
if ( world_rank == 0 ) {
printf("AcceleratorHipInit: using default device \n");
printf("AcceleratorHipInit: assume user either uses a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding \n");
printf("AcceleratorHipInit: Configure options --enable-summit, --enable-select-gpu=no \n");
printf("AcceleratorHipInit: assume user or srun sets ROCR_VISIBLE_DEVICES and numa binding \n");
printf("AcceleratorHipInit: Configure options --enable-setdevice=no \n");
}
int device = 0;
#else
if ( world_rank == 0 ) {
printf("AcceleratorHipInit: rank %d setting device to node rank %d\n",world_rank,rank);
printf("AcceleratorHipInit: Configure options --enable-select-gpu=yes \n");
printf("AcceleratorHipInit: Configure options --enable-setdevice=yes \n");
}
hipSetDevice(rank);
int device = rank;
#endif
hipSetDevice(device);
hipStreamCreate(&copyStream);
const int len=64;
char busid[len];
if( rank == world_rank ) {
hipDeviceGetPCIBusId(busid, len, device);
printf("local rank %d device %d bus id: %s\n", rank, device, busid);
}
if ( world_rank == 0 ) printf("AcceleratorHipInit: ================================================\n");
}
#endif

71
Grid/threads/Accelerator.h
View File

@ -95,6 +95,7 @@ void acceleratorInit(void);
//////////////////////////////////////////////
#ifdef GRID_CUDA
#include <cuda.h>
#ifdef __CUDA_ARCH__
@ -115,6 +116,14 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
#endif
} // CUDA specific
inline void cuda_mem(void)
{
size_t free_t,total_t,used_t;
cudaMemGetInfo(&free_t,&total_t);
used_t=total_t-free_t;
std::cout << " MemoryManager : GPU used "<<used_t<<" free "<<free_t<< " total "<<total_t<<std::endl;
}
#define accelerator_for2dNB( iter1, num1, iter2, num2, nsimd, ... ) \
{ \
int nt=acceleratorThreads(); \
@ -254,6 +263,7 @@ inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes
cudaMemcpyAsync(to,from,bytes, cudaMemcpyDeviceToDevice,copyStream);
}
inline void acceleratorCopySynchronise(void) { cudaStreamSynchronize(copyStream); };
inline int acceleratorIsCommunicable(void *ptr)
{
// int uvm=0;
@ -330,7 +340,7 @@ inline void acceleratorFreeDevice(void *ptr){free(ptr,*theGridAccelerator);};
inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) {
theGridAccelerator->memcpy(to,from,bytes);
}
inline void acceleratorCopySynchronise(void) { theGridAccelerator->wait(); }
inline void acceleratorCopySynchronise(void) { theGridAccelerator->wait(); std::cout<<"acceleratorCopySynchronise() wait "<<std::endl; }
inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes) { theGridAccelerator->memcpy(to,from,bytes); theGridAccelerator->wait();}
inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ theGridAccelerator->memcpy(to,from,bytes); theGridAccelerator->wait();}
inline void acceleratorMemSet(void *base,int value,size_t bytes) { theGridAccelerator->memset(base,value,bytes); theGridAccelerator->wait();}
@ -361,10 +371,11 @@ NAMESPACE_BEGIN(Grid);
#define accelerator __host__ __device__
#define accelerator_inline __host__ __device__ inline
extern hipStream_t copyStream;
/*These routines define mapping from thread grid to loop & vector lane indexing */
accelerator_inline int acceleratorSIMTlane(int Nsimd) {
#ifdef GRID_SIMT
return hipThreadIdx_z;
return hipThreadIdx_x;
#else
return 0;
#endif
@ -378,19 +389,41 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
{ __VA_ARGS__;} \
}; \
int nt=acceleratorThreads(); \
dim3 hip_threads(nt,1,nsimd); \
dim3 hip_blocks ((num1+nt-1)/nt,num2,1); \
hipLaunchKernelGGL(LambdaApply,hip_blocks,hip_threads, \
0,0, \
num1,num2,nsimd,lambda); \
dim3 hip_threads(nsimd, nt, 1); \
dim3 hip_blocks ((num1+nt-1)/nt,num2,1); \
if(hip_threads.x * hip_threads.y * hip_threads.z <= 64){ \
hipLaunchKernelGGL(LambdaApply64,hip_blocks,hip_threads, \
0,0, \
num1,num2,nsimd, lambda); \
} else { \
hipLaunchKernelGGL(LambdaApply,hip_blocks,hip_threads, \
0,0, \
num1,num2,nsimd, lambda); \
} \
}
template<typename lambda> __global__
__launch_bounds__(64,1)
void LambdaApply64(uint64_t numx, uint64_t numy, uint64_t numz, lambda Lambda)
{
// Following the same scheme as CUDA for now
uint64_t x = threadIdx.y + blockDim.y*blockIdx.x;
uint64_t y = threadIdx.z + blockDim.z*blockIdx.y;
uint64_t z = threadIdx.x;
if ( (x < numx) && (y<numy) && (z<numz) ) {
Lambda(x,y,z);
}
}
template<typename lambda> __global__
__launch_bounds__(1024,1)
void LambdaApply(uint64_t numx, uint64_t numy, uint64_t numz, lambda Lambda)
{
uint64_t x = hipThreadIdx_x + hipBlockDim_x*hipBlockIdx_x;
uint64_t y = hipThreadIdx_y + hipBlockDim_y*hipBlockIdx_y;
uint64_t z = hipThreadIdx_z ;//+ hipBlockDim_z*hipBlockIdx_z;
// Following the same scheme as CUDA for now
uint64_t x = threadIdx.y + blockDim.y*blockIdx.x;
uint64_t y = threadIdx.z + blockDim.z*blockIdx.y;
uint64_t z = threadIdx.x;
if ( (x < numx) && (y<numy) && (z<numz) ) {
Lambda(x,y,z);
}
@ -435,10 +468,16 @@ inline void acceleratorFreeShared(void *ptr){ hipFree(ptr);};
inline void acceleratorFreeDevice(void *ptr){ hipFree(ptr);};
inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes) { hipMemcpy(to,from,bytes, hipMemcpyHostToDevice);}
inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ hipMemcpy(to,from,bytes, hipMemcpyDeviceToHost);}
inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) { hipMemcpy(to,from,bytes, hipMemcpyDeviceToDevice);}
inline void acceleratorCopySynchronise(void) { }
//inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) { hipMemcpy(to,from,bytes, hipMemcpyDeviceToDevice);}
//inline void acceleratorCopySynchronise(void) { }
inline void acceleratorMemSet(void *base,int value,size_t bytes) { hipMemset(base,value,bytes);}
inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) // Asynch
{
hipMemcpyAsync(to,from,bytes, hipMemcpyDeviceToDevice,copyStream);
}
inline void acceleratorCopySynchronise(void) { hipStreamSynchronize(copyStream); };
#endif
//////////////////////////////////////////////
@ -509,18 +548,12 @@ inline void acceleratorFreeCpu (void *ptr){free(ptr);};
///////////////////////////////////////////////////
// Synchronise across local threads for divergence resynch
///////////////////////////////////////////////////
accelerator_inline void acceleratorSynchronise(void)
accelerator_inline void acceleratorSynchronise(void) // Only Nvidia needs
{
#ifdef GRID_SIMT
#ifdef GRID_CUDA
__syncwarp();
#endif
#ifdef GRID_SYCL
//cl::sycl::detail::workGroupBarrier();
#endif
#ifdef GRID_HIP
__syncthreads();
#endif
#endif
return;
}

2
Grid/util/Coordinate.h
View File

@ -88,7 +88,7 @@ public:
// Coordinate class, maxdims = 8 for now.
////////////////////////////////////////////////////////////////
#define GRID_MAX_LATTICE_DIMENSION (8)
#define GRID_MAX_SIMD (16)
#define GRID_MAX_SIMD (32)
static constexpr int MaxDims = GRID_MAX_LATTICE_DIMENSION;

8
Grid/util/Init.cc
View File

@ -167,6 +167,13 @@ void GridCmdOptionInt(std::string &str,int & val)
return;
}
void GridCmdOptionFloat(std::string &str,float & val)
{
std::stringstream ss(str);
ss>>val;
return;
}
void GridParseLayout(char **argv,int argc,
Coordinate &latt_c,
@ -527,6 +534,7 @@ void Grid_init(int *argc,char ***argv)
void Grid_finalize(void)
{
#if defined (GRID_COMMS_MPI) || defined (GRID_COMMS_MPI3) || defined (GRID_COMMS_MPIT)
MPI_Barrier(MPI_COMM_WORLD);
MPI_Finalize();
Grid_unquiesce_nodes();
#endif

1
Grid/util/Init.h
View File

@ -57,6 +57,7 @@ void GridCmdOptionCSL(std::string str,std::vector<std::string> & vec);
template<class VectorInt>
void GridCmdOptionIntVector(const std::string &str,VectorInt & vec);
void GridCmdOptionInt(std::string &str,int & val);
void GridCmdOptionFloat(std::string &str,float & val);
void GridParseLayout(char **argv,int argc,

6
benchmarks/Benchmark_IO.cc
View File

@ -137,7 +137,7 @@ int main (int argc, char ** argv)
Eigen::MatrixXd mean(nVol, 4), stdDev(nVol, 4), rob(nVol, 4);
Eigen::VectorXd avMean(4), avStdDev(4), avRob(4);
double n = BENCH_IO_NPASS;
// double n = BENCH_IO_NPASS;
stats(mean, stdDev, perf);
stats(avMean, avStdDev, avPerf);
@ -164,7 +164,7 @@ int main (int argc, char ** argv)
mean(volInd(l), gWrite), stdDev(volInd(l), gWrite));
}
MSG << std::endl;
MSG << "Robustness of individual results, in \%. (rob = 100\% - std dev / mean)" << std::endl;
MSG << "Robustness of individual results, in %. (rob = 100% - std dev / mean)" << std::endl;
MSG << std::endl;
grid_printf("%4s %12s %12s %12s %12s\n",
"L", "std read", "std write", "Grid read", "Grid write");
@ -185,7 +185,7 @@ int main (int argc, char ** argv)
avMean(sRead), avStdDev(sRead), avMean(sWrite), avStdDev(sWrite),
avMean(gRead), avStdDev(gRead), avMean(gWrite), avStdDev(gWrite));
MSG << std::endl;
MSG << "Robustness of volume-averaged results, in \%. (rob = 100\% - std dev / mean)" << std::endl;
MSG << "Robustness of volume-averaged results, in %. (rob = 100% - std dev / mean)" << std::endl;
MSG << std::endl;
grid_printf("%12s %12s %12s %12s\n",
"std read", "std write", "Grid read", "Grid write");

4
benchmarks/Benchmark_ITT.cc
View File

@ -142,7 +142,7 @@ public:
// bzero((void *)rbuf[d],lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
}
int ncomm;
// int ncomm;
double dbytes;
for(int dir=0;dir<8;dir++) {
@ -290,7 +290,7 @@ public:
LatticeSU4 z(&Grid); z=Zero();
LatticeSU4 x(&Grid); x=Zero();
LatticeSU4 y(&Grid); y=Zero();
double a=2.0;
// double a=2.0;
uint64_t Nloop=NLOOP;

2
benchmarks/Benchmark_comms_host_device.cc
View File

@ -72,7 +72,7 @@ int main (int argc, char ** argv)
std::cout << GridLogMessage << "Number of iterations to average: "<< Nloop << std::endl;
std::vector<double> t_time(Nloop);
time_statistics timestat;
// time_statistics timestat;
std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
std::cout<<GridLogMessage << "= Benchmarking sequential halo exchange from host memory "<<std::endl;

63
benchmarks/Benchmark_dwf_fp32.cc
View File

@ -126,19 +126,10 @@ int main (int argc, char ** argv)
// Naive wilson implementation
////////////////////////////////////
// replicate across fifth dimension
LatticeGaugeFieldF Umu5d(FGrid);
std::vector<LatticeColourMatrixF> U(4,FGrid);
{
autoView( Umu5d_v, Umu5d, CpuWrite);
autoView( Umu_v , Umu , CpuRead);
for(int ss=0;ss<Umu.Grid()->oSites();ss++){
for(int s=0;s<Ls;s++){
Umu5d_v[Ls*ss+s] = Umu_v[ss];
}
}
}
// LatticeGaugeFieldF Umu5d(FGrid);
std::vector<LatticeColourMatrixF> U(4,UGrid);
for(int mu=0;mu<Nd;mu++){
U[mu] = PeekIndex<LorentzIndex>(Umu5d,mu);
U[mu] = PeekIndex<LorentzIndex>(Umu,mu);
}
std::cout << GridLogMessage << "Setting up Cshift based reference " << std::endl;
@ -147,10 +138,28 @@ int main (int argc, char ** argv)
ref = Zero();
for(int mu=0;mu<Nd;mu++){
tmp = U[mu]*Cshift(src,mu+1,1);
tmp = Cshift(src,mu+1,1);
{
autoView( tmp_v , tmp , CpuWrite);
autoView( U_v , U[mu] , CpuRead);
for(int ss=0;ss<U[mu].Grid()->oSites();ss++){
for(int s=0;s<Ls;s++){
tmp_v[Ls*ss+s] = U_v[ss]*tmp_v[Ls*ss+s];
}
}
}
ref=ref + tmp - Gamma(Gmu[mu])*tmp;
tmp =adj(U[mu])*src;
{
autoView( tmp_v , tmp , CpuWrite);
autoView( U_v , U[mu] , CpuRead);
autoView( src_v, src , CpuRead);
for(int ss=0;ss<U[mu].Grid()->oSites();ss++){
for(int s=0;s<Ls;s++){
tmp_v[Ls*ss+s] = adj(U_v[ss])*src_v[Ls*ss+s];
}
}
}
tmp =Cshift(tmp,mu+1,-1);
ref=ref + tmp + Gamma(Gmu[mu])*tmp;
}
@ -182,7 +191,7 @@ int main (int argc, char ** argv)
std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
DomainWallFermionF Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
int ncall =3000;
int ncall =300;
if (1) {
FGrid->Barrier();
@ -242,16 +251,30 @@ int main (int argc, char ** argv)
for(int mu=0;mu<Nd;mu++){
// ref = src - Gamma(Gamma::Algebra::GammaX)* src ; // 1+gamma_x
tmp = U[mu]*Cshift(src,mu+1,1);
tmp = Cshift(src,mu+1,1);
{
autoView( ref_v, ref, CpuWrite);
autoView( tmp_v, tmp, CpuRead);
for(int i=0;i<ref_v.size();i++){
ref_v[i]+= tmp_v[i] + Gamma(Gmu[mu])*tmp_v[i]; ;
autoView( U_v , U[mu] , CpuRead);
for(int ss=0;ss<U[mu].Grid()->oSites();ss++){
for(int s=0;s<Ls;s++){
int i=s+Ls*ss;
ref_v[i]+= U_v[ss]*(tmp_v[i] + Gamma(Gmu[mu])*tmp_v[i]); ;
}
}
}
tmp =adj(U[mu])*src;
{
autoView( tmp_v , tmp , CpuWrite);
autoView( U_v , U[mu] , CpuRead);
autoView( src_v, src , CpuRead);
for(int ss=0;ss<U[mu].Grid()->oSites();ss++){
for(int s=0;s<Ls;s++){
tmp_v[Ls*ss+s] = adj(U_v[ss])*src_v[Ls*ss+s];
}
}
}
// tmp =adj(U[mu])*src;
tmp =Cshift(tmp,mu+1,-1);
{
autoView( ref_v, ref, CpuWrite);

6
benchmarks/Benchmark_memory_bandwidth.cc
View File

@ -184,8 +184,10 @@ int main (int argc, char ** argv)
double bytes=1.0*vol*Nvec*sizeof(Real);
double flops=vol*Nvec*2;// mul,add
std::cout<<GridLogMessage<<std::setprecision(3) << lat<<"\t\t"<<bytes<<" \t\t"<<bytes/time<<"\t\t"<<flops/time<< "\t\t"<<(stop-start)/1000./1000.<< "\t\t " <<std::endl;
std::cout<<GridLogMessage<<std::setprecision(3) << lat<<"\t\t"
<<bytes<<" \t\t"<<bytes/time<<"\t\t"<<flops/time<< "\t\t"
<<(stop-start)/1000./1000.<< "\t\t " <<std::endl;
assert(nn==nn);
}
Grid_finalize();

2
examples/Example_Laplacian_solver.cc
View File

@ -4,7 +4,7 @@ using namespace Grid;
template<class Field>
void SimpleConjugateGradient(LinearOperatorBase<Field> &HPDop,const Field &b, Field &x)
{
RealD cp, c, alpha, d, beta, ssq, qq;
RealD cp, c, alpha, d, beta, ssq;
RealD Tolerance=1.0e-10;
int MaxIterations=10000;

539
examples/Example_wall_wall_3pt.cc Normal file
View File

@ -0,0 +1,539 @@
/*
* Warning: This code illustrative only: not well tested, and not meant for production use
* without regression / tests being applied
*/
#include <Grid/Grid.h>
using namespace std;
using namespace Grid;
typedef SpinColourMatrix Propagator;
typedef SpinColourVector Fermion;
typedef PeriodicGimplR GimplR;
template<class Gimpl,class Field> class CovariantLaplacianCshift : public SparseMatrixBase<Field>
{
public:
INHERIT_GIMPL_TYPES(Gimpl);
GridBase *grid;
GaugeField U;
CovariantLaplacianCshift(GaugeField &_U) :
grid(_U.Grid()),
U(_U) { };
virtual GridBase *Grid(void) { return grid; };
virtual void M (const Field &in, Field &out)
{
out=Zero();
for(int mu=0;mu<Nd-1;mu++) {
GaugeLinkField Umu = PeekIndex<LorentzIndex>(U, mu); // NB: Inefficent
out = out - Gimpl::CovShiftForward(Umu,mu,in);
out = out - Gimpl::CovShiftBackward(Umu,mu,in);
out = out + 2.0*in;
}
};
virtual void Mdag (const Field &in, Field &out) { M(in,out);}; // Laplacian is hermitian
virtual void Mdiag (const Field &in, Field &out) {assert(0);}; // Unimplemented need only for multigrid
virtual void Mdir (const Field &in, Field &out,int dir, int disp){assert(0);}; // Unimplemented need only for multigrid
virtual void MdirAll (const Field &in, std::vector<Field> &out) {assert(0);}; // Unimplemented need only for multigrid
};
void MakePhase(Coordinate mom,LatticeComplex &phase)
{
GridBase *grid = phase.Grid();
auto latt_size = grid->GlobalDimensions();
ComplexD ci(0.0,1.0);
phase=Zero();
LatticeComplex coor(phase.Grid());
for(int mu=0;mu<Nd;mu++){
RealD TwoPiL = M_PI * 2.0/ latt_size[mu];
LatticeCoordinate(coor,mu);
phase = phase + (TwoPiL * mom[mu]) * coor;
}
phase = exp(phase*ci);
}
void LinkSmear(int nstep, RealD rho,LatticeGaugeField &Uin,LatticeGaugeField &Usmr)
{
Smear_Stout<GimplR> Stout(rho);
LatticeGaugeField Utmp(Uin.Grid());
Utmp = Uin;
for(int i=0;i<nstep;i++){
Stout.smear(Usmr,Utmp);
Utmp = Usmr;
}
}
void PointSource(Coordinate &coor,LatticePropagator &source)
{
// Coordinate coor({0,0,0,0});
source=Zero();
SpinColourMatrix kronecker; kronecker=1.0;
pokeSite(kronecker,source,coor);
}
void GFWallSource(int tslice,LatticePropagator &source)
{
GridBase *grid = source.Grid();
LatticeComplex one(grid); one = ComplexD(1.0,0.0);
LatticeComplex zz(grid); zz=Zero();
LatticeInteger t(grid);
LatticeCoordinate(t,Tdir);
one = where(t==Integer(tslice), one, zz);
source = 1.0;
source = source * one;
}
void Z2WallSource(GridParallelRNG &RNG,int tslice,LatticePropagator &source)
{
GridBase *grid = source.Grid();
LatticeComplex noise(grid);
LatticeComplex zz(grid); zz=Zero();
LatticeInteger t(grid);
RealD nrm=1.0/sqrt(2);
bernoulli(RNG, noise); // 0,1 50:50
noise = (2.*noise - Complex(1,1))*nrm;
LatticeCoordinate(t,Tdir);
noise = where(t==Integer(tslice), noise, zz);
source = 1.0;
source = source*noise;
std::cout << " Z2 wall " << norm2(source) << std::endl;
}
void GaugeFix(LatticeGaugeField &U,LatticeGaugeField &Ufix)
{
Real alpha=0.05;
Real plaq=WilsonLoops<GimplR>::avgPlaquette(U);
std::cout << " Initial plaquette "<<plaq << std::endl;
LatticeColourMatrix xform(U.Grid());
Ufix = U;
int orthog=Nd-1;
FourierAcceleratedGaugeFixer<GimplR>::SteepestDescentGaugeFix(Ufix,xform,alpha,100000,1.0e-14, 1.0e-14,true,orthog);
plaq=WilsonLoops<GimplR>::avgPlaquette(Ufix);
std::cout << " Final plaquette "<<plaq << std::endl;
}
template<class Field>
void GaussianSmear(LatticeGaugeField &U,Field &unsmeared,Field &smeared)
{
typedef CovariantLaplacianCshift <GimplR,Field> Laplacian_t;
Laplacian_t Laplacian(U);
Integer Iterations = 40;
Real width = 2.0;
Real coeff = (width*width) / Real(4*Iterations);
Field tmp(U.Grid());
smeared=unsmeared;
// chi = (1-p^2/2N)^N kronecker
for(int n = 0; n < Iterations; ++n) {
Laplacian.M(smeared,tmp);
smeared = smeared - coeff*tmp;
std::cout << " smear iter " << n<<" " <<norm2(smeared)<<std::endl;
}
}
void GaussianSource(Coordinate &site,LatticeGaugeField &U,LatticePropagator &source)
{
LatticePropagator tmp(source.Grid());
PointSource(site,source);
std::cout << " GaussianSource Kronecker "<< norm2(source)<<std::endl;
tmp = source;
GaussianSmear(U,tmp,source);
std::cout << " GaussianSource Smeared "<< norm2(source)<<std::endl;
}
void GaussianWallSource(GridParallelRNG &RNG,int tslice,LatticeGaugeField &U,LatticePropagator &source)
{
Z2WallSource(RNG,tslice,source);
auto tmp = source;
GaussianSmear(U,tmp,source);
}
void SequentialSource(int tslice,Coordinate &mom,LatticePropagator &spectator,LatticePropagator &source)
{
assert(mom.size()==Nd);
assert(mom[Tdir] == 0);
GridBase * grid = spectator.Grid();
LatticeInteger ts(grid);
LatticeCoordinate(ts,Tdir);
source = Zero();
source = where(ts==Integer(tslice),spectator,source); // Stick in a slice of the spectator, zero everywhere else
LatticeComplex phase(grid);
MakePhase(mom,phase);
source = source *phase;
}
template<class Action>
void Solve(Action &D,LatticePropagator &source,LatticePropagator &propagator)
{
GridBase *UGrid = D.GaugeGrid();
GridBase *FGrid = D.FermionGrid();
LatticeFermion src4 (UGrid);
LatticeFermion src5 (FGrid);
LatticeFermion result5(FGrid);
LatticeFermion result4(UGrid);
ConjugateGradient<LatticeFermion> CG(1.0e-12,100000);
SchurRedBlackDiagTwoSolve<LatticeFermion> schur(CG);
ZeroGuesser<LatticeFermion> ZG; // Could be a DeflatedGuesser if have eigenvectors
for(int s=0;s<Nd;s++){
for(int c=0;c<Nc;c++){
PropToFerm<Action>(src4,source,s,c);
D.ImportPhysicalFermionSource(src4,src5);
result5=Zero();
schur(D,src5,result5,ZG);
std::cout<<GridLogMessage
<<"spin "<<s<<" color "<<c
<<" norm2(src5d) " <<norm2(src5)
<<" norm2(result5d) "<<norm2(result5)<<std::endl;
D.ExportPhysicalFermionSolution(result5,result4);
FermToProp<Action>(propagator,result4,s,c);
}
}
}
class MesonFile: Serializable {
public:
GRID_SERIALIZABLE_CLASS_MEMBERS(MesonFile, std::vector<std::vector<Complex> >, data);
};
void MesonTrace(std::string file,LatticePropagator &q1,LatticePropagator &q2,LatticeComplex &phase)
{
const int nchannel=4;
Gamma::Algebra Gammas[nchannel][2] = {
{Gamma::Algebra::Gamma5 ,Gamma::Algebra::Gamma5},
{Gamma::Algebra::GammaTGamma5,Gamma::Algebra::GammaTGamma5},
{Gamma::Algebra::GammaTGamma5,Gamma::Algebra::Gamma5},
{Gamma::Algebra::Gamma5 ,Gamma::Algebra::GammaTGamma5}
};
Gamma G5(Gamma::Algebra::Gamma5);
LatticeComplex meson_CF(q1.Grid());
MesonFile MF;
for(int ch=0;ch<nchannel;ch++){
Gamma Gsrc(Gammas[ch][0]);
Gamma Gsnk(Gammas[ch][1]);
meson_CF = trace(G5*adj(q1)*G5*Gsnk*q2*adj(Gsrc));
std::vector<TComplex> meson_T;
sliceSum(meson_CF,meson_T, Tdir);
int nt=meson_T.size();
std::vector<Complex> corr(nt);
for(int t=0;t<nt;t++){
corr[t] = TensorRemove(meson_T[t]); // Yes this is ugly, not figured a work around
std::cout << " channel "<<ch<<" t "<<t<<" " <<corr[t]<<std::endl;
}
MF.data.push_back(corr);
}
{
XmlWriter WR(file);
write(WR,"MesonFile",MF);
}
}
void Meson3pt(std::string file,LatticePropagator &q1,LatticePropagator &q2,LatticeComplex &phase)
{
const int nchannel=4;
Gamma::Algebra Gammas[nchannel][2] = {
{Gamma::Algebra::Gamma5 ,Gamma::Algebra::GammaX},
{Gamma::Algebra::Gamma5 ,Gamma::Algebra::GammaY},
{Gamma::Algebra::Gamma5 ,Gamma::Algebra::GammaZ},
{Gamma::Algebra::Gamma5 ,Gamma::Algebra::GammaT}
};
Gamma G5(Gamma::Algebra::Gamma5);
LatticeComplex meson_CF(q1.Grid());
MesonFile MF;
for(int ch=0;ch<nchannel;ch++){
Gamma Gsrc(Gammas[ch][0]);
Gamma Gsnk(Gammas[ch][1]);
meson_CF = trace(G5*adj(q1)*G5*Gsnk*q2*adj(Gsrc));
std::vector<TComplex> meson_T;
sliceSum(meson_CF,meson_T, Tdir);
int nt=meson_T.size();
std::vector<Complex> corr(nt);
for(int t=0;t<nt;t++){
corr[t] = TensorRemove(meson_T[t]); // Yes this is ugly, not figured a work around
std::cout << " channel "<<ch<<" t "<<t<<" " <<corr[t]<<std::endl;
}
MF.data.push_back(corr);
}
{
XmlWriter WR(file);
write(WR,"MesonFile",MF);
}
}
void WallSinkMesonTrace(std::string file,std::vector<Propagator> &q1,std::vector<Propagator> &q2)
{
const int nchannel=4;
Gamma::Algebra Gammas[nchannel][2] = {
{Gamma::Algebra::Gamma5 ,Gamma::Algebra::Gamma5},
{Gamma::Algebra::GammaTGamma5,Gamma::Algebra::GammaTGamma5},
{Gamma::Algebra::GammaTGamma5,Gamma::Algebra::Gamma5},
{Gamma::Algebra::Gamma5 ,Gamma::Algebra::GammaTGamma5}
};
Gamma G5(Gamma::Algebra::Gamma5);
int nt=q1.size();
std::vector<Complex> meson_CF(nt);
MesonFile MF;
for(int ch=0;ch<nchannel;ch++){
Gamma Gsrc(Gammas[ch][0]);
Gamma Gsnk(Gammas[ch][1]);
std::vector<Complex> corr(nt);
for(int t=0;t<nt;t++){
meson_CF[t] = trace(G5*adj(q1[t])*G5*Gsnk*q2[t]*adj(Gsrc));
corr[t] = TensorRemove(meson_CF[t]); // Yes this is ugly, not figured a work around
std::cout << " channel "<<ch<<" t "<<t<<" " <<corr[t]<<std::endl;
}
MF.data.push_back(corr);
}
{
XmlWriter WR(file);
write(WR,"MesonFile",MF);
}
}
int make_idx(int p, int m,int nmom)
{
if (m==0) return p;
assert(p==0);
return nmom + m - 1;
}
int main (int argc, char ** argv)
{
Grid_init(&argc,&argv);
// Double precision grids
auto latt = GridDefaultLatt();
GridCartesian * UGrid = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(),
GridDefaultSimd(Nd,vComplex::Nsimd()),
GridDefaultMpi());
GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
LatticeGaugeField Umu(UGrid);
LatticeGaugeField Utmp(UGrid);
LatticeGaugeField Usmr(UGrid);
std::string config;
if( argc > 1 && argv[1][0] != '-' )
{
std::cout<<GridLogMessage <<"Loading configuration from "<<argv[1]<<std::endl;
FieldMetaData header;
NerscIO::readConfiguration(Umu, header, argv[1]);
config=argv[1];
}
else
{
std::cout<<GridLogMessage <<"Using hot configuration"<<std::endl;
SU<Nc>::ColdConfiguration(Umu);
config="ColdConfig";
}
// GaugeFix(Umu,Utmp);
// Umu=Utmp;
int nsmr=3;
RealD rho=0.1;
LinkSmear(nsmr,rho,Umu,Usmr);
std::vector<int> smeared_link({ 0,0,1} );
std::vector<RealD> masses({ 0.004,0.02477,0.447} ); // u/d, s, c ??
std::vector<RealD> M5s ({ 1.8,1.8,1.0} );
std::vector<RealD> bs ({ 1.0,1.0,1.5} ); // DDM
std::vector<RealD> cs ({ 0.0,0.0,0.5} ); // DDM
std::vector<int> Ls_s ({ 16,16,12} );
std::vector<GridCartesian *> FGrids;
std::vector<GridRedBlackCartesian *> FrbGrids;
std::vector<Coordinate> momenta;
momenta.push_back(Coordinate({0,0,0,0}));
momenta.push_back(Coordinate({1,0,0,0}));
momenta.push_back(Coordinate({2,0,0,0}));
int nmass = masses.size();
int nmom = momenta.size();
std::vector<MobiusFermionR *> FermActs;
std::cout<<GridLogMessage <<"======================"<<std::endl;
std::cout<<GridLogMessage <<"MobiusFermion action as Scaled Shamir kernel"<<std::endl;
std::cout<<GridLogMessage <<"======================"<<std::endl;
std::vector<Complex> boundary = {1,1,1,-1};
typedef MobiusFermionR FermionAction;
FermionAction::ImplParams Params(boundary);
for(int m=0;m<masses.size();m++) {
RealD mass = masses[m];
RealD M5 = M5s[m];
RealD b = bs[m];
RealD c = cs[m];
int Ls = Ls_s[m];
if ( smeared_link[m] ) Utmp = Usmr;
else Utmp = Umu;
FGrids.push_back(SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid));
FrbGrids.push_back(SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid));
FermActs.push_back(new MobiusFermionR(Utmp,*FGrids[m],*FrbGrids[m],*UGrid,*UrbGrid,mass,M5,b,c,Params));
}
LatticePropagator z2wall_source(UGrid);
LatticePropagator gfwall_source(UGrid);
LatticePropagator phased_prop(UGrid);
int tslice = 0;
int tseq=(tslice+16)%latt[Nd-1];
//////////////////////////////////////////////////////////////////////
// RNG seeded for Z2 wall
//////////////////////////////////////////////////////////////////////
// You can manage seeds however you like.
// Recommend SeedUniqueString.
//////////////////////////////////////////////////////////////////////
GridParallelRNG RNG4(UGrid); RNG4.SeedUniqueString("Study2-Source_Z2_p_0_0_0_t_0-880");
Z2WallSource (RNG4,tslice,z2wall_source);
GFWallSource (tslice,gfwall_source);
std::vector<LatticeComplex> phase(nmom,UGrid);
for(int m=0;m<nmom;m++){
MakePhase(momenta[m],phase[m]);
}
std::vector<LatticePropagator> Z2Props (nmom+nmass-1,UGrid);
std::vector<LatticePropagator> GFProps (nmom+nmass-1,UGrid);
for(int p=0;p<nmom;p++) {
int m=0;
int idx = make_idx(p,m,nmom);
phased_prop = z2wall_source * phase[p];
Solve(*FermActs[m],phased_prop ,Z2Props[idx]);
phased_prop = gfwall_source * phase[p];
Solve(*FermActs[m],phased_prop ,GFProps[idx]);
}
for(int m=1;m<nmass;m++) {
int p=0;
int idx = make_idx(p,m,nmom);
phased_prop = z2wall_source;
Solve(*FermActs[m],phased_prop ,Z2Props[idx]);
phased_prop = gfwall_source;
Solve(*FermActs[m],phased_prop ,GFProps[idx]);
}
std::vector<std::vector<Propagator> > wsnk_z2Props(nmom+nmass-1);
std::vector<std::vector<Propagator> > wsnk_gfProps(nmom+nmass-1);
// Non-zero kaon and point and D two point
// WW stick momentum on m1 (lighter)
// zero momentum on m2
for(int m1=0;m1<nmass;m1++) {
for(int m2=m1;m2<nmass;m2++) {
int pmax = (m1==0)? nmom:1;
for(int p=0;p<pmax;p++){
std::stringstream ssg,ssz;
std::stringstream wssg,wssz;
int idx1 = make_idx(p,m1,nmom);
int idx2 = make_idx(0,m2,nmom);
/// Point sinks
ssg<<config<<"_p"<<p<< "_m" << m1 << "_m"<< m2 << "_p_gf_meson.xml";
ssz<<config<<"_p"<<p<< "_m" << m1 << "_m"<< m2 << "_p_z2_meson.xml";
MesonTrace(ssz.str(),Z2Props[idx1],Z2Props[idx2],phase[p]); // Q1 is conjugated
MesonTrace(ssg.str(),GFProps[idx1],GFProps[idx2],phase[p]);
/// Wall sinks
wssg<<config<<"_p"<<p<< "_m" << m1 << "_m"<< m2 << "_w_gf_meson.xml";
wssz<<config<<"_p"<<p<< "_m" << m1 << "_m"<< m2 << "_w_z2_meson.xml";
phased_prop = GFProps[m2] * phase[p];
sliceSum(phased_prop,wsnk_gfProps[m1],Tdir);
sliceSum(GFProps[m1],wsnk_gfProps[m2],Tdir);
WallSinkMesonTrace(wssg.str(),wsnk_gfProps[m1],wsnk_gfProps[m2]);
phased_prop = Z2Props[m2] * phase[p];
sliceSum(phased_prop,wsnk_gfProps[m1],Tdir);
sliceSum(Z2Props[m1],wsnk_gfProps[m2],Tdir);
WallSinkMesonTrace(wssz.str(),wsnk_z2Props[m1],wsnk_z2Props[m2]);
}
}}
/////////////////////////////////////
// Sequential solves
/////////////////////////////////////
LatticePropagator seq_wsnk_z2src(UGrid);
LatticePropagator seq_wsnk_gfsrc(UGrid);
LatticePropagator seq_psnk_z2src(UGrid);
LatticePropagator seq_psnk_gfsrc(UGrid);
LatticePropagator source(UGrid);
for(int m=0;m<nmass-1;m++){
int spect_idx = make_idx(0,m,nmom);
int charm=nmass-1;
SequentialSource(tseq,momenta[0],GFProps[spect_idx],source);
Solve(*FermActs[charm],source,seq_psnk_gfsrc);
SequentialSource(tseq,momenta[0],Z2Props[spect_idx],source);
Solve(*FermActs[charm],source,seq_psnk_z2src);
// Todo need wall sequential solve
for(int p=0;p<nmom;p++){
int active_idx = make_idx(p,0,nmom);
std::stringstream seq_3pt_p_z2;
std::stringstream seq_3pt_p_gf;
std::stringstream seq_3pt_w_z2;
std::stringstream seq_3pt_w_gf;
seq_3pt_p_z2 <<config<<"_3pt_p"<<p<< "_m" << m << "_p_z2_meson.xml";
seq_3pt_p_gf <<config<<"_3pt_p"<<p<< "_m" << m << "_p_gf_meson.xml";
seq_3pt_w_z2 <<config<<"_3pt_p"<<p<< "_m" << m << "_w_z2_meson.xml";
seq_3pt_w_gf <<config<<"_3pt_p"<<p<< "_m" << m << "_w_gf_meson.xml";
Meson3pt(seq_3pt_p_gf.str(),GFProps[active_idx],seq_psnk_gfsrc,phase[p]);
Meson3pt(seq_3pt_p_z2.str(),Z2Props[active_idx],seq_psnk_z2src,phase[p]);
}
}
Grid_finalize();
}

95
examples/Example_wall_wall_spectrum.cc
View File

@ -9,6 +9,7 @@ using namespace std;
using namespace Grid;
typedef SpinColourMatrix Propagator;
typedef SpinColourVector Fermion;
typedef PeriodicGimplR GimplR;
template<class Gimpl,class Field> class CovariantLaplacianCshift : public SparseMatrixBase<Field>
{
@ -55,6 +56,16 @@ void MakePhase(Coordinate mom,LatticeComplex &phase)
}
phase = exp(phase*ci);
}
void LinkSmear(int nstep, RealD rho,LatticeGaugeField &Uin,LatticeGaugeField &Usmr)
{
Smear_Stout<GimplR> Stout(rho);
LatticeGaugeField Utmp(Uin.Grid());
Utmp = Uin;
for(int i=0;i<nstep;i++){
Stout.smear(Usmr,Utmp);
Utmp = Usmr;
}
}
void PointSource(Coordinate &coor,LatticePropagator &source)
{
// Coordinate coor({0,0,0,0});
@ -97,23 +108,23 @@ void GaugeFix(LatticeGaugeField &U,LatticeGaugeField &Ufix)
{
Real alpha=0.05;
Real plaq=WilsonLoops<PeriodicGimplR>::avgPlaquette(U);
Real plaq=WilsonLoops<GimplR>::avgPlaquette(U);
std::cout << " Initial plaquette "<<plaq << std::endl;
LatticeColourMatrix xform(U.Grid());
Ufix = U;
int orthog=Nd-1;
FourierAcceleratedGaugeFixer<PeriodicGimplR>::SteepestDescentGaugeFix(Ufix,xform,alpha,10000,1.0e-12, 1.0e-12,true,orthog);
FourierAcceleratedGaugeFixer<GimplR>::SteepestDescentGaugeFix(Ufix,xform,alpha,100000,1.0e-14, 1.0e-14,true,orthog);
plaq=WilsonLoops<PeriodicGimplR>::avgPlaquette(Ufix);
plaq=WilsonLoops<GimplR>::avgPlaquette(Ufix);
std::cout << " Final plaquette "<<plaq << std::endl;
}
template<class Field>
void GaussianSmear(LatticeGaugeField &U,Field &unsmeared,Field &smeared)
{
typedef CovariantLaplacianCshift <PeriodicGimplR,Field> Laplacian_t;
typedef CovariantLaplacianCshift <GimplR,Field> Laplacian_t;
Laplacian_t Laplacian(U);
Integer Iterations = 40;
@ -167,19 +178,21 @@ void Solve(Action &D,LatticePropagator &source,LatticePropagator &propagator)
GridBase *UGrid = D.GaugeGrid();
GridBase *FGrid = D.FermionGrid();
LatticeFermion src4 (UGrid);
LatticeFermion src4 (UGrid); src4 = Zero();
LatticeFermion src5 (FGrid);
LatticeFermion result5(FGrid);
LatticeFermion result4(UGrid);
ConjugateGradient<LatticeFermion> CG(1.0e-8,100000);
SchurRedBlackDiagMooeeSolve<LatticeFermion> schur(CG);
ConjugateGradient<LatticeFermion> CG(1.0e-12,100000);
SchurRedBlackDiagTwoSolve<LatticeFermion> schur(CG);
ZeroGuesser<LatticeFermion> ZG; // Could be a DeflatedGuesser if have eigenvectors
std::cout<<GridLogMessage<< " source4 "<<norm2(source)<<std::endl;
for(int s=0;s<Nd;s++){
for(int c=0;c<Nc;c++){
PropToFerm<Action>(src4,source,s,c);
std::cout<<GridLogMessage<< s<<c<<" src4 "<<norm2(src4)<<std::endl;
D.ImportPhysicalFermionSource(src4,src5);
std::cout<<GridLogMessage<< s<<c<<" src5 "<<norm2(src5)<<std::endl;
result5=Zero();
schur(D,src5,result5,ZG);
@ -287,15 +300,10 @@ int main (int argc, char ** argv)
GridDefaultMpi());
GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
//////////////////////////////////////////////////////////////////////
// You can manage seeds however you like.
// Recommend SeedUniqueString.
//////////////////////////////////////////////////////////////////////
std::vector<int> seeds4({1,2,3,4});
GridParallelRNG RNG4(UGrid); RNG4.SeedFixedIntegers(seeds4);
LatticeGaugeField Umu(UGrid);
LatticeGaugeField Ufixed(UGrid);
LatticeGaugeField Utmp(UGrid);
LatticeGaugeField Usmr(UGrid);
std::string config;
if( argc > 1 && argv[1][0] != '-' )
{
@ -308,13 +316,20 @@ int main (int argc, char ** argv)
{
std::cout<<GridLogMessage <<"Using hot configuration"<<std::endl;
SU<Nc>::ColdConfiguration(Umu);
// SU<Nc>::HotConfiguration(RNG4,Umu);
config="HotConfig";
config="ColdConfig";
}
GaugeFix(Umu,Ufixed);
Umu=Ufixed;
// GaugeFix(Umu,Utmp);
// Umu=Utmp;
int nsmr=3;
RealD rho=0.1;
RealD plaq_gf =WilsonLoops<GimplR>::avgPlaquette(Umu);
LinkSmear(nsmr,rho,Umu,Usmr);
RealD plaq_smr=WilsonLoops<GimplR>::avgPlaquette(Usmr);
std::cout << GridLogMessage << " GF Plaquette " <<plaq_gf<<std::endl;
std::cout << GridLogMessage << " SM Plaquette " <<plaq_smr<<std::endl;
std::vector<int> smeared_link({ 0,0,1} );
std::vector<RealD> masses({ 0.004,0.02477,0.447} ); // u/d, s, c ??
std::vector<RealD> M5s ({ 1.8,1.8,1.0} );
std::vector<RealD> bs ({ 1.0,1.0,1.5} ); // DDM
@ -330,6 +345,9 @@ int main (int argc, char ** argv)
std::cout<<GridLogMessage <<"======================"<<std::endl;
std::cout<<GridLogMessage <<"MobiusFermion action as Scaled Shamir kernel"<<std::endl;
std::cout<<GridLogMessage <<"======================"<<std::endl;
std::vector<Complex> boundary = {1,1,1,-1};
typedef MobiusFermionR FermionAction;
FermionAction::ImplParams Params(boundary);
for(int m=0;m<masses.size();m++) {
@ -339,30 +357,40 @@ int main (int argc, char ** argv)
RealD c = cs[m];
int Ls = Ls_s[m];
if ( smeared_link[m] ) Utmp = Usmr;
else Utmp = Umu;
FGrids.push_back(SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid));
FrbGrids.push_back(SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid));
FermActs.push_back(new MobiusFermionR(Umu,*FGrids[m],*FrbGrids[m],*UGrid,*UrbGrid,mass,M5,b,c));
FermActs.push_back(new MobiusFermionR(Utmp,*FGrids[m],*FrbGrids[m],*UGrid,*UrbGrid,mass,M5,b,c,Params));
}
LatticePropagator point_source(UGrid);
LatticePropagator z2wall_source(UGrid);
LatticePropagator gfwall_source(UGrid);
Coordinate Origin({0,0,0,0});
PointSource (Origin,point_source);
Z2WallSource (RNG4,0,z2wall_source);
GFWallSource (0,gfwall_source);
std::vector<LatticePropagator> PointProps(nmass,UGrid);
std::vector<LatticePropagator> GaussProps(nmass,UGrid);
int tslice = 0;
//////////////////////////////////////////////////////////////////////
// RNG seeded for Z2 wall
//////////////////////////////////////////////////////////////////////
// You can manage seeds however you like.
// Recommend SeedUniqueString.
//////////////////////////////////////////////////////////////////////
GridParallelRNG RNG4(UGrid); RNG4.SeedUniqueString("Study2-Source_Z2_p_0_0_0_t_0-880");
Z2WallSource (RNG4,tslice,z2wall_source);
GFWallSource (tslice,gfwall_source);
std::vector<LatticePropagator> Z2Props (nmass,UGrid);
std::vector<LatticePropagator> GFProps (nmass,UGrid);
for(int m=0;m<nmass;m++) {
std::cout << GridLogMessage << " Mass " <<m << " z2wall source "<<norm2(z2wall_source)<<std::endl;
Solve(*FermActs[m],z2wall_source ,Z2Props[m]);
std::cout << GridLogMessage << " Mass " <<m << " gfwall source "<<norm2(gfwall_source)<<std::endl;
Solve(*FermActs[m],gfwall_source ,GFProps[m]);
std::cout << GridLogMessage << " Mass " <<m << " z2wall source "<<norm2(z2wall_source)<< " " << norm2(gfwall_source)<<std::endl;
}
@ -383,14 +411,15 @@ int main (int argc, char ** argv)
std::stringstream wssg,wssz;
/// Point sinks
ssg<<config<< "_m" << m1 << "_m"<< m2 << "p_gf_meson.xml";
ssz<<config<< "_m" << m1 << "_m"<< m2 << "p_z2_meson.xml";
ssg<<config<< "_m" << m1 << "_m"<< m2 << "_p_gf_meson.xml";
ssz<<config<< "_m" << m1 << "_m"<< m2 << "_p_z2_meson.xml";
MesonTrace(ssz.str(),Z2Props[m1],Z2Props[m2],phase);
MesonTrace(ssg.str(),GFProps[m1],GFProps[m2],phase);
/// Wall sinks
wssg<<config<< "_m" << m1 << "_m"<< m2 << "w_gf_meson.xml";
wssz<<config<< "_m" << m1 << "_m"<< m2 << "w_z2_meson.xml";
wssg<<config<< "_m" << m1 << "_m"<< m2 << "_w_gf_meson.xml";
wssz<<config<< "_m" << m1 << "_m"<< m2 << "_w_z2_meson.xml";
WallSinkMesonTrace(wssg.str(),wsnk_gfProps[m1],wsnk_gfProps[m2]);
WallSinkMesonTrace(wssz.str(),wsnk_z2Props[m1],wsnk_z2Props[m2]);

12
systems/Crusher/config-command Normal file
View File

@ -0,0 +1,12 @@
../../configure --enable-comms=mpi-auto \
--enable-unified=no \
--enable-shm=nvlink \
--enable-accelerator=hip \
--enable-gen-simd-width=64 \
--enable-simd=GPU \
--disable-fermion-reps \
--disable-gparity \
CXX=hipcc MPICXX=mpicxx \
CXXFLAGS="-fPIC -I/opt/rocm-4.5.0/include/ -std=c++14 -I${MPICH_DIR}/include " \
LDFLAGS=" -L${MPICH_DIR}/lib -lmpi -L${CRAY_MPICH_ROOTDIR}/gtl/lib -lmpi_gtl_hsa "
HIPFLAGS = --amdgpu-target=gfx90a

30
systems/Crusher/dwf.slurm Normal file
View File

@ -0,0 +1,30 @@
#!/bin/bash
# Begin LSF Directives
#SBATCH -A LGT104
#SBATCH -t 01:00:00
##SBATCH -U openmpThu
##SBATCH -p ecp
#SBATCH -J DWF
#SBATCH -o DWF.%J
#SBATCH -e DWF.%J
#SBATCH -N 1
#SBATCH -n 1
#SBATCH --exclusive
DIR=.
module list
#export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
export MPICH_GPU_SUPPORT_ENABLED=1
export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
#export MPICH_SMP_SINGLE_COPY_MODE=NONE
#export MPICH_SMP_SINGLE_COPY_MODE=CMA
export OMP_NUM_THREADS=1
AT=8
echo MPICH_SMP_SINGLE_COPY_MODE $MPICH_SMP_SINGLE_COPY_MODE
PARAMS=" --accelerator-threads ${AT} --grid 24.24.24.24 --shm-mpi 0 --mpi 1.1.1.1"
srun --gpus-per-task 1 -n1 ./benchmarks/Benchmark_dwf_fp32 $PARAMS

27
systems/Crusher/dwf4.slurm Normal file
View File

@ -0,0 +1,27 @@
#!/bin/bash
# Begin LSF Directives
#SBATCH -A LGT104
#SBATCH -t 01:00:00
##SBATCH -U openmpThu
#SBATCH -J DWF
#SBATCH -o DWF.%J
#SBATCH -e DWF.%J
#SBATCH -N 1
#SBATCH -n 4
#SBATCH --exclusive
DIR=.
module list
export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
export MPICH_GPU_SUPPORT_ENABLED=1
#export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
export MPICH_SMP_SINGLE_COPY_MODE=NONE
#export MPICH_SMP_SINGLE_COPY_MODE=CMA
export OMP_NUM_THREADS=4
echo MPICH_SMP_SINGLE_COPY_MODE $MPICH_SMP_SINGLE_COPY_MODE
PARAMS=" --accelerator-threads 8 --grid 32.32.64.64 --mpi 1.1.2.2 --comms-overlap --shm 2048 --shm-mpi 0"
srun --gpus-per-task 1 -n4 ./mpiwrapper.sh ./benchmarks/Benchmark_dwf_fp32 $PARAMS

27
systems/Crusher/dwf8.slurm Normal file
View File

@ -0,0 +1,27 @@
#!/bin/bash
# Begin LSF Directives
#SBATCH -A LGT104
#SBATCH -t 01:00:00
##SBATCH -U openmpThu
#SBATCH -J DWF
#SBATCH -o DWF.%J
#SBATCH -e DWF.%J
#SBATCH -N 1
#SBATCH -n 8
#SBATCH --exclusive
DIR=.
module list
export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
export MPICH_GPU_SUPPORT_ENABLED=1
export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
#export MPICH_SMP_SINGLE_COPY_MODE=NONE
#export MPICH_SMP_SINGLE_COPY_MODE=CMA
export OMP_NUM_THREADS=1
echo MPICH_SMP_SINGLE_COPY_MODE $MPICH_SMP_SINGLE_COPY_MODE
PARAMS=" --accelerator-threads 8 --grid 32.64.64.64 --mpi 1.2.2.2 --comms-overlap --shm 2048 --shm-mpi 0"
srun --gpus-per-task 1 -n8 ./mpiwrapper.sh ./benchmarks/Benchmark_dwf_fp32 $PARAMS

12
systems/Crusher/mpiwrapper.sh Executable file
View File

@ -0,0 +1,12 @@
#!/bin/bash
lrank=$SLURM_LOCALID
export ROCR_VISIBLE_DEVICES=$SLURM_LOCALID
echo "`hostname` - $lrank device=$ROCR_VISIBLE_DEVICES binding=$BINDING"
$*

5
systems/Crusher/sourceme.sh Normal file
View File

@ -0,0 +1,5 @@
module load PrgEnv-gnu
module load rocm/4.5.0
module load gmp
module load cray-fftw
module load craype-accel-amd-gfx90a

26
systems/Spock/comms.slurm Normal file
View File

@ -0,0 +1,26 @@
#!/bin/bash
# Begin LSF Directives
#SBATCH -A LGT104
#SBATCH -t 01:00:00
##SBATCH -U openmpThu
#SBATCH -p ecp
#SBATCH -J comms
#SBATCH -o comms.%J
#SBATCH -e comms.%J
#SBATCH -N 1
#SBATCH -n 2
DIR=.
module list
export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
export MPICH_GPU_SUPPORT_ENABLED=1
#export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
#export MPICH_SMP_SINGLE_COPY_MODE=CMA
export MPICH_SMP_SINGLE_COPY_MODE=NONE
export OMP_NUM_THREADS=8
AT=8
echo MPICH_SMP_SINGLE_COPY_MODE $MPICH_SMP_SINGLE_COPY_MODE
PARAMS=" --accelerator-threads ${AT} --grid 64.64.32.32 --mpi 2.1.1.1 "
srun -n2 --label -c$OMP_NUM_THREADS --gpus-per-task=1 ./mpiwrapper.sh ./benchmarks/Benchmark_comms_host_device $PARAMS

12
systems/Spock/config-command Normal file
View File

@ -0,0 +1,12 @@
../../configure --enable-comms=mpi-auto \
--enable-unified=no \
--enable-shm=nvlink \
--enable-accelerator=hip \
--enable-gen-simd-width=64 \
--enable-simd=GPU \
--disable-fermion-reps \
--disable-gparity \
CXX=hipcc MPICXX=mpicxx \
CXXFLAGS="-fPIC -I/opt/rocm-4.3.0/include/ -std=c++14 -I${MPICH_DIR}/include " \
--prefix=/ccs/home/chulwoo/Grid \
LDFLAGS=" -L${MPICH_DIR}/lib -lmpi -L${CRAY_MPICH_ROOTDIR}/gtl/lib -lmpi_gtl_hsa "

26
systems/Spock/dwf.slurm Normal file
View File

@ -0,0 +1,26 @@
#!/bin/bash
# Begin LSF Directives
#SBATCH -A LGT104
#SBATCH -t 01:00:00
##SBATCH -U openmpThu
#SBATCH -p ecp
#SBATCH -J DWF
#SBATCH -o DWF.%J
#SBATCH -e DWF.%J
#SBATCH -N 1
#SBATCH -n 1
DIR=.
module list
export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
export MPICH_GPU_SUPPORT_ENABLED=1
#export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
#export MPICH_SMP_SINGLE_COPY_MODE=NONE
export MPICH_SMP_SINGLE_COPY_MODE=CMA
export OMP_NUM_THREADS=8
AT=8
echo MPICH_SMP_SINGLE_COPY_MODE $MPICH_SMP_SINGLE_COPY_MODE
PARAMS=" --accelerator-threads ${AT} --grid 32.32.32.32 --mpi 1.1.1.1 --comms-overlap"
srun -n1 --label -c$OMP_NUM_THREADS --gpus-per-task=1 ./mpiwrapper.sh ./benchmarks/Benchmark_dwf_fp32 $PARAMS

26
systems/Spock/dwf4.slurm Normal file
View File

@ -0,0 +1,26 @@
#!/bin/bash
# Begin LSF Directives
#SBATCH -A LGT104
#SBATCH -t 01:00:00
##SBATCH -U openmpThu
#SBATCH -p ecp
#SBATCH -J DWF
#SBATCH -o DWF.%J
#SBATCH -e DWF.%J
#SBATCH -N 1
#SBATCH -n 4
DIR=.
module list
export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
export MPICH_GPU_SUPPORT_ENABLED=1
#export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
export MPICH_SMP_SINGLE_COPY_MODE=NONE
#export MPICH_SMP_SINGLE_COPY_MODE=CMA
export OMP_NUM_THREADS=8
AT=8
echo MPICH_SMP_SINGLE_COPY_MODE $MPICH_SMP_SINGLE_COPY_MODE
PARAMS=" --accelerator-threads ${AT} --grid 32.32.64.64 --mpi 1.1.2.2 --comms-overlap --shm 2048 --shm-mpi 0"
srun -n4 --label -c$OMP_NUM_THREADS --gpus-per-task=1 ./mpiwrapper.sh ./benchmarks/Benchmark_dwf_fp32 $PARAMS

26
systems/Spock/dwf8.slurm Normal file
View File

@ -0,0 +1,26 @@
#!/bin/bash
# Begin LSF Directives
#SBATCH -A LGT104
#SBATCH -t 01:00:00
##SBATCH -U openmpThu
#SBATCH -p ecp
#SBATCH -J DWF
#SBATCH -o DWF.%J
#SBATCH -e DWF.%J
#SBATCH -N 2
#SBATCH -n 8
DIR=.
module list
export MPIR_CVAR_GPU_EAGER_DEVICE_MEM=0
export MPICH_GPU_SUPPORT_ENABLED=1
#export MPICH_SMP_SINGLE_COPY_MODE=XPMEM
export MPICH_SMP_SINGLE_COPY_MODE=NONE
#export MPICH_SMP_SINGLE_COPY_MODE=CMA
export OMP_NUM_THREADS=8
AT=8
echo MPICH_SMP_SINGLE_COPY_MODE $MPICH_SMP_SINGLE_COPY_MODE
PARAMS=" --accelerator-threads ${AT} --grid 32.64.64.64 --mpi 1.2.2.2 --comms-overlap --shm 2048 --shm-mpi 0"
srun -n8 --label -c$OMP_NUM_THREADS --gpus-per-task=1 ./mpiwrapper.sh ./benchmarks/Benchmark_dwf_fp32 $PARAMS

12
systems/Spock/mpiwrapper.sh Executable file
View File

@ -0,0 +1,12 @@
#!/bin/bash
lrank=$SLURM_LOCALID
export ROCR_VISIBLE_DEVICES=$SLURM_LOCALID
echo "`hostname` - $lrank device=$ROCR_VISIBLE_DEVICES binding=$BINDING"
$*

5
systems/Spock/sourceme.sh Normal file
View File

@ -0,0 +1,5 @@
module load PrgEnv-gnu
module load rocm/4.3.0
module load gmp
module load cray-fftw
module load craype-accel-amd-gfx908

179
systems/Summit/comms.4node Normal file
View File

@ -0,0 +1,179 @@
OPENMPI detected
AcceleratorCudaInit[0]: ========================
AcceleratorCudaInit[0]: Device Number : 0
AcceleratorCudaInit[0]: ========================
AcceleratorCudaInit[0]: Device identifier: Tesla V100-SXM2-16GB
AcceleratorCudaInit[0]: totalGlobalMem: 16911433728
AcceleratorCudaInit[0]: managedMemory: 1
AcceleratorCudaInit[0]: isMultiGpuBoard: 0
AcceleratorCudaInit[0]: warpSize: 32
AcceleratorCudaInit[0]: pciBusID: 4
AcceleratorCudaInit[0]: pciDeviceID: 0
AcceleratorCudaInit[0]: maxGridSize (2147483647,65535,65535)
AcceleratorCudaInit: rank 0 setting device to node rank 0
AcceleratorCudaInit: Configure options --enable-setdevice=yes
local rank 0 device 0 bus id: 0004:04:00.0
AcceleratorCudaInit: ================================================
SharedMemoryMpi: World communicator of size 24
SharedMemoryMpi: Node communicator of size 6
0SharedMemoryMpi: SharedMemoryMPI.cc acceleratorAllocDevice 1073741824bytes at 0x200060000000 for comms buffers
Setting up IPC
__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
__|_ | | | | | | | | | | | | _|__
__|_ _|__
__|_ GGGG RRRR III DDDD _|__
__|_ G R R I D D _|__
__|_ G R R I D D _|__
__|_ G GG RRRR I D D _|__
__|_ G G R R I D D _|__
__|_ GGGG R R III DDDD _|__
__|_ _|__
__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
| | | | | | | | | | | | | |
Copyright (C) 2015 Peter Boyle, Azusa Yamaguchi, Guido Cossu, Antonin Portelli and other authors
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.
Current Grid git commit hash=7cb1ff7395a5833ded6526c43891bd07a0436290: (HEAD -> develop, origin/develop, origin/HEAD) clean
Grid : Message : ================================================
Grid : Message : MPI is initialised and logging filters activated
Grid : Message : ================================================
Grid : Message : Requested 1073741824 byte stencil comms buffers
AcceleratorCudaInit: rank 1 setting device to node rank 1
AcceleratorCudaInit: Configure options --enable-setdevice=yes
local rank 1 device 1 bus id: 0004:05:00.0
AcceleratorCudaInit: rank 2 setting device to node rank 2
AcceleratorCudaInit: Configure options --enable-setdevice=yes
local rank 2 device 2 bus id: 0004:06:00.0
AcceleratorCudaInit: rank 5 setting device to node rank 5
AcceleratorCudaInit: Configure options --enable-setdevice=yes
local rank 5 device 5 bus id: 0035:05:00.0
AcceleratorCudaInit: rank 4 setting device to node rank 4
AcceleratorCudaInit: Configure options --enable-setdevice=yes
local rank 4 device 4 bus id: 0035:04:00.0
AcceleratorCudaInit: rank 3 setting device to node rank 3
AcceleratorCudaInit: Configure options --enable-setdevice=yes
local rank 3 device 3 bus id: 0035:03:00.0
Grid : Message : MemoryManager Cache 13529146982 bytes
Grid : Message : MemoryManager::Init() setting up
Grid : Message : MemoryManager::Init() cache pool for recent allocations: SMALL 8 LARGE 2
Grid : Message : MemoryManager::Init() Non unified: Caching accelerator data in dedicated memory
Grid : Message : MemoryManager::Init() Using cudaMalloc
Grid : Message : 2.137929 s : Grid is setup to use 6 threads
Grid : Message : 2.137941 s : Number of iterations to average: 250
Grid : Message : 2.137950 s : ====================================================================================================
Grid : Message : 2.137958 s : = Benchmarking sequential halo exchange from host memory
Grid : Message : 2.137966 s : ====================================================================================================
Grid : Message : 2.137974 s : L Ls bytes MB/s uni MB/s bidi
AcceleratorCudaInit: rank 22 setting device to node rank 4
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 10 setting device to node rank 4
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 15 setting device to node rank 3
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 21 setting device to node rank 3
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 20 setting device to node rank 2
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 7 setting device to node rank 1
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 9 setting device to node rank 3
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 11 setting device to node rank 5
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 8 setting device to node rank 2
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 6 setting device to node rank 0
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 19 setting device to node rank 1
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 23 setting device to node rank 5
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 18 setting device to node rank 0
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 12 setting device to node rank 0
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 16 setting device to node rank 4
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 13 setting device to node rank 1
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 14 setting device to node rank 2
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 17 setting device to node rank 5
AcceleratorCudaInit: Configure options --enable-setdevice=yes
Grid : Message : 2.604949 s : 8 8 393216 89973.9 179947.8
Grid : Message : 2.668249 s : 8 8 393216 18650.3 37300.5
Grid : Message : 2.732288 s : 8 8 393216 18428.5 36857.1
Grid : Message : 2.753565 s : 8 8 393216 55497.2 110994.4
Grid : Message : 2.808960 s : 12 8 1327104 100181.5 200363.0
Grid : Message : 3.226900 s : 12 8 1327104 20600.5 41201.0
Grid : Message : 3.167459 s : 12 8 1327104 24104.6 48209.2
Grid : Message : 3.227660 s : 12 8 1327104 66156.7 132313.5
Grid : Message : 3.413570 s : 16 8 3145728 56174.4 112348.8
Grid : Message : 3.802697 s : 16 8 3145728 24255.9 48511.7
Grid : Message : 4.190498 s : 16 8 3145728 24336.7 48673.4
Grid : Message : 4.385171 s : 16 8 3145728 48484.1 96968.2
Grid : Message : 4.805284 s : 20 8 6144000 46380.5 92761.1
Grid : Message : 5.562975 s : 20 8 6144000 24328.5 48656.9
Grid : Message : 6.322562 s : 20 8 6144000 24266.7 48533.4
Grid : Message : 6.773598 s : 20 8 6144000 40868.5 81736.9
Grid : Message : 7.600999 s : 24 8 10616832 40198.3 80396.6
Grid : Message : 8.912917 s : 24 8 10616832 24279.5 48559.1
Grid : Message : 10.220961 s : 24 8 10616832 24350.2 48700.4
Grid : Message : 11.728250 s : 24 8 10616832 37390.9 74781.8
Grid : Message : 12.497258 s : 28 8 16859136 36792.2 73584.5
Grid : Message : 14.585387 s : 28 8 16859136 24222.2 48444.3
Grid : Message : 16.664783 s : 28 8 16859136 24323.4 48646.8
Grid : Message : 17.955238 s : 28 8 16859136 39194.7 78389.4
Grid : Message : 20.136479 s : 32 8 25165824 35718.3 71436.5
Grid : Message : 23.241958 s : 32 8 25165824 24311.4 48622.9
Grid : Message : 26.344810 s : 32 8 25165824 24331.9 48663.7
Grid : Message : 28.384420 s : 32 8 25165824 37016.3 74032.7
Grid : Message : 28.388879 s : ====================================================================================================
Grid : Message : 28.388894 s : = Benchmarking sequential halo exchange from GPU memory
Grid : Message : 28.388909 s : ====================================================================================================
Grid : Message : 28.388924 s : L Ls bytes MB/s uni MB/s bidi
Grid : Message : 28.553993 s : 8 8 393216 8272.4 16544.7
Grid : Message : 28.679592 s : 8 8 393216 9395.4 18790.8
Grid : Message : 28.811112 s : 8 8 393216 8971.0 17942.0
Grid : Message : 28.843770 s : 8 8 393216 36145.6 72291.2
Grid : Message : 28.981754 s : 12 8 1327104 49591.6 99183.2
Grid : Message : 29.299764 s : 12 8 1327104 12520.8 25041.7
Grid : Message : 29.620288 s : 12 8 1327104 12422.2 24844.4
Grid : Message : 29.657645 s : 12 8 1327104 106637.5 213275.1
Grid : Message : 29.952933 s : 16 8 3145728 43939.2 87878.5
Grid : Message : 30.585411 s : 16 8 3145728 14922.1 29844.2
Grid : Message : 31.219781 s : 16 8 3145728 14877.2 29754.4
Grid : Message : 31.285017 s : 16 8 3145728 144724.3 289448.7
Grid : Message : 31.706443 s : 20 8 6144000 54676.2 109352.4
Grid : Message : 32.739205 s : 20 8 6144000 17848.0 35696.1
Grid : Message : 33.771852 s : 20 8 6144000 17849.9 35699.7
Grid : Message : 33.871981 s : 20 8 6144000 184141.4 368282.8
Grid : Message : 34.536808 s : 24 8 10616832 55784.3 111568.6
Grid : Message : 36.275648 s : 24 8 10616832 18317.6 36635.3
Grid : Message : 37.997181 s : 24 8 10616832 18501.7 37003.4
Grid : Message : 38.140442 s : 24 8 10616832 222383.9 444767.9
Grid : Message : 39.177222 s : 28 8 16859136 56609.7 113219.4
Grid : Message : 41.874755 s : 28 8 16859136 18749.9 37499.8
Grid : Message : 44.529381 s : 28 8 16859136 19052.9 38105.8
Grid : Message : 44.742192 s : 28 8 16859136 237717.1 475434.2
Grid : Message : 46.184000 s : 32 8 25165824 57091.2 114182.4
Grid : Message : 50.734740 s : 32 8 25165824 19411.0 38821.9
Grid : Message : 53.931228 s : 32 8 25165824 19570.6 39141.2
Grid : Message : 54.238467 s : 32 8 25165824 245765.6 491531.2
Grid : Message : 54.268664 s : ====================================================================================================
Grid : Message : 54.268680 s : = All done; Bye Bye
Grid : Message : 54.268691 s : ====================================================================================================

14
systems/Summit/config-command Normal file
View File

@ -0,0 +1,14 @@
../../configure --enable-comms=mpi \
--enable-simd=GPU \
--enable-gen-simd-width=32 \
--enable-unified=no \
--enable-shm=nvlink \
--disable-gparity \
--enable-setdevice \
--disable-fermion-reps \
--enable-accelerator=cuda \
--prefix /ccs/home/paboyle/prefix \
CXX=nvcc \
LDFLAGS=-L/ccs/home/paboyle/prefix/lib/ \
CXXFLAGS="-ccbin mpicxx -gencode arch=compute_70,code=sm_70 -I/ccs/home/paboyle/prefix/include/ -std=c++14"

206
systems/Summit/dwf.24.4node Normal file
View File

@ -0,0 +1,206 @@
OPENMPI detected
AcceleratorCudaInit[0]: ========================
AcceleratorCudaInit[0]: Device Number : 0
AcceleratorCudaInit[0]: ========================
AcceleratorCudaInit[0]: Device identifier: Tesla V100-SXM2-16GB
AcceleratorCudaInit[0]: totalGlobalMem: 16911433728
AcceleratorCudaInit[0]: managedMemory: 1
AcceleratorCudaInit[0]: isMultiGpuBoard: 0
AcceleratorCudaInit[0]: warpSize: 32
AcceleratorCudaInit[0]: pciBusID: 4
AcceleratorCudaInit[0]: pciDeviceID: 0
AcceleratorCudaInit[0]: maxGridSize (2147483647,65535,65535)
AcceleratorCudaInit: rank 0 setting device to node rank 0
AcceleratorCudaInit: Configure options --enable-setdevice=yes
local rank 0 device 0 bus id: 0004:04:00.0
AcceleratorCudaInit: ================================================
SharedMemoryMpi: World communicator of size 24
SharedMemoryMpi: Node communicator of size 6
0SharedMemoryMpi: SharedMemoryMPI.cc acceleratorAllocDevice 2147483648bytes at 0x200080000000 for comms buffers
AcceleratorCudaInit: rank 3 setting device to node rank 3
AcceleratorCudaInit: Configure options --enable-setdevice=yes
local rank 3 device 3 bus id: 0035:03:00.0
AcceleratorCudaInit: rank 5 setting device to node rank 5
AcceleratorCudaInit: Configure options --enable-setdevice=yes
local rank 5 device 5 bus id: 0035:05:00.0
Setting up IPC
__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
__|_ | | | | | | | | | | | | _|__
__|_ _|__
__|_ GGGG RRRR III DDDD _|__
__|_ G R R I D D _|__
__|_ G R R I D D _|__
__|_ G GG RRRR I D D _|__
__|_ G G R R I D D _|__
__|_ GGGG R R III DDDD _|__
__|_ _|__
__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
| | | | | | | | | | | | | |
Copyright (C) 2015 Peter Boyle, Azusa Yamaguchi, Guido Cossu, Antonin Portelli and other authors
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
AcceleratorCudaInit: rank 4 setting device to node rank 4
AcceleratorCudaInit: Configure options --enable-setdevice=yes
local rank 4 device 4 bus id: 0035:04:00.0
AcceleratorCudaInit: rank 1 setting device to node rank 1
AcceleratorCudaInit: Configure options --enable-setdevice=yes
local rank 1 device 1 bus id: 0004:05:00.0
AcceleratorCudaInit: rank 2 setting device to node rank 2
AcceleratorCudaInit: Configure options --enable-setdevice=yes
local rank 2 device 2 bus id: 0004:06:00.0
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.
Current Grid git commit hash=7cb1ff7395a5833ded6526c43891bd07a0436290: (HEAD -> develop, origin/develop, origin/HEAD) clean
Grid : Message : ================================================
Grid : Message : MPI is initialised and logging filters activated
Grid : Message : ================================================
Grid : Message : Requested 2147483648 byte stencil comms buffers
Grid : Message : MemoryManager Cache 8388608000 bytes
Grid : Message : MemoryManager::Init() setting up
Grid : Message : MemoryManager::Init() cache pool for recent allocations: SMALL 8 LARGE 2
Grid : Message : MemoryManager::Init() Non unified: Caching accelerator data in dedicated memory
Grid : Message : MemoryManager::Init() Using cudaMalloc
Grid : Message : 1.731905 s : Grid Layout
Grid : Message : 1.731915 s : Global lattice size : 48 48 48 72
Grid : Message : 1.731928 s : OpenMP threads : 6
Grid : Message : 1.731938 s : MPI tasks : 2 2 2 3
AcceleratorCudaInit: rank 9 setting device to node rank 3
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 23 setting device to node rank 5
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 22 setting device to node rank 4
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 21 setting device to node rank 3
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 18 setting device to node rank 0
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 6 setting device to node rank 0
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 7 setting device to node rank 1
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 10 setting device to node rank 4
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 8 setting device to node rank 2
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 11 setting device to node rank 5
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 20 setting device to node rank 2
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 19 setting device to node rank 1
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 13 setting device to node rank 1
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 12 setting device to node rank 0
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 14 setting device to node rank 2
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 16 setting device to node rank 4
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 15 setting device to node rank 3
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 17 setting device to node rank 5
AcceleratorCudaInit: Configure options --enable-setdevice=yes
Grid : Message : 2.683494 s : Making s innermost grids
Grid : Message : 2.780034 s : Initialising 4d RNG
Grid : Message : 2.833099 s : Intialising parallel RNG with unique string 'The 4D RNG'
Grid : Message : 2.833121 s : Seed SHA256: 49db4542db694e3b1a74bf2592a8c1b83bfebbe18401693c2609a4c3af1
Grid : Message : 2.916841 s : Initialising 5d RNG
Grid : Message : 3.762880 s : Intialising parallel RNG with unique string 'The 5D RNG'
Grid : Message : 3.762902 s : Seed SHA256: b6316f2fac44ce14111f93e0296389330b077bfd0a7b359f781c58589f8a
Grid : Message : 5.264345 s : Initialised RNGs
Grid : Message : 6.489904 s : Drawing gauge field
Grid : Message : 6.729262 s : Random gauge initialised
Grid : Message : 7.781273 s : Setting up Cshift based reference
Grid : Message : 8.725313 s : *****************************************************************
Grid : Message : 8.725332 s : * Kernel options --dslash-generic, --dslash-unroll, --dslash-asm
Grid : Message : 8.725342 s : *****************************************************************
Grid : Message : 8.725352 s : *****************************************************************
Grid : Message : 8.725362 s : * Benchmarking DomainWallFermionR::Dhop
Grid : Message : 8.725372 s : * Vectorising space-time by 4
Grid : Message : 8.725383 s : * VComplexF size is 32 B
Grid : Message : 8.725395 s : * SINGLE precision
Grid : Message : 8.725405 s : * Using Overlapped Comms/Compute
Grid : Message : 8.725415 s : * Using GENERIC Nc WilsonKernels
Grid : Message : 8.725425 s : *****************************************************************
Grid : Message : 9.465229 s : Called warmup
Grid : Message : 58.646066 s : Called Dw 3000 times in 4.91764e+07 us
Grid : Message : 58.646121 s : mflop/s = 1.02592e+07
Grid : Message : 58.646134 s : mflop/s per rank = 427468
Grid : Message : 58.646145 s : mflop/s per node = 2.56481e+06
Grid : Message : 58.646156 s : RF GiB/s (base 2) = 20846.5
Grid : Message : 58.646166 s : mem GiB/s (base 2) = 13029.1
Grid : Message : 58.648008 s : norm diff 1.04778e-13
Grid : Message : 58.734885 s : #### Dhop calls report
Grid : Message : 58.734897 s : WilsonFermion5D Number of DhopEO Calls : 6002
Grid : Message : 58.734909 s : WilsonFermion5D TotalTime /Calls : 8217.71 us
Grid : Message : 58.734922 s : WilsonFermion5D CommTime /Calls : 7109.5 us
Grid : Message : 58.734933 s : WilsonFermion5D FaceTime /Calls : 446.623 us
Grid : Message : 58.734943 s : WilsonFermion5D ComputeTime1/Calls : 18.0558 us
Grid : Message : 58.734953 s : WilsonFermion5D ComputeTime2/Calls : 731.097 us
Grid : Message : 58.734979 s : Average mflops/s per call : 4.8157e+09
Grid : Message : 58.734989 s : Average mflops/s per call per rank : 2.00654e+08
Grid : Message : 58.734999 s : Average mflops/s per call per node : 1.20393e+09
Grid : Message : 58.735008 s : Average mflops/s per call (full) : 1.04183e+07
Grid : Message : 58.735017 s : Average mflops/s per call per rank (full): 434094
Grid : Message : 58.735026 s : Average mflops/s per call per node (full): 2.60456e+06
Grid : Message : 58.735035 s : WilsonFermion5D Stencil
Grid : Message : 58.735043 s : WilsonFermion5D StencilEven
Grid : Message : 58.735051 s : WilsonFermion5D StencilOdd
Grid : Message : 58.735059 s : WilsonFermion5D Stencil Reporti()
Grid : Message : 58.735067 s : WilsonFermion5D StencilEven Reporti()
Grid : Message : 58.735075 s : WilsonFermion5D StencilOdd Reporti()
Grid : Message : 64.934380 s : Compare to naive wilson implementation Dag to verify correctness
Grid : Message : 64.934740 s : Called DwDag
Grid : Message : 64.934870 s : norm dag result 12.0422
Grid : Message : 64.120756 s : norm dag ref 12.0422
Grid : Message : 64.149389 s : norm dag diff 7.6644e-14
Grid : Message : 64.317786 s : Calling Deo and Doe and //assert Deo+Doe == Dunprec
Grid : Message : 64.465331 s : src_e0.499995
Grid : Message : 64.524653 s : src_o0.500005
Grid : Message : 64.558706 s : *********************************************************
Grid : Message : 64.558717 s : * Benchmarking DomainWallFermionF::DhopEO
Grid : Message : 64.558727 s : * Vectorising space-time by 4
Grid : Message : 64.558737 s : * SINGLE precision
Grid : Message : 64.558745 s : * Using Overlapped Comms/Compute
Grid : Message : 64.558753 s : * Using GENERIC Nc WilsonKernels
Grid : Message : 64.558761 s : *********************************************************
Grid : Message : 92.702145 s : Deo mflop/s = 8.97692e+06
Grid : Message : 92.702185 s : Deo mflop/s per rank 374038
Grid : Message : 92.702198 s : Deo mflop/s per node 2.24423e+06
Grid : Message : 92.702209 s : #### Dhop calls report
Grid : Message : 92.702223 s : WilsonFermion5D Number of DhopEO Calls : 3001
Grid : Message : 92.702240 s : WilsonFermion5D TotalTime /Calls : 9377.88 us
Grid : Message : 92.702257 s : WilsonFermion5D CommTime /Calls : 8221.84 us
Grid : Message : 92.702277 s : WilsonFermion5D FaceTime /Calls : 543.548 us
Grid : Message : 92.702301 s : WilsonFermion5D ComputeTime1/Calls : 20.936 us
Grid : Message : 92.702322 s : WilsonFermion5D ComputeTime2/Calls : 732.33 us
Grid : Message : 92.702376 s : Average mflops/s per call : 4.13001e+09
Grid : Message : 92.702387 s : Average mflops/s per call per rank : 1.72084e+08
Grid : Message : 92.702397 s : Average mflops/s per call per node : 1.0325e+09
Grid : Message : 92.702407 s : Average mflops/s per call (full) : 9.12937e+06
Grid : Message : 92.702416 s : Average mflops/s per call per rank (full): 380391
Grid : Message : 92.702426 s : Average mflops/s per call per node (full): 2.28234e+06
Grid : Message : 92.702435 s : WilsonFermion5D Stencil
Grid : Message : 92.702443 s : WilsonFermion5D StencilEven
Grid : Message : 92.702451 s : WilsonFermion5D StencilOdd
Grid : Message : 92.702459 s : WilsonFermion5D Stencil Reporti()
Grid : Message : 92.702467 s : WilsonFermion5D StencilEven Reporti()
Grid : Message : 92.702475 s : WilsonFermion5D StencilOdd Reporti()
Grid : Message : 92.772983 s : r_e6.02121
Grid : Message : 92.786384 s : r_o6.02102
Grid : Message : 92.799622 s : res12.0422
Grid : Message : 93.860500 s : norm diff 0
Grid : Message : 93.162026 s : norm diff even 0
Grid : Message : 93.197529 s : norm diff odd 0

206
systems/Summit/dwf.32.4node Normal file
View File

@ -0,0 +1,206 @@
OPENMPI detected
AcceleratorCudaInit[0]: ========================
AcceleratorCudaInit[0]: Device Number : 0
AcceleratorCudaInit[0]: ========================
AcceleratorCudaInit[0]: Device identifier: Tesla V100-SXM2-16GB
AcceleratorCudaInit[0]: totalGlobalMem: 16911433728
AcceleratorCudaInit[0]: managedMemory: 1
AcceleratorCudaInit[0]: isMultiGpuBoard: 0
AcceleratorCudaInit[0]: warpSize: 32
AcceleratorCudaInit[0]: pciBusID: 4
AcceleratorCudaInit[0]: pciDeviceID: 0
AcceleratorCudaInit[0]: maxGridSize (2147483647,65535,65535)
AcceleratorCudaInit: rank 0 setting device to node rank 0
AcceleratorCudaInit: Configure options --enable-setdevice=yes
local rank 0 device 0 bus id: 0004:04:00.0
AcceleratorCudaInit: ================================================
SharedMemoryMpi: World communicator of size 24
SharedMemoryMpi: Node communicator of size 6
0SharedMemoryMpi: SharedMemoryMPI.cc acceleratorAllocDevice 2147483648bytes at 0x200080000000 for comms buffers
Setting up IPC
__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
__|_ | | | | | | | | | | | | _|__
__|_ _|__
__|_ GGGG RRRR III DDDD _|__
__|_ G R R I D D _|__
__|_ G R R I D D _|__
__|_ G GG RRRR I D D _|__
__|_ G G R R I D D _|__
__|_ GGGG R R III DDDD _|__
__|_ _|__
__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
__|__|__|__|__|__|__|__|__|__|__|__|__|__|__
| | | | | | | | | | | | | |
Copyright (C) 2015 Peter Boyle, Azusa Yamaguchi, Guido Cossu, Antonin Portelli and other authors
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
AcceleratorCudaInit: rank 2 setting device to node rank 2
AcceleratorCudaInit: Configure options --enable-setdevice=yes
local rank 2 device 2 bus id: 0004:06:00.0
AcceleratorCudaInit: rank 1 setting device to node rank 1
AcceleratorCudaInit: Configure options --enable-setdevice=yes
local rank 1 device 1 bus id: 0004:05:00.0
AcceleratorCudaInit: rank 4 setting device to node rank 4
AcceleratorCudaInit: Configure options --enable-setdevice=yes
local rank 4 device 4 bus id: 0035:04:00.0
AcceleratorCudaInit: rank 3 setting device to node rank 3
AcceleratorCudaInit: Configure options --enable-setdevice=yes
local rank 3 device 3 bus id: 0035:03:00.0
AcceleratorCudaInit: rank 5 setting device to node rank 5
AcceleratorCudaInit: Configure options --enable-setdevice=yes
local rank 5 device 5 bus id: 0035:05:00.0
GNU General Public License for more details.
Current Grid git commit hash=7cb1ff7395a5833ded6526c43891bd07a0436290: (HEAD -> develop, origin/develop, origin/HEAD) clean
Grid : Message : ================================================
Grid : Message : MPI is initialised and logging filters activated
Grid : Message : ================================================
Grid : Message : Requested 2147483648 byte stencil comms buffers
Grid : Message : MemoryManager Cache 8388608000 bytes
Grid : Message : MemoryManager::Init() setting up
Grid : Message : MemoryManager::Init() cache pool for recent allocations: SMALL 8 LARGE 2
Grid : Message : MemoryManager::Init() Non unified: Caching accelerator data in dedicated memory
Grid : Message : MemoryManager::Init() Using cudaMalloc
Grid : Message : 1.544984 s : Grid Layout
Grid : Message : 1.544992 s : Global lattice size : 64 64 64 96
Grid : Message : 1.545003 s : OpenMP threads : 6
Grid : Message : 1.545011 s : MPI tasks : 2 2 2 3
AcceleratorCudaInit: rank 8 setting device to node rank 2
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 6 setting device to node rank 0
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 11 setting device to node rank 5
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 16 setting device to node rank 4
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 17 setting device to node rank 5
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 13 setting device to node rank 1
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 12 setting device to node rank 0
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 21 setting device to node rank 3
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 23 setting device to node rank 5
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 22 setting device to node rank 4
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 19 setting device to node rank 1
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 18 setting device to node rank 0
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 7 setting device to node rank 1
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 10 setting device to node rank 4
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 9 setting device to node rank 3
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 14 setting device to node rank 2
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 15 setting device to node rank 3
AcceleratorCudaInit: Configure options --enable-setdevice=yes
AcceleratorCudaInit: rank 20 setting device to node rank 2
AcceleratorCudaInit: Configure options --enable-setdevice=yes
Grid : Message : 2.994920 s : Making s innermost grids
Grid : Message : 2.232502 s : Initialising 4d RNG
Grid : Message : 2.397047 s : Intialising parallel RNG with unique string 'The 4D RNG'
Grid : Message : 2.397069 s : Seed SHA256: 49db4542db694e3b1a74bf2592a8c1b83bfebbe18401693c2609a4c3af1
Grid : Message : 2.653140 s : Initialising 5d RNG
Grid : Message : 5.285347 s : Intialising parallel RNG with unique string 'The 5D RNG'
Grid : Message : 5.285369 s : Seed SHA256: b6316f2fac44ce14111f93e0296389330b077bfd0a7b359f781c58589f8a
Grid : Message : 9.994738 s : Initialised RNGs
Grid : Message : 13.153426 s : Drawing gauge field
Grid : Message : 13.825697 s : Random gauge initialised
Grid : Message : 18.537657 s : Setting up Cshift based reference
Grid : Message : 22.296755 s : *****************************************************************
Grid : Message : 22.296781 s : * Kernel options --dslash-generic, --dslash-unroll, --dslash-asm
Grid : Message : 22.296791 s : *****************************************************************
Grid : Message : 22.296800 s : *****************************************************************
Grid : Message : 22.296809 s : * Benchmarking DomainWallFermionR::Dhop
Grid : Message : 22.296818 s : * Vectorising space-time by 4
Grid : Message : 22.296828 s : * VComplexF size is 32 B
Grid : Message : 22.296838 s : * SINGLE precision
Grid : Message : 22.296847 s : * Using Overlapped Comms/Compute
Grid : Message : 22.296855 s : * Using GENERIC Nc WilsonKernels
Grid : Message : 22.296863 s : *****************************************************************
Grid : Message : 24.746452 s : Called warmup
Grid : Message : 137.525756 s : Called Dw 3000 times in 1.12779e+08 us
Grid : Message : 137.525818 s : mflop/s = 1.41383e+07
Grid : Message : 137.525831 s : mflop/s per rank = 589097
Grid : Message : 137.525843 s : mflop/s per node = 3.53458e+06
Grid : Message : 137.525854 s : RF GiB/s (base 2) = 28728.7
Grid : Message : 137.525864 s : mem GiB/s (base 2) = 17955.5
Grid : Message : 137.693645 s : norm diff 1.04885e-13
Grid : Message : 137.965585 s : #### Dhop calls report
Grid : Message : 137.965598 s : WilsonFermion5D Number of DhopEO Calls : 6002
Grid : Message : 137.965612 s : WilsonFermion5D TotalTime /Calls : 18899.7 us
Grid : Message : 137.965624 s : WilsonFermion5D CommTime /Calls : 16041.4 us
Grid : Message : 137.965634 s : WilsonFermion5D FaceTime /Calls : 859.705 us
Grid : Message : 137.965644 s : WilsonFermion5D ComputeTime1/Calls : 70.5881 us
Grid : Message : 137.965654 s : WilsonFermion5D ComputeTime2/Calls : 2094.8 us
Grid : Message : 137.965682 s : Average mflops/s per call : 3.87638e+09
Grid : Message : 137.965692 s : Average mflops/s per call per rank : 1.61516e+08
Grid : Message : 137.965702 s : Average mflops/s per call per node : 9.69095e+08
Grid : Message : 137.965712 s : Average mflops/s per call (full) : 1.43168e+07
Grid : Message : 137.965721 s : Average mflops/s per call per rank (full): 596533
Grid : Message : 137.965730 s : Average mflops/s per call per node (full): 3.5792e+06
Grid : Message : 137.965740 s : WilsonFermion5D Stencil
Grid : Message : 137.965748 s : WilsonFermion5D StencilEven
Grid : Message : 137.965756 s : WilsonFermion5D StencilOdd
Grid : Message : 137.965764 s : WilsonFermion5D Stencil Reporti()
Grid : Message : 137.965772 s : WilsonFermion5D StencilEven Reporti()
Grid : Message : 137.965780 s : WilsonFermion5D StencilOdd Reporti()
Grid : Message : 156.554605 s : Compare to naive wilson implementation Dag to verify correctness
Grid : Message : 156.554632 s : Called DwDag
Grid : Message : 156.554642 s : norm dag result 12.0421
Grid : Message : 156.639265 s : norm dag ref 12.0421
Grid : Message : 156.888281 s : norm dag diff 7.62057e-14
Grid : Message : 157.609797 s : Calling Deo and Doe and //assert Deo+Doe == Dunprec
Grid : Message : 158.208630 s : src_e0.499996
Grid : Message : 158.162447 s : src_o0.500004
Grid : Message : 158.267780 s : *********************************************************
Grid : Message : 158.267791 s : * Benchmarking DomainWallFermionF::DhopEO
Grid : Message : 158.267801 s : * Vectorising space-time by 4
Grid : Message : 158.267811 s : * SINGLE precision
Grid : Message : 158.267820 s : * Using Overlapped Comms/Compute
Grid : Message : 158.267828 s : * Using GENERIC Nc WilsonKernels
Grid : Message : 158.267836 s : *********************************************************
Grid : Message : 216.487829 s : Deo mflop/s = 1.37283e+07
Grid : Message : 216.487869 s : Deo mflop/s per rank 572011
Grid : Message : 216.487881 s : Deo mflop/s per node 3.43206e+06
Grid : Message : 216.487893 s : #### Dhop calls report
Grid : Message : 216.487903 s : WilsonFermion5D Number of DhopEO Calls : 3001
Grid : Message : 216.487913 s : WilsonFermion5D TotalTime /Calls : 19399.6 us
Grid : Message : 216.487923 s : WilsonFermion5D CommTime /Calls : 16475.4 us
Grid : Message : 216.487933 s : WilsonFermion5D FaceTime /Calls : 972.393 us
Grid : Message : 216.487943 s : WilsonFermion5D ComputeTime1/Calls : 49.8474 us
Grid : Message : 216.487953 s : WilsonFermion5D ComputeTime2/Calls : 2089.93 us
Grid : Message : 216.488001 s : Average mflops/s per call : 5.39682e+09
Grid : Message : 216.488011 s : Average mflops/s per call per rank : 2.24867e+08
Grid : Message : 216.488020 s : Average mflops/s per call per node : 1.3492e+09
Grid : Message : 216.488030 s : Average mflops/s per call (full) : 1.39479e+07
Grid : Message : 216.488039 s : Average mflops/s per call per rank (full): 581162
Grid : Message : 216.488048 s : Average mflops/s per call per node (full): 3.48697e+06
Grid : Message : 216.488057 s : WilsonFermion5D Stencil
Grid : Message : 216.488065 s : WilsonFermion5D StencilEven
Grid : Message : 216.488073 s : WilsonFermion5D StencilOdd
Grid : Message : 216.488081 s : WilsonFermion5D Stencil Reporti()
Grid : Message : 216.488089 s : WilsonFermion5D StencilEven Reporti()
Grid : Message : 216.488097 s : WilsonFermion5D StencilOdd Reporti()
Grid : Message : 217.384495 s : r_e6.02113
Grid : Message : 217.426121 s : r_o6.02096
Grid : Message : 217.472636 s : res12.0421
Grid : Message : 218.200068 s : norm diff 0
Grid : Message : 218.645673 s : norm diff even 0
Grid : Message : 218.816561 s : norm diff odd 0

25
systems/Summit/dwf16.lsf Normal file
View File

@ -0,0 +1,25 @@
#!/bin/bash
#BSUB -P LGT104
#BSUB -W 2:00
#BSUB -nnodes 16
#BSUB -J DWF
export OMP_NUM_THREADS=6
export PAMI_IBV_ADAPTER_AFFINITY=1
export PAMI_ENABLE_STRIPING=1
export OPT="--comms-concurrent --comms-overlap "
APP="./benchmarks/Benchmark_comms_host_device --mpi 4.4.4.3 "
jsrun --nrs 16 -a6 -g6 -c42 -dpacked -b packed:7 --latency_priority gpu-cpu --smpiargs=-gpu $APP > comms.16node.log
APP="./benchmarks/Benchmark_dwf_fp32 --grid 96.96.96.72 --mpi 4.4.4.3 --shm 2048 --shm-force-mpi 1 --device-mem 8000 --shm-force-mpi 1 $OPT "
jsrun --nrs 16 -a6 -g6 -c42 -dpacked -b packed:7 --latency_priority gpu-cpu --smpiargs=-gpu $APP > dwf.16node.24.log
APP="./benchmarks/Benchmark_dwf_fp32 --grid 128.128.128.96 --mpi 4.4.4.3 --shm 2048 --shm-force-mpi 1 --device-mem 8000 --shm-force-mpi 1 $OPT "
jsrun --nrs 16 -a6 -g6 -c42 -dpacked -b packed:7 --latency_priority gpu-cpu --smpiargs=-gpu $APP > dwf.16node.32.log

25
systems/Summit/dwf4.lsf Normal file
View File

@ -0,0 +1,25 @@
#!/bin/bash
#BSUB -P LGT104
#BSUB -W 2:00
#BSUB -nnodes 4
#BSUB -J DWF
export OMP_NUM_THREADS=6
export PAMI_IBV_ADAPTER_AFFINITY=1
export PAMI_ENABLE_STRIPING=1
export OPT="--comms-concurrent --comms-overlap "
#export GRID_ALLOC_NCACHE_LARGE=1
export APP="./benchmarks/Benchmark_comms_host_device --mpi 2.2.2.3 "
jsrun --nrs 4 -a6 -g6 -c42 -dpacked -b packed:7 --latency_priority gpu-cpu --smpiargs=-gpu $APP > comms.4node
APP="./benchmarks/Benchmark_dwf_fp32 --grid 48.48.48.72 --mpi 2.2.2.3 --shm 2048 --shm-force-mpi 1 --device-mem 8000 --shm-force-mpi 1 $OPT "
jsrun --nrs 4 -a6 -g6 -c42 -dpacked -b packed:7 --latency_priority gpu-cpu --smpiargs=-gpu $APP > dwf.24.4node
APP="./benchmarks/Benchmark_dwf_fp32 --grid 64.64.64.96 --mpi 2.2.2.3 --shm 2048 --shm-force-mpi 1 --device-mem 8000 --shm-force-mpi 1 $OPT "
jsrun --nrs 4 -a6 -g6 -c42 -dpacked -b packed:7 --latency_priority gpu-cpu --smpiargs=-gpu $APP > dwf.32.4node

8
systems/Summit/sourceme-cuda10.sh Normal file
View File

@ -0,0 +1,8 @@
export UCX_GDR_COPY_RCACHE=no
export UCX_MEMTYPE_CACHE=n
export UCX_RNDV_SCHEME=put_zcopy
module load gcc/7.5.0
module load cuda/10.2.89
#cuda/11.4.0
export LD_LIBRARY_PATH=/ccs/home/paboyle/prefix/lib/:$LD_LIBRARY_PATH

2
systems/Tursa/config-command
View File

@ -5,7 +5,7 @@
--enable-gen-simd-width=64 \
--enable-accelerator=cuda \
--with-lime=/mnt/lustre/tursafs1/home/tc002/tc002/dc-boyl1/spack/spack/opt/spack/linux-rhel8-zen/gcc-8.4.1/c-lime-2-3-9-e6wxqrid6rqmd45z7n32dxkvkykpvyez \
--disable-accelerator-cshift \
--enable-accelerator-cshift \
--disable-unified \
CXX=nvcc \
LDFLAGS="-cudart shared " \

8
systems/Tursa/sourceme.sh
View File

@ -1,2 +1,6 @@
spack load c-lime
module load cuda/11.4.1 openmpi/4.1.1 ucx/1.10.1
module load cuda/11.4.1 openmpi/4.1.1-cuda11.4.1 ucx/1.12.0-cuda11.4.1
#module load cuda/11.4.1 openmpi/4.1.1 ucx/1.10.1
export PREFIX=/home/tc002/tc002/shared/env/prefix/
export LD_LIBRARY_PATH=$PREFIX/lib/:$LD_LIBRARY_PATH
unset SBATCH_EXPORT

226
tests/core/Test_compact_wilson_clover_speedup.cc Normal file
View File

@ -0,0 +1,226 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./tests/core/Test_compact_wilson_clover_speedup.cc
Copyright (C) 2020 - 2022
Author: Daniel Richtmann <daniel.richtmann@gmail.com>
Author: Nils Meyer <nils.meyer@ur.de>
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 Grid;
NAMESPACE_BEGIN(CommandlineHelpers);
static bool checkPresent(int* argc, char*** argv, const std::string& option) {
return GridCmdOptionExists(*argv, *argv + *argc, option);
}
static std::string getContent(int* argc, char*** argv, const std::string& option) {
return GridCmdOptionPayload(*argv, *argv + *argc, option);
}
static int readInt(int* argc, char*** argv, std::string&& option, int defaultValue) {
std::string arg;
int ret = defaultValue;
if(checkPresent(argc, argv, option)) {
arg = getContent(argc, argv, option);
GridCmdOptionInt(arg, ret);
}
return ret;
}
static float readFloat(int* argc, char*** argv, std::string&& option, float defaultValue) {
std::string arg;
float ret = defaultValue;
if(checkPresent(argc, argv, option)) {
arg = getContent(argc, argv, option);
GridCmdOptionFloat(arg, ret);
}
return ret;
}
NAMESPACE_END(CommandlineHelpers);
#define _grid_printf(LOGGER, ...) \
{ \
if((LOGGER).isActive()) { /* this makes it safe to put, e.g., norm2 in the calling code w.r.t. performance */ \
char _printf_buf[1024]; \
std::sprintf(_printf_buf, __VA_ARGS__); \
std::cout << (LOGGER) << _printf_buf; \
fflush(stdout); \
} \
}
#define grid_printf_msg(...) _grid_printf(GridLogMessage, __VA_ARGS__)
template<typename Field>
bool resultsAgree(const Field& ref, const Field& res, const std::string& name) {
RealD checkTolerance = (getPrecision<Field>::value == 2) ? 1e-15 : 1e-7;
Field diff(ref.Grid());
diff = ref - res;
auto absDev = norm2(diff);
auto relDev = absDev / norm2(ref);
std::cout << GridLogMessage
<< "norm2(reference), norm2(" << name << "), abs. deviation, rel. deviation: " << norm2(ref) << " "
<< norm2(res) << " " << absDev << " " << relDev << " -> check "
<< ((relDev < checkTolerance) ? "passed" : "failed") << std::endl;
return relDev <= checkTolerance;
}
template<typename vCoeff_t>
void runBenchmark(int* argc, char*** argv) {
// read from command line
const int nIter = CommandlineHelpers::readInt( argc, argv, "--niter", 1000);
const RealD mass = CommandlineHelpers::readFloat( argc, argv, "--mass", 0.5);
const RealD csw = CommandlineHelpers::readFloat( argc, argv, "--csw", 1.0);
const RealD cF = CommandlineHelpers::readFloat( argc, argv, "--cF", 1.0);
const bool antiPeriodic = CommandlineHelpers::checkPresent(argc, argv, "--antiperiodic");
// precision
static_assert(getPrecision<vCoeff_t>::value == 2 || getPrecision<vCoeff_t>::value == 1, "Incorrect precision"); // double or single
std::string precision = (getPrecision<vCoeff_t>::value == 2 ? "double" : "single");
// setup grids
GridCartesian* UGrid = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd, vCoeff_t::Nsimd()), GridDefaultMpi());
GridRedBlackCartesian* UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
// clang-format on
// setup rng
std::vector<int> seeds({1, 2, 3, 4});
GridParallelRNG pRNG(UGrid);
pRNG.SeedFixedIntegers(seeds);
// type definitions
typedef WilsonImpl<vCoeff_t, FundamentalRepresentation, CoeffReal> WImpl;
typedef WilsonCloverFermion<WImpl> WilsonCloverOperator;
typedef CompactWilsonCloverFermion<WImpl> CompactWilsonCloverOperator;
typedef typename WilsonCloverOperator::FermionField Fermion;
typedef typename WilsonCloverOperator::GaugeField Gauge;
// setup fields
Fermion src(UGrid); random(pRNG, src);
Fermion ref(UGrid); ref = Zero();
Fermion res(UGrid); res = Zero();
Fermion hop(UGrid); hop = Zero();
Fermion diff(UGrid); diff = Zero();
Gauge Umu(UGrid); SU3::HotConfiguration(pRNG, Umu);
// setup boundary phases
typename WilsonCloverOperator::ImplParams implParams;
std::vector<Complex> boundary_phases(Nd, 1.);
if(antiPeriodic) boundary_phases[Nd-1] = -1.;
implParams.boundary_phases = boundary_phases;
WilsonAnisotropyCoefficients anisParams;
// misc stuff needed for benchmarks
double volume=1.0; for(int mu=0; mu<Nd; mu++) volume*=UGrid->_fdimensions[mu];
// setup fermion operators
WilsonCloverOperator Dwc( Umu, *UGrid, *UrbGrid, mass, csw, csw, anisParams, implParams);
CompactWilsonCloverOperator Dwc_compact(Umu, *UGrid, *UrbGrid, mass, csw, csw, cF, anisParams, implParams);
// now test the conversions
typename CompactWilsonCloverOperator::CloverField tmp_ref(UGrid); tmp_ref = Dwc.CloverTerm;
typename CompactWilsonCloverOperator::CloverField tmp_res(UGrid); tmp_res = Zero();
typename CompactWilsonCloverOperator::CloverField tmp_diff(UGrid); tmp_diff = Zero();
typename CompactWilsonCloverOperator::CloverDiagonalField diagonal(UGrid); diagonal = Zero();
typename CompactWilsonCloverOperator::CloverTriangleField triangle(UGrid); diagonal = Zero();
CompactWilsonCloverOperator::CompactHelpers::ConvertLayout(tmp_ref, diagonal, triangle);
CompactWilsonCloverOperator::CompactHelpers::ConvertLayout(diagonal, triangle, tmp_res);
tmp_diff = tmp_ref - tmp_res;
std::cout << GridLogMessage << "conversion: ref, res, diff, eps"
<< " " << norm2(tmp_ref)
<< " " << norm2(tmp_res)
<< " " << norm2(tmp_diff)
<< " " << norm2(tmp_diff) / norm2(tmp_ref)
<< std::endl;
// performance per site (use minimal values necessary)
double hop_flop_per_site = 1320; // Rich's Talk + what Peter uses
double hop_byte_per_site = (8 * 9 + 9 * 12) * 2 * getPrecision<vCoeff_t>::value * 4;
double clov_flop_per_site = 504; // Rich's Talk and 1412.2629
double clov_byte_per_site = (2 * 18 + 12 + 12) * 2 * getPrecision<vCoeff_t>::value * 4;
double clov_flop_per_site_performed = 1128;
double clov_byte_per_site_performed = (12 * 12 + 12 + 12) * 2 * getPrecision<vCoeff_t>::value * 4;
// total performance numbers
double hop_gflop_total = volume * nIter * hop_flop_per_site / 1e9;
double hop_gbyte_total = volume * nIter * hop_byte_per_site / 1e9;
double clov_gflop_total = volume * nIter * clov_flop_per_site / 1e9;
double clov_gbyte_total = volume * nIter * clov_byte_per_site / 1e9;
double clov_gflop_performed_total = volume * nIter * clov_flop_per_site_performed / 1e9;
double clov_gbyte_performed_total = volume * nIter * clov_byte_per_site_performed / 1e9;
// warmup + measure dhop
for(auto n : {1, 2, 3, 4, 5}) Dwc.Dhop(src, hop, 0);
double t0 = usecond();
for(int n = 0; n < nIter; n++) Dwc.Dhop(src, hop, 0);
double t1 = usecond();
double secs_hop = (t1-t0)/1e6;
grid_printf_msg("Performance(%35s, %s): %2.4f s, %6.0f GFlop/s, %6.0f GByte/s, speedup vs ref = %.2f, fraction of hop = %.2f\n",
"hop", precision.c_str(), secs_hop, hop_gflop_total/secs_hop, hop_gbyte_total/secs_hop, 0.0, secs_hop/secs_hop);
#define BENCH_CLOVER_KERNEL(KERNEL) { \
/* warmup + measure reference clover */ \
for(auto n : {1, 2, 3, 4, 5}) Dwc.KERNEL(src, ref); \
double t2 = usecond(); \
for(int n = 0; n < nIter; n++) Dwc.KERNEL(src, ref); \
double t3 = usecond(); \
double secs_ref = (t3-t2)/1e6; \
grid_printf_msg("Performance(%35s, %s): %2.4f s, %6.0f GFlop/s, %6.0f GByte/s, speedup vs ref = %.2f, fraction of hop = %.2f\n", \
"reference_"#KERNEL, precision.c_str(), secs_ref, clov_gflop_total/secs_ref, clov_gbyte_total/secs_ref, secs_ref/secs_ref, secs_ref/secs_hop); \
grid_printf_msg("Performance(%35s, %s): %2.4f s, %6.0f GFlop/s, %6.0f GByte/s, speedup vs ref = %.2f, fraction of hop = %.2f\n", /* to see how well the ET performs */ \
"reference_"#KERNEL"_performed", precision.c_str(), secs_ref, clov_gflop_performed_total/secs_ref, clov_gbyte_performed_total/secs_ref, secs_ref/secs_ref, secs_ref/secs_hop); \
\
/* warmup + measure compact clover */ \
for(auto n : {1, 2, 3, 4, 5}) Dwc_compact.KERNEL(src, res); \
double t4 = usecond(); \
for(int n = 0; n < nIter; n++) Dwc_compact.KERNEL(src, res); \
double t5 = usecond(); \
double secs_res = (t5-t4)/1e6; \
grid_printf_msg("Performance(%35s, %s): %2.4f s, %6.0f GFlop/s, %6.0f GByte/s, speedup vs ref = %.2f, fraction of hop = %.2f\n", \
"compact_"#KERNEL, precision.c_str(), secs_res, clov_gflop_total/secs_res, clov_gbyte_total/secs_res, secs_ref/secs_res, secs_res/secs_hop); \
assert(resultsAgree(ref, res, #KERNEL)); \
}
BENCH_CLOVER_KERNEL(Mooee);
BENCH_CLOVER_KERNEL(MooeeDag);
BENCH_CLOVER_KERNEL(MooeeInv);
BENCH_CLOVER_KERNEL(MooeeInvDag);
grid_printf_msg("finalize %s\n", precision.c_str());
}
int main(int argc, char** argv) {
Grid_init(&argc, &argv);
runBenchmark<vComplexD>(&argc, &argv);
runBenchmark<vComplexF>(&argc, &argv);
Grid_finalize();
}

1
tests/core/Test_contfrac_even_odd.cc
View File

@ -235,7 +235,6 @@ void TestWhat(What & Ddwf,
pickCheckerboard(Odd ,chi_o,chi);
pickCheckerboard(Even,phi_e,phi);
pickCheckerboard(Odd ,phi_o,phi);
RealD t1,t2;
SchurDiagMooeeOperator<What,LatticeFermion> HermOpEO(Ddwf);
HermOpEO.MpcDagMpc(chi_e,dchi_e);

1
tests/core/Test_dwf_eofa_even_odd.cc
View File

@ -215,7 +215,6 @@ int main (int argc, char ** argv)
pickCheckerboard(Odd , chi_o, chi);
pickCheckerboard(Even, phi_e, phi);
pickCheckerboard(Odd , phi_o, phi);
RealD t1,t2;
SchurDiagMooeeOperator<DomainWallEOFAFermionR,LatticeFermion> HermOpEO(Ddwf);
HermOpEO.MpcDagMpc(chi_e, dchi_e);

2
tests/core/Test_dwf_even_odd.cc
View File

@ -212,8 +212,6 @@ int main (int argc, char ** argv)
pickCheckerboard(Odd ,chi_o,chi);
pickCheckerboard(Even,phi_e,phi);
pickCheckerboard(Odd ,phi_o,phi);
RealD t1,t2;
SchurDiagMooeeOperator<DomainWallFermionR,LatticeFermion> HermOpEO(Ddwf);
HermOpEO.MpcDagMpc(chi_e,dchi_e);

4
tests/core/Test_gamma.cc
View File

@ -181,8 +181,8 @@ void checkAdj(const Gamma::Algebra a)
void checkProject(GridSerialRNG &rng)
{
SpinVector rv, recon, full;
HalfSpinVector hsp, hsm;
SpinVector rv, recon;
HalfSpinVector hsm;
random(rng, rv);

1
tests/core/Test_gpwilson_even_odd.cc
View File

@ -198,7 +198,6 @@ int main (int argc, char ** argv)
pickCheckerboard(Odd ,chi_o,chi);
pickCheckerboard(Even,phi_e,phi);
pickCheckerboard(Odd ,phi_o,phi);
RealD t1,t2;
SchurDiagMooeeOperator<GparityWilsonFermionR,FermionField> HermOpEO(Dw);
HermOpEO.MpcDagMpc(chi_e,dchi_e);

12
tests/core/Test_main.cc
View File

@ -364,14 +364,12 @@ int main(int argc, char **argv) {
{ // Peek-ology and Poke-ology, with a little app-ology
Complex c;
ColourMatrix c_m;
SpinMatrix s_m;
SpinColourMatrix sc_m;
ColourMatrix c_m = Zero();
SpinMatrix s_m = Zero();
SpinColourMatrix sc_m = Zero();
s_m = TensorIndexRecursion<ColourIndex>::traceIndex(
sc_m); // Map to traceColour
c_m = TensorIndexRecursion<SpinIndex>::traceIndex(
sc_m); // map to traceSpin
s_m = TensorIndexRecursion<ColourIndex>::traceIndex(sc_m); // Map to traceColour
c_m = TensorIndexRecursion<SpinIndex>::traceIndex(sc_m); // map to traceSpin
c = TensorIndexRecursion<SpinIndex>::traceIndex(s_m);
c = TensorIndexRecursion<ColourIndex>::traceIndex(c_m);

1
tests/core/Test_mobius_eofa_even_odd.cc
View File

@ -217,7 +217,6 @@ int main (int argc, char ** argv)
pickCheckerboard(Odd , chi_o, chi);
pickCheckerboard(Even, phi_e, phi);
pickCheckerboard(Odd , phi_o, phi);
RealD t1,t2;
SchurDiagMooeeOperator<MobiusEOFAFermionR,LatticeFermion> HermOpEO(Ddwf);
HermOpEO.MpcDagMpc(chi_e, dchi_e);

1
tests/core/Test_mobius_even_odd.cc
View File

@ -262,7 +262,6 @@ int main (int argc, char ** argv)
pickCheckerboard(Odd ,chi_o,chi);
pickCheckerboard(Even,phi_e,phi);
pickCheckerboard(Odd ,phi_o,phi);
RealD t1,t2;
SchurDiagMooeeOperator<MobiusFermionR,LatticeFermion> HermOpEO(Ddwf);

2
tests/core/Test_staggered.cc
View File

@ -144,7 +144,7 @@ int main (int argc, char ** argv)
Ds.Dhop(src,result,0);
}
double t1=usecond();
double t2;
double flops=(16*(3*(6+8+8)) + 15*3*2)*volume*ncall; // == 66*16 + == 1146
std::cout<<GridLogMessage << "Called Ds"<<std::endl;

1
tests/core/Test_staggered5D.cc
View File

@ -162,7 +162,6 @@ int main (int argc, char ** argv)
}
double t1=usecond();
double t2;
double flops=(16*(3*(6+8+8)) + 15*3*2)*volume*ncall; // == 66*16 + == 1146
std::cout<<GridLogMessage << "Called Ds"<<std::endl;

2
tests/core/Test_staggered_naive.cc
View File

@ -30,7 +30,6 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
using namespace std;
using namespace Grid;
;
int main (int argc, char ** argv)
{
@ -135,7 +134,6 @@ int main (int argc, char ** argv)
Ds.Dhop(src,result,0);
}
double t1=usecond();
double t2;
double flops=(16*(3*(6+8+8)) + 15*3*2)*volume*ncall; // == 66*16 + == 1146
std::cout<<GridLogMessage << "Called Ds"<<std::endl;

1
tests/core/Test_wilson_even_odd.cc
View File

@ -204,7 +204,6 @@ int main (int argc, char ** argv)
pickCheckerboard(Odd ,chi_o,chi);
pickCheckerboard(Even,phi_e,phi);
pickCheckerboard(Odd ,phi_o,phi);
RealD t1,t2;
SchurDiagMooeeOperator<WilsonFermionR,LatticeFermion> HermOpEO(Dw);
HermOpEO.MpcDagMpc(chi_e,dchi_e);

1
tests/core/Test_wilson_twisted_mass_even_odd.cc
View File

@ -205,7 +205,6 @@ int main (int argc, char ** argv)
pickCheckerboard(Odd ,chi_o,chi);
pickCheckerboard(Even,phi_e,phi);
pickCheckerboard(Odd ,phi_o,phi);
RealD t1,t2;
SchurDiagMooeeOperator<WilsonTMFermionR,LatticeFermion> HermOpEO(Dw);
HermOpEO.MpcDagMpc(chi_e,dchi_e);

1
tests/core/Test_zmobius_even_odd.cc
View File

@ -276,7 +276,6 @@ int main (int argc, char ** argv)
pickCheckerboard(Odd ,chi_o,chi);
pickCheckerboard(Even,phi_e,phi);
pickCheckerboard(Odd ,phi_o,phi);
RealD t1,t2;
SchurDiagMooeeOperator<ZMobiusFermionR,LatticeFermion> HermOpEO(Ddwf);

1
tests/forces/Test_rect_force.cc
View File

@ -57,7 +57,6 @@ int main (int argc, char ** argv)
SU<Nc>::HotConfiguration(pRNG,U);
double beta = 1.0;
double c1 = -0.331;
IwasakiGaugeActionR Action(beta);
// PlaqPlusRectangleActionR Action(beta,c1);

3
tests/lanczos/Test_dwf_compressed_lanczos_reorg_synthetic.cc
View File

@ -40,6 +40,7 @@ using namespace Grid;
template<class Fobj,class CComplex,int nbasis>
class ProjectedHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > {
public:
using LinearFunction<Lattice<iVector<CComplex,nbasis > > >::operator();
typedef iVector<CComplex,nbasis > CoarseSiteVector;
typedef Lattice<CoarseSiteVector> CoarseField;
typedef Lattice<CComplex> CoarseScalar; // used for inner products on fine field
@ -67,6 +68,8 @@ public:
template<class Fobj,class CComplex,int nbasis>
class ProjectedFunctionHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > {
public:
using LinearFunction<Lattice<iVector<CComplex,nbasis > > >::operator ();
typedef iVector<CComplex,nbasis > CoarseSiteVector;
typedef Lattice<CoarseSiteVector> CoarseField;
typedef Lattice<CComplex> CoarseScalar; // used for inner products on fine field

3
tests/solver/Test_dwf_hdcr.cc
View File

@ -55,6 +55,7 @@ RealD InverseApproximation(RealD x){
template<class Field,class Matrix> class ChebyshevSmoother : public LinearFunction<Field>
{
public:
using LinearFunction<Field>::operator();
typedef LinearOperatorBase<Field> FineOperator;
Matrix & _SmootherMatrix;
FineOperator & _SmootherOperator;
@ -78,6 +79,7 @@ public:
template<class Field,class Matrix> class MirsSmoother : public LinearFunction<Field>
{
public:
using LinearFunction<Field>::operator();
typedef LinearOperatorBase<Field> FineOperator;
Matrix & SmootherMatrix;
FineOperator & SmootherOperator;
@ -108,6 +110,7 @@ public:
template<class Fobj,class CComplex,int nbasis, class Matrix, class Guesser, class CoarseSolver>
class MultiGridPreconditioner : public LinearFunction< Lattice<Fobj> > {
public:
using LinearFunction<Lattice<Fobj> >::operator();
typedef Aggregation<Fobj,CComplex,nbasis> Aggregates;
typedef CoarsenedMatrix<Fobj,CComplex,nbasis> CoarseOperator;

9
tests/solver/Test_dwf_hdcr_16_rb.cc
View File

@ -56,9 +56,9 @@ template<class Field> class SolverWrapper : public LinearFunction<Field> {
private:
CheckerBoardedSparseMatrixBase<Field> & _Matrix;
SchurRedBlackBase<Field> & _Solver;
public:
/////////////////////////////////////////////////////
public:
using LinearFunction<Field>::operator();
/////////////////////////////////////////////////////
// Wrap the usual normal equations trick
/////////////////////////////////////////////////////
SolverWrapper(CheckerBoardedSparseMatrixBase<Field> &Matrix,
@ -75,6 +75,7 @@ public:
template<class Field,class Matrix> class ChebyshevSmoother : public LinearFunction<Field>
{
public:
using LinearFunction<Field>::operator();
typedef LinearOperatorBase<Field> FineOperator;
Matrix & _SmootherMatrix;
FineOperator & _SmootherOperator;
@ -98,6 +99,7 @@ public:
template<class Field,class Matrix> class MirsSmoother : public LinearFunction<Field>
{
public:
using LinearFunction<Field>::operator();
typedef LinearOperatorBase<Field> FineOperator;
Matrix & SmootherMatrix;
FineOperator & SmootherOperator;
@ -128,6 +130,7 @@ public:
template<class Fobj,class CComplex,int nbasis, class Matrix, class Guesser, class CoarseSolver>
class MultiGridPreconditioner : public LinearFunction< Lattice<Fobj> > {
public:
using LinearFunction<Lattice<Fobj> >::operator();
typedef Aggregation<Fobj,CComplex,nbasis> Aggregates;
typedef CoarsenedMatrix<Fobj,CComplex,nbasis> CoarseOperator;

4
tests/solver/Test_dwf_hdcr_24_regression.cc
View File

@ -55,6 +55,7 @@ RealD InverseApproximation(RealD x){
template<class Field,class Matrix> class ChebyshevSmoother : public LinearFunction<Field>
{
public:
using LinearFunction<Field>::operator();
typedef LinearOperatorBase<Field> FineOperator;
Matrix & _SmootherMatrix;
FineOperator & _SmootherOperator;
@ -78,6 +79,7 @@ public:
template<class Field,class Matrix> class MirsSmoother : public LinearFunction<Field>
{
public:
using LinearFunction<Field>::operator();
typedef LinearOperatorBase<Field> FineOperator;
Matrix & SmootherMatrix;
FineOperator & SmootherOperator;
@ -108,6 +110,8 @@ public:
template<class Fobj,class CComplex,int nbasis, class Matrix, class Guesser, class CoarseSolver>
class MultiGridPreconditioner : public LinearFunction< Lattice<Fobj> > {
public:
using LinearFunction<Lattice<Fobj> >::operator();
typedef Aggregation<Fobj,CComplex,nbasis> Aggregates;
typedef CoarsenedMatrix<Fobj,CComplex,nbasis> CoarseOperator;

8
tests/solver/Test_dwf_hdcr_2level.cc
View File

@ -56,6 +56,7 @@ RealD InverseApproximation(RealD x){
template<class Field,class Matrix> class ChebyshevSmoother : public LinearFunction<Field>
{
public:
using LinearFunction<Field>::operator();
typedef LinearOperatorBase<Field> FineOperator;
Matrix & _SmootherMatrix;
FineOperator & _SmootherOperator;
@ -79,6 +80,7 @@ public:
template<class Field,class Matrix> class MirsSmoother : public LinearFunction<Field>
{
public:
using LinearFunction<Field>::operator();
typedef LinearOperatorBase<Field> FineOperator;
Matrix & SmootherMatrix;
FineOperator & SmootherOperator;
@ -108,6 +110,7 @@ public:
template<class Field,class Matrix> class RedBlackSmoother : public LinearFunction<Field>
{
public:
using LinearFunction<Field>::operator();
typedef LinearOperatorBase<Field> FineOperator;
Matrix & SmootherMatrix;
RealD tol;
@ -134,6 +137,7 @@ public:
template<class Fobj,class CComplex,int nbasis, class Matrix, class Guesser, class CoarseSolver>
class MultiGridPreconditioner : public LinearFunction< Lattice<Fobj> > {
public:
using LinearFunction<Lattice<Fobj> >::operator();
typedef Aggregation<Fobj,CComplex,nbasis> Aggregates;
typedef CoarsenedMatrix<Fobj,CComplex,nbasis> CoarseOperator;
@ -241,7 +245,7 @@ int main (int argc, char ** argv)
Grid_init(&argc,&argv);
const int Ls=16;
const int rLs=8;
// const int rLs=8;
GridCartesian * UGrid = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
@ -388,7 +392,7 @@ int main (int argc, char ** argv)
// RedBlackSmoother<LatticeFermion,DomainWallFermionR> FineRBSmoother(0.00,0.001,100,Ddwf);
// Wrap the 2nd level solver in a MultiGrid preconditioner acting on the fine space
ZeroGuesser<CoarseVector> CoarseZeroGuesser;
// ZeroGuesser<CoarseVector> CoarseZeroGuesser;
TwoLevelMG TwoLevelPrecon(Aggregates, LDOp,
HermIndefOp,Ddwf,
FineSmoother,

5
tests/solver/Test_dwf_hdcr_48_rb.cc
View File

@ -57,7 +57,7 @@ private:
CheckerBoardedSparseMatrixBase<Field> & _Matrix;
SchurRedBlackBase<Field> & _Solver;
public:
using LinearFunction<Field>::operator();
/////////////////////////////////////////////////////
// Wrap the usual normal equations trick
/////////////////////////////////////////////////////
@ -75,6 +75,7 @@ public:
template<class Field,class Matrix> class ChebyshevSmoother : public LinearFunction<Field>
{
public:
using LinearFunction<Field>::operator();
typedef LinearOperatorBase<Field> FineOperator;
Matrix & _SmootherMatrix;
FineOperator & _SmootherOperator;
@ -98,6 +99,7 @@ public:
template<class Field,class Matrix> class MirsSmoother : public LinearFunction<Field>
{
public:
using LinearFunction<Field>::operator();
typedef LinearOperatorBase<Field> FineOperator;
Matrix & SmootherMatrix;
FineOperator & SmootherOperator;
@ -128,6 +130,7 @@ public:
template<class Fobj,class CComplex,int nbasis, class Matrix, class Guesser, class CoarseSolver>
class MultiGridPreconditioner : public LinearFunction< Lattice<Fobj> > {
public:
using LinearFunction<Lattice<Fobj> >::operator();
typedef Aggregation<Fobj,CComplex,nbasis> Aggregates;
typedef CoarsenedMatrix<Fobj,CComplex,nbasis> CoarseOperator;

3
tests/solver/Test_dwf_hdcr_48_regression.cc
View File

@ -55,6 +55,7 @@ RealD InverseApproximation(RealD x){
template<class Field,class Matrix> class ChebyshevSmoother : public LinearFunction<Field>
{
public:
using LinearFunction<Field>::operator();
typedef LinearOperatorBase<Field> FineOperator;
Matrix & _SmootherMatrix;
FineOperator & _SmootherOperator;
@ -78,6 +79,7 @@ public:
template<class Field,class Matrix> class MirsSmoother : public LinearFunction<Field>
{
public:
using LinearFunction<Field>::operator();
typedef LinearOperatorBase<Field> FineOperator;
Matrix & SmootherMatrix;
FineOperator & SmootherOperator;
@ -108,6 +110,7 @@ public:
template<class Fobj,class CComplex,int nbasis, class Matrix, class Guesser, class CoarseSolver>
class MultiGridPreconditioner : public LinearFunction< Lattice<Fobj> > {
public:
using LinearFunction<Lattice<Fobj> >::operator();
typedef Aggregation<Fobj,CComplex,nbasis> Aggregates;
typedef CoarsenedMatrix<Fobj,CComplex,nbasis> CoarseOperator;

3
tests/solver/Test_dwf_multigrid.cc
View File

@ -57,6 +57,7 @@ private:
OperatorFunction<Field> & _Solver;
LinearFunction<Field> & _Guess;
public:
using LinearFunction<Field>::operator();
/////////////////////////////////////////////////////
// Wrap the usual normal equations trick
@ -118,6 +119,7 @@ RealD InverseApproximation(RealD x){
template<class Field,class Matrix> class ChebyshevSmoother : public LinearFunction<Field>
{
public:
using LinearFunction<Field>::operator();
typedef LinearOperatorBase<Field> FineOperator;
Matrix & _SmootherMatrix;
FineOperator & _SmootherOperator;
@ -174,6 +176,7 @@ public:
template<class Fobj,class CComplex,int nbasis, class CoarseSolver>
class HDCRPreconditioner : public LinearFunction< Lattice<Fobj> > {
public:
using LinearFunction<Lattice<Fobj> >::operator();
typedef Aggregation<Fobj,CComplex,nbasis> Aggregates;
typedef CoarsenedMatrix<Fobj,CComplex,nbasis> CoarseOperator;

21
tests/solver/Test_hw_multigrid_mixed_48.cc
View File

@ -456,8 +456,8 @@ public:
siteVector *CBp=Stencil.CommBuf();
int ptype;
int nb2=nbasis/2;
// int ptype;
// int nb2=nbasis/2;
autoView(in_v , in, AcceleratorRead);
autoView(st, Stencil, AcceleratorRead);
@ -471,7 +471,7 @@ public:
typedef decltype(coalescedRead(in_v[0])) calcVector;
typedef decltype(coalescedRead(in_v[0](0))) calcComplex;
int sU = sF/Ls;
int s = sF%Ls;
// int s = sF%Ls;
calcComplex res = Zero();
calcVector nbr;
@ -517,14 +517,14 @@ public:
autoView(st, Stencil, AcceleratorRead);
siteVector *CBp=Stencil.CommBuf();
int ptype;
int nb2=nbasis/2;
// int ptype;
// int nb2=nbasis/2;
accelerator_for2d(sF, Coarse5D->oSites(), b, nbasis, Nsimd, {
typedef decltype(coalescedRead(in_v[0])) calcVector;
typedef decltype(coalescedRead(in_v[0](0))) calcComplex;
int sU = sF/Ls;
int s = sF%Ls;
// int s = sF%Ls;
calcComplex res = Zero();
@ -650,7 +650,7 @@ private:
OperatorFunction<Field> & _Solver;
LinearFunction<Field> & _Guess;
public:
using LinearFunction<Field>::operator();
/////////////////////////////////////////////////////
// Wrap the usual normal equations trick
/////////////////////////////////////////////////////
@ -712,6 +712,7 @@ RealD InverseApproximation(RealD x){
template<class Field,class Matrix> class ChebyshevSmoother : public LinearFunction<Field>
{
public:
using LinearFunction<Field>::operator();
typedef LinearOperatorBase<Field> FineOperator;
Matrix & _SmootherMatrix;
FineOperator & _SmootherOperator;
@ -735,6 +736,7 @@ public:
template<class Fobj,class CComplex,int nbasis, class CoarseSolver>
class MGPreconditioner : public LinearFunction< Lattice<Fobj> > {
public:
using LinearFunction<Lattice<Fobj> >::operator();
typedef Aggregation<Fobj,CComplex,nbasis> Aggregates;
typedef typename Aggregation<Fobj,CComplex,nbasis>::CoarseVector CoarseVector;
@ -831,6 +833,7 @@ public:
template<class Fobj,class CComplex,int nbasis, class CoarseSolver>
class HDCRPreconditioner : public LinearFunction< Lattice<Fobj> > {
public:
using LinearFunction<Lattice<Fobj> >::operator();
typedef Aggregation<Fobj,CComplex,nbasis> Aggregates;
typedef typename Aggregation<Fobj,CComplex,nbasis>::CoarseVector CoarseVector;
@ -1174,18 +1177,18 @@ int main (int argc, char ** argv)
PlainHermOp<CoarseCoarseVector> IRLOpL2 (IRLHermOpL2);
ImplicitlyRestartedLanczos<CoarseCoarseVector> IRLL2(IRLOpChebyL2,IRLOpL2,cNstop,cNk,cNm,1.0e-3,20);
int cNconv;
cNm=0;
std::vector<RealD> eval2(cNm);
std::vector<CoarseCoarseVector> evec2(cNm,CoarseCoarse5d);
cc_src=1.0;
// int cNconv;
// IRLL2.calc(eval2,evec2,cc_src,cNconv);
ConjugateGradient<CoarseCoarseVector> CoarseCoarseCG(0.02,10000);
DeflatedGuesser<CoarseCoarseVector> DeflCoarseCoarseGuesser(evec2,eval2);
NormalEquations<CoarseCoarseVector> DeflCoarseCoarseCGNE(cc_Dwf,CoarseCoarseCG,DeflCoarseCoarseGuesser);
ZeroGuesser<CoarseVector> CoarseZeroGuesser;
// ZeroGuesser<CoarseVector> CoarseZeroGuesser;
ZeroGuesser<CoarseCoarseVector> CoarseCoarseZeroGuesser;
std::cout<<GridLogMessage << "**************************************************"<< std::endl;

28
tests/solver/Test_hw_multigrid_mixed_48_rb.cc
View File

@ -456,8 +456,8 @@ public:
siteVector *CBp=Stencil.CommBuf();
int ptype;
int nb2=nbasis/2;
//int ptype;
// int nb2=nbasis/2;
autoView(in_v , in, AcceleratorRead);
autoView(st, Stencil, AcceleratorRead);
@ -471,7 +471,7 @@ public:
typedef decltype(coalescedRead(in_v[0])) calcVector;
typedef decltype(coalescedRead(in_v[0](0))) calcComplex;
int sU = sF/Ls;
int s = sF%Ls;
// int s = sF%Ls;
calcComplex res = Zero();
calcVector nbr;
@ -517,14 +517,14 @@ public:
autoView(st, Stencil, AcceleratorRead);
siteVector *CBp=Stencil.CommBuf();
int ptype;
int nb2=nbasis/2;
// int ptype;
// int nb2=nbasis/2;
accelerator_for2d(sF, Coarse5D->oSites(), b, nbasis, Nsimd, {
typedef decltype(coalescedRead(in_v[0])) calcVector;
typedef decltype(coalescedRead(in_v[0](0))) calcComplex;
int sU = sF/Ls;
int s = sF%Ls;
// int s = sF%Ls;
calcComplex res = Zero();
@ -648,7 +648,7 @@ private:
CheckerBoardedSparseMatrixBase<Field> & _Matrix;
SchurRedBlackBase<Field> & _Solver;
public:
using LinearFunction<Field>::operator();
/////////////////////////////////////////////////////
// Wrap the usual normal equations trick
/////////////////////////////////////////////////////
@ -669,6 +669,7 @@ private:
OperatorFunction<Field> & _Solver;
LinearFunction<Field> & _Guess;
public:
using LinearFunction<Field>::operator();
/////////////////////////////////////////////////////
// Wrap the usual normal equations trick
@ -731,6 +732,7 @@ RealD InverseApproximation(RealD x){
template<class Field,class Matrix> class ChebyshevSmoother : public LinearFunction<Field>
{
public:
using LinearFunction<Field>::operator();
typedef LinearOperatorBase<Field> FineOperator;
Matrix & _SmootherMatrix;
FineOperator & _SmootherOperator;
@ -754,6 +756,7 @@ public:
template<class Fobj,class CComplex,int nbasis, class CoarseSolver>
class MGPreconditioner : public LinearFunction< Lattice<Fobj> > {
public:
using LinearFunction<Lattice<Fobj> >::operator();
typedef Aggregation<Fobj,CComplex,nbasis> Aggregates;
typedef typename Aggregation<Fobj,CComplex,nbasis>::CoarseVector CoarseVector;
@ -850,7 +853,8 @@ public:
template<class Fobj,class CComplex,int nbasis, class CoarseSolver>
class HDCRPreconditioner : public LinearFunction< Lattice<Fobj> > {
public:
using LinearFunction<Lattice<Fobj> >::operator();
typedef Aggregation<Fobj,CComplex,nbasis> Aggregates;
typedef typename Aggregation<Fobj,CComplex,nbasis>::CoarseVector CoarseVector;
typedef typename Aggregation<Fobj,CComplex,nbasis>::CoarseMatrix CoarseMatrix;
@ -1194,11 +1198,11 @@ int main (int argc, char ** argv)
PlainHermOp<CoarseCoarseVector> IRLOpL2 (IRLHermOpL2);
ImplicitlyRestartedLanczos<CoarseCoarseVector> IRLL2(IRLOpChebyL2,IRLOpL2,cNstop,cNk,cNm,1.0e-3,20);
int cNconv;
cNm=0;
std::vector<RealD> eval2(cNm);
std::vector<CoarseCoarseVector> evec2(cNm,CoarseCoarse5d);
cc_src=1.0;
// int cNconv;
// IRLL2.calc(eval2,evec2,cc_src,cNconv);
std::vector<RealD> tols ({0.005,0.001});
@ -1218,10 +1222,10 @@ int main (int argc, char ** argv)
for(auto c_hi : c_his ) {
for(auto f_lo : f_los ) {
for(auto f_hi : f_his ) {
ZeroGuesser<CoarseVector> CoarseZeroGuesser;
ZeroGuesser<CoarseCoarseVector> CoarseCoarseZeroGuesser;
// ZeroGuesser<CoarseVector> CoarseZeroGuesser;
// ZeroGuesser<CoarseCoarseVector> CoarseCoarseZeroGuesser;
ConjugateGradient<CoarseCoarseVector> CoarseCoarseCG(tol,10000);
ZeroGuesser<CoarseCoarseVector> CoarseCoarseGuesser;
// ZeroGuesser<CoarseCoarseVector> CoarseCoarseGuesser;
SchurRedBlackDiagMooeeSolve<CoarseCoarseVector> CoarseCoarseRBCG(CoarseCoarseCG);
SchurSolverWrapper<CoarseCoarseVector> CoarseCoarseSolver(cc_Dwf,CoarseCoarseRBCG);

3
tests/solver/Test_multigrid_common.h
View File

@ -143,6 +143,7 @@ public:
template<class Field> class MultiGridPreconditionerBase : public LinearFunction<Field> {
public:
using LinearFunction<Field>::operator();
virtual ~MultiGridPreconditionerBase() = default;
virtual void setup() = 0;
virtual void operator()(Field const &in, Field &out) = 0;
@ -156,6 +157,7 @@ public:
/////////////////////////////////////////////
// Type Definitions
/////////////////////////////////////////////
using MultiGridPreconditionerBase<Lattice<Fobj>>::operator();
// clang-format off
typedef Aggregation<Fobj, CComplex, nBasis> Aggregates;
@ -568,6 +570,7 @@ public:
/////////////////////////////////////////////
// Type Definitions
/////////////////////////////////////////////
using MultiGridPreconditionerBase<Lattice<Fobj>>::operator();
typedef Matrix FineDiracMatrix;
typedef Lattice<Fobj> FineVector;

1
tests/solver/Test_wilson_qmr_unprec.cc
View File

@ -56,7 +56,6 @@ int main (int argc, char ** argv)
QuasiMinimalResidual<LatticeFermion> QMR(1.0e-8,10000);
RealD mass=0.0;
RealD M5=1.8;
WilsonFermionR Dw(Umu,*Grid,*rbGrid,mass);
NonHermitianLinearOperator<WilsonFermionR,LatticeFermion> NonHermOp(Dw);