portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-17 23:37:06 +01:00
Code Releases Activity

Merge branch 'develop' of https://github.com/paboyle/Grid into develop

Browse Source
This commit is contained in:
James Richings
2021-10-26 10:45:46 +01:00
parent c144b32368 ba7e371b90
commit fd13a3f2be
31 changed files with 1083 additions and 236 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
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(){};
};

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/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;

5
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; };
};
@ -57,6 +60,7 @@ private:
const unsigned int N;
public:
using LinearFunction<Field>::operator();
DeflatedGuesser(const std::vector<Field> & _evec,const std::vector<RealD> & _eval)
: DeflatedGuesser(_evec, _eval, _evec.size())
@ -101,6 +105,7 @@ 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)

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;
}
}

5
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:

3
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;

2
Grid/communicator/Communicator_mpi3.cc
View File

@ -389,7 +389,6 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
void *shm = (void *) this->ShmBufferTranslate(dest,recv);
assert(shm!=NULL);
acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
acceleratorCopySynchronise(); // MPI prob slower
}
if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
@ -405,6 +404,7 @@ void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsReque
if (nreq==0) return;
std::vector<MPI_Status> status(nreq);
acceleratorCopySynchronise();
int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
assert(ierr==0);
list.resize(0);

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

1
Grid/parallelIO/IldgIO.h
View File

@ -576,6 +576,7 @@ 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()) ) ) {
skipPastObjectRecord(std::string(GRID_FIELD_NORM));
skipPastObjectRecord(std::string(SCIDAC_CHECKSUM));
return;
}

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);

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;

14
Grid/threads/Accelerator.cc
View File

@ -74,11 +74,13 @@ 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");
@ -87,10 +89,20 @@ void acceleratorInit(void)
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

9
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(); \