portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-10 07:55:35 +00:00
Code Releases Activity

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

Browse Source
This commit is contained in:
Peter Boyle 2020-09-03 20:30:49 -04:00
commit 6dbd117aa5
11 changed files with 119 additions and 173 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

57
Grid/allocator/AlignedAllocator.h
View File

@ -65,8 +65,7 @@ public:
MemoryManager::CpuFree((void *)__p,bytes);
}
// FIXME: hack for the copy constructor, eventually it must be avoided
//void construct(pointer __p, const _Tp& __val) { new((void *)__p) _Tp(__val); };
// FIXME: hack for the copy constructor: it must be avoided to avoid single thread loop
void construct(pointer __p, const _Tp& __val) { assert(0);};
void construct(pointer __p) { };
void destroy(pointer __p) { };
@ -74,6 +73,9 @@ public:
template<typename _Tp> inline bool operator==(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return true; }
template<typename _Tp> inline bool operator!=(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return false; }
//////////////////////////////////////////////////////////////////////////////////////
// Unified virtual memory
//////////////////////////////////////////////////////////////////////////////////////
template<typename _Tp>
class uvmAllocator {
public:
@ -109,22 +111,63 @@ public:
MemoryManager::SharedFree((void *)__p,bytes);
}
// FIXME: hack for the copy constructor, eventually it must be avoided
void construct(pointer __p, const _Tp& __val) { new((void *)__p) _Tp(__val); };
//void construct(pointer __p, const _Tp& __val) { };
void construct(pointer __p) { };
void destroy(pointer __p) { };
};
template<typename _Tp> inline bool operator==(const uvmAllocator<_Tp>&, const uvmAllocator<_Tp>&){ return true; }
template<typename _Tp> inline bool operator!=(const uvmAllocator<_Tp>&, const uvmAllocator<_Tp>&){ return false; }
////////////////////////////////////////////////////////////////////////////////
// Device memory
////////////////////////////////////////////////////////////////////////////////
template<typename _Tp>
class devAllocator {
public:
typedef std::size_t size_type;
typedef std::ptrdiff_t difference_type;
typedef _Tp* pointer;
typedef const _Tp* const_pointer;
typedef _Tp& reference;
typedef const _Tp& const_reference;
typedef _Tp value_type;
template<typename _Tp1> struct rebind { typedef devAllocator<_Tp1> other; };
devAllocator() throw() { }
devAllocator(const devAllocator&) throw() { }
template<typename _Tp1> devAllocator(const devAllocator<_Tp1>&) throw() { }
~devAllocator() throw() { }
pointer address(reference __x) const { return &__x; }
size_type max_size() const throw() { return size_t(-1) / sizeof(_Tp); }
pointer allocate(size_type __n, const void* _p= 0)
{
size_type bytes = __n*sizeof(_Tp);
profilerAllocate(bytes);
_Tp *ptr = (_Tp*) MemoryManager::AcceleratorAllocate(bytes);
assert( ( (_Tp*)ptr != (_Tp *)NULL ) );
return ptr;
}
void deallocate(pointer __p, size_type __n)
{
size_type bytes = __n * sizeof(_Tp);
profilerFree(bytes);
MemoryManager::AcceleratorFree((void *)__p,bytes);
}
void construct(pointer __p, const _Tp& __val) { };
void construct(pointer __p) { };
void destroy(pointer __p) { };
};
template<typename _Tp> inline bool operator==(const devAllocator<_Tp>&, const devAllocator<_Tp>&){ return true; }
template<typename _Tp> inline bool operator!=(const devAllocator<_Tp>&, const devAllocator<_Tp>&){ return false; }
////////////////////////////////////////////////////////////////////////////////
// Template typedefs
////////////////////////////////////////////////////////////////////////////////
template<class T> using commAllocator = uvmAllocator<T>;
//template<class T> using commAllocator = devAllocator<T>;
template<class T> using Vector = std::vector<T,uvmAllocator<T> >;
template<class T> using commVector = std::vector<T,uvmAllocator<T> >;
//template<class T> using Matrix = std::vector<std::vector<T,alignedAllocator<T> > >;
template<class T> using commVector = std::vector<T,devAllocator<T> >;
NAMESPACE_END(Grid);

4
Grid/allocator/MemoryManager.h
View File

@ -93,12 +93,12 @@ private:
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 *AcceleratorAllocate(size_t bytes);
static void AcceleratorFree (void *ptr,size_t bytes);
static void PrintBytes(void);
public:
static void Init(void);
static void InitMessage(void);
static void *AcceleratorAllocate(size_t bytes);
static void AcceleratorFree (void *ptr,size_t bytes);
static void *SharedAllocate(size_t bytes);
static void SharedFree (void *ptr,size_t bytes);
static void *CpuAllocate(size_t bytes);

81
Grid/communicator/Communicator_mpi3.cc
View File

@ -302,60 +302,28 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
int bytes)
{
std::vector<CommsRequest_t> reqs(0);
// unsigned long xcrc = crc32(0L, Z_NULL, 0);
// unsigned long rcrc = crc32(0L, Z_NULL, 0);
// xcrc = crc32(xcrc,(unsigned char *)xmit,bytes);
SendToRecvFromBegin(reqs,xmit,dest,recv,from,bytes);
SendToRecvFromComplete(reqs);
// rcrc = crc32(rcrc,(unsigned char *)recv,bytes);
// printf("proc %d SendToRecvFrom %d bytes %lx %lx\n",_processor,bytes,xcrc,rcrc);
}
void CartesianCommunicator::SendRecvPacket(void *xmit,
void *recv,
int sender,
int receiver,
int bytes)
{
MPI_Status stat;
assert(sender != receiver);
int tag = sender;
if ( _processor == sender ) {
MPI_Send(xmit, bytes, MPI_CHAR,receiver,tag,communicator);
}
if ( _processor == receiver ) {
MPI_Recv(recv, bytes, MPI_CHAR,sender,tag,communicator,&stat);
}
}
// Basic Halo comms primitive
void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void *xmit,
int dest,
void *recv,
int from,
int bytes)
{
unsigned long xcrc = crc32(0L, Z_NULL, 0);
unsigned long rcrc = crc32(0L, Z_NULL, 0);
int myrank = _processor;
int ierr;
if ( CommunicatorPolicy == CommunicatorPolicyConcurrent ) {
MPI_Request xrq;
MPI_Request rrq;
// Enforce no UVM in comms, device or host OK
assert(acceleratorIsCommunicable(xmit));
assert(acceleratorIsCommunicable(recv));
ierr =MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
ierr|=MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
// Give the CPU to MPI immediately; can use threads to overlap optionally
// printf("proc %d SendToRecvFrom %d bytes Sendrecv \n",_processor,bytes);
ierr=MPI_Sendrecv(xmit,bytes,MPI_CHAR,dest,myrank,
recv,bytes,MPI_CHAR,from, from,
communicator,MPI_STATUS_IGNORE);
assert(ierr==0);
assert(ierr==0);
list.push_back(xrq);
list.push_back(rrq);
} else {
// Give the CPU to MPI immediately; can use threads to overlap optionally
ierr=MPI_Sendrecv(xmit,bytes,MPI_CHAR,dest,myrank,
recv,bytes,MPI_CHAR,from, from,
communicator,MPI_STATUS_IGNORE);
assert(ierr==0);
}
// xcrc = crc32(xcrc,(unsigned char *)xmit,bytes);
// rcrc = crc32(rcrc,(unsigned char *)recv,bytes);
// printf("proc %d SendToRecvFrom %d bytes xcrc %lx rcrc %lx\n",_processor,bytes,xcrc,rcrc); fflush
}
// Basic Halo comms primitive
double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
int dest,
void *recv,
@ -411,15 +379,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
return off_node_bytes;
}
void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
{
SendToRecvFromComplete(waitall);
}
void CartesianCommunicator::StencilBarrier(void)
{
MPI_Barrier (ShmComm);
}
void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
{
int nreq=list.size();
@ -430,6 +390,13 @@ void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &
assert(ierr==0);
list.resize(0);
}
void CartesianCommunicator::StencilBarrier(void)
{
MPI_Barrier (ShmComm);
}
//void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
//{
//}
void CartesianCommunicator::Barrier(void)
{
int ierr = MPI_Barrier(communicator);

1
Grid/cshift/Cshift_common.h
View File

@ -222,6 +222,7 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerA
// Test_cshift_red_black code.
// std::cout << "Scatter_plane merge assert(0); think this is buggy FIXME "<< std::endl;// think this is buggy FIXME
std::cout<<" Unthreaded warning -- buffer is not densely packed ??"<<std::endl;
assert(0); // This will fail if hit on GPU
autoView( rhs_v, rhs, CpuWrite);
for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){

2
Grid/qcd/action/fermion/ImprovedStaggeredFermion5D.h
View File

@ -208,7 +208,7 @@ public:
LebesgueOrder LebesgueEvenOdd;
// Comms buffer
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > comm_buf;
// std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > comm_buf;
///////////////////////////////////////////////////////////////
// Conserved current utilities

2
Grid/qcd/action/fermion/WilsonFermion5D.h
View File

@ -215,7 +215,7 @@ public:
LebesgueOrder LebesgueEvenOdd;
// Comms buffer
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > comm_buf;
// std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > comm_buf;
};

23
Grid/threads/Accelerator.h
View File

@ -70,6 +70,7 @@ NAMESPACE_BEGIN(Grid);
//
// Memory management:
//
// int acceleratorIsCommunicable(void *pointer);
// void *acceleratorAllocShared(size_t bytes);
// void acceleratorFreeShared(void *ptr);
//
@ -90,6 +91,7 @@ void acceleratorInit(void);
//////////////////////////////////////////////
#ifdef GRID_CUDA
#include <cuda.h>
#ifdef __CUDA_ARCH__
#define GRID_SIMT
@ -165,6 +167,16 @@ inline void acceleratorFreeShared(void *ptr){ cudaFree(ptr);};
inline void acceleratorFreeDevice(void *ptr){ cudaFree(ptr);};
inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes) { cudaMemcpy(to,from,bytes, cudaMemcpyHostToDevice);}
inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ cudaMemcpy(to,from,bytes, cudaMemcpyDeviceToHost);}
inline int acceleratorIsCommunicable(void *ptr)
{
int uvm;
auto
cuerr = cuPointerGetAttribute( &uvm, CU_POINTER_ATTRIBUTE_IS_MANAGED, (CUdeviceptr) ptr);
assert(cuerr == cudaSuccess );
if(uvm) return 0;
else return 1;
}
#endif
//////////////////////////////////////////////
@ -219,6 +231,15 @@ inline void acceleratorFreeShared(void *ptr){free(ptr,*theGridAccelerator);};
inline void acceleratorFreeDevice(void *ptr){free(ptr,*theGridAccelerator);};
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 int acceleratorIsCommunicable(void *ptr)
{
#if 0
auto uvm = cl::sycl::usm::get_pointer_type(ptr, theGridAccelerator->get_context());
if ( uvm = cl::sycl::usm::alloc::shared ) return 1;
else return 0;
#endif
return 1;
}
#endif
@ -298,6 +319,7 @@ inline void *acceleratorAllocShared(size_t bytes)
return malloc(bytes);
#endif
};
inline int acceleratorIsCommunicable(void *ptr){ return 1; }
inline void *acceleratorAllocDevice(size_t bytes)
{
@ -352,6 +374,7 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) { return 0; } // CUDA spec
inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes) { memcpy(to,from,bytes);}
inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ memcpy(to,from,bytes);}
inline int acceleratorIsCommunicable(void *ptr){ return 1; }
#ifdef HAVE_MM_MALLOC_H
inline void *acceleratorAllocShared(size_t bytes){return _mm_malloc(bytes,GRID_ALLOC_ALIGN);};
inline void *acceleratorAllocDevice(size_t bytes){return _mm_malloc(bytes,GRID_ALLOC_ALIGN);};

8
Grid/util/Coordinate.h
View File

@ -99,10 +99,10 @@ inline std::ostream & operator<<(std::ostream &os, const AcceleratorVector<T,_nd
{
os << "[";
for(int s=0;s<v.size();s++) {
os << v[s] << " ";
}
if (v.size() > 0) {
os << "\b";
os << v[s];
if( s < (v.size()-1) ){
os << " ";
}
}
os << "]";
return os;

108
benchmarks/Benchmark_comms.cc
View File

@ -74,90 +74,6 @@ int main (int argc, char ** argv)
std::vector<double> t_time(Nloop);
time_statistics timestat;
std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
std::cout<<GridLogMessage << "= Benchmarking concurrent halo exchange in "<<nmu<<" dimensions"<<std::endl;
std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
header();
for(int lat=8;lat<=maxlat;lat+=4){
for(int Ls=8;Ls<=8;Ls*=2){
Coordinate latt_size ({lat*mpi_layout[0],
lat*mpi_layout[1],
lat*mpi_layout[2],
lat*mpi_layout[3]});
GridCartesian Grid(latt_size,simd_layout,mpi_layout);
RealD Nrank = Grid._Nprocessors;
RealD Nnode = Grid.NodeCount();
RealD ppn = Nrank/Nnode;
std::vector<Vector<HalfSpinColourVectorD> > xbuf(8);
std::vector<Vector<HalfSpinColourVectorD> > rbuf(8);
int ncomm;
int bytes=lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD);
for(int mu=0;mu<8;mu++){
xbuf[mu].resize(lat*lat*lat*Ls);
rbuf[mu].resize(lat*lat*lat*Ls);
// std::cout << " buffers " << std::hex << (uint64_t)&xbuf[mu][0] <<" " << (uint64_t)&rbuf[mu][0] <<std::endl;
}
for(int i=0;i<Nloop;i++){
double start=usecond();
std::vector<CommsRequest_t> requests;
ncomm=0;
for(int mu=0;mu<4;mu++){
if (mpi_layout[mu]>1 ) {
ncomm++;
int comm_proc=1;
int xmit_to_rank;
int recv_from_rank;
Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
Grid.SendToRecvFromBegin(requests,
(void *)&xbuf[mu][0],
xmit_to_rank,
(void *)&rbuf[mu][0],
recv_from_rank,
bytes);
comm_proc = mpi_layout[mu]-1;
Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
Grid.SendToRecvFromBegin(requests,
(void *)&xbuf[mu+4][0],
xmit_to_rank,
(void *)&rbuf[mu+4][0],
recv_from_rank,
bytes);
}
}
Grid.SendToRecvFromComplete(requests);
Grid.Barrier();
double stop=usecond();
t_time[i] = stop-start; // microseconds
}
timestat.statistics(t_time);
double dbytes = bytes*ppn;
double xbytes = dbytes*2.0*ncomm;
double rbytes = xbytes;
double bidibytes = xbytes+rbytes;
std::cout<<GridLogMessage << std::setw(4) << lat<<"\t"<<Ls<<"\t"
<<std::setw(11) << bytes<< std::fixed << std::setprecision(1) << std::setw(7)
<<std::right<< xbytes/timestat.mean<<" "<< xbytes*timestat.err/(timestat.mean*timestat.mean)<< " "
<<xbytes/timestat.max <<" "<< xbytes/timestat.min
<< "\t\t"<<std::setw(7)<< bidibytes/timestat.mean<< " " << bidibytes*timestat.err/(timestat.mean*timestat.mean) << " "
<< bidibytes/timestat.max << " " << bidibytes/timestat.min << std::endl;
}
}
std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
@ -206,26 +122,22 @@ int main (int argc, char ** argv)
{
std::vector<CommsRequest_t> requests;
Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
Grid.SendToRecvFromBegin(requests,
(void *)&xbuf[mu][0],
xmit_to_rank,
(void *)&rbuf[mu][0],
recv_from_rank,
bytes);
Grid.SendToRecvFromComplete(requests);
Grid.SendToRecvFrom((void *)&xbuf[mu][0],
xmit_to_rank,
(void *)&rbuf[mu][0],
recv_from_rank,
bytes);
}
comm_proc = mpi_layout[mu]-1;
{
std::vector<CommsRequest_t> requests;
Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
Grid.SendToRecvFromBegin(requests,
(void *)&xbuf[mu+4][0],
xmit_to_rank,
(void *)&rbuf[mu+4][0],
recv_from_rank,
bytes);
Grid.SendToRecvFromComplete(requests);
Grid.SendToRecvFrom((void *)&xbuf[mu+4][0],
xmit_to_rank,
(void *)&rbuf[mu+4][0],
recv_from_rank,
bytes);
}
}
}

1
benchmarks/Benchmark_wilson_sweep.cc
View File

@ -89,6 +89,7 @@ int main (int argc, char ** argv)
std::cout << GridLogMessage;
std::cout << latt_size;
std::cout << "\t\t";
GridCartesian Grid(latt_size,simd_layout,mpi_layout);
GridRedBlackCartesian RBGrid(&Grid);

5
configure.ac
View File

@ -154,6 +154,7 @@ AC_ARG_ENABLE([accelerator],
case ${ac_ACCELERATOR} in
cuda)
echo CUDA acceleration
LIBS="${LIBS} -lcuda"
AC_DEFINE([GRID_CUDA],[1],[Use CUDA offload]);;
sycl)
echo SYCL acceleration
@ -323,7 +324,6 @@ case ${CXXTEST} in
# CXXLD="nvcc -v -link"
CXX="${CXXBASE} -x cu "
CXXLD="${CXXBASE} -link"
# CXXFLAGS="$CXXFLAGS -Xcompiler -fno-strict-aliasing -Xcompiler -Wno-unusable-partial-specialization --expt-extended-lambda --expt-relaxed-constexpr"
CXXFLAGS="$CXXFLAGS -Xcompiler -fno-strict-aliasing --expt-extended-lambda --expt-relaxed-constexpr"
if test $ac_openmp = yes; then
CXXFLAGS="$CXXFLAGS -Xcompiler -fopenmp"
@ -483,8 +483,7 @@ case ${ac_SHM} in
LDFLAGS_CPY=$LDFLAGS
CXXFLAGS="$AM_CXXFLAGS $CXXFLAGS"
LDFLAGS="$AM_LDFLAGS $LDFLAGS"
AC_SEARCH_LIBS([shm_unlink], [rt], [],
[AC_MSG_ERROR("no library found for shm_unlink")])
AC_SEARCH_LIBS([shm_unlink], [rt], [],[AC_MSG_ERROR("no library found for shm_unlink")])
CXXFLAGS=$CXXFLAGS_CPY
LDFLAGS=$LDFLAGS_CPY
;;