portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-10 11:26:56 +01:00
Code Releases Activity

Merge remote-tracking branch 'upstream/develop' into gauge_action_deriv

Browse Source
This commit is contained in:
Mashy Green
2025-03-24 10:12:42 +00:00
parent d41542c64b 199818bd6c
commit e465fce201
10 changed files with 141 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
Grid/algorithms/FFT.h
View File

@ -191,7 +191,7 @@ public:
Lattice<sobj> pgbuf(&pencil_g);
autoView(pgbuf_v , pgbuf, CpuWrite);
std::cout << "CPU view" << std::endl;
//std::cout << "CPU view" << std::endl;
typedef typename FFTW<scalar>::FFTW_scalar FFTW_scalar;
typedef typename FFTW<scalar>::FFTW_plan FFTW_plan;
@ -215,7 +215,7 @@ public:
else if ( sign == forward ) div = 1.0;
else assert(0);
std::cout << GridLogPerformance<<"Making FFTW plan" << std::endl;
//std::cout << GridLogPerformance<<"Making FFTW plan" << std::endl;
FFTW_plan p;
{
FFTW_scalar *in = (FFTW_scalar *)&pgbuf_v[0];
@ -229,7 +229,7 @@ public:
}
// Barrel shift and collect global pencil
std::cout << GridLogPerformance<<"Making pencil" << std::endl;
//std::cout << GridLogPerformance<<"Making pencil" << std::endl;
Coordinate lcoor(Nd), gcoor(Nd);
result = source;
int pc = processor_coor[dim];
@ -251,7 +251,7 @@ public:
}
}
std::cout <<GridLogPerformance<< "Looping orthog" << std::endl;
//std::cout <<GridLogPerformance<< "Looping orthog" << std::endl;
// Loop over orthog coords
int NN=pencil_g.lSites();
GridStopWatch timer;
@ -274,7 +274,7 @@ public:
usec += timer.useconds();
flops+= flops_call*NN;
std::cout <<GridLogPerformance<< "Writing back results " << std::endl;
//std::cout <<GridLogPerformance<< "Writing back results " << std::endl;
// writing out result
{
autoView(pgbuf_v,pgbuf,CpuRead);
@ -291,7 +291,7 @@ public:
}
result = result*div;
std::cout <<GridLogPerformance<< "Destroying plan " << std::endl;
//std::cout <<GridLogPerformance<< "Destroying plan " << std::endl;
// destroying plan
FFTW<scalar>::fftw_destroy_plan(p);
#endif

147
Grid/algorithms/blas/BatchedBlas.h
View File

@ -208,8 +208,8 @@ public:
assert(Bkn.size()==batchCount);
assert(Cmn.size()==batchCount);
assert(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
assert(OpB!=GridBLAS_OP_T);
//assert(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
//assert(OpB!=GridBLAS_OP_T);
int lda = m; // m x k column major
int ldb = k; // k x n column major
@ -367,28 +367,67 @@ public:
Eigen::Map<Eigen::MatrixXcd> eAmk(Amk[p],m,k);
Eigen::Map<Eigen::MatrixXcd> eBkn(Bkn[p],k,n);
Eigen::Map<Eigen::MatrixXcd> eCmn(Cmn[p],m,n);
eCmn = beta * eCmn + alpha * eAmk * eBkn ;
if (std::abs(beta) != 0.0)
eCmn = beta * eCmn + alpha * eAmk * eBkn ;
else
eCmn = alpha * eAmk * eBkn ;
});
} else if ( (OpA == GridBLAS_OP_C ) && (OpB == GridBLAS_OP_N) ) {
thread_for (p, batchCount, {
Eigen::Map<Eigen::MatrixXcd> eAmk(Amk[p],k,m);
Eigen::Map<Eigen::MatrixXcd> eBkn(Bkn[p],k,n);
Eigen::Map<Eigen::MatrixXcd> eCmn(Cmn[p],m,n);
eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn ;
if (std::abs(beta) != 0.0)
eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn ;
else
eCmn = alpha * eAmk.adjoint() * eBkn ;
});
} else if ( (OpA == GridBLAS_OP_T ) && (OpB == GridBLAS_OP_N) ) {
thread_for (p, batchCount, {
Eigen::Map<Eigen::MatrixXcd> eAmk(Amk[p],k,m);
Eigen::Map<Eigen::MatrixXcd> eBkn(Bkn[p],k,n);
Eigen::Map<Eigen::MatrixXcd> eCmn(Cmn[p],m,n);
if (std::abs(beta) != 0.0)
eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn ;
else
eCmn = alpha * eAmk.transpose() * eBkn ;
});
} else if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_C) ) {
thread_for (p, batchCount, {
Eigen::Map<Eigen::MatrixXcd> eAmk(Amk[p],m,k);
Eigen::Map<Eigen::MatrixXcd> eBkn(Bkn[p],n,k);
Eigen::Map<Eigen::MatrixXcd> eCmn(Cmn[p],m,n);
eCmn = beta * eCmn + alpha * eAmk * eBkn.adjoint() ;
if (std::abs(beta) != 0.0)
eCmn = beta * eCmn + alpha * eAmk * eBkn.adjoint() ;
else
eCmn = alpha * eAmk * eBkn.adjoint() ;
});
} else if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_T) ) {
thread_for (p, batchCount, {
Eigen::Map<Eigen::MatrixXcd> eAmk(Amk[p],m,k);
Eigen::Map<Eigen::MatrixXcd> eBkn(Bkn[p],n,k);
Eigen::Map<Eigen::MatrixXcd> eCmn(Cmn[p],m,n);
eCmn = beta * eCmn + alpha * eAmk * eBkn.transpose() ;
});
} else if ( (OpA == GridBLAS_OP_C ) && (OpB == GridBLAS_OP_C) ) {
thread_for (p, batchCount, {
Eigen::Map<Eigen::MatrixXcd> eAmk(Amk[p],k,m);
Eigen::Map<Eigen::MatrixXcd> eBkn(Bkn[p],n,k);
Eigen::Map<Eigen::MatrixXcd> eCmn(Cmn[p],m,n);
eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn.adjoint() ;
if (std::abs(beta) != 0.0)
eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn.adjoint() ;
else
eCmn = alpha * eAmk.adjoint() * eBkn.adjoint() ;
} );
} else if ( (OpA == GridBLAS_OP_T ) && (OpB == GridBLAS_OP_T) ) {
thread_for (p, batchCount, {
Eigen::Map<Eigen::MatrixXcd> eAmk(Amk[p],k,m);
Eigen::Map<Eigen::MatrixXcd> eBkn(Bkn[p],n,k);
Eigen::Map<Eigen::MatrixXcd> eCmn(Cmn[p],m,n);
if (std::abs(beta) != 0.0)
eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn.transpose() ;
else
eCmn = alpha * eAmk.transpose() * eBkn.transpose() ;
} );
} else {
assert(0);
@ -414,8 +453,8 @@ public:
RealD t2=usecond();
int32_t batchCount = Amk.size();
assert(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
assert(OpB!=GridBLAS_OP_T);
//assert(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
//assert(OpB!=GridBLAS_OP_T);
int lda = m; // m x k column major
int ldb = k; // k x n column major
@ -514,28 +553,70 @@ public:
Eigen::Map<Eigen::MatrixXcf> eAmk(Amk[p],m,k);
Eigen::Map<Eigen::MatrixXcf> eBkn(Bkn[p],k,n);
Eigen::Map<Eigen::MatrixXcf> eCmn(Cmn[p],m,n);
eCmn = beta * eCmn + alpha * eAmk * eBkn ;
if (std::abs(beta) != 0.0)
eCmn = beta * eCmn + alpha * eAmk * eBkn ;
else
eCmn = alpha * eAmk * eBkn ;
});
} else if ( (OpA == GridBLAS_OP_C ) && (OpB == GridBLAS_OP_N) ) {
thread_for (p, batchCount, {
Eigen::Map<Eigen::MatrixXcf> eAmk(Amk[p],k,m);
Eigen::Map<Eigen::MatrixXcf> eBkn(Bkn[p],k,n);
Eigen::Map<Eigen::MatrixXcf> eCmn(Cmn[p],m,n);
eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn ;
if (std::abs(beta) != 0.0)
eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn ;
else
eCmn = alpha * eAmk.adjoint() * eBkn ;
});
} else if ( (OpA == GridBLAS_OP_T ) && (OpB == GridBLAS_OP_N) ) {
thread_for (p, batchCount, {
Eigen::Map<Eigen::MatrixXcf> eAmk(Amk[p],k,m);
Eigen::Map<Eigen::MatrixXcf> eBkn(Bkn[p],k,n);
Eigen::Map<Eigen::MatrixXcf> eCmn(Cmn[p],m,n);
if (std::abs(beta) != 0.0)
eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn ;
else
eCmn = alpha * eAmk.transpose() * eBkn ;
});
} else if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_C) ) {
thread_for (p, batchCount, {
Eigen::Map<Eigen::MatrixXcf> eAmk(Amk[p],m,k);
Eigen::Map<Eigen::MatrixXcf> eBkn(Bkn[p],n,k);
Eigen::Map<Eigen::MatrixXcf> eCmn(Cmn[p],m,n);
eCmn = beta * eCmn + alpha * eAmk * eBkn.adjoint() ;
if (std::abs(beta) != 0.0)
eCmn = beta * eCmn + alpha * eAmk * eBkn.adjoint() ;
else
eCmn = alpha * eAmk * eBkn.adjoint() ;
});
} else if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_T) ) {
thread_for (p, batchCount, {
Eigen::Map<Eigen::MatrixXcf> eAmk(Amk[p],m,k);
Eigen::Map<Eigen::MatrixXcf> eBkn(Bkn[p],n,k);
Eigen::Map<Eigen::MatrixXcf> eCmn(Cmn[p],m,n);
if (std::abs(beta) != 0.0)
eCmn = beta * eCmn + alpha * eAmk * eBkn.transpose() ;
else
eCmn = alpha * eAmk * eBkn.transpose() ;
});
} else if ( (OpA == GridBLAS_OP_C ) && (OpB == GridBLAS_OP_C) ) {
thread_for (p, batchCount, {
Eigen::Map<Eigen::MatrixXcf> eAmk(Amk[p],k,m);
Eigen::Map<Eigen::MatrixXcf> eBkn(Bkn[p],n,k);
Eigen::Map<Eigen::MatrixXcf> eCmn(Cmn[p],m,n);
eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn.adjoint() ;
if (std::abs(beta) != 0.0)
eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn.adjoint() ;
else
eCmn = alpha * eAmk.adjoint() * eBkn.adjoint() ;
} );
} else if ( (OpA == GridBLAS_OP_T ) && (OpB == GridBLAS_OP_T) ) {
thread_for (p, batchCount, {
Eigen::Map<Eigen::MatrixXcf> eAmk(Amk[p],k,m);
Eigen::Map<Eigen::MatrixXcf> eBkn(Bkn[p],n,k);
Eigen::Map<Eigen::MatrixXcf> eCmn(Cmn[p],m,n);
if (std::abs(beta) != 0.0)
eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn.transpose() ;
else
eCmn = alpha * eAmk.transpose() * eBkn.transpose() ;
} );
} else {
assert(0);
@ -661,29 +742,41 @@ public:
Eigen::Map<Eigen::MatrixXf> eAmk(Amk[p],m,k);
Eigen::Map<Eigen::MatrixXf> eBkn(Bkn[p],k,n);
Eigen::Map<Eigen::MatrixXf> eCmn(Cmn[p],m,n);
eCmn = beta * eCmn + alpha * eAmk * eBkn ;
if (std::abs(beta) != 0.0)
eCmn = beta * eCmn + alpha * eAmk * eBkn ;
else
eCmn = alpha * eAmk * eBkn ;
});
} else if ( (OpA == GridBLAS_OP_T ) && (OpB == GridBLAS_OP_N) ) {
thread_for (p, batchCount, {
Eigen::Map<Eigen::MatrixXf> eAmk(Amk[p],k,m);
Eigen::Map<Eigen::MatrixXf> eBkn(Bkn[p],k,n);
Eigen::Map<Eigen::MatrixXf> eCmn(Cmn[p],m,n);
eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn ;
if (std::abs(beta) != 0.0)
eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn ;
else
eCmn = alpha * eAmk.transpose() * eBkn ;
});
} else if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_T) ) {
thread_for (p, batchCount, {
Eigen::Map<Eigen::MatrixXf> eAmk(Amk[p],m,k);
Eigen::Map<Eigen::MatrixXf> eBkn(Bkn[p],n,k);
Eigen::Map<Eigen::MatrixXf> eCmn(Cmn[p],m,n);
eCmn = beta * eCmn + alpha * eAmk * eBkn.transpose() ;
if (std::abs(beta) != 0.0)
eCmn = beta * eCmn + alpha * eAmk * eBkn.transpose() ;
else
eCmn = alpha * eAmk * eBkn.transpose() ;
});
} else if ( (OpA == GridBLAS_OP_T ) && (OpB == GridBLAS_OP_T) ) {
thread_for (p, batchCount, {
Eigen::Map<Eigen::MatrixXf> eAmk(Amk[p],k,m);
Eigen::Map<Eigen::MatrixXf> eBkn(Bkn[p],n,k);
Eigen::Map<Eigen::MatrixXf> eCmn(Cmn[p],m,n);
eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn.transpose() ;
} );
if (std::abs(beta) != 0.0)
eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn.transpose() ;
else
eCmn = alpha * eAmk.transpose() * eBkn.transpose() ;
});
} else {
assert(0);
}
@ -809,28 +902,40 @@ public:
Eigen::Map<Eigen::MatrixXd> eAmk(Amk[p],m,k);
Eigen::Map<Eigen::MatrixXd> eBkn(Bkn[p],k,n);
Eigen::Map<Eigen::MatrixXd> eCmn(Cmn[p],m,n);
eCmn = beta * eCmn + alpha * eAmk * eBkn ;
if (std::abs(beta) != 0.0)
eCmn = beta * eCmn + alpha * eAmk * eBkn ;
else
eCmn = alpha * eAmk * eBkn ;
});
} else if ( (OpA == GridBLAS_OP_T ) && (OpB == GridBLAS_OP_N) ) {
thread_for (p, batchCount, {
Eigen::Map<Eigen::MatrixXd> eAmk(Amk[p],k,m);
Eigen::Map<Eigen::MatrixXd> eBkn(Bkn[p],k,n);
Eigen::Map<Eigen::MatrixXd> eCmn(Cmn[p],m,n);
eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn ;
if (std::abs(beta) != 0.0)
eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn ;
else
eCmn = alpha * eAmk.transpose() * eBkn ;
});
} else if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_T) ) {
thread_for (p, batchCount, {
Eigen::Map<Eigen::MatrixXd> eAmk(Amk[p],m,k);
Eigen::Map<Eigen::MatrixXd> eBkn(Bkn[p],n,k);
Eigen::Map<Eigen::MatrixXd> eCmn(Cmn[p],m,n);
eCmn = beta * eCmn + alpha * eAmk * eBkn.transpose() ;
if (std::abs(beta) != 0.0)
eCmn = beta * eCmn + alpha * eAmk * eBkn.transpose() ;
else
eCmn = alpha * eAmk * eBkn.transpose() ;
});
} else if ( (OpA == GridBLAS_OP_T ) && (OpB == GridBLAS_OP_T) ) {
thread_for (p, batchCount, {
Eigen::Map<Eigen::MatrixXd> eAmk(Amk[p],k,m);
Eigen::Map<Eigen::MatrixXd> eBkn(Bkn[p],n,k);
Eigen::Map<Eigen::MatrixXd> eCmn(Cmn[p],m,n);
eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn.transpose() ;
if (std::abs(beta) != 0.0)
eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn.transpose() ;
else
eCmn = alpha * eAmk.transpose() * eBkn.transpose() ;
});
} else {
assert(0);

2
Grid/allocator/AlignedAllocator.h
View File

@ -69,7 +69,7 @@ public:
}
// 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, const _Tp& __val) { };
void construct(pointer __p) { };
void destroy(pointer __p) { };
};

3
Grid/allocator/MemoryManagerCache.cc
View File

@ -234,6 +234,9 @@ void *MemoryManager::ViewOpen(void* _CpuPtr,size_t bytes,ViewMode mode,ViewAdvis
}
void MemoryManager::EvictVictims(uint64_t bytes)
{
if(bytes>=DeviceMaxBytes) {
printf("EvictVictims bytes %ld DeviceMaxBytes %ld\n",bytes,DeviceMaxBytes);
}
assert(bytes<DeviceMaxBytes);
while(bytes+DeviceLRUBytes > DeviceMaxBytes){
if ( DeviceLRUBytes > 0){

3
Grid/communicator/Communicator_base.h
View File

@ -149,7 +149,8 @@ public:
sizeof(obj),d*100+p);
}
CommsComplete(list);
if (!list.empty()) // avoid triggering assert in comms == none
CommsComplete(list);
for(int p=1;p<_processors[d];p++){
accum = accum + column[p];
}

3
Grid/communicator/Communicator_mpi3.cc
View File

@ -759,9 +759,6 @@ void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsReque
int nreq=MpiRequests.size();
std::cout << GridLogMessage << " StencilSendToRevFromComplete "<<nreq<<" Mpi Requests"<<std::endl;
if (nreq>0) {
status.resize(MpiRequests.size());
int ierr = MPI_Waitall(MpiRequests.size(),&MpiRequests[0],&status[0]); // Sends are guaranteed in order. No harm in not completing.

2
Grid/communicator/SharedMemoryMPI.cc
View File

@ -542,7 +542,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
// Each MPI rank should allocate our own buffer
///////////////////////////////////////////////////////////////////////////////////////////////////////////
#ifndef ACCELERATOR_AWARE_MPI
printf("Host buffer allocate for GPU non-aware MPI\n");
// printf("Host buffer allocate for GPU non-aware MPI\n");
#if 0
HostCommBuf= acceleratorAllocHost(bytes);
#else

5
Grid/cshift/Cshift_mpi.h
View File

@ -126,8 +126,8 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
static deviceVector<vobj> send_buf; send_buf.resize(buffer_size);
static deviceVector<vobj> recv_buf; recv_buf.resize(buffer_size);
#ifndef ACCELERATOR_AWARE_MPI
static hostVector<vobj> hsend_buf; hsend_buf.resize(buffer_size);
static hostVector<vobj> hrecv_buf; hrecv_buf.resize(buffer_size);
static hostVector<vobj> hsend_buf; hsend_buf.resize(buffer_size);
static hostVector<vobj> hrecv_buf; hrecv_buf.resize(buffer_size);
#endif
int cb= (cbmask==0x2)? Odd : Even;
@ -244,7 +244,6 @@ template<class vobj> void Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
scalar_object * recv_buf_extract_mpi;
scalar_object * send_buf_extract_mpi;
for(int s=0;s<Nsimd;s++){
send_buf_extract[s].resize(buffer_size);
recv_buf_extract[s].resize(buffer_size);

1
Grid/qcd/action/fermion/WilsonCompressor.h
View File

@ -485,7 +485,6 @@ public:
assert(this->u_comm_offset==this->_unified_buffer_size);
accelerator_barrier();
#ifdef NVLINK_GET
#warning "NVLINK_GET"
this->_grid->StencilBarrier(); // He can now get mu local gather, I can get his
// Synch shared memory on a single nodes; could use an asynchronous barrier here and defer check
// Or issue barrier AFTER the DMA is running

1
Grid/stencil/Stencil.h
View File

@ -518,7 +518,6 @@ public:
}
accelerator_barrier(); // All my local gathers are complete
#ifdef NVLINK_GET
#warning "NVLINK_GET"
_grid->StencilBarrier(); // He can now get mu local gather, I can get his
// Synch shared memory on a single nodes; could use an asynchronous barrier here and defer check
// Or issue barrier AFTER the DMA is running