mirror of
https://github.com/paboyle/Grid.git
synced 2025-06-17 15:27:06 +01:00
Merge branch 'develop' into feature/multi-communicator
This commit is contained in:
@ -11,7 +11,7 @@ int PointerCache::victim;
|
||||
|
||||
void *PointerCache::Insert(void *ptr,size_t bytes) {
|
||||
|
||||
if (bytes < 4096 ) return NULL;
|
||||
if (bytes < 4096 ) return ptr;
|
||||
|
||||
#ifdef GRID_OMP
|
||||
assert(omp_in_parallel()==0);
|
||||
|
@ -185,17 +185,18 @@ public:
|
||||
////////////////////////////////////////////////////////////////
|
||||
|
||||
void show_decomposition(){
|
||||
std::cout << GridLogMessage << "Full Dimensions : " << _fdimensions << std::endl;
|
||||
std::cout << GridLogMessage << "Global Dimensions : " << _gdimensions << std::endl;
|
||||
std::cout << GridLogMessage << "Local Dimensions : " << _ldimensions << std::endl;
|
||||
std::cout << GridLogMessage << "Reduced Dimensions : " << _rdimensions << std::endl;
|
||||
std::cout << GridLogMessage << "Outer strides : " << _ostride << std::endl;
|
||||
std::cout << GridLogMessage << "Inner strides : " << _istride << std::endl;
|
||||
std::cout << GridLogMessage << "iSites : " << _isites << std::endl;
|
||||
std::cout << GridLogMessage << "oSites : " << _osites << std::endl;
|
||||
std::cout << GridLogMessage << "lSites : " << lSites() << std::endl;
|
||||
std::cout << GridLogMessage << "gSites : " << gSites() << std::endl;
|
||||
std::cout << GridLogMessage << "Nd : " << _ndimension << std::endl;
|
||||
std::cout << GridLogMessage << "\tFull Dimensions : " << _fdimensions << std::endl;
|
||||
std::cout << GridLogMessage << "\tSIMD layout : " << _simd_layout << std::endl;
|
||||
std::cout << GridLogMessage << "\tGlobal Dimensions : " << _gdimensions << std::endl;
|
||||
std::cout << GridLogMessage << "\tLocal Dimensions : " << _ldimensions << std::endl;
|
||||
std::cout << GridLogMessage << "\tReduced Dimensions : " << _rdimensions << std::endl;
|
||||
std::cout << GridLogMessage << "\tOuter strides : " << _ostride << std::endl;
|
||||
std::cout << GridLogMessage << "\tInner strides : " << _istride << std::endl;
|
||||
std::cout << GridLogMessage << "\tiSites : " << _isites << std::endl;
|
||||
std::cout << GridLogMessage << "\toSites : " << _osites << std::endl;
|
||||
std::cout << GridLogMessage << "\tlSites : " << lSites() << std::endl;
|
||||
std::cout << GridLogMessage << "\tgSites : " << gSites() << std::endl;
|
||||
std::cout << GridLogMessage << "\tNd : " << _ndimension << std::endl;
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////
|
||||
|
@ -62,77 +62,81 @@ public:
|
||||
return shift;
|
||||
}
|
||||
GridCartesian(const std::vector<int> &dimensions,
|
||||
const std::vector<int> &simd_layout,
|
||||
const std::vector<int> &processor_grid
|
||||
) : GridBase(processor_grid)
|
||||
const std::vector<int> &simd_layout,
|
||||
const std::vector<int> &processor_grid) : GridBase(processor_grid)
|
||||
{
|
||||
///////////////////////
|
||||
// Grid information
|
||||
///////////////////////
|
||||
_ndimension = dimensions.size();
|
||||
|
||||
_fdimensions.resize(_ndimension);
|
||||
_gdimensions.resize(_ndimension);
|
||||
_ldimensions.resize(_ndimension);
|
||||
_rdimensions.resize(_ndimension);
|
||||
_simd_layout.resize(_ndimension);
|
||||
_lstart.resize(_ndimension);
|
||||
_lend.resize(_ndimension);
|
||||
|
||||
_ostride.resize(_ndimension);
|
||||
_istride.resize(_ndimension);
|
||||
|
||||
_fsites = _gsites = _osites = _isites = 1;
|
||||
///////////////////////
|
||||
// Grid information
|
||||
///////////////////////
|
||||
_ndimension = dimensions.size();
|
||||
|
||||
for(int d=0;d<_ndimension;d++){
|
||||
_fdimensions[d] = dimensions[d]; // Global dimensions
|
||||
_gdimensions[d] = _fdimensions[d]; // Global dimensions
|
||||
_simd_layout[d] = simd_layout[d];
|
||||
_fsites = _fsites * _fdimensions[d];
|
||||
_gsites = _gsites * _gdimensions[d];
|
||||
_fdimensions.resize(_ndimension);
|
||||
_gdimensions.resize(_ndimension);
|
||||
_ldimensions.resize(_ndimension);
|
||||
_rdimensions.resize(_ndimension);
|
||||
_simd_layout.resize(_ndimension);
|
||||
_lstart.resize(_ndimension);
|
||||
_lend.resize(_ndimension);
|
||||
|
||||
//FIXME check for exact division
|
||||
_ostride.resize(_ndimension);
|
||||
_istride.resize(_ndimension);
|
||||
|
||||
// Use a reduced simd grid
|
||||
_ldimensions[d]= _gdimensions[d]/_processors[d]; //local dimensions
|
||||
_rdimensions[d]= _ldimensions[d]/_simd_layout[d]; //overdecomposition
|
||||
_lstart[d] = _processor_coor[d]*_ldimensions[d];
|
||||
_lend[d] = _processor_coor[d]*_ldimensions[d]+_ldimensions[d]-1;
|
||||
_osites *= _rdimensions[d];
|
||||
_isites *= _simd_layout[d];
|
||||
|
||||
// Addressing support
|
||||
if ( d==0 ) {
|
||||
_ostride[d] = 1;
|
||||
_istride[d] = 1;
|
||||
} else {
|
||||
_ostride[d] = _ostride[d-1]*_rdimensions[d-1];
|
||||
_istride[d] = _istride[d-1]*_simd_layout[d-1];
|
||||
}
|
||||
_fsites = _gsites = _osites = _isites = 1;
|
||||
|
||||
for (int d = 0; d < _ndimension; d++)
|
||||
{
|
||||
_fdimensions[d] = dimensions[d]; // Global dimensions
|
||||
_gdimensions[d] = _fdimensions[d]; // Global dimensions
|
||||
_simd_layout[d] = simd_layout[d];
|
||||
_fsites = _fsites * _fdimensions[d];
|
||||
_gsites = _gsites * _gdimensions[d];
|
||||
|
||||
// Use a reduced simd grid
|
||||
_ldimensions[d] = _gdimensions[d] / _processors[d]; //local dimensions
|
||||
assert(_ldimensions[d] * _processors[d] == _gdimensions[d]);
|
||||
|
||||
_rdimensions[d] = _ldimensions[d] / _simd_layout[d]; //overdecomposition
|
||||
assert(_rdimensions[d] * _simd_layout[d] == _ldimensions[d]);
|
||||
|
||||
_lstart[d] = _processor_coor[d] * _ldimensions[d];
|
||||
_lend[d] = _processor_coor[d] * _ldimensions[d] + _ldimensions[d] - 1;
|
||||
_osites *= _rdimensions[d];
|
||||
_isites *= _simd_layout[d];
|
||||
|
||||
// Addressing support
|
||||
if (d == 0)
|
||||
{
|
||||
_ostride[d] = 1;
|
||||
_istride[d] = 1;
|
||||
}
|
||||
|
||||
///////////////////////
|
||||
// subplane information
|
||||
///////////////////////
|
||||
_slice_block.resize(_ndimension);
|
||||
_slice_stride.resize(_ndimension);
|
||||
_slice_nblock.resize(_ndimension);
|
||||
|
||||
int block =1;
|
||||
int nblock=1;
|
||||
for(int d=0;d<_ndimension;d++) nblock*=_rdimensions[d];
|
||||
|
||||
for(int d=0;d<_ndimension;d++){
|
||||
nblock/=_rdimensions[d];
|
||||
_slice_block[d] =block;
|
||||
_slice_stride[d]=_ostride[d]*_rdimensions[d];
|
||||
_slice_nblock[d]=nblock;
|
||||
block = block*_rdimensions[d];
|
||||
else
|
||||
{
|
||||
_ostride[d] = _ostride[d - 1] * _rdimensions[d - 1];
|
||||
_istride[d] = _istride[d - 1] * _simd_layout[d - 1];
|
||||
}
|
||||
}
|
||||
|
||||
///////////////////////
|
||||
// subplane information
|
||||
///////////////////////
|
||||
_slice_block.resize(_ndimension);
|
||||
_slice_stride.resize(_ndimension);
|
||||
_slice_nblock.resize(_ndimension);
|
||||
|
||||
int block = 1;
|
||||
int nblock = 1;
|
||||
for (int d = 0; d < _ndimension; d++)
|
||||
nblock *= _rdimensions[d];
|
||||
|
||||
for (int d = 0; d < _ndimension; d++)
|
||||
{
|
||||
nblock /= _rdimensions[d];
|
||||
_slice_block[d] = block;
|
||||
_slice_stride[d] = _ostride[d] * _rdimensions[d];
|
||||
_slice_nblock[d] = nblock;
|
||||
block = block * _rdimensions[d];
|
||||
}
|
||||
};
|
||||
};
|
||||
|
||||
|
||||
}
|
||||
#endif
|
||||
|
@ -131,21 +131,21 @@ public:
|
||||
Init(dimensions,simd_layout,processor_grid,checker_dim_mask,0);
|
||||
}
|
||||
void Init(const std::vector<int> &dimensions,
|
||||
const std::vector<int> &simd_layout,
|
||||
const std::vector<int> &processor_grid,
|
||||
const std::vector<int> &checker_dim_mask,
|
||||
int checker_dim)
|
||||
const std::vector<int> &simd_layout,
|
||||
const std::vector<int> &processor_grid,
|
||||
const std::vector<int> &checker_dim_mask,
|
||||
int checker_dim)
|
||||
{
|
||||
///////////////////////
|
||||
// Grid information
|
||||
///////////////////////
|
||||
///////////////////////
|
||||
// Grid information
|
||||
///////////////////////
|
||||
_checker_dim = checker_dim;
|
||||
assert(checker_dim_mask[checker_dim]==1);
|
||||
assert(checker_dim_mask[checker_dim] == 1);
|
||||
_ndimension = dimensions.size();
|
||||
assert(checker_dim_mask.size()==_ndimension);
|
||||
assert(processor_grid.size()==_ndimension);
|
||||
assert(simd_layout.size()==_ndimension);
|
||||
|
||||
assert(checker_dim_mask.size() == _ndimension);
|
||||
assert(processor_grid.size() == _ndimension);
|
||||
assert(simd_layout.size() == _ndimension);
|
||||
|
||||
_fdimensions.resize(_ndimension);
|
||||
_gdimensions.resize(_ndimension);
|
||||
_ldimensions.resize(_ndimension);
|
||||
@ -153,114 +153,133 @@ public:
|
||||
_simd_layout.resize(_ndimension);
|
||||
_lstart.resize(_ndimension);
|
||||
_lend.resize(_ndimension);
|
||||
|
||||
|
||||
_ostride.resize(_ndimension);
|
||||
_istride.resize(_ndimension);
|
||||
|
||||
|
||||
_fsites = _gsites = _osites = _isites = 1;
|
||||
|
||||
_checker_dim_mask=checker_dim_mask;
|
||||
|
||||
for(int d=0;d<_ndimension;d++){
|
||||
_fdimensions[d] = dimensions[d];
|
||||
_gdimensions[d] = _fdimensions[d];
|
||||
_fsites = _fsites * _fdimensions[d];
|
||||
_gsites = _gsites * _gdimensions[d];
|
||||
|
||||
if (d==_checker_dim) {
|
||||
_gdimensions[d] = _gdimensions[d]/2; // Remove a checkerboard
|
||||
}
|
||||
_ldimensions[d] = _gdimensions[d]/_processors[d];
|
||||
_lstart[d] = _processor_coor[d]*_ldimensions[d];
|
||||
_lend[d] = _processor_coor[d]*_ldimensions[d]+_ldimensions[d]-1;
|
||||
_checker_dim_mask = checker_dim_mask;
|
||||
|
||||
// Use a reduced simd grid
|
||||
_simd_layout[d] = simd_layout[d];
|
||||
_rdimensions[d]= _ldimensions[d]/_simd_layout[d];
|
||||
assert(_rdimensions[d]>0);
|
||||
for (int d = 0; d < _ndimension; d++)
|
||||
{
|
||||
_fdimensions[d] = dimensions[d];
|
||||
_gdimensions[d] = _fdimensions[d];
|
||||
_fsites = _fsites * _fdimensions[d];
|
||||
_gsites = _gsites * _gdimensions[d];
|
||||
|
||||
// all elements of a simd vector must have same checkerboard.
|
||||
// If Ls vectorised, this must still be the case; e.g. dwf rb5d
|
||||
if ( _simd_layout[d]>1 ) {
|
||||
if ( checker_dim_mask[d] ) {
|
||||
assert( (_rdimensions[d]&0x1) == 0 );
|
||||
}
|
||||
}
|
||||
if (d == _checker_dim)
|
||||
{
|
||||
assert((_gdimensions[d] & 0x1) == 0);
|
||||
_gdimensions[d] = _gdimensions[d] / 2; // Remove a checkerboard
|
||||
}
|
||||
_ldimensions[d] = _gdimensions[d] / _processors[d];
|
||||
assert(_ldimensions[d] * _processors[d] == _gdimensions[d]);
|
||||
_lstart[d] = _processor_coor[d] * _ldimensions[d];
|
||||
_lend[d] = _processor_coor[d] * _ldimensions[d] + _ldimensions[d] - 1;
|
||||
|
||||
_osites *= _rdimensions[d];
|
||||
_isites *= _simd_layout[d];
|
||||
|
||||
// Addressing support
|
||||
if ( d==0 ) {
|
||||
_ostride[d] = 1;
|
||||
_istride[d] = 1;
|
||||
} else {
|
||||
_ostride[d] = _ostride[d-1]*_rdimensions[d-1];
|
||||
_istride[d] = _istride[d-1]*_simd_layout[d-1];
|
||||
}
|
||||
// Use a reduced simd grid
|
||||
_simd_layout[d] = simd_layout[d];
|
||||
_rdimensions[d] = _ldimensions[d] / _simd_layout[d]; // this is not checking if this is integer
|
||||
assert(_rdimensions[d] * _simd_layout[d] == _ldimensions[d]);
|
||||
assert(_rdimensions[d] > 0);
|
||||
|
||||
// all elements of a simd vector must have same checkerboard.
|
||||
// If Ls vectorised, this must still be the case; e.g. dwf rb5d
|
||||
if (_simd_layout[d] > 1)
|
||||
{
|
||||
if (checker_dim_mask[d])
|
||||
{
|
||||
assert((_rdimensions[d] & 0x1) == 0);
|
||||
}
|
||||
}
|
||||
|
||||
_osites *= _rdimensions[d];
|
||||
_isites *= _simd_layout[d];
|
||||
|
||||
// Addressing support
|
||||
if (d == 0)
|
||||
{
|
||||
_ostride[d] = 1;
|
||||
_istride[d] = 1;
|
||||
}
|
||||
else
|
||||
{
|
||||
_ostride[d] = _ostride[d - 1] * _rdimensions[d - 1];
|
||||
_istride[d] = _istride[d - 1] * _simd_layout[d - 1];
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// subplane information
|
||||
////////////////////////////////////////////////////////////////////////////////////////////
|
||||
_slice_block.resize(_ndimension);
|
||||
_slice_stride.resize(_ndimension);
|
||||
_slice_nblock.resize(_ndimension);
|
||||
|
||||
int block =1;
|
||||
int nblock=1;
|
||||
for(int d=0;d<_ndimension;d++) nblock*=_rdimensions[d];
|
||||
|
||||
for(int d=0;d<_ndimension;d++){
|
||||
nblock/=_rdimensions[d];
|
||||
_slice_block[d] =block;
|
||||
_slice_stride[d]=_ostride[d]*_rdimensions[d];
|
||||
_slice_nblock[d]=nblock;
|
||||
block = block*_rdimensions[d];
|
||||
|
||||
int block = 1;
|
||||
int nblock = 1;
|
||||
for (int d = 0; d < _ndimension; d++)
|
||||
nblock *= _rdimensions[d];
|
||||
|
||||
for (int d = 0; d < _ndimension; d++)
|
||||
{
|
||||
nblock /= _rdimensions[d];
|
||||
_slice_block[d] = block;
|
||||
_slice_stride[d] = _ostride[d] * _rdimensions[d];
|
||||
_slice_nblock[d] = nblock;
|
||||
block = block * _rdimensions[d];
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////
|
||||
// Create a checkerboard lookup table
|
||||
////////////////////////////////////////////////
|
||||
int rvol = 1;
|
||||
for(int d=0;d<_ndimension;d++){
|
||||
rvol=rvol * _rdimensions[d];
|
||||
for (int d = 0; d < _ndimension; d++)
|
||||
{
|
||||
rvol = rvol * _rdimensions[d];
|
||||
}
|
||||
_checker_board.resize(rvol);
|
||||
for(int osite=0;osite<_osites;osite++){
|
||||
_checker_board[osite] = CheckerBoardFromOindex (osite);
|
||||
for (int osite = 0; osite < _osites; osite++)
|
||||
{
|
||||
_checker_board[osite] = CheckerBoardFromOindex(osite);
|
||||
}
|
||||
|
||||
};
|
||||
protected:
|
||||
|
||||
protected:
|
||||
virtual int oIndex(std::vector<int> &coor)
|
||||
{
|
||||
int idx=0;
|
||||
for(int d=0;d<_ndimension;d++) {
|
||||
if( d==_checker_dim ) {
|
||||
idx+=_ostride[d]*((coor[d]/2)%_rdimensions[d]);
|
||||
} else {
|
||||
idx+=_ostride[d]*(coor[d]%_rdimensions[d]);
|
||||
}
|
||||
int idx = 0;
|
||||
for (int d = 0; d < _ndimension; d++)
|
||||
{
|
||||
if (d == _checker_dim)
|
||||
{
|
||||
idx += _ostride[d] * ((coor[d] / 2) % _rdimensions[d]);
|
||||
}
|
||||
else
|
||||
{
|
||||
idx += _ostride[d] * (coor[d] % _rdimensions[d]);
|
||||
}
|
||||
}
|
||||
return idx;
|
||||
};
|
||||
|
||||
|
||||
virtual int iIndex(std::vector<int> &lcoor)
|
||||
{
|
||||
int idx=0;
|
||||
for(int d=0;d<_ndimension;d++) {
|
||||
if( d==_checker_dim ) {
|
||||
idx+=_istride[d]*(lcoor[d]/(2*_rdimensions[d]));
|
||||
} else {
|
||||
idx+=_istride[d]*(lcoor[d]/_rdimensions[d]);
|
||||
}
|
||||
}
|
||||
return idx;
|
||||
int idx = 0;
|
||||
for (int d = 0; d < _ndimension; d++)
|
||||
{
|
||||
if (d == _checker_dim)
|
||||
{
|
||||
idx += _istride[d] * (lcoor[d] / (2 * _rdimensions[d]));
|
||||
}
|
||||
else
|
||||
{
|
||||
idx += _istride[d] * (lcoor[d] / _rdimensions[d]);
|
||||
}
|
||||
}
|
||||
return idx;
|
||||
}
|
||||
};
|
||||
|
||||
}
|
||||
#endif
|
||||
|
@ -98,35 +98,39 @@ class BinaryIO {
|
||||
|
||||
NerscChecksum(grid,scalardata,nersc_csum);
|
||||
}
|
||||
|
||||
template<class fobj> static inline void NerscChecksum(GridBase *grid,std::vector<fobj> &fbuf,uint32_t &nersc_csum)
|
||||
|
||||
template <class fobj>
|
||||
static inline void NerscChecksum(GridBase *grid, std::vector<fobj> &fbuf, uint32_t &nersc_csum)
|
||||
{
|
||||
const uint64_t size32 = sizeof(fobj)/sizeof(uint32_t);
|
||||
const uint64_t size32 = sizeof(fobj) / sizeof(uint32_t);
|
||||
|
||||
|
||||
uint64_t lsites =grid->lSites();
|
||||
if (fbuf.size()==1) {
|
||||
lsites=1;
|
||||
uint64_t lsites = grid->lSites();
|
||||
if (fbuf.size() == 1)
|
||||
{
|
||||
lsites = 1;
|
||||
}
|
||||
|
||||
#pragma omp parallel
|
||||
{
|
||||
uint32_t nersc_csum_thr=0;
|
||||
#pragma omp parallel
|
||||
{
|
||||
uint32_t nersc_csum_thr = 0;
|
||||
|
||||
#pragma omp for
|
||||
for(uint64_t local_site=0;local_site<lsites;local_site++){
|
||||
uint32_t * site_buf = (uint32_t *)&fbuf[local_site];
|
||||
for(uint64_t j=0;j<size32;j++){
|
||||
nersc_csum_thr=nersc_csum_thr+site_buf[j];
|
||||
}
|
||||
#pragma omp for
|
||||
for (uint64_t local_site = 0; local_site < lsites; local_site++)
|
||||
{
|
||||
uint32_t *site_buf = (uint32_t *)&fbuf[local_site];
|
||||
for (uint64_t j = 0; j < size32; j++)
|
||||
{
|
||||
nersc_csum_thr = nersc_csum_thr + site_buf[j];
|
||||
}
|
||||
}
|
||||
|
||||
#pragma omp critical
|
||||
#pragma omp critical
|
||||
{
|
||||
nersc_csum += nersc_csum_thr;
|
||||
nersc_csum += nersc_csum_thr;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
template<class fobj> static inline void ScidacChecksum(GridBase *grid,std::vector<fobj> &fbuf,uint32_t &scidac_csuma,uint32_t &scidac_csumb)
|
||||
{
|
||||
const uint64_t size32 = sizeof(fobj)/sizeof(uint32_t);
|
||||
@ -266,7 +270,7 @@ class BinaryIO {
|
||||
grid->Barrier();
|
||||
GridStopWatch timer;
|
||||
GridStopWatch bstimer;
|
||||
|
||||
|
||||
nersc_csum=0;
|
||||
scidac_csuma=0;
|
||||
scidac_csumb=0;
|
||||
@ -362,18 +366,22 @@ class BinaryIO {
|
||||
#else
|
||||
assert(0);
|
||||
#endif
|
||||
} else {
|
||||
std::cout<< GridLogMessage<< "C++ read I/O "<< file<<" : "
|
||||
<< iodata.size()*sizeof(fobj)<<" bytes"<<std::endl;
|
||||
std::ifstream fin;
|
||||
fin.open(file,std::ios::binary|std::ios::in);
|
||||
if ( control & BINARYIO_MASTER_APPEND ) {
|
||||
fin.seekg(-sizeof(fobj),fin.end);
|
||||
} else {
|
||||
fin.seekg(offset+myrank*lsites*sizeof(fobj));
|
||||
}
|
||||
fin.read((char *)&iodata[0],iodata.size()*sizeof(fobj));assert( fin.fail()==0);
|
||||
fin.close();
|
||||
} else {
|
||||
std::cout << GridLogMessage << "C++ read I/O " << file << " : "
|
||||
<< iodata.size() * sizeof(fobj) << " bytes" << std::endl;
|
||||
std::ifstream fin;
|
||||
fin.open(file, std::ios::binary | std::ios::in);
|
||||
if (control & BINARYIO_MASTER_APPEND)
|
||||
{
|
||||
fin.seekg(-sizeof(fobj), fin.end);
|
||||
}
|
||||
else
|
||||
{
|
||||
fin.seekg(offset + myrank * lsites * sizeof(fobj));
|
||||
}
|
||||
fin.read((char *)&iodata[0], iodata.size() * sizeof(fobj));
|
||||
assert(fin.fail() == 0);
|
||||
fin.close();
|
||||
}
|
||||
timer.Stop();
|
||||
|
||||
@ -405,30 +413,78 @@ class BinaryIO {
|
||||
timer.Start();
|
||||
if ( (control & BINARYIO_LEXICOGRAPHIC) && (nrank > 1) ) {
|
||||
#ifdef USE_MPI_IO
|
||||
std::cout<< GridLogMessage<< "MPI write I/O "<< file<< std::endl;
|
||||
ierr=MPI_File_open(grid->communicator,(char *) file.c_str(), MPI_MODE_RDWR|MPI_MODE_CREATE,MPI_INFO_NULL, &fh); assert(ierr==0);
|
||||
ierr=MPI_File_set_view(fh, disp, mpiObject, fileArray, "native", MPI_INFO_NULL); assert(ierr==0);
|
||||
ierr=MPI_File_write_all(fh, &iodata[0], 1, localArray, &status); assert(ierr==0);
|
||||
MPI_File_close(&fh);
|
||||
MPI_Type_free(&fileArray);
|
||||
MPI_Type_free(&localArray);
|
||||
std::cout << GridLogMessage << "MPI write I/O " << file << std::endl;
|
||||
ierr = MPI_File_open(grid->communicator, (char *)file.c_str(), MPI_MODE_RDWR | MPI_MODE_CREATE, MPI_INFO_NULL, &fh);
|
||||
std::cout << GridLogMessage << "Checking for errors" << std::endl;
|
||||
if (ierr != MPI_SUCCESS)
|
||||
{
|
||||
char error_string[BUFSIZ];
|
||||
int length_of_error_string, error_class;
|
||||
|
||||
MPI_Error_class(ierr, &error_class);
|
||||
MPI_Error_string(error_class, error_string, &length_of_error_string);
|
||||
fprintf(stderr, "%3d: %s\n", myrank, error_string);
|
||||
MPI_Error_string(ierr, error_string, &length_of_error_string);
|
||||
fprintf(stderr, "%3d: %s\n", myrank, error_string);
|
||||
MPI_Abort(MPI_COMM_WORLD, 1); //assert(ierr == 0);
|
||||
}
|
||||
|
||||
std::cout << GridLogDebug << "MPI read I/O set view " << file << std::endl;
|
||||
ierr = MPI_File_set_view(fh, disp, mpiObject, fileArray, "native", MPI_INFO_NULL);
|
||||
assert(ierr == 0);
|
||||
|
||||
std::cout << GridLogDebug << "MPI read I/O write all " << file << std::endl;
|
||||
ierr = MPI_File_write_all(fh, &iodata[0], 1, localArray, &status);
|
||||
assert(ierr == 0);
|
||||
|
||||
MPI_File_close(&fh);
|
||||
MPI_Type_free(&fileArray);
|
||||
MPI_Type_free(&localArray);
|
||||
#else
|
||||
assert(0);
|
||||
#endif
|
||||
} else {
|
||||
std::ofstream fout; fout.open(file,std::ios::binary|std::ios::out|std::ios::in);
|
||||
std::cout<< GridLogMessage<< "C++ write I/O "<< file<<" : "
|
||||
<< iodata.size()*sizeof(fobj)<<" bytes"<<std::endl;
|
||||
if ( control & BINARYIO_MASTER_APPEND ) {
|
||||
|
||||
std::ofstream fout;
|
||||
fout.exceptions ( std::fstream::failbit | std::fstream::badbit );
|
||||
try {
|
||||
fout.open(file,std::ios::binary|std::ios::out|std::ios::in);
|
||||
} catch (const std::fstream::failure& exc) {
|
||||
std::cout << GridLogError << "Error in opening the file " << file << " for output" <<std::endl;
|
||||
std::cout << GridLogError << "Exception description: " << exc.what() << std::endl;
|
||||
std::cout << GridLogError << "Probable cause: wrong path, inaccessible location "<< std::endl;
|
||||
#ifdef USE_MPI_IO
|
||||
MPI_Abort(MPI_COMM_WORLD,1);
|
||||
#else
|
||||
exit(1);
|
||||
#endif
|
||||
}
|
||||
std::cout << GridLogMessage<< "C++ write I/O "<< file<<" : "
|
||||
<< iodata.size()*sizeof(fobj)<<" bytes"<<std::endl;
|
||||
|
||||
if ( control & BINARYIO_MASTER_APPEND ) {
|
||||
fout.seekp(0,fout.end);
|
||||
} else {
|
||||
fout.seekp(offset+myrank*lsites*sizeof(fobj));
|
||||
}
|
||||
fout.write((char *)&iodata[0],iodata.size()*sizeof(fobj));assert( fout.fail()==0);
|
||||
|
||||
try {
|
||||
fout.write((char *)&iodata[0],iodata.size()*sizeof(fobj));//assert( fout.fail()==0);
|
||||
}
|
||||
catch (const std::fstream::failure& exc) {
|
||||
std::cout << "Exception in writing file " << file << std::endl;
|
||||
std::cout << GridLogError << "Exception description: "<< exc.what() << std::endl;
|
||||
#ifdef USE_MPI_IO
|
||||
MPI_Abort(MPI_COMM_WORLD,1);
|
||||
#else
|
||||
exit(1);
|
||||
#endif
|
||||
}
|
||||
|
||||
fout.close();
|
||||
}
|
||||
timer.Stop();
|
||||
}
|
||||
}
|
||||
timer.Stop();
|
||||
}
|
||||
|
||||
std::cout<<GridLogMessage<<"IOobject: ";
|
||||
if ( control & BINARYIO_READ) std::cout << " read ";
|
||||
@ -442,11 +498,14 @@ class BinaryIO {
|
||||
//////////////////////////////////////////////////////////////////////////////
|
||||
// Safety check
|
||||
//////////////////////////////////////////////////////////////////////////////
|
||||
grid->Barrier();
|
||||
grid->GlobalSum(nersc_csum);
|
||||
grid->GlobalXOR(scidac_csuma);
|
||||
grid->GlobalXOR(scidac_csumb);
|
||||
grid->Barrier();
|
||||
// if the data size is 1 we do not want to sum over the MPI ranks
|
||||
if (iodata.size() != 1){
|
||||
grid->Barrier();
|
||||
grid->GlobalSum(nersc_csum);
|
||||
grid->GlobalXOR(scidac_csuma);
|
||||
grid->GlobalXOR(scidac_csumb);
|
||||
grid->Barrier();
|
||||
}
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////
|
||||
@ -546,9 +605,9 @@ class BinaryIO {
|
||||
int gsites = grid->gSites();
|
||||
int lsites = grid->lSites();
|
||||
|
||||
uint32_t nersc_csum_tmp;
|
||||
uint32_t scidac_csuma_tmp;
|
||||
uint32_t scidac_csumb_tmp;
|
||||
uint32_t nersc_csum_tmp = 0;
|
||||
uint32_t scidac_csuma_tmp = 0;
|
||||
uint32_t scidac_csumb_tmp = 0;
|
||||
|
||||
GridStopWatch timer;
|
||||
|
||||
|
@ -165,7 +165,7 @@ class HMCResourceManager {
|
||||
// Grids
|
||||
//////////////////////////////////////////////////////////////
|
||||
|
||||
void AddGrid(std::string s, GridModule& M) {
|
||||
void AddGrid(const std::string s, GridModule& M) {
|
||||
// Check for name clashes
|
||||
auto search = Grids.find(s);
|
||||
if (search != Grids.end()) {
|
||||
@ -174,14 +174,24 @@ class HMCResourceManager {
|
||||
exit(1);
|
||||
}
|
||||
Grids[s] = std::move(M);
|
||||
std::cout << GridLogMessage << "::::::::::::::::::::::::::::::::::::::::" <<std::endl;
|
||||
std::cout << GridLogMessage << "HMCResourceManager:" << std::endl;
|
||||
std::cout << GridLogMessage << "Created grid set with name '" << s << "' and decomposition for the full cartesian " << std::endl;
|
||||
Grids[s].show_full_decomposition();
|
||||
std::cout << GridLogMessage << "::::::::::::::::::::::::::::::::::::::::" <<std::endl;
|
||||
}
|
||||
|
||||
// Add a named grid set, 4d shortcut
|
||||
void AddFourDimGrid(std::string s) {
|
||||
void AddFourDimGrid(const std::string s) {
|
||||
GridFourDimModule<vComplex> Mod;
|
||||
AddGrid(s, Mod);
|
||||
}
|
||||
|
||||
// Add a named grid set, 4d shortcut + tweak simd lanes
|
||||
void AddFourDimGrid(const std::string s, const std::vector<int> simd_decomposition) {
|
||||
GridFourDimModule<vComplex> Mod(simd_decomposition);
|
||||
AddGrid(s, Mod);
|
||||
}
|
||||
|
||||
|
||||
GridCartesian* GetCartesian(std::string s = "") {
|
||||
|
@ -33,28 +33,29 @@ directory
|
||||
namespace Grid {
|
||||
|
||||
// Resources
|
||||
// Modules for grids
|
||||
// Modules for grids
|
||||
|
||||
// Introduce another namespace HMCModules?
|
||||
|
||||
class GridModuleParameters: Serializable{
|
||||
class GridModuleParameters: Serializable{
|
||||
public:
|
||||
GRID_SERIALIZABLE_CLASS_MEMBERS(GridModuleParameters,
|
||||
std::string, lattice,
|
||||
std::string, mpi);
|
||||
|
||||
std::vector<int> getLattice(){return strToVec<int>(lattice);}
|
||||
std::vector<int> getMpi() {return strToVec<int>(mpi);}
|
||||
std::vector<int> getLattice() const {return strToVec<int>(lattice);}
|
||||
std::vector<int> getMpi() const {return strToVec<int>(mpi);}
|
||||
|
||||
void check(){
|
||||
if (getLattice().size() != getMpi().size()) {
|
||||
std::cout << GridLogError
|
||||
|
||||
void check() const {
|
||||
if (getLattice().size() != getMpi().size() ) {
|
||||
std::cout << GridLogError
|
||||
<< "Error in GridModuleParameters: lattice and mpi dimensions "
|
||||
"do not match"
|
||||
<< std::endl;
|
||||
exit(1);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
template <class ReaderClass>
|
||||
GridModuleParameters(Reader<ReaderClass>& Reader, std::string n = "LatticeGrid"):name(n) {
|
||||
@ -75,51 +76,94 @@ private:
|
||||
// Lower level class
|
||||
class GridModule {
|
||||
public:
|
||||
GridCartesian* get_full() {
|
||||
GridCartesian* get_full() {
|
||||
std::cout << GridLogDebug << "Getting cartesian in module"<< std::endl;
|
||||
return grid_.get(); }
|
||||
GridRedBlackCartesian* get_rb() {
|
||||
GridRedBlackCartesian* get_rb() {
|
||||
std::cout << GridLogDebug << "Getting rb-cartesian in module"<< std::endl;
|
||||
return rbgrid_.get(); }
|
||||
|
||||
void set_full(GridCartesian* grid) { grid_.reset(grid); }
|
||||
void set_rb(GridRedBlackCartesian* rbgrid) { rbgrid_.reset(rbgrid); }
|
||||
void show_full_decomposition(){ grid_->show_decomposition(); }
|
||||
void show_rb_decomposition(){ rbgrid_->show_decomposition(); }
|
||||
|
||||
protected:
|
||||
std::unique_ptr<GridCartesian> grid_;
|
||||
std::unique_ptr<GridRedBlackCartesian> rbgrid_;
|
||||
|
||||
|
||||
};
|
||||
|
||||
////////////////////////////////////
|
||||
// Classes for the user
|
||||
////////////////////////////////////
|
||||
// Note: the space time grid should be out of the QCD namespace
|
||||
template< class vector_type>
|
||||
class GridFourDimModule : public GridModule {
|
||||
public:
|
||||
GridFourDimModule() {
|
||||
template <class vector_type>
|
||||
class GridFourDimModule : public GridModule
|
||||
{
|
||||
public:
|
||||
GridFourDimModule()
|
||||
{
|
||||
using namespace QCD;
|
||||
set_full(SpaceTimeGrid::makeFourDimGrid(
|
||||
GridDefaultLatt(), GridDefaultSimd(4, vector_type::Nsimd()),
|
||||
GridDefaultLatt(),
|
||||
GridDefaultSimd(4, vector_type::Nsimd()),
|
||||
GridDefaultMpi()));
|
||||
set_rb(SpaceTimeGrid::makeFourDimRedBlackGrid(grid_.get()));
|
||||
}
|
||||
|
||||
GridFourDimModule(GridModuleParameters Params) {
|
||||
GridFourDimModule(const std::vector<int> tweak_simd)
|
||||
{
|
||||
using namespace QCD;
|
||||
if (tweak_simd.size() != 4)
|
||||
{
|
||||
std::cout << GridLogError
|
||||
<< "Error in GridFourDimModule: SIMD size different from 4"
|
||||
<< std::endl;
|
||||
exit(1);
|
||||
}
|
||||
|
||||
// Checks that the product agrees with the expectation
|
||||
int simd_sum = 1;
|
||||
for (auto &n : tweak_simd)
|
||||
simd_sum *= n;
|
||||
std::cout << GridLogDebug << "TweakSIMD: " << tweak_simd << " Sum: " << simd_sum << std::endl;
|
||||
|
||||
if (simd_sum == vector_type::Nsimd())
|
||||
{
|
||||
set_full(SpaceTimeGrid::makeFourDimGrid(
|
||||
GridDefaultLatt(),
|
||||
tweak_simd,
|
||||
GridDefaultMpi()));
|
||||
set_rb(SpaceTimeGrid::makeFourDimRedBlackGrid(grid_.get()));
|
||||
}
|
||||
else
|
||||
{
|
||||
std::cout << GridLogError
|
||||
<< "Error in GridFourDimModule: SIMD lanes must sum to "
|
||||
<< vector_type::Nsimd()
|
||||
<< std::endl;
|
||||
}
|
||||
}
|
||||
|
||||
GridFourDimModule(const GridModuleParameters Params)
|
||||
{
|
||||
using namespace QCD;
|
||||
Params.check();
|
||||
std::vector<int> lattice_v = Params.getLattice();
|
||||
std::vector<int> mpi_v = Params.getMpi();
|
||||
if (lattice_v.size() == 4) {
|
||||
if (lattice_v.size() == 4)
|
||||
{
|
||||
set_full(SpaceTimeGrid::makeFourDimGrid(
|
||||
lattice_v, GridDefaultSimd(4, vector_type::Nsimd()),
|
||||
lattice_v,
|
||||
GridDefaultSimd(4, vector_type::Nsimd()),
|
||||
mpi_v));
|
||||
set_rb(SpaceTimeGrid::makeFourDimRedBlackGrid(grid_.get()));
|
||||
} else {
|
||||
std::cout << GridLogError
|
||||
<< "Error in GridFourDimModule: lattice dimension different from 4"
|
||||
<< std::endl;
|
||||
}
|
||||
else
|
||||
{
|
||||
std::cout << GridLogError
|
||||
<< "Error in GridFourDimModule: lattice dimension different from 4"
|
||||
<< std::endl;
|
||||
exit(1);
|
||||
}
|
||||
}
|
||||
|
@ -381,7 +381,7 @@ void Grid_init(int *argc,char ***argv)
|
||||
std::cout << GridLogDebug << "Requesting "<< CartesianCommunicator::MAX_MPI_SHM_BYTES <<" byte stencil comms buffers "<<std::endl;
|
||||
|
||||
if( GridCmdOptionExists(*argv,*argv+*argc,"--decomposition") ){
|
||||
std::cout<<GridLogMessage<<"Grid Decomposition\n";
|
||||
std::cout<<GridLogMessage<<"Grid Default Decomposition patterns\n";
|
||||
std::cout<<GridLogMessage<<"\tOpenMP threads : "<<GridThread::GetThreads()<<std::endl;
|
||||
std::cout<<GridLogMessage<<"\tMPI tasks : "<<GridCmdVectorIntToString(GridDefaultMpi())<<std::endl;
|
||||
std::cout<<GridLogMessage<<"\tvRealF : "<<sizeof(vRealF)*8 <<"bits ; " <<GridCmdVectorIntToString(GridDefaultSimd(4,vRealF::Nsimd()))<<std::endl;
|
||||
|
Reference in New Issue
Block a user