portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-13 20:57:06 +01:00
Code Releases Activity

Merge branch 'develop' into feature/json-fix

Browse Source
This commit is contained in:
Guido Cossu
2017-09-08 13:42:20 +01:00
parent 7cb2b11f26 1184ed29ae
commit 13fa70ac1a
40 changed files with 2426 additions and 563 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
lib/Makefile.am
View File

@ -10,8 +10,8 @@ if BUILD_COMMS_MPI3
extra_sources+=communicator/Communicator_base.cc
endif
if BUILD_COMMS_MPI3L
extra_sources+=communicator/Communicator_mpi3_leader.cc
if BUILD_COMMS_MPIT
extra_sources+=communicator/Communicator_mpit.cc
extra_sources+=communicator/Communicator_base.cc
endif

29
lib/algorithms/iterative/BlockConjugateGradient.h
View File

@ -87,15 +87,22 @@ void ThinQRfact (Eigen::MatrixXcd &m_rr,
////////////////////////////////////////////////////////////////////////////////////////////////////
sliceInnerProductMatrix(m_rr,R,R,Orthog);
////////////////////////////////////////////////////////////////////////////////////////////////////
// Cholesky from Eigen
// There exists a ldlt that is documented as more stable
////////////////////////////////////////////////////////////////////////////////////////////////////
Eigen::MatrixXcd L = m_rr.llt().matrixL();
// Force manifest hermitian to avoid rounding related
m_rr = 0.5*(m_rr+m_rr.adjoint());
#if 0
std::cout << " Calling Cholesky ldlt on m_rr " << m_rr <<std::endl;
Eigen::MatrixXcd L_ldlt = m_rr.ldlt().matrixL();
std::cout << " Called Cholesky ldlt on m_rr " << L_ldlt <<std::endl;
auto D_ldlt = m_rr.ldlt().vectorD();
std::cout << " Called Cholesky ldlt on m_rr " << D_ldlt <<std::endl;
#endif
// std::cout << " Calling Cholesky llt on m_rr " <<std::endl;
Eigen::MatrixXcd L = m_rr.llt().matrixL();
// std::cout << " Called Cholesky llt on m_rr " << L <<std::endl;
C = L.adjoint();
Cinv = C.inverse();
////////////////////////////////////////////////////////////////////////////////////////////////////
// Q = R C^{-1}
//
@ -103,7 +110,6 @@ void ThinQRfact (Eigen::MatrixXcd &m_rr,
//
// NB maddMatrix conventions are Right multiplication X[j] a[j,i] already
////////////////////////////////////////////////////////////////////////////////////////////////////
// FIXME:: make a sliceMulMatrix to avoid zero vector
sliceMulMatrix(Q,Cinv,R,Orthog);
}
////////////////////////////////////////////////////////////////////////////////////////////////////
@ -199,7 +205,12 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
Linop.HermOp(X, AD);
tmp = B - AD;
//std::cout << GridLogMessage << " initial tmp " << norm2(tmp)<< std::endl;
ThinQRfact (m_rr, m_C, m_Cinv, Q, tmp);
//std::cout << GridLogMessage << " initial Q " << norm2(Q)<< std::endl;
//std::cout << GridLogMessage << " m_rr " << m_rr<<std::endl;
//std::cout << GridLogMessage << " m_C " << m_C<<std::endl;
//std::cout << GridLogMessage << " m_Cinv " << m_Cinv<<std::endl;
D=Q;
std::cout << GridLogMessage<<"BlockCGrQ computed initial residual and QR fact " <<std::endl;
@ -221,13 +232,15 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
MatrixTimer.Start();
Linop.HermOp(D, Z);
MatrixTimer.Stop();
//std::cout << GridLogMessage << " norm2 Z " <<norm2(Z)<<std::endl;
//4. M = [D^dag Z]^{-1}
sliceInnerTimer.Start();
sliceInnerProductMatrix(m_DZ,D,Z,Orthog);
sliceInnerTimer.Stop();
m_M = m_DZ.inverse();
//std::cout << GridLogMessage << " m_DZ " <<m_DZ<<std::endl;
//5. X = X + D MC
m_tmp = m_M * m_C;
sliceMaddTimer.Start();

39
lib/allocator/AlignedAllocator.cc
View File

@ -1,7 +1,5 @@
#include <Grid/GridCore.h>
#include <fcntl.h>
namespace Grid {
@ -11,7 +9,7 @@ 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);
@ -63,4 +61,37 @@ void *PointerCache::Lookup(size_t bytes) {
return NULL;
}
void check_huge_pages(void *Buf,uint64_t BYTES)
{
#ifdef __linux__
int fd = open("/proc/self/pagemap", O_RDONLY);
assert(fd >= 0);
const int page_size = 4096;
uint64_t virt_pfn = (uint64_t)Buf / page_size;
off_t offset = sizeof(uint64_t) * virt_pfn;
uint64_t npages = (BYTES + page_size-1) / page_size;
uint64_t pagedata[npages];
uint64_t ret = lseek(fd, offset, SEEK_SET);
assert(ret == offset);
ret = ::read(fd, pagedata, sizeof(uint64_t)*npages);
assert(ret == sizeof(uint64_t) * npages);
int nhugepages = npages / 512;
int n4ktotal, nnothuge;
n4ktotal = 0;
nnothuge = 0;
for (int i = 0; i < nhugepages; ++i) {
uint64_t baseaddr = (pagedata[i*512] & 0x7fffffffffffffULL) * page_size;
for (int j = 0; j < 512; ++j) {
uint64_t pageaddr = (pagedata[i*512+j] & 0x7fffffffffffffULL) * page_size;
++n4ktotal;
if (pageaddr != baseaddr + j * page_size)
++nnothuge;
}
}
int rank = CartesianCommunicator::RankWorld();
printf("rank %d Allocated %d 4k pages, %d not in huge pages\n", rank, n4ktotal, nnothuge);
#endif
}
}

27
lib/allocator/AlignedAllocator.h
View File

@ -64,6 +64,8 @@ namespace Grid {
};
void check_huge_pages(void *Buf,uint64_t BYTES);
////////////////////////////////////////////////////////////////////
// A lattice of something, but assume the something is SIMDized.
////////////////////////////////////////////////////////////////////
@ -92,12 +94,20 @@ public:
size_type bytes = __n*sizeof(_Tp);
_Tp *ptr = (_Tp *) PointerCache::Lookup(bytes);
// if ( ptr != NULL )
// std::cout << "alignedAllocator "<<__n << " cache hit "<< std::hex << ptr <<std::dec <<std::endl;
//////////////////
// Hack 2MB align; could make option probably doesn't need configurability
//////////////////
//define GRID_ALLOC_ALIGN (128)
#define GRID_ALLOC_ALIGN (2*1024*1024)
#ifdef HAVE_MM_MALLOC_H
if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) _mm_malloc(bytes,128);
if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) _mm_malloc(bytes,GRID_ALLOC_ALIGN);
#else
if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) memalign(128,bytes);
if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) memalign(GRID_ALLOC_ALIGN,bytes);
#endif
// std::cout << "alignedAllocator " << std::hex << ptr <<std::dec <<std::endl;
// First touch optimise in threaded loop
uint8_t *cp = (uint8_t *)ptr;
#ifdef GRID_OMP
@ -111,6 +121,7 @@ public:
void deallocate(pointer __p, size_type __n) {
size_type bytes = __n * sizeof(_Tp);
pointer __freeme = (pointer)PointerCache::Insert((void *)__p,bytes);
#ifdef HAVE_MM_MALLOC_H
@ -189,16 +200,18 @@ public:
pointer allocate(size_type __n, const void* _p= 0)
{
#ifdef HAVE_MM_MALLOC_H
_Tp * ptr = (_Tp *) _mm_malloc(__n*sizeof(_Tp),128);
_Tp * ptr = (_Tp *) _mm_malloc(__n*sizeof(_Tp),GRID_ALLOC_ALIGN);
#else
_Tp * ptr = (_Tp *) memalign(128,__n*sizeof(_Tp));
_Tp * ptr = (_Tp *) memalign(GRID_ALLOC_ALIGN,__n*sizeof(_Tp));
#endif
size_type bytes = __n*sizeof(_Tp);
uint8_t *cp = (uint8_t *)ptr;
if ( ptr ) {
// One touch per 4k page, static OMP loop to catch same loop order
#pragma omp parallel for schedule(static)
for(size_type n=0;n<bytes;n+=4096){
cp[n]=0;
for(size_type n=0;n<bytes;n+=4096){
cp[n]=0;
}
}
return ptr;
}

23
lib/cartesian/Cartesian_base.h
View File

@ -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;
}
////////////////////////////////////////////////////////////////

130
lib/cartesian/Cartesian_full.h
View File

@ -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

191
lib/cartesian/Cartesian_red_black.h
View File

@ -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

64
lib/communicator/Communicator_base.cc
View File

@ -26,6 +26,10 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
*************************************************************************************/
/* END LEGAL */
#include <Grid/GridCore.h>
#include <fcntl.h>
#include <unistd.h>
#include <limits.h>
#include <sys/mman.h>
namespace Grid {
@ -33,8 +37,11 @@ namespace Grid {
// Info that is setup once and indept of cartesian layout
///////////////////////////////////////////////////////////////
void * CartesianCommunicator::ShmCommBuf;
uint64_t CartesianCommunicator::MAX_MPI_SHM_BYTES = 128*1024*1024;
CartesianCommunicator::CommunicatorPolicy_t CartesianCommunicator::CommunicatorPolicy= CartesianCommunicator::CommunicatorPolicyConcurrent;
uint64_t CartesianCommunicator::MAX_MPI_SHM_BYTES = 1024LL*1024LL*1024LL;
CartesianCommunicator::CommunicatorPolicy_t
CartesianCommunicator::CommunicatorPolicy= CartesianCommunicator::CommunicatorPolicyConcurrent;
int CartesianCommunicator::nCommThreads = -1;
int CartesianCommunicator::Hugepages = 0;
/////////////////////////////////
// Alloc, free shmem region
@ -89,25 +96,43 @@ void CartesianCommunicator::GlobalSumVector(ComplexD *c,int N)
GlobalSumVector((double *)c,2*N);
}
#if !defined( GRID_COMMS_MPI3) && !defined (GRID_COMMS_MPI3L)
#if !defined( GRID_COMMS_MPI3)
int CartesianCommunicator::NodeCount(void) { return ProcessorCount();};
int CartesianCommunicator::RankCount(void) { return ProcessorCount();};
double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void *xmit,
int xmit_to_rank,
void *recv,
int recv_from_rank,
int bytes)
#endif
#if !defined( GRID_COMMS_MPI3) && !defined (GRID_COMMS_MPIT)
double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
int xmit_to_rank,
void *recv,
int recv_from_rank,
int bytes, int dir)
{
std::vector<CommsRequest_t> list;
// Discard the "dir"
SendToRecvFromBegin (list,xmit,xmit_to_rank,recv,recv_from_rank,bytes);
SendToRecvFromComplete(list);
return 2.0*bytes;
}
double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void *xmit,
int xmit_to_rank,
void *recv,
int recv_from_rank,
int bytes, int dir)
{
// Discard the "dir"
SendToRecvFromBegin(list,xmit,xmit_to_rank,recv,recv_from_rank,bytes);
return 2.0*bytes;
}
void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall)
void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
{
SendToRecvFromComplete(waitall);
}
#endif
#if !defined( GRID_COMMS_MPI3)
void CartesianCommunicator::StencilBarrier(void){};
commVector<uint8_t> CartesianCommunicator::ShmBufStorageVector;
@ -121,8 +146,25 @@ void *CartesianCommunicator::ShmBufferTranslate(int rank,void * local_p) {
return NULL;
}
void CartesianCommunicator::ShmInitGeneric(void){
#if 1
int mmap_flag = MAP_SHARED | MAP_ANONYMOUS;
#ifdef MAP_HUGETLB
if ( Hugepages ) mmap_flag |= MAP_HUGETLB;
#endif
ShmCommBuf =(void *) mmap(NULL, MAX_MPI_SHM_BYTES, PROT_READ | PROT_WRITE, mmap_flag, -1, 0);
if (ShmCommBuf == (void *)MAP_FAILED) {
perror("mmap failed ");
exit(EXIT_FAILURE);
}
#ifdef MADV_HUGEPAGE
if (!Hugepages ) madvise(ShmCommBuf,MAX_MPI_SHM_BYTES,MADV_HUGEPAGE);
#endif
#else
ShmBufStorageVector.resize(MAX_MPI_SHM_BYTES);
ShmCommBuf=(void *)&ShmBufStorageVector[0];
#endif
bzero(ShmCommBuf,MAX_MPI_SHM_BYTES);
}
#endif

55
lib/communicator/Communicator_base.h
View File

@ -38,7 +38,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#ifdef GRID_COMMS_MPI3
#include <mpi.h>
#endif
#ifdef GRID_COMMS_MPI3L
#ifdef GRID_COMMS_MPIT
#include <mpi.h>
#endif
#ifdef GRID_COMMS_SHMEM
@ -50,12 +50,24 @@ namespace Grid {
class CartesianCommunicator {
public:
// 65536 ranks per node adequate for now
////////////////////////////////////////////
// Isend/Irecv/Wait, or Sendrecv blocking
////////////////////////////////////////////
enum CommunicatorPolicy_t { CommunicatorPolicyConcurrent, CommunicatorPolicySequential };
static CommunicatorPolicy_t CommunicatorPolicy;
static void SetCommunicatorPolicy(CommunicatorPolicy_t policy ) { CommunicatorPolicy = policy; }
///////////////////////////////////////////
// Up to 65536 ranks per node adequate for now
// 128MB shared memory for comms enought for 48^4 local vol comms
// Give external control (command line override?) of this
static const int MAXLOG2RANKSPERNODE = 16;
static uint64_t MAX_MPI_SHM_BYTES;
///////////////////////////////////////////
static const int MAXLOG2RANKSPERNODE = 16;
static uint64_t MAX_MPI_SHM_BYTES;
static int nCommThreads;
// use explicit huge pages
static int Hugepages;
// Communicator should know nothing of the physics grid, only processor grid.
int _Nprocessors; // How many in all
@ -64,14 +76,18 @@ class CartesianCommunicator {
std::vector<int> _processor_coor; // linear processor coordinate
unsigned long _ndimension;
#if defined (GRID_COMMS_MPI) || defined (GRID_COMMS_MPI3) || defined (GRID_COMMS_MPI3L)
#if defined (GRID_COMMS_MPI) || defined (GRID_COMMS_MPI3) || defined (GRID_COMMS_MPIT)
static MPI_Comm communicator_world;
MPI_Comm communicator;
MPI_Comm communicator;
std::vector<MPI_Comm> communicator_halo;
typedef MPI_Request CommsRequest_t;
#else
typedef int CommsRequest_t;
#endif
////////////////////////////////////////////////////////////////////
// Helper functionality for SHM Windows common to all other impls
////////////////////////////////////////////////////////////////////
@ -117,11 +133,7 @@ class CartesianCommunicator {
/////////////////////////////////
static void * ShmCommBuf;
// Isend/Irecv/Wait, or Sendrecv blocking
enum CommunicatorPolicy_t { CommunicatorPolicyConcurrent, CommunicatorPolicySequential };
static CommunicatorPolicy_t CommunicatorPolicy;
static void SetCommunicatorPolicy(CommunicatorPolicy_t policy ) { CommunicatorPolicy = policy; }
size_t heap_top;
size_t heap_bytes;
@ -211,14 +223,21 @@ class CartesianCommunicator {
void SendToRecvFromComplete(std::vector<CommsRequest_t> &waitall);
double StencilSendToRecvFrom(void *xmit,
int xmit_to_rank,
void *recv,
int recv_from_rank,
int bytes,int dir);
double StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void *xmit,
int xmit_to_rank,
void *recv,
int recv_from_rank,
int bytes);
void *xmit,
int xmit_to_rank,
void *recv,
int recv_from_rank,
int bytes,int dir);
void StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall);
void StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int i);
void StencilBarrier(void);
////////////////////////////////////////////////////////////

125
lib/communicator/Communicator_mpi3.cc
View File

@ -41,9 +41,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#ifdef HAVE_NUMAIF_H
#include <numaif.h>
#endif
#ifndef SHM_HUGETLB
#define SHM_HUGETLB 04000
#endif
namespace Grid {
@ -200,7 +198,46 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
ShmCommBuf = 0;
ShmCommBufs.resize(ShmSize);
#if 1
////////////////////////////////////////////////////////////////////////////////////////////
// Hugetlbf and others map filesystems as mappable huge pages
////////////////////////////////////////////////////////////////////////////////////////////
#ifdef GRID_MPI3_SHMMMAP
char shm_name [NAME_MAX];
for(int r=0;r<ShmSize;r++){
size_t size = CartesianCommunicator::MAX_MPI_SHM_BYTES;
sprintf(shm_name,GRID_SHM_PATH "/Grid_mpi3_shm_%d_%d",GroupRank,r);
//sprintf(shm_name,"/var/lib/hugetlbfs/group/wheel/pagesize-2MB/" "Grid_mpi3_shm_%d_%d",GroupRank,r);
// printf("Opening file %s \n",shm_name);
int fd=open(shm_name,O_RDWR|O_CREAT,0666);
if ( fd == -1) {
printf("open %s failed\n",shm_name);
perror("open hugetlbfs");
exit(0);
}
int mmap_flag = MAP_SHARED ;
#ifdef MAP_POPULATE
mmap_flag|=MAP_POPULATE;
#endif
#ifdef MAP_HUGETLB
if ( Hugepages ) mmap_flag |= MAP_HUGETLB;
#endif
void *ptr = (void *) mmap(NULL, MAX_MPI_SHM_BYTES, PROT_READ | PROT_WRITE, mmap_flag,fd, 0);
if ( ptr == (void *)MAP_FAILED ) {
printf("mmap %s failed\n",shm_name);
perror("failed mmap"); assert(0);
}
assert(((uint64_t)ptr&0x3F)==0);
ShmCommBufs[r] =ptr;
}
#endif
////////////////////////////////////////////////////////////////////////////////////////////
// POSIX SHMOPEN ; as far as I know Linux does not allow EXPLICIT HugePages with this case
// tmpfs (Larry Meadows says) does not support explicit huge page, and this is used for
// the posix shm virtual file system
////////////////////////////////////////////////////////////////////////////////////////////
#ifdef GRID_MPI3_SHMOPEN
char shm_name [NAME_MAX];
if ( ShmRank == 0 ) {
for(int r=0;r<ShmSize;r++){
@ -213,13 +250,22 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
int fd=shm_open(shm_name,O_RDWR|O_CREAT,0666);
if ( fd < 0 ) { perror("failed shm_open"); assert(0); }
ftruncate(fd, size);
int mmap_flag = MAP_SHARED;
#ifdef MAP_POPULATE
mmap_flag |= MAP_POPULATE;
#endif
#ifdef MAP_HUGETLB
if (Hugepages) mmap_flag |= MAP_HUGETLB;
#endif
void * ptr = mmap(NULL,size, PROT_READ | PROT_WRITE, mmap_flag, fd, 0);
void * ptr = mmap(NULL,size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
if ( ptr == MAP_FAILED ) { perror("failed mmap"); assert(0); }
if ( ptr == (void * )MAP_FAILED ) { perror("failed mmap"); assert(0); }
assert(((uint64_t)ptr&0x3F)==0);
// Try to force numa domain on the shm segment if we have numaif.h
#ifdef HAVE_NUMAIF_H
// Experiments; Experiments; Try to force numa domain on the shm segment if we have numaif.h
#if 0
//#ifdef HAVE_NUMAIF_H
int status;
int flags=MPOL_MF_MOVE;
#ifdef KNL
@ -236,7 +282,7 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
if (ierr && (page==0)) perror("numa relocate command failed");
}
#endif
ShmCommBufs[r] =ptr;
ShmCommBufs[r] =ptr;
}
}
@ -258,21 +304,32 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
ShmCommBufs[r] =ptr;
}
}
#else
#endif
////////////////////////////////////////////////////////////////////////////////////////////
// SHMGET SHMAT and SHM_HUGETLB flag
////////////////////////////////////////////////////////////////////////////////////////////
#ifdef GRID_MPI3_SHMGET
std::vector<int> shmids(ShmSize);
if ( ShmRank == 0 ) {
for(int r=0;r<ShmSize;r++){
size_t size = CartesianCommunicator::MAX_MPI_SHM_BYTES;
key_t key = 0x4545 + r;
if ((shmids[r]= shmget(key,size, SHM_HUGETLB | IPC_CREAT | SHM_R | SHM_W)) < 0) {
key_t key = IPC_PRIVATE;
int flags = IPC_CREAT | SHM_R | SHM_W;
#ifdef SHM_HUGETLB
if (Hugepages) flags|=SHM_HUGETLB;
#endif
if ((shmids[r]= shmget(key,size, flags)) ==-1) {
int errsv = errno;
printf("Errno %d\n",errsv);
printf("key %d\n",key);
printf("size %lld\n",size);
printf("flags %d\n",flags);
perror("shmget");
exit(1);
} else {
printf("shmid: 0x%x\n", shmids[r]);
}
printf("shmid: 0x%x\n", shmids[r]);
}
}
MPI_Barrier(ShmComm);
@ -397,8 +454,14 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
{
int ierr;
communicator=communicator_world;
_ndimension = processors.size();
communicator_halo.resize (2*_ndimension);
for(int i=0;i<_ndimension*2;i++){
MPI_Comm_dup(communicator,&communicator_halo[i]);
}
////////////////////////////////////////////////////////////////
// Assert power of two shm_size.
////////////////////////////////////////////////////////////////
@ -621,13 +684,27 @@ void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &lis
}
}
double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void *xmit,
int dest,
void *recv,
int from,
int bytes)
double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
int dest,
void *recv,
int from,
int bytes,int dir)
{
std::vector<CommsRequest_t> list;
double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,recv,from,bytes,dir);
StencilSendToRecvFromComplete(list,dir);
return offbytes;
}
double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void *xmit,
int dest,
void *recv,
int from,
int bytes,int dir)
{
assert(dir < communicator_halo.size());
MPI_Request xrq;
MPI_Request rrq;
@ -646,26 +723,26 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
gfrom = MPI_UNDEFINED;
#endif
if ( gfrom ==MPI_UNDEFINED) {
ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator_halo[dir],&rrq);
assert(ierr==0);
list.push_back(rrq);
off_node_bytes+=bytes;
}
if ( gdest == MPI_UNDEFINED ) {
ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator_halo[dir],&xrq);
assert(ierr==0);
list.push_back(xrq);
off_node_bytes+=bytes;
}
if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
this->StencilSendToRecvFromComplete(list);
this->StencilSendToRecvFromComplete(list,dir);
}
return off_node_bytes;
}
void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall)
void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
{
SendToRecvFromComplete(waitall);
}

286
lib/communicator/Communicator_mpit.cc Normal file
View File

@ -0,0 +1,286 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/communicator/Communicator_mpi.cc
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/GridCore.h>
#include <Grid/GridQCDcore.h>
#include <Grid/qcd/action/ActionCore.h>
#include <mpi.h>
namespace Grid {
///////////////////////////////////////////////////////////////////////////////////////////////////
// Info that is setup once and indept of cartesian layout
///////////////////////////////////////////////////////////////////////////////////////////////////
MPI_Comm CartesianCommunicator::communicator_world;
// Should error check all MPI calls.
void CartesianCommunicator::Init(int *argc, char ***argv) {
int flag;
int provided;
MPI_Initialized(&flag); // needed to coexist with other libs apparently
if ( !flag ) {
MPI_Init_thread(argc,argv,MPI_THREAD_MULTIPLE,&provided);
if ( provided != MPI_THREAD_MULTIPLE ) {
QCD::WilsonKernelsStatic::Comms = QCD::WilsonKernelsStatic::CommsThenCompute;
}
}
MPI_Comm_dup (MPI_COMM_WORLD,&communicator_world);
ShmInitGeneric();
}
CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
{
_ndimension = processors.size();
std::vector<int> periodic(_ndimension,1);
_Nprocessors=1;
_processors = processors;
_processor_coor.resize(_ndimension);
MPI_Cart_create(communicator_world, _ndimension,&_processors[0],&periodic[0],1,&communicator);
MPI_Comm_rank(communicator,&_processor);
MPI_Cart_coords(communicator,_processor,_ndimension,&_processor_coor[0]);
for(int i=0;i<_ndimension;i++){
_Nprocessors*=_processors[i];
}
communicator_halo.resize (2*_ndimension);
for(int i=0;i<_ndimension*2;i++){
MPI_Comm_dup(communicator,&communicator_halo[i]);
}
int Size;
MPI_Comm_size(communicator,&Size);
assert(Size==_Nprocessors);
}
void CartesianCommunicator::GlobalSum(uint32_t &u){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalSum(uint64_t &u){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalXOR(uint32_t &u){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_BXOR,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalXOR(uint64_t &u){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_BXOR,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalSum(float &f){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalSumVector(float *f,int N)
{
int ierr=MPI_Allreduce(MPI_IN_PLACE,f,N,MPI_FLOAT,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalSum(double &d)
{
int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalSumVector(double *d,int N)
{
int ierr = MPI_Allreduce(MPI_IN_PLACE,d,N,MPI_DOUBLE,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
{
int ierr=MPI_Cart_shift(communicator,dim,shift,&source,&dest);
assert(ierr==0);
}
int CartesianCommunicator::RankFromProcessorCoor(std::vector<int> &coor)
{
int rank;
int ierr=MPI_Cart_rank (communicator, &coor[0], &rank);
assert(ierr==0);
return rank;
}
void CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor)
{
coor.resize(_ndimension);
int ierr=MPI_Cart_coords (communicator, rank, _ndimension,&coor[0]);
assert(ierr==0);
}
// Basic Halo comms primitive
void CartesianCommunicator::SendToRecvFrom(void *xmit,
int dest,
void *recv,
int from,
int bytes)
{
std::vector<CommsRequest_t> reqs(0);
SendToRecvFromBegin(reqs,xmit,dest,recv,from,bytes);
SendToRecvFromComplete(reqs);
}
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)
{
int myrank = _processor;
int ierr;
if ( CommunicatorPolicy == CommunicatorPolicyConcurrent ) {
MPI_Request xrq;
MPI_Request rrq;
ierr =MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
ierr|=MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
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);
}
}
void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
{
if ( CommunicatorPolicy == CommunicatorPolicyConcurrent ) {
int nreq=list.size();
std::vector<MPI_Status> status(nreq);
int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
assert(ierr==0);
}
}
void CartesianCommunicator::Barrier(void)
{
int ierr = MPI_Barrier(communicator);
assert(ierr==0);
}
void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
{
int ierr=MPI_Bcast(data,
bytes,
MPI_BYTE,
root,
communicator);
assert(ierr==0);
}
///////////////////////////////////////////////////////
// Should only be used prior to Grid Init finished.
// Check for this?
///////////////////////////////////////////////////////
int CartesianCommunicator::RankWorld(void){
int r;
MPI_Comm_rank(communicator_world,&r);
return r;
}
void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
{
int ierr= MPI_Bcast(data,
bytes,
MPI_BYTE,
root,
communicator_world);
assert(ierr==0);
}
double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void *xmit,
int xmit_to_rank,
void *recv,
int recv_from_rank,
int bytes,int dir)
{
int myrank = _processor;
int ierr;
assert(dir < communicator_halo.size());
// std::cout << " sending on communicator "<<dir<<" " <<communicator_halo[dir]<<std::endl;
// Give the CPU to MPI immediately; can use threads to overlap optionally
MPI_Request req[2];
MPI_Irecv(recv,bytes,MPI_CHAR,recv_from_rank,recv_from_rank, communicator_halo[dir],&req[1]);
MPI_Isend(xmit,bytes,MPI_CHAR,xmit_to_rank ,myrank , communicator_halo[dir],&req[0]);
list.push_back(req[0]);
list.push_back(req[1]);
return 2.0*bytes;
}
void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
{
int nreq=waitall.size();
MPI_Waitall(nreq, &waitall[0], MPI_STATUSES_IGNORE);
};
double CartesianCommunicator::StencilSendToRecvFrom(void *xmit,
int xmit_to_rank,
void *recv,
int recv_from_rank,
int bytes,int dir)
{
int myrank = _processor;
int ierr;
assert(dir < communicator_halo.size());
// std::cout << " sending on communicator "<<dir<<" " <<communicator_halo[dir]<<std::endl;
// Give the CPU to MPI immediately; can use threads to overlap optionally
MPI_Request req[2];
MPI_Irecv(recv,bytes,MPI_CHAR,recv_from_rank,recv_from_rank, communicator_halo[dir],&req[1]);
MPI_Isend(xmit,bytes,MPI_CHAR,xmit_to_rank ,myrank , communicator_halo[dir],&req[0]);
MPI_Waitall(2, req, MPI_STATUSES_IGNORE);
return 2.0*bytes;
}
}

2
lib/cshift/Cshift.h
View File

@ -42,7 +42,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#include <Grid/cshift/Cshift_mpi.h>
#endif
#ifdef GRID_COMMS_MPI3L
#ifdef GRID_COMMS_MPIT
#include <Grid/cshift/Cshift_mpi.h>
#endif

38
lib/lattice/Lattice_reduction.h
View File

@ -369,6 +369,7 @@ static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice
}
};
/*
inline GridBase *makeSubSliceGrid(const GridBase *BlockSolverGrid,int Orthog)
{
int NN = BlockSolverGrid->_ndimension;
@ -387,6 +388,7 @@ inline GridBase *makeSubSliceGrid(const GridBase *BlockSolverGrid,int Or
}
return (GridBase *)new GridCartesian(latt_phys,simd_phys,mpi_phys);
}
*/
template<class vobj>
static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,const Lattice<vobj> &Y,int Orthog,RealD scale=1.0)
@ -398,14 +400,15 @@ static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice
int Nblock = X._grid->GlobalDimensions()[Orthog];
GridBase *FullGrid = X._grid;
GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
// GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
Lattice<vobj> Xslice(SliceGrid);
Lattice<vobj> Rslice(SliceGrid);
// Lattice<vobj> Xslice(SliceGrid);
// Lattice<vobj> Rslice(SliceGrid);
assert( FullGrid->_simd_layout[Orthog]==1);
int nh = FullGrid->_ndimension;
int nl = SliceGrid->_ndimension;
// int nl = SliceGrid->_ndimension;
int nl = nh-1;
//FIXME package in a convenient iterator
//Should loop over a plane orthogonal to direction "Orthog"
@ -448,14 +451,14 @@ static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<
int Nblock = X._grid->GlobalDimensions()[Orthog];
GridBase *FullGrid = X._grid;
GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
Lattice<vobj> Xslice(SliceGrid);
Lattice<vobj> Rslice(SliceGrid);
// GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
// Lattice<vobj> Xslice(SliceGrid);
// Lattice<vobj> Rslice(SliceGrid);
assert( FullGrid->_simd_layout[Orthog]==1);
int nh = FullGrid->_ndimension;
int nl = SliceGrid->_ndimension;
// int nl = SliceGrid->_ndimension;
int nl=1;
//FIXME package in a convenient iterator
//Should loop over a plane orthogonal to direction "Orthog"
@ -498,18 +501,19 @@ static void sliceInnerProductMatrix( Eigen::MatrixXcd &mat, const Lattice<vobj>
typedef typename vobj::vector_type vector_type;
GridBase *FullGrid = lhs._grid;
GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
// GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
int Nblock = FullGrid->GlobalDimensions()[Orthog];
Lattice<vobj> Lslice(SliceGrid);
Lattice<vobj> Rslice(SliceGrid);
// Lattice<vobj> Lslice(SliceGrid);
// Lattice<vobj> Rslice(SliceGrid);
mat = Eigen::MatrixXcd::Zero(Nblock,Nblock);
assert( FullGrid->_simd_layout[Orthog]==1);
int nh = FullGrid->_ndimension;
int nl = SliceGrid->_ndimension;
// int nl = SliceGrid->_ndimension;
int nl = nh-1;
//FIXME package in a convenient iterator
//Should loop over a plane orthogonal to direction "Orthog"
@ -550,6 +554,14 @@ static void sliceInnerProductMatrix( Eigen::MatrixXcd &mat, const Lattice<vobj>
mat += mat_thread;
}
}
for(int i=0;i<Nblock;i++){
for(int j=0;j<Nblock;j++){
ComplexD sum = mat(i,j);
FullGrid->GlobalSum(sum);
mat(i,j)=sum;
}}
return;
}

2
lib/log/Log.cc
View File

@ -95,7 +95,7 @@ void GridLogConfigure(std::vector<std::string> &logstreams) {
////////////////////////////////////////////////////////////
void Grid_quiesce_nodes(void) {
int me = 0;
#if defined(GRID_COMMS_MPI) || defined(GRID_COMMS_MPI3) || defined(GRID_COMMS_MPI3L)
#if defined(GRID_COMMS_MPI) || defined(GRID_COMMS_MPI3) || defined(GRID_COMMS_MPIT)
MPI_Comm_rank(MPI_COMM_WORLD, &me);
#endif
#ifdef GRID_COMMS_SHMEM

169
lib/parallelIO/BinaryIO.h
View File

@ -29,7 +29,7 @@
#ifndef GRID_BINARY_IO_H
#define GRID_BINARY_IO_H
#if defined(GRID_COMMS_MPI) || defined(GRID_COMMS_MPI3)
#if defined(GRID_COMMS_MPI) || defined(GRID_COMMS_MPI3) || defined(GRID_COMMS_MPIT)
#define USE_MPI_IO
#else
#undef USE_MPI_IO
@ -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;

14
lib/qcd/action/fermion/CayleyFermion5D.cc
View File

@ -414,7 +414,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,std::vector<Co
for(int i=0; i < Ls; i++){
as[i] = 1.0;
omega[i] = gamma[i]*zolo_hi; //NB reciprocal relative to Chroma NEF code
// assert(fabs(omega[i])>0.0);
assert(omega[i]!=Coeff_t(0.0));
bs[i] = 0.5*(bpc/omega[i] + bmc);
cs[i] = 0.5*(bpc/omega[i] - bmc);
}
@ -429,7 +429,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,std::vector<Co
for(int i=0;i<Ls;i++){
bee[i]=as[i]*(bs[i]*(4.0-this->M5) +1.0);
// assert(fabs(bee[i])>0.0);
assert(bee[i]!=Coeff_t(0.0));
cee[i]=as[i]*(1.0-cs[i]*(4.0-this->M5));
beo[i]=as[i]*bs[i];
ceo[i]=-as[i]*cs[i];
@ -455,11 +455,17 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,std::vector<Co
dee[i] = bee[i];
if ( i < Ls-1 ) {
assert(bee[i]!=Coeff_t(0.0));
assert(bee[0]!=Coeff_t(0.0));
lee[i] =-cee[i+1]/bee[i]; // sub-diag entry on the ith column
leem[i]=mass*cee[Ls-1]/bee[0];
for(int j=0;j<i;j++) leem[i]*= aee[j]/bee[j+1];
for(int j=0;j<i;j++) {
assert(bee[j+1]!=Coeff_t(0.0));
leem[i]*= aee[j]/bee[j+1];
}
uee[i] =-aee[i]/bee[i]; // up-diag entry on the ith row
@ -478,7 +484,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,std::vector<Co
{
Coeff_t delta_d=mass*cee[Ls-1];
for(int j=0;j<Ls-1;j++) {
// assert(fabs(bee[j])>0.0);
assert(bee[j] != Coeff_t(0.0));
delta_d *= cee[j]/bee[j];
}
dee[Ls-1] += delta_d;

60
lib/qcd/action/fermion/ImprovedStaggeredFermion5D.cc
View File

@ -230,8 +230,15 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOr
{
Compressor compressor;
int LLs = in._grid->_rdimensions[0];
DhopTotalTime -= usecond();
DhopCommTime -= usecond();
st.HaloExchange(in,compressor);
DhopCommTime += usecond();
DhopComputeTime -= usecond();
// Dhop takes the 4d grid from U, and makes a 5d index for fermion
if (dag == DaggerYes) {
parallel_for (int ss = 0; ss < U._grid->oSites(); ss++) {
@ -244,12 +251,15 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOr
Kernels::DhopSite(st,lo,U,UUU,st.CommBuf(),LLs,sU,in,out);
}
}
DhopComputeTime += usecond();
DhopTotalTime += usecond();
}
template<class Impl>
void ImprovedStaggeredFermion5D<Impl>::DhopOE(const FermionField &in, FermionField &out,int dag)
{
DhopCalls+=1;
conformable(in._grid,FermionRedBlackGrid()); // verifies half grid
conformable(in._grid,out._grid); // drops the cb check
@ -261,6 +271,7 @@ void ImprovedStaggeredFermion5D<Impl>::DhopOE(const FermionField &in, FermionFie
template<class Impl>
void ImprovedStaggeredFermion5D<Impl>::DhopEO(const FermionField &in, FermionField &out,int dag)
{
DhopCalls+=1;
conformable(in._grid,FermionRedBlackGrid()); // verifies half grid
conformable(in._grid,out._grid); // drops the cb check
@ -272,6 +283,7 @@ void ImprovedStaggeredFermion5D<Impl>::DhopEO(const FermionField &in, FermionFie
template<class Impl>
void ImprovedStaggeredFermion5D<Impl>::Dhop(const FermionField &in, FermionField &out,int dag)
{
DhopCalls+=2;
conformable(in._grid,FermionGrid()); // verifies full grid
conformable(in._grid,out._grid);
@ -280,6 +292,54 @@ void ImprovedStaggeredFermion5D<Impl>::Dhop(const FermionField &in, FermionField
DhopInternal(Stencil,Lebesgue,Umu,UUUmu,in,out,dag);
}
template<class Impl>
void ImprovedStaggeredFermion5D<Impl>::Report(void)
{
std::vector<int> latt = GridDefaultLatt();
RealD volume = Ls; for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
RealD NP = _FourDimGrid->_Nprocessors;
RealD NN = _FourDimGrid->NodeCount();
std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion5D Number of DhopEO Calls : "
<< DhopCalls << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion5D TotalTime /Calls : "
<< DhopTotalTime / DhopCalls << " us" << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion5D CommTime /Calls : "
<< DhopCommTime / DhopCalls << " us" << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion5D ComputeTime/Calls : "
<< DhopComputeTime / DhopCalls << " us" << std::endl;
// Average the compute time
_FourDimGrid->GlobalSum(DhopComputeTime);
DhopComputeTime/=NP;
RealD mflops = 1154*volume*DhopCalls/DhopComputeTime/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call : " << mflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per rank : " << mflops/NP << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per node : " << mflops/NN << std::endl;
RealD Fullmflops = 1154*volume*DhopCalls/(DhopTotalTime)/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call (full) : " << Fullmflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per rank (full): " << Fullmflops/NP << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per node (full): " << Fullmflops/NN << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion5D Stencil" <<std::endl; Stencil.Report();
std::cout << GridLogMessage << "ImprovedStaggeredFermion5D StencilEven"<<std::endl; StencilEven.Report();
std::cout << GridLogMessage << "ImprovedStaggeredFermion5D StencilOdd" <<std::endl; StencilOdd.Report();
}
template<class Impl>
void ImprovedStaggeredFermion5D<Impl>::ZeroCounters(void)
{
DhopCalls = 0;
DhopTotalTime = 0;
DhopCommTime = 0;
DhopComputeTime = 0;
Stencil.ZeroCounters();
StencilEven.ZeroCounters();
StencilOdd.ZeroCounters();
}
/////////////////////////////////////////////////////////////////////////
// Implement the general interface. Here we use SAME mass on all slices

10
lib/qcd/action/fermion/ImprovedStaggeredFermion5D.h
View File

@ -55,6 +55,16 @@ namespace QCD {
FermionField _tmp;
FermionField &tmp(void) { return _tmp; }
////////////////////////////////////////
// Performance monitoring
////////////////////////////////////////
void Report(void);
void ZeroCounters(void);
double DhopTotalTime;
double DhopCalls;
double DhopCommTime;
double DhopComputeTime;
///////////////////////////////////////////////////////////////
// Implement the abstract base
///////////////////////////////////////////////////////////////

48
lib/qcd/action/fermion/WilsonCompressor.h
View File

@ -238,7 +238,33 @@ template<typename HCS,typename HS,typename S> using WilsonCompressor = WilsonCom
template<class vobj,class cobj>
class WilsonStencil : public CartesianStencil<vobj,cobj> {
public:
double timer0;
double timer1;
double timer2;
double timer3;
double timer4;
double timer5;
double timer6;
uint64_t callsi;
void ZeroCountersi(void)
{
timer0=0;
timer1=0;
timer2=0;
timer3=0;
timer4=0;
timer5=0;
timer6=0;
callsi=0;
}
void Reporti(int calls)
{
if ( timer0 ) std::cout << GridLogMessage << " timer0 (HaloGatherOpt) " <<timer0/calls <<std::endl;
if ( timer1 ) std::cout << GridLogMessage << " timer1 (Communicate) " <<timer1/calls <<std::endl;
if ( timer2 ) std::cout << GridLogMessage << " timer2 (CommsMerge ) " <<timer2/calls <<std::endl;
if ( timer3 ) std::cout << GridLogMessage << " timer3 (commsMergeShm) " <<timer3/calls <<std::endl;
if ( timer4 ) std::cout << GridLogMessage << " timer4 " <<timer4 <<std::endl;
}
typedef CartesianCommunicator::CommsRequest_t CommsRequest_t;
std::vector<int> same_node;
@ -252,6 +278,7 @@ public:
: CartesianStencil<vobj,cobj> (grid,npoints,checkerboard,directions,distances) ,
same_node(npoints)
{
ZeroCountersi();
surface_list.resize(0);
};
@ -261,7 +288,6 @@ public:
// Here we know the distance is 1 for WilsonStencil
for(int point=0;point<this->_npoints;point++){
same_node[point] = this->SameNode(point);
// std::cout << " dir " <<point<<" same_node " <<same_node[point]<<std::endl;
}
for(int site = 0 ;site< vol4;site++){
@ -282,17 +308,28 @@ public:
{
std::vector<std::vector<CommsRequest_t> > reqs;
this->HaloExchangeOptGather(source,compress);
this->CommunicateBegin(reqs);
this->CommunicateComplete(reqs);
double t1=usecond();
// Asynchronous MPI calls multidirectional, Isend etc...
// this->CommunicateBegin(reqs);
// this->CommunicateComplete(reqs);
// Non-overlapped directions within a thread. Asynchronous calls except MPI3, threaded up to comm threads ways.
this->Communicate();
double t2=usecond(); timer1 += t2-t1;
this->CommsMerge(compress);
double t3=usecond(); timer2 += t3-t2;
this->CommsMergeSHM(compress);
double t4=usecond(); timer3 += t4-t3;
}
template <class compressor>
void HaloExchangeOptGather(const Lattice<vobj> &source,compressor &compress)
{
this->Prepare();
double t0=usecond();
this->HaloGatherOpt(source,compress);
double t1=usecond();
timer0 += t1-t0;
callsi++;
}
template <class compressor>
@ -304,7 +341,9 @@ public:
typedef typename compressor::SiteHalfSpinor SiteHalfSpinor;
typedef typename compressor::SiteHalfCommSpinor SiteHalfCommSpinor;
this->mpi3synctime_g-=usecond();
this->_grid->StencilBarrier();
this->mpi3synctime_g+=usecond();
assert(source._grid==this->_grid);
this->halogtime-=usecond();
@ -323,7 +362,6 @@ public:
int dag = compress.dag;
int face_idx=0;
if ( dag ) {
// std::cout << " Optimised Dagger compress " <<std::endl;
assert(same_node[Xp]==this->HaloGatherDir(source,XpCompress,Xp,face_idx));
assert(same_node[Yp]==this->HaloGatherDir(source,YpCompress,Yp,face_idx));
assert(same_node[Zp]==this->HaloGatherDir(source,ZpCompress,Zp,face_idx));

107
lib/qcd/action/fermion/WilsonFermion5D.cc
View File

@ -123,22 +123,24 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
int vol4;
vol4=FourDimGrid.oSites();
Stencil.BuildSurfaceList(LLs,vol4);
vol4=FourDimRedBlackGrid.oSites();
StencilEven.BuildSurfaceList(LLs,vol4);
StencilOdd.BuildSurfaceList(LLs,vol4);
std::cout << GridLogMessage << " SurfaceLists "<< Stencil.surface_list.size()
<<" " << StencilEven.surface_list.size()<<std::endl;
// std::cout << GridLogMessage << " SurfaceLists "<< Stencil.surface_list.size()
// <<" " << StencilEven.surface_list.size()<<std::endl;
}
template<class Impl>
void WilsonFermion5D<Impl>::Report(void)
{
std::vector<int> latt = GridDefaultLatt();
RealD volume = Ls; for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
RealD NP = _FourDimGrid->_Nprocessors;
RealD NN = _FourDimGrid->NodeCount();
RealD NP = _FourDimGrid->_Nprocessors;
RealD NN = _FourDimGrid->NodeCount();
RealD volume = Ls;
std::vector<int> latt = _FourDimGrid->GlobalDimensions();
for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
if ( DhopCalls > 0 ) {
std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
@ -184,6 +186,11 @@ void WilsonFermion5D<Impl>::Report(void)
std::cout << GridLogMessage << "WilsonFermion5D StencilEven"<<std::endl; StencilEven.Report();
std::cout << GridLogMessage << "WilsonFermion5D StencilOdd" <<std::endl; StencilOdd.Report();
}
if ( DhopCalls > 0){
std::cout << GridLogMessage << "WilsonFermion5D Stencil Reporti()" <<std::endl; Stencil.Reporti(DhopCalls);
std::cout << GridLogMessage << "WilsonFermion5D StencilEven Reporti()"<<std::endl; StencilEven.Reporti(DhopCalls);
std::cout << GridLogMessage << "WilsonFermion5D StencilOdd Reporti()" <<std::endl; StencilOdd.Reporti(DhopCalls);
}
}
template<class Impl>
@ -203,6 +210,9 @@ void WilsonFermion5D<Impl>::ZeroCounters(void) {
Stencil.ZeroCounters();
StencilEven.ZeroCounters();
StencilOdd.ZeroCounters();
Stencil.ZeroCountersi();
StencilEven.ZeroCountersi();
StencilOdd.ZeroCountersi();
}
@ -379,7 +389,6 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
{
#ifdef GRID_OMP
// assert((dag==DaggerNo) ||(dag==DaggerYes));
typedef CartesianCommunicator::CommsRequest_t CommsRequest_t;
Compressor compressor(dag);
@ -388,46 +397,70 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
DhopFaceTime-=usecond();
st.HaloExchangeOptGather(in,compressor);
DhopFaceTime+=usecond();
std::vector<std::vector<CommsRequest_t> > reqs;
// Rely on async comms; start comms before merge of local data
DhopCommTime-=usecond();
st.CommunicateBegin(reqs);
DhopFaceTime-=usecond();
st.CommsMergeSHM(compressor);
st.CommsMergeSHM(compressor);// Could do this inside parallel region overlapped with comms
DhopFaceTime+=usecond();
// Perhaps use omp task and region
#pragma omp parallel
double ctime=0;
double ptime=0;
//////////////////////////////////////////////////////////////////////////////////////////////////////
// Ugly explicit thread mapping introduced for OPA reasons.
//////////////////////////////////////////////////////////////////////////////////////////////////////
#pragma omp parallel reduction(max:ctime) reduction(max:ptime)
{
int tid = omp_get_thread_num();
int nthreads = omp_get_num_threads();
int me = omp_get_thread_num();
int myoff, mywork;
GridThread::GetWork(len,me-1,mywork,myoff,nthreads-1);
int sF = LLs * myoff;
if ( me == 0 ) {
st.CommunicateComplete(reqs);
DhopCommTime+=usecond();
} else {
// Interior links in stencil
if ( me==1 ) DhopComputeTime-=usecond();
if (dag == DaggerYes) Kernels::DhopSiteDag(st,lo,U,st.CommBuf(),sF,myoff,LLs,mywork,in,out,1,0);
else Kernels::DhopSite(st,lo,U,st.CommBuf(),sF,myoff,LLs,mywork,in,out,1,0);
if ( me==1 ) DhopComputeTime+=usecond();
int ncomms = CartesianCommunicator::nCommThreads;
if (ncomms == -1) ncomms = 1;
assert(nthreads > ncomms);
if (tid >= ncomms) {
double start = usecond();
nthreads -= ncomms;
int ttid = tid - ncomms;
int n = U._grid->oSites();
int chunk = n / nthreads;
int rem = n % nthreads;
int myblock, myn;
if (ttid < rem) {
myblock = ttid * chunk + ttid;
myn = chunk+1;
} else {
myblock = ttid*chunk + rem;
myn = chunk;
}
// do the compute
if (dag == DaggerYes) {
for (int ss = myblock; ss < myblock+myn; ++ss) {
int sU = ss;
int sF = LLs * sU;
Kernels::DhopSiteDag(st,lo,U,st.CommBuf(),sF,sU,LLs,1,in,out,1,0);
}
} else {
for (int ss = myblock; ss < myblock+myn; ++ss) {
int sU = ss;
int sF = LLs * sU;
Kernels::DhopSite(st,lo,U,st.CommBuf(),sF,sU,LLs,1,in,out,1,0);
}
}
ptime = usecond() - start;
}
{
double start = usecond();
st.CommunicateThreaded();
ctime = usecond() - start;
}
}
DhopCommTime += ctime;
DhopComputeTime+=ptime;
// First to enter, last to leave timing
st.CollateThreads();
DhopFaceTime-=usecond();
st.CommsMerge(compressor);
DhopFaceTime+=usecond();
// Load imbalance alert. Should use dynamic schedule OMP for loop
// Perhaps create a list of only those sites with face work, and
// load balance process the list.
DhopComputeTime2-=usecond();
if (dag == DaggerYes) {
int sz=st.surface_list.size();
@ -448,11 +481,9 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
#else
assert(0);
#endif
}
template<class Impl>
void WilsonFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st, LebesgueOrder &lo,
DoubledGaugeField & U,

14
lib/qcd/hmc/HMCResourceManager.h
View File

@ -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 = "") {

92
lib/qcd/hmc/HMC_GridModules.h
View File

@ -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);
}
}

15
lib/qcd/modules/ObservableModules.h
View File

@ -84,8 +84,6 @@ class PlaquetteMod: public ObservableModule<PlaquetteLogger<Impl>, NoParameters>
typedef ObservableModule<PlaquetteLogger<Impl>, NoParameters> ObsBase;
using ObsBase::ObsBase; // for constructors
// acquire resource
virtual void initialize(){
this->ObservablePtr.reset(new PlaquetteLogger<Impl>());
@ -94,23 +92,22 @@ class PlaquetteMod: public ObservableModule<PlaquetteLogger<Impl>, NoParameters>
PlaquetteMod(): ObsBase(NoParameters()){}
};
template < class Impl >
class TopologicalChargeMod: public ObservableModule<TopologicalCharge<Impl>, NoParameters>{
typedef ObservableModule<TopologicalCharge<Impl>, NoParameters> ObsBase;
class TopologicalChargeMod: public ObservableModule<TopologicalCharge<Impl>, TopologyObsParameters>{
typedef ObservableModule<TopologicalCharge<Impl>, TopologyObsParameters> ObsBase;
using ObsBase::ObsBase; // for constructors
// acquire resource
virtual void initialize(){
this->ObservablePtr.reset(new TopologicalCharge<Impl>());
this->ObservablePtr.reset(new TopologicalCharge<Impl>(this->Par_));
}
public:
TopologicalChargeMod(): ObsBase(NoParameters()){}
TopologicalChargeMod(TopologyObsParameters Par): ObsBase(Par){}
TopologicalChargeMod(): ObsBase(){}
};
}// QCD temporarily here

69
lib/qcd/observables/topological_charge.h
View File

@ -33,9 +33,45 @@ directory
namespace Grid {
namespace QCD {
struct TopologySmearingParameters : Serializable {
GRID_SERIALIZABLE_CLASS_MEMBERS(TopologySmearingParameters,
int, steps,
float, step_size,
int, meas_interval,
float, maxTau);
TopologySmearingParameters(int s = 0, float ss = 0.0f, int mi = 0, float mT = 0.0f):
steps(s), step_size(ss), meas_interval(mi), maxTau(mT){}
template < class ReaderClass >
TopologySmearingParameters(Reader<ReaderClass>& Reader){
read(Reader, "Smearing", *this);
}
};
struct TopologyObsParameters : Serializable {
GRID_SERIALIZABLE_CLASS_MEMBERS(TopologyObsParameters,
int, interval,
bool, do_smearing,
TopologySmearingParameters, Smearing);
TopologyObsParameters(int interval = 1, bool smearing = false):
interval(interval), Smearing(smearing){}
template <class ReaderClass >
TopologyObsParameters(Reader<ReaderClass>& Reader){
read(Reader, "TopologyMeasurement", *this);
}
};
// this is only defined for a gauge theory
template <class Impl>
class TopologicalCharge : public HmcObservable<typename Impl::Field> {
TopologyObsParameters Pars;
public:
// here forces the Impl to be of gauge fields
// if not the compiler will complain
@ -44,20 +80,39 @@ class TopologicalCharge : public HmcObservable<typename Impl::Field> {
// necessary for HmcObservable compatibility
typedef typename Impl::Field Field;
TopologicalCharge(int interval = 1, bool do_smearing = false):
Pars(interval, do_smearing){}
TopologicalCharge(TopologyObsParameters P):Pars(P){
std::cout << GridLogDebug << "Creating TopologicalCharge " << std::endl;
}
void TrajectoryComplete(int traj,
Field &U,
GridSerialRNG &sRNG,
GridParallelRNG &pRNG) {
Real q = WilsonLoops<Impl>::TopologicalCharge(U);
if (traj%Pars.interval == 0){
// Smearing
Field Usmear = U;
int def_prec = std::cout.precision();
if (Pars.do_smearing){
// using wilson flow by default here
WilsonFlow<PeriodicGimplR> WF(Pars.Smearing.steps, Pars.Smearing.step_size, Pars.Smearing.meas_interval);
WF.smear_adaptive(Usmear, U, Pars.Smearing.maxTau);
Real T0 = WF.energyDensityPlaquette(Usmear);
std::cout << GridLogMessage << std::setprecision(std::numeric_limits<Real>::digits10 + 1)
<< "T0 : [ " << traj << " ] "<< T0 << std::endl;
}
int def_prec = std::cout.precision();
Real q = WilsonLoops<Impl>::TopologicalCharge(Usmear);
std::cout << GridLogMessage
<< std::setprecision(std::numeric_limits<Real>::digits10 + 1)
<< "Topological Charge: [ " << traj << " ] "<< q << std::endl;
std::cout << GridLogMessage
<< std::setprecision(std::numeric_limits<Real>::digits10 + 1)
<< "Topological Charge: [ " << traj << " ] "<< q << std::endl;
std::cout.precision(def_prec);
std::cout.precision(def_prec);
}
}
};

11
lib/qcd/utils/GaugeFix.h
View File

@ -26,12 +26,14 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
/* END LEGAL */
//#include <Grid/Grid.h>
using namespace Grid;
using namespace Grid::QCD;
#ifndef GRID_QCD_GAUGE_FIX_H
#define GRID_QCD_GAUGE_FIX_H
namespace Grid {
namespace QCD {
template <class Gimpl>
class FourierAcceleratedGaugeFixer : public Gimpl {
public:
public:
INHERIT_GIMPL_TYPES(Gimpl);
typedef typename Gimpl::GaugeLinkField GaugeMat;
@ -186,3 +188,6 @@ class FourierAcceleratedGaugeFixer : public Gimpl {
}
};
}
}
#endif

43
lib/simd/Grid_neon.h
View File

@ -82,11 +82,11 @@ namespace Optimization {
double tmp[2]={a,b};
return vld1q_f64(tmp);
}
//Real double // N:tbc
//Real double
inline float64x2_t operator()(double a){
return vdupq_n_f64(a);
}
//Integer // N:tbc
//Integer
inline uint32x4_t operator()(Integer a){
return vdupq_n_u32(a);
}
@ -124,33 +124,32 @@ namespace Optimization {
// Nils: Vset untested; not used currently in Grid at all;
// git commit 4a8c4ccfba1d05159348d21a9698028ea847e77b
struct Vset{
// Complex float // N:ok
// Complex float
inline float32x4_t operator()(Grid::ComplexF *a){
float tmp[4]={a[1].imag(),a[1].real(),a[0].imag(),a[0].real()};
return vld1q_f32(tmp);
}
// Complex double // N:ok
// Complex double
inline float64x2_t operator()(Grid::ComplexD *a){
double tmp[2]={a[0].imag(),a[0].real()};
return vld1q_f64(tmp);
}
// Real float // N:ok
// Real float
inline float32x4_t operator()(float *a){
float tmp[4]={a[3],a[2],a[1],a[0]};
return vld1q_f32(tmp);
}
// Real double // N:ok
// Real double
inline float64x2_t operator()(double *a){
double tmp[2]={a[1],a[0]};
return vld1q_f64(tmp);
}
// Integer // N:ok
// Integer
inline uint32x4_t operator()(Integer *a){
return vld1q_dup_u32(a);
}
};
// N:leaving as is
template <typename Out_type, typename In_type>
struct Reduce{
//Need templated class to overload output type
@ -249,9 +248,9 @@ namespace Optimization {
return vfmaq_f32(r4, r0, a); // ar*br-ai*bi ai*br+ar*bi ...
// no fma, use mul and add
//float32x4_t r5;
//r5 = vmulq_f32(r0, a);
//return vaddq_f32(r4, r5);
// float32x4_t r5;
// r5 = vmulq_f32(r0, a);
// return vaddq_f32(r4, r5);
}
// Complex double
inline float64x2_t operator()(float64x2_t a, float64x2_t b){
@ -272,9 +271,9 @@ namespace Optimization {
return vfmaq_f64(r4, r0, a); // ar*br-ai*bi ai*br+ar*bi
// no fma, use mul and add
//float64x2_t r5;
//r5 = vmulq_f64(r0, a);
//return vaddq_f64(r4, r5);
// float64x2_t r5;
// r5 = vmulq_f64(r0, a);
// return vaddq_f64(r4, r5);
}
};
@ -421,11 +420,6 @@ namespace Optimization {
}
}
// working, but no restriction on n
// template<int n> static inline float32x4_t tRotate(float32x4_t in){ return vextq_f32(in,in,n); };
// template<int n> static inline float64x2_t tRotate(float64x2_t in){ return vextq_f64(in,in,n); };
// restriction on n
template<int n> static inline float32x4_t tRotate(float32x4_t in){ return vextq_f32(in,in,n%4); };
template<int n> static inline float64x2_t tRotate(float64x2_t in){ return vextq_f64(in,in,n%2); };
@ -441,7 +435,7 @@ namespace Optimization {
sb = vcvt_high_f32_f16(h);
// there is no direct conversion from lower float32x4_t to float64x2_t
// vextq_f16 not supported by clang 3.8 / 4.0 / arm clang
//float16x8_t h1 = vextq_f16(h, h, 4); // correct, but not supported by clang
// float16x8_t h1 = vextq_f16(h, h, 4); // correct, but not supported by clang
// workaround for clang
uint32x4_t h1u = reinterpret_cast<uint32x4_t>(h);
float16x8_t h1 = reinterpret_cast<float16x8_t>(vextq_u32(h1u, h1u, 2));
@ -547,7 +541,7 @@ namespace Optimization {
//Complex double Reduce
template<> // N:by Boyle
template<>
inline Grid::ComplexD Reduce<Grid::ComplexD, float64x2_t>::operator()(float64x2_t in){
u128d conv; conv.v = in;
return Grid::ComplexD(conv.f[0],conv.f[1]);
@ -562,9 +556,7 @@ namespace Optimization {
//Integer Reduce
template<>
inline Integer Reduce<Integer, uint32x4_t>::operator()(uint32x4_t in){
// FIXME unimplemented
printf("Reduce : Missing integer implementation -> FIX\n");
assert(0);
return vaddvq_u32(in);
}
}
@ -603,4 +595,5 @@ namespace Optimization {
typedef Optimization::TimesMinusI TimesMinusISIMD;
typedef Optimization::TimesI TimesISIMD;
}
}

145
lib/stencil/Stencil.h
View File

@ -176,6 +176,9 @@ class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal
// Timing info; ugly; possibly temporary
/////////////////////////////////////////
double commtime;
double mpi3synctime;
double mpi3synctime_g;
double shmmergetime;
double gathertime;
double gathermtime;
double halogtime;
@ -185,6 +188,10 @@ class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal
double splicetime;
double nosplicetime;
double calls;
std::vector<double> comm_bytes_thr;
std::vector<double> comm_time_thr;
std::vector<double> comm_enter_thr;
std::vector<double> comm_leave_thr;
////////////////////////////////////////
// Stencil query
@ -248,35 +255,120 @@ class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal
//////////////////////////////////////////
// Comms packet queue for asynch thread
//////////////////////////////////////////
void CommunicateThreaded()
{
#ifdef GRID_OMP
// must be called in parallel region
int mythread = omp_get_thread_num();
int nthreads = CartesianCommunicator::nCommThreads;
#else
int mythread = 0;
int nthreads = 1;
#endif
if (nthreads == -1) nthreads = 1;
if (mythread < nthreads) {
comm_enter_thr[mythread] = usecond();
for (int i = mythread; i < Packets.size(); i += nthreads) {
uint64_t bytes = _grid->StencilSendToRecvFrom(Packets[i].send_buf,
Packets[i].to_rank,
Packets[i].recv_buf,
Packets[i].from_rank,
Packets[i].bytes,i);
comm_bytes_thr[mythread] += bytes;
}
comm_leave_thr[mythread]= usecond();
comm_time_thr[mythread] += comm_leave_thr[mythread] - comm_enter_thr[mythread];
}
}
void CollateThreads(void)
{
int nthreads = CartesianCommunicator::nCommThreads;
double first=0.0;
double last =0.0;
for(int t=0;t<nthreads;t++) {
double t0 = comm_enter_thr[t];
double t1 = comm_leave_thr[t];
comms_bytes+=comm_bytes_thr[t];
comm_enter_thr[t] = 0.0;
comm_leave_thr[t] = 0.0;
comm_time_thr[t] = 0.0;
comm_bytes_thr[t]=0;
if ( first == 0.0 ) first = t0; // first is t0
if ( (t0 > 0.0) && ( t0 < first ) ) first = t0; // min time seen
if ( t1 > last ) last = t1; // max time seen
}
commtime+= last-first;
}
void CommunicateBegin(std::vector<std::vector<CommsRequest_t> > &reqs)
{
reqs.resize(Packets.size());
commtime-=usecond();
for(int i=0;i<Packets.size();i++){
comms_bytes+=_grid->StencilSendToRecvFromBegin(reqs[i],
Packets[i].send_buf,
Packets[i].to_rank,
Packets[i].recv_buf,
Packets[i].from_rank,
Packets[i].bytes);
Packets[i].send_buf,
Packets[i].to_rank,
Packets[i].recv_buf,
Packets[i].from_rank,
Packets[i].bytes,i);
}
}
void CommunicateComplete(std::vector<std::vector<CommsRequest_t> > &reqs)
{
for(int i=0;i<Packets.size();i++){
_grid->StencilSendToRecvFromComplete(reqs[i]);
_grid->StencilSendToRecvFromComplete(reqs[i],i);
}
commtime+=usecond();
}
void Communicate(void)
{
#ifdef GRID_OMP
#pragma omp parallel
{
// must be called in parallel region
int mythread = omp_get_thread_num();
int maxthreads= omp_get_max_threads();
int nthreads = CartesianCommunicator::nCommThreads;
assert(nthreads <= maxthreads);
if (nthreads == -1) nthreads = 1;
#else
int mythread = 0;
int nthreads = 1;
#endif
if (mythread < nthreads) {
for (int i = mythread; i < Packets.size(); i += nthreads) {
double start = usecond();
comm_bytes_thr[mythread] += _grid->StencilSendToRecvFrom(Packets[i].send_buf,
Packets[i].to_rank,
Packets[i].recv_buf,
Packets[i].from_rank,
Packets[i].bytes,i);
comm_time_thr[mythread] += usecond() - start;
}
}
#ifdef GRID_OMP
}
#endif
}
template<class compressor> void HaloExchange(const Lattice<vobj> &source,compressor &compress)
{
std::vector<std::vector<CommsRequest_t> > reqs;
Prepare();
HaloGather(source,compress);
CommunicateBegin(reqs);
CommunicateComplete(reqs);
// Concurrent
//CommunicateBegin(reqs);
//CommunicateComplete(reqs);
// Sequential, possibly threaded
Communicate();
CommsMergeSHM(compress);
CommsMerge(compress);
}
@ -337,7 +429,9 @@ class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal
template<class compressor>
void HaloGather(const Lattice<vobj> &source,compressor &compress)
{
mpi3synctime_g-=usecond();
_grid->StencilBarrier();// Synch shared memory on a single nodes
mpi3synctime_g+=usecond();
// conformable(source._grid,_grid);
assert(source._grid==_grid);
@ -397,8 +491,12 @@ class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal
CommsMerge(decompress,Mergers,Decompressions);
}
template<class decompressor> void CommsMergeSHM(decompressor decompress) {
mpi3synctime-=usecond();
_grid->StencilBarrier();// Synch shared memory on a single nodes
mpi3synctime+=usecond();
shmmergetime-=usecond();
CommsMerge(decompress,MergersSHM,DecompressionsSHM);
shmmergetime+=usecond();
}
template<class decompressor>
@ -442,7 +540,12 @@ class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal
int checkerboard,
const std::vector<int> &directions,
const std::vector<int> &distances)
: _permute_type(npoints), _comm_buf_size(npoints)
: _permute_type(npoints),
_comm_buf_size(npoints),
comm_bytes_thr(npoints),
comm_enter_thr(npoints),
comm_leave_thr(npoints),
comm_time_thr(npoints)
{
face_table_computed=0;
_npoints = npoints;
@ -996,6 +1099,15 @@ class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal
void ZeroCounters(void) {
gathertime = 0.;
commtime = 0.;
mpi3synctime=0.;
mpi3synctime_g=0.;
shmmergetime=0.;
for(int i=0;i<_npoints;i++){
comm_time_thr[i]=0;
comm_bytes_thr[i]=0;
comm_enter_thr[i]=0;
comm_leave_thr[i]=0;
}
halogtime = 0.;
mergetime = 0.;
decompresstime = 0.;
@ -1011,6 +1123,18 @@ class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal
#define PRINTIT(A) AVERAGE(A); std::cout << GridLogMessage << " Stencil " << #A << " "<< A/calls<<std::endl;
RealD NP = _grid->_Nprocessors;
RealD NN = _grid->NodeCount();
double t = 0;
// if comm_time_thr is set they were all done in parallel so take the max
// but add up the bytes
int threaded = 0 ;
for (int i = 0; i < 8; ++i) {
if ( comm_time_thr[i]>0.0 ) {
threaded = 1;
comms_bytes += comm_bytes_thr[i];
if (t < comm_time_thr[i]) t = comm_time_thr[i];
}
}
if (threaded) commtime += t;
_grid->GlobalSum(commtime); commtime/=NP;
if ( calls > 0. ) {
@ -1026,6 +1150,9 @@ class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal
std::cout << GridLogMessage << " Stencil " << comms_bytes/commtime/1000. << " GB/s per rank"<<std::endl;
std::cout << GridLogMessage << " Stencil " << comms_bytes/commtime/1000.*NP/NN << " GB/s per node"<<std::endl;
}
PRINTIT(mpi3synctime);
PRINTIT(mpi3synctime_g);
PRINTIT(shmmergetime);
PRINTIT(splicetime);
PRINTIT(nosplicetime);
}

4
lib/tensors/Tensor_arith_mul.h
View File

@ -98,7 +98,9 @@ template<class rtype,class vtype,class mtype,int N>
strong_inline void mult(iVector<rtype,N> * __restrict__ ret,
const iVector<vtype,N> * __restrict__ rhs,
const iScalar<mtype> * __restrict__ lhs){
mult(ret,lhs,rhs);
for(int c1=0;c1<N;c1++){
mult(&ret->_internal[c1],&rhs->_internal[c1],&lhs->_internal);
}
}

27
lib/util/Init.cc
View File

@ -219,9 +219,15 @@ void Grid_init(int *argc,char ***argv)
int MB;
arg= GridCmdOptionPayload(*argv,*argv+*argc,"--shm");
GridCmdOptionInt(arg,MB);
CartesianCommunicator::MAX_MPI_SHM_BYTES = MB*1024*1024;
uint64_t MB64 = MB;
CartesianCommunicator::MAX_MPI_SHM_BYTES = MB64*1024LL*1024LL;
}
if( GridCmdOptionExists(*argv,*argv+*argc,"--shm-hugepages") ){
CartesianCommunicator::Hugepages = 1;
}
if( GridCmdOptionExists(*argv,*argv+*argc,"--debug-signals") ){
Grid_debug_handler_init();
}
@ -304,6 +310,7 @@ void Grid_init(int *argc,char ***argv)
std::cout<<GridLogMessage<<" --threads n : default number of OMP threads"<<std::endl;
std::cout<<GridLogMessage<<" --grid n.n.n.n : default Grid size"<<std::endl;
std::cout<<GridLogMessage<<" --shm M : allocate M megabytes of shared memory for comms"<<std::endl;
std::cout<<GridLogMessage<<" --shm-hugepages : use explicit huge pages in mmap call "<<std::endl;
std::cout<<GridLogMessage<<std::endl;
std::cout<<GridLogMessage<<"Verbose and debug:"<<std::endl;
std::cout<<GridLogMessage<<std::endl;
@ -317,7 +324,7 @@ void Grid_init(int *argc,char ***argv)
std::cout<<GridLogMessage<<std::endl;
std::cout<<GridLogMessage<<" --comms-concurrent : Asynchronous MPI calls; several dirs at a time "<<std::endl;
std::cout<<GridLogMessage<<" --comms-sequential : Synchronous MPI calls; one dirs at a time "<<std::endl;
std::cout<<GridLogMessage<<" --comms-overlap : Overlap comms with compute "<<std::endl;
std::cout<<GridLogMessage<<" --comms-overlap : Overlap comms with compute "<<std::endl;
std::cout<<GridLogMessage<<std::endl;
std::cout<<GridLogMessage<<" --dslash-generic: Wilson kernel for generic Nc"<<std::endl;
std::cout<<GridLogMessage<<" --dslash-unroll : Wilson kernel for Nc=3"<<std::endl;
@ -356,10 +363,15 @@ void Grid_init(int *argc,char ***argv)
if( GridCmdOptionExists(*argv,*argv+*argc,"--comms-sequential") ){
CartesianCommunicator::SetCommunicatorPolicy(CartesianCommunicator::CommunicatorPolicySequential);
}
if( GridCmdOptionExists(*argv,*argv+*argc,"--lebesgue") ){
LebesgueOrder::UseLebesgueOrder=1;
}
CartesianCommunicator::nCommThreads = -1;
if( GridCmdOptionExists(*argv,*argv+*argc,"--comms-threads") ){
arg= GridCmdOptionPayload(*argv,*argv+*argc,"--comms-threads");
GridCmdOptionInt(arg,CartesianCommunicator::nCommThreads);
}
if( GridCmdOptionExists(*argv,*argv+*argc,"--cacheblocking") ){
arg= GridCmdOptionPayload(*argv,*argv+*argc,"--cacheblocking");
GridCmdOptionIntVector(arg,LebesgueOrder::Block);
@ -374,10 +386,13 @@ void Grid_init(int *argc,char ***argv)
Grid_default_latt,
Grid_default_mpi);
std::cout << GridLogDebug << "Requesting "<< CartesianCommunicator::MAX_MPI_SHM_BYTES <<" byte stencil comms buffers "<<std::endl;
std::cout << GridLogMessage << "Requesting "<< CartesianCommunicator::MAX_MPI_SHM_BYTES <<" byte stencil comms buffers "<<std::endl;
if ( CartesianCommunicator::Hugepages) {
std::cout << GridLogMessage << "Mapped stencil comms buffers as MAP_HUGETLB "<<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;
@ -393,7 +408,7 @@ void Grid_init(int *argc,char ***argv)
void Grid_finalize(void)
{
#if defined (GRID_COMMS_MPI) || defined (GRID_COMMS_MPI3)
#if defined (GRID_COMMS_MPI) || defined (GRID_COMMS_MPI3) || defined (GRID_COMMS_MPIT)
MPI_Finalize();
Grid_unquiesce_nodes();
#endif