mirror of
https://github.com/paboyle/Grid.git
synced 2025-06-25 11:12:02 +01:00
Hadrons: big update abstracting the block meson field routine, tested & working, performance counters broken and code dirty
This commit is contained in:
@ -7,6 +7,7 @@ Source file: Hadrons/A2AMatrix.hpp
|
||||
Copyright (C) 2015-2018
|
||||
|
||||
Author: Antonin Portelli <antonin.portelli@me.com>
|
||||
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
|
||||
@ -29,10 +30,28 @@ See the full license in the file "LICENSE" in the top level distribution directo
|
||||
#define A2A_Matrix_hpp_
|
||||
|
||||
#include <Hadrons/Global.hpp>
|
||||
#include <Hadrons/TimerArray.hpp>
|
||||
#include <Grid/Eigen/unsupported/CXX11/Tensor>
|
||||
|
||||
#ifndef HADRONS_A2AM_NAME
|
||||
#define HADRONS_A2AM_NAME "a2aMatrix"
|
||||
#endif
|
||||
|
||||
#define HADRONS_A2AM_PARALLEL_IO
|
||||
|
||||
BEGIN_HADRONS_NAMESPACE
|
||||
|
||||
template <typename T, typename MetadataType>
|
||||
// general A2A matrix set class based on Eigen tensors and Grid-allocated memory
|
||||
// Dimensions:
|
||||
// 0 - ext - external field (momentum, EM field, ...)
|
||||
// 1 - str - spin-color structure
|
||||
// 2 - t - timeslice
|
||||
// 3 - i - left A2A mode index
|
||||
// 4 - j - right A2A mode index
|
||||
template <typename T>
|
||||
using A2AMatrixSet = Eigen::TensorMap<Eigen::Tensor<T, 5, Eigen::RowMajor>>;
|
||||
|
||||
template <typename T>
|
||||
class A2AMatrixIo
|
||||
{
|
||||
public:
|
||||
@ -41,26 +60,265 @@ public:
|
||||
const unsigned int nt, const unsigned int ni,
|
||||
const unsigned int nj);
|
||||
~A2AMatrixIo(void) = default;
|
||||
template <typename MetadataType>
|
||||
void initFile(const MetadataType &d, const unsigned int chunkSize);
|
||||
void saveBlock(const T *data, const unsigned int i, const unsigned int j,
|
||||
const unsigned int blockSizei, const unsigned int blockSizej);
|
||||
void saveBlock(const A2AMatrixSet<T> &m, const unsigned int ext, const unsigned int str,
|
||||
const unsigned int i, const unsigned int j);
|
||||
private:
|
||||
std::string filename_, dataname_;
|
||||
unsigned int nt_, ni_, nj_;
|
||||
};
|
||||
|
||||
template <typename T, typename MetadataType>
|
||||
A2AMatrixIo<T, MetadataType>::A2AMatrixIo(std::string filename,
|
||||
std::string dataname,
|
||||
const unsigned int nt,
|
||||
const unsigned int ni,
|
||||
const unsigned int nj)
|
||||
template <typename T, typename Field>
|
||||
class A2AKernel
|
||||
{
|
||||
public:
|
||||
A2AKernel(void) = default;
|
||||
virtual ~A2AKernel(void) = default;
|
||||
virtual void operator()(A2AMatrixSet<T> &m, const Field *left, const Field *right,
|
||||
const unsigned int orthogDim, double &time) = 0;
|
||||
virtual double flops(const unsigned int blockSizei, const unsigned int blockSizej) = 0;
|
||||
virtual double bytes(const unsigned int blockSizei, const unsigned int blockSizej) = 0;
|
||||
};
|
||||
|
||||
template <typename T, typename Field, typename MetadataType, typename TIo = T>
|
||||
class A2AMatrixBlockComputation
|
||||
{
|
||||
private:
|
||||
struct IoHelper
|
||||
{
|
||||
A2AMatrixIo<TIo> io;
|
||||
MetadataType md;
|
||||
unsigned int e, s, i, j;
|
||||
};
|
||||
typedef std::function<std::string(const unsigned int, const unsigned int)> FilenameFn;
|
||||
typedef std::function<MetadataType(const unsigned int, const unsigned int)> MetadataFn;
|
||||
public:
|
||||
A2AMatrixBlockComputation(GridBase *grid,
|
||||
const unsigned int orthogDim,
|
||||
const unsigned int next,
|
||||
const unsigned int nstr,
|
||||
const unsigned int blockSize,
|
||||
const unsigned int cacheBlockSize,
|
||||
TimerArray *tArray = nullptr);
|
||||
void execute(const std::vector<Field> &left,
|
||||
const std::vector<Field> &right,
|
||||
A2AKernel<T, Field> &kernel,
|
||||
const FilenameFn &ionameFn,
|
||||
const FilenameFn &filenameFn,
|
||||
const MetadataFn &metadataFn);
|
||||
private:
|
||||
void saveBlock(const A2AMatrixSet<TIo> &m, IoHelper &h);
|
||||
private:
|
||||
TimerArray *tArray_;
|
||||
GridBase *grid_;
|
||||
unsigned int orthogDim_, nt_, next_, nstr_, blockSize_, cacheBlockSize_;
|
||||
Vector<T> mCache_;
|
||||
Vector<TIo> mBuf_;
|
||||
std::vector<IoHelper> nodeIo_;
|
||||
};
|
||||
|
||||
template <typename T, typename Field, typename MetadataType, typename TIo>
|
||||
A2AMatrixBlockComputation<T, Field, MetadataType, TIo>
|
||||
::A2AMatrixBlockComputation(GridBase *grid,
|
||||
const unsigned int orthogDim,
|
||||
const unsigned int next,
|
||||
const unsigned int nstr,
|
||||
const unsigned int blockSize,
|
||||
const unsigned int cacheBlockSize,
|
||||
TimerArray *tArray)
|
||||
: grid_(grid), nt_(grid->GlobalDimensions()[orthogDim]), orthogDim_(orthogDim)
|
||||
, next_(next), nstr_(nstr), blockSize_(blockSize), cacheBlockSize_(cacheBlockSize)
|
||||
, tArray_(tArray)
|
||||
{
|
||||
std::cout << nt_ << std::endl;
|
||||
std::cout << next_ << std::endl;
|
||||
std::cout << nstr_ << std::endl;
|
||||
std::cout << cacheBlockSize_ << std::endl;
|
||||
std::cout << blockSize_ << std::endl;
|
||||
mCache_.resize(nt_*next_*nstr_*cacheBlockSize_*cacheBlockSize_);
|
||||
mBuf_.resize(nt_*next_*nstr_*blockSize_*blockSize_);
|
||||
std::cout << mCache_.size() << std::endl;
|
||||
std::cout << mBuf_.size() << std::endl;
|
||||
std::cout << blockSize << std::endl;
|
||||
std::cout << cacheBlockSize << std::endl;
|
||||
}
|
||||
|
||||
#define START_TIMER(name) if (tArray_) tArray_->startTimer(name)
|
||||
#define STOP_TIMER(name) if (tArray_) tArray_->stopTimer(name)
|
||||
#define GET_TIMER(name) (tArray_ != nullptr) ? tArray_->getDTimer(name) : 0.
|
||||
|
||||
template <typename T, typename Field, typename MetadataType, typename TIo>
|
||||
void A2AMatrixBlockComputation<T, Field, MetadataType, TIo>
|
||||
::execute(const std::vector<Field> &left, const std::vector<Field> &right,
|
||||
A2AKernel<T, Field> &kernel, const FilenameFn &ionameFn,
|
||||
const FilenameFn &filenameFn, const MetadataFn &metadataFn)
|
||||
{
|
||||
//////////////////////////////////////////////////////////////////////////
|
||||
// i,j is first loop over SchurBlock factors reusing 5D matrices
|
||||
// ii,jj is second loop over cacheBlock factors for high perf contractoin
|
||||
// iii,jjj are loops within cacheBlock
|
||||
// Total index is sum of these i+ii+iii etc...
|
||||
//////////////////////////////////////////////////////////////////////////
|
||||
int N_i = left.size();
|
||||
int N_j = right.size();
|
||||
double flops;
|
||||
double bytes;
|
||||
double t_kernel = 0.;
|
||||
double nodes = grid_->NodeCount();
|
||||
double tot_kernel;
|
||||
|
||||
double t0 = usecond();
|
||||
int NBlock_i = N_i/blockSize_ + (((N_i % blockSize_) != 0) ? 1 : 0);
|
||||
int NBlock_j = N_j/blockSize_ + (((N_j % blockSize_) != 0) ? 1 : 0);
|
||||
|
||||
for(int i=0;i<N_i;i+=blockSize_)
|
||||
for(int j=0;j<N_j;j+=blockSize_)
|
||||
{
|
||||
// Get the W and V vectors for this block^2 set of terms
|
||||
int N_ii = MIN(N_i-i,blockSize_);
|
||||
int N_jj = MIN(N_j-j,blockSize_);
|
||||
|
||||
LOG(Message) << "All-to-all matrix block "
|
||||
<< j/blockSize_ + NBlock_j*i/blockSize_ + 1
|
||||
<< "/" << NBlock_i*NBlock_j << " [" << i <<" .. "
|
||||
<< i+N_ii-1 << ", " << j <<" .. " << j+N_jj-1 << "]"
|
||||
<< std::endl;
|
||||
|
||||
A2AMatrixSet<TIo> mBlock(mBuf_.data(), next_, nstr_, nt_, N_ii, N_jj);
|
||||
|
||||
// Series of cache blocked chunks of the contractions within this block
|
||||
flops = 0.0;
|
||||
bytes = 0.0;
|
||||
for(int ii=0;ii<N_ii;ii+=cacheBlockSize_)
|
||||
for(int jj=0;jj<N_jj;jj+=cacheBlockSize_)
|
||||
{
|
||||
int N_iii = MIN(N_ii-ii,cacheBlockSize_);
|
||||
int N_jjj = MIN(N_jj-jj,cacheBlockSize_);
|
||||
A2AMatrixSet<T> mCacheBlock(mCache_.data(), next_, nstr_, nt_, N_iii, N_jjj);
|
||||
|
||||
// makeMesonFieldBlock(mfCacheBlock, &w[i+ii], &v[j+jj], gamma_, ph,
|
||||
// env().getNd() - 1, this);
|
||||
START_TIMER("kernel");
|
||||
kernel(mCacheBlock, &left[i+ii], &right[j+jj], orthogDim_, tot_kernel);
|
||||
STOP_TIMER("kernel");
|
||||
|
||||
// flops for general N_c & N_s
|
||||
// flops += vol * ( 2 * 8.0 + 6.0 + 8.0*nmom) * N_iii*N_jjj*ngamma;
|
||||
// bytes += vol * (12.0 * sizeof(Complex) ) * N_iii*N_jjj
|
||||
// + vol * ( 2.0 * sizeof(Complex) *nmom ) * N_iii*N_jjj* ngamma;
|
||||
flops += kernel.flops(N_iii, N_jjj);
|
||||
bytes += kernel.bytes(N_iii, N_jjj);
|
||||
|
||||
START_TIMER("cache copy");
|
||||
parallel_for_nest5(int e =0;e<next_;e++)
|
||||
for(int s =0;s< nstr_;s++)
|
||||
for(int t =0;t< nt_;t++)
|
||||
for(int iii=0;iii< N_iii;iii++)
|
||||
for(int jjj=0;jjj< N_jjj;jjj++)
|
||||
{
|
||||
mBlock(e,s,t,ii+iii,jj+jjj) = mCacheBlock(e,s,t,iii,jjj);
|
||||
}
|
||||
STOP_TIMER("cache copy");
|
||||
}
|
||||
|
||||
// perf
|
||||
// tot_kernel = getDTimer("contraction: colour trace & mom.")
|
||||
// + getDTimer("contraction: local space sum");
|
||||
t_kernel = tot_kernel - t_kernel;
|
||||
LOG(Message) << "Kernel perf " << flops/t_kernel/1.0e3/nodes
|
||||
<< " Gflop/s/node " << std::endl;
|
||||
LOG(Message) << "Kernel perf " << bytes/t_kernel*1.0e6/1024/1024/1024/nodes
|
||||
<< " GB/s/node " << std::endl;
|
||||
t_kernel = tot_kernel;
|
||||
|
||||
// IO
|
||||
double blockSize, ioTime;
|
||||
unsigned int myRank = grid_->ThisRank(),
|
||||
nRank = grid_->RankCount();
|
||||
|
||||
LOG(Message) << "Writing block to disk" << std::endl;
|
||||
ioTime = -GET_TIMER("IO: write block");
|
||||
START_TIMER("IO: total");
|
||||
makeFileDir(filenameFn(0, 0), grid_);
|
||||
#ifdef HADRONS_A2AM_PARALLEL_IO
|
||||
grid_->Barrier();
|
||||
// make task list for current node
|
||||
nodeIo_.clear();
|
||||
for(int f = myRank; f < next_*nstr_; f += nRank)
|
||||
{
|
||||
IoHelper h;
|
||||
|
||||
h.i = i;
|
||||
h.j = j;
|
||||
h.e = f/nstr_;
|
||||
h.s = f % nstr_;
|
||||
h.io = A2AMatrixIo<TIo>(filenameFn(h.e, h.s),
|
||||
ionameFn(h.e, h.s), nt_, N_i, N_j);
|
||||
h.md = metadataFn(h.e, h.s);
|
||||
nodeIo_.push_back(h);
|
||||
}
|
||||
// parallel IO
|
||||
for (auto &h: nodeIo_)
|
||||
{
|
||||
saveBlock(mBlock, h);
|
||||
}
|
||||
grid_->Barrier();
|
||||
#else
|
||||
// serial IO, for testing purposes only
|
||||
for(int e = 0; e < next_; e++)
|
||||
for(int s = 0; s < nstr_; s++)
|
||||
{
|
||||
IoHelper h;
|
||||
|
||||
h.i = i;
|
||||
h.j = j;
|
||||
h.e = e;
|
||||
h.s = s;
|
||||
h.io = A2AMatrixIo<TIo>(filenameFn(h.e, h.s),
|
||||
ionameFn(h.e, h.s), nt_, N_i, N_j);
|
||||
h.md = metadataFn(h.e, h.s);
|
||||
saveBlock(mfBlock, h);
|
||||
}
|
||||
#endif
|
||||
STOP_TIMER("IO: total");
|
||||
blockSize = static_cast<double>(next_*nstr_*nt_*N_ii*N_jj*sizeof(TIo));
|
||||
ioTime += GET_TIMER("IO: write block");
|
||||
LOG(Message) << "HDF5 IO done " << sizeString(blockSize) << " in "
|
||||
<< ioTime << " us ("
|
||||
<< blockSize/ioTime*1.0e6/1024/1024
|
||||
<< " MB/s)" << std::endl;
|
||||
}
|
||||
}
|
||||
|
||||
template <typename T, typename Field, typename MetadataType, typename TIo>
|
||||
void A2AMatrixBlockComputation<T, Field, MetadataType, TIo>
|
||||
::saveBlock(const A2AMatrixSet<TIo> &m, IoHelper &h)
|
||||
{
|
||||
if ((h.i == 0) and (h.j == 0))
|
||||
{
|
||||
START_TIMER("IO: file creation");
|
||||
h.io.initFile(h.md, blockSize_);
|
||||
STOP_TIMER("IO: file creation");
|
||||
}
|
||||
START_TIMER("IO: write block");
|
||||
h.io.saveBlock(m, h.e, h.s, h.i, h.j);
|
||||
STOP_TIMER("IO: write block");
|
||||
}
|
||||
|
||||
template <typename T>
|
||||
A2AMatrixIo<T>::A2AMatrixIo(std::string filename, std::string dataname,
|
||||
const unsigned int nt, const unsigned int ni,
|
||||
const unsigned int nj)
|
||||
: filename_(filename), dataname_(dataname)
|
||||
, nt_(nt), ni_(ni), nj_(nj)
|
||||
{}
|
||||
|
||||
template <typename T, typename MetadataType>
|
||||
void A2AMatrixIo<T, MetadataType>::initFile(const MetadataType &d, const unsigned int chunkSize)
|
||||
template <typename T>
|
||||
template <typename MetadataType>
|
||||
void A2AMatrixIo<T>::initFile(const MetadataType &d, const unsigned int chunkSize)
|
||||
{
|
||||
#ifdef HAVE_HDF5
|
||||
std::vector<hsize_t> dim = {static_cast<hsize_t>(nt_),
|
||||
@ -85,18 +343,18 @@ void A2AMatrixIo<T, MetadataType>::initFile(const MetadataType &d, const unsigne
|
||||
push(reader, dataname_);
|
||||
auto &group = reader.getGroup();
|
||||
plist.setChunk(chunk.size(), chunk.data());
|
||||
dataset = group.createDataSet("data", Hdf5Type<T>::type(), dataspace, plist);
|
||||
dataset = group.createDataSet(HADRONS_A2AM_NAME, Hdf5Type<T>::type(), dataspace, plist);
|
||||
#else
|
||||
HADRONS_ERROR(Implementation, "all-to-all matrix I/O needs HDF5 library");
|
||||
#endif
|
||||
}
|
||||
|
||||
template <typename T, typename MetadataType>
|
||||
void A2AMatrixIo<T, MetadataType>::saveBlock(const T *data,
|
||||
const unsigned int i,
|
||||
const unsigned int j,
|
||||
const unsigned int blockSizei,
|
||||
const unsigned int blockSizej)
|
||||
template <typename T>
|
||||
void A2AMatrixIo<T>::saveBlock(const T *data,
|
||||
const unsigned int i,
|
||||
const unsigned int j,
|
||||
const unsigned int blockSizei,
|
||||
const unsigned int blockSizej)
|
||||
{
|
||||
#ifdef HAVE_HDF5
|
||||
Hdf5Reader reader(filename_);
|
||||
@ -111,7 +369,7 @@ void A2AMatrixIo<T, MetadataType>::saveBlock(const T *data,
|
||||
|
||||
push(reader, dataname_);
|
||||
auto &group = reader.getGroup();
|
||||
dataset = group.openDataSet("data");
|
||||
dataset = group.openDataSet(HADRONS_A2AM_NAME);
|
||||
dataspace = dataset.getSpace();
|
||||
dataspace.selectHyperslab(H5S_SELECT_SET, count.data(), offset.data(),
|
||||
stride.data(), block.data());
|
||||
@ -121,6 +379,21 @@ void A2AMatrixIo<T, MetadataType>::saveBlock(const T *data,
|
||||
#endif
|
||||
}
|
||||
|
||||
template <typename T>
|
||||
void A2AMatrixIo<T>::saveBlock(const A2AMatrixSet<T> &m,
|
||||
const unsigned int ext, const unsigned int str,
|
||||
const unsigned int i, const unsigned int j)
|
||||
{
|
||||
unsigned int blockSizei = m.dimension(3);
|
||||
unsigned int blockSizej = m.dimension(4);
|
||||
unsigned int nstr = m.dimension(1);
|
||||
size_t offset = (ext*nstr + str)*nt_*blockSizei*blockSizej;
|
||||
|
||||
saveBlock(m.data() + offset, i, j, blockSizei, blockSizej);
|
||||
}
|
||||
|
||||
|
||||
|
||||
END_HADRONS_NAMESPACE
|
||||
|
||||
#endif // A2A_Matrix_hpp_
|
||||
|
||||
Reference in New Issue
Block a user