/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/lattice/Lattice_rng.h
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Guido Cossu <guido.cossu@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 */
#ifndef GRID_LATTICE_RNG_H
#define GRID_LATTICE_RNG_H
#include <random>
#ifdef RNG_SITMO
#include <Grid/sitmo_rng/sitmo_prng_engine.hpp>
#endif
#if defined(RNG_SITMO)
#define RNG_FAST_DISCARD
#else
#undef RNG_FAST_DISCARD
#endif
NAMESPACE_BEGIN(Grid);
//////////////////////////////////////////////////////////////
// Allow the RNG state to be less dense than the fine grid
//////////////////////////////////////////////////////////////
inline int RNGfillable(GridBase *coarse,GridBase *fine)
{
int rngdims = coarse->_ndimension;
// trivially extended in higher dims, with locality guaranteeing RNG state is local to node
int lowerdims = fine->_ndimension - coarse->_ndimension;
assert(lowerdims >= 0);
for(int d=0;d<lowerdims;d++){
assert(fine->_simd_layout[d]==1);
assert(fine->_processors[d]==1);
}
int multiplicity=1;
for(int d=0;d<lowerdims;d++){
multiplicity=multiplicity*fine->_rdimensions[d];
}
// local and global volumes subdivide cleanly after SIMDization
for(int d=0;d<rngdims;d++){
int fd= d+lowerdims;
assert(coarse->_processors[d] == fine->_processors[fd]);
assert(coarse->_simd_layout[d] == fine->_simd_layout[fd]);
assert(((fine->_rdimensions[fd] / coarse->_rdimensions[d])* coarse->_rdimensions[d])==fine->_rdimensions[fd]);
multiplicity = multiplicity *fine->_rdimensions[fd] / coarse->_rdimensions[d];
}
return multiplicity;
}
// merge of April 11 2017
// this function is necessary for the LS vectorised field
inline int RNGfillable_general(GridBase *coarse,GridBase *fine)
{
int rngdims = coarse->_ndimension;
// trivially extended in higher dims, with locality guaranteeing RNG state is local to node
int lowerdims = fine->_ndimension - coarse->_ndimension; assert(lowerdims >= 0);
// assumes that the higher dimensions are not using more processors
// all further divisions are local
for(int d=0;d<lowerdims;d++) assert(fine->_processors[d]==1);
for(int d=0;d<rngdims;d++) assert(coarse->_processors[d] == fine->_processors[d+lowerdims]);
// then divide the number of local sites
// check that the total number of sims agree, meanse the iSites are the same
assert(fine->Nsimd() == coarse->Nsimd());
// check that the two grids divide cleanly
assert( (fine->lSites() / coarse->lSites() ) * coarse->lSites() == fine->lSites() );
return fine->lSites() / coarse->lSites();
}
// real scalars are one component
template<class scalar,class distribution,class generator>
void fillScalar(scalar &s,distribution &dist,generator & gen)
{
s=dist(gen);
}
template<class distribution,class generator>
void fillScalar(ComplexF &s,distribution &dist, generator &gen)
{
// s=ComplexF(dist(gen),dist(gen));
s.real(dist(gen));
s.imag(dist(gen));
}
template<class distribution,class generator>
void fillScalar(ComplexD &s,distribution &dist,generator &gen)
{
// s=ComplexD(dist(gen),dist(gen));
s.real(dist(gen));
s.imag(dist(gen));
}
class GridRNGbase {
public:
// One generator per site.
// Uniform and Gaussian distributions from these generators.
#ifdef RNG_RANLUX
typedef std::ranlux48 RngEngine;
typedef uint64_t RngStateType;
static const int RngStateCount = 15;
#endif
#ifdef RNG_MT19937
typedef std::mt19937 RngEngine;
typedef uint32_t RngStateType;
static const int RngStateCount = std::mt19937::state_size;
#endif
#ifdef RNG_SITMO
typedef sitmo::prng_engine RngEngine;
typedef uint64_t RngStateType;
static const int RngStateCount = 13;
#endif
std::vector<RngEngine> _generators;
std::vector<std::uniform_real_distribution<RealD> > _uniform;
std::vector<std::normal_distribution<RealD> > _gaussian;
std::vector<std::discrete_distribution<int32_t> > _bernoulli;
std::vector<std::uniform_int_distribution<uint32_t> > _uid;
///////////////////////
// support for parallel init
///////////////////////
#ifdef RNG_FAST_DISCARD
static void Skip(RngEngine &eng,uint64_t site)
{
/////////////////////////////////////////////////////////////////////////////////////
// Skip by 2^40 elements between successive lattice sites
// This goes by 10^12.
// Consider quenched updating; likely never exceeding rate of 1000 sweeps
// per second on any machine. This gives us of order 10^9 seconds, or 100 years
// skip ahead.
// For HMC unlikely to go at faster than a solve per second, and
// tens of seconds per trajectory so this is clean in all reasonable cases,
// and margin of safety is orders of magnitude.
// We could hack Sitmo to skip in the higher order words of state if necessary
//
// Replace with 2^30 ; avoid problem on large volumes
//
/////////////////////////////////////////////////////////////////////////////////////
// uint64_t skip = site+1; // Old init Skipped then drew. Checked compat with faster init
const int shift = 30;
////////////////////////////////////////////////////////////////////
// Weird compiler bug in Intel 2018.1 under O3 was generating 32bit and not 64 bit left shift.
////////////////////////////////////////////////////////////////////
volatile uint64_t skip = site;
skip = skip<<shift;
assert((skip >> shift)==site); // check for overflow
eng.discard(skip);
// std::cout << " Engine " <<site << " state " <<eng<<std::endl;
}
#endif
static RngEngine Reseed(RngEngine &eng)
{
std::vector<uint32_t> newseed;
std::uniform_int_distribution<uint32_t> uid;
return Reseed(eng,newseed,uid);
}
static RngEngine Reseed(RngEngine &eng,std::vector<uint32_t> & newseed,
std::uniform_int_distribution<uint32_t> &uid)
{
const int reseeds=4;
newseed.resize(reseeds);
for(int i=0;i<reseeds;i++){
newseed[i] = uid(eng);
}
std::seed_seq sseq(newseed.begin(),newseed.end());
return RngEngine(sseq);
}
void GetState(std::vector<RngStateType> & saved,RngEngine &eng) {
saved.resize(RngStateCount);
std::stringstream ss;
ss<<eng;
ss.seekg(0,ss.beg);
for(int i=0;i<RngStateCount;i++){
ss>>saved[i];
}
}
void GetState(std::vector<RngStateType> & saved,int gen) {
GetState(saved,_generators[gen]);
}
void SetState(std::vector<RngStateType> & saved,RngEngine &eng){
assert(saved.size()==RngStateCount);
std::stringstream ss;
for(int i=0;i<RngStateCount;i++){
ss<< saved[i]<<" ";
}
ss.seekg(0,ss.beg);
ss>>eng;
}
void SetState(std::vector<RngStateType> & saved,int gen){
SetState(saved,_generators[gen]);
}
void SetEngine(RngEngine &Eng, int gen){
_generators[gen]=Eng;
}
void GetEngine(RngEngine &Eng, int gen){
Eng=_generators[gen];
}
template<class source> void Seed(source &src, int gen)
{
_generators[gen] = RngEngine(src);
}
};
class GridSerialRNG : public GridRNGbase {
public:
GridSerialRNG() : GridRNGbase() {
_generators.resize(1);
_uniform.resize(1,std::uniform_real_distribution<RealD>{0,1});
_gaussian.resize(1,std::normal_distribution<RealD>(0.0,1.0) );
_bernoulli.resize(1,std::discrete_distribution<int32_t>{1,1});
_uid.resize(1,std::uniform_int_distribution<uint32_t>() );
}
template <class sobj,class distribution> inline void fill(sobj &l,std::vector<distribution> &dist){
typedef typename sobj::scalar_type scalar_type;
int words = sizeof(sobj)/sizeof(scalar_type);
scalar_type *buf = (scalar_type *) & l;
dist[0].reset();
for(int idx=0;idx<words;idx++){
fillScalar(buf[idx],dist[0],_generators[0]);
}
CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
}
template <class distribution> inline void fill(ComplexF &l,std::vector<distribution> &dist){
dist[0].reset();
fillScalar(l,dist[0],_generators[0]);
CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
}
template <class distribution> inline void fill(ComplexD &l,std::vector<distribution> &dist){
dist[0].reset();
fillScalar(l,dist[0],_generators[0]);
CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
}
template <class distribution> inline void fill(RealF &l,std::vector<distribution> &dist){
dist[0].reset();
fillScalar(l,dist[0],_generators[0]);
CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
}
template <class distribution> inline void fill(RealD &l,std::vector<distribution> &dist){
dist[0].reset();
fillScalar(l,dist[0],_generators[0]);
CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
}
// vector fill
template <class distribution> inline void fill(vComplexF &l,std::vector<distribution> &dist){
RealF *pointer=(RealF *)&l;
dist[0].reset();
for(int i=0;i<2*vComplexF::Nsimd();i++){
fillScalar(pointer[i],dist[0],_generators[0]);
}
CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
}
template <class distribution> inline void fill(vComplexD &l,std::vector<distribution> &dist){
RealD *pointer=(RealD *)&l;
dist[0].reset();
for(int i=0;i<2*vComplexD::Nsimd();i++){
fillScalar(pointer[i],dist[0],_generators[0]);
}
CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
}
template <class distribution> inline void fill(vRealF &l,std::vector<distribution> &dist){
RealF *pointer=(RealF *)&l;
dist[0].reset();
for(int i=0;i<vRealF::Nsimd();i++){
fillScalar(pointer[i],dist[0],_generators[0]);
}
CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
}
template <class distribution> inline void fill(vRealD &l,std::vector<distribution> &dist){
RealD *pointer=(RealD *)&l;
dist[0].reset();
for(int i=0;i<vRealD::Nsimd();i++){
fillScalar(pointer[i],dist[0],_generators[0]);
}
CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
}
void SeedFixedIntegers(const std::vector<int> &seeds){
CartesianCommunicator::BroadcastWorld(0,(void *)&seeds[0],sizeof(int)*seeds.size());
std::seed_seq src(seeds.begin(),seeds.end());
Seed(src,0);
}
void SeedUniqueString(const std::string &s){
std::vector<int> seeds;
std::stringstream sha;
seeds = GridChecksum::sha256_seeds(s);
for(int i=0;i<seeds.size();i++) {
sha << std::hex << seeds[i];
}
std::cout << GridLogMessage << "Intialising serial RNG with unique string '"
<< s << "'" << std::endl;
std::cout << GridLogMessage << "Seed SHA256: " << sha.str() << std::endl;
SeedFixedIntegers(seeds);
}
};
class GridParallelRNG : public GridRNGbase {
private:
double _time_counter;
GridBase *_grid;
unsigned int _vol;
public:
GridBase *Grid(void) const { return _grid; }
int generator_idx(int os,int is) {
return is*_grid->oSites()+os;
}
GridParallelRNG(GridBase *grid) : GridRNGbase() {
_grid = grid;
_vol =_grid->iSites()*_grid->oSites();
_generators.resize(_vol);
_uniform.resize(_vol,std::uniform_real_distribution<RealD>{0,1});
_gaussian.resize(_vol,std::normal_distribution<RealD>(0.0,1.0) );
_bernoulli.resize(_vol,std::discrete_distribution<int32_t>{1,1});
_uid.resize(_vol,std::uniform_int_distribution<uint32_t>() );
}
template <class vobj,class distribution> inline void fill(Lattice<vobj> &l,std::vector<distribution> &dist){
typedef typename vobj::scalar_object scalar_object;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
double inner_time_counter = usecond();
int multiplicity = RNGfillable_general(_grid, l.Grid()); // l has finer or same grid
int Nsimd = _grid->Nsimd(); // guaranteed to be the same for l.Grid() too
int osites = _grid->oSites(); // guaranteed to be <= l.Grid()->oSites() by a factor multiplicity
int words = sizeof(scalar_object) / sizeof(scalar_type);
autoView(l_v, l, CpuWrite);
thread_for( ss, osites, {
ExtractBuffer<scalar_object> buf(Nsimd);
for (int m = 0; m < multiplicity; m++) { // Draw from same generator multiplicity times
int sm = multiplicity * ss + m; // Maps the generator site to the fine site
for (int si = 0; si < Nsimd; si++) {
int gdx = generator_idx(ss, si); // index of generator state
scalar_type *pointer = (scalar_type *)&buf[si];
dist[gdx].reset();
for (int idx = 0; idx < words; idx++)
fillScalar(pointer[idx], dist[gdx], _generators[gdx]);
}
// merge into SIMD lanes, FIXME suboptimal implementation
merge(l_v[sm], buf);
}
});
// });
_time_counter += usecond()- inner_time_counter;
}
void SeedUniqueString(const std::string &s){
std::vector<int> seeds;
seeds = GridChecksum::sha256_seeds(s);
std::cout << GridLogMessage << "Intialising parallel RNG with unique string '"
<< s << "'" << std::endl;
std::cout << GridLogMessage << "Seed SHA256: " << GridChecksum::sha256_string(seeds) << std::endl;
SeedFixedIntegers(seeds);
}
void SeedFixedIntegers(const std::vector<int> &seeds){
// Everyone generates the same seed_seq based on input seeds
CartesianCommunicator::BroadcastWorld(0,(void *)&seeds[0],sizeof(int)*seeds.size());
std::seed_seq source(seeds.begin(),seeds.end());
RngEngine master_engine(source);
#ifdef RNG_FAST_DISCARD
////////////////////////////////////////////////
// Skip ahead through a single stream.
// Applicable to SITMO and other has based/crypto RNGs
// Should be applicable to Mersenne Twister, but the C++11
// MT implementation does not implement fast discard even though
// in principle this is possible
////////////////////////////////////////////////
// Everybody loops over global volume.
thread_for( gidx, _grid->_gsites, {
// Where is it?
int rank;
int o_idx;
int i_idx;
Coordinate gcoor;
_grid->GlobalIndexToGlobalCoor(gidx,gcoor);
_grid->GlobalCoorToRankIndex(rank,o_idx,i_idx,gcoor);
// If this is one of mine we take it
if( rank == _grid->ThisRank() ){
int l_idx=generator_idx(o_idx,i_idx);
_generators[l_idx] = master_engine;
Skip(_generators[l_idx],gidx); // Skip to next RNG sequence
}
});
#else
////////////////////////////////////////////////////////////////
// Machine and thread decomposition dependent seeding is efficient
// and maximally parallel; but NOT reproducible from machine to machine.
// Not ideal, but fastest way to reseed all nodes.
////////////////////////////////////////////////////////////////
{
// Obtain one Reseed per processor
int Nproc = _grid->ProcessorCount();
std::vector<RngEngine> seeders(Nproc);
int me= _grid->ThisRank();
for(int p=0;p<Nproc;p++){
seeders[p] = Reseed(master_engine);
}
master_engine = seeders[me];
}
{
// Obtain one reseeded generator per thread
int Nthread = 32; // Hardwire a good level or parallelism
std::vector<RngEngine> seeders(Nthread);
for(int t=0;t<Nthread;t++){
seeders[t] = Reseed(master_engine);
}
thread_for( t, Nthread, {
// set up one per local site in threaded fashion
std::vector<uint32_t> newseeds;
std::uniform_int_distribution<uint32_t> uid;
for(int l=0;l<_grid->lSites();l++) {
if ( (l%Nthread)==t ) {
_generators[l] = Reseed(seeders[t],newseeds,uid);
}
}
});
}
#endif
}
void Report(){
std::cout << GridLogMessage << "Time spent in the fill() routine by GridParallelRNG: "<< _time_counter/1e3 << " ms" << std::endl;
}
////////////////////////////////////////////////////////////////////////
// Support for rigorous test of RNG's
// Return uniform random uint32_t from requested site generator
////////////////////////////////////////////////////////////////////////
uint32_t GlobalU01(int gsite){
uint32_t the_number;
// who
int rank,o_idx,i_idx;
Coordinate gcoor;
_grid->GlobalIndexToGlobalCoor(gsite,gcoor);
_grid->GlobalCoorToRankIndex(rank,o_idx,i_idx,gcoor);
// draw
int l_idx=generator_idx(o_idx,i_idx);
if( rank == _grid->ThisRank() ){
the_number = _uid[l_idx](_generators[l_idx]);
}
// share & return
_grid->Broadcast(rank,(void *)&the_number,sizeof(the_number));
return the_number;
}
};
template <class vobj> inline void random(GridParallelRNG &rng,Lattice<vobj> &l) { rng.fill(l,rng._uniform); }
template <class vobj> inline void gaussian(GridParallelRNG &rng,Lattice<vobj> &l) { rng.fill(l,rng._gaussian); }
template <class vobj> inline void bernoulli(GridParallelRNG &rng,Lattice<vobj> &l){ rng.fill(l,rng._bernoulli);}
template <class sobj> inline void random(GridSerialRNG &rng,sobj &l) { rng.fill(l,rng._uniform ); }
template <class sobj> inline void gaussian(GridSerialRNG &rng,sobj &l) { rng.fill(l,rng._gaussian ); }
template <class sobj> inline void bernoulli(GridSerialRNG &rng,sobj &l){ rng.fill(l,rng._bernoulli); }
NAMESPACE_END(Grid);
#endif