Improved the gamma quite a bit.

Serial rng's which are set on node zero and broadcaste

lib/Grid_communicator.h

@@ -99,6 +99,8 @@ class CartesianCommunicator {
       Broadcast(root,(void *)&data,sizeof(data));
     };
 
+    static void BroadcastWorld(int root,void* data, int bytes);
+
 }; 
 }
 

lib/communicator/Grid_communicator_fake.cc

@@ -37,6 +37,9 @@ void CartesianCommunicator::Barrier(void)
 void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
 {
 }
+void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
+{
+}
 
 
 void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)

lib/communicator/Grid_communicator_mpi.cc

@@ -92,5 +92,14 @@ void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
 	    communicator);
 }
 
+void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
+{
+  MPI_Bcast(data,
+	    bytes,
+	    MPI_BYTE,
+	    root,
+	    MPI_COMM_WORLD);
+}
+
 }
 

lib/lattice/Grid_lattice_rng.h

@@ -26,7 +26,70 @@ namespace Grid {
     }
 
   };
-  class GridRNG {
+  class GridRNGbase {
+
+  public:
+
+   GridRNGbase() : _uniform{0,1}, _gaussian(0.0,1.0) {};
+
+    int _seeded;
+    // One generator per site.
+    // Uniform and Gaussian distributions from these generators.
+    std::vector<std::ranlux48>             _generators;
+    std::uniform_real_distribution<double> _uniform;
+    std::normal_distribution<double>       _gaussian;
+
+
+  };
+
+  class GridSerialRNG : public GridRNGbase {
+  public:
+
+    // FIXME ... do we require lockstep draws of randoms 
+    // from all nodes keeping seeds consistent.
+    // place a barrier/broadcast in the fill routine
+    template<class source> void Seed(source &src)
+    {
+      typename source::result_type init = src();
+      CartesianCommunicator::BroadcastWorld(0,(void *)&init,sizeof(init));
+      _generators[0] = std::ranlux48(init);
+      _seeded=1;
+    }    
+
+    GridSerialRNG() : GridRNGbase() {
+      _generators.resize(1);
+      _seeded=0;
+    }
+
+
+    template <class sobj,class distribution> inline void fill(sobj &l,distribution &dist){
+
+      typedef typename sobj::scalar_type scalar_type;
+ 
+      int words = sizeof(sobj)/sizeof(scalar_type);
+
+      scalar_type *buf = (scalar_type *) & l;
+
+      for(int idx=0;idx<words;idx++){
+	buf[idx] = dist(_generators[0]);
+      }
+      
+      CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
+
+    };
+
+    void SeedRandomDevice(void){
+      std::random_device rd;
+      Seed(rd);
+    }
+    void SeedFixedIntegers(std::vector<int> &seeds){
+      fixedSeed src(seeds);
+      Seed(src);
+    }
+
+  };
+
+  class GridParallelRNG : public GridRNGbase {
   public:
     // One generator per site.
     std::vector<std::ranlux48>             _generators;
@@ -42,25 +105,13 @@ namespace Grid {
       return is*_grid->oSites()+os;
     }
 
-    GridRNG(GridBase *grid) : _uniform{0,1}, _gaussian(0.0,1.0) {
+    GridParallelRNG(GridBase *grid) : GridRNGbase() {
       _grid=grid;
       _vol =_grid->iSites()*_grid->oSites();
       _generators.resize(_vol);
-      //      SeedFixedIntegers(seeds);
-      // worst case we seed properly but non-deterministically
-      SeedRandomDevice();
+      _seeded=0;
     }
 
-    // FIXME: drive seeding from node zero and transmit to all
-    // to get unique randoms on each node
-    void SeedRandomDevice(void){
-      std::random_device rd;
-      Seed(rd);
-    }
-    void SeedFixedIntegers(std::vector<int> &seeds){
-      fixedSeed src(seeds);
-      Seed(src);
-    }
 
     // This loop could be made faster to avoid the Ahmdahl by
     // i)  seed generators on each timeslice, for x=y=z=0;
@@ -86,6 +137,7 @@ namespace Grid {
 	  _generators[l_idx] = std::ranlux48(init);
 	}
       }
+      _seeded=1;
     }    
 
     //FIXME implement generic IO and create state save/restore
@@ -122,15 +174,34 @@ namespace Grid {
 	merge(l._odata[ss],pointers);
       }
     };
+
+    void SeedRandomDevice(void){
+      std::random_device rd;
+      Seed(rd);
+    }
+    void SeedFixedIntegers(std::vector<int> &seeds){
+      fixedSeed src(seeds);
+      Seed(src);
+    }
+
   };
 
-  // FIXME Implement a consistent seed management strategy
-  template <class vobj> inline void random(GridRNG &rng,Lattice<vobj> &l){
+
+  template <class vobj> inline void random(GridParallelRNG &rng,Lattice<vobj> &l){
     rng.fill(l,rng._uniform);
   }
 
-  template <class vobj> inline void gaussian(GridRNG &rng,Lattice<vobj> &l){
+  template <class vobj> inline void gaussian(GridParallelRNG &rng,Lattice<vobj> &l){
     rng.fill(l,rng._gaussian);
   }
+
+
+  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);
+  }
+
 }
 #endif

lib/math/Grid_math_inner.h

@@ -6,6 +6,13 @@ namespace Grid {
     // innerProduct Vector x Vector -> Scalar
     // innerProduct Matrix x Matrix -> Scalar
     ///////////////////////////////////////////////////////////////////////////////////////
+    template<class sobj> inline RealD norm2l(sobj &arg){
+      typedef typename sobj::scalar_type scalar;
+      decltype(innerProduct(arg,arg)) nrm;
+      nrm = innerProduct(arg,arg);
+      return real(nrm);
+    }
+
     template<class l,class r,int N> inline
     auto innerProduct (const iVector<l,N>& lhs,const iVector<r,N>& rhs) -> iScalar<decltype(innerProduct(lhs._internal[0],rhs._internal[0]))>
     {

lib/qcd/Grid_qcd_dirac.h

@@ -76,15 +76,15 @@ namespace QCD {
 	ret(i,0) = timesMinusI(rhs(i,3));
 	ret(i,1) = timesMinusI(rhs(i,2));
 	ret(i,2) = timesI(rhs(i,1));
-	ret(i,3) = timesI(rhs(i,1));
+	ret(i,3) = timesI(rhs(i,0));
       }
     };
     template<class vtype> inline void multGammaX(iMatrix<vtype,Ns> &ret, const iMatrix<vtype,Ns> &rhs){
       for(int i=0;i<Ns;i++){
 	ret(0,i) = timesI(rhs(3,i));
-	ret(1,i) = timesI(rhs._internal[2][i]);
-	ret(2,i) = timesMinusI(rhs._internal[1][i]);
-	ret(3,i) = timesMinusI(rhs._internal[0][i]);
+	ret(1,i) = timesI(rhs(2,i));
+	ret(2,i) = timesMinusI(rhs(1,i));
+	ret(3,i) = timesMinusI(rhs(0,i));
       }
     };
     template<class vtype> inline void multMinusGammaX(iMatrix<vtype,Ns> &ret, const iMatrix<vtype,Ns> &rhs){

tests/Grid_cshift.cc

@@ -14,7 +14,8 @@ int main (int argc, char ** argv)
   std::vector<int> latt_size  ({8,8,8,16});
     
   GridCartesian     Fine(latt_size,simd_layout,mpi_layout);
-  GridRNG           FineRNG(&Fine);
+  GridParallelRNG           FineRNG(&Fine);
+  FineRNG.SeedRandomDevice();
 
   LatticeComplex U(&Fine);
   LatticeComplex ShiftU(&Fine);

tests/Grid_gamma.cc

@@ -16,9 +16,15 @@ int main (int argc, char ** argv)
     
   GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
 
-  GridRNG           RNG(&Grid);
+  GridParallelRNG          pRNG(&Grid);
+  pRNG.SeedRandomDevice();
+
+  GridSerialRNG            sRNG;
+  sRNG.SeedRandomDevice();
 
   SpinMatrix ident=zero;
+  SpinMatrix rnd  ; random(sRNG,rnd);
+
   SpinMatrix ll=zero;
   SpinMatrix rr=zero;
   SpinMatrix result;
@@ -94,6 +100,29 @@ int main (int argc, char ** argv)
 
   }
 
-  
+  std::cout << "Testing Gamma^2 - 1 = 0"<<std::endl;
+  for(int mu=0;mu<6;mu++){
+    result =  Gamma(g[mu])* ident * Gamma(g[mu]);
+    result = result - ident;
+    double mag = TensorRemove(norm2l(result));
+    std::cout << list[mu]<<" " << mag<<std::endl;
+  }
+
+  std::cout << "Testing (MinusGamma + G )M = 0"<<std::endl;
+  for(int mu=0;mu<6;mu++){
+    result =          rnd * Gamma(g[mu]);
+    result = result + rnd * Gamma(g[mu+6]);
+    double mag = TensorRemove(norm2l(result));
+    std::cout << list[mu]<<" " << mag<<std::endl;
+  }
+
+  std::cout << "Testing M(MinusGamma + G )  = 0"<<std::endl;
+  for(int mu=0;mu<6;mu++){
+    result =           Gamma(g[mu])  *rnd;
+    result = result +  Gamma(g[mu+6])*rnd;
+    double mag = TensorRemove(norm2l(result));
+    std::cout << list[mu]<<" " << mag<<std::endl;
+  }
+
   Grid_finalize();
 }

tests/Grid_main.cc

@@ -57,7 +57,8 @@ int main (int argc, char ** argv)
     
     GridCartesian Fine(latt_size,simd_layout,mpi_layout);
     GridRedBlackCartesian rbFine(latt_size,simd_layout,mpi_layout);
-    GridRNG       FineRNG(&Fine);
+    GridParallelRNG       FineRNG(&Fine);
+    FineRNG.SeedRandomDevice();
 
     LatticeColourMatrix Foo(&Fine);
     LatticeColourMatrix Bar(&Fine);

tests/Grid_nersc_io.cc

@@ -20,7 +20,6 @@ int main (int argc, char ** argv)
   GridCartesian     Fine(latt_size,simd_layout,mpi_layout);
   GridCartesian     Coarse(clatt_size,simd_layout,mpi_layout);
 
-  GridRNG           FineRNG(&Fine);
   LatticeGaugeField Umu(&Fine);
 
   std::vector<LatticeColourMatrix> U(4,&Fine);

tests/Grid_stencil.cc

@@ -48,7 +48,8 @@ int main (int argc, char ** argv)
     
     GridCartesian Fine(latt_size,simd_layout,mpi_layout);
     GridRedBlackCartesian rbFine(latt_size,simd_layout,mpi_layout);
-    GridRNG       fRNG(&Fine);
+    GridParallelRNG       fRNG(&Fine);
+    fRNG.SeedRandomDevice();
 
     LatticeColourMatrix Foo(&Fine);
     LatticeColourMatrix Bar(&Fine);