Merge branch 'develop' into feature/mpi3

Conflicts:
	lib/simd/Grid_avx512.h

.gitignore

@@ -5,7 +5,6 @@
 *.o
 *.obj
 
-
 # Editor files #
 ################
 *~
@@ -48,6 +47,7 @@ Config.h.in
 config.log
 config.status
 .deps
+*.inc
 
 # http://www.gnu.org/software/autoconf #
 ########################################
@@ -62,19 +62,8 @@ stamp-h1
 config.sub
 config.guess
 INSTALL
-
+.dirstamp
-# Packages #
+ltmain.sh
-############
-# it's better to unpack these files and commit the raw source
-# git has its own built in compression methods
-*.7z
-*.dmg
-*.gz
-*.iso
-*.jar
-*.rar
-*.tar
-*.zip
  
 # Logs and databases #
 ######################
@@ -100,3 +89,16 @@ build*/*
 #####################
 *.xcodeproj/*
 build.sh
+
+# Eigen source #
+################
+lib/Eigen/*
+
+# FFTW source #
+################
+lib/fftw/*
+
+# libtool macros #
+##################
+m4/lt*
+m4/libtool.m4

.travis.yml

@@ -1,14 +1,14 @@
 language: cpp
 
+cache:
+  directories:
+    - clang
+
 matrix:
   include:
     - os:        osx
       osx_image: xcode7.2
       compiler: clang
-    - os:        osx
-      osx_image: xcode7.2
-      compiler: gcc
-      env: VERSION=-5
     - compiler: gcc
       addons:
         apt:
@@ -19,6 +19,8 @@ matrix:
             - libmpfr-dev
             - libgmp-dev
             - libmpc-dev
+            - libopenmpi-dev
+            - openmpi-bin
             - binutils-dev
       env: VERSION=-4.9
     - compiler: gcc
@@ -31,6 +33,8 @@ matrix:
             - libmpfr-dev
             - libgmp-dev
             - libmpc-dev
+            - libopenmpi-dev
+            - openmpi-bin
             - binutils-dev
       env: VERSION=-5
     - compiler: clang
@@ -38,42 +42,65 @@ matrix:
         apt:
           sources:
             - ubuntu-toolchain-r-test
-            - llvm-toolchain-precise-3.7
           packages:
-            - clang-3.7
+            - g++-4.8
             - libmpfr-dev
             - libgmp-dev
             - libmpc-dev
+            - libopenmpi-dev
+            - openmpi-bin
             - binutils-dev
-      env: VERSION=-3.7
+      env: CLANG_LINK=http://llvm.org/releases/3.8.0/clang+llvm-3.8.0-x86_64-linux-gnu-ubuntu-14.04.tar.xz
     - compiler: clang
       addons:
         apt:
           sources:
             - ubuntu-toolchain-r-test
-            - llvm-toolchain-precise-3.8
           packages:
-            - clang-3.8
+            - g++-4.8
             - libmpfr-dev
             - libgmp-dev
             - libmpc-dev
+            - libopenmpi-dev
+            - openmpi-bin
             - binutils-dev
-      env: VERSION=-3.8
+      env: CLANG_LINK=http://llvm.org/releases/3.7.0/clang+llvm-3.7.0-x86_64-linux-gnu-ubuntu-14.04.tar.xz
       
 before_install:
+    - export GRIDDIR=`pwd`
+    - if [[ "$TRAVIS_OS_NAME" == "linux" ]] && [[ "$CC" == "clang" ]] && [ ! -e clang/bin ]; then wget $CLANG_LINK; tar -xf `basename $CLANG_LINK`; mkdir clang; mv clang+*/* clang/; fi
+    - if [[ "$TRAVIS_OS_NAME" == "linux" ]] && [[ "$CC" == "clang" ]]; then export PATH="${GRIDDIR}/clang/bin:${PATH}"; fi
+    - if [[ "$TRAVIS_OS_NAME" == "linux" ]] && [[ "$CC" == "clang" ]]; then export LD_LIBRARY_PATH="${GRIDDIR}/clang/lib:${LD_LIBRARY_PATH}"; fi
     - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then brew update; fi
     - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then brew install libmpc; fi
+    - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then brew install openmpi; fi
     - if [[ "$TRAVIS_OS_NAME" == "osx" ]] && [[ "$CC" == "gcc" ]]; then brew install gcc5; fi
     
 install:
     - export CC=$CC$VERSION
     - export CXX=$CXX$VERSION
+    - echo $PATH
+    - which $CC
+    - $CC  --version
+    - which $CXX
+    - $CXX --version
     - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then export LDFLAGS='-L/usr/local/lib'; fi
     
 script:
-    - ./scripts/reconfigure_script
+    - ./bootstrap.sh
     - mkdir build
     - cd build
-    - ../configure CXXFLAGS="-msse4.2 -O3 -std=c++11" LIBS="-lmpfr -lgmp" --enable-precision=single --enable-simd=SSE4 --enable-comms=none
+    - ../configure --enable-precision=single --enable-simd=SSE4 --enable-comms=none
     - make -j4 
     - ./benchmarks/Benchmark_dwf --threads 1
+    - echo make clean
+    - ../configure --enable-precision=double --enable-simd=SSE4 --enable-comms=none
+    - make -j4
+    - ./benchmarks/Benchmark_dwf --threads 1
+    - echo make clean
+    - if [[ "$TRAVIS_OS_NAME" == "linux" ]]; then export CXXFLAGS='-DMPI_UINT32_T=MPI_UNSIGNED -DMPI_UINT64_T=MPI_UNSIGNED_LONG'; fi
+    - ../configure --enable-precision=single --enable-simd=SSE4 --enable-comms=mpi-auto
+    - make -j4
+    - if [[ "$TRAVIS_OS_NAME" == "linux" ]]; then mpirun.openmpi -n 2 ./benchmarks/Benchmark_dwf --threads 1 --mpi 2.1.1.1; fi
+    - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then mpirun -n 2 ./benchmarks/Benchmark_dwf --threads 1 --mpi 2.1.1.1; fi
+

Makefile.am

@@ -1,5 +1,5 @@
 # additional include paths necessary to compile the C++ library
-AM_CXXFLAGS = -I$(top_srcdir)/
+SUBDIRS = lib benchmarks tests
-SUBDIRS = lib tests benchmarks
 
-filelist: $(SUBDIRS)
+AM_CXXFLAGS += -I$(top_builddir)/include
+ACLOCAL_AMFLAGS = -I m4

README.md

@@ -1,8 +1,28 @@
-# Grid [![Build Status](https://travis-ci.org/paboyle/Grid.svg?branch=master)](https://travis-ci.org/paboyle/Grid)
+# Grid
-Data parallel C++ mathematical object library
+<table>
+<tr>
+    <td>Last stable release</td>
+    <td><a href="https://travis-ci.org/paboyle/Grid">
+    <img src="https://travis-ci.org/paboyle/Grid.svg?branch=master"></a>
+    </td>
+</tr>
+<tr>
+    <td>Development branch</td>
+    <td><a href="https://travis-ci.org/paboyle/Grid">
+    <img src="https://travis-ci.org/paboyle/Grid.svg?branch=develop"></a>
+    </td>
+</tr>
+</table>
 
-Last update 2015/7/30
+**Data parallel C++ mathematical object library.**
 
+Please send all pull requests to the `develop` branch.
+
+License: GPL v2.
+
+Last update 2016/08/03.
+
+### Description
 This library provides data parallel C++ container classes with internal memory layout
 that is transformed to map efficiently to SIMD architectures. CSHIFT facilities
 are provided, similar to HPF and cmfortran, and user control is given over the mapping of
@@ -22,37 +42,75 @@ optimally use MPI, OpenMP and SIMD parallelism under the hood. This is a signifi
 for most programmers.
 
 The layout transformations are parametrised by the SIMD vector length. This adapts according to the architecture.
-Presently SSE4 (128 bit) AVX, AVX2 (256 bit) and IMCI and AVX512 (512 bit) targets are supported (ARM NEON on the way).
+Presently SSE4 (128 bit) AVX, AVX2 (256 bit) and IMCI and AVX512 (512 bit) targets are supported (ARM NEON and BG/Q QPX on the way).
 
-These are presented as 
+These are presented as `vRealF`, `vRealD`, `vComplexF`, and `vComplexD` internal vector data types. These may be useful in themselves for other programmers.
-
+The corresponding scalar types are named `RealF`, `RealD`, `ComplexF` and `ComplexD`.
-     vRealF, vRealD, vComplexF, vComplexD 
-
-internal vector data types. These may be useful in themselves for other programmers.
-The corresponding scalar types are named
-
-     RealF, RealD, ComplexF, ComplexD
 
 MPI, OpenMP, and SIMD parallelism are present in the library.
+Please see https://arxiv.org/abs/1512.03487 for more detail.
 
-   You can give `configure' initial values for configuration parameters
+### Installation
-by setting variables in the command line or in the environment.  Here
+First, start by cloning the repository:
-are examples:
 
-     ./configure CXX=clang++ CXXFLAGS="-std=c++11 -O3 -msse4" --enable-simd=SSE4
+``` bash
+git clone https://github.com/paboyle/Grid.git
+```
 
-     ./configure CXX=clang++ CXXFLAGS="-std=c++11 -O3 -mavx" --enable-simd=AVX
+Then enter the cloned directory and set up the build system:
 
-     ./configure CXX=clang++ CXXFLAGS="-std=c++11 -O3 -mavx2" --enable-simd=AVX2
+``` bash
+cd Grid
+./bootstrap.sh
+```
 
-     ./configure CXX=icpc CXXFLAGS="-std=c++11 -O3 -mmic" --enable-simd=AVX512 --host=none
+Now you can execute the `configure` script to generate makefiles (here from a build directory):
 
-Note: Before running configure it could be necessary to execute the script 
+``` bash
+mkdir build; cd build
+../configure --enable-precision=double --enable-simd=AVX --enable-comms=mpi-auto --prefix=<path>
+```
 
-       script/filelist
+where `--enable-precision=` set the default precision (`single` or `double`),
+`--enable-simd=` set the SIMD type (see possible values below), `--enable-
+comms=` set the protocol used for communications (`none`, `mpi`, `mpi-auto` or
+`shmem`), and `<path>` should be replaced by the prefix path where you want to
+install Grid. The `mpi-auto` communication option set `configure` to determine
+automatically how to link to MPI. Other options are available, use `configure
+--help` to display them. Like with any other program using GNU autotool, the
+`CXX`, `CXXFLAGS`, `LDFLAGS`, ... environment variables can be modified to
+customise the build.
 
+Finally, you can build and install Grid:
 
+``` bash
+make; make install
+```
 
-For developers:
+To minimise the build time, only the tests at the root of the `tests` directory are built by default. If you want to build tests in the sub-directory `<subdir>` you can execute:
-Use reconfigure_script in the scripts/ directory to create the autotools environment 
 
+``` bash
+make -C tests/<subdir> tests
+```
+
+### Possible SIMD types
+
+The following options can be use with the `--enable-simd=` option to target different SIMD instruction sets:
+
+| String      | Description                            |
+| ----------- | -------------------------------------- |
+| `GEN`       | generic portable vector code           |
+| `SSE4`      | SSE 4.2 (128 bit)                      |
+| `AVX`       | AVX (256 bit)                          |
+| `AVXFMA4`   | AVX (256 bit) + FMA                    |
+| `AVX2`      | AVX 2 (256 bit)                        |
+| `AVX512`    | AVX 512 bit                            |
+| `AVX512MIC` | AVX 512 bit for Intel MIC architecture |
+| `ICMI`      | Intel ICMI instructions (512 bit)      |
+
+Alternatively, some CPU codenames can be directly used:
+
+| String      | Description                            |
+| ----------- | -------------------------------------- |
+| `KNC`       | [Intel Knights Corner](http://ark.intel.com/products/codename/57721/Knights-Corner) |
+| `KNL`       | [Intel Knights Landing](http://ark.intel.com/products/codename/48999/Knights-Landing) |

VERSION

@@ -0,0 +1,4 @@
+Version : 0.5.0
+
+- AVX512, AVX2, AVX, SSE good
+- Clang 3.5 and above, ICPC v16 and above, GCC 4.9 and above

benchmarks/Benchmark_comms.cc

@@ -25,7 +25,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#include <Grid.h>
+#include <Grid/Grid.h>
 
 using namespace std;
 using namespace Grid;
@@ -194,7 +194,128 @@ int main (int argc, char ** argv)
     }
   }  
 
+#if 0
 
+  std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
+  std::cout<<GridLogMessage << "= Benchmarking sequential persistent halo exchange in "<<nmu<<" dimensions"<<std::endl;
+  std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
+  std::cout<<GridLogMessage << "  L  "<<"\t\t"<<" Ls  "<<"\t\t"<<"bytes"<<"\t\t"<<"MB/s uni"<<"\t\t"<<"MB/s bidi"<<std::endl;
+
+
+  for(int lat=4;lat<=32;lat+=2){
+    for(int Ls=1;Ls<=16;Ls*=2){
+
+      std::vector<int> latt_size  ({lat,lat,lat,lat});
+
+      GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
+
+      std::vector<std::vector<HalfSpinColourVectorD> > xbuf(8,std::vector<HalfSpinColourVectorD>(lat*lat*lat*Ls));
+      std::vector<std::vector<HalfSpinColourVectorD> > rbuf(8,std::vector<HalfSpinColourVectorD>(lat*lat*lat*Ls));
+
+
+      int ncomm;
+      int bytes=lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD);
+
+
+      std::vector<CartesianCommunicator::CommsRequest_t> empty;
+      std::vector<std::vector<CartesianCommunicator::CommsRequest_t> > requests_fwd(Nd,empty);
+      std::vector<std::vector<CartesianCommunicator::CommsRequest_t> > requests_bwd(Nd,empty);
+
+      for(int mu=0;mu<4;mu++){
+	ncomm=0;
+	if (mpi_layout[mu]>1 ) {
+	  ncomm++;
+
+	  int comm_proc;
+	  int xmit_to_rank;
+	  int recv_from_rank;
+
+	  comm_proc=1;
+	  Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
+	  Grid.SendToRecvFromInit(requests_fwd[mu],
+				  (void *)&xbuf[mu][0],
+				  xmit_to_rank,
+				  (void *)&rbuf[mu][0],
+				  recv_from_rank,
+				  bytes);
+
+	  comm_proc = mpi_layout[mu]-1;
+	  Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
+	  Grid.SendToRecvFromInit(requests_bwd[mu],
+				  (void *)&xbuf[mu+4][0],
+				  xmit_to_rank,
+				  (void *)&rbuf[mu+4][0],
+				  recv_from_rank,
+				  bytes);
+
+	}
+      }
+
+      {
+	double start=usecond();
+	for(int i=0;i<Nloop;i++){
+	  
+	  for(int mu=0;mu<4;mu++){
+	    
+	    if (mpi_layout[mu]>1 ) {
+	      
+	      Grid.SendToRecvFromBegin(requests_fwd[mu]);
+	      Grid.SendToRecvFromComplete(requests_fwd[mu]);
+	      Grid.SendToRecvFromBegin(requests_bwd[mu]);
+	      Grid.SendToRecvFromComplete(requests_bwd[mu]);
+	    }
+	  }
+	  Grid.Barrier();
+	}
+	
+	double stop=usecond();
+	
+	double dbytes    = bytes;
+	double xbytes    = Nloop*dbytes*2.0*ncomm;
+	double rbytes    = xbytes;
+	double bidibytes = xbytes+rbytes;
+	
+	double time = stop-start;
+	
+	std::cout<<GridLogMessage << lat<<"\t\t"<<Ls<<"\t\t"<<bytes<<"\t\t"<<xbytes/time<<"\t\t"<<bidibytes/time<<std::endl;
+
+      }
+
+
+      {
+	double start=usecond();
+	for(int i=0;i<Nloop;i++){
+	  
+	  for(int mu=0;mu<4;mu++){
+	    
+	    if (mpi_layout[mu]>1 ) {
+	      
+	      Grid.SendToRecvFromBegin(requests_fwd[mu]);
+	      Grid.SendToRecvFromBegin(requests_bwd[mu]);
+	      Grid.SendToRecvFromComplete(requests_fwd[mu]);
+	      Grid.SendToRecvFromComplete(requests_bwd[mu]);
+	    }
+	  }
+	  Grid.Barrier();
+	}
+	
+	double stop=usecond();
+	
+	double dbytes    = bytes;
+	double xbytes    = Nloop*dbytes*2.0*ncomm;
+	double rbytes    = xbytes;
+	double bidibytes = xbytes+rbytes;
+	
+	double time = stop-start;
+	
+	std::cout<<GridLogMessage << lat<<"\t\t"<<Ls<<"\t\t"<<bytes<<"\t\t"<<xbytes/time<<"\t\t"<<bidibytes/time<<std::endl;
+
+      }
+
+    }
+  }
+
+#endif
 
   Grid_finalize();
 }

benchmarks/Benchmark_dwf.cc

@@ -26,7 +26,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#include <Grid.h>
+#include <Grid/Grid.h>
 
 using namespace std;
 using namespace Grid;
@@ -45,6 +45,10 @@ struct scal {
   };
 
 bool overlapComms = false;
+typedef WilsonFermion5D<DomainWallVec5dImplR> WilsonFermion5DR;
+typedef WilsonFermion5D<DomainWallVec5dImplF> WilsonFermion5DF;
+typedef WilsonFermion5D<DomainWallVec5dImplD> WilsonFermion5DD;
+
 
 int main (int argc, char ** argv)
 {
@@ -58,12 +62,18 @@ int main (int argc, char ** argv)
   std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;
 
   std::vector<int> latt4 = GridDefaultLatt();
-  const int Ls=8;
+  const int Ls=16;
   GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
   GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
   GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
   GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
 
+  std::cout << GridLogMessage << "Making s innermost grids"<<std::endl;
+  GridCartesian         * sUGrid   = SpaceTimeGrid::makeFourDimDWFGrid(GridDefaultLatt(),GridDefaultMpi());
+  GridRedBlackCartesian * sUrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(sUGrid);
+  GridCartesian         * sFGrid   = SpaceTimeGrid::makeFiveDimDWFGrid(Ls,UGrid);
+  GridRedBlackCartesian * sFrbGrid = SpaceTimeGrid::makeFiveDimDWFRedBlackGrid(Ls,UGrid);
+
   std::vector<int> seeds4({1,2,3,4});
   std::vector<int> seeds5({5,6,7,8});
 
@@ -76,9 +86,9 @@ int main (int argc, char ** argv)
   LatticeFermion    tmp(FGrid);
   LatticeFermion    err(FGrid);
 
-  ColourMatrix cm = Complex(1.0,0.0);
+  LatticeGaugeField Umu(UGrid); 
+  random(RNG4,Umu);
 
-  LatticeGaugeField Umu(UGrid); random(RNG4,Umu);
   LatticeGaugeField Umu5d(FGrid); 
 
   // replicate across fifth dimension
@@ -114,19 +124,25 @@ int main (int argc, char ** argv)
   RealD mass=0.1;
   RealD M5  =1.8;
 
-  typename DomainWallFermionR::ImplParams params; 
-  params.overlapCommsCompute = overlapComms;
-  
   RealD NP = UGrid->_Nprocessors;
 
-  DomainWallFermionR Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5,params);
+  for(int doasm=1;doasm<2;doasm++){
 
+    QCD::WilsonKernelsStatic::AsmOpt=doasm;
+
+  DomainWallFermionR Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+  
+  std::cout<<GridLogMessage << "Naive wilson implementation "<<std::endl;
   std::cout << GridLogMessage<< "Calling Dw"<<std::endl;
   int ncall =100;
-  {
+  if (1) {
+
+    Dw.ZeroCounters();
     double t0=usecond();
     for(int i=0;i<ncall;i++){
+      __SSC_START;
       Dw.Dhop(src,result,0);
+      __SSC_STOP;
     }
     double t1=usecond();
     
@@ -137,13 +153,140 @@ int main (int argc, char ** argv)
     std::cout<<GridLogMessage << "norm result "<< norm2(result)<<std::endl;
     std::cout<<GridLogMessage << "norm ref    "<< norm2(ref)<<std::endl;
     std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t1-t0)<<std::endl;
-    std::cout<<GridLogMessage << "mflop/s per node =  "<< flops/(t1-t0)/NP<<std::endl;
+    std::cout<<GridLogMessage << "mflop/s per rank =  "<< flops/(t1-t0)/NP<<std::endl;
     err = ref-result; 
     std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
+    assert (norm2(err)< 1.0e-5 );
     Dw.Report();
   }
 
-  exit(0);
+  if (1)
+  {
+    typedef WilsonFermion5D<DomainWallVec5dImplR> WilsonFermion5DR;
+    LatticeFermion ssrc(sFGrid);
+    LatticeFermion sref(sFGrid);
+    LatticeFermion sresult(sFGrid);
+
+    WilsonFermion5DR sDw(Umu,*sFGrid,*sFrbGrid,*sUGrid,*sUrbGrid,M5);
+  
+    for(int x=0;x<latt4[0];x++){
+    for(int y=0;y<latt4[1];y++){
+    for(int z=0;z<latt4[2];z++){
+    for(int t=0;t<latt4[3];t++){
+    for(int s=0;s<Ls;s++){
+      std::vector<int> site({s,x,y,z,t});
+      SpinColourVector tmp;
+      peekSite(tmp,src,site);
+      pokeSite(tmp,ssrc,site);
+    }}}}}
+    std::cout<<GridLogMessage<< "src norms "<< norm2(src)<<" " <<norm2(ssrc)<<std::endl;
+    double t0=usecond();
+    sDw.ZeroCounters();
+    for(int i=0;i<ncall;i++){
+      __SSC_START;
+      sDw.Dhop(ssrc,sresult,0);
+      __SSC_STOP;
+    }
+    double t1=usecond();
+    double volume=Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
+    double flops=1344*volume*ncall;
+
+    std::cout<<GridLogMessage << "Called Dw s_inner "<<ncall<<" times in "<<t1-t0<<" us"<<std::endl;
+    std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t1-t0)<<std::endl;
+    std::cout<<GridLogMessage << "mflop/s per rank =  "<< flops/(t1-t0)/NP<<std::endl;
+    sDw.Report();
+  
+    if(0){
+      for(int i=0;i< PerformanceCounter::NumTypes(); i++ ){
+  sDw.Dhop(ssrc,sresult,0);
+  PerformanceCounter Counter(i);
+  Counter.Start();
+  sDw.Dhop(ssrc,sresult,0);
+  Counter.Stop();
+  Counter.Report();
+      }
+    }
+
+    std::cout<<GridLogMessage<< "res norms "<< norm2(result)<<" " <<norm2(sresult)<<std::endl;
+
+    RealD sum=0;
+    for(int x=0;x<latt4[0];x++){
+    for(int y=0;y<latt4[1];y++){
+    for(int z=0;z<latt4[2];z++){
+    for(int t=0;t<latt4[3];t++){
+    for(int s=0;s<Ls;s++){
+      std::vector<int> site({s,x,y,z,t});
+      SpinColourVector normal, simd;
+      peekSite(normal,result,site);
+      peekSite(simd,sresult,site);
+      sum=sum+norm2(normal-simd);
+      if (norm2(normal-simd) > 1.0e-6 ) {
+	std::cout << "site "<<x<<","<<y<<","<<z<<","<<t<<","<<s<<" "<<norm2(normal-simd)<<std::endl;
+	std::cout << "site "<<x<<","<<y<<","<<z<<","<<t<<","<<s<<" normal "<<normal<<std::endl;
+	std::cout << "site "<<x<<","<<y<<","<<z<<","<<t<<","<<s<<" simd   "<<simd<<std::endl;
+      }
+    }}}}}
+    std::cout<<GridLogMessage<<" difference between normal and simd is "<<sum<<std::endl;
+    assert (sum< 1.0e-5 );
+
+
+    if (1) {
+
+      LatticeFermion sr_eo(sFGrid);
+
+      LatticeFermion ssrc_e (sFrbGrid);
+      LatticeFermion ssrc_o (sFrbGrid);
+      LatticeFermion sr_e   (sFrbGrid);
+      LatticeFermion sr_o   (sFrbGrid);
+
+      pickCheckerboard(Even,ssrc_e,ssrc);
+      pickCheckerboard(Odd,ssrc_o,ssrc);
+
+      setCheckerboard(sr_eo,ssrc_o);
+      setCheckerboard(sr_eo,ssrc_e);
+
+      sr_e = zero;
+      sr_o = zero;
+
+      sDw.ZeroCounters();
+      sDw.stat.init("DhopEO");
+      double t0=usecond();
+      for (int i = 0; i < ncall; i++) {
+        sDw.DhopEO(ssrc_o, sr_e, DaggerNo);
+      }
+      double t1=usecond();
+      sDw.stat.print();
+
+      double volume=Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
+      double flops=(1344.0*volume*ncall)/2;
+
+      std::cout<<GridLogMessage << "sDeo mflop/s =   "<< flops/(t1-t0)<<std::endl;
+      std::cout<<GridLogMessage << "sDeo mflop/s per rank   "<< flops/(t1-t0)/NP<<std::endl;
+      sDw.Report();
+
+      sDw.DhopEO(ssrc_o,sr_e,DaggerNo);
+      sDw.DhopOE(ssrc_e,sr_o,DaggerNo);
+      sDw.Dhop  (ssrc  ,sresult,DaggerNo);
+
+      pickCheckerboard(Even,ssrc_e,sresult);
+      pickCheckerboard(Odd ,ssrc_o,sresult);
+      ssrc_e = ssrc_e - sr_e;
+      RealD error = norm2(ssrc_e);
+
+      std::cout<<GridLogMessage << "sE norm diff   "<< norm2(ssrc_e)<< "  vec nrm"<<norm2(sr_e) <<std::endl;
+      ssrc_o = ssrc_o - sr_o;
+
+      error+= norm2(ssrc_o);
+      std::cout<<GridLogMessage << "sO norm diff   "<< norm2(ssrc_o)<< "  vec nrm"<<norm2(sr_o) <<std::endl;
+      if(error>1.0e-5) { 
+	setCheckerboard(ssrc,ssrc_o);
+	setCheckerboard(ssrc,ssrc_e);
+	std::cout<< ssrc << std::endl;
+      }
+    }
+
+
+  }
 
   if (1)
   { // Naive wilson dag implementation
@@ -165,12 +308,13 @@ int main (int argc, char ** argv)
     ref = -0.5*ref;
   }
   Dw.Dhop(src,result,1);
+  std::cout << GridLogMessage << "Naive wilson implementation Dag" << std::endl;
   std::cout<<GridLogMessage << "Called DwDag"<<std::endl;
   std::cout<<GridLogMessage << "norm result "<< norm2(result)<<std::endl;
   std::cout<<GridLogMessage << "norm ref    "<< norm2(ref)<<std::endl;
   err = ref-result; 
   std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
-
+  assert(norm2(err)<1.0e-5);
   LatticeFermion src_e (FrbGrid);
   LatticeFermion src_o (FrbGrid);
   LatticeFermion r_e   (FrbGrid);
@@ -186,6 +330,7 @@ int main (int argc, char ** argv)
   std::cout<<GridLogMessage << "src_o"<<norm2(src_o)<<std::endl;
 
   {
+    Dw.ZeroCounters();
     double t0=usecond();
     for(int i=0;i<ncall;i++){
       Dw.DhopEO(src_o,r_e,DaggerNo);
@@ -196,7 +341,8 @@ int main (int argc, char ** argv)
     double flops=(1344.0*volume*ncall)/2;
 
     std::cout<<GridLogMessage << "Deo mflop/s =   "<< flops/(t1-t0)<<std::endl;
-    std::cout<<GridLogMessage << "Deo mflop/s per node   "<< flops/(t1-t0)/NP<<std::endl;
+    std::cout<<GridLogMessage << "Deo mflop/s per rank   "<< flops/(t1-t0)/NP<<std::endl;
+    Dw.Report();
   }
   Dw.DhopEO(src_o,r_e,DaggerNo);
   Dw.DhopOE(src_e,r_o,DaggerNo);
@@ -211,11 +357,17 @@ int main (int argc, char ** argv)
 
   err = r_eo-result; 
   std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
+  assert(norm2(err)<1.0e-5);
 
   pickCheckerboard(Even,src_e,err);
   pickCheckerboard(Odd,src_o,err);
   std::cout<<GridLogMessage << "norm diff even  "<< norm2(src_e)<<std::endl;
   std::cout<<GridLogMessage << "norm diff odd   "<< norm2(src_o)<<std::endl;
+  assert(norm2(src_e)<1.0e-5);
+  assert(norm2(src_o)<1.0e-5);
+
+
+  }
 
   Grid_finalize();
 }

benchmarks/Benchmark_dwf_ntpf.cc

@@ -0,0 +1,153 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./benchmarks/Benchmark_dwf.cc
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: paboyle <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/Grid.h>
+
+using namespace std;
+using namespace Grid;
+using namespace Grid::QCD;
+
+template<class d>
+struct scal {
+  d internal;
+};
+
+  Gamma::GammaMatrix Gmu [] = {
+    Gamma::GammaX,
+    Gamma::GammaY,
+    Gamma::GammaZ,
+    Gamma::GammaT
+  };
+
+bool overlapComms = false;
+
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  if( GridCmdOptionExists(argv,argv+argc,"--asynch") ){
+    overlapComms = true;
+  }
+
+  int threads = GridThread::GetThreads();
+  std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;
+
+  std::vector<int> latt4 = GridDefaultLatt();
+  const int Ls=16;
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
+  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
+
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+
+  GridParallelRNG          RNG4(UGrid);  RNG4.SeedFixedIntegers(seeds4);
+  GridParallelRNG          RNG5(FGrid);  RNG5.SeedFixedIntegers(seeds5);
+
+  LatticeFermion src   (FGrid); random(RNG5,src);
+  LatticeFermion result(FGrid); result=zero;
+  LatticeFermion    ref(FGrid);    ref=zero;
+  LatticeFermion    tmp(FGrid);
+  LatticeFermion    err(FGrid);
+
+  ColourMatrix cm = Complex(1.0,0.0);
+
+  LatticeGaugeField Umu(UGrid); 
+  random(RNG4,Umu);
+
+  LatticeGaugeField Umu5d(FGrid); 
+
+  // replicate across fifth dimension
+  for(int ss=0;ss<Umu._grid->oSites();ss++){
+    for(int s=0;s<Ls;s++){
+      Umu5d._odata[Ls*ss+s] = Umu._odata[ss];
+    }
+  }
+
+  ////////////////////////////////////
+  // Naive wilson implementation
+  ////////////////////////////////////
+  std::vector<LatticeColourMatrix> U(4,FGrid);
+  for(int mu=0;mu<Nd;mu++){
+    U[mu] = PeekIndex<LorentzIndex>(Umu5d,mu);
+  }
+
+  if (1)
+  {
+    ref = zero;
+    for(int mu=0;mu<Nd;mu++){
+
+      tmp = U[mu]*Cshift(src,mu+1,1);
+      ref=ref + tmp - Gamma(Gmu[mu])*tmp;
+
+      tmp =adj(U[mu])*src;
+      tmp =Cshift(tmp,mu+1,-1);
+      ref=ref + tmp + Gamma(Gmu[mu])*tmp;
+    }
+    ref = -0.5*ref;
+  }
+
+  RealD mass=0.1;
+  RealD M5  =1.8;
+
+  typename DomainWallFermionR::ImplParams params; 
+  params.overlapCommsCompute = overlapComms;
+  
+  RealD NP = UGrid->_Nprocessors;
+
+
+  QCD::WilsonKernelsStatic::AsmOpt=1;
+
+  DomainWallFermionR Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5,params);
+  
+  std::cout<<GridLogMessage << "Calling Dw"<<std::endl;
+  int ncall =50;
+  if (1) {
+
+    double t0=usecond();
+    for(int i=0;i<ncall;i++){
+      Dw.Dhop(src,result,0);
+    }
+    double t1=usecond();
+    
+    double volume=Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
+    double flops=1344*volume*ncall;
+
+    std::cout<<GridLogMessage << "Called Dw "<<ncall<<" times in "<<t1-t0<<" us"<<std::endl;
+    std::cout<<GridLogMessage << "norm result "<< norm2(result)<<std::endl;
+    std::cout<<GridLogMessage << "norm ref    "<< norm2(ref)<<std::endl;
+    std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t1-t0)<<std::endl;
+    std::cout<<GridLogMessage << "mflop/s per node =  "<< flops/(t1-t0)/NP<<std::endl;
+    err = ref-result; 
+    std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
+    //    Dw.Report();
+  }
+  Grid_finalize();
+}

benchmarks/Benchmark_dwf_sweep.cc

@@ -0,0 +1,364 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./benchmarks/Benchmark_dwf.cc
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: paboyle <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/Grid.h>
+
+using namespace std;
+using namespace Grid;
+using namespace Grid::QCD;
+
+template<class d>
+struct scal {
+  d internal;
+};
+
+  Gamma::GammaMatrix Gmu [] = {
+    Gamma::GammaX,
+    Gamma::GammaY,
+    Gamma::GammaZ,
+    Gamma::GammaT
+  };
+
+void benchDw(std::vector<int> & L, int Ls, int threads, int report =0 );
+void benchsDw(std::vector<int> & L, int Ls, int threads, int report=0 );
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  const int Ls=8;
+  int threads = GridThread::GetThreads();
+  std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;
+
+  if ( getenv("ASMOPT") )  {
+    QCD::WilsonKernelsStatic::AsmOpt=1;
+  } else { 
+    QCD::WilsonKernelsStatic::AsmOpt=0;
+  }
+
+  std::cout<<GridLogMessage << "=========================================================================="<<std::endl;
+  std::cout<<GridLogMessage << "= Benchmarking DWF"<<std::endl;
+  std::cout<<GridLogMessage << "=========================================================================="<<std::endl;
+  std::cout<<GridLogMessage << "Volume \t\t\tProcs \t Dw \t eoDw \t sDw \t eosDw (Mflop/s)  "<<std::endl;
+  std::cout<<GridLogMessage << "=========================================================================="<<std::endl;
+
+  int Lmax=32;
+  int dmin=0;
+  if ( getenv("LMAX") ) Lmax=atoi(getenv("LMAX"));
+  if ( getenv("DMIN") ) dmin=atoi(getenv("DMIN"));
+  for (int L=8;L<=Lmax;L*=2){
+    std::vector<int> latt4(4,L);
+    for(int d=4;d>dmin;d--){
+      if ( d<=3 ) latt4[d]*=2;
+      std::cout << GridLogMessage <<"\t";
+      for(int d=0;d<Nd;d++){
+	std::cout<<latt4[d]<<"x";
+      }
+      std::cout <<Ls<<"\t" ;
+      benchDw (latt4,Ls,threads,0);
+      benchsDw(latt4,Ls,threads,0);
+      std::cout<<std::endl;
+    }
+  }
+  std::cout<<GridLogMessage << "=========================================================================="<<std::endl;
+  {
+    std::vector<int> latt4(4,16);
+    std::cout<<GridLogMessage << "16^4 Dw miss rate"<<std::endl;
+    benchDw (latt4,Ls,threads,1);
+    std::cout<<GridLogMessage << "16^4 sDw miss rate"<<std::endl;
+    benchsDw(latt4,Ls,threads,1);
+  }
+
+  Grid_finalize();
+}
+
+#undef CHECK
+
+void benchDw(std::vector<int> & latt4, int Ls, int threads,int report )
+{
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(latt4, GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
+  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
+
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+
+#ifdef CHECK 
+  GridParallelRNG          RNG4(UGrid);  RNG4.SeedFixedIntegers(seeds4);
+  GridParallelRNG          RNG5(FGrid);  RNG5.SeedFixedIntegers(seeds5);
+  LatticeFermion src   (FGrid); random(RNG5,src);
+  LatticeGaugeField Umu(UGrid); 
+  random(RNG4,Umu);
+#else 
+  LatticeFermion src   (FGrid); src=zero;
+  LatticeGaugeField Umu(UGrid); Umu=zero;
+#endif
+
+  LatticeFermion result(FGrid); result=zero;
+  LatticeFermion    ref(FGrid);    ref=zero;
+  LatticeFermion    tmp(FGrid);
+  LatticeFermion    err(FGrid);
+
+  ColourMatrix cm = Complex(1.0,0.0);
+
+  LatticeGaugeField Umu5d(FGrid); 
+
+  // replicate across fifth dimension
+  for(int ss=0;ss<Umu._grid->oSites();ss++){
+    for(int s=0;s<Ls;s++){
+      Umu5d._odata[Ls*ss+s] = Umu._odata[ss];
+    }
+  }
+
+  ////////////////////////////////////
+  // Naive wilson implementation
+  ////////////////////////////////////
+  std::vector<LatticeColourMatrix> U(4,FGrid);
+  for(int mu=0;mu<Nd;mu++){
+    U[mu] = PeekIndex<LorentzIndex>(Umu5d,mu);
+  }
+
+#ifdef CHECK
+  if (1) {
+
+    ref = zero;
+    for(int mu=0;mu<Nd;mu++){
+      tmp = U[mu]*Cshift(src,mu+1,1);
+      ref=ref + tmp - Gamma(Gmu[mu])*tmp;
+
+      tmp =adj(U[mu])*src;
+      tmp =Cshift(tmp,mu+1,-1);
+      ref=ref + tmp + Gamma(Gmu[mu])*tmp;
+    }
+    ref = -0.5*ref;
+  }
+#endif
+
+  RealD mass=0.1;
+  RealD M5  =1.8;
+  RealD NP = UGrid->_Nprocessors;
+
+  DomainWallFermionR Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+  
+  double t0=usecond();
+  Dw.Dhop(src,result,0);
+  double t1=usecond();
+
+#ifdef TIMERS_OFF
+    int ncall =10;
+#else
+  int ncall =1+(int) ((5.0*1000*1000)/(t1-t0));
+#endif
+
+  if (ncall < 5 ) exit(0);
+
+  Dw.Dhop(src,result,0);
+
+  PerformanceCounter Counter(8);
+  Counter.Start();
+  t0=usecond();
+  for(int i=0;i<ncall;i++){
+    Dw.Dhop(src,result,0);
+  }
+  t1=usecond();
+  Counter.Stop();
+  if ( report ) {
+    Counter.Report();
+  }
+  
+  if ( ! report ) {
+    double volume=Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
+    double flops=1344*volume*ncall;
+    std::cout <<"\t"<<NP<< "\t"<<flops/(t1-t0)<< "\t";
+  }
+  
+#ifdef CHECK
+  err = ref-result; 
+  RealD errd = norm2(err);
+  if ( errd> 1.0e-4 ) {
+    std::cout<<GridLogMessage << "oops !!! norm diff   "<< norm2(err)<<std::endl;
+    exit(-1);
+  }
+#endif
+    
+  LatticeFermion src_e (FrbGrid);
+  LatticeFermion src_o (FrbGrid);
+  LatticeFermion r_e   (FrbGrid);
+  LatticeFermion r_o   (FrbGrid);
+  LatticeFermion r_eo  (FGrid);
+  
+  pickCheckerboard(Even,src_e,src);
+  pickCheckerboard(Odd,src_o,src);
+  
+  {
+    Dw.DhopEO(src_o,r_e,DaggerNo);
+    double t0=usecond();
+    for(int i=0;i<ncall;i++){
+      Dw.DhopEO(src_o,r_e,DaggerNo);
+    }
+    double t1=usecond();
+    
+    if(!report){
+      double volume=Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
+      double flops=(1344.0*volume*ncall)/2;
+      std::cout<< flops/(t1-t0);
+    }
+  }
+}
+
+#define CHECK_SDW
+void benchsDw(std::vector<int> & latt4, int Ls, int threads, int report )
+{
+
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(latt4, GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
+  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
+
+  GridCartesian         * sUGrid   = SpaceTimeGrid::makeFourDimDWFGrid(latt4,GridDefaultMpi());
+  GridRedBlackCartesian * sUrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(sUGrid);
+  GridCartesian         * sFGrid   = SpaceTimeGrid::makeFiveDimDWFGrid(Ls,UGrid);
+  GridRedBlackCartesian * sFrbGrid = SpaceTimeGrid::makeFiveDimDWFRedBlackGrid(Ls,UGrid);
+
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+
+#ifdef CHECK_SDW
+  GridParallelRNG          RNG4(UGrid);  RNG4.SeedFixedIntegers(seeds4);
+  GridParallelRNG          RNG5(FGrid);  RNG5.SeedFixedIntegers(seeds5);
+  LatticeFermion src   (FGrid); random(RNG5,src);
+  LatticeGaugeField Umu(UGrid); 
+  random(RNG4,Umu);
+#else 
+  LatticeFermion src   (FGrid); src=zero;
+  LatticeGaugeField Umu(UGrid); Umu=zero;
+#endif
+
+  LatticeFermion result(FGrid); result=zero;
+  LatticeFermion    ref(FGrid);    ref=zero;
+  LatticeFermion    tmp(FGrid);
+  LatticeFermion    err(FGrid);
+
+  ColourMatrix cm = Complex(1.0,0.0);
+
+  LatticeGaugeField Umu5d(FGrid); 
+
+  // replicate across fifth dimension
+  for(int ss=0;ss<Umu._grid->oSites();ss++){
+    for(int s=0;s<Ls;s++){
+      Umu5d._odata[Ls*ss+s] = Umu._odata[ss];
+    }
+  }
+
+  RealD mass=0.1;
+  RealD M5  =1.8;
+
+  typedef WilsonFermion5D<DomainWallVec5dImplR> WilsonFermion5DR;
+  LatticeFermion ssrc(sFGrid);
+  LatticeFermion sref(sFGrid);
+  LatticeFermion sresult(sFGrid);
+  WilsonFermion5DR sDw(Umu,*sFGrid,*sFrbGrid,*sUGrid,*sUrbGrid,M5);
+  
+  for(int x=0;x<latt4[0];x++){
+  for(int y=0;y<latt4[1];y++){
+  for(int z=0;z<latt4[2];z++){
+  for(int t=0;t<latt4[3];t++){
+  for(int s=0;s<Ls;s++){
+    std::vector<int> site({s,x,y,z,t});
+    SpinColourVector tmp;
+    peekSite(tmp,src,site);
+    pokeSite(tmp,ssrc,site);
+  }}}}}
+
+  double t0=usecond();
+  sDw.Dhop(ssrc,sresult,0);
+  double t1=usecond();
+
+#ifdef TIMERS_OFF
+  int ncall =10;
+#else 
+  int ncall =1+(int) ((5.0*1000*1000)/(t1-t0));
+#endif
+
+  PerformanceCounter Counter(8);
+  Counter.Start();
+  t0=usecond();
+  for(int i=0;i<ncall;i++){
+    sDw.Dhop(ssrc,sresult,0);
+  }
+  t1=usecond();
+  Counter.Stop();
+  
+  if ( report ) {
+    Counter.Report();
+  } else { 
+    double volume=Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
+    double flops=1344*volume*ncall;
+    std::cout<<"\t"<< flops/(t1-t0);
+  }
+
+  LatticeFermion sr_eo(sFGrid);
+  LatticeFermion serr(sFGrid);
+  
+  LatticeFermion ssrc_e (sFrbGrid);
+  LatticeFermion ssrc_o (sFrbGrid);
+  LatticeFermion sr_e   (sFrbGrid);
+  LatticeFermion sr_o   (sFrbGrid);
+      
+  pickCheckerboard(Even,ssrc_e,ssrc);
+  pickCheckerboard(Odd,ssrc_o,ssrc);
+  
+  setCheckerboard(sr_eo,ssrc_o);
+  setCheckerboard(sr_eo,ssrc_e);
+    
+  sr_e = zero;
+  sr_o = zero;
+  
+  sDw.DhopEO(ssrc_o,sr_e,DaggerNo);
+  PerformanceCounter CounterSdw(8);
+  CounterSdw.Start();
+  t0=usecond();
+  for(int i=0;i<ncall;i++){
+    __SSC_START;
+    sDw.DhopEO(ssrc_o,sr_e,DaggerNo);
+    __SSC_STOP;
+  }
+  t1=usecond();
+  CounterSdw.Stop();
+
+  if ( report ) { 
+    CounterSdw.Report();
+  } else {
+    double volume=Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
+    double flops=(1344.0*volume*ncall)/2;
+    std::cout<<"\t"<< flops/(t1-t0);
+  }
+}
+
+

benchmarks/Benchmark_memory_asynch.cc

@@ -26,7 +26,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#include <Grid.h>
+#include <Grid/Grid.h>
 
 using namespace std;
 using namespace Grid;

benchmarks/Benchmark_memory_bandwidth.cc

@@ -26,7 +26,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#include <Grid.h>
+#include <Grid/Grid.h>
 
 using namespace std;
 using namespace Grid;

benchmarks/Benchmark_su3.cc

@@ -26,7 +26,7 @@ Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#include <Grid.h>
+#include <Grid/Grid.h>
 
 using namespace std;
 using namespace Grid;

benchmarks/Benchmark_wilson.cc

@@ -26,7 +26,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#include <Grid.h>
+#include <Grid/Grid.h>
 
 using namespace std;
 using namespace Grid;

benchmarks/Benchmark_wilson_sweep.cc

@@ -0,0 +1,117 @@
+/*************************************************************************************
+    Grid physics library, www.github.com/paboyle/Grid 
+    Source file: ./benchmarks/Benchmark_wilson.cc
+    Copyright (C) 2015
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Richard Rollins <rprollins@users.noreply.github.com>
+    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/Grid.h>
+
+using namespace std;
+using namespace Grid;
+using namespace Grid::QCD;
+
+template<class d>
+struct scal {
+  d internal;
+};
+
+Gamma::GammaMatrix Gmu [] = {
+  Gamma::GammaX,
+  Gamma::GammaY,
+  Gamma::GammaZ,
+  Gamma::GammaT
+};
+
+bool overlapComms = false;
+
+void bench_wilson (
+		   LatticeFermion &    src,
+		   LatticeFermion & result,
+		   WilsonFermionR &     Dw,
+		   double const     volume,
+		   int const           dag );
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+  if( GridCmdOptionExists(argv,argv+argc,"--asynch") ){ overlapComms = true; }
+  typename WilsonFermionR::ImplParams params;
+  params.overlapCommsCompute = overlapComms;
+
+  std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
+  std::vector<int> mpi_layout  = GridDefaultMpi();
+  std::vector<int> seeds({1,2,3,4});
+  RealD mass = 0.1;
+
+  std::cout<<GridLogMessage << "============================================================================="<< std::endl;
+  std::cout<<GridLogMessage << "= Benchmarking Wilson" << std::endl;
+  std::cout<<GridLogMessage << "============================================================================="<< std::endl;
+  std::cout<<GridLogMessage << "Volume\t\t\tWilson/MFLOPs\tWilsonDag/MFLOPs" << std::endl;
+  std::cout<<GridLogMessage << "============================================================================="<< std::endl;
+
+  int Lmax = 32;
+  int dmin = 0;
+  if ( getenv("LMAX") ) Lmax=atoi(getenv("LMAX"));
+  if ( getenv("DMIN") ) dmin=atoi(getenv("DMIN"));
+  for (int L=8; L<=Lmax; L*=2)
+    {
+      std::vector<int> latt_size = std::vector<int>(4,L);
+      for(int d=4; d>dmin; d--)
+	{
+	  if ( d<=3 ) { latt_size[d] *= 2; }
+
+	  std::cout << GridLogMessage;
+	  std::copy( latt_size.begin(), --latt_size.end(), std::ostream_iterator<int>( std::cout, std::string("x").c_str() ) );
+	  std::cout << latt_size.back() << "\t\t";
+
+	  GridCartesian           Grid(latt_size,simd_layout,mpi_layout);
+	  GridRedBlackCartesian RBGrid(latt_size,simd_layout,mpi_layout);
+
+	  GridParallelRNG  pRNG(&Grid); pRNG.SeedFixedIntegers(seeds);
+	  LatticeGaugeField Umu(&Grid); random(pRNG,Umu);
+	  LatticeFermion    src(&Grid); random(pRNG,src);
+	  LatticeFermion result(&Grid); result=zero;
+
+	  double volume = std::accumulate(latt_size.begin(),latt_size.end(),1,std::multiplies<int>());
+
+	  WilsonFermionR Dw(Umu,Grid,RBGrid,mass,params);
+      
+	  bench_wilson(src,result,Dw,volume,DaggerNo);
+	  bench_wilson(src,result,Dw,volume,DaggerYes);
+	  std::cout << std::endl;
+	}
+    }
+
+  std::cout<<GridLogMessage << "============================================================================="<< std::endl;
+  Grid_finalize();
+}
+
+void bench_wilson (
+		   LatticeFermion &    src,
+		   LatticeFermion & result,
+		   WilsonFermionR &     Dw,
+		   double const     volume,
+		   int const           dag )
+{
+  int ncall    = 1000;
+  double t0    = usecond();
+  for(int i=0; i<ncall; i++) { Dw.Dhop(src,result,dag); }
+  double t1    = usecond();
+  double flops = 1344 * volume * ncall;
+  std::cout << flops/(t1-t0) << "\t\t";
+}

benchmarks/Benchmark_zmm.cc

@@ -25,8 +25,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#include <Grid.h>
+#include <Grid/Grid.h>
-#include <PerfCount.h>
 
 
 using namespace Grid;
@@ -41,14 +40,20 @@ int main(int argc,char **argv)
   std::ofstream os("zmm.dat");
 
   os << "#V Ls Lxy Lzt C++ Asm OMP L1 " <<std::endl;
+  std::cout<<GridLogMessage << "====================================================================="<<std::endl;
+  std::cout<<GridLogMessage << "= Benchmarking ZMM"<<std::endl;
+  std::cout<<GridLogMessage << "====================================================================="<<std::endl;
+  std::cout<<GridLogMessage << "Volume \t\t\t\tC++DW/MFLOPs\tASM-DW/MFLOPs\tdiff"<<std::endl;
+  std::cout<<GridLogMessage << "====================================================================="<<std::endl;
   for(int L=4;L<=32;L+=4){
     for(int m=1;m<=2;m++){
       for(int Ls=8;Ls<=16;Ls+=8){
 	std::vector<int> grid({L,L,m*L,m*L});
+  std::cout << GridLogMessage <<"\t";
 	for(int i=0;i<4;i++) { 
 	  std::cout << grid[i]<<"x";
 	}
-	std::cout << Ls<<std::endl;
+	std::cout << Ls<<"\t\t";
 	bench(os,grid,Ls);
       }
     }
@@ -105,7 +110,6 @@ int bench(std::ofstream &os, std::vector<int> &latt4,int Ls)
   RealD M5  =1.8;
   DomainWallFermionR Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
 
-  std::cout<<GridLogMessage << "Calling Dw"<<std::endl;
   int ncall=50;
   double t0=usecond();
   for(int i=0;i<ncall;i++){
@@ -117,16 +121,16 @@ int bench(std::ofstream &os, std::vector<int> &latt4,int Ls)
   double flops=1344*volume/2;
 
   mfc = flops*ncall/(t1-t0);
-  std::cout<<GridLogMessage << "Called C++ Dw"<< " mflop/s =   "<< mfc<<std::endl;
+  std::cout<<mfc<<"\t\t";
 
-  QCD::WilsonFermion5DStatic::AsmOptDslash=1;
+  QCD::WilsonKernelsStatic::AsmOpt=1;
   t0=usecond();
   for(int i=0;i<ncall;i++){
     Dw.DhopOE(srce,resulta,0);
   }
   t1=usecond();
   mfa = flops*ncall/(t1-t0);
-  std::cout<<GridLogMessage << "Called ASM Dw"<< " mflop/s =   "<< mfa<<std::endl;
+  std::cout<<mfa<<"\t\t";
   /*
   int dag=DaggerNo;
   t0=usecond();
@@ -164,8 +168,7 @@ int bench(std::ofstream &os, std::vector<int> &latt4,int Ls)
   //resulta = (-0.5) * resulta;
 
   diff = resulto-resulta;
-  std::cout<<GridLogMessage << "diff "<< norm2(diff)<<std::endl;
+  std::cout<<norm2(diff)<<std::endl;
-  std::cout<<std::endl;
   return 0;
 }
 

benchmarks/Make.inc

@@ -1,31 +0,0 @@
-
-bin_PROGRAMS = Benchmark_comms Benchmark_dwf Benchmark_memory_asynch Benchmark_memory_bandwidth Benchmark_su3 Benchmark_wilson Benchmark_zmm
-
-
-Benchmark_comms_SOURCES=Benchmark_comms.cc
-Benchmark_comms_LDADD=-lGrid
-
-
-Benchmark_dwf_SOURCES=Benchmark_dwf.cc
-Benchmark_dwf_LDADD=-lGrid
-
-
-Benchmark_memory_asynch_SOURCES=Benchmark_memory_asynch.cc
-Benchmark_memory_asynch_LDADD=-lGrid
-
-
-Benchmark_memory_bandwidth_SOURCES=Benchmark_memory_bandwidth.cc
-Benchmark_memory_bandwidth_LDADD=-lGrid
-
-
-Benchmark_su3_SOURCES=Benchmark_su3.cc
-Benchmark_su3_LDADD=-lGrid
-
-
-Benchmark_wilson_SOURCES=Benchmark_wilson.cc
-Benchmark_wilson_LDADD=-lGrid
-
-
-Benchmark_zmm_SOURCES=Benchmark_zmm.cc
-Benchmark_zmm_LDADD=-lGrid
-

benchmarks/Makefile.am

@@ -1,8 +1 @@
-# additional include paths necessary to compile the C++ library
-AM_CXXFLAGS = -I$(top_srcdir)/lib
-AM_LDFLAGS = -L$(top_builddir)/lib
-
-#
-# Test code
-#
 include Make.inc

bootstrap.sh

@@ -0,0 +1,19 @@
+#!/usr/bin/env bash
+
+EIGEN_URL='http://bitbucket.org/eigen/eigen/get/3.2.9.tar.bz2'
+FFTW_URL=http://www.fftw.org/fftw-3.3.4.tar.gz
+
+echo "-- deploying Eigen source..."
+wget ${EIGEN_URL} --no-check-certificate
+./scripts/update_eigen.sh `basename ${EIGEN_URL}`
+rm `basename ${EIGEN_URL}`
+
+echo "-- copying fftw prototypes..."
+wget ${FFTW_URL}
+./scripts/update_fftw.sh `basename ${FFTW_URL}`
+rm `basename ${FFTW_URL}`
+
+echo '-- generating Make.inc files...'
+./scripts/filelist
+echo '-- generating configure script...'
+autoreconf -fvi

configure.ac

@@ -1,193 +1,269 @@
-#                         -*- Autoconf -*-
-# Process this file with autoconf to produce a configure script.
-#
-# Project Grid package  
-# 
-# Time-stamp: <2015-07-10 17:46:21 neo>
-
 AC_PREREQ([2.63])
-AC_INIT([Grid], [1.0], [paboyle@ph.ed.ac.uk])
+AC_INIT([Grid], [0.5.1-dev], [https://github.com/paboyle/Grid], [Grid])
-AC_CANONICAL_SYSTEM
+AC_CANONICAL_BUILD
+AC_CANONICAL_HOST
+AC_CANONICAL_TARGET
 AM_INIT_AUTOMAKE(subdir-objects)
 AC_CONFIG_MACRO_DIR([m4])
 AC_CONFIG_SRCDIR([lib/Grid.h])
 AC_CONFIG_HEADERS([lib/Config.h])
 m4_ifdef([AM_SILENT_RULES], [AM_SILENT_RULES([yes])])
 
-AC_MSG_NOTICE([
 
-:::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::
+############### Checks for programs
-Configuring $PACKAGE v$VERSION  for $host
-:::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::
-])
-
-# Checks for programs.
 AC_LANG(C++)
+CXXFLAGS="-O3 $CXXFLAGS"
 AC_PROG_CXX
-AC_OPENMP
 AC_PROG_RANLIB
-#AX_CXX_COMPILE_STDCXX_11(noext, mandatory)
-AX_EXT
 
-# Checks for libraries.
+############ openmp  ###############
-#AX_GCC_VAR_ATTRIBUTE(aligned)
+AC_OPENMP
 
-# Checks for header files.
+ac_openmp=no
+
+if test "${OPENMP_CXXFLAGS}X" != "X"; then
+ac_openmp=yes
+AM_CXXFLAGS="$OPENMP_CXXFLAGS $AM_CXXFLAGS"
+AM_LDFLAGS="$OPENMP_CXXFLAGS $AM_LDFLAGS"
+fi
+
+############### Checks for header files
 AC_CHECK_HEADERS(stdint.h)
 AC_CHECK_HEADERS(mm_malloc.h)
 AC_CHECK_HEADERS(malloc/malloc.h)
 AC_CHECK_HEADERS(malloc.h)
 AC_CHECK_HEADERS(endian.h)
 AC_CHECK_HEADERS(execinfo.h)
-AC_CHECK_HEADERS(gmp.h)
 AC_CHECK_DECLS([ntohll],[], [], [[#include <arpa/inet.h>]])
 AC_CHECK_DECLS([be64toh],[], [], [[#include <arpa/inet.h>]])
 
-# Checks for typedefs, structures, and compiler characteristics.
+############### Checks for typedefs, structures, and compiler characteristics
 AC_TYPE_SIZE_T
 AC_TYPE_UINT32_T
 AC_TYPE_UINT64_T
 
-# Checks for library functions.
+############### GMP and MPFR #################
-echo
+AC_ARG_WITH([gmp],
-echo Checking libraries 
+    [AS_HELP_STRING([--with-gmp=prefix],
-echo :::::::::::::::::::::::::::::::::::::::::::
+    [try this for a non-standard install prefix of the GMP library])],
+    [AM_CXXFLAGS="-I$with_gmp/include $AM_CXXFLAGS"]
+    [AM_LDFLAGS="-L$with_gmp/lib $AM_LDFLAGS"])
+AC_ARG_WITH([mpfr],
+    [AS_HELP_STRING([--with-mpfr=prefix],
+    [try this for a non-standard install prefix of the MPFR library])],
+    [AM_CXXFLAGS="-I$with_mpfr/include $AM_CXXFLAGS"]
+    [AM_LDFLAGS="-L$with_mpfr/lib $AM_LDFLAGS"])
 
+################## lapack ####################
+AC_ARG_ENABLE([lapack],
+    [AC_HELP_STRING([--enable-lapack=yes|no|prefix], [enable LAPACK])], 
+    [ac_LAPACK=${enable_lapack}],[ac_LAPACK=no])
+
+case ${ac_LAPACK} in
+    no)
+        ;;
+    yes)
+        AC_DEFINE([USE_LAPACK],[1],[use LAPACK]);;
+    *)
+        AM_CXXFLAGS="-I$ac_LAPACK/include $AM_CXXFLAGS"
+        AM_LDFLAGS="-L$ac_LAPACK/lib $AM_LDFLAGS"
+        AC_DEFINE([USE_LAPACK],[1],[use LAPACK])
+esac
+
+################## first-touch ####################
+AC_ARG_ENABLE([numa],
+    [AC_HELP_STRING([--enable-numa=yes|no|prefix], [enable first touch numa opt])], 
+    [ac_NUMA=${enable_NUMA}],[ac_NUMA=no])
+
+case ${ac_NUMA} in
+    no)
+        ;;
+    yes)
+        AC_DEFINE([GRID_NUMA],[1],[First touch numa locality]);;
+    *)
+        AC_DEFINE([GRID_NUMA],[1],[First touch numa locality]);;
+esac
+
+################## FFTW3 ####################
+AC_ARG_WITH([fftw],    
+            [AS_HELP_STRING([--with-fftw=prefix],
+            [try this for a non-standard install prefix of the FFTW3 library])],
+            [AM_CXXFLAGS="-I$with_fftw/include $AM_CXXFLAGS"]
+            [AM_LDFLAGS="-L$with_fftw/lib $AM_LDFLAGS"])
+
+################ Get compiler informations
+AC_LANG([C++])
+AX_CXX_COMPILE_STDCXX_11([noext],[mandatory])
+AX_COMPILER_VENDOR
+AC_DEFINE_UNQUOTED([CXX_COMP_VENDOR],["$ax_cv_cxx_compiler_vendor"],
+      [vendor of C++ compiler that will compile the code])
+AX_GXX_VERSION
+AC_DEFINE_UNQUOTED([GXX_VERSION],["$GXX_VERSION"],
+      [version of g++ that will compile the code])
+
+############### Checks for library functions
+CXXFLAGS_CPY=$CXXFLAGS
+LDFLAGS_CPY=$LDFLAGS
+CXXFLAGS="$AM_CXXFLAGS $CXXFLAGS"
+LDFLAGS="$AM_LDFLAGS $LDFLAGS"
 AC_CHECK_FUNCS([gettimeofday])
+AC_CHECK_LIB([gmp],[__gmpf_init],
+             [AC_CHECK_LIB([mpfr],[mpfr_init],
+                 [AC_DEFINE([HAVE_LIBMPFR], [1], [Define to 1 if you have the `MPFR' library (-lmpfr).])]
+                 [have_mpfr=true]
+                 [LIBS="$LIBS -lmpfr"],
+                 [AC_MSG_ERROR([MPFR library not found])])]
+   	     [AC_DEFINE([HAVE_LIBGMP], [1], [Define to 1 if you have the `GMP' library (-lgmp).])]
+             [have_gmp=true]
+             [LIBS="$LIBS -lgmp"],
+             [AC_MSG_WARN([**** GMP library not found, Grid can still compile but RHMC will not work ****])])
 
+if test "${ac_LAPACK}x" != "nox"; then
+    AC_CHECK_LIB([lapack],[LAPACKE_sbdsdc],[],
+                 [AC_MSG_ERROR("LAPACK enabled but library not found")])
+fi
+AC_CHECK_LIB([fftw3],[fftw_execute],
+  [AC_DEFINE([HAVE_FFTW],[1],[Define to 1 if you have the `FFTW' library (-lfftw3).])]
+  [have_fftw=true]
+  [LIBS="$LIBS -lfftw3 -lfftw3f"],
+  [AC_MSG_WARN([**** FFTW library not found, Grid can still compile but FFT-based routines will not work ****])])
+CXXFLAGS=$CXXFLAGS_CPY
+LDFLAGS=$LDFLAGS_CPY
 
-#
+############### SIMD instruction selection
-# SIMD instructions selection
+AC_ARG_ENABLE([simd],[AC_HELP_STRING([--enable-simd=SSE4|AVX|AVXFMA4|AVXFMA|AVX2|AVX512|AVX512MIC|IMCI|KNL|KNC],\
-#
-
-AC_ARG_ENABLE([simd],[AC_HELP_STRING([--enable-simd=SSE4|AVX|AVXFMA4|AVX2|AVX512|IMCI],\
 	[Select instructions to be SSE4.0, AVX 1.0, AVX 2.0+FMA, AVX 512, IMCI])],\
-	[ac_SIMD=${enable_simd}],[ac_SIMD=DEBUG])
+	[ac_SIMD=${enable_simd}],[ac_SIMD=GEN])
-
-supported=no
-
-ac_ZMM=no;
 
+case ${ax_cv_cxx_compiler_vendor} in
+  clang|gnu)
     case ${ac_SIMD} in
       SSE4)
-       echo Configuring for SSE4
+        AC_DEFINE([SSE4],[1],[SSE4 intrinsics])
-       AC_DEFINE([SSE4],[1],[SSE4 Intrinsics] )
+        SIMD_FLAGS='-msse4.2';;
-       if test x"$ax_cv_support_ssse3_ext" = x"yes"; then  dnl minimal support for SSE4
-         supported=yes
-       else
-  	AC_MSG_WARN([Your processor does not support SSE4 instructions])
-       fi
-     ;;
       AVX)
-       echo Configuring for AVX
+        AC_DEFINE([AVX1],[1],[AVX intrinsics])
-       AC_DEFINE([AVX1],[1],[AVX Intrinsics] )
+        SIMD_FLAGS='-mavx';;
-       if test x"$ax_cv_support_avx_ext" = x"yes"; then  dnl minimal support for AVX
-       supported=yes			  
-       else
-       	AC_MSG_WARN([Your processor does not support AVX instructions])
-       fi
-     ;;
       AVXFMA4)
-       echo Configuring for AVX
+        AC_DEFINE([AVXFMA4],[1],[AVX intrinsics with FMA4])
-       AC_DEFINE([AVXFMA4],[1],[AVX Intrinsics with FMA4] )
+        SIMD_FLAGS='-mavx -mfma4';;
-       if test x"$ax_cv_support_avx_ext" = x"yes"; then  dnl minimal support for AVX
+      AVXFMA)
-       supported=yes			  
+        AC_DEFINE([AVXFMA],[1],[AVX intrinsics with FMA3])
-       else
+        SIMD_FLAGS='-mavx -mfma';;
-       	AC_MSG_WARN([Your processor does not support AVX instructions])
-       fi
-     ;;
       AVX2)
-       echo Configuring for AVX2
+        AC_DEFINE([AVX2],[1],[AVX2 intrinsics])
-       AC_DEFINE([AVX2],[1],[AVX2 Intrinsics] )
+        SIMD_FLAGS='-mavx2 -mfma';;
-       if test x"$ax_cv_support_avx2_ext" = x"yes"; then  dnl minimal support for AVX2
+      AVX512|AVX512MIC|KNL)
-       supported=yes
+        AC_DEFINE([AVX512],[1],[AVX512 intrinsics])
-       else
+        SIMD_FLAGS='-mavx512f -mavx512pf -mavx512er -mavx512cd';;
-       AC_MSG_WARN([Your processor does not support AVX2 instructions])
+      IMCI|KNC)
-       fi
+        AC_DEFINE([IMCI],[1],[IMCI intrinsics for Knights Corner])
-     ;;
+        SIMD_FLAGS='';;
-     AVX512)
+      GEN)
-       echo Configuring for AVX512 
+        AC_DEFINE([GENERIC_VEC],[1],[generic vector code])
-       AC_DEFINE([AVX512],[1],[AVX512 Intrinsics for Knights Landing] )
+        SIMD_FLAGS='';;
-       supported="cross compilation"
+      QPX|BGQ)
-       ac_ZMM=yes;
+        AC_DEFINE([QPX],[1],[QPX intrinsics for BG/Q])
-     ;;
+        SIMD_FLAGS='';;
-     IMCI)
+      *)
-       echo Configuring for IMCI
+        AC_MSG_ERROR(["SIMD option ${ac_SIMD} not supported by the GCC/Clang compiler"]);;
-       AC_DEFINE([IMCI],[1],[IMCI Intrinsics for Knights Corner] )
+    esac;;
-       supported="cross compilation"
+  intel)
-       ac_ZMM=no;
+    case ${ac_SIMD} in
-     ;;
+      SSE4)
-     NEONv8)
+        AC_DEFINE([SSE4],[1],[SSE4 intrinsics])
-       echo Configuring for experimental ARMv8a support 
+        SIMD_FLAGS='-msse4.2 -xsse4.2';;
-       AC_DEFINE([NEONv8],[1],[NEON ARMv8 Experimental support ] )
+      AVX)
-       supported="cross compilation"
+        AC_DEFINE([AVX1],[1],[AVX intrinsics])
-     ;;
+        SIMD_FLAGS='-mavx -xavx';;
-     DEBUG)
+      AVXFMA4)
-       echo Configuring without SIMD support - only for compiler DEBUGGING!
+        AC_DEFINE([AVXFMA4],[1],[AVX intrinsics with FMA4])
-       AC_DEFINE([EMPTY_SIMD],[1],[EMPTY_SIMD only for DEBUGGING] )
+        SIMD_FLAGS='-mavx -mfma';;
-      ;;     
+      AVXFMA)
+        AC_DEFINE([AVXFMA],[1],[AVX intrinsics with FMA4])
+        SIMD_FLAGS='-mavx -mfma';;
+      AVX2)
+        AC_DEFINE([AVX2],[1],[AVX2 intrinsics])
+        SIMD_FLAGS='-march=core-avx2 -xcore-avx2';;
+      AVX512)
+        AC_DEFINE([AVX512],[1],[AVX512 intrinsics])
+        SIMD_FLAGS='-xcore-avx512';;
+      AVX512MIC|KNL)
+        AC_DEFINE([AVX512],[1],[AVX512 intrinsics for Knights Landing])
+        SIMD_FLAGS='-xmic-avx512';;
+      IMCI|KNC)
+        AC_DEFINE([IMCI],[1],[IMCI Intrinsics for Knights Corner])
+        SIMD_FLAGS='';;
+      GEN)
+        AC_DEFINE([GENERIC_VEC],[1],[generic vector code])
+        SIMD_FLAGS='';;
+      *)
+        AC_MSG_ERROR(["SIMD option ${ac_SIMD} not supported by the Intel compiler"]);;
+    esac;;
+  *)
+    AC_MSG_WARN([Compiler unknown, using generic vector code])
+    AC_DEFINE([GENERIC_VEC],[1],[generic vector code]);;
+esac
+AM_CXXFLAGS="$SIMD_FLAGS $AM_CXXFLAGS"
+AM_CFLAGS="$SIMD_FLAGS $AM_CFLAGS"
+
+case ${ac_SIMD} in
+  AVX512|AVX512MIC|KNL)
+    AC_DEFINE([TEST_ZMM],[1],[compile ZMM test]);;
   *)
-     AC_MSG_ERROR([${ac_SIMD} flag unsupported as --enable-simd option\nRun ./configure --help for the list of options]); 
 	;;
 esac
 
-case ${ac_ZMM} in
+############### precision selection
-yes)
-	echo Enabling ZMM source code
-;;
-no)
-	echo Disabling ZMM source code
-;;
-esac
-
-AM_CONDITIONAL(BUILD_ZMM,[ test "X${ac_ZMM}X" == "XyesX" ])
-
 AC_ARG_ENABLE([precision],[AC_HELP_STRING([--enable-precision=single|double],[Select default word size of Real])],[ac_PRECISION=${enable_precision}],[ac_PRECISION=double])
 case ${ac_PRECISION} in
      single)
-       echo default precision is single
        AC_DEFINE([GRID_DEFAULT_PRECISION_SINGLE],[1],[GRID_DEFAULT_PRECISION is SINGLE] )
      ;;
      double)
-       echo default precision is double
        AC_DEFINE([GRID_DEFAULT_PRECISION_DOUBLE],[1],[GRID_DEFAULT_PRECISION is DOUBLE] )
      ;;
 esac
 
-#
+############### communication type selection
-# Comms selection
+AC_ARG_ENABLE([comms],[AC_HELP_STRING([--enable-comms=none|mpi|mpi-auto|shmem],[Select communications])],[ac_COMMS=${enable_comms}],[ac_COMMS=none])
-#
-
-AC_ARG_ENABLE([comms],[AC_HELP_STRING([--enable-comms=none|mpi],[Select communications])],[ac_COMMS=${enable_comms}],[ac_COMMS=none])
 
 case ${ac_COMMS} in
      none)
-       echo Configuring for NO communications
        AC_DEFINE([GRID_COMMS_NONE],[1],[GRID_COMMS_NONE] )
      ;;
+     mpi-auto)
+       AC_DEFINE([GRID_COMMS_MPI],[1],[GRID_COMMS_MPI] )
+       LX_FIND_MPI
+       if test "x$have_CXX_mpi" = 'xno'; then AC_MSG_ERROR(["MPI not found"]); fi
+       AM_CXXFLAGS="$MPI_CXXFLAGS $AM_CXXFLAGS"
+       AM_CFLAGS="$MPI_CFLAGS $AM_CFLAGS"
+       AM_LDFLAGS="`echo $MPI_CXXLDFLAGS | sed -E 's/-l@<:@^ @:>@+//g'` $AM_LDFLAGS"
+       LIBS="`echo $MPI_CXXLDFLAGS | sed -E 's/-L@<:@^ @:>@+//g'` $LIBS"
+     ;;
      mpi)
-       echo Configuring for MPI communications
        AC_DEFINE([GRID_COMMS_MPI],[1],[GRID_COMMS_MPI] )
      ;;
+     mpi3)
+       AC_DEFINE([GRID_COMMS_MPI3],[1],[GRID_COMMS_MPI3] )
+     ;;
      shmem)
-       echo Configuring for SHMEM communications
        AC_DEFINE([GRID_COMMS_SHMEM],[1],[GRID_COMMS_SHMEM] )
      ;;
      *)
      AC_MSG_ERROR([${ac_COMMS} unsupported --enable-comms option]); 
      ;;
 esac
-
 AM_CONDITIONAL(BUILD_COMMS_SHMEM,[ test "X${ac_COMMS}X" == "XshmemX" ])
-AM_CONDITIONAL(BUILD_COMMS_MPI,[ test "X${ac_COMMS}X" == "XmpiX" ])
+AM_CONDITIONAL(BUILD_COMMS_MPI,[ test "X${ac_COMMS}X" == "XmpiX" || test "X${ac_COMMS}X" == "Xmpi-autoX" ])
+AM_CONDITIONAL(BUILD_COMMS_MPI3,[ test "X${ac_COMMS}X" == "Xmpi3X"] )
 AM_CONDITIONAL(BUILD_COMMS_NONE,[ test "X${ac_COMMS}X" == "XnoneX" ])
 
-#
+############### RNG selection
-# RNG selection
-#
 AC_ARG_ENABLE([rng],[AC_HELP_STRING([--enable-rng=ranlux48|mt19937],\
 	[Select Random Number Generator to be used])],\
 	[ac_RNG=${enable_rng}],[ac_RNG=ranlux48])
+
 case ${ac_RNG} in
      ranlux48)
       AC_DEFINE([RNG_RANLUX],[1],[RNG_RANLUX] )
@@ -199,89 +275,92 @@ case ${ac_RNG} in
       AC_MSG_ERROR([${ac_RNG} unsupported --enable-rng option]); 
      ;;
 esac
-#
-# Chroma regression tests
-#
-AC_ARG_ENABLE([chroma],[AC_HELP_STRING([--enable-chroma],[Expect chroma compiled under c++11 ])],ac_CHROMA=yes,ac_CHROMA=no)
 
-case ${ac_CHROMA} in
+############### timer option
+AC_ARG_ENABLE([timers],[AC_HELP_STRING([--enable-timers],\
+	[Enable system dependent high res timers])],\
+	[ac_TIMERS=${enable_timers}],[ac_TIMERS=yes])
+case ${ac_TIMERS} in
      yes)
-       echo Enabling tests regressing to Chroma
+      AC_DEFINE([TIMERS_ON],[1],[TIMERS_ON] )
      ;;
      no)
-       echo Disabling tests regressing to Chroma
+      AC_DEFINE([TIMERS_OFF],[1],[TIMERS_OFF] )
+     ;;
+     *)
+      AC_MSG_ERROR([${ac_TIMERS} unsupported --enable-timers option]); 
+     ;;
+esac
+
+############### Chroma regression test
+AC_ARG_ENABLE([chroma],[AC_HELP_STRING([--enable-chroma],[Expect chroma compiled under c++11 ])],ac_CHROMA=yes,ac_CHROMA=no)
+case ${ac_CHROMA} in
+     yes|no)
      ;;
      *)
        AC_MSG_ERROR([${ac_CHROMA} unsupported --enable-chroma option]); 
      ;;
 esac
-
 AM_CONDITIONAL(BUILD_CHROMA_REGRESSION,[ test "X${ac_CHROMA}X" == "XyesX" ])
 
-#
+############### Doxygen
-# Lapack
+AC_PROG_DOXYGEN
-#
-AC_ARG_ENABLE([lapack],[AC_HELP_STRING([--enable-lapack],[Enable lapack yes/no ])],[ac_LAPACK=${enable_lapack}],[ac_LAPACK=no])
 
-case ${ac_LAPACK} in
+if test -n "$DOXYGEN"
-     yes)
+then
-       echo Enabling lapack
+AC_CONFIG_FILES([docs/doxy.cfg])
-     ;;
+fi
-     no)
-       echo Disabling lapack
-     ;;
-     *)
-       echo Enabling lapack at ${ac_LAPACK}
-     ;;
-esac
 
-AM_CONDITIONAL(USE_LAPACK,[ test "X${ac_LAPACK}X" != "XnoX" ])
+############### Ouput
-AM_CONDITIONAL(USE_LAPACK_LIB,[ test "X${ac_LAPACK}X" != "XyesX" ])
+cwd=`pwd -P`; cd ${srcdir}; abs_srcdir=`pwd -P`; cd ${cwd}
-
+AM_CXXFLAGS="-I${abs_srcdir}/include $AM_CXXFLAGS"
-###################################################################
+AM_CFLAGS="-I${abs_srcdir}/include $AM_CFLAGS"
-# Checks for doxygen support
+AM_LDFLAGS="-L${cwd}/lib $AM_LDFLAGS"
-# if present enables the "make doxyfile" command
+AC_SUBST([AM_CFLAGS])
-#echo
+AC_SUBST([AM_CXXFLAGS])
-#echo Checking doxygen support 
+AC_SUBST([AM_LDFLAGS])
-#echo :::::::::::::::::::::::::::::::::::::::::::
-#AC_PROG_DOXYGEN
-
-#if test -n "$DOXYGEN"
-#then
-#AC_CONFIG_FILES([docs/doxy.cfg])
-#fi
-
-echo
-echo Creating configuration files
-echo :::::::::::::::::::::::::::::::::::::::::::
 AC_CONFIG_FILES(Makefile)
 AC_CONFIG_FILES(lib/Makefile)
 AC_CONFIG_FILES(tests/Makefile)
+AC_CONFIG_FILES(tests/IO/Makefile)
+AC_CONFIG_FILES(tests/core/Makefile)
+AC_CONFIG_FILES(tests/debug/Makefile)
+AC_CONFIG_FILES(tests/forces/Makefile)
+AC_CONFIG_FILES(tests/hmc/Makefile)
+AC_CONFIG_FILES(tests/solver/Makefile)
 AC_CONFIG_FILES(tests/qdpxx/Makefile)
 AC_CONFIG_FILES(benchmarks/Makefile)
 AC_OUTPUT
 
-
 echo "
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 Summary of configuration for $PACKAGE v$VERSION
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-The following features are enabled:
+----- PLATFORM ----------------------------------------
-
 - architecture (build)          : $build_cpu
 - os (build)                    : $build_os
 - architecture (target)         : $target_cpu
 - os (target)                   : $target_os
+- compiler vendor               : ${ax_cv_cxx_compiler_vendor}
+- compiler version              : ${ax_cv_gxx_version}
+----- BUILD OPTIONS -----------------------------------
+- SIMD                          : ${ac_SIMD}
+- Threading                     : ${ac_openmp} 
+- Communications type           : ${ac_COMMS}
+- Default precision             : ${ac_PRECISION}
+- RNG choice                    : ${ac_RNG} 
+- GMP                           : `if test "x$have_gmp" = xtrue; then echo yes; else echo no; fi`
+- LAPACK                        : ${ac_LAPACK}
+- FFTW                          : `if test "x$have_fftw" = xtrue; then echo yes; else echo no; fi`
 - build DOXYGEN documentation   : `if test "x$enable_doc" = xyes; then echo yes; else echo no; fi`
 - graphs and diagrams           : `if test "x$enable_dot" = xyes; then echo yes; else echo no; fi`
-- Supported SIMD flags          : $SIMD_FLAGS
+----- BUILD FLAGS -------------------------------------
-----------------------------------------------------------
+- CXXFLAGS:
-- enabled simd support          : ${ac_SIMD}   (config macro says supported: $supported )
+`echo ${AM_CXXFLAGS} ${CXXFLAGS} | tr ' ' '\n' | sed 's/^-/    -/g'`
-- communications type           : ${ac_COMMS}
+- LDFLAGS:
-- default precision             : ${ac_PRECISION}
+`echo ${AM_LDFLAGS} ${LDFLAGS} | tr ' ' '\n' | sed 's/^-/    -/g'`
-- RNG choice                    : ${ac_RNG} 
+- LIBS:
-- LAPACK	                : ${ac_LAPACK} 
+`echo ${LIBS} | tr ' ' '\n' | sed 's/^-/    -/g'`
-
+-------------------------------------------------------
-
 "

include/Grid

@@ -0,0 +1 @@
+../lib

lib/Algorithms.h

@@ -29,27 +29,28 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_ALGORITHMS_H
 #define GRID_ALGORITHMS_H
 
-#include <algorithms/SparseMatrix.h>
+#include <Grid/algorithms/SparseMatrix.h>
-#include <algorithms/LinearOperator.h>
+#include <Grid/algorithms/LinearOperator.h>
-#include <algorithms/Preconditioner.h>
+#include <Grid/algorithms/Preconditioner.h>
 
-#include <algorithms/approx/Zolotarev.h>
+#include <Grid/algorithms/approx/Zolotarev.h>
-#include <algorithms/approx/Chebyshev.h>
+#include <Grid/algorithms/approx/Chebyshev.h>
-#include <algorithms/approx/Remez.h>
+#include <Grid/algorithms/approx/Remez.h>
-#include <algorithms/approx/MultiShiftFunction.h>
+#include <Grid/algorithms/approx/MultiShiftFunction.h>
 
-#include <algorithms/iterative/ConjugateGradient.h>
+#include <Grid/algorithms/iterative/ConjugateGradient.h>
-#include <algorithms/iterative/ConjugateResidual.h>
+#include <Grid/algorithms/iterative/ConjugateResidual.h>
-#include <algorithms/iterative/NormalEquations.h>
+#include <Grid/algorithms/iterative/NormalEquations.h>
-#include <algorithms/iterative/SchurRedBlack.h>
+#include <Grid/algorithms/iterative/SchurRedBlack.h>
 
-#include <algorithms/iterative/ConjugateGradientMultiShift.h>
+#include <Grid/algorithms/iterative/ConjugateGradientMultiShift.h>
+#include <Grid/algorithms/iterative/ConjugateGradientMixedPrec.h>
 
 // Lanczos support
-#include <algorithms/iterative/MatrixUtils.h>
+#include <Grid/algorithms/iterative/MatrixUtils.h>
-#include <algorithms/iterative/ImplicitlyRestartedLanczos.h>
+#include <Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h>
 
-#include <algorithms/CoarsenedMatrix.h>
+#include <Grid/algorithms/CoarsenedMatrix.h>
 
 // Eigen/lanczos
 // EigCg

lib/AlignedAllocator.h

@@ -40,14 +40,6 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <mm_malloc.h>
 #endif
 
-#ifdef GRID_COMMS_SHMEM
-extern "C" { 
-#include <mpp/shmem.h>
-extern void * shmem_align(size_t, size_t);
-extern void  shmem_free(void *);
-}
-#endif
-
 namespace Grid {
 
 ////////////////////////////////////////////////////////////////////
@@ -65,28 +57,85 @@ public:
   typedef _Tp        value_type;
 
   template<typename _Tp1>  struct rebind { typedef alignedAllocator<_Tp1> other; };
-
   alignedAllocator() throw() { }
-
   alignedAllocator(const alignedAllocator&) throw() { }
-
   template<typename _Tp1> alignedAllocator(const alignedAllocator<_Tp1>&) throw() { }
-
   ~alignedAllocator() throw() { }
-
   pointer       address(reference __x)       const { return &__x; }
-  //  const_pointer address(const_reference __x) const { return &__x; }
-
   size_type  max_size() const throw() { return size_t(-1) / sizeof(_Tp); }
 
   pointer allocate(size_type __n, const void* _p= 0)
   { 
+#ifdef HAVE_MM_MALLOC_H
+    _Tp * ptr = (_Tp *) _mm_malloc(__n*sizeof(_Tp),128);
+#else
+    _Tp * ptr = (_Tp *) memalign(128,__n*sizeof(_Tp));
+#endif
+
+    _Tp tmp;
+#ifdef GRID_NUMA
+#pragma omp parallel for schedule(static)
+  for(int i=0;i<__n;i++){
+    ptr[i]=tmp;
+  }
+#endif 
+    return ptr;
+  }
+
+  void deallocate(pointer __p, size_type) { 
+#ifdef HAVE_MM_MALLOC_H
+    _mm_free((void *)__p); 
+#else
+    free((void *)__p);
+#endif
+  }
+  void construct(pointer __p, const _Tp& __val) { };
+  void construct(pointer __p) { };
+  void destroy(pointer __p) { };
+};
+template<typename _Tp>  inline bool operator==(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return true; }
+template<typename _Tp>  inline bool operator!=(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return false; }
+
+//////////////////////////////////////////////////////////////////////////////////////////
+// MPI3 : comms must use shm region
+// SHMEM: comms must use symmetric heap
+//////////////////////////////////////////////////////////////////////////////////////////
 #ifdef GRID_COMMS_SHMEM
-
+extern "C" { 
-    _Tp *ptr = (_Tp *) shmem_align(__n*sizeof(_Tp),64);
+#include <mpp/shmem.h>
-
+extern void * shmem_align(size_t, size_t);
-
+extern void  shmem_free(void *);
+}
 #define PARANOID_SYMMETRIC_HEAP
+#endif
+
+template<typename _Tp>
+class commAllocator {
+public: 
+  typedef std::size_t     size_type;
+  typedef std::ptrdiff_t  difference_type;
+  typedef _Tp*       pointer;
+  typedef const _Tp* const_pointer;
+  typedef _Tp&       reference;
+  typedef const _Tp& const_reference;
+  typedef _Tp        value_type;
+
+  template<typename _Tp1>  struct rebind { typedef commAllocator<_Tp1> other; };
+  commAllocator() throw() { }
+  commAllocator(const commAllocator&) throw() { }
+  template<typename _Tp1> commAllocator(const commAllocator<_Tp1>&) throw() { }
+  ~commAllocator() throw() { }
+  pointer       address(reference __x)       const { return &__x; }
+  size_type  max_size() const throw() { return size_t(-1) / sizeof(_Tp); }
+
+#ifdef GRID_COMMS_SHMEM
+  pointer allocate(size_type __n, const void* _p= 0)
+  {
+#ifdef CRAY
+    _Tp *ptr = (_Tp *) shmem_align(__n*sizeof(_Tp),64);
+#else
+    _Tp *ptr = (_Tp *) shmem_align(64,__n*sizeof(_Tp));
+#endif
 #ifdef PARANOID_SYMMETRIC_HEAP
     static void * bcast;
     static long  psync[_SHMEM_REDUCE_SYNC_SIZE];
@@ -96,55 +145,47 @@ public:
 
     if ( bcast != ptr ) {
       std::printf("inconsistent alloc pe %d %lx %lx \n",shmem_my_pe(),bcast,ptr);std::fflush(stdout);
-      BACKTRACEFILE();
+      //      BACKTRACEFILE();
       exit(0);
     }
-
     assert( bcast == (void *) ptr);
-
 #endif 
+    return ptr;
+  }
+  void deallocate(pointer __p, size_type) { 
+    shmem_free((void *)__p);
+  }
 #else
-
+  pointer allocate(size_type __n, const void* _p= 0) 
+  {
 #ifdef HAVE_MM_MALLOC_H
     _Tp * ptr = (_Tp *) _mm_malloc(__n*sizeof(_Tp),128);
 #else
     _Tp * ptr = (_Tp *) memalign(128,__n*sizeof(_Tp));
 #endif
-
-#endif
-    _Tp tmp;
-#undef FIRST_TOUCH_OPTIMISE
-#ifdef FIRST_TOUCH_OPTIMISE
-#pragma omp parallel for 
-  for(int i=0;i<__n;i++){
-    ptr[i]=tmp;
-  }
-#endif 
     return ptr;
   }
-
   void deallocate(pointer __p, size_type) { 
-#ifdef GRID_COMMS_SHMEM
-    shmem_free((void *)__p);
-#else
 #ifdef HAVE_MM_MALLOC_H
     _mm_free((void *)__p); 
 #else
     free((void *)__p);
-#endif
 #endif
   }
+#endif
   void construct(pointer __p, const _Tp& __val) { };
   void construct(pointer __p) { };
-
   void destroy(pointer __p) { };
 };
+template<typename _Tp>  inline bool operator==(const commAllocator<_Tp>&, const commAllocator<_Tp>&){ return true; }
+template<typename _Tp>  inline bool operator!=(const commAllocator<_Tp>&, const commAllocator<_Tp>&){ return false; }
 
-template<typename _Tp>  inline bool
+////////////////////////////////////////////////////////////////////////////////
-operator==(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return true; }
+// Template typedefs
-
+////////////////////////////////////////////////////////////////////////////////
-template<typename _Tp>  inline bool
+template<class T> using Vector     = std::vector<T,alignedAllocator<T> >;           
-operator!=(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return false; }
+template<class T> using commVector = std::vector<T,commAllocator<T> >;              
+template<class T> using Matrix     = std::vector<std::vector<T,alignedAllocator<T> > >;
     
 }; // namespace Grid
 #endif

lib/Cartesian.h

@@ -28,8 +28,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_CARTESIAN_H
 #define GRID_CARTESIAN_H
 
-#include <cartesian/Cartesian_base.h>
+#include <Grid/cartesian/Cartesian_base.h>
-#include <cartesian/Cartesian_full.h>
+#include <Grid/cartesian/Cartesian_full.h>
-#include <cartesian/Cartesian_red_black.h> 
+#include <Grid/cartesian/Cartesian_red_black.h> 
 
 #endif

lib/Communicator.h

@@ -28,6 +28,6 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_COMMUNICATOR_H
 #define GRID_COMMUNICATOR_H
 
-#include <communicator/Communicator_base.h>
+#include <Grid/communicator/Communicator_base.h>
 
 #endif

lib/Cshift.h

@@ -28,17 +28,21 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef _GRID_CSHIFT_H_
 #define _GRID_CSHIFT_H_
 
-#include <cshift/Cshift_common.h>
+#include <Grid/cshift/Cshift_common.h>
 
 #ifdef GRID_COMMS_NONE
-#include <cshift/Cshift_none.h>
+#include <Grid/cshift/Cshift_none.h>
 #endif
 
 #ifdef GRID_COMMS_MPI
-#include <cshift/Cshift_mpi.h>
+#include <Grid/cshift/Cshift_mpi.h>
+#endif 
+
+#ifdef GRID_COMMS_MPI3
+#include <Grid/cshift/Cshift_mpi.h>
 #endif 
 
 #ifdef GRID_COMMS_SHMEM
-#include <cshift/Cshift_mpi.h> // uses same implementation of communicator
+#include <Grid/cshift/Cshift_mpi.h> // uses same implementation of communicator
 #endif 
 #endif

lib/FFT.h

@@ -0,0 +1,271 @@
+
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/Cshift.h
+
+    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 */
+#ifndef _GRID_FFT_H_
+#define _GRID_FFT_H_
+
+#ifdef HAVE_FFTW	
+#include <fftw3.h>
+#endif
+namespace Grid {
+
+  template<class scalar> struct FFTW { };
+
+#ifdef HAVE_FFTW	
+  template<> struct FFTW<ComplexD> {
+  public:
+
+    typedef fftw_complex FFTW_scalar;
+    typedef fftw_plan    FFTW_plan;
+
+    static FFTW_plan fftw_plan_many_dft(int rank, const int *n,int howmany,
+					FFTW_scalar *in, const int *inembed,		
+					int istride, int idist,		
+					FFTW_scalar *out, const int *onembed,		
+					int ostride, int odist,		
+					int sign, unsigned flags) {
+      return ::fftw_plan_many_dft(rank,n,howmany,in,inembed,istride,idist,out,onembed,ostride,odist,sign,flags);
+    }	  
+    
+    static void fftw_flops(const FFTW_plan p,double *add, double *mul, double *fmas){
+      ::fftw_flops(p,add,mul,fmas);
+    }
+
+    inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out) {
+      ::fftw_execute_dft(p,in,out);
+    }
+    inline static void fftw_destroy_plan(const FFTW_plan p) {
+      ::fftw_destroy_plan(p);
+    }
+  };
+
+  template<> struct FFTW<ComplexF> {
+  public:
+
+    typedef fftwf_complex FFTW_scalar;
+    typedef fftwf_plan    FFTW_plan;
+
+    static FFTW_plan fftw_plan_many_dft(int rank, const int *n,int howmany,
+					FFTW_scalar *in, const int *inembed,		
+					int istride, int idist,		
+					FFTW_scalar *out, const int *onembed,		
+					int ostride, int odist,		
+					int sign, unsigned flags) {
+      return ::fftwf_plan_many_dft(rank,n,howmany,in,inembed,istride,idist,out,onembed,ostride,odist,sign,flags);
+    }	  
+    
+    static void fftw_flops(const FFTW_plan p,double *add, double *mul, double *fmas){
+      ::fftwf_flops(p,add,mul,fmas);
+    }
+
+    inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out) {
+      ::fftwf_execute_dft(p,in,out);
+    }
+    inline static void fftw_destroy_plan(const FFTW_plan p) {
+      ::fftwf_destroy_plan(p);
+    }
+  };
+
+#endif
+
+#ifndef FFTW_FORWARD
+#define FFTW_FORWARD (-1)
+#define FFTW_BACKWARD (+1)
+#endif
+
+  class FFT { 
+  private:
+
+    GridCartesian *vgrid;
+    GridCartesian *sgrid;
+
+    int Nd;
+    double flops;
+    double flops_call;
+    uint64_t usec;
+
+    std::vector<int> dimensions;
+    std::vector<int> processors;
+    std::vector<int> processor_coor;
+
+  public:
+
+    static const int forward=FFTW_FORWARD;
+    static const int backward=FFTW_BACKWARD;
+
+    double Flops(void) {return flops;}
+    double MFlops(void) {return flops/usec;}
+
+    FFT ( GridCartesian * grid ) : 
+      vgrid(grid),
+      Nd(grid->_ndimension),
+      dimensions(grid->_fdimensions),
+      processors(grid->_processors),
+      processor_coor(grid->_processor_coor)
+    {
+      flops=0;
+      usec =0;
+      std::vector<int> layout(Nd,1);
+      sgrid = new GridCartesian(dimensions,layout,processors);
+    };
+
+    ~FFT ( void)  { 
+      delete sgrid; 
+    }
+    
+    template<class vobj>
+    void FFT_dim(Lattice<vobj> &result,const Lattice<vobj> &source,int dim, int inverse){
+
+      conformable(result._grid,vgrid);
+      conformable(source._grid,vgrid);
+
+      int L = vgrid->_ldimensions[dim];
+      int G = vgrid->_fdimensions[dim];
+
+      std::vector<int> layout(Nd,1);
+      std::vector<int> pencil_gd(vgrid->_fdimensions);
+
+      pencil_gd[dim] = G*processors[dim];    
+
+      // Pencil global vol LxLxGxLxL per node
+      GridCartesian pencil_g(pencil_gd,layout,processors);
+
+      // Construct pencils
+      typedef typename vobj::scalar_object sobj;
+      typedef typename sobj::scalar_type   scalar;
+
+      Lattice<vobj> ssource(vgrid); ssource =source;
+      Lattice<sobj> pgsource(&pencil_g);
+      Lattice<sobj> pgresult(&pencil_g); pgresult=zero;
+
+#ifndef HAVE_FFTW	
+      assert(0);
+#else 
+      typedef typename FFTW<scalar>::FFTW_scalar FFTW_scalar;
+      typedef typename FFTW<scalar>::FFTW_plan   FFTW_plan;
+
+      {
+	int Ncomp = sizeof(sobj)/sizeof(scalar);
+	int Nlow  = 1;
+	for(int d=0;d<dim;d++){
+	  Nlow*=vgrid->_ldimensions[d];
+	}
+
+	int rank = 1;  /* 1d transforms */
+	int n[] = {G}; /* 1d transforms of length G */
+	int howmany = Ncomp;
+	int odist,idist,istride,ostride;
+	idist   = odist   = 1;          /* Distance between consecutive FT's */
+	istride = ostride = Ncomp*Nlow; /* distance between two elements in the same FT */
+	int *inembed = n, *onembed = n;
+
+	
+	int sign = FFTW_FORWARD;
+	if (inverse) sign = FFTW_BACKWARD;
+
+	FFTW_plan p;
+	{
+	  FFTW_scalar *in = (FFTW_scalar *)&pgsource._odata[0];
+	  FFTW_scalar *out= (FFTW_scalar *)&pgresult._odata[0];
+	  p = FFTW<scalar>::fftw_plan_many_dft(rank,n,howmany,
+					       in,inembed,
+					       istride,idist,
+					       out,onembed,
+					       ostride, odist,
+					       sign,FFTW_ESTIMATE);
+	}
+
+    std::vector<int> lcoor(Nd), gcoor(Nd);
+
+	// Barrel shift and collect global pencil
+	for(int p=0;p<processors[dim];p++) { 
+
+	  for(int idx=0;idx<sgrid->lSites();idx++) { 
+
+	    
+    	    sgrid->LocalIndexToLocalCoor(idx,lcoor);
+
+	    sobj s;
+
+	    peekLocalSite(s,ssource,lcoor);
+
+	    lcoor[dim]+=p*L;
+	   
+	    pokeLocalSite(s,pgsource,lcoor);
+	  }
+
+	  ssource = Cshift(ssource,dim,L);
+	}
+	
+	// Loop over orthog coords
+	int NN=pencil_g.lSites();
+	GridStopWatch timer;
+	timer.Start();
+
+PARALLEL_FOR_LOOP
+	for(int idx=0;idx<NN;idx++) {
+	  pencil_g.LocalIndexToLocalCoor(idx,lcoor);
+
+	  if ( lcoor[dim] == 0 ) {  // restricts loop to plane at lcoor[dim]==0
+	    FFTW_scalar *in = (FFTW_scalar *)&pgsource._odata[idx];
+	    FFTW_scalar *out= (FFTW_scalar *)&pgresult._odata[idx];
+	    FFTW<scalar>::fftw_execute_dft(p,in,out);
+	  }
+	}
+
+        timer.Stop();
+
+          double add,mul,fma;
+          FFTW<scalar>::fftw_flops(p,&add,&mul,&fma);
+          flops_call = add+mul+2.0*fma;
+          usec += timer.useconds();
+          flops+= flops_call*NN;
+        int pc = processor_coor[dim];
+        for(int idx=0;idx<sgrid->lSites();idx++) {
+	  sgrid->LocalIndexToLocalCoor(idx,lcoor);
+	  gcoor = lcoor;
+	  // extract the result
+	  sobj s;
+	  gcoor[dim] = lcoor[dim]+L*pc;
+	  peekLocalSite(s,pgresult,gcoor);
+	  pokeLocalSite(s,result,lcoor);
+	}
+      	  
+	FFTW<scalar>::fftw_destroy_plan(p);
+      }
+#endif
+
+
+    }
+
+  };
+
+
+}
+
+#endif

lib/Grid.h

@@ -59,29 +59,31 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 ///////////////////
 // Grid headers
 ///////////////////
-#include <serialisation/Serialisation.h>
+#include <Grid/serialisation/Serialisation.h>
-#include <Config.h>
+#include "Config.h"
-#include <Timer.h>
+#include <Grid/Timer.h>
-#include <PerfCount.h>
+#include <Grid/PerfCount.h>
-#include <Log.h>
+#include <Grid/Log.h>
-#include <AlignedAllocator.h>
+#include <Grid/AlignedAllocator.h>
-#include <Simd.h>
+#include <Grid/Simd.h>
-#include <Threads.h>
+#include <Grid/Threads.h>
-#include <Lexicographic.h>
+#include <Grid/Lexicographic.h>
-#include <Communicator.h> 
+#include <Grid/Init.h>
-#include <Cartesian.h>    
+#include <Grid/Communicator.h> 
-#include <Tensors.h>      
+#include <Grid/Cartesian.h>    
-#include <Lattice.h>      
+#include <Grid/Tensors.h>      
-#include <Cshift.h>       
+#include <Grid/Lattice.h>      
-#include <Stencil.h>      
+#include <Grid/Cshift.h>       
-#include <Algorithms.h>   
+#include <Grid/Stencil.h>      
-#include <parallelIO/BinaryIO.h>
+#include <Grid/Algorithms.h>   
-#include <qcd/QCD.h>
+#include <Grid/parallelIO/BinaryIO.h>
-#include <parallelIO/NerscIO.h>
+#include <Grid/qcd/QCD.h>
-#include <Init.h>
+#include <Grid/parallelIO/NerscIO.h>
 
-#include <qcd/hmc/NerscCheckpointer.h>
+#include <Grid/FFT.h>
-#include <qcd/hmc/HmcRunner.h>
+
+#include <Grid/qcd/hmc/NerscCheckpointer.h>
+#include <Grid/qcd/hmc/HmcRunner.h>
 
 
 

lib/Init.cc

@@ -153,6 +153,7 @@ void GridParseLayout(char **argv,int argc,
     assert(ompthreads.size()==1);
     GridThread::SetThreads(ompthreads[0]);
   }
+
   if( GridCmdOptionExists(argv,argv+argc,"--cores") ){
     std::vector<int> cores(0);
     arg= GridCmdOptionPayload(argv,argv+argc,"--cores");
@@ -170,14 +171,17 @@ std::string GridCmdVectorIntToString(const std::vector<int> & vec){
 /////////////////////////////////////////////////////////
 //
 /////////////////////////////////////////////////////////
+static int Grid_is_initialised = 0;
+
+
 void Grid_init(int *argc,char ***argv)
 {
+  GridLogger::StopWatch.Start();
+
   CartesianCommunicator::Init(argc,argv);
 
   // Parse command line args.
 
-  GridLogger::StopWatch.Start();
-
   std::string arg;
   std::vector<std::string> logstreams;
   std::string defaultLog("Error,Warning,Message,Performance");
@@ -193,7 +197,7 @@ void Grid_init(int *argc,char ***argv)
     std::cout<<GridLogMessage<<"--mpi n.n.n.n   : default MPI decomposition"<<std::endl;    
     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<<"--log list      : comma separted list of streams from Error,Warning,Message,Performance,Iterative,Integrator,Debug"<<std::endl;
+    std::cout<<GridLogMessage<<"--log list      : comma separted list of streams from Error,Warning,Message,Performance,Iterative,Integrator,Debug,Colours"<<std::endl;
     exit(EXIT_SUCCESS);
   }
 
@@ -203,7 +207,6 @@ void Grid_init(int *argc,char ***argv)
     GridLogConfigure(logstreams);
   }
 
-
   if( GridCmdOptionExists(*argv,*argv+*argc,"--debug-signals") ){
     Grid_debug_handler_init();
   }
@@ -211,17 +214,19 @@ void Grid_init(int *argc,char ***argv)
     Grid_quiesce_nodes();
   }
   if( GridCmdOptionExists(*argv,*argv+*argc,"--dslash-opt") ){
-    QCD::WilsonFermionStatic::HandOptDslash=1;
+    QCD::WilsonKernelsStatic::HandOpt=1;
-    QCD::WilsonFermion5DStatic::HandOptDslash=1;
   }
   if( GridCmdOptionExists(*argv,*argv+*argc,"--lebesgue") ){
     LebesgueOrder::UseLebesgueOrder=1;
   }
-
   if( GridCmdOptionExists(*argv,*argv+*argc,"--cacheblocking") ){
     arg= GridCmdOptionPayload(*argv,*argv+*argc,"--cacheblocking");
     GridCmdOptionIntVector(arg,LebesgueOrder::Block);
   }
+  if( GridCmdOptionExists(*argv,*argv+*argc,"--timestamp") ){
+    GridLogTimestamp(1);
+  }
+
   GridParseLayout(*argv,*argc,
 		  Grid_default_latt,
 		  Grid_default_mpi);
@@ -235,26 +240,34 @@ void Grid_init(int *argc,char ***argv)
     std::cout<<GridLogMessage<<"\tvComplexD      : "<<sizeof(vComplexD)*8 <<"bits ; " <<GridCmdVectorIntToString(GridDefaultSimd(4,vComplexD::Nsimd()))<<std::endl;
   }
 
+  std::string COL_RED    = GridLogColours.colour["RED"];
+  std::string COL_PURPLE = GridLogColours.colour["PURPLE"];
+  std::string COL_BLACK  = GridLogColours.colour["BLACK"];
+  std::string COL_GREEN  = GridLogColours.colour["GREEN"];
+  std::string COL_BLUE   = GridLogColours.colour["BLUE"];
+  std::string COL_YELLOW = GridLogColours.colour["YELLOW"];
+  std::string COL_BACKGROUND = GridLogColours.colour["NORMAL"];
+
+  
   std::cout <<std::endl;
-  std::cout <<Logger::RED  << "__|__|__|__|__"<<             "|__|__|_"<<Logger::PURPLE<<"_|__|__|"<<                "__|__|__|__|__"<<std::endl; 
+  std::cout <<COL_RED  << "__|__|__|__|__"<<             "|__|__|_"<<COL_PURPLE<<"_|__|__|"<<                "__|__|__|__|__"<<std::endl; 
-  std::cout <<Logger::RED  << "__|__|__|__|__"<<             "|__|__|_"<<Logger::PURPLE<<"_|__|__|"<<                "__|__|__|__|__"<<std::endl; 
+  std::cout <<COL_RED  << "__|__|__|__|__"<<             "|__|__|_"<<COL_PURPLE<<"_|__|__|"<<                "__|__|__|__|__"<<std::endl; 
-  std::cout <<Logger::RED  << "__|__|  |  |  "<<             "|  |  | "<<Logger::PURPLE<<" |  |  |"<<                "  |  |  | _|__"<<std::endl; 
+  std::cout <<COL_RED  << "__|_ |  |  |  "<<             "|  |  | "<<COL_PURPLE<<" |  |  |"<<                "  |  |  | _|__"<<std::endl; 
-  std::cout <<Logger::RED  << "__|__         "<<             "        "<<Logger::PURPLE<<"        "<<                "          _|__"<<std::endl; 
+  std::cout <<COL_RED  << "__|_          "<<             "        "<<COL_PURPLE<<"        "<<                "          _|__"<<std::endl; 
-  std::cout <<Logger::RED  << "__|_  "<<Logger::GREEN<<" GGGG   "<<Logger::RED<<" RRRR   "<<Logger::BLUE  <<" III    "<<Logger::PURPLE<<"DDDD  "<<Logger::PURPLE<<"    _|__"<<std::endl;
+  std::cout <<COL_RED  << "__|_  "<<COL_GREEN<<" GGGG   "<<COL_RED<<" RRRR   "<<COL_BLUE  <<" III    "<<COL_PURPLE<<"DDDD  "<<COL_PURPLE<<"    _|__"<<std::endl;
-  std::cout <<Logger::RED  << "__|_  "<<Logger::GREEN<<"G       "<<Logger::RED<<" R   R  "<<Logger::BLUE  <<"  I     "<<Logger::PURPLE<<"D   D "<<Logger::PURPLE<<"    _|__"<<std::endl;
+  std::cout <<COL_RED  << "__|_  "<<COL_GREEN<<"G       "<<COL_RED<<" R   R  "<<COL_BLUE  <<"  I     "<<COL_PURPLE<<"D   D "<<COL_PURPLE<<"    _|__"<<std::endl;
-  std::cout <<Logger::RED  << "__|_  "<<Logger::GREEN<<"G       "<<Logger::RED<<" R   R  "<<Logger::BLUE  <<"  I     "<<Logger::PURPLE<<"D    D"<<Logger::PURPLE<<"    _|__"<<std::endl;
+  std::cout <<COL_RED  << "__|_  "<<COL_GREEN<<"G       "<<COL_RED<<" R   R  "<<COL_BLUE  <<"  I     "<<COL_PURPLE<<"D    D"<<COL_PURPLE<<"    _|__"<<std::endl;
-  std::cout <<Logger::BLUE << "__|_  "<<Logger::GREEN<<"G  GG   "<<Logger::RED<<" RRRR   "<<Logger::BLUE  <<"  I     "<<Logger::PURPLE<<"D    D"<<Logger::GREEN <<"    _|__"<<std::endl;
+  std::cout <<COL_BLUE << "__|_  "<<COL_GREEN<<"G  GG   "<<COL_RED<<" RRRR   "<<COL_BLUE  <<"  I     "<<COL_PURPLE<<"D    D"<<COL_GREEN <<"    _|__"<<std::endl;
-  std::cout <<Logger::BLUE << "__|_  "<<Logger::GREEN<<"G   G   "<<Logger::RED<<" R  R   "<<Logger::BLUE  <<"  I     "<<Logger::PURPLE<<"D   D "<<Logger::GREEN <<"    _|__"<<std::endl;
+  std::cout <<COL_BLUE << "__|_  "<<COL_GREEN<<"G   G   "<<COL_RED<<" R  R   "<<COL_BLUE  <<"  I     "<<COL_PURPLE<<"D   D "<<COL_GREEN <<"    _|__"<<std::endl;
-  std::cout <<Logger::BLUE << "__|_  "<<Logger::GREEN<<" GGGG   "<<Logger::RED<<" R   R  "<<Logger::BLUE  <<" III    "<<Logger::PURPLE<<"DDDD  "<<Logger::GREEN <<"    _|__"<<std::endl;
+  std::cout <<COL_BLUE << "__|_  "<<COL_GREEN<<" GGGG   "<<COL_RED<<" R   R  "<<COL_BLUE  <<" III    "<<COL_PURPLE<<"DDDD  "<<COL_GREEN <<"    _|__"<<std::endl;
-  std::cout <<Logger::BLUE << "__|__         "<<             "        "<<Logger::GREEN <<"        "<<                "          _|__"<<std::endl; 
+  std::cout <<COL_BLUE << "__|_          "<<             "        "<<COL_GREEN <<"        "<<                "          _|__"<<std::endl; 
-  std::cout <<Logger::BLUE << "__|__|__|__|__"<<             "|__|__|_"<<Logger::GREEN <<"_|__|__|"<<                "__|__|__|__|__"<<std::endl; 
+  std::cout <<COL_BLUE << "__|__|__|__|__"<<             "|__|__|_"<<COL_GREEN <<"_|__|__|"<<                "__|__|__|__|__"<<std::endl; 
-  std::cout <<Logger::BLUE << "__|__|__|__|__"<<             "|__|__|_"<<Logger::GREEN <<"_|__|__|"<<                "__|__|__|__|__"<<std::endl; 
+  std::cout <<COL_BLUE << "__|__|__|__|__"<<             "|__|__|_"<<COL_GREEN <<"_|__|__|"<<                "__|__|__|__|__"<<std::endl; 
-  std::cout <<Logger::BLUE << "  |  |  |  |  "<<             "|  |  | "<<Logger::GREEN <<" |  |  |"<<                "  |  |  |  |  "<<std::endl; 
+  std::cout <<COL_BLUE << "  |  |  |  |  "<<             "|  |  | "<<COL_GREEN <<" |  |  |"<<                "  |  |  |  |  "<<std::endl; 
   std::cout << std::endl;
   std::cout << std::endl;
-  std::cout <<Logger::YELLOW<< std::endl;
+  std::cout <<COL_YELLOW<< std::endl;
   std::cout << "Copyright (C) 2015 Peter Boyle, Azusa Yamaguchi, Guido Cossu, Antonin Portelli and other authors"<<std::endl;
-  std::cout << "Colours by Tadahito Boyle "<<std::endl;
   std::cout << std::endl;
   std::cout << "This program is free software; you can redistribute it and/or modify"<<std::endl;
   std::cout << "it under the terms of the GNU General Public License as published by"<<std::endl;
@@ -265,22 +278,20 @@ void Grid_init(int *argc,char ***argv)
   std::cout << "but WITHOUT ANY WARRANTY; without even the implied warranty of"<<std::endl;
   std::cout << "MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the"<<std::endl;
   std::cout << "GNU General Public License for more details."<<std::endl;
-  std::cout << Logger::BLACK <<std::endl;
+  std::cout << COL_BACKGROUND <<std::endl;
+  std::cout << std::endl;
+
+  Grid_is_initialised = 1;
 }
 
   
 void Grid_finalize(void)
 {
-#ifdef GRID_COMMS_MPI
+#if defined (GRID_COMMS_MPI) || defined (GRID_COMMS_MPI3)
   MPI_Finalize();
   Grid_unquiesce_nodes();
 #endif
 }
-double usecond(void) {
-  struct timeval tv;
-  gettimeofday(&tv,NULL);
-  return 1.0*tv.tv_usec + 1.0e6*tv.tv_sec;
-}
 
 void * Grid_backtrace_buffer[_NBACKTRACE];
 

lib/Init.h

@@ -33,6 +33,7 @@ namespace Grid {
 
   void Grid_init(int *argc,char ***argv);
   void Grid_finalize(void);
+
   // internal, controled with --handle
   void Grid_sa_signal_handler(int sig,siginfo_t *si,void * ptr);
   void Grid_debug_handler_init(void);
@@ -44,6 +45,7 @@ namespace Grid {
   const std::vector<int> &GridDefaultMpi(void);
   const int              &GridThreads(void)  ;
   void                    GridSetThreads(int t) ;
+  void GridLogTimestamp(int);
 
   // Common parsing chores
   std::string GridCmdOptionPayload(char ** begin, char ** end, const std::string & option);

lib/Lattice.h

@@ -28,6 +28,6 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_LATTICE_H
 #define GRID_LATTICE_H
 
-#include <lattice/Lattice_base.h>
+#include <Grid/lattice/Lattice_base.h>
 
 #endif

lib/Log.cc

@@ -25,7 +25,8 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 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
+See the full license in the file "LICENSE" in the top level distribution
+directory
 *************************************************************************************/
 /*  END LEGAL */
 #include <Grid.h>
@@ -33,77 +34,51 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 namespace Grid {
 
 GridStopWatch Logger::StopWatch;
+int Logger::timestamp;
 std::ostream Logger::devnull(0);
-std::string Logger::BLACK("\033[30m");
-std::string Logger::RED("\033[31m");
-std::string Logger::GREEN("\033[32m");
-std::string Logger::YELLOW("\033[33m");
-std::string Logger::BLUE("\033[34m");
-std::string Logger::PURPLE("\033[35m");
-std::string Logger::CYAN("\033[36m");
-std::string Logger::WHITE("\033[37m");
-std::string Logger::NORMAL("\033[0;39m");
-std::string EMPTY("");
 
-#if 0  
+void GridLogTimestamp(int on){
-  GridLogger GridLogError      (1,"Error",Logger::RED);
+  Logger::Timestamp(on);
-  GridLogger GridLogWarning    (1,"Warning",Logger::YELLOW);
+}
-  GridLogger GridLogMessage    (1,"Message",Logger::BLACK);
-  GridLogger GridLogDebug      (1,"Debug",Logger::PURPLE);
-  GridLogger GridLogPerformance(1,"Performance",Logger::GREEN);
-  GridLogger GridLogIterative  (1,"Iterative",Logger::BLUE);
-  GridLogger GridLogIntegrator (1,"Integrator",Logger::BLUE);
-#else
-  GridLogger GridLogError      (1,"Error",EMPTY);
-  GridLogger GridLogWarning    (1,"Warning",EMPTY);
-  GridLogger GridLogMessage    (1,"Message",EMPTY);
-  GridLogger GridLogDebug      (1,"Debug",EMPTY);
-  GridLogger GridLogPerformance(1,"Performance",EMPTY);
-  GridLogger GridLogIterative  (1,"Iterative",EMPTY);
-  GridLogger GridLogIntegrator (1,"Integrator",EMPTY);
-#endif
 
-void GridLogConfigure(std::vector<std::string> &logstreams)
+Colours GridLogColours(0);
-{
+GridLogger GridLogError(1, "Error", GridLogColours, "RED");
+GridLogger GridLogWarning(1, "Warning", GridLogColours, "YELLOW");
+GridLogger GridLogMessage(1, "Message", GridLogColours, "NORMAL");
+GridLogger GridLogDebug(1, "Debug", GridLogColours, "PURPLE");
+GridLogger GridLogPerformance(1, "Performance", GridLogColours, "GREEN");
+GridLogger GridLogIterative(1, "Iterative", GridLogColours, "BLUE");
+GridLogger GridLogIntegrator(1, "Integrator", GridLogColours, "BLUE");
+
+void GridLogConfigure(std::vector<std::string> &logstreams) {
   GridLogError.Active(0);
   GridLogWarning.Active(0);
-  GridLogMessage.Active(0);
+  GridLogMessage.Active(1); // at least the messages should be always on
   GridLogIterative.Active(0);
   GridLogDebug.Active(0);
   GridLogPerformance.Active(0);
   GridLogIntegrator.Active(0);
-
+  GridLogColours.Active(0);
-  int blackAndWhite = 1;
-  if(blackAndWhite){
-    Logger::BLACK = std::string("");
-    Logger::RED    =Logger::BLACK;
-    Logger::GREEN  =Logger::BLACK;
-    Logger::YELLOW =Logger::BLACK;
-    Logger::BLUE   =Logger::BLACK;
-    Logger::PURPLE =Logger::BLACK;
-    Logger::CYAN   =Logger::BLACK;
-    Logger::WHITE  =Logger::BLACK;
-    Logger::NORMAL =Logger::BLACK;
-  }
 
   for (int i = 0; i < logstreams.size(); i++) {
     if (logstreams[i] == std::string("Error")) GridLogError.Active(1);
     if (logstreams[i] == std::string("Warning")) GridLogWarning.Active(1);
-    if ( logstreams[i]== std::string("Message")     ) GridLogMessage.Active(1);
+    if (logstreams[i] == std::string("NoMessage")) GridLogMessage.Active(0);
     if (logstreams[i] == std::string("Iterative")) GridLogIterative.Active(1);
     if (logstreams[i] == std::string("Debug")) GridLogDebug.Active(1);
-    if ( logstreams[i]== std::string("Performance") ) GridLogPerformance.Active(1);
+    if (logstreams[i] == std::string("Performance"))
+      GridLogPerformance.Active(1);
     if (logstreams[i] == std::string("Integrator")) GridLogIntegrator.Active(1);
+    if (logstreams[i] == std::string("Colours")) GridLogColours.Active(1);
   }
 }
 
 ////////////////////////////////////////////////////////////
 // Verbose limiter on MPI tasks
 ////////////////////////////////////////////////////////////
-void Grid_quiesce_nodes(void)
+void Grid_quiesce_nodes(void) {
-{
   int me = 0;
-#ifdef GRID_COMMS_MPI
+#if defined(GRID_COMMS_MPI) || defined(GRID_COMMS_MPI3)
   MPI_Comm_rank(MPI_COMM_WORLD, &me);
 #endif
 #ifdef GRID_COMMS_SHMEM
@@ -114,13 +89,9 @@ void Grid_quiesce_nodes(void)
   }
 }
 
-void Grid_unquiesce_nodes(void)
+void Grid_unquiesce_nodes(void) {
-{
 #ifdef GRID_COMMS_MPI
   std::cout.clear();
 #endif
 }
-
-
 }
-

lib/Log.h

@@ -27,6 +27,9 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+
+#include <map>
+
 #ifndef GRID_LOG_H
 #define GRID_LOG_H
 
@@ -36,43 +39,86 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 namespace Grid {
 
+//////////////////////////////////////////////////////////////////////////////////////////////////
 // Dress the output; use std::chrono for time stamping via the StopWatch class
-int Rank(void); // used for early stage debug before library init
+//////////////////////////////////////////////////////////////////////////////////////////////////
+
+
+class Colours{
+protected:
+  bool is_active;
+public:
+  std::map<std::string, std::string> colour;
+
+  Colours(bool activate=false){
+    Active(activate);
+  };
+
+  void Active(bool activate){
+    is_active=activate;
+    if (is_active){
+     colour["BLACK"]  ="\033[30m";
+     colour["RED"]    ="\033[31m";
+     colour["GREEN"]  ="\033[32m";
+     colour["YELLOW"] ="\033[33m";
+     colour["BLUE"]   ="\033[34m";
+     colour["PURPLE"] ="\033[35m";
+     colour["CYAN"]   ="\033[36m";
+     colour["WHITE"]  ="\033[37m";
+     colour["NORMAL"] ="\033[0;39m";
+    } else {
+      colour["BLACK"] ="";
+      colour["RED"]   ="";
+      colour["GREEN"] ="";
+      colour["YELLOW"]="";
+      colour["BLUE"]  ="";
+      colour["PURPLE"]="";
+      colour["CYAN"]  ="";
+      colour["WHITE"] ="";
+      colour["NORMAL"]="";
+    }
+  };
+};
 
 
 class Logger {
 protected:
+  Colours &Painter;
   int active;
-    std::string name, topName, COLOUR;
+  static int timestamp;
+  std::string name, topName;
+  std::string COLOUR;
+
 public:
   static GridStopWatch StopWatch;
   static std::ostream devnull;
 
-    static std::string BLACK;
+  std::string background() {return Painter.colour["NORMAL"];}
-    static std::string RED  ;
+  std::string evidence() {return Painter.colour["YELLOW"];}
-    static std::string GREEN;
+  std::string colour() {return Painter.colour[COLOUR];}
-    static std::string YELLOW;
-    static std::string BLUE  ;
-    static std::string PURPLE;
-    static std::string CYAN  ;
-    static std::string WHITE ;
-    static std::string NORMAL;
 
- Logger(std::string topNm, int on, std::string nm,std::string col)
+  Logger(std::string topNm, int on, std::string nm, Colours& col_class, std::string col)  : active(on),
-   : active(on), name(nm), topName(topNm), COLOUR(col) {};
+    name(nm),
+    topName(topNm),
+    Painter(col_class),
+    COLOUR(col) {} ;
   
   void Active(int on) {active = on;};
   int  isActive(void) {return active;};
+  static void Timestamp(int on) {timestamp = on;};
+  
+  friend std::ostream& operator<< (std::ostream& stream, Logger& log){
 
-    friend std::ostream& operator<< (std::ostream& stream, const Logger& log){
     if ( log.active ) {
+      stream << log.background()<< log.topName << log.background()<< " : ";
+      stream << log.colour() <<std::setw(14) << std::left << log.name << log.background() << " : ";
+      if ( log.timestamp ) {
 	StopWatch.Stop();
 	GridTime now = StopWatch.Elapsed();
 	StopWatch.Start();
-            stream << BLACK <<std::setw(8) << std::left << log.topName << BLACK<< " : ";
+	stream << log.evidence()<< now << log.background() << " : " ;
-            stream << log.COLOUR <<std::setw(11)  << log.name << BLACK << " : ";
+      }
-            stream << YELLOW <<std::setw(6) << now <<BLACK << " : " ;
+      stream << log.colour();
-            stream << log.COLOUR;
       return stream;
     } else { 
       return devnull;
@@ -83,7 +129,8 @@ public:
 
 class GridLogger: public Logger {
 public:
- GridLogger(int on, std::string nm, std::string col = Logger::BLACK): Logger("Grid", on, nm, col){};
+  GridLogger(int on, std::string nm, Colours&col_class, std::string col_key = "NORMAL"):
+  Logger("Grid", on, nm, col_class, col_key){};
 };
 
 void GridLogConfigure(std::vector<std::string> &logstreams);
@@ -95,6 +142,7 @@ extern GridLogger GridLogDebug  ;
 extern GridLogger GridLogPerformance;
 extern GridLogger GridLogIterative  ;
 extern GridLogger GridLogIntegrator  ;
+extern Colours    GridLogColours;
 
 
 #define _NBACKTRACE (256)
@@ -102,7 +150,7 @@ extern void * Grid_backtrace_buffer[_NBACKTRACE];
 
 #define BACKTRACEFILE() {\
 char string[20];					\
-    std::sprintf(string,"backtrace.%d",Rank());				\
+std::sprintf(string,"backtrace.%d",CartesianCommunicator::RankWorld()); \
 std::FILE * fp = std::fopen(string,"w");				\
 BACKTRACEFP(fp)\
 std::fclose(fp);	    \
@@ -128,5 +176,6 @@ extern void * Grid_backtrace_buffer[_NBACKTRACE];
 
 #define BACKTRACE() BACKTRACEFP(stdout) 
 
+
 }
 #endif

lib/Makefile.am

@@ -1,32 +1,32 @@
-# additional include paths necessary to compile the C++ library
-AM_CXXFLAGS = -I$(top_srcdir)/
-
 extra_sources=
 if BUILD_COMMS_MPI
   extra_sources+=communicator/Communicator_mpi.cc
+  extra_sources+=communicator/Communicator_base.cc
+endif
+
+if BUILD_COMMS_MPI3
+  extra_sources+=communicator/Communicator_mpi3.cc
+  extra_sources+=communicator/Communicator_base.cc
 endif
 
 if BUILD_COMMS_SHMEM
   extra_sources+=communicator/Communicator_shmem.cc
+  extra_sources+=communicator/Communicator_base.cc
 endif
 
 if BUILD_COMMS_NONE
   extra_sources+=communicator/Communicator_none.cc
+  extra_sources+=communicator/Communicator_base.cc
 endif
 
 #
 # Libraries
 #
-
 include Make.inc
+include Eigen.inc
 
 lib_LIBRARIES = libGrid.a
+
 libGrid_a_SOURCES              = $(CCFILES) $(extra_sources)
-
+libGrid_adir                   = $(pkgincludedir)
-
+nobase_dist_pkginclude_HEADERS = $(HFILES) $(eigen_files) Config.h
-#	qcd/action/fermion/PartialFractionFermion5D.cc\	\
-#
-# Include files
-#
-nobase_include_HEADERS=$(HFILES)
-

lib/PerfCount.cc

@@ -32,28 +32,44 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 namespace Grid {
 
 #define CacheControl(L,O,R) ((PERF_COUNT_HW_CACHE_##L)|(PERF_COUNT_HW_CACHE_OP_##O<<8)| (PERF_COUNT_HW_CACHE_RESULT_##R<<16))
-
+#define RawConfig(A,B) (A<<8|B)
 const PerformanceCounter::PerformanceCounterConfig PerformanceCounter::PerformanceCounterConfigs [] = {
 #ifdef __linux__
-  { PERF_TYPE_HARDWARE, PERF_COUNT_HW_CPU_CYCLES          ,  "CPUCYCLES.........." },
+  { PERF_TYPE_HARDWARE, PERF_COUNT_HW_CACHE_REFERENCES    ,  "CACHE_REFERENCES..." , INSTRUCTIONS},
-  { PERF_TYPE_HARDWARE, PERF_COUNT_HW_INSTRUCTIONS        ,  "INSTRUCTIONS......." },
+  { PERF_TYPE_HARDWARE, PERF_COUNT_HW_CACHE_MISSES        ,  "CACHE_MISSES......." , CACHE_REFERENCES},
-  { PERF_TYPE_HARDWARE, PERF_COUNT_HW_CACHE_REFERENCES    ,  "CACHE_REFERENCES..." },
+  { PERF_TYPE_HARDWARE, PERF_COUNT_HW_CPU_CYCLES          ,  "CPUCYCLES.........." , INSTRUCTIONS},
-  { PERF_TYPE_HARDWARE, PERF_COUNT_HW_CACHE_MISSES        ,  "CACHE_MISSES......." },
+  { PERF_TYPE_HARDWARE, PERF_COUNT_HW_INSTRUCTIONS        ,  "INSTRUCTIONS......." , CPUCYCLES   },
-  { PERF_TYPE_HW_CACHE, CacheControl(L1D,READ,MISS)       ,  "L1D_READ_MISS......"},
+    // 4
-  { PERF_TYPE_HW_CACHE, CacheControl(L1D,READ,ACCESS)     ,  "L1D_READ_ACCESS...."},
+#ifdef AVX512
-  { PERF_TYPE_HW_CACHE, CacheControl(L1D,WRITE,MISS)      ,  "L1D_WRITE_MISS....."},
+    { PERF_TYPE_RAW, RawConfig(0x40,0x04), "ALL_LOADS..........", CPUCYCLES    },
-  { PERF_TYPE_HW_CACHE, CacheControl(L1D,WRITE,ACCESS)    ,  "L1D_WRITE_ACCESS..."},
+    { PERF_TYPE_RAW, RawConfig(0x01,0x04), "L1_MISS_LOADS......", L1D_READ_ACCESS  },
-  { PERF_TYPE_HW_CACHE, CacheControl(L1D,PREFETCH,MISS)   ,  "L1D_PREFETCH_MISS.."},
+    { PERF_TYPE_RAW, RawConfig(0x40,0x04), "ALL_LOADS..........", L1D_READ_ACCESS    },
-  { PERF_TYPE_HW_CACHE, CacheControl(L1D,PREFETCH,ACCESS) ,  "L1D_PREFETCH_ACCESS"},
+    { PERF_TYPE_RAW, RawConfig(0x02,0x04), "L2_HIT_LOADS.......", L1D_READ_ACCESS  },
-  { PERF_TYPE_HW_CACHE, CacheControl(LL,READ,MISS)        ,  "LL_READ_MISS......."},
+    { PERF_TYPE_RAW, RawConfig(0x04,0x04), "L2_MISS_LOADS......", L1D_READ_ACCESS  },
-  //  { PERF_TYPE_HW_CACHE, CacheControl(LL,READ,ACCESS)      ,  "LL_READ_ACCESS....."},
+    { PERF_TYPE_RAW, RawConfig(0x10,0x04), "UTLB_MISS_LOADS....", L1D_READ_ACCESS },
-  { PERF_TYPE_HW_CACHE, CacheControl(LL,WRITE,MISS)       ,  "LL_WRITE_MISS......"},
+    { PERF_TYPE_RAW, RawConfig(0x08,0x04), "DTLB_MISS_LOADS....", L1D_READ_ACCESS },
-  { PERF_TYPE_HW_CACHE, CacheControl(LL,WRITE,ACCESS)     ,  "LL_WRITE_ACCESS...."},
+    // 11
-  { PERF_TYPE_HW_CACHE, CacheControl(LL,PREFETCH,MISS)    ,  "LL_PREFETCH_MISS..."},
+#else
-  { PERF_TYPE_HW_CACHE, CacheControl(LL,PREFETCH,ACCESS)  ,  "LL_PREFETCH_ACCESS."},
+  { PERF_TYPE_HW_CACHE, CacheControl(L1D,READ,ACCESS)     ,  "L1D_READ_ACCESS....",INSTRUCTIONS},
-  { PERF_TYPE_HW_CACHE, CacheControl(L1I,READ,MISS)       ,  "L1I_READ_MISS......"},
+  { PERF_TYPE_HW_CACHE, CacheControl(L1D,READ,MISS)       ,  "L1D_READ_MISS......",L1D_READ_ACCESS},
-  { PERF_TYPE_HW_CACHE, CacheControl(L1I,READ,ACCESS)     ,  "L1I_READ_ACCESS...."}
+  { PERF_TYPE_HW_CACHE, CacheControl(L1D,WRITE,MISS)      ,  "L1D_WRITE_MISS.....",L1D_READ_ACCESS},
+  { PERF_TYPE_HW_CACHE, CacheControl(L1D,WRITE,ACCESS)    ,  "L1D_WRITE_ACCESS...",L1D_READ_ACCESS},
+  { PERF_TYPE_HW_CACHE, CacheControl(L1D,PREFETCH,MISS)   ,  "L1D_PREFETCH_MISS..",L1D_READ_ACCESS},
+  { PERF_TYPE_HW_CACHE, CacheControl(L1D,PREFETCH,ACCESS) ,  "L1D_PREFETCH_ACCESS",L1D_READ_ACCESS},
+  { PERF_TYPE_HW_CACHE, CacheControl(L1D,PREFETCH,ACCESS) ,  "L1D_PREFETCH_ACCESS",L1D_READ_ACCESS},
+    // 11
 #endif
+  { PERF_TYPE_HW_CACHE, CacheControl(LL,READ,MISS)        ,  "LL_READ_MISS.......",L1D_READ_ACCESS},
+  { PERF_TYPE_HW_CACHE, CacheControl(LL,READ,ACCESS)      ,  "LL_READ_ACCESS.....",L1D_READ_ACCESS},
+  { PERF_TYPE_HW_CACHE, CacheControl(LL,WRITE,MISS)       ,  "LL_WRITE_MISS......",L1D_READ_ACCESS},
+  { PERF_TYPE_HW_CACHE, CacheControl(LL,WRITE,ACCESS)     ,  "LL_WRITE_ACCESS....",L1D_READ_ACCESS},
+    //15
+  { PERF_TYPE_HW_CACHE, CacheControl(LL,PREFETCH,MISS)    ,  "LL_PREFETCH_MISS...",L1D_READ_ACCESS},
+  { PERF_TYPE_HW_CACHE, CacheControl(LL,PREFETCH,ACCESS)  ,  "LL_PREFETCH_ACCESS.",L1D_READ_ACCESS},
+  { PERF_TYPE_HW_CACHE, CacheControl(L1I,READ,MISS)       ,  "L1I_READ_MISS......",INSTRUCTIONS},
+  { PERF_TYPE_HW_CACHE, CacheControl(L1I,READ,ACCESS)     ,  "L1I_READ_ACCESS....",INSTRUCTIONS}
+    //19
   //  { PERF_TYPE_HARDWARE, PERF_COUNT_HW_STALLED_CYCLES_FRONTEND, "STALL_CYCLES" },
+#endif
 };
 }

lib/PerfCount.h

@@ -58,6 +58,27 @@ static long perf_event_open(struct perf_event_attr *hw_event, pid_t pid,
 }
 #endif
 
+#ifdef TIMERS_OFF
+
+
+inline uint64_t cyclecount(void){ 
+  return 0;
+}
+#define __SSC_MARK(mark) __asm__ __volatile__ ("movl %0, %%ebx; .byte 0x64, 0x67, 0x90 " ::"i"(mark):"%ebx")
+#define __SSC_STOP  __SSC_MARK(0x110)
+#define __SSC_START __SSC_MARK(0x111)
+
+
+#else
+
+#define __SSC_MARK(mark) 
+#define __SSC_STOP  
+#define __SSC_START 
+
+/*
+ * cycle counters arch dependent
+ */
+
 #ifdef __bgq__
 inline uint64_t cyclecount(void){ 
    uint64_t tmp;
@@ -65,18 +86,20 @@ inline uint64_t cyclecount(void){
    return tmp;
 }
 #elif defined __x86_64__
-#include <immintrin.h>
-#ifndef __INTEL_COMPILER
 #include <x86intrin.h>
-#endif
 inline uint64_t cyclecount(void){ 
   return __rdtsc();
+  //  unsigned int dummy;
+  // return __rdtscp(&dummy);
 }
 #else
-#warning No cycle counter implemented for this architecture
+
 inline uint64_t cyclecount(void){ 
    return 0;
 }
+
+#endif
+
 #endif
 
 class PerformanceCounter {
@@ -87,6 +110,7 @@ private:
     uint32_t type;
     uint64_t config;
     const char *name;
+    int normalisation;
   } PerformanceCounterConfig; 
   
   static const PerformanceCounterConfig PerformanceCounterConfigs [];
@@ -94,26 +118,12 @@ private:
 public:
 
   enum PerformanceCounterType {
-    CPUCYCLES=0,
+    CACHE_REFERENCES=0,
-    INSTRUCTIONS,
+    CACHE_MISSES=1,
-    //    STALL_CYCLES,
+    CPUCYCLES=2,
-    CACHE_REFERENCES,
+    INSTRUCTIONS=3,
-    CACHE_MISSES,
+    L1D_READ_ACCESS=4,
-    L1D_READ_MISS,
+    PERFORMANCE_COUNTER_NUM_TYPES=19
-    L1D_READ_ACCESS,
-    L1D_WRITE_MISS,
-    L1D_WRITE_ACCESS,
-    L1D_PREFETCH_MISS,
-    L1D_PREFETCH_ACCESS,
-    LL_READ_MISS,
-    //    LL_READ_ACCESS,
-    LL_WRITE_MISS,
-    LL_WRITE_ACCESS,
-    LL_PREFETCH_MISS,
-    LL_PREFETCH_ACCESS,
-    L1I_READ_MISS,
-    L1I_READ_ACCESS,
-    PERFORMANCE_COUNTER_NUM_TYPES
   };
 
 public:
@@ -121,7 +131,9 @@ public:
   int PCT;
 
   long long count;
+  long long cycles;
   int fd;
+  int cyclefd;
   unsigned long long elapsed;
   uint64_t begin;
 
@@ -134,7 +146,9 @@ public:
     assert(_pct>=0);
     assert(_pct<PERFORMANCE_COUNTER_NUM_TYPES);
     fd=-1;
+    cyclefd=-1;
     count=0;
+    cycles=0;
     PCT =_pct;
     Open();
 #endif
@@ -159,6 +173,15 @@ public:
       fprintf(stderr, "Error opening leader %llx for event %s\n", pe.config,name);
       perror("Error is");
     }
+    int norm = PerformanceCounterConfigs[PCT].normalisation;
+    pe.type  = PerformanceCounterConfigs[norm].type;
+    pe.config= PerformanceCounterConfigs[norm].config;
+    name = PerformanceCounterConfigs[norm].name;
+    cyclefd = perf_event_open(&pe, 0, -1, -1, 0); // pid 0, cpu -1 current process any cpu. group -1
+    if (cyclefd == -1) {
+      fprintf(stderr, "Error opening leader %llx for event %s\n", pe.config,name);
+      perror("Error is");
+    }
 #endif
   }
 
@@ -168,6 +191,8 @@ public:
     if ( fd!= -1) {
       ::ioctl(fd, PERF_EVENT_IOC_RESET, 0);
       ::ioctl(fd, PERF_EVENT_IOC_ENABLE, 0);
+      ::ioctl(cyclefd, PERF_EVENT_IOC_RESET, 0);
+      ::ioctl(cyclefd, PERF_EVENT_IOC_ENABLE, 0);
     }
     begin  =cyclecount();
 #else
@@ -177,10 +202,13 @@ public:
 
   void Stop(void) {
     count=0;
+    cycles=0;
 #ifdef __linux__
     if ( fd!= -1) {
       ::ioctl(fd, PERF_EVENT_IOC_DISABLE, 0);
+      ::ioctl(cyclefd, PERF_EVENT_IOC_DISABLE, 0);
       ::read(fd, &count, sizeof(long long));
+      ::read(cyclefd, &cycles, sizeof(long long));
     }
     elapsed = cyclecount() - begin;
 #else
@@ -190,7 +218,11 @@ public:
   }
   void Report(void) {
 #ifdef __linux__
-    std::printf("%llu cycles %s = %20llu\n", elapsed , PerformanceCounterConfigs[PCT].name, count);
+    int N = PerformanceCounterConfigs[PCT].normalisation;
+    const char * sn = PerformanceCounterConfigs[N].name ;
+    const char * sc = PerformanceCounterConfigs[PCT].name;
+      std::printf("tsc = %llu %s = %llu  %s = %20llu\n (%s/%s) rate = %lf\n", elapsed,sn ,cycles, 
+		  sc, count, sc,sn, (double)count/(double)cycles);
 #else
     std::printf("%llu cycles \n", elapsed );
 #endif
@@ -199,7 +231,7 @@ public:
   ~PerformanceCounter()
   {
 #ifdef __linux__
-    ::close(fd);
+    ::close(fd);    ::close(cyclefd);
 #endif
   }
 

lib/Simd.h

@@ -24,7 +24,8 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 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
+See the full license in the file "LICENSE" in the top level distribution
+directory
 *************************************************************************************/
 /*  END LEGAL */
 #ifndef GRID_SIMD_H
@@ -118,6 +119,14 @@ namespace Grid {
   inline ComplexD timesI(const ComplexD &r)     { return(r*ComplexD(0.0,1.0));}
   inline ComplexF timesMinusI(const ComplexF &r){ return(r*ComplexF(0.0,-1.0));}
   inline ComplexD timesMinusI(const ComplexD &r){ return(r*ComplexD(0.0,-1.0));}
+
+  // define projections to real and imaginay parts
+  inline ComplexF projReal(const ComplexF &r){return( ComplexF(std::real(r), 0.0));}
+  inline ComplexD projReal(const ComplexD &r){return( ComplexD(std::real(r), 0.0));}
+  inline ComplexF projImag(const ComplexF &r){return (ComplexF(std::imag(r), 0.0 ));}
+  inline ComplexD projImag(const ComplexD &r){return (ComplexD(std::imag(r), 0.0));}
+
+  // define auxiliary functions for complex computations
   inline void timesI(ComplexF &ret,const ComplexF &r)     { ret = timesI(r);}
   inline void timesI(ComplexD &ret,const ComplexD &r)     { ret = timesI(r);}
   inline void timesMinusI(ComplexF &ret,const ComplexF &r){ ret = timesMinusI(r);}
@@ -163,8 +172,8 @@ namespace Grid {
 
 };
 
-#include <simd/Grid_vector_types.h>
+#include "simd/Grid_vector_types.h"
-#include <simd/Grid_vector_unops.h>
+#include "simd/Grid_vector_unops.h"
 
 namespace Grid {
   // Default precision

lib/Stat.cc

@@ -0,0 +1,247 @@
+#include <Grid.h>
+#include <PerfCount.h>
+#include <Stat.h>
+
+
+namespace Grid { 
+
+
+bool PmuStat::pmu_initialized=false;
+
+
+void PmuStat::init(const char *regname)
+{
+#ifdef __x86_64__
+  name = regname;
+  if (!pmu_initialized)
+    {
+      std::cout<<"initialising pmu"<<std::endl;
+      pmu_initialized = true;
+      pmu_init();
+    }
+  clear();
+#endif
+}
+void PmuStat::clear(void)
+{
+#ifdef __x86_64__
+  count = 0;
+  tregion = 0;
+  pmc0 = 0;
+  pmc1 = 0;
+  inst = 0;
+  cyc = 0;
+  ref = 0;
+  tcycles = 0;
+  reads = 0;
+  writes = 0;
+#endif
+}
+void PmuStat::print(void)
+{
+#ifdef __x86_64__
+  std::cout <<"Reg "<<std::string(name)<<":\n";
+  std::cout <<"  region "<<tregion<<std::endl;
+  std::cout <<"  cycles "<<tcycles<<std::endl;
+  std::cout <<"  inst   "<<inst   <<std::endl;
+  std::cout <<"  cyc    "<<cyc    <<std::endl;
+  std::cout <<"  ref    "<<ref    <<std::endl;
+  std::cout <<"  pmc0   "<<pmc0   <<std::endl;
+  std::cout <<"  pmc1   "<<pmc1   <<std::endl;
+  std::cout <<"  count  "<<count  <<std::endl;
+  std::cout <<"  reads  "<<reads  <<std::endl;
+  std::cout <<"  writes "<<writes <<std::endl;
+#endif
+}
+void PmuStat::start(void)
+{
+#ifdef __x86_64__
+  pmu_start();
+  ++count;
+  xmemctrs(&mrstart, &mwstart);
+  tstart = __rdtsc();
+#endif
+}
+void PmuStat::enter(int t)
+{
+#ifdef __x86_64__
+  counters[0][t] = __rdpmc(0);
+  counters[1][t] = __rdpmc(1);
+  counters[2][t] = __rdpmc((1<<30)|0);
+  counters[3][t] = __rdpmc((1<<30)|1);
+  counters[4][t] = __rdpmc((1<<30)|2);
+  counters[5][t] = __rdtsc();
+#endif
+}
+void PmuStat::exit(int t)
+{
+#ifdef __x86_64__
+  counters[0][t] = __rdpmc(0) - counters[0][t];
+  counters[1][t] = __rdpmc(1) - counters[1][t];
+  counters[2][t] = __rdpmc((1<<30)|0) - counters[2][t];
+  counters[3][t] = __rdpmc((1<<30)|1) - counters[3][t];
+  counters[4][t] = __rdpmc((1<<30)|2) - counters[4][t];
+  counters[5][t] = __rdtsc() - counters[5][t];
+#endif
+}
+void PmuStat::accum(int nthreads)
+{
+#ifdef __x86_64__
+  tend = __rdtsc();
+  xmemctrs(&mrend, &mwend);
+  pmu_stop();
+  for (int t = 0; t < nthreads; ++t) {
+    pmc0 += counters[0][t];
+    pmc1 += counters[1][t];
+    inst += counters[2][t];
+    cyc += counters[3][t];
+    ref += counters[4][t];
+    tcycles += counters[5][t];
+  }
+  uint64_t region = tend - tstart;
+  tregion += region;
+  uint64_t mreads = mrend - mrstart;
+  reads += mreads;
+  uint64_t mwrites = mwend - mwstart;
+  writes += mwrites;
+#endif
+}
+
+
+void PmuStat::pmu_fini(void) {}
+void PmuStat::pmu_start(void) {};
+void PmuStat::pmu_stop(void) {};
+void PmuStat::pmu_init(void)
+{
+#ifdef _KNIGHTS_LANDING_
+  KNLsetup();
+#endif
+}
+void PmuStat::xmemctrs(uint64_t *mr, uint64_t *mw)
+{
+#ifdef _KNIGHTS_LANDING_
+  ctrs c;
+  KNLreadctrs(c);
+  uint64_t emr = 0, emw = 0;
+  for (int i = 0; i < NEDC; ++i)
+    {
+      emr += c.edcrd[i];
+      emw += c.edcwr[i];
+    }
+  *mr = emr;
+  *mw = emw;
+#else
+  *mr = *mw = 0;
+#endif
+}
+
+#ifdef _KNIGHTS_LANDING_
+
+struct knl_gbl_ PmuStat::gbl;
+
+#define PMU_MEM
+
+void PmuStat::KNLevsetup(const char *ename, int &fd, int event, int umask)
+{
+  char fname[1024];
+  snprintf(fname, sizeof(fname), "%s/type", ename);
+  FILE *fp = fopen(fname, "r");
+  if (fp == 0) {
+    ::printf("open %s", fname);
+    ::exit(0);
+  }
+  int type;
+  int ret = fscanf(fp, "%d", &type);
+  assert(ret == 1);
+  fclose(fp);
+  //  std::cout << "Using PMU type "<<type<<" from " << std::string(ename) <<std::endl;
+
+  struct perf_event_attr hw = {};
+  hw.size = sizeof(hw);
+  hw.type = type;
+  // see /sys/devices/uncore_*/format/*
+  // All of the events we are interested in are configured the same way, but
+  // that isn't always true. Proper code would parse the format files
+  hw.config = event | (umask << 8);
+  //hw.read_format = PERF_FORMAT_GROUP;
+  // unfortunately the above only works within a single PMU; might
+  // as well just read them one at a time
+  int cpu = 0;
+  fd = perf_event_open(&hw, -1, cpu, -1, 0);
+  if (fd == -1) {
+    ::printf("CPU %d, box %s, event 0x%lx", cpu, ename, hw.config);
+    ::exit(0);
+  } else { 
+    //    std::cout << "event "<<std::string(ename)<<" set up for fd "<<fd<<" hw.config "<<hw.config <<std::endl;
+  }
+}
+
+
+ void PmuStat::KNLsetup(void){
+
+   int ret;
+   char fname[1024];
+
+   // MC RPQ inserts and WPQ inserts (reads & writes)
+   for (int mc = 0; mc < NMC; ++mc)
+     {
+       ::snprintf(fname, sizeof(fname), "/sys/devices/uncore_imc_%d",mc);
+       // RPQ Inserts
+       KNLevsetup(fname, gbl.mc_rd[mc], 0x1, 0x1);
+       // WPQ Inserts
+       KNLevsetup(fname, gbl.mc_wr[mc], 0x2, 0x1);
+     }
+   // EDC RPQ inserts and WPQ inserts
+   for (int edc=0; edc < NEDC; ++edc)
+     {
+       ::snprintf(fname, sizeof(fname), "/sys/devices/uncore_edc_eclk_%d",edc);
+       // RPQ inserts
+       KNLevsetup(fname, gbl.edc_rd[edc], 0x1, 0x1);
+       // WPQ inserts
+       KNLevsetup(fname, gbl.edc_wr[edc], 0x2, 0x1);
+     }
+   // EDC HitE, HitM, MissE, MissM
+   for (int edc=0; edc < NEDC; ++edc)
+     {
+       ::snprintf(fname, sizeof(fname), "/sys/devices/uncore_edc_uclk_%d", edc);
+       KNLevsetup(fname, gbl.edc_hite[edc], 0x2, 0x1);
+       KNLevsetup(fname, gbl.edc_hitm[edc], 0x2, 0x2);
+       KNLevsetup(fname, gbl.edc_misse[edc], 0x2, 0x4);
+       KNLevsetup(fname, gbl.edc_missm[edc], 0x2, 0x8);
+     }
+ }
+
+uint64_t PmuStat::KNLreadctr(int fd)
+{
+  uint64_t data;
+  size_t s = ::read(fd, &data, sizeof(data));
+  if (s != sizeof(uint64_t)){
+    ::printf("read counter %lu", s);
+    ::exit(0);
+  }
+  return data;
+}
+
+void PmuStat::KNLreadctrs(ctrs &c)
+{
+  for (int i = 0; i < NMC; ++i)
+    {
+      c.mcrd[i] = KNLreadctr(gbl.mc_rd[i]);
+      c.mcwr[i] = KNLreadctr(gbl.mc_wr[i]);
+    }
+  for (int i = 0; i < NEDC; ++i)
+    {
+      c.edcrd[i] = KNLreadctr(gbl.edc_rd[i]);
+      c.edcwr[i] = KNLreadctr(gbl.edc_wr[i]);
+    }
+  for (int i = 0; i < NEDC; ++i)
+    {
+      c.edchite[i] = KNLreadctr(gbl.edc_hite[i]);
+      c.edchitm[i] = KNLreadctr(gbl.edc_hitm[i]);
+      c.edcmisse[i] = KNLreadctr(gbl.edc_misse[i]);
+      c.edcmissm[i] = KNLreadctr(gbl.edc_missm[i]);
+    }
+}
+
+#endif
+}

lib/Stat.h

@@ -0,0 +1,104 @@
+#ifndef _GRID_STAT_H
+#define _GRID_STAT_H
+
+#ifdef AVX512
+#define _KNIGHTS_LANDING_ROOTONLY
+#endif
+
+namespace Grid { 
+
+///////////////////////////////////////////////////////////////////////////////
+// Extra KNL counters from MCDRAM
+///////////////////////////////////////////////////////////////////////////////
+#ifdef _KNIGHTS_LANDING_
+#define NMC 6
+#define NEDC 8
+struct ctrs
+{
+    uint64_t mcrd[NMC];
+    uint64_t mcwr[NMC];
+    uint64_t edcrd[NEDC]; 
+    uint64_t edcwr[NEDC];
+    uint64_t edchite[NEDC];
+    uint64_t edchitm[NEDC];
+    uint64_t edcmisse[NEDC];
+    uint64_t edcmissm[NEDC];
+};
+// Peter/Azusa:
+// Our modification of a code provided by Larry Meadows from Intel
+// Verified by email exchange non-NDA, ok for github. Should be as uses /sys/devices/ FS
+// so is already public and in the linux kernel for KNL.
+struct knl_gbl_
+{
+  int mc_rd[NMC];
+  int mc_wr[NMC];
+  int edc_rd[NEDC];
+  int edc_wr[NEDC];
+  int edc_hite[NEDC];
+  int edc_hitm[NEDC];
+  int edc_misse[NEDC];
+  int edc_missm[NEDC];
+};
+#endif
+///////////////////////////////////////////////////////////////////////////////
+
+class PmuStat
+{
+    uint64_t counters[8][256];
+#ifdef _KNIGHTS_LANDING_
+    static struct knl_gbl_ gbl;
+#endif
+    const char *name;
+
+    uint64_t reads;     // memory reads
+    uint64_t writes;    // memory writes
+    uint64_t mrstart;   // memory read counter at start of parallel region
+    uint64_t mrend;     // memory read counter at end of parallel region
+    uint64_t mwstart;   // memory write counter at start of parallel region
+    uint64_t mwend;     // memory write counter at end of parallel region
+
+    // cumulative counters
+    uint64_t count;     // number of invocations
+    uint64_t tregion;   // total time in parallel region (from thread 0)
+    uint64_t tcycles;   // total cycles inside parallel region
+    uint64_t inst, ref, cyc;   // fixed counters
+    uint64_t pmc0, pmc1;// pmu
+    // add memory counters here
+    // temp variables
+    uint64_t tstart;    // tsc at start of parallel region
+    uint64_t tend;      // tsc at end of parallel region
+    // map for ctrs values
+    // 0 pmc0 start
+    // 1 pmc0 end
+    // 2 pmc1 start
+    // 3 pmc1 end
+    // 4 tsc start
+    // 5 tsc end
+    static bool pmu_initialized;
+public:
+    static bool is_init(void){ return pmu_initialized;}
+    static void pmu_init(void);
+    static void pmu_fini(void);
+    static void pmu_start(void);
+    static void pmu_stop(void);
+    void accum(int nthreads);
+    static void xmemctrs(uint64_t *mr, uint64_t *mw);
+    void start(void);
+    void enter(int t);
+    void exit(int t);
+    void print(void);
+    void init(const char *regname);
+    void clear(void);
+#ifdef _KNIGHTS_LANDING_
+    static void     KNLsetup(void);
+    static uint64_t KNLreadctr(int fd);
+    static void     KNLreadctrs(ctrs &c);
+    static void     KNLevsetup(const char *ename, int &fd, int event, int umask);
+#endif
+    
+  };
+
+}
+#endif
+
+

lib/Stencil.h

@@ -30,7 +30,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
  #include <thread>
 
-#include <stencil/Lebesgue.h>   // subdir aggregate
+ #include <Grid/stencil/Lebesgue.h>   // subdir aggregate
 
  //////////////////////////////////////////////////////////////////////////////////////////
  // Must not lose sight that goal is to be able to construct really efficient
@@ -70,17 +70,66 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 namespace Grid {
 
+inline void Gather_plane_simple_table_compute (GridBase *grid,int dimension,int plane,int cbmask,
+					       int off,std::vector<std::pair<int,int> > & table)
+{
+  table.resize(0);
+  int rd = grid->_rdimensions[dimension];
+
+  if ( !grid->CheckerBoarded(dimension) ) {
+    cbmask = 0x3;
+  }
+  int so= plane*grid->_ostride[dimension]; // base offset for start of plane 
+  int e1=grid->_slice_nblock[dimension];
+  int e2=grid->_slice_block[dimension];
+
+  int stride=grid->_slice_stride[dimension];
+  if ( cbmask == 0x3 ) { 
+    table.resize(e1*e2);
+    for(int n=0;n<e1;n++){
+      for(int b=0;b<e2;b++){
+	int o  = n*stride;
+	int bo = n*e2;
+	table[bo+b]=std::pair<int,int>(bo+b,o+b);
+      }
+    }
+  } else { 
+     int bo=0;
+     table.resize(e1*e2/2);
+     for(int n=0;n<e1;n++){
+       for(int b=0;b<e2;b++){
+	 int o  = n*stride;
+	 int ocb=1<<grid->CheckerBoardFromOindexTable(o+b);
+	 if ( ocb &cbmask ) {
+	   table[bo]=std::pair<int,int>(bo,o+b); bo++;
+	 }
+       }
+     }
+  }
+}
+
+template<class vobj,class cobj,class compressor> void 
+Gather_plane_simple_table (std::vector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,cobj *buffer,compressor &compress, int off,int so)
+{
+PARALLEL_FOR_LOOP     
+     for(int i=0;i<table.size();i++){
+       buffer[off+table[i].first]=compress(rhs._odata[so+table[i].second]);
+     }
+}
+
 struct StencilEntry { 
-    int _offset;
+  uint64_t _offset;
-    int _is_local;
+  uint64_t _byte_offset;
-    int _permute;
+  uint16_t _is_local;
-    int _around_the_world;
+  uint16_t _permute;
+  uint32_t _around_the_world; //256 bits, 32 bytes, 1/2 cacheline
 };
 
 template<class vobj,class cobj>
 class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal fill in.
  public:
 
+  typedef CartesianCommunicator::CommsRequest_t CommsRequest_t;
   typedef uint32_t StencilInteger;
   typedef typename cobj::vector_type vector_type;
   typedef typename cobj::scalar_type scalar_type;
@@ -96,87 +145,60 @@ namespace Grid {
     Integer to_rank;
     Integer from_rank;
     Integer bytes;
-	volatile Integer done;
   };
 
   std::vector<Packet> Packets;
 
-#define SEND_IMMEDIATE
+  int face_table_computed;
-#define SERIAL_SENDS
+  std::vector<std::vector<std::pair<int,int> > > face_table ;
   
   void AddPacket(void *xmit,void * rcv, Integer to,Integer from,Integer bytes){
-	comms_bytes+=2.0*bytes;
-#ifdef SEND_IMMEDIATE
-	commtime-=usecond();
-	_grid->SendToRecvFrom(xmit,to,rcv,from,bytes);
-	commtime+=usecond();
-#endif
     Packet p;
     p.send_buf = xmit;
     p.recv_buf = rcv;
     p.to_rank  = to;
     p.from_rank= from;
     p.bytes    = bytes;
-	p.done     = 0;
     comms_bytes+=2.0*bytes;
     Packets.push_back(p);
-
   }
 
-#ifdef SERIAL_SENDS
+  void CommunicateBegin(std::vector<std::vector<CommsRequest_t> > &reqs)
-      void Communicate(void ) { 
+  {
+    reqs.resize(Packets.size());
     commtime-=usecond();
     for(int i=0;i<Packets.size();i++){
-#ifndef SEND_IMMEDIATE
+	_grid->StencilSendToRecvFromBegin(reqs[i],
-	  _grid->SendToRecvFrom(
 					  Packets[i].send_buf,
 					  Packets[i].to_rank,
 					  Packets[i].recv_buf,
 					  Packets[i].from_rank,
 					  Packets[i].bytes);
-#endif
+	/*
-	  Packets[i].done = 1;
+      }else{
+	_grid->SendToRecvFromBegin(reqs[i],
+				   Packets[i].send_buf,
+				   Packets[i].to_rank,
+				   Packets[i].recv_buf,
+				   Packets[i].from_rank,
+				   Packets[i].bytes);
+      }
+	*/
     }
     commtime+=usecond();
   }
-#else
+  void CommunicateComplete(std::vector<std::vector<CommsRequest_t> > &reqs)
-      void Communicate(void ) { 
+  {
-	typedef CartesianCommunicator::CommsRequest_t CommsRequest_t;
-	std::vector<std::vector<CommsRequest_t> > reqs(Packets.size());
     commtime-=usecond();
-	const int concurrency=2;
+
-	for(int i=0;i<Packets.size();i+=concurrency){
+    for(int i=0;i<Packets.size();i++){
-	  for(int ii=0;ii<concurrency;ii++){
+      //      if( ShmDirectCopy ) 
-	    int j = i+ii;
+	_grid->StencilSendToRecvFromComplete(reqs[i]);
-	    if ( j<Packets.size() ) {
+	//      else 
-#ifndef SEND_IMMEDIATE
+	//	_grid->SendToRecvFromComplete(reqs[i]);
-	      _grid->SendToRecvFromBegin(reqs[j],
-					 Packets[j].send_buf,
-					 Packets[j].to_rank,
-					 Packets[j].recv_buf,
-					 Packets[j].from_rank,
-					 Packets[j].bytes);
-#endif
-	    }
-	  }
-	  for(int ii=0;ii<concurrency;ii++){
-	    int j = i+ii;
-	    if ( j<Packets.size() ) {
-#ifndef SEND_IMMEDIATE
-	      _grid->SendToRecvFromComplete(reqs[i]);
-#endif
-	    }
-	  }
-	  for(int ii=0;ii<concurrency;ii++){
-	    int j = i+ii;
-	    if ( j<Packets.size() ) {
-	      Packets[j].done = 1;
-	    }
-	  }
     }
     commtime+=usecond();
   }
-#endif
 
   ///////////////////////////////////////////
   // Simd merge queue for asynch comms
@@ -196,36 +218,19 @@ namespace Grid {
     m.rpointers= rpointers;
     m.buffer_size = buffer_size;
     m.packet_id   = packet_id;
-#ifdef SEND_IMMEDIATE
-	mergetime-=usecond();
-PARALLEL_FOR_LOOP
-        for(int o=0;o<m.buffer_size;o++){
-	  merge1(m.mpointer[o],m.rpointers,o);
-	}
-	mergetime+=usecond();
-#else
     Mergers.push_back(m);
-#endif
-
   }
 
   void CommsMerge(void ) { 
-	//PARALLEL_NESTED_LOOP2 
+
     for(int i=0;i<Mergers.size();i++){	
       
-	spintime-=usecond();
-	int packet_id = Mergers[i].packet_id;
-	while(! Packets[packet_id].done ); // spin for completion
-	spintime+=usecond();
-
-#ifndef SEND_IMMEDIATE
       mergetime-=usecond();
 PARALLEL_FOR_LOOP
       for(int o=0;o<Mergers[i].buffer_size;o++){
 	merge1(Mergers[i].mpointer[o],Mergers[i].rpointers,o);
       }
       mergetime+=usecond();
-#endif
 
     }
   }
@@ -245,17 +250,59 @@ PARALLEL_FOR_LOOP
   std::vector<int>                  _permute_type;
   
   // npoints x Osites() of these
-      std::vector<std::vector<StencilEntry> > _entries;
+  // Flat vector, change layout for cache friendly.
-      inline StencilEntry * GetEntry(int &ptype,int point,int osite) { ptype = _permute_type[point]; return & _entries[point][osite]; }
+  Vector<StencilEntry>  _entries;
+  
+  void PrecomputeByteOffsets(void){
+    for(int i=0;i<_entries.size();i++){
+      if( _entries[i]._is_local ) {
+	_entries[i]._byte_offset = _entries[i]._offset*sizeof(vobj);
+      } else { 
+	_entries[i]._byte_offset = _entries[i]._offset*sizeof(cobj);
+      }
+    }
+  };
+
+  inline StencilEntry * GetEntry(int &ptype,int point,int osite) { ptype = _permute_type[point]; return & _entries[point+_npoints*osite]; }
+  inline uint64_t GetInfo(int &ptype,int &local,int &perm,int point,int ent,uint64_t base) {
+    uint64_t cbase = (uint64_t)&u_recv_buf_p[0];
+    local = _entries[ent]._is_local;
+    perm  = _entries[ent]._permute;
+    if (perm)  ptype = _permute_type[point]; 
+    if (local) {
+      return  base + _entries[ent]._byte_offset;
+    } else {
+      return cbase + _entries[ent]._byte_offset;
+    }
+  }
+  inline uint64_t GetPFInfo(int ent,uint64_t base) {
+    uint64_t cbase = (uint64_t)&u_recv_buf_p[0];
+    int local = _entries[ent]._is_local;
+    if (local) return  base + _entries[ent]._byte_offset;
+    else       return cbase + _entries[ent]._byte_offset;
+  }
+  
+  ///////////////////////////////////////////////////////////
+  // Unified Comms buffers for all directions
+  ///////////////////////////////////////////////////////////
+  // Vectors that live on the symmetric heap in case of SHMEM
+  //  std::vector<commVector<scalar_object> > u_simd_send_buf_hide;
+  //  std::vector<commVector<scalar_object> > u_simd_recv_buf_hide;
+  //  commVector<cobj>          u_send_buf_hide;
+  //  commVector<cobj>          u_recv_buf_hide;
+
+  // These are used; either SHM objects or refs to the above symmetric heap vectors
+  // depending on comms target
+  cobj* u_recv_buf_p;
+  cobj* u_send_buf_p;
+  std::vector<scalar_object *> u_simd_send_buf;
+  std::vector<scalar_object *> u_simd_recv_buf;
 
-      // Comms buffers
-      std::vector<Vector<scalar_object> > u_simd_send_buf;
-      std::vector<Vector<scalar_object> > u_simd_recv_buf;
-      Vector<cobj>          u_send_buf;
-      Vector<cobj>          comm_buf;
   int u_comm_offset;
   int _unified_buffer_size;
   
+  cobj *CommBuf(void) { return u_recv_buf_p; }
+
   /////////////////////////////////////////
   // Timing info; ugly; possibly temporary
   /////////////////////////////////////////
@@ -271,6 +318,48 @@ PARALLEL_FOR_LOOP
   double gathermtime;
   double splicetime;
   double nosplicetime;
+  double t_data;
+  double t_table;
+  double calls;
+  
+  void ZeroCounters(void) {
+    gathertime = 0.;
+    jointime = 0.;
+    commtime = 0.;
+    halogtime = 0.;
+    mergetime = 0.;
+    spintime = 0.;
+    gathermtime = 0.;
+    splicetime = 0.;
+    nosplicetime = 0.;
+    t_data = 0.0;
+    t_table= 0.0;
+    comms_bytes = 0.;
+    calls = 0.;
+  };
+  
+  void Report(void) {
+#define PRINTIT(A)	\
+ std::cout << GridLogMessage << " Stencil " << #A << " "<< A/calls<<std::endl;
+    if ( calls > 0. ) {
+      std::cout << GridLogMessage << " Stencil calls "<<calls<<std::endl;
+      PRINTIT(halogtime);
+      PRINTIT(gathertime);
+      PRINTIT(gathermtime);
+      PRINTIT(mergetime);
+      if(comms_bytes>1.0){
+	PRINTIT(comms_bytes);
+	PRINTIT(commtime);
+	std::cout << GridLogMessage << " Stencil " << comms_bytes/commtime/1000. << " GB/s "<<std::endl;
+      }
+      PRINTIT(jointime);
+      PRINTIT(spintime);
+      PRINTIT(splicetime);
+      PRINTIT(nosplicetime);
+      PRINTIT(t_table);
+	   PRINTIT(t_data);
+    }
+  };
  #endif
 
  CartesianStencil(GridBase *grid,
@@ -278,20 +367,9 @@ PARALLEL_FOR_LOOP
 		  int checkerboard,
 		  const std::vector<int> &directions,
 		  const std::vector<int> &distances) 
-    :   _entries(npoints), _permute_type(npoints), _comm_buf_size(npoints)
+   :   _permute_type(npoints), _comm_buf_size(npoints)
   {
-#ifdef TIMING_HACK
+    face_table_computed=0;
-      gathertime=0;
-      jointime=0;
-      commtime=0;
-      halogtime=0;
-      mergetime=0;
-      spintime=0;
-      gathermtime=0;
-      splicetime=0;
-      nosplicetime=0;
-      comms_bytes=0;
-#endif
     _npoints = npoints;
     _grid    = grid;
     _directions = directions;
@@ -300,13 +378,12 @@ PARALLEL_FOR_LOOP
 
     int osites  = _grid->oSites();
     
+    _entries.resize(_npoints* osites);
     for(int ii=0;ii<npoints;ii++){
       
       int i = ii; // reverse direction to get SIMD comms done first
       int point = i;
       
-	_entries[i].resize( osites);
-
       int dimension    = directions[i];
       int displacement = distances[i];
       int shift = displacement;
@@ -353,16 +430,25 @@ PARALLEL_FOR_LOOP
 	}
       }
     }
-      u_send_buf.resize(_unified_buffer_size);
-      comm_buf.resize(_unified_buffer_size);
 
+    /////////////////////////////////////////////////////////////////////////////////
+    // Try to allocate for receiving in a shared memory region, fall back to buffer
+    /////////////////////////////////////////////////////////////////////////////////
     const int Nsimd = grid->Nsimd();
+
+    _grid->ShmBufferFreeAll();
+
     u_simd_send_buf.resize(Nsimd);
     u_simd_recv_buf.resize(Nsimd);
+
+    u_send_buf_p=(cobj *)_grid->ShmBufferMalloc(_unified_buffer_size*sizeof(cobj));
+    u_recv_buf_p=(cobj *)_grid->ShmBufferMalloc(_unified_buffer_size*sizeof(cobj));
     for(int l=0;l<Nsimd;l++){
-	u_simd_send_buf[l].resize(_unified_buffer_size);
+      u_simd_recv_buf[l] = (scalar_object *)_grid->ShmBufferMalloc(_unified_buffer_size*sizeof(scalar_object));
-	u_simd_recv_buf[l].resize(_unified_buffer_size);
+      u_simd_send_buf[l] = (scalar_object *)_grid->ShmBufferMalloc(_unified_buffer_size*sizeof(scalar_object));
     }
+
+    PrecomputeByteOffsets();
   }
 
   void Local     (int point, int dimension,int shiftpm,int cbmask)
@@ -506,10 +592,11 @@ PARALLEL_FOR_LOOP
       // Simple block stride gather of SIMD objects
       for(int n=0;n<_grid->_slice_nblock[dimension];n++){
 	for(int b=0;b<_grid->_slice_block[dimension];b++){
-	    _entries[point][lo+o+b]._offset  =ro+o+b;
+	  int idx=point+(lo+o+b)*_npoints;
-	    _entries[point][lo+o+b]._is_local=1;
+	  _entries[idx]._offset  =ro+o+b;
-	    _entries[point][lo+o+b]._permute=permute;
+	  _entries[idx]._permute=permute;
-	    _entries[point][lo+o+b]._around_the_world=wrap;
+	  _entries[idx]._is_local=1;
+	  _entries[idx]._around_the_world=wrap;
 	}
 	o +=_grid->_slice_stride[dimension];
       }
@@ -526,10 +613,11 @@ PARALLEL_FOR_LOOP
 	  int ocb=1<<_grid->CheckerBoardFromOindex(o+b);
 	  
 	  if ( ocb&cbmask ) {
-	      _entries[point][lo+o+b]._offset =ro+o+b;
+	    int idx = point+(lo+o+b)*_npoints;
-	      _entries[point][lo+o+b]._is_local=1;
+	    _entries[idx]._offset =ro+o+b;
-	      _entries[point][lo+o+b]._permute=permute;
+	    _entries[idx]._is_local=1;
-	      _entries[point][lo+o+b]._around_the_world=wrap;
+	    _entries[idx]._permute=permute;
+	    _entries[idx]._around_the_world=wrap;
 	  }
 	  
 	}
@@ -552,10 +640,11 @@ PARALLEL_FOR_LOOP
       // Simple block stride gather of SIMD objects
       for(int n=0;n<_grid->_slice_nblock[dimension];n++){
 	for(int b=0;b<_grid->_slice_block[dimension];b++){
-	    _entries[point][so+o+b]._offset  =offset+(bo++);
+	  int idx=point+(so+o+b)*_npoints;
-	    _entries[point][so+o+b]._is_local=0;
+	  _entries[idx]._offset  =offset+(bo++);
-	    _entries[point][so+o+b]._permute=0;
+	  _entries[idx]._is_local=0;
-	    _entries[point][so+o+b]._around_the_world=wrap;
+	  _entries[idx]._permute=0;
+	  _entries[idx]._around_the_world=wrap;
 	}
 	o +=_grid->_slice_stride[dimension];
       }
@@ -571,10 +660,11 @@ PARALLEL_FOR_LOOP
 	  
 	  int ocb=1<<_grid->CheckerBoardFromOindex(o+b);// Could easily be a table lookup
 	  if ( ocb & cbmask ) {
-	      _entries[point][so+o+b]._offset  =offset+(bo++);
+	    int idx = point+(so+o+b)*_npoints;
-	      _entries[point][so+o+b]._is_local=0;
+	    _entries[idx]._offset  =offset+(bo++);
-	      _entries[point][so+o+b]._permute =0;
+	    _entries[idx]._is_local=0;
-	      _entries[point][so+o+b]._around_the_world=wrap;
+	    _entries[idx]._permute =0;
+	    _entries[idx]._around_the_world=wrap;
 	  }
 	}
 	o +=_grid->_slice_stride[dimension];
@@ -582,35 +672,22 @@ PARALLEL_FOR_LOOP
     }
   }
   
-
+  template<class compressor> void HaloExchange(const Lattice<vobj> &source,compressor &compress) 
-      template<class compressor>
+  {
-      std::thread HaloExchangeBegin(const Lattice<vobj> &source,compressor &compress) {
+    std::vector<std::vector<CommsRequest_t> > reqs;
+    calls++;
     Mergers.resize(0);
     Packets.resize(0);
+    _grid->StencilBarrier();
     HaloGather(source,compress);
-        return std::thread([&] { this->Communicate(); });
+    this->CommunicateBegin(reqs);
-      }
+    _grid->StencilBarrier();
-
+    this->CommunicateComplete(reqs);
-      template<class compressor>
+    _grid->StencilBarrier();
-      void HaloExchange(const Lattice<vobj> &source,compressor &compress) 
-      {
-	Mergers.resize(0); 
-	Packets.resize(0);
-	HaloGather(source,compress);
-	Communicate();
-	CommsMerge();
-      }
-
-      void HaloExchangeComplete(std::thread &thr) 
-      {
     CommsMerge(); // spins
-	jointime-=usecond();
-	thr.join();
-	jointime+=usecond();
   }
   
-      template<class compressor>
+  template<class compressor> void HaloGatherDir(const Lattice<vobj> &source,compressor &compress,int point,int & face_idx)
-      void HaloGatherDir(const Lattice<vobj> &source,compressor &compress,int point)
   {
     int dimension    = _directions[point];
     int displacement = _distances[point];
@@ -638,23 +715,23 @@ PARALLEL_FOR_LOOP
       if ( sshift[0] == sshift[1] ) {
 	if (splice_dim) {
 	  splicetime-=usecond();
-		GatherSimd(source,dimension,shift,0x3,compress);
+	  GatherSimd(source,dimension,shift,0x3,compress,face_idx);
 	  splicetime+=usecond();
 	} else { 
 	  nosplicetime-=usecond();
-		Gather(source,dimension,shift,0x3,compress);
+	  Gather(source,dimension,shift,0x3,compress,face_idx);
 	  nosplicetime+=usecond();
 	}
       } else {
 	if(splice_dim){
 	  splicetime-=usecond();
-		GatherSimd(source,dimension,shift,0x1,compress);// if checkerboard is unfavourable take two passes
+	  GatherSimd(source,dimension,shift,0x1,compress,face_idx);// if checkerboard is unfavourable take two passes
-		GatherSimd(source,dimension,shift,0x2,compress);// both with block stride loop iteration
+	  GatherSimd(source,dimension,shift,0x2,compress,face_idx);// both with block stride loop iteration
 	  splicetime+=usecond();
 	} else {
 	  nosplicetime-=usecond();
-		Gather(source,dimension,shift,0x1,compress);
+	  Gather(source,dimension,shift,0x1,compress,face_idx);
-		Gather(source,dimension,shift,0x2,compress);
+	  Gather(source,dimension,shift,0x2,compress,face_idx);
 	  nosplicetime+=usecond();
 	}
       }
@@ -668,21 +745,22 @@ PARALLEL_FOR_LOOP
     assert(source._grid==_grid);
     halogtime-=usecond();
     
-	assert (comm_buf.size() == _unified_buffer_size );
     u_comm_offset=0;
     
     // Gather all comms buffers
+    int face_idx=0;
     for(int point = 0 ; point < _npoints; point++) {
       compress.Point(point);
-	  HaloGatherDir(source,compress,point);
+      HaloGatherDir(source,compress,point,face_idx);
     }
+    face_table_computed=1;
     
     assert(u_comm_offset==_unified_buffer_size);
     halogtime+=usecond();
   }
   
   template<class compressor>
-        void Gather(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor & compress)
+  void Gather(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor & compress,int &face_idx)
   {
     typedef typename cobj::vector_type vector_type;
     typedef typename cobj::scalar_type scalar_type;
@@ -719,31 +797,54 @@ PARALLEL_FOR_LOOP
 	int bytes = words * sizeof(cobj);
 	
 	gathertime-=usecond();
-	      Gather_plane_simple (rhs,u_send_buf,dimension,sx,cbmask,compress,u_comm_offset);
+	int so  = sx*rhs._grid->_ostride[dimension]; // base offset for start of plane 
-	      gathertime+=usecond();
+	if ( !face_table_computed ) {
+	  t_table-=usecond();
+	  face_table.resize(face_idx+1);
+	  Gather_plane_simple_table_compute ((GridBase *)_grid,dimension,sx,cbmask,u_comm_offset,
+					     face_table[face_idx]);
+	  t_table+=usecond();
+	}
+	
 	
 	int rank           = _grid->_processor;
 	int recv_from_rank;
 	int xmit_to_rank;
 	_grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
+	
 	assert (xmit_to_rank   != _grid->ThisRank());
 	assert (recv_from_rank != _grid->ThisRank());
 	
-	      //      FIXME Implement asynchronous send & also avoid buffer copy
+	/////////////////////////////////////////////////////////
-	      AddPacket((void *)&u_send_buf[u_comm_offset],
+	// try the direct copy if possible
-			(void *)  &comm_buf[u_comm_offset],
+	/////////////////////////////////////////////////////////
+
+
+	cobj *send_buf = (cobj *)_grid->ShmBufferTranslate(xmit_to_rank,u_recv_buf_p);
+	if ( send_buf==NULL ) { 
+	  send_buf = u_send_buf_p;
+	}
+	//	std::cout << " send_bufs  "<<std::hex<< send_buf <<" ubp "<<u_send_buf_p <<std::dec<<std::endl;
+	t_data-=usecond();
+	assert(u_send_buf_p!=NULL);
+	assert(send_buf!=NULL);
+	Gather_plane_simple_table         (face_table[face_idx],rhs,send_buf,compress,u_comm_offset,so);  face_idx++;
+	t_data+=usecond();
+	
+	AddPacket((void *)&send_buf[u_comm_offset],
+		  (void *)&u_recv_buf_p[u_comm_offset],
 		  xmit_to_rank,
 		  recv_from_rank,
 		  bytes);
 
+	gathertime+=usecond();
 	u_comm_offset+=words;
       }
     }
   }
   
-
   template<class compressor>
-	void  GatherSimd(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor &compress)
+  void  GatherSimd(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor &compress,int & face_idx)
   {
     const int Nsimd = _grid->Nsimd();
 
@@ -822,10 +923,6 @@ PARALLEL_FOR_LOOP
 	  auto rp = &u_simd_recv_buf[i       ][u_comm_offset];
 	  auto sp = &u_simd_send_buf[nbr_lane][u_comm_offset];
 	  
-		void *vrp = (void *)rp;
-		void *vsp = (void *)sp;
-
-
 	  if(nbr_proc){
 	    
 	    int recv_from_rank;
@@ -833,9 +930,17 @@ PARALLEL_FOR_LOOP
 	    
 	    _grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank); 
  
-		  AddPacket( vsp,vrp,xmit_to_rank,recv_from_rank,bytes);
+	    scalar_object *shm = (scalar_object *) _grid->ShmBufferTranslate(recv_from_rank,sp);
+	    //	    if ((ShmDirectCopy==0)||(shm==NULL)) { 
+	    if (shm==NULL) { 
+	      shm = rp;
+	    } 
 	    
-		  rpointers[i] = rp;
+	    // if Direct, StencilSendToRecvFrom will suppress copy to a peer on node
+	    // assuming above pointer flip
+	    AddPacket((void *)sp,(void *)rp,xmit_to_rank,recv_from_rank,bytes);
+	    
+	    rpointers[i] = shm;
 	    
 	  } else { 
 	    
@@ -844,7 +949,7 @@ PARALLEL_FOR_LOOP
 	  }
 	}
 
-	      AddMerge(&comm_buf[u_comm_offset],rpointers,buffer_size,Packets.size()-1);
+	AddMerge(&u_recv_buf_p[u_comm_offset],rpointers,buffer_size,Packets.size()-1);
 	
 	u_comm_offset     +=buffer_size;
       }

lib/Tensors.h

@@ -30,22 +30,22 @@ Author: neo <cossu@post.kek.jp>
 #ifndef GRID_MATH_H
 #define GRID_MATH_H
 
-#include <tensors/Tensor_traits.h>
+#include <Grid/tensors/Tensor_traits.h>
-#include <tensors/Tensor_class.h>
+#include <Grid/tensors/Tensor_class.h>
-#include <tensors/Tensor_arith.h>
+#include <Grid/tensors/Tensor_arith.h>
-#include <tensors/Tensor_inner.h>
+#include <Grid/tensors/Tensor_inner.h>
-#include <tensors/Tensor_outer.h>
+#include <Grid/tensors/Tensor_outer.h>
-#include <tensors/Tensor_transpose.h>
+#include <Grid/tensors/Tensor_transpose.h>
-#include <tensors/Tensor_trace.h>
+#include <Grid/tensors/Tensor_trace.h>
-#include <tensors/Tensor_index.h>
+#include <Grid/tensors/Tensor_index.h>
-#include <tensors/Tensor_Ta.h>
+#include <Grid/tensors/Tensor_Ta.h>
-#include <tensors/Tensor_determinant.h>
+#include <Grid/tensors/Tensor_determinant.h>
-#include <tensors/Tensor_exp.h>
+#include <Grid/tensors/Tensor_exp.h>
-//#include <tensors/Tensor_peek.h>
+//#include <Grid/tensors/Tensor_peek.h>
-//#include <tensors/Tensor_poke.h>
+//#include <Grid/tensors/Tensor_poke.h>
-#include <tensors/Tensor_reality.h>
+#include <Grid/tensors/Tensor_reality.h>
-#include <tensors/Tensor_unary.h>
+#include <Grid/tensors/Tensor_unary.h>
-#include <tensors/Tensor_extract_merge.h>
+#include <Grid/tensors/Tensor_extract_merge.h>
-#include <tensors/Tensor_logical.h>
+#include <Grid/tensors/Tensor_logical.h>
 
 #endif

lib/Threads.h

@@ -37,7 +37,11 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 
 #ifdef GRID_OMP
 #include <omp.h>
-#define PARALLEL_FOR_LOOP _Pragma("omp parallel for ")
+#ifdef GRID_NUMA
+#define PARALLEL_FOR_LOOP _Pragma("omp parallel for schedule(static)")
+#else
+#define PARALLEL_FOR_LOOP _Pragma("omp parallel for schedule(runtime)")
+#endif
 #define PARALLEL_NESTED_LOOP2 _Pragma("omp parallel for collapse(2)")
 #else
 #define PARALLEL_FOR_LOOP 
@@ -123,6 +127,22 @@ class GridThread {
     ThreadBarrier();
   };
 
+  static void bcopy(const void *src, void *dst, size_t len) {
+#ifdef GRID_OMP
+#pragma omp parallel 
+    {
+      const char *c_src =(char *) src;
+      char *c_dest=(char *) dst;
+      int me,mywork,myoff;
+      GridThread::GetWorkBarrier(len,me, mywork,myoff);
+      bcopy(&c_src[myoff],&c_dest[myoff],mywork);
+    }
+#else 
+    bcopy(src,dst,len);
+#endif
+  }
+
+
 };
 
 }

lib/Timer.h

@@ -39,7 +39,13 @@ namespace Grid {
   // Dress the output; use std::chrono
 
 // C++11 time facilities better?
-double usecond(void);
+inline double usecond(void) {
+  struct timeval tv;
+#ifdef TIMERS_ON
+  gettimeofday(&tv,NULL);
+#endif
+  return 1.0*tv.tv_usec + 1.0e6*tv.tv_sec;
+}
 
 typedef  std::chrono::system_clock          GridClock;
 typedef  std::chrono::time_point<GridClock> GridTimePoint;
@@ -63,17 +69,23 @@ public:
   }
   void     Start(void) { 
     assert(running == false);
+#ifdef TIMERS_ON
     start = GridClock::now(); 
+#endif
     running = true;
   }
   void     Stop(void)  { 
     assert(running == true);
+#ifdef TIMERS_ON
     accumulator+= std::chrono::duration_cast<GridUsecs>(GridClock::now()-start); 
+#endif
     running = false; 
   };
   void     Reset(void){
     running = false;
+#ifdef TIMERS_ON
     start = GridClock::now();
+#endif
     accumulator = std::chrono::duration_cast<GridUsecs>(start-start); 
   }
   GridTime Elapsed(void) {

lib/algorithms/CoarsenedMatrix.h

@@ -31,7 +31,6 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #ifndef  GRID_ALGORITHM_COARSENED_MATRIX_H
 #define  GRID_ALGORITHM_COARSENED_MATRIX_H
 
-#include <Grid.h>
 
 namespace Grid {
 

lib/algorithms/SparseMatrix.h

@@ -28,7 +28,6 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef  GRID_ALGORITHM_SPARSE_MATRIX_H
 #define  GRID_ALGORITHM_SPARSE_MATRIX_H
 
-#include <Grid.h>
 
 namespace Grid {
 

lib/algorithms/approx/Chebyshev.h

@@ -29,8 +29,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_CHEBYSHEV_H
 #define GRID_CHEBYSHEV_H
 
-#include<Grid.h>
+#include <Grid/algorithms/LinearOperator.h>
-#include<algorithms/LinearOperator.h>
 
 namespace Grid {
 

lib/algorithms/approx/Remez.h

@@ -18,10 +18,10 @@
 #include <stddef.h>
 #include <Config.h>
 
-#ifdef HAVE_GMP_H
+#ifdef HAVE_LIBGMP
-#include <algorithms/approx/bigfloat.h>
+#include "bigfloat.h"
 #else
-#include <algorithms/approx/bigfloat_double.h>
+#include "bigfloat_double.h"
 #endif
 
 #define JMAX 10000 //Maximum number of iterations of Newton's approximation

lib/algorithms/iterative/ConjugateGradient.h

@@ -24,7 +24,8 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 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
+See the full license in the file "LICENSE" in the top level distribution
+directory
 *************************************************************************************/
 /*  END LEGAL */
 #ifndef GRID_CONJUGATE_GRADIENT_H
@@ -40,14 +41,17 @@ namespace Grid {
 template <class Field>
 class ConjugateGradient : public OperatorFunction<Field> {
  public:
+  bool ErrorOnNoConverge;  // throw an assert when the CG fails to converge.
+                           // Defaults true.
   RealD Tolerance;
   Integer MaxIterations;
-    ConjugateGradient(RealD tol,Integer maxit) : Tolerance(tol), MaxIterations(maxit) { 
+  ConjugateGradient(RealD tol, Integer maxit, bool err_on_no_conv = true)
-    };
+      : Tolerance(tol),
-
+        MaxIterations(maxit),
-
+        ErrorOnNoConverge(err_on_no_conv){};
-    void operator() (LinearOperatorBase<Field> &Linop,const Field &src, Field &psi){
 
+  void operator()(LinearOperatorBase<Field> &Linop, const Field &src,
+                  Field &psi) {
     psi.checkerboard = src.checkerboard;
     conformable(psi, src);
 
@@ -61,8 +65,10 @@ public:
     RealD guess = norm2(psi);
     assert(std::isnan(guess) == 0);
 
+    
     Linop.HermOpAndNorm(psi, mmp, d, b);
     
+
     r = src - mmp;
     p = r;
 
@@ -70,12 +76,18 @@ public:
     cp = a;
     ssq = norm2(src);
 
-      std::cout<<GridLogIterative <<std::setprecision(4)<< "ConjugateGradient: guess "<<guess<<std::endl;
+    std::cout << GridLogIterative << std::setprecision(4)
-      std::cout<<GridLogIterative <<std::setprecision(4)<< "ConjugateGradient:   src "<<ssq  <<std::endl;
+              << "ConjugateGradient: guess " << guess << std::endl;
-      std::cout<<GridLogIterative <<std::setprecision(4)<< "ConjugateGradient:    mp "<<d    <<std::endl;
+    std::cout << GridLogIterative << std::setprecision(4)
-      std::cout<<GridLogIterative <<std::setprecision(4)<< "ConjugateGradient:   mmp "<<b    <<std::endl;
+              << "ConjugateGradient:   src " << ssq << std::endl;
-      std::cout<<GridLogIterative <<std::setprecision(4)<< "ConjugateGradient:  cp,r "<<cp   <<std::endl;
+    std::cout << GridLogIterative << std::setprecision(4)
-      std::cout<<GridLogIterative <<std::setprecision(4)<< "ConjugateGradient:     p "<<a    <<std::endl;
+              << "ConjugateGradient:    mp " << d << std::endl;
+    std::cout << GridLogIterative << std::setprecision(4)
+              << "ConjugateGradient:   mmp " << b << std::endl;
+    std::cout << GridLogIterative << std::setprecision(4)
+              << "ConjugateGradient:  cp,r " << cp << std::endl;
+    std::cout << GridLogIterative << std::setprecision(4)
+              << "ConjugateGradient:     p " << a << std::endl;
 
     RealD rsq = Tolerance * Tolerance * ssq;
 
@@ -84,7 +96,9 @@ public:
       return;
     }
 
-      std::cout<<GridLogIterative << std::setprecision(4)<< "ConjugateGradient: k=0 residual "<<cp<<" target "<<rsq<<std::endl;
+    std::cout << GridLogIterative << std::setprecision(4)
+              << "ConjugateGradient: k=0 residual " << cp << " target " << rsq
+              << std::endl;
 
     GridStopWatch LinalgTimer;
     GridStopWatch MatrixTimer;
@@ -93,7 +107,6 @@ public:
     SolverTimer.Start();
     int k;
     for (k = 1; k <= MaxIterations; k++) {
-	
       c = cp;
 
       MatrixTimer.Start();
@@ -115,11 +128,11 @@ public:
       p = p * b + r;
 
       LinalgTimer.Stop();
-	std::cout<<GridLogIterative<<"ConjugateGradient: Iteration " <<k<<" residual "<<cp<< " target "<< rsq<<std::endl;
+      std::cout << GridLogIterative << "ConjugateGradient: Iteration " << k
+                << " residual " << cp << " target " << rsq << std::endl;
 
       // Stopping condition
       if (cp <= rsq) {
-	  
         SolverTimer.Stop();
         Linop.HermOpAndNorm(psi, mmp, d, qq);
         p = mmp - src;
@@ -130,20 +143,25 @@ public:
         RealD resnorm = sqrt(norm2(p));
         RealD true_residual = resnorm / srcnorm;
 
-	  std::cout<<GridLogMessage<<"ConjugateGradient: Converged on iteration " <<k
+        std::cout << GridLogMessage
-		   <<" computed residual "<<sqrt(cp/ssq)
+                  << "ConjugateGradient: Converged on iteration " << k << std::endl;
-		   <<" true residual "    <<true_residual
+        std::cout << GridLogMessage << "Computed residual " << sqrt(cp / ssq)
-		   <<" target "<<Tolerance<<std::endl;
+                  << " true residual " << true_residual << " target "
-	  std::cout<<GridLogMessage<<"Time elapsed: Total "<< SolverTimer.Elapsed() << " Matrix  "<<MatrixTimer.Elapsed() << " Linalg "<<LinalgTimer.Elapsed();
+                  << Tolerance << std::endl;
+        std::cout << GridLogMessage << "Time elapsed: Iterations "
+                  << SolverTimer.Elapsed() << " Matrix  "
+                  << MatrixTimer.Elapsed() << " Linalg "
+                  << LinalgTimer.Elapsed();
         std::cout << std::endl;
 
-	  assert(true_residual/Tolerance < 1000.0);
+        if (ErrorOnNoConverge) assert(true_residual / Tolerance < 1000.0);
 
         return;
       }
     }
-      std::cout<<GridLogMessage<<"ConjugateGradient did NOT converge"<<std::endl;
+    std::cout << GridLogMessage << "ConjugateGradient did NOT converge"
-      assert(0);
+              << std::endl;
+    if (ErrorOnNoConverge) assert(0);
   }
 };
 }

lib/algorithms/iterative/ConjugateGradientMixedPrec.h

@@ -0,0 +1,142 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/ConjugateGradientMixedPrec.h
+
+    Copyright (C) 2015
+
+Author: Christopher Kelly <ckelly@phys.columbia.edu>
+
+    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_CONJUGATE_GRADIENT_MIXED_PREC_H
+#define GRID_CONJUGATE_GRADIENT_MIXED_PREC_H
+
+namespace Grid {
+
+  //Mixed precision restarted defect correction CG
+  template<class FieldD,class FieldF, typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
+  class MixedPrecisionConjugateGradient : public LinearFunction<FieldD> {
+  public:                                                
+    RealD   Tolerance;
+    Integer MaxInnerIterations;
+    Integer MaxOuterIterations;
+    GridBase* SinglePrecGrid; //Grid for single-precision fields
+    RealD OuterLoopNormMult; //Stop the outer loop and move to a final double prec solve when the residual is OuterLoopNormMult * Tolerance
+    LinearOperatorBase<FieldF> &Linop_f;
+    LinearOperatorBase<FieldD> &Linop_d;
+
+    //Option to speed up *inner single precision* solves using a LinearFunction that produces a guess
+    LinearFunction<FieldF> *guesser;
+    
+    MixedPrecisionConjugateGradient(RealD tol, Integer maxinnerit, Integer maxouterit, GridBase* _sp_grid, LinearOperatorBase<FieldF> &_Linop_f, LinearOperatorBase<FieldD> &_Linop_d) :
+      Linop_f(_Linop_f), Linop_d(_Linop_d),
+      Tolerance(tol), MaxInnerIterations(maxinnerit), MaxOuterIterations(maxouterit), SinglePrecGrid(_sp_grid),
+      OuterLoopNormMult(100.), guesser(NULL){ };
+
+    void useGuesser(LinearFunction<FieldF> &g){
+      guesser = &g;
+    }
+  
+    void operator() (const FieldD &src_d_in, FieldD &sol_d){
+      GridStopWatch TotalTimer;
+      TotalTimer.Start();
+    
+      int cb = src_d_in.checkerboard;
+      sol_d.checkerboard = cb;
+    
+      RealD src_norm = norm2(src_d_in);
+      RealD stop = src_norm * Tolerance*Tolerance;
+
+      GridBase* DoublePrecGrid = src_d_in._grid;
+      FieldD tmp_d(DoublePrecGrid);
+      tmp_d.checkerboard = cb;
+    
+      FieldD tmp2_d(DoublePrecGrid);
+      tmp2_d.checkerboard = cb;
+    
+      FieldD src_d(DoublePrecGrid);
+      src_d = src_d_in; //source for next inner iteration, computed from residual during operation
+    
+      RealD inner_tol = Tolerance;
+    
+      FieldF src_f(SinglePrecGrid);
+      src_f.checkerboard = cb;
+    
+      FieldF sol_f(SinglePrecGrid);
+      sol_f.checkerboard = cb;
+    
+      ConjugateGradient<FieldF> CG_f(inner_tol, MaxInnerIterations);
+      CG_f.ErrorOnNoConverge = false;
+
+      GridStopWatch InnerCGtimer;
+
+      GridStopWatch PrecChangeTimer;
+    
+      for(Integer outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++){
+	//Compute double precision rsd and also new RHS vector.
+	Linop_d.HermOp(sol_d, tmp_d);
+	RealD norm = axpy_norm(src_d, -1., tmp_d, src_d_in); //src_d is residual vector
+      
+	std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Outer iteration " <<outer_iter<<" residual "<< norm<< " target "<< stop<<std::endl;
+
+	if(norm < OuterLoopNormMult * stop){
+	  std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Outer iteration converged on iteration " <<outer_iter <<std::endl;
+	  break;
+	}
+	while(norm * inner_tol * inner_tol < stop) inner_tol *= 2;  // inner_tol = sqrt(stop/norm) ??
+
+	PrecChangeTimer.Start();
+	precisionChange(src_f, src_d);
+	PrecChangeTimer.Stop();
+      
+	zeroit(sol_f);
+
+	//Optionally improve inner solver guess (eg using known eigenvectors)
+	if(guesser != NULL)
+	  (*guesser)(src_f, sol_f);
+
+	//Inner CG
+	CG_f.Tolerance = inner_tol;
+	InnerCGtimer.Start();
+	CG_f(Linop_f, src_f, sol_f);
+	InnerCGtimer.Stop();
+      
+	//Convert sol back to double and add to double prec solution
+	PrecChangeTimer.Start();
+	precisionChange(tmp_d, sol_f);
+	PrecChangeTimer.Stop();
+      
+	axpy(sol_d, 1.0, tmp_d, sol_d);
+      }
+    
+      //Final trial CG
+      std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Starting final patch-up double-precision solve"<<std::endl;
+    
+      ConjugateGradient<FieldD> CG_d(Tolerance, MaxInnerIterations);
+      CG_d(Linop_d, src_d_in, sol_d);
+
+      TotalTimer.Stop();
+      std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Total " << TotalTimer.Elapsed() << " Precision change " << PrecChangeTimer.Elapsed() << " Inner CG total " << InnerCGtimer.Elapsed() << std::endl;
+    }
+  };
+
+}
+
+#endif

lib/algorithms/iterative/ConjugateGradientMultiShift.h

@@ -243,8 +243,6 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector
       if ( (!converged[s]) ){
 	
 	RealD css  = c * z[s][iz]* z[s][iz];
-    if((k%100)==0 && (s==0) )
-	    std::cout<<GridLogMessage<<"ConjugateGradientMultiShift k="<<k<<" css " <<css <<std::endl;
 	
 	if(css<rsq[s]){
 	  if ( ! converged[s] )

lib/algorithms/iterative/DenseMatrix.h

@@ -130,8 +130,8 @@ DenseMatrix<T> GetSubMtx(DenseMatrix<T> &A,int row_st, int row_end, int col_st,
 
 }
 
-#include <algorithms/iterative/Householder.h>
+#include "Householder.h"
-#include <algorithms/iterative/Francis.h>
+#include "Francis.h"
 
 #endif
 

lib/algorithms/iterative/ImplicitlyRestartedLanczos.h

@@ -33,8 +33,8 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #ifdef USE_LAPACK
 #include <lapacke.h>
 #endif
-#include <algorithms/iterative/DenseMatrix.h>
+#include "DenseMatrix.h"
-#include <algorithms/iterative/EigenSort.h>
+#include "EigenSort.h"
 
 namespace Grid {
 

lib/cartesian/Cartesian_base.h

@@ -29,7 +29,6 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_CARTESIAN_BASE_H
 #define GRID_CARTESIAN_BASE_H
 
-#include <Grid.h>
 
 namespace Grid{
 
@@ -78,15 +77,12 @@ public:
     // GridCartesian / GridRedBlackCartesian
     ////////////////////////////////////////////////////////////////
     virtual int CheckerBoarded(int dim)=0;
-    virtual int CheckerBoard(std::vector<int> site)=0;
+    virtual int CheckerBoard(std::vector<int> &site)=0;
     virtual int CheckerBoardDestination(int source_cb,int shift,int dim)=0;
     virtual int CheckerBoardShift(int source_cb,int dim,int shift,int osite)=0;
     virtual int CheckerBoardShiftForCB(int source_cb,int dim,int shift,int cb)=0;
-    int  CheckerBoardFromOindex (int Oindex){
+    virtual int CheckerBoardFromOindex (int Oindex)=0;
-      std::vector<int> ocoor;
+    virtual int CheckerBoardFromOindexTable (int Oindex)=0;
-      oCoorFromOindex(ocoor,Oindex); 
-      return CheckerBoard(ocoor);
-    }
 
     //////////////////////////////////////////////////////////////////////////////////////////////
     // Local layout calculations
@@ -107,6 +103,12 @@ public:
         for(int d=0;d<_ndimension;d++) 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++) idx+=_istride[d]*(lcoor[d]/_rdimensions[d]);
+        return idx;
+    }
     inline int oIndexReduced(std::vector<int> &ocoor)
     {
       int idx=0; 
@@ -123,12 +125,6 @@ public:
     //////////////////////////////////////////////////////////
     // SIMD lane addressing
     //////////////////////////////////////////////////////////
-    inline int iIndex(std::vector<int> &lcoor)
-    {
-        int idx=0;
-        for(int d=0;d<_ndimension;d++) idx+=_istride[d]*(lcoor[d]/_rdimensions[d]);
-        return idx;
-    }
     inline void iCoorFromIindex(std::vector<int> &coor,int lane)
     {
       Lexicographic::CoorFromIndex(coor,lane,_simd_layout);

lib/cartesian/Cartesian_full.h

@@ -39,10 +39,17 @@ class GridCartesian: public GridBase {
 
 public:
 
+    virtual int  CheckerBoardFromOindexTable (int Oindex) {
+      return 0;
+    }
+    virtual int  CheckerBoardFromOindex (int Oindex)
+    {
+      return 0;
+    }
     virtual int CheckerBoarded(int dim){
       return 0;
     }
-    virtual int CheckerBoard(std::vector<int> site){
+    virtual int CheckerBoard(std::vector<int> &site){
         return 0;
     }
     virtual int CheckerBoardDestination(int cb,int shift,int dim){

lib/cartesian/Cartesian_red_black.h

@@ -37,24 +37,19 @@ namespace Grid {
   static const int Even   =CbRed;
   static const int Odd    =CbBlack;
     
-    // Perhaps these are misplaced and 
-    // should be in sparse matrix.
-    // Also should make these a named enum type
-    static const int DaggerNo=0;
-    static const int DaggerYes=1;
-
 // Specialise this for red black grids storing half the data like a chess board.
 class GridRedBlackCartesian : public GridBase
 {
 public:
     std::vector<int> _checker_dim_mask;
     int              _checker_dim;
+    std::vector<int> _checker_board;
 
     virtual int CheckerBoarded(int dim){
       if( dim==_checker_dim) return 1;
       else return 0;
     }
-    virtual int CheckerBoard(std::vector<int> site){
+    virtual int CheckerBoard(std::vector<int> &site){
       int linear=0;
       assert(site.size()==_ndimension);
       for(int d=0;d<_ndimension;d++){ 
@@ -78,12 +73,20 @@ public:
       // or by looping over x,y,z and multiply rather than computing checkerboard.
 	  
       if ( (source_cb+ocb)&1 ) {
-
 	return (shift)/2;
       } else {
 	return (shift+1)/2;
       }
     }
+    virtual int  CheckerBoardFromOindexTable (int Oindex) {
+      return _checker_board[Oindex];
+    }
+    virtual int  CheckerBoardFromOindex (int Oindex)
+    {
+      std::vector<int> ocoor;
+      oCoorFromOindex(ocoor,Oindex);
+      return CheckerBoard(ocoor);
+    }
     virtual int CheckerBoardShift(int source_cb,int dim,int shift,int osite){
 
       if(dim != _checker_dim) return shift;
@@ -170,9 +173,15 @@ public:
 	// Use a reduced simd grid
 	_simd_layout[d] = simd_layout[d];
 	_rdimensions[d]= _ldimensions[d]/_simd_layout[d];
+	assert(_rdimensions[d]>0);
 
 	// all elements of a simd vector must have same checkerboard.
-	if ( simd_layout[d]>1 ) assert((_rdimensions[d]&0x1)==0); 
+	// 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];
@@ -185,6 +194,8 @@ public:
 	  _ostride[d] = _ostride[d-1]*_rdimensions[d-1];
 	  _istride[d] = _istride[d-1]*_simd_layout[d-1];
 	}
+
+
       }
             
       ////////////////////////////////////////////////////////////////////////////////////////////
@@ -206,6 +217,18 @@ public:
 	block = block*_rdimensions[d];
       }
 
+      ////////////////////////////////////////////////
+      // Create a checkerboard lookup table
+      ////////////////////////////////////////////////
+      int rvol = 1;
+      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);
+      }
+      
     };
 protected:
     virtual int oIndex(std::vector<int> &coor)
@@ -221,6 +244,18 @@ protected:
       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;
+    }
 };
 
 }

lib/communicator/Communicator_base.cc

@@ -0,0 +1,132 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/communicator/Communicator_none.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.h"
+namespace Grid {
+
+///////////////////////////////////////////////////////////////
+// Info that is setup once and indept of cartesian layout
+///////////////////////////////////////////////////////////////
+int CartesianCommunicator::ShmRank;
+int CartesianCommunicator::ShmSize;
+int CartesianCommunicator::GroupRank;
+int CartesianCommunicator::GroupSize;
+int CartesianCommunicator::WorldRank;
+int CartesianCommunicator::WorldSize;
+int CartesianCommunicator::Slave;
+void *              CartesianCommunicator::ShmCommBuf;
+
+/////////////////////////////////
+// Alloc, free shmem region
+/////////////////////////////////
+void *CartesianCommunicator::ShmBufferMalloc(size_t bytes){
+  //  bytes = (bytes+sizeof(vRealD))&(~(sizeof(vRealD)-1));// align up bytes
+  void *ptr = (void *)heap_top;
+  heap_top  += bytes;
+  heap_bytes+= bytes;
+  std::cout <<"Shm alloc "<<ptr<<std::endl;
+  assert(heap_bytes < MAX_MPI_SHM_BYTES);
+  return ptr;
+}
+void CartesianCommunicator::ShmBufferFreeAll(void) { 
+  heap_top  =(size_t)ShmBufferSelf();
+  heap_bytes=0;
+}
+
+/////////////////////////////////
+// Grid information queries
+/////////////////////////////////
+int                      CartesianCommunicator::IsBoss(void)            { return _processor==0; };
+int                      CartesianCommunicator::BossRank(void)          { return 0; };
+int                      CartesianCommunicator::ThisRank(void)          { return _processor; };
+const std::vector<int> & CartesianCommunicator::ThisProcessorCoor(void) { return _processor_coor; };
+const std::vector<int> & CartesianCommunicator::ProcessorGrid(void)     { return _processors; };
+int                      CartesianCommunicator::ProcessorCount(void)    { return _Nprocessors; };
+
+////////////////////////////////////////////////////////////////////////////////
+// very VERY rarely (Log, serial RNG) we need world without a grid
+////////////////////////////////////////////////////////////////////////////////
+int  CartesianCommunicator::RankWorld(void){ return WorldRank; };
+int CartesianCommunicator::Ranks    (void) { return WorldSize; };
+int CartesianCommunicator::Nodes    (void) { return GroupSize; };
+int CartesianCommunicator::Cores    (void) { return ShmSize;   };
+int CartesianCommunicator::NodeRank (void) { return GroupRank; };
+int CartesianCommunicator::CoreRank (void) { return ShmRank;   };
+
+void CartesianCommunicator::GlobalSum(ComplexF &c)
+{
+  GlobalSumVector((float *)&c,2);
+}
+void CartesianCommunicator::GlobalSumVector(ComplexF *c,int N)
+{
+  GlobalSumVector((float *)c,2*N);
+}
+void CartesianCommunicator::GlobalSum(ComplexD &c)
+{
+  GlobalSumVector((double *)&c,2);
+}
+void CartesianCommunicator::GlobalSumVector(ComplexD *c,int N)
+{
+  GlobalSumVector((double *)c,2*N);
+}
+
+#ifndef GRID_COMMS_MPI3
+
+void CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+						       void *xmit,
+						       int xmit_to_rank,
+						       void *recv,
+						       int recv_from_rank,
+						       int bytes)
+{
+  SendToRecvFromBegin(list,xmit,xmit_to_rank,recv,recv_from_rank,bytes);
+}
+void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall)
+{
+  SendToRecvFromComplete(waitall);
+}
+void CartesianCommunicator::StencilBarrier(void){};
+
+commVector<uint8_t> CartesianCommunicator::ShmBufStorageVector;
+
+void *CartesianCommunicator::ShmBufferSelf(void) { return ShmCommBuf; }
+
+void *CartesianCommunicator::ShmBuffer(int rank) {
+  return NULL;
+}
+void *CartesianCommunicator::ShmBufferTranslate(int rank,void * local_p) { 
+  return NULL;
+}
+void CartesianCommunicator::ShmInitGeneric(void){
+  ShmBufStorageVector.resize(MAX_MPI_SHM_BYTES);
+  ShmCommBuf=(void *)&ShmBufStorageVector[0];
+}
+
+#endif
+  
+}
+

lib/communicator/Communicator_base.h

@@ -34,77 +34,137 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifdef GRID_COMMS_MPI
 #include <mpi.h>
 #endif
+#ifdef GRID_COMMS_MPI3
+#include <mpi.h>
+#endif
 #ifdef GRID_COMMS_SHMEM
 #include <mpp/shmem.h>
 #endif
+
 namespace Grid {
+
 class CartesianCommunicator {
   public:    
 
-  // Communicator should know nothing of the physics grid, only processor grid.
+  // 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 const uint64_t MAX_MPI_SHM_BYTES   = 128*1024*1024; 
+
+  // Communicator should know nothing of the physics grid, only processor grid.
   int              _Nprocessors;     // How many in all
   std::vector<int> _processors;      // Which dimensions get relayed out over processors lanes.
   int              _processor;       // linear processor rank
   std::vector<int> _processor_coor;  // linear processor coordinate
   unsigned long _ndimension;
 
-#ifdef GRID_COMMS_MPI
+#if defined (GRID_COMMS_MPI) || defined (GRID_COMMS_MPI3)
   MPI_Comm communicator;
+  static MPI_Comm communicator_world;
   typedef MPI_Request CommsRequest_t;
 #else 
   typedef int CommsRequest_t;
 #endif
 
+  ////////////////////////////////////////////////////////////////////
+  // Helper functionality for SHM Windows common to all other impls
+  ////////////////////////////////////////////////////////////////////
+  // Longer term; drop this in favour of a master / slave model with 
+  // cartesian communicator on a subset of ranks, slave ranks controlled
+  // by group leader with data xfer via shared memory
+  ////////////////////////////////////////////////////////////////////
+#ifdef  GRID_COMMS_MPI3
+  std::vector<int>  WorldDims;
+  std::vector<int>  GroupDims;
+  std::vector<int>  ShmDims;
+  
+  std::vector<int> GroupCoor;
+  std::vector<int> ShmCoor;
+  std::vector<int> WorldCoor;
+  
+  static std::vector<int> GroupRanks; 
+  static std::vector<int> MyGroup;
+  static int ShmSetup;
+  static MPI_Win ShmWindow; 
+  static MPI_Comm ShmComm;
+  
+  std::vector<int>  LexicographicToWorldRank;
+  
+  static std::vector<void *> ShmCommBufs;
+#else 
+  static void ShmInitGeneric(void);
+  static commVector<uint8_t> ShmBufStorageVector;
+#endif 
+  static void * ShmCommBuf;
+  size_t heap_top;
+  size_t heap_bytes;
+  void *ShmBufferSelf(void);
+  void *ShmBuffer(int rank);
+  void *ShmBufferTranslate(int rank,void * local_p);
+  void *ShmBufferMalloc(size_t bytes);
+  void ShmBufferFreeAll(void) ;
+  
+  ////////////////////////////////////////////////
+  // Must call in Grid startup
+  ////////////////////////////////////////////////
   static void Init(int *argc, char ***argv);
   
-    // Constructor
+  ////////////////////////////////////////////////
+  // Constructor of any given grid
+  ////////////////////////////////////////////////
   CartesianCommunicator(const std::vector<int> &pdimensions_in);
   
-    // Wraps MPI_Cart routines
+  ////////////////////////////////////////////////////////////////////////////////////////
+  // Wraps MPI_Cart routines, or implements equivalent on other impls
+  ////////////////////////////////////////////////////////////////////////////////////////
   void ShiftedRanks(int dim,int shift,int & source, int & dest);
   int  RankFromProcessorCoor(std::vector<int> &coor);
   void ProcessorCoorFromRank(int rank,std::vector<int> &coor);
   
   /////////////////////////////////
-    // Grid information queries
+  // Grid information and queries
   /////////////////////////////////
-    int                      IsBoss(void)            { return _processor==0; };
+  static int ShmRank;
-    int                      BossRank(void)          { return 0; };
+  static int ShmSize;
-    int                      ThisRank(void)          { return _processor; };
+  static int GroupSize;
-    const std::vector<int> & ThisProcessorCoor(void) { return _processor_coor; };
+  static int GroupRank;
-    const std::vector<int> & ProcessorGrid(void)     { return _processors; };
+  static int WorldRank;
-    int                      ProcessorCount(void)    { return _Nprocessors; };
+  static int WorldSize;
+  static int Slave;
+  
+  int                      IsBoss(void)            ;
+  int                      BossRank(void)          ;
+  int                      ThisRank(void)          ;
+  const std::vector<int> & ThisProcessorCoor(void) ;
+  const std::vector<int> & ProcessorGrid(void)     ;
+  int                      ProcessorCount(void)    ;
+  static int Ranks    (void);
+  static int Nodes    (void);
+  static int Cores    (void);
+  static int NodeRank (void);
+  static int CoreRank (void);
+
+  ////////////////////////////////////////////////////////////////////////////////
+  // very VERY rarely (Log, serial RNG) we need world without a grid
+  ////////////////////////////////////////////////////////////////////////////////
+  static int  RankWorld(void) ;
+  static void BroadcastWorld(int root,void* data, int bytes);
   
   ////////////////////////////////////////////////////////////
   // Reduction
   ////////////////////////////////////////////////////////////
   void GlobalSum(RealF &);
   void GlobalSumVector(RealF *,int N);
-
   void GlobalSum(RealD &);
   void GlobalSumVector(RealD *,int N);
-
   void GlobalSum(uint32_t &);
   void GlobalSum(uint64_t &);
-
+  void GlobalSum(ComplexF &c);
-    void GlobalSum(ComplexF &c)
+  void GlobalSumVector(ComplexF *c,int N);
-    {
+  void GlobalSum(ComplexD &c);
-      GlobalSumVector((float *)&c,2);
+  void GlobalSumVector(ComplexD *c,int N);
-    }
-    void GlobalSumVector(ComplexF *c,int N)
-    {
-      GlobalSumVector((float *)c,2*N);
-    }
-
-    void GlobalSum(ComplexD &c)
-    {
-      GlobalSumVector((double *)&c,2);
-    }
-    void GlobalSumVector(ComplexD *c,int N)
-    {
-      GlobalSumVector((double *)c,2*N);
-    }
   
   template<class obj> void GlobalSum(obj &o){
     typedef typename obj::scalar_type scalar_type;
@@ -112,6 +172,7 @@ class CartesianCommunicator {
     scalar_type * ptr = (scalar_type *)& o;
     GlobalSumVector(ptr,words);
   }
+  
   ////////////////////////////////////////////////////////////
   // Face exchange, buffer swap in translational invariant way
   ////////////////////////////////////////////////////////////
@@ -133,8 +194,19 @@ class CartesianCommunicator {
 			   void *recv,
 			   int recv_from_rank,
 			   int bytes);
+  
   void SendToRecvFromComplete(std::vector<CommsRequest_t> &waitall);
 
+  void StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+				  void *xmit,
+				  int xmit_to_rank,
+				  void *recv,
+				  int recv_from_rank,
+				  int bytes);
+  
+  void StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall);
+  void StencilBarrier(void);
+
   ////////////////////////////////////////////////////////////
   // Barrier
   ////////////////////////////////////////////////////////////
@@ -144,13 +216,12 @@ class CartesianCommunicator {
   // Broadcast a buffer and composite larger
   ////////////////////////////////////////////////////////////
   void Broadcast(int root,void* data, int bytes);
+  
   template<class obj> void Broadcast(int root,obj &data)
     {
       Broadcast(root,(void *)&data,sizeof(data));
     };
 
-    static void BroadcastWorld(int root,void* data, int bytes);
-
 }; 
 }
 

lib/communicator/Communicator_mpi.cc

@@ -30,6 +30,12 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 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;
@@ -37,12 +43,15 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
   if ( !flag ) {
     MPI_Init(argc,argv);
   }
-}
+  MPI_Comm_dup (MPI_COMM_WORLD,&communicator_world);
-
+  MPI_Comm_rank(communicator_world,&WorldRank);
-  int Rank(void) {
+  MPI_Comm_size(communicator_world,&WorldSize);
-    int pe;
+  ShmRank=0;
-    MPI_Comm_rank(MPI_COMM_WORLD,&pe);
+  ShmSize=1;
-    return pe;
+  GroupRank=WorldRank;
+  GroupSize=WorldSize;
+  Slave    =0;
+  ShmInitGeneric();
 }
 
 CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
@@ -53,9 +62,8 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
   _Nprocessors=1;
   _processors = processors;
   _processor_coor.resize(_ndimension);
-  std::cout << processors << std::endl;
   
-  MPI_Cart_create(MPI_COMM_WORLD, _ndimension,&_processors[0],&periodic[0],1,&communicator);
+  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]);
 
@@ -68,7 +76,6 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
   
   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);
@@ -169,7 +176,6 @@ void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &
   int nreq=list.size();
   std::vector<MPI_Status> status(nreq);
   int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
-
   assert(ierr==0);
 }
 
@@ -188,14 +194,17 @@ void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
 		     communicator);
   assert(ierr==0);
 }
-
+  ///////////////////////////////////////////////////////
+  // Should only be used prior to Grid Init finished.
+  // Check for this?
+  ///////////////////////////////////////////////////////
 void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
 {
   int ierr= MPI_Bcast(data,
 		      bytes,
 		      MPI_BYTE,
 		      root,
-		      MPI_COMM_WORLD);
+		      communicator_world);
   assert(ierr==0);
 }
 

lib/communicator/Communicator_mpi3.cc

@@ -0,0 +1,574 @@
+    /*************************************************************************************
+
+    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.h"
+#include <mpi.h>
+
+namespace Grid {
+
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+// Info that is setup once and indept of cartesian layout
+///////////////////////////////////////////////////////////////////////////////////////////////////
+int CartesianCommunicator::ShmSetup = 0;
+
+MPI_Comm CartesianCommunicator::communicator_world;
+MPI_Comm CartesianCommunicator::ShmComm;
+MPI_Win  CartesianCommunicator::ShmWindow;
+
+std::vector<int> CartesianCommunicator::GroupRanks;  
+std::vector<int> CartesianCommunicator::MyGroup;
+std::vector<void *> CartesianCommunicator::ShmCommBufs;
+
+void *CartesianCommunicator::ShmBufferSelf(void)
+{
+  return ShmCommBufs[ShmRank];
+}
+void *CartesianCommunicator::ShmBuffer(int rank)
+{
+  int gpeer = GroupRanks[rank];
+  if (gpeer == MPI_UNDEFINED){
+    return NULL;
+  } else { 
+    return ShmCommBufs[gpeer];
+  }
+}
+void *CartesianCommunicator::ShmBufferTranslate(int rank,void * local_p)
+{
+  int gpeer = GroupRanks[rank];
+  if (gpeer == MPI_UNDEFINED){
+    return NULL;
+  } else { 
+    uint64_t offset = (uint64_t)local_p - (uint64_t)ShmCommBufs[ShmRank];
+    uint64_t remote = (uint64_t)ShmCommBufs[gpeer]+offset;
+    return (void *) remote;
+  }
+}
+
+void CartesianCommunicator::Init(int *argc, char ***argv) {
+  int flag;
+  MPI_Initialized(&flag); // needed to coexist with other libs apparently
+  if ( !flag ) {
+    MPI_Init(argc,argv);
+  }
+
+  MPI_Comm_dup (MPI_COMM_WORLD,&communicator_world);
+  MPI_Comm_rank(communicator_world,&WorldRank);
+  MPI_Comm_size(communicator_world,&WorldSize);
+
+  /////////////////////////////////////////////////////////////////////
+  // Split into groups that can share memory
+  /////////////////////////////////////////////////////////////////////
+  MPI_Comm_split_type(communicator_world, MPI_COMM_TYPE_SHARED, 0, MPI_INFO_NULL,&ShmComm);
+  MPI_Comm_rank(ShmComm     ,&ShmRank);
+  MPI_Comm_size(ShmComm     ,&ShmSize);
+  GroupSize = WorldSize/ShmSize;
+
+  /////////////////////////////////////////////////////////////////////
+  // find world ranks in our SHM group (i.e. which ranks are on our node)
+  /////////////////////////////////////////////////////////////////////
+  MPI_Group WorldGroup, ShmGroup;
+  MPI_Comm_group (communicator_world, &WorldGroup); 
+  MPI_Comm_group (ShmComm, &ShmGroup);
+  
+  std::vector<int> world_ranks(WorldSize); 
+  GroupRanks.resize(WorldSize); 
+  MyGroup.resize(ShmSize);
+  for(int r=0;r<WorldSize;r++) world_ranks[r]=r;
+  
+  MPI_Group_translate_ranks (WorldGroup,WorldSize,&world_ranks[0],ShmGroup, &GroupRanks[0]); 
+
+  ///////////////////////////////////////////////////////////////////
+  // Identify who is in my group and noninate the leader
+    ///////////////////////////////////////////////////////////////////
+  int g=0;
+  for(int rank=0;rank<WorldSize;rank++){
+    if(GroupRanks[rank]!=MPI_UNDEFINED){
+      assert(g<ShmSize);
+      MyGroup[g++] = rank;
+    }
+  }
+  
+  std::sort(MyGroup.begin(),MyGroup.end(),std::less<int>());
+  int myleader = MyGroup[0];
+  
+  std::vector<int> leaders_1hot(WorldSize,0);
+  std::vector<int> leaders_group(GroupSize,0);
+  leaders_1hot [ myleader ] = 1;
+    
+  ///////////////////////////////////////////////////////////////////
+  // global sum leaders over comm world
+  ///////////////////////////////////////////////////////////////////
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&leaders_1hot[0],WorldSize,MPI_INT,MPI_SUM,communicator_world);
+  assert(ierr==0);
+  
+  ///////////////////////////////////////////////////////////////////
+  // find the group leaders world rank
+  ///////////////////////////////////////////////////////////////////
+  int group=0;
+  for(int l=0;l<WorldSize;l++){
+    if(leaders_1hot[l]){
+      leaders_group[group++] = l;
+    }
+  }
+  
+  ///////////////////////////////////////////////////////////////////
+  // Identify the rank of the group in which I (and my leader) live
+  ///////////////////////////////////////////////////////////////////
+  GroupRank=-1;
+  for(int g=0;g<GroupSize;g++){
+    if (myleader == leaders_group[g]){
+      GroupRank=g;
+    }
+  }
+  assert(GroupRank!=-1);
+  
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  // allocate the shared window for our group
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  
+  ShmCommBuf = 0;
+  ierr = MPI_Win_allocate_shared(MAX_MPI_SHM_BYTES,1,MPI_INFO_NULL,ShmComm,&ShmCommBuf,&ShmWindow);
+  assert(ierr==0);
+  // KNL hack -- force to numa-domain 1 in flat
+#if 0
+  //#include <numaif.h>
+  for(uint64_t page=0;page<MAX_MPI_SHM_BYTES;page+=4096){
+    void *pages = (void *) ( page + ShmCommBuf );
+    int status;
+    int flags=MPOL_MF_MOVE_ALL;
+    int nodes=1; // numa domain == MCDRAM
+    unsigned long count=1;
+    ierr= move_pages(0,count, &pages,&nodes,&status,flags);
+    if (ierr && (page==0)) perror("numa relocate command failed");
+  }
+#endif
+  MPI_Win_lock_all (MPI_MODE_NOCHECK, ShmWindow);
+  
+  /////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Plan: allocate a fixed SHM region. Scratch that is just used via some scheme during stencil comms, with no allocate free.
+  /////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
+  ShmCommBufs.resize(ShmSize);
+  for(int r=0;r<ShmSize;r++){
+    MPI_Aint sz;
+    int dsp_unit;
+    MPI_Win_shared_query (ShmWindow, r, &sz, &dsp_unit, &ShmCommBufs[r]);
+  }
+
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Verbose for now
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  if (WorldRank == 0){
+    std::cout<<GridLogMessage<< "Grid MPI-3 configuration: detected ";
+    std::cout<< WorldSize << " Ranks " ;
+    std::cout<< GroupSize << " Nodes " ;
+    std::cout<<  ShmSize  << " with ranks-per-node "<<std::endl;
+    
+    std::cout<<GridLogMessage     <<"Grid MPI-3 configuration: allocated shared memory region of size ";
+    std::cout<<std::hex << MAX_MPI_SHM_BYTES <<" ShmCommBuf address = "<<ShmCommBuf << std::dec<<std::endl;
+
+    for(int g=0;g<GroupSize;g++){
+      std::cout<<GridLogMessage<<" Node "<<g<<" led by MPI rank "<<leaders_group[g]<<std::endl;
+    }
+
+    std::cout<<GridLogMessage<<" Boss Node Shm Pointers are {";
+    for(int g=0;g<ShmSize;g++){
+      std::cout<<std::hex<<ShmCommBufs[g]<<std::dec;
+      if(g!=ShmSize-1) std::cout<<",";
+      else std::cout<<"}"<<std::endl;
+    }
+
+  }
+  
+  for(int g=0;g<GroupSize;g++){
+    if ( (ShmRank == 0) && (GroupRank==g) )  std::cout<<GridLogMessage<<"["<<g<<"] Node Group "<<g<<" is ranks {";
+    for(int r=0;r<ShmSize;r++){
+      if ( (ShmRank == 0) && (GroupRank==g) ) {
+	std::cout<<MyGroup[r];
+	if(r<ShmSize-1) std::cout<<",";
+	else std::cout<<"}"<<std::endl;
+      }
+      MPI_Barrier(communicator_world);
+    }
+  }
+  
+  assert(ShmSetup==0);  ShmSetup=1;
+}
+
+
+////////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Want to implement some magic ... Group sub-cubes into those on same node
+////////////////////////////////////////////////////////////////////////////////////////////////////////////
+void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
+{
+  std::vector<int> coor = _processor_coor;
+
+  assert(std::abs(shift) <_processors[dim]);
+
+  coor[dim] = (_processor_coor[dim] + shift + _processors[dim])%_processors[dim];
+  Lexicographic::IndexFromCoor(coor,source,_processors);
+  source = LexicographicToWorldRank[source];
+
+  coor[dim] = (_processor_coor[dim] - shift + _processors[dim])%_processors[dim];
+  Lexicographic::IndexFromCoor(coor,dest,_processors);
+  dest = LexicographicToWorldRank[dest];
+}
+int CartesianCommunicator::RankFromProcessorCoor(std::vector<int> &coor)
+{
+  int rank;
+  Lexicographic::IndexFromCoor(coor,rank,_processors);
+  rank = LexicographicToWorldRank[rank];
+  return rank;
+}
+void  CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor)
+{
+  Lexicographic::CoorFromIndex(coor,rank,_processors);
+  rank = LexicographicToWorldRank[rank];
+}
+
+CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
+{ 
+  int ierr;
+
+  communicator=communicator_world;
+
+  _ndimension = processors.size();
+  
+  ////////////////////////////////////////////////////////////////
+  // Assert power of two shm_size.
+  ////////////////////////////////////////////////////////////////
+  int log2size = -1;
+  for(int i=0;i<=MAXLOG2RANKSPERNODE;i++){  
+    if ( (0x1<<i) == ShmSize ) {
+      log2size = i;
+      break;
+    }
+  }
+  assert(log2size != -1);
+  
+  ////////////////////////////////////////////////////////////////
+  // Identify subblock of ranks on node spreading across dims
+  // in a maximally symmetrical way
+  ////////////////////////////////////////////////////////////////
+  int dim = 0;
+  
+  std::vector<int> WorldDims = processors;
+
+  ShmDims.resize(_ndimension,1);
+  GroupDims.resize(_ndimension);
+    
+  ShmCoor.resize(_ndimension);
+  GroupCoor.resize(_ndimension);
+  WorldCoor.resize(_ndimension);
+
+  for(int l2=0;l2<log2size;l2++){
+    while ( WorldDims[dim] / ShmDims[dim] <= 1 ) dim=(dim+1)%_ndimension;
+    ShmDims[dim]*=2;
+    dim=(dim+1)%_ndimension;
+  }
+
+  ////////////////////////////////////////////////////////////////
+  // Establish torus of processes and nodes with sub-blockings
+  ////////////////////////////////////////////////////////////////
+  for(int d=0;d<_ndimension;d++){
+    GroupDims[d] = WorldDims[d]/ShmDims[d];
+  }
+
+  ////////////////////////////////////////////////////////////////
+  // Check processor counts match
+  ////////////////////////////////////////////////////////////////
+  _Nprocessors=1;
+  _processors = processors;
+  _processor_coor.resize(_ndimension);
+  for(int i=0;i<_ndimension;i++){
+    _Nprocessors*=_processors[i];
+  }
+  assert(WorldSize==_Nprocessors);
+      
+  ////////////////////////////////////////////////////////////////
+  // Establish mapping between lexico physics coord and WorldRank
+  // 
+  ////////////////////////////////////////////////////////////////
+  LexicographicToWorldRank.resize(WorldSize,0);
+  Lexicographic::CoorFromIndex(GroupCoor,GroupRank,GroupDims);
+  Lexicographic::CoorFromIndex(ShmCoor,ShmRank,ShmDims);
+  for(int d=0;d<_ndimension;d++){
+    WorldCoor[d] = GroupCoor[d]*ShmDims[d]+ShmCoor[d];
+  }
+  _processor_coor = WorldCoor;
+
+  int lexico;
+  Lexicographic::IndexFromCoor(WorldCoor,lexico,WorldDims);
+  LexicographicToWorldRank[lexico]=WorldRank;
+  _processor = lexico;
+
+  ///////////////////////////////////////////////////////////////////
+  // global sum Lexico to World mapping
+  ///////////////////////////////////////////////////////////////////
+  ierr=MPI_Allreduce(MPI_IN_PLACE,&LexicographicToWorldRank[0],WorldSize,MPI_INT,MPI_SUM,communicator);
+  assert(ierr==0);
+  
+};
+
+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::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);
+}
+
+
+// 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)
+{
+#if 0
+  this->StencilBarrier();
+
+  MPI_Request xrq;
+  MPI_Request rrq;
+  
+  static int sequence;
+
+  int ierr;
+  int tag;
+  int check;
+
+  assert(dest != _processor);
+  assert(from != _processor);
+  
+  int gdest = GroupRanks[dest];
+  int gfrom = GroupRanks[from];
+  int gme   = GroupRanks[_processor];
+
+  sequence++;
+  
+  char *from_ptr = (char *)ShmCommBufs[ShmRank];
+
+  int small = (bytes<MAX_MPI_SHM_BYTES);
+
+  typedef uint64_t T;
+  int words = bytes/sizeof(T);
+
+  assert(((size_t)bytes &(sizeof(T)-1))==0);
+  assert(gme == ShmRank);
+
+  if ( small && (gdest !=MPI_UNDEFINED) ) {
+
+    char *to_ptr   = (char *)ShmCommBufs[gdest];
+
+    assert(gme != gdest);
+
+    T *ip = (T *)xmit;
+    T *op = (T *)to_ptr;
+PARALLEL_FOR_LOOP 
+    for(int w=0;w<words;w++) {
+      op[w]=ip[w];
+    }
+
+    bcopy(&_processor,&to_ptr[bytes],sizeof(_processor));
+    bcopy(&  sequence,&to_ptr[bytes+4],sizeof(sequence));
+  } else { 
+    ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
+    assert(ierr==0);
+    list.push_back(xrq);
+  }
+
+  this->StencilBarrier();
+  
+  if (small && (gfrom !=MPI_UNDEFINED) ) {
+    T *ip = (T *)from_ptr;
+    T *op = (T *)recv;
+PARALLEL_FOR_LOOP 
+    for(int w=0;w<words;w++) {
+      op[w]=ip[w];
+    }
+    bcopy(&from_ptr[bytes]  ,&tag  ,sizeof(tag));
+    bcopy(&from_ptr[bytes+4],&check,sizeof(check));
+    assert(check==sequence);
+    assert(tag==from);
+  } else { 
+    ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
+    assert(ierr==0);
+    list.push_back(rrq);
+  }
+
+  this->StencilBarrier();
+
+#else
+  MPI_Request xrq;
+  MPI_Request rrq;
+  int rank = _processor;
+  int ierr;
+  ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
+  ierr|=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
+  
+  assert(ierr==0);
+
+  list.push_back(xrq);
+  list.push_back(rrq);
+#endif
+}
+
+void CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+						       void *xmit,
+						       int dest,
+						       void *recv,
+						       int from,
+						       int bytes)
+{
+  MPI_Request xrq;
+  MPI_Request rrq;
+
+  int ierr;
+
+  assert(dest != _processor);
+  assert(from != _processor);
+  
+  int gdest = GroupRanks[dest];
+  int gfrom = GroupRanks[from];
+  int gme   = GroupRanks[_processor];
+
+  assert(gme == ShmRank);
+
+  if ( gdest == MPI_UNDEFINED ) {
+    ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
+    assert(ierr==0);
+    list.push_back(xrq);
+  }
+  
+  if ( gfrom ==MPI_UNDEFINED) {
+    ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
+    assert(ierr==0);
+    list.push_back(rrq);
+  }
+
+}
+
+
+void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list)
+{
+  SendToRecvFromComplete(list);
+}
+
+void CartesianCommunicator::StencilBarrier(void)
+{
+  MPI_Win_sync (ShmWindow);   
+  MPI_Barrier  (ShmComm);
+  MPI_Win_sync (ShmWindow);   
+}
+
+void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
+{
+  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);
+}
+
+void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
+{
+  int ierr= MPI_Bcast(data,
+		      bytes,
+		      MPI_BYTE,
+		      root,
+		      communicator_world);
+  assert(ierr==0);
+}
+
+}
+

lib/communicator/Communicator_none.cc

@@ -28,12 +28,22 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include "Grid.h"
 namespace Grid {
 
+///////////////////////////////////////////////////////////////////////////////////////////////////
+// Info that is setup once and indept of cartesian layout
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
 void CartesianCommunicator::Init(int *argc, char *** arv)
 {
+  WorldRank = 0;
+  WorldSize = 1;
+  ShmRank=0;
+  ShmSize=1;
+  GroupRank=WorldRank;
+  GroupSize=WorldSize;
+  Slave    =0;
+  ShmInitGeneric();
 }
 
-int Rank(void ){ return 0; };
-
 CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
 {
   _processors = processors;
@@ -89,30 +99,16 @@ void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &
   assert(0);
 }
 
-void CartesianCommunicator::Barrier(void)
+void CartesianCommunicator::Barrier(void){}
-{
+void CartesianCommunicator::Broadcast(int root,void* data, int bytes) {}
-}
+void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes) { }
-
+int  CartesianCommunicator::RankFromProcessorCoor(std::vector<int> &coor) {  return 0;}
-void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
+void CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor){  assert(0);}
-{
-}
-void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
-{
-}
-
-
 void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
 {
   source =0;
   dest=0;
 }
-int CartesianCommunicator::RankFromProcessorCoor(std::vector<int> &coor)
-{
-  return 0;
-}
-void  CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor)
-{
-}
 
 
 }

lib/communicator/Communicator_shmem.cc

@@ -39,17 +39,22 @@ namespace Grid {
     BACKTRACEFILE();		   \
   }\
 }
-int Rank(void) {
+
-  return shmem_my_pe();
+
-}
+///////////////////////////////////////////////////////////////////////////////////////////////////
+// Info that is setup once and indept of cartesian layout
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
 typedef struct HandShake_t { 
   uint64_t seq_local;
   uint64_t seq_remote;
 } HandShake;
 
+
 static Vector< HandShake > XConnections;
 static Vector< HandShake > RConnections;
 
+
 void CartesianCommunicator::Init(int *argc, char ***argv) {
   shmem_init();
   XConnections.resize(shmem_n_pes());
@@ -60,8 +65,17 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
     RConnections[pe].seq_local = 0;
     RConnections[pe].seq_remote= 0;
   }
+  WorldSize = shmem_n_pes();
+  WorldRank = shmem_my_pe();
+  ShmRank=0;
+  ShmSize=1;
+  GroupRank=WorldRank;
+  GroupSize=WorldSize;
+  Slave    =0;
   shmem_barrier_all();
+  ShmInitGeneric();
 }
+
 CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
 {
   _ndimension = processors.size();
@@ -230,12 +244,9 @@ void CartesianCommunicator::SendRecvPacket(void *xmit,
 
   if ( _processor == sender ) {
 
-    printf("Sender SHMEM pt2pt %d -> %d\n",sender,receiver);
     // Check he has posted a receive
     while(SendSeq->seq_remote == SendSeq->seq_local);
 
-    printf("Sender receive %d posted\n",sender,receiver);
-
     // Advance our send count
     seq = ++(SendSeq->seq_local);
     
@@ -244,26 +255,19 @@ void CartesianCommunicator::SendRecvPacket(void *xmit,
     shmem_putmem(recv,xmit,bytes,receiver);
     shmem_fence();
 
-    printf("Sender sent payload %d\n",seq);
     //Notify him we're done
     shmem_putmem((void *)&(RecvSeq->seq_remote),&seq,sizeof(seq),receiver);
     shmem_fence();
-    printf("Sender ringing door bell  %d\n",seq);
   }
   if ( _processor == receiver ) {
 
-    printf("Receiver SHMEM pt2pt %d->%d\n",sender,receiver);
     // Post a receive
     seq = ++(RecvSeq->seq_local);
     shmem_putmem((void *)&(SendSeq->seq_remote),&seq,sizeof(seq),sender);
 
-    printf("Receiver Opening letter box %d\n",seq);
-
-    
     // Now wait until he has advanced our reception counter
     while(RecvSeq->seq_remote != RecvSeq->seq_local);
 
-    printf("Receiver Got the mail %d\n",seq);
   }
 }
 

lib/cshift/Cshift_common.h

@@ -1,3 +1,4 @@
+
     /*************************************************************************************
 
     Grid physics library, www.github.com/paboyle/Grid 
@@ -44,7 +45,7 @@ public:
 // Gather for when there is no need to SIMD split with compression
 ///////////////////////////////////////////////////////////////////
 template<class vobj,class cobj,class compressor> void 
-Gather_plane_simple (const Lattice<vobj> &rhs,std::vector<cobj,alignedAllocator<cobj> > &buffer,int dimension,int plane,int cbmask,compressor &compress, int off=0)
+Gather_plane_simple (const Lattice<vobj> &rhs,commVector<cobj> &buffer,int dimension,int plane,int cbmask,compressor &compress, int off=0)
 {
   int rd = rhs._grid->_rdimensions[dimension];
 
@@ -56,6 +57,7 @@ Gather_plane_simple (const Lattice<vobj> &rhs,std::vector<cobj,alignedAllocator<
   
   int e1=rhs._grid->_slice_nblock[dimension];
   int e2=rhs._grid->_slice_block[dimension];
+
   int stride=rhs._grid->_slice_stride[dimension];
   if ( cbmask == 0x3 ) { 
 PARALLEL_NESTED_LOOP2
@@ -68,15 +70,20 @@ PARALLEL_NESTED_LOOP2
     }
   } else { 
      int bo=0;
+     std::vector<std::pair<int,int> > table;
      for(int n=0;n<e1;n++){
        for(int b=0;b<e2;b++){
 	 int o  = n*stride;
-	 int ocb=1<<rhs._grid->CheckerBoardFromOindex(o+b);// Could easily be a table lookup
+	 int ocb=1<<rhs._grid->CheckerBoardFromOindexTable(o+b);
 	 if ( ocb &cbmask ) {
-	   buffer[off+bo++]=compress(rhs._odata[so+o+b]);
+	   table.push_back(std::pair<int,int> (bo++,o+b));
 	 }
        }
      }
+PARALLEL_FOR_LOOP     
+     for(int i=0;i<table.size();i++){
+       buffer[off+table[i].first]=compress(rhs._odata[so+table[i].second]);
+     }
   }
 }
 
@@ -107,6 +114,7 @@ PARALLEL_NESTED_LOOP2
 	int o      =   n*n1;
 	int offset = b+n*n2;
 	cobj temp =compress(rhs._odata[so+o+b]);
+
 	extract<cobj>(temp,pointers,offset);
 
       }
@@ -114,6 +122,7 @@ PARALLEL_NESTED_LOOP2
   } else { 
 
     assert(0); //Fixme think this is buggy
+
     for(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
 	int o=n*rhs._grid->_slice_stride[dimension];
@@ -132,7 +141,7 @@ PARALLEL_NESTED_LOOP2
 //////////////////////////////////////////////////////
 // Gather for when there is no need to SIMD split
 //////////////////////////////////////////////////////
-template<class vobj> void Gather_plane_simple (const Lattice<vobj> &rhs,std::vector<vobj,alignedAllocator<vobj> > &buffer,             int dimension,int plane,int cbmask)
+template<class vobj> void Gather_plane_simple (const Lattice<vobj> &rhs,commVector<vobj> &buffer, int dimension,int plane,int cbmask)
 {
   SimpleCompressor<vobj> dontcompress;
   Gather_plane_simple (rhs,buffer,dimension,plane,cbmask,dontcompress);
@@ -150,7 +159,7 @@ template<class vobj> void Gather_plane_extract(const Lattice<vobj> &rhs,std::vec
 //////////////////////////////////////////////////////
 // Scatter for when there is no need to SIMD split
 //////////////////////////////////////////////////////
-template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,std::vector<vobj,alignedAllocator<vobj> > &buffer, int dimension,int plane,int cbmask)
+template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,commVector<vobj> &buffer, int dimension,int plane,int cbmask)
 {
   int rd = rhs._grid->_rdimensions[dimension];
 

lib/cshift/Cshift_mpi.h

@@ -119,8 +119,8 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
   assert(shift<fd);
   
   int buffer_size = rhs._grid->_slice_nblock[dimension]*rhs._grid->_slice_block[dimension];
-  std::vector<vobj,alignedAllocator<vobj> > send_buf(buffer_size);
+  commVector<vobj> send_buf(buffer_size);
-  std::vector<vobj,alignedAllocator<vobj> > recv_buf(buffer_size);
+  commVector<vobj> recv_buf(buffer_size);
 
   int cb= (cbmask==0x2)? Odd : Even;
   int sshift= rhs._grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,cb);
@@ -191,8 +191,8 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
   int buffer_size = grid->_slice_nblock[dimension]*grid->_slice_block[dimension];
   int words = sizeof(vobj)/sizeof(vector_type);
 
-  std::vector<Vector<scalar_object> >   send_buf_extract(Nsimd,Vector<scalar_object>(buffer_size) );
+  std::vector<commVector<scalar_object> >   send_buf_extract(Nsimd,commVector<scalar_object>(buffer_size) );
-  std::vector<Vector<scalar_object> >   recv_buf_extract(Nsimd,Vector<scalar_object>(buffer_size) );
+  std::vector<commVector<scalar_object> >   recv_buf_extract(Nsimd,commVector<scalar_object>(buffer_size) );
 
   int bytes = buffer_size*sizeof(scalar_object);
 

lib/fftw/fftw3.h

@@ -0,0 +1,412 @@
+/*
+ * Copyright (c) 2003, 2007-14 Matteo Frigo
+ * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
+ *
+ * The following statement of license applies *only* to this header file,
+ * and *not* to the other files distributed with FFTW or derived therefrom:
+ * 
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *
+ * 1. Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ *
+ * 2. Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in the
+ *    documentation and/or other materials provided with the distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS
+ * OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY
+ * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+ * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE
+ * GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+ * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
+ * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+ * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+/***************************** NOTE TO USERS *********************************
+ *
+ *                 THIS IS A HEADER FILE, NOT A MANUAL
+ *
+ *    If you want to know how to use FFTW, please read the manual,
+ *    online at http://www.fftw.org/doc/ and also included with FFTW.
+ *    For a quick start, see the manual's tutorial section.
+ *
+ *   (Reading header files to learn how to use a library is a habit
+ *    stemming from code lacking a proper manual.  Arguably, it's a
+ *    *bad* habit in most cases, because header files can contain
+ *    interfaces that are not part of the public, stable API.)
+ *
+ ****************************************************************************/
+
+#ifndef FFTW3_H
+#define FFTW3_H
+
+#include <stdio.h>
+
+#ifdef __cplusplus
+extern "C"
+{
+#endif /* __cplusplus */
+
+/* If <complex.h> is included, use the C99 complex type.  Otherwise
+   define a type bit-compatible with C99 complex */
+#if !defined(FFTW_NO_Complex) && defined(_Complex_I) && defined(complex) && defined(I)
+#  define FFTW_DEFINE_COMPLEX(R, C) typedef R _Complex C
+#else
+#  define FFTW_DEFINE_COMPLEX(R, C) typedef R C[2]
+#endif
+
+#define FFTW_CONCAT(prefix, name) prefix ## name
+#define FFTW_MANGLE_DOUBLE(name) FFTW_CONCAT(fftw_, name)
+#define FFTW_MANGLE_FLOAT(name) FFTW_CONCAT(fftwf_, name)
+#define FFTW_MANGLE_LONG_DOUBLE(name) FFTW_CONCAT(fftwl_, name)
+#define FFTW_MANGLE_QUAD(name) FFTW_CONCAT(fftwq_, name)
+
+/* IMPORTANT: for Windows compilers, you should add a line
+        #define FFTW_DLL
+   here and in kernel/ifftw.h if you are compiling/using FFTW as a
+   DLL, in order to do the proper importing/exporting, or
+   alternatively compile with -DFFTW_DLL or the equivalent
+   command-line flag.  This is not necessary under MinGW/Cygwin, where
+   libtool does the imports/exports automatically. */
+#if defined(FFTW_DLL) && (defined(_WIN32) || defined(__WIN32__))
+   /* annoying Windows syntax for shared-library declarations */
+#  if defined(COMPILING_FFTW) /* defined in api.h when compiling FFTW */
+#    define FFTW_EXTERN extern __declspec(dllexport) 
+#  else /* user is calling FFTW; import symbol */
+#    define FFTW_EXTERN extern __declspec(dllimport) 
+#  endif
+#else
+#  define FFTW_EXTERN extern
+#endif
+
+enum fftw_r2r_kind_do_not_use_me {
+     FFTW_R2HC=0, FFTW_HC2R=1, FFTW_DHT=2,
+     FFTW_REDFT00=3, FFTW_REDFT01=4, FFTW_REDFT10=5, FFTW_REDFT11=6,
+     FFTW_RODFT00=7, FFTW_RODFT01=8, FFTW_RODFT10=9, FFTW_RODFT11=10
+};
+
+struct fftw_iodim_do_not_use_me {
+     int n;                     /* dimension size */
+     int is;			/* input stride */
+     int os;			/* output stride */
+};
+
+#include <stddef.h> /* for ptrdiff_t */
+struct fftw_iodim64_do_not_use_me {
+     ptrdiff_t n;                     /* dimension size */
+     ptrdiff_t is;			/* input stride */
+     ptrdiff_t os;			/* output stride */
+};
+
+typedef void (*fftw_write_char_func_do_not_use_me)(char c, void *);
+typedef int (*fftw_read_char_func_do_not_use_me)(void *);
+
+/*
+  huge second-order macro that defines prototypes for all API
+  functions.  We expand this macro for each supported precision
+ 
+  X: name-mangling macro
+  R: real data type
+  C: complex data type
+*/
+
+#define FFTW_DEFINE_API(X, R, C)					   \
+									   \
+FFTW_DEFINE_COMPLEX(R, C);						   \
+									   \
+typedef struct X(plan_s) *X(plan);					   \
+									   \
+typedef struct fftw_iodim_do_not_use_me X(iodim);			   \
+typedef struct fftw_iodim64_do_not_use_me X(iodim64);			   \
+									   \
+typedef enum fftw_r2r_kind_do_not_use_me X(r2r_kind);			   \
+									   \
+typedef fftw_write_char_func_do_not_use_me X(write_char_func);		   \
+typedef fftw_read_char_func_do_not_use_me X(read_char_func);		   \
+									   \
+FFTW_EXTERN void X(execute)(const X(plan) p);				   \
+									   \
+FFTW_EXTERN X(plan) X(plan_dft)(int rank, const int *n,			   \
+		    C *in, C *out, int sign, unsigned flags);		   \
+									   \
+FFTW_EXTERN X(plan) X(plan_dft_1d)(int n, C *in, C *out, int sign,	   \
+		       unsigned flags);					   \
+FFTW_EXTERN X(plan) X(plan_dft_2d)(int n0, int n1,			   \
+		       C *in, C *out, int sign, unsigned flags);	   \
+FFTW_EXTERN X(plan) X(plan_dft_3d)(int n0, int n1, int n2,		   \
+		       C *in, C *out, int sign, unsigned flags);	   \
+									   \
+FFTW_EXTERN X(plan) X(plan_many_dft)(int rank, const int *n,		   \
+                         int howmany,					   \
+                         C *in, const int *inembed,			   \
+                         int istride, int idist,			   \
+                         C *out, const int *onembed,			   \
+                         int ostride, int odist,			   \
+                         int sign, unsigned flags);			   \
+									   \
+FFTW_EXTERN X(plan) X(plan_guru_dft)(int rank, const X(iodim) *dims,	   \
+			 int howmany_rank,				   \
+			 const X(iodim) *howmany_dims,			   \
+			 C *in, C *out,					   \
+			 int sign, unsigned flags);			   \
+FFTW_EXTERN X(plan) X(plan_guru_split_dft)(int rank, const X(iodim) *dims, \
+			 int howmany_rank,				   \
+			 const X(iodim) *howmany_dims,			   \
+			 R *ri, R *ii, R *ro, R *io,			   \
+			 unsigned flags);				   \
+									   \
+FFTW_EXTERN X(plan) X(plan_guru64_dft)(int rank,			   \
+                         const X(iodim64) *dims,			   \
+			 int howmany_rank,				   \
+			 const X(iodim64) *howmany_dims,		   \
+			 C *in, C *out,					   \
+			 int sign, unsigned flags);			   \
+FFTW_EXTERN X(plan) X(plan_guru64_split_dft)(int rank,			   \
+                         const X(iodim64) *dims,			   \
+			 int howmany_rank,				   \
+			 const X(iodim64) *howmany_dims,		   \
+			 R *ri, R *ii, R *ro, R *io,			   \
+			 unsigned flags);				   \
+									   \
+FFTW_EXTERN void X(execute_dft)(const X(plan) p, C *in, C *out);	   \
+FFTW_EXTERN void X(execute_split_dft)(const X(plan) p, R *ri, R *ii,	   \
+                                      R *ro, R *io);			   \
+									   \
+FFTW_EXTERN X(plan) X(plan_many_dft_r2c)(int rank, const int *n,	   \
+                             int howmany,				   \
+                             R *in, const int *inembed,			   \
+                             int istride, int idist,			   \
+                             C *out, const int *onembed,		   \
+                             int ostride, int odist,			   \
+                             unsigned flags);				   \
+									   \
+FFTW_EXTERN X(plan) X(plan_dft_r2c)(int rank, const int *n,		   \
+                        R *in, C *out, unsigned flags);			   \
+									   \
+FFTW_EXTERN X(plan) X(plan_dft_r2c_1d)(int n,R *in,C *out,unsigned flags); \
+FFTW_EXTERN X(plan) X(plan_dft_r2c_2d)(int n0, int n1,			   \
+			   R *in, C *out, unsigned flags);		   \
+FFTW_EXTERN X(plan) X(plan_dft_r2c_3d)(int n0, int n1,			   \
+			   int n2,					   \
+			   R *in, C *out, unsigned flags);		   \
+									   \
+									   \
+FFTW_EXTERN X(plan) X(plan_many_dft_c2r)(int rank, const int *n,	   \
+			     int howmany,				   \
+			     C *in, const int *inembed,			   \
+			     int istride, int idist,			   \
+			     R *out, const int *onembed,		   \
+			     int ostride, int odist,			   \
+			     unsigned flags);				   \
+									   \
+FFTW_EXTERN X(plan) X(plan_dft_c2r)(int rank, const int *n,		   \
+                        C *in, R *out, unsigned flags);			   \
+									   \
+FFTW_EXTERN X(plan) X(plan_dft_c2r_1d)(int n,C *in,R *out,unsigned flags); \
+FFTW_EXTERN X(plan) X(plan_dft_c2r_2d)(int n0, int n1,			   \
+			   C *in, R *out, unsigned flags);		   \
+FFTW_EXTERN X(plan) X(plan_dft_c2r_3d)(int n0, int n1,			   \
+			   int n2,					   \
+			   C *in, R *out, unsigned flags);		   \
+									   \
+FFTW_EXTERN X(plan) X(plan_guru_dft_r2c)(int rank, const X(iodim) *dims,   \
+			     int howmany_rank,				   \
+			     const X(iodim) *howmany_dims,		   \
+			     R *in, C *out,				   \
+			     unsigned flags);				   \
+FFTW_EXTERN X(plan) X(plan_guru_dft_c2r)(int rank, const X(iodim) *dims,   \
+			     int howmany_rank,				   \
+			     const X(iodim) *howmany_dims,		   \
+			     C *in, R *out,				   \
+			     unsigned flags);				   \
+									   \
+FFTW_EXTERN X(plan) X(plan_guru_split_dft_r2c)(				   \
+                             int rank, const X(iodim) *dims,		   \
+			     int howmany_rank,				   \
+			     const X(iodim) *howmany_dims,		   \
+			     R *in, R *ro, R *io,			   \
+			     unsigned flags);				   \
+FFTW_EXTERN X(plan) X(plan_guru_split_dft_c2r)(				   \
+                             int rank, const X(iodim) *dims,		   \
+			     int howmany_rank,				   \
+			     const X(iodim) *howmany_dims,		   \
+			     R *ri, R *ii, R *out,			   \
+			     unsigned flags);				   \
+									   \
+FFTW_EXTERN X(plan) X(plan_guru64_dft_r2c)(int rank,			   \
+                             const X(iodim64) *dims,			   \
+			     int howmany_rank,				   \
+			     const X(iodim64) *howmany_dims,		   \
+			     R *in, C *out,				   \
+			     unsigned flags);				   \
+FFTW_EXTERN X(plan) X(plan_guru64_dft_c2r)(int rank,			   \
+                             const X(iodim64) *dims,			   \
+			     int howmany_rank,				   \
+			     const X(iodim64) *howmany_dims,		   \
+			     C *in, R *out,				   \
+			     unsigned flags);				   \
+									   \
+FFTW_EXTERN X(plan) X(plan_guru64_split_dft_r2c)(			   \
+                             int rank, const X(iodim64) *dims,		   \
+			     int howmany_rank,				   \
+			     const X(iodim64) *howmany_dims,		   \
+			     R *in, R *ro, R *io,			   \
+			     unsigned flags);				   \
+FFTW_EXTERN X(plan) X(plan_guru64_split_dft_c2r)(			   \
+                             int rank, const X(iodim64) *dims,		   \
+			     int howmany_rank,				   \
+			     const X(iodim64) *howmany_dims,		   \
+			     R *ri, R *ii, R *out,			   \
+			     unsigned flags);				   \
+									   \
+FFTW_EXTERN void X(execute_dft_r2c)(const X(plan) p, R *in, C *out);	   \
+FFTW_EXTERN void X(execute_dft_c2r)(const X(plan) p, C *in, R *out);	   \
+									   \
+FFTW_EXTERN void X(execute_split_dft_r2c)(const X(plan) p,		   \
+                                          R *in, R *ro, R *io);		   \
+FFTW_EXTERN void X(execute_split_dft_c2r)(const X(plan) p,		   \
+                                          R *ri, R *ii, R *out);	   \
+									   \
+FFTW_EXTERN X(plan) X(plan_many_r2r)(int rank, const int *n,		   \
+                         int howmany,					   \
+                         R *in, const int *inembed,			   \
+                         int istride, int idist,			   \
+                         R *out, const int *onembed,			   \
+                         int ostride, int odist,			   \
+                         const X(r2r_kind) *kind, unsigned flags);	   \
+									   \
+FFTW_EXTERN X(plan) X(plan_r2r)(int rank, const int *n, R *in, R *out,	   \
+                    const X(r2r_kind) *kind, unsigned flags);		   \
+									   \
+FFTW_EXTERN X(plan) X(plan_r2r_1d)(int n, R *in, R *out,		   \
+                       X(r2r_kind) kind, unsigned flags);		   \
+FFTW_EXTERN X(plan) X(plan_r2r_2d)(int n0, int n1, R *in, R *out,	   \
+                       X(r2r_kind) kind0, X(r2r_kind) kind1,		   \
+                       unsigned flags);					   \
+FFTW_EXTERN X(plan) X(plan_r2r_3d)(int n0, int n1, int n2,		   \
+                       R *in, R *out, X(r2r_kind) kind0,		   \
+                       X(r2r_kind) kind1, X(r2r_kind) kind2,		   \
+                       unsigned flags);					   \
+									   \
+FFTW_EXTERN X(plan) X(plan_guru_r2r)(int rank, const X(iodim) *dims,	   \
+                         int howmany_rank,				   \
+                         const X(iodim) *howmany_dims,			   \
+                         R *in, R *out,					   \
+                         const X(r2r_kind) *kind, unsigned flags);	   \
+									   \
+FFTW_EXTERN X(plan) X(plan_guru64_r2r)(int rank, const X(iodim64) *dims,   \
+                         int howmany_rank,				   \
+                         const X(iodim64) *howmany_dims,		   \
+                         R *in, R *out,					   \
+                         const X(r2r_kind) *kind, unsigned flags);	   \
+									   \
+FFTW_EXTERN void X(execute_r2r)(const X(plan) p, R *in, R *out);	   \
+									   \
+FFTW_EXTERN void X(destroy_plan)(X(plan) p);				   \
+FFTW_EXTERN void X(forget_wisdom)(void);				   \
+FFTW_EXTERN void X(cleanup)(void);					   \
+									   \
+FFTW_EXTERN void X(set_timelimit)(double t);				   \
+									   \
+FFTW_EXTERN void X(plan_with_nthreads)(int nthreads);			   \
+FFTW_EXTERN int X(init_threads)(void);					   \
+FFTW_EXTERN void X(cleanup_threads)(void);				   \
+									   \
+FFTW_EXTERN int X(export_wisdom_to_filename)(const char *filename);	   \
+FFTW_EXTERN void X(export_wisdom_to_file)(FILE *output_file);		   \
+FFTW_EXTERN char *X(export_wisdom_to_string)(void);			   \
+FFTW_EXTERN void X(export_wisdom)(X(write_char_func) write_char,   	   \
+                                  void *data);				   \
+FFTW_EXTERN int X(import_system_wisdom)(void);				   \
+FFTW_EXTERN int X(import_wisdom_from_filename)(const char *filename);	   \
+FFTW_EXTERN int X(import_wisdom_from_file)(FILE *input_file);		   \
+FFTW_EXTERN int X(import_wisdom_from_string)(const char *input_string);	   \
+FFTW_EXTERN int X(import_wisdom)(X(read_char_func) read_char, void *data); \
+									   \
+FFTW_EXTERN void X(fprint_plan)(const X(plan) p, FILE *output_file);	   \
+FFTW_EXTERN void X(print_plan)(const X(plan) p);			   \
+FFTW_EXTERN char *X(sprint_plan)(const X(plan) p);			   \
+									   \
+FFTW_EXTERN void *X(malloc)(size_t n);					   \
+FFTW_EXTERN R *X(alloc_real)(size_t n);					   \
+FFTW_EXTERN C *X(alloc_complex)(size_t n);				   \
+FFTW_EXTERN void X(free)(void *p);					   \
+									   \
+FFTW_EXTERN void X(flops)(const X(plan) p,				   \
+                          double *add, double *mul, double *fmas);	   \
+FFTW_EXTERN double X(estimate_cost)(const X(plan) p);			   \
+FFTW_EXTERN double X(cost)(const X(plan) p);				   \
+									   \
+FFTW_EXTERN int X(alignment_of)(R *p);                                     \
+FFTW_EXTERN const char X(version)[];                                       \
+FFTW_EXTERN const char X(cc)[];						   \
+FFTW_EXTERN const char X(codelet_optim)[];
+
+
+/* end of FFTW_DEFINE_API macro */
+
+FFTW_DEFINE_API(FFTW_MANGLE_DOUBLE, double, fftw_complex)
+FFTW_DEFINE_API(FFTW_MANGLE_FLOAT, float, fftwf_complex)
+FFTW_DEFINE_API(FFTW_MANGLE_LONG_DOUBLE, long double, fftwl_complex)
+
+/* __float128 (quad precision) is a gcc extension on i386, x86_64, and ia64
+   for gcc >= 4.6 (compiled in FFTW with --enable-quad-precision) */
+#if (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 6)) \
+ && !(defined(__ICC) || defined(__INTEL_COMPILER)) \
+ && (defined(__i386__) || defined(__x86_64__) || defined(__ia64__))
+#  if !defined(FFTW_NO_Complex) && defined(_Complex_I) && defined(complex) && defined(I)
+/* note: __float128 is a typedef, which is not supported with the _Complex
+         keyword in gcc, so instead we use this ugly __attribute__ version.
+         However, we can't simply pass the __attribute__ version to
+         FFTW_DEFINE_API because the __attribute__ confuses gcc in pointer
+         types.  Hence redefining FFTW_DEFINE_COMPLEX.  Ugh. */
+#    undef FFTW_DEFINE_COMPLEX
+#    define FFTW_DEFINE_COMPLEX(R, C) typedef _Complex float __attribute__((mode(TC))) C
+#  endif
+FFTW_DEFINE_API(FFTW_MANGLE_QUAD, __float128, fftwq_complex)
+#endif
+
+#define FFTW_FORWARD (-1)
+#define FFTW_BACKWARD (+1)
+
+#define FFTW_NO_TIMELIMIT (-1.0)
+
+/* documented flags */
+#define FFTW_MEASURE (0U)
+#define FFTW_DESTROY_INPUT (1U << 0)
+#define FFTW_UNALIGNED (1U << 1)
+#define FFTW_CONSERVE_MEMORY (1U << 2)
+#define FFTW_EXHAUSTIVE (1U << 3) /* NO_EXHAUSTIVE is default */
+#define FFTW_PRESERVE_INPUT (1U << 4) /* cancels FFTW_DESTROY_INPUT */
+#define FFTW_PATIENT (1U << 5) /* IMPATIENT is default */
+#define FFTW_ESTIMATE (1U << 6)
+#define FFTW_WISDOM_ONLY (1U << 21)
+
+/* undocumented beyond-guru flags */
+#define FFTW_ESTIMATE_PATIENT (1U << 7)
+#define FFTW_BELIEVE_PCOST (1U << 8)
+#define FFTW_NO_DFT_R2HC (1U << 9)
+#define FFTW_NO_NONTHREADED (1U << 10)
+#define FFTW_NO_BUFFERING (1U << 11)
+#define FFTW_NO_INDIRECT_OP (1U << 12)
+#define FFTW_ALLOW_LARGE_GENERIC (1U << 13) /* NO_LARGE_GENERIC is default */
+#define FFTW_NO_RANK_SPLITS (1U << 14)
+#define FFTW_NO_VRANK_SPLITS (1U << 15)
+#define FFTW_NO_VRECURSE (1U << 16)
+#define FFTW_NO_SIMD (1U << 17)
+#define FFTW_NO_SLOW (1U << 18)
+#define FFTW_NO_FIXED_RADIX_LARGE_N (1U << 19)
+#define FFTW_ALLOW_PRUNING (1U << 20)
+
+#ifdef __cplusplus
+}  /* extern "C" */
+#endif /* __cplusplus */
+
+#endif /* FFTW3_H */

lib/lattice/Lattice_ET.h

@@ -24,16 +24,17 @@ Author: neo <cossu@post.kek.jp>
 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
+See the full license in the file "LICENSE" in the top level distribution
+directory
 *************************************************************************************/
 /*  END LEGAL */
 #ifndef GRID_LATTICE_ET_H
 #define GRID_LATTICE_ET_H
 
 #include <iostream>
-#include <vector>
 #include <tuple>
 #include <typeinfo>
+#include <vector>
 
 namespace Grid {
 
@@ -41,8 +42,8 @@ namespace Grid {
 // Predicated where support
 ////////////////////////////////////////////////////
 template <class iobj, class vobj, class robj>
-    inline vobj predicatedWhere(const iobj &predicate,const vobj &iftrue,const robj &iffalse) {
+inline vobj predicatedWhere(const iobj &predicate, const vobj &iftrue,
-
+                            const robj &iffalse) {
   typename std::remove_const<vobj>::type ret;
 
   typedef typename vobj::scalar_object scalar_object;
@@ -75,50 +76,69 @@ namespace Grid {
 // from tuple is hideous; C++14 introduces std::make_index_sequence for this
 ////////////////////////////////////////////
 
-
 // leaf eval of lattice ; should enable if protect using traits
 
-template <typename T> using is_lattice      = std::is_base_of<LatticeBase,T >;
+template <typename T>
+using is_lattice = std::is_base_of<LatticeBase, T>;
+
+template <typename T>
+using is_lattice_expr = std::is_base_of<LatticeExpressionBase, T>;
 
 template <typename T> using is_lattice_expr = std::is_base_of<LatticeExpressionBase,T >;
 
+//Specialization of getVectorType for lattices
+template<typename T>
+struct getVectorType<Lattice<T> >{
+  typedef typename Lattice<T>::vector_object type;
+};
+ 
 template<class sobj>
 inline sobj eval(const unsigned int ss, const sobj &arg)
 {
   return arg;
 }
 template <class lobj>
-inline const lobj &eval(const unsigned int ss, const Lattice<lobj> &arg)
+inline const lobj &eval(const unsigned int ss, const Lattice<lobj> &arg) {
-{
   return arg._odata[ss];
 }
 
 // handle nodes in syntax tree
 template <typename Op, typename T1>
-auto inline eval(const unsigned int ss, const LatticeUnaryExpression<Op,T1 > &expr) // eval one operand
+auto inline eval(
-  -> decltype(expr.first.func(eval(ss,std::get<0>(expr.second))))
+    const unsigned int ss,
-{
+    const LatticeUnaryExpression<Op, T1> &expr)  // eval one operand
+    -> decltype(expr.first.func(eval(ss, std::get<0>(expr.second)))) {
   return expr.first.func(eval(ss, std::get<0>(expr.second)));
 }
 
 template <typename Op, typename T1, typename T2>
-auto inline eval(const unsigned int ss, const LatticeBinaryExpression<Op,T1,T2> &expr) // eval two operands
+auto inline eval(
-  -> decltype(expr.first.func(eval(ss,std::get<0>(expr.second)),eval(ss,std::get<1>(expr.second))))
+    const unsigned int ss,
-{
+    const LatticeBinaryExpression<Op, T1, T2> &expr)  // eval two operands
-  return expr.first.func(eval(ss,std::get<0>(expr.second)),eval(ss,std::get<1>(expr.second)));
+    -> decltype(expr.first.func(eval(ss, std::get<0>(expr.second)),
+                                eval(ss, std::get<1>(expr.second)))) {
+  return expr.first.func(eval(ss, std::get<0>(expr.second)),
+                         eval(ss, std::get<1>(expr.second)));
 }
 
 template <typename Op, typename T1, typename T2, typename T3>
-auto inline eval(const unsigned int ss, const LatticeTrinaryExpression<Op,T1,T2,T3 > &expr) // eval three operands
+auto inline eval(const unsigned int ss,
-  -> decltype(expr.first.func(eval(ss,std::get<0>(expr.second)),eval(ss,std::get<1>(expr.second)),eval(ss,std::get<2>(expr.second))))
+                 const LatticeTrinaryExpression<Op, T1, T2, T3>
-{
+                     &expr)  // eval three operands
-  return expr.first.func(eval(ss,std::get<0>(expr.second)),eval(ss,std::get<1>(expr.second)),eval(ss,std::get<2>(expr.second)) );
+    -> decltype(expr.first.func(eval(ss, std::get<0>(expr.second)),
+                                eval(ss, std::get<1>(expr.second)),
+                                eval(ss, std::get<2>(expr.second)))) {
+  return expr.first.func(eval(ss, std::get<0>(expr.second)),
+                         eval(ss, std::get<1>(expr.second)),
+                         eval(ss, std::get<2>(expr.second)));
 }
 
 //////////////////////////////////////////////////////////////////////////
-// Obtain the grid from an expression, ensuring conformable. This must follow a tree recursion
+// Obtain the grid from an expression, ensuring conformable. This must follow a
+// tree recursion
 //////////////////////////////////////////////////////////////////////////
-template<class T1, typename std::enable_if<is_lattice<T1>::value, T1>::type * =nullptr >
+template <class T1,
+          typename std::enable_if<is_lattice<T1>::value, T1>::type * = nullptr>
 inline void GridFromExpression(GridBase *&grid, const T1 &lat)  // Lattice leaf
 {
   if (grid) {
@@ -126,35 +146,37 @@ inline void GridFromExpression(GridBase * &grid,const T1& lat)   // Lattice leaf
   }
   grid = lat._grid;
 }
-template<class T1,typename std::enable_if<!is_lattice<T1>::value, T1>::type * = nullptr >
+template <class T1,
-inline void GridFromExpression(GridBase * &grid,const T1& notlat)   // non-lattice leaf
+          typename std::enable_if<!is_lattice<T1>::value, T1>::type * = nullptr>
-{
+inline void GridFromExpression(GridBase *&grid,
-}
+                               const T1 &notlat)  // non-lattice leaf
+{}
 template <typename Op, typename T1>
-inline void GridFromExpression(GridBase * &grid,const LatticeUnaryExpression<Op,T1 > &expr)
+inline void GridFromExpression(GridBase *&grid,
-{
+                               const LatticeUnaryExpression<Op, T1> &expr) {
   GridFromExpression(grid, std::get<0>(expr.second));  // recurse
 }
 
 template <typename Op, typename T1, typename T2>
-inline void GridFromExpression(GridBase * &grid,const LatticeBinaryExpression<Op,T1,T2> &expr) 
+inline void GridFromExpression(
-{
+    GridBase *&grid, const LatticeBinaryExpression<Op, T1, T2> &expr) {
   GridFromExpression(grid, std::get<0>(expr.second));  // recurse
   GridFromExpression(grid, std::get<1>(expr.second));
 }
 template <typename Op, typename T1, typename T2, typename T3>
-inline void GridFromExpression( GridBase * &grid,const LatticeTrinaryExpression<Op,T1,T2,T3 > &expr) 
+inline void GridFromExpression(
-{
+    GridBase *&grid, const LatticeTrinaryExpression<Op, T1, T2, T3> &expr) {
   GridFromExpression(grid, std::get<0>(expr.second));  // recurse
   GridFromExpression(grid, std::get<1>(expr.second));
   GridFromExpression(grid, std::get<2>(expr.second));
 }
 
-
 //////////////////////////////////////////////////////////////////////////
-// Obtain the CB from an expression, ensuring conformable. This must follow a tree recursion
+// Obtain the CB from an expression, ensuring conformable. This must follow a
+// tree recursion
 //////////////////////////////////////////////////////////////////////////
-template<class T1, typename std::enable_if<is_lattice<T1>::value, T1>::type * =nullptr >
+template <class T1,
+          typename std::enable_if<is_lattice<T1>::value, T1>::type * = nullptr>
 inline void CBFromExpression(int &cb, const T1 &lat)  // Lattice leaf
 {
   if ((cb == Odd) || (cb == Even)) {
@@ -163,28 +185,29 @@ inline void CBFromExpression(int &cb,const T1& lat)   // Lattice leaf
   cb = lat.checkerboard;
   //  std::cout<<GridLogMessage<<"Lattice leaf cb "<<cb<<std::endl;
 }
-template<class T1,typename std::enable_if<!is_lattice<T1>::value, T1>::type * = nullptr >
+template <class T1,
+          typename std::enable_if<!is_lattice<T1>::value, T1>::type * = nullptr>
 inline void CBFromExpression(int &cb, const T1 &notlat)  // non-lattice leaf
 {
   //  std::cout<<GridLogMessage<<"Non lattice leaf cb"<<cb<<std::endl;
 }
 template <typename Op, typename T1>
-inline void CBFromExpression(int &cb,const LatticeUnaryExpression<Op,T1 > &expr)
+inline void CBFromExpression(int &cb,
-{
+                             const LatticeUnaryExpression<Op, T1> &expr) {
   CBFromExpression(cb, std::get<0>(expr.second));  // recurse
   //  std::cout<<GridLogMessage<<"Unary node cb "<<cb<<std::endl;
 }
 
 template <typename Op, typename T1, typename T2>
-inline void CBFromExpression(int &cb,const LatticeBinaryExpression<Op,T1,T2> &expr) 
+inline void CBFromExpression(int &cb,
-{
+                             const LatticeBinaryExpression<Op, T1, T2> &expr) {
   CBFromExpression(cb, std::get<0>(expr.second));  // recurse
   CBFromExpression(cb, std::get<1>(expr.second));
   //  std::cout<<GridLogMessage<<"Binary node cb "<<cb<<std::endl;
 }
 template <typename Op, typename T1, typename T2, typename T3>
-inline void CBFromExpression( int &cb,const LatticeTrinaryExpression<Op,T1,T2,T3 > &expr) 
+inline void CBFromExpression(
-{
+    int &cb, const LatticeTrinaryExpression<Op, T1, T2, T3> &expr) {
   CBFromExpression(cb, std::get<0>(expr.second));  // recurse
   CBFromExpression(cb, std::get<1>(expr.second));
   CBFromExpression(cb, std::get<2>(expr.second));
@@ -195,8 +218,8 @@ inline void CBFromExpression( int &cb,const LatticeTrinaryExpression<Op,T1,T2,T3
 // Unary operators and funcs
 ////////////////////////////////////////////
 #define GridUnopClass(name, ret)                                          \
-template <class arg> struct name\
+  template <class arg>                                                    \
-{\
+  struct name {                                                           \
     static auto inline func(const arg a) -> decltype(ret) { return ret; } \
   };
 
@@ -212,11 +235,15 @@ GridUnopClass(UnaryReal,real(a));
 GridUnopClass(UnaryImag, imag(a));
 GridUnopClass(UnaryToReal, toReal(a));
 GridUnopClass(UnaryToComplex, toComplex(a));
+GridUnopClass(UnaryTimesI, timesI(a));
+GridUnopClass(UnaryTimesMinusI, timesMinusI(a));
 GridUnopClass(UnaryAbs, abs(a));
 GridUnopClass(UnarySqrt, sqrt(a));
 GridUnopClass(UnaryRsqrt, rsqrt(a));
 GridUnopClass(UnarySin, sin(a));
 GridUnopClass(UnaryCos, cos(a));
+GridUnopClass(UnaryAsin, asin(a));
+GridUnopClass(UnaryAcos, acos(a));
 GridUnopClass(UnaryLog, log(a));
 GridUnopClass(UnaryExp, exp(a));
 
@@ -225,10 +252,9 @@ GridUnopClass(UnaryExp,exp(a));
 ////////////////////////////////////////////
 #define GridBinOpClass(name, combination)                      \
   template <class left, class right>                           \
-struct name\
+  struct name {                                                \
-{\
+    static auto inline func(const left &lhs, const right &rhs) \
-  static auto inline func(const left &lhs,const right &rhs)-> decltype(combination) const \
+        -> decltype(combination) const {                       \
-    {\
       return combination;                                      \
     }                                                          \
   }
@@ -246,17 +272,18 @@ GridBinOpClass(BinaryOrOr  ,lhs||rhs);
 ////////////////////////////////////////////////////
 #define GridTrinOpClass(name, combination)                                     \
   template <class predicate, class left, class right>                          \
-struct name\
+  struct name {                                                                \
-{\
+    static auto inline func(const predicate &pred, const left &lhs,            \
-  static auto inline func(const predicate &pred,const left &lhs,const right &rhs)-> decltype(combination) const \
+                            const right &rhs) -> decltype(combination) const { \
-    {\
       return combination;                                                      \
     }                                                                          \
   }
 
-GridTrinOpClass(TrinaryWhere,(predicatedWhere<predicate, \
+GridTrinOpClass(
-			       typename std::remove_reference<left>::type, \
+    TrinaryWhere,
-			       typename std::remove_reference<right>::type> (pred,lhs,rhs)));
+    (predicatedWhere<predicate, typename std::remove_reference<left>::type,
+                     typename std::remove_reference<right>::type>(pred, lhs,
+                                                                  rhs)));
 
 ////////////////////////////////////////////
 // Operator syntactical glue
@@ -264,49 +291,66 @@ GridTrinOpClass(TrinaryWhere,(predicatedWhere<predicate, \
 
 #define GRID_UNOP(name) name<decltype(eval(0, arg))>
 #define GRID_BINOP(name) name<decltype(eval(0, lhs)), decltype(eval(0, rhs))>
-#define GRID_TRINOP(name) name<decltype(eval(0, pred)), decltype(eval(0, lhs)), decltype(eval(0, rhs))>
+#define GRID_TRINOP(name) \
+  name<decltype(eval(0, pred)), decltype(eval(0, lhs)), decltype(eval(0, rhs))>
 
 #define GRID_DEF_UNOP(op, name)                                             \
   template <typename T1,                                                    \
-  typename std::enable_if<is_lattice<T1>::value||is_lattice_expr<T1>::value, T1>::type* = nullptr> inline auto op(const T1 &arg) \
+            typename std::enable_if<is_lattice<T1>::value ||                \
-  -> decltype(LatticeUnaryExpression<GRID_UNOP(name),const T1&>(std::make_pair(GRID_UNOP(name)(),std::forward_as_tuple(arg)))) \
+                                        is_lattice_expr<T1>::value,         \
-{ return LatticeUnaryExpression<GRID_UNOP(name), const T1 &>(std::make_pair(GRID_UNOP(name)(),std::forward_as_tuple(arg))); }
+                                    T1>::type * = nullptr>                  \
+  inline auto op(const T1 &arg)                                             \
+      ->decltype(LatticeUnaryExpression<GRID_UNOP(name), const T1 &>(       \
+          std::make_pair(GRID_UNOP(name)(), std::forward_as_tuple(arg)))) { \
+    return LatticeUnaryExpression<GRID_UNOP(name), const T1 &>(             \
+        std::make_pair(GRID_UNOP(name)(), std::forward_as_tuple(arg)));     \
+  }
 
 #define GRID_BINOP_LEFT(op, name)                                             \
   template <typename T1, typename T2,                                         \
-          typename std::enable_if<is_lattice<T1>::value||is_lattice_expr<T1>::value, T1>::type* = nullptr>\
+            typename std::enable_if<is_lattice<T1>::value ||                  \
+                                        is_lattice_expr<T1>::value,           \
+                                    T1>::type * = nullptr>                    \
   inline auto op(const T1 &lhs, const T2 &rhs)                                \
-  -> decltype(LatticeBinaryExpression<GRID_BINOP(name),const T1&,const T2 &>(std::make_pair(GRID_BINOP(name)(),\
+      ->decltype(                                                             \
-											    std::forward_as_tuple(lhs, rhs)))) \
+          LatticeBinaryExpression<GRID_BINOP(name), const T1 &, const T2 &>(  \
-{\
+              std::make_pair(GRID_BINOP(name)(),                              \
- return LatticeBinaryExpression<GRID_BINOP(name), const T1 &, const T2 &>(std::make_pair(GRID_BINOP(name)(),\
+                             std::forward_as_tuple(lhs, rhs)))) {             \
-									  std::forward_as_tuple(lhs, rhs))); \
+    return LatticeBinaryExpression<GRID_BINOP(name), const T1 &, const T2 &>( \
+        std::make_pair(GRID_BINOP(name)(), std::forward_as_tuple(lhs, rhs))); \
   }
 
 #define GRID_BINOP_RIGHT(op, name)                                            \
   template <typename T1, typename T2,                                         \
-           typename std::enable_if<!is_lattice<T1>::value && !is_lattice_expr<T1>::value, T1>::type* = nullptr,\
+            typename std::enable_if<!is_lattice<T1>::value &&                 \
-           typename std::enable_if< is_lattice<T2>::value ||  is_lattice_expr<T2>::value, T2>::type* = nullptr> \
+                                        !is_lattice_expr<T1>::value,          \
+                                    T1>::type * = nullptr,                    \
+            typename std::enable_if<is_lattice<T2>::value ||                  \
+                                        is_lattice_expr<T2>::value,           \
+                                    T2>::type * = nullptr>                    \
   inline auto op(const T1 &lhs, const T2 &rhs)                                \
-  -> decltype(LatticeBinaryExpression<GRID_BINOP(name),const T1&,const T2 &>(std::make_pair(GRID_BINOP(name)(),\
+      ->decltype(                                                             \
-											    std::forward_as_tuple(lhs, rhs)))) \
+          LatticeBinaryExpression<GRID_BINOP(name), const T1 &, const T2 &>(  \
-{\
+              std::make_pair(GRID_BINOP(name)(),                              \
- return LatticeBinaryExpression<GRID_BINOP(name), const T1 &, const T2 &>(std::make_pair(GRID_BINOP(name)(),\
+                             std::forward_as_tuple(lhs, rhs)))) {             \
-								          std::forward_as_tuple(lhs, rhs))); \
+    return LatticeBinaryExpression<GRID_BINOP(name), const T1 &, const T2 &>( \
+        std::make_pair(GRID_BINOP(name)(), std::forward_as_tuple(lhs, rhs))); \
   }
 
 #define GRID_DEF_BINOP(op, name) \
   GRID_BINOP_LEFT(op, name);     \
   GRID_BINOP_RIGHT(op, name);
 
-
 #define GRID_DEF_TRINOP(op, name)                                              \
-template <typename T1,typename T2,typename T3> inline auto op(const T1 &pred,const T2&lhs,const T3 &rhs) \
+  template <typename T1, typename T2, typename T3>                             \
-  -> decltype(LatticeTrinaryExpression<GRID_TRINOP(name),const T1&,const T2 &,const T3&>(std::make_pair(GRID_TRINOP(name)(),\
+  inline auto op(const T1 &pred, const T2 &lhs, const T3 &rhs)                 \
-										   std::forward_as_tuple(pred,lhs,rhs)))) \
+      ->decltype(                                                              \
-{\
+          LatticeTrinaryExpression<GRID_TRINOP(name), const T1 &, const T2 &,  \
-  return LatticeTrinaryExpression<GRID_TRINOP(name), const T1 &, const T2 &,const T3&>(std::make_pair(GRID_TRINOP(name)(), \
+                                   const T3 &>(std::make_pair(                 \
-										 std::forward_as_tuple(pred,lhs, rhs))); \
+              GRID_TRINOP(name)(), std::forward_as_tuple(pred, lhs, rhs)))) {  \
+    return LatticeTrinaryExpression<GRID_TRINOP(name), const T1 &, const T2 &, \
+                                    const T3 &>(std::make_pair(                \
+        GRID_TRINOP(name)(), std::forward_as_tuple(pred, lhs, rhs)));          \
   }
 ////////////////////////
 // Operator definitions
@@ -325,11 +369,16 @@ GRID_DEF_UNOP(real,UnaryReal);
 GRID_DEF_UNOP(imag, UnaryImag);
 GRID_DEF_UNOP(toReal, UnaryToReal);
 GRID_DEF_UNOP(toComplex, UnaryToComplex);
-GRID_DEF_UNOP(abs  ,UnaryAbs); //abs overloaded in cmath C++98; DON'T do the abs-fabs-dabs-labs thing
+GRID_DEF_UNOP(timesI, UnaryTimesI);
+GRID_DEF_UNOP(timesMinusI, UnaryTimesMinusI);
+GRID_DEF_UNOP(abs, UnaryAbs);  // abs overloaded in cmath C++98; DON'T do the
+                               // abs-fabs-dabs-labs thing
 GRID_DEF_UNOP(sqrt, UnarySqrt);
 GRID_DEF_UNOP(rsqrt, UnaryRsqrt);
 GRID_DEF_UNOP(sin, UnarySin);
 GRID_DEF_UNOP(cos, UnaryCos);
+GRID_DEF_UNOP(asin, UnaryAsin);
+GRID_DEF_UNOP(acos, UnaryAcos);
 GRID_DEF_UNOP(log, UnaryLog);
 GRID_DEF_UNOP(exp, UnaryExp);
 
@@ -349,29 +398,29 @@ GRID_DEF_TRINOP(where,TrinaryWhere);
 /////////////////////////////////////////////////////////////
 template <class Op, class T1>
 auto closure(const LatticeUnaryExpression<Op, T1> &expr)
-  -> Lattice<decltype(expr.first.func(eval(0,std::get<0>(expr.second))))>
+    -> Lattice<decltype(expr.first.func(eval(0, std::get<0>(expr.second))))> {
-{
+  Lattice<decltype(expr.first.func(eval(0, std::get<0>(expr.second))))> ret(
-  Lattice<decltype(expr.first.func(eval(0,std::get<0>(expr.second))))> ret(expr);
+      expr);
   return ret;
 }
 template <class Op, class T1, class T2>
 auto closure(const LatticeBinaryExpression<Op, T1, T2> &expr)
     -> Lattice<decltype(expr.first.func(eval(0, std::get<0>(expr.second)),
-				      eval(0,std::get<1>(expr.second))))>
+                                        eval(0, std::get<1>(expr.second))))> {
-{
   Lattice<decltype(expr.first.func(eval(0, std::get<0>(expr.second)),
-				   eval(0,std::get<1>(expr.second))))> ret(expr);
+                                   eval(0, std::get<1>(expr.second))))>
+      ret(expr);
   return ret;
 }
 template <class Op, class T1, class T2, class T3>
 auto closure(const LatticeTrinaryExpression<Op, T1, T2, T3> &expr)
     -> Lattice<decltype(expr.first.func(eval(0, std::get<0>(expr.second)),
                                         eval(0, std::get<1>(expr.second)),
-				      eval(0,std::get<2>(expr.second))))>
+                                        eval(0, std::get<2>(expr.second))))> {
-{
   Lattice<decltype(expr.first.func(eval(0, std::get<0>(expr.second)),
                                    eval(0, std::get<1>(expr.second)),
-				   eval(0,std::get<2>(expr.second))))> ret(expr);
+                                   eval(0, std::get<2>(expr.second))))>
+      ret(expr);
   return ret;
 }
 
@@ -382,7 +431,6 @@ template<class Op,class T1, class T2, class T3>
 #undef GRID_DEF_UNOP
 #undef GRID_DEF_BINOP
 #undef GRID_DEF_TRINOP
-
 }
 
 #if 0

lib/lattice/Lattice_base.h

@@ -24,7 +24,8 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 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
+See the full license in the file "LICENSE" in the top level distribution
+directory
 *************************************************************************************/
 /*  END LEGAL */
 #ifndef GRID_LATTICE_BASE_H
@@ -64,9 +65,6 @@ public:
     
 class LatticeExpressionBase {};
 
-template<class T> using Vector = std::vector<T,alignedAllocator<T> >;               // Aligned allocator??
-template<class T> using Matrix = std::vector<std::vector<T,alignedAllocator<T> > >; // Aligned allocator??
-
 template <typename Op, typename T1>                           
 class LatticeUnaryExpression  : public std::pair<Op,std::tuple<T1> > , public LatticeExpressionBase {
  public:
@@ -101,6 +99,7 @@ public:
     int begin(void) { return 0;};
     int end(void)   { return _odata.size(); }
     vobj & operator[](int i) { return _odata[i]; };
+    const vobj & operator[](int i) const { return _odata[i]; };
 
 public:
     typedef typename vobj::scalar_type scalar_type;
@@ -255,6 +254,18 @@ PARALLEL_FOR_LOOP
         checkerboard=0;
     }
 
+    Lattice(const Lattice& r){ // copy constructor
+    	_grid = r._grid;
+    	checkerboard = r.checkerboard;
+    	_odata.resize(_grid->oSites());// essential
+  		PARALLEL_FOR_LOOP
+        for(int ss=0;ss<_grid->oSites();ss++){
+            _odata[ss]=r._odata[ss];
+        }  	
+    }
+
+
+
     virtual ~Lattice(void) = default;
     
     template<class sobj> strong_inline Lattice<vobj> & operator = (const sobj & r){
@@ -267,7 +278,7 @@ PARALLEL_FOR_LOOP
     template<class robj> strong_inline Lattice<vobj> & operator = (const Lattice<robj> & r){
       this->checkerboard = r.checkerboard;
       conformable(*this,r);
-      std::cout<<GridLogMessage<<"Lattice operator ="<<std::endl;
+      
 PARALLEL_FOR_LOOP
         for(int ss=0;ss<_grid->oSites();ss++){
             this->_odata[ss]=r._odata[ss];
@@ -324,27 +335,27 @@ PARALLEL_FOR_LOOP
 
 
 
-#include <lattice/Lattice_conformable.h>
+#include "Lattice_conformable.h"
 #define GRID_LATTICE_EXPRESSION_TEMPLATES
 #ifdef  GRID_LATTICE_EXPRESSION_TEMPLATES
-#include <lattice/Lattice_ET.h>
+#include "Lattice_ET.h"
 #else 
-#include <lattice/Lattice_overload.h>
+#include "Lattice_overload.h"
 #endif
-#include <lattice/Lattice_arith.h>
+#include "Lattice_arith.h"
-#include <lattice/Lattice_trace.h>
+#include "Lattice_trace.h"
-#include <lattice/Lattice_transpose.h>
+#include "Lattice_transpose.h"
-#include <lattice/Lattice_local.h>
+#include "Lattice_local.h"
-#include <lattice/Lattice_reduction.h>
+#include "Lattice_reduction.h"
-#include <lattice/Lattice_peekpoke.h>
+#include "Lattice_peekpoke.h"
-#include <lattice/Lattice_reality.h>
+#include "Lattice_reality.h"
-#include <lattice/Lattice_comparison_utils.h>
+#include "Lattice_comparison_utils.h"
-#include <lattice/Lattice_comparison.h>
+#include "Lattice_comparison.h"
-#include <lattice/Lattice_coordinate.h>
+#include "Lattice_coordinate.h"
-#include <lattice/Lattice_where.h>
+#include "Lattice_where.h"
-#include <lattice/Lattice_rng.h>
+#include "Lattice_rng.h"
-#include <lattice/Lattice_unary.h>
+#include "Lattice_unary.h"
-#include <lattice/Lattice_transfer.h>
+#include "Lattice_transfer.h"
 
 
 #endif

lib/lattice/Lattice_peekpoke.h

@@ -164,15 +164,17 @@ PARALLEL_FOR_LOOP
       assert( l.checkerboard== l._grid->CheckerBoard(site));
       assert( sizeof(sobj)*Nsimd == sizeof(vobj));
 
+      static const int words=sizeof(vobj)/sizeof(vector_type);
       int odx,idx;
       idx= grid->iIndex(site);
       odx= grid->oIndex(site);
 
-      std::vector<sobj> buf(Nsimd);
+      scalar_type * vp = (scalar_type *)&l._odata[odx];
+      scalar_type * pt = (scalar_type *)&s;
       
-      extract(l._odata[odx],buf);
+      for(int w=0;w<words;w++){
-      
+        pt[w] = vp[idx+w*Nsimd];
-      s = buf[idx];
+      }
       
       return;
     };
@@ -190,18 +192,17 @@ PARALLEL_FOR_LOOP
       assert( l.checkerboard== l._grid->CheckerBoard(site));
       assert( sizeof(sobj)*Nsimd == sizeof(vobj));
 
+      static const int words=sizeof(vobj)/sizeof(vector_type);
       int odx,idx;
       idx= grid->iIndex(site);
       odx= grid->oIndex(site);
 
-      std::vector<sobj> buf(Nsimd);
+      scalar_type * vp = (scalar_type *)&l._odata[odx];
+      scalar_type * pt = (scalar_type *)&s;
       
-      // extract-modify-merge cycle is easiest way and this is not perf critical
+      for(int w=0;w<words;w++){
-      extract(l._odata[odx],buf);
+        vp[idx+w*Nsimd] = pt[w];
-      
+      }
-      buf[idx] = s;
-
-      merge(l._odata[odx],buf);
 
       return;
     };

lib/lattice/Lattice_reduction.h

@@ -40,7 +40,7 @@ namespace Grid {
     ////////////////////////////////////////////////////////////////////////////////////////////////////
   template<class vobj> inline RealD norm2(const Lattice<vobj> &arg){
     ComplexD nrm = innerProduct(arg,arg);
-    return real(nrm); 
+    return std::real(nrm); 
   }
 
     template<class vobj>

lib/lattice/Lattice_rng.h

@@ -297,8 +297,9 @@ namespace Grid {
 
 	int l_idx=generator_idx(o_idx,i_idx);
 
-	std::vector<int> site_seeds(4);
+	const int num_rand_seed=16;
-	for(int i=0;i<4;i++){
+	std::vector<int> site_seeds(num_rand_seed);
+	for(int i=0;i<site_seeds.size();i++){
 	  site_seeds[i]= ui(pseeder);
 	}
 

lib/lattice/Lattice_transfer.h

@@ -349,7 +349,7 @@ void localConvert(const Lattice<vobj> &in,Lattice<vvobj> &out)
     assert(ig->_ldimensions[d] == og->_ldimensions[d]);
   }
 
-PARALLEL_FOR_LOOP
+  //PARALLEL_FOR_LOOP
   for(int idx=0;idx<ig->lSites();idx++){
     std::vector<int> lcoor(ni);
     ig->LocalIndexToLocalCoor(idx,lcoor);
@@ -386,7 +386,7 @@ void InsertSlice(Lattice<vobj> &lowDim,Lattice<vobj> & higherDim,int slice, int
   }
 
   // the above should guarantee that the operations are local
-PARALLEL_FOR_LOOP
+  //PARALLEL_FOR_LOOP
   for(int idx=0;idx<lg->lSites();idx++){
     std::vector<int> lcoor(nl);
     std::vector<int> hcoor(nh);
@@ -428,7 +428,7 @@ void ExtractSlice(Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slice, in
     }
   }
   // the above should guarantee that the operations are local
-PARALLEL_FOR_LOOP
+  //PARALLEL_FOR_LOOP
   for(int idx=0;idx<lg->lSites();idx++){
     std::vector<int> lcoor(nl);
     std::vector<int> hcoor(nh);
@@ -446,6 +446,79 @@ PARALLEL_FOR_LOOP
 
 }
 
+
+template<class vobj>
+void InsertSliceLocal(Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slice_lo,int slice_hi, int orthog)
+{
+  typedef typename vobj::scalar_object sobj;
+  sobj s;
+
+  GridBase *lg = lowDim._grid;
+  GridBase *hg = higherDim._grid;
+  int nl = lg->_ndimension;
+  int nh = hg->_ndimension;
+
+  assert(nl == nh);
+  assert(orthog<nh);
+  assert(orthog>=0);
+
+  for(int d=0;d<nh;d++){
+    assert(lg->_processors[d]  == hg->_processors[d]);
+    assert(lg->_ldimensions[d] == hg->_ldimensions[d]);
+  }
+
+  // the above should guarantee that the operations are local
+  //PARALLEL_FOR_LOOP
+  for(int idx=0;idx<lg->lSites();idx++){
+    std::vector<int> lcoor(nl);
+    std::vector<int> hcoor(nh);
+    lg->LocalIndexToLocalCoor(idx,lcoor);
+    if( lcoor[orthog] == slice_lo ) { 
+      hcoor=lcoor;
+      hcoor[orthog] = slice_hi;
+      peekLocalSite(s,lowDim,lcoor);
+      pokeLocalSite(s,higherDim,hcoor);
+    }
+  }
+}
+
+
+template<class vobj>
+void ExtractSliceLocal(Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slice_lo,int slice_hi, int orthog)
+{
+  typedef typename vobj::scalar_object sobj;
+  sobj s;
+
+  GridBase *lg = lowDim._grid;
+  GridBase *hg = higherDim._grid;
+  int nl = lg->_ndimension;
+  int nh = hg->_ndimension;
+
+  assert(nl == nh);
+  assert(orthog<nh);
+  assert(orthog>=0);
+
+  for(int d=0;d<nh;d++){
+    assert(lg->_processors[d]  == hg->_processors[d]);
+    assert(lg->_ldimensions[d] == hg->_ldimensions[d]);
+  }
+
+  // the above should guarantee that the operations are local
+  //PARALLEL_FOR_LOOP
+  for(int idx=0;idx<lg->lSites();idx++){
+    std::vector<int> lcoor(nl);
+    std::vector<int> hcoor(nh);
+    lg->LocalIndexToLocalCoor(idx,lcoor);
+    if( lcoor[orthog] == slice_lo ) { 
+      hcoor=lcoor;
+      hcoor[orthog] = slice_hi;
+      peekLocalSite(s,higherDim,hcoor);
+      pokeLocalSite(s,lowDim,lcoor);
+    }
+  }
+}
+
+
 template<class vobj>
 void Replicate(Lattice<vobj> &coarse,Lattice<vobj> & fine)
 {
@@ -482,6 +555,96 @@ void Replicate(Lattice<vobj> &coarse,Lattice<vobj> & fine)
 
 }
 
+//Copy SIMD-vectorized lattice to array of scalar objects in lexicographic order
+template<typename vobj, typename sobj>
+typename std::enable_if<isSIMDvectorized<vobj>::value && !isSIMDvectorized<sobj>::value, void>::type unvectorizeToLexOrdArray(std::vector<sobj> &out, const Lattice<vobj> &in){
+  typedef typename vobj::vector_type vtype;
+  
+  GridBase* in_grid = in._grid;
+  out.resize(in_grid->lSites());
+  
+  int ndim = in_grid->Nd();
+  int in_nsimd = vtype::Nsimd();
+
+  std::vector<std::vector<int> > in_icoor(in_nsimd);
+      
+  for(int lane=0; lane < in_nsimd; lane++){
+    in_icoor[lane].resize(ndim);
+    in_grid->iCoorFromIindex(in_icoor[lane], lane);
+  }
+  
+PARALLEL_FOR_LOOP
+  for(int in_oidx = 0; in_oidx < in_grid->oSites(); in_oidx++){ //loop over outer index
+    //Assemble vector of pointers to output elements
+    std::vector<sobj*> out_ptrs(in_nsimd);
+
+    std::vector<int> in_ocoor(ndim);
+    in_grid->oCoorFromOindex(in_ocoor, in_oidx);
+
+    std::vector<int> lcoor(in_grid->Nd());
+      
+    for(int lane=0; lane < in_nsimd; lane++){
+      for(int mu=0;mu<ndim;mu++)
+	lcoor[mu] = in_ocoor[mu] + in_grid->_rdimensions[mu]*in_icoor[lane][mu];
+
+      int lex;
+      Lexicographic::IndexFromCoor(lcoor, lex, in_grid->_ldimensions);
+      out_ptrs[lane] = &out[lex];
+    }
+    
+    //Unpack into those ptrs
+    const vobj & in_vobj = in._odata[in_oidx];
+    extract1(in_vobj, out_ptrs, 0);
+  }
+}
+
+//Convert a Lattice from one precision to another
+template<class VobjOut, class VobjIn>
+void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in){
+  assert(out._grid->Nd() == in._grid->Nd());
+  out.checkerboard = in.checkerboard;
+  GridBase *in_grid=in._grid;
+  GridBase *out_grid = out._grid;
+
+  typedef typename VobjOut::scalar_object SobjOut;
+  typedef typename VobjIn::scalar_object SobjIn;
+
+  int ndim = out._grid->Nd();
+  int out_nsimd = out_grid->Nsimd();
+    
+  std::vector<std::vector<int> > out_icoor(out_nsimd);
+      
+  for(int lane=0; lane < out_nsimd; lane++){
+    out_icoor[lane].resize(ndim);
+    out_grid->iCoorFromIindex(out_icoor[lane], lane);
+  }
+        
+  std::vector<SobjOut> in_slex_conv(in_grid->lSites());
+  unvectorizeToLexOrdArray(in_slex_conv, in);
+    
+  PARALLEL_FOR_LOOP
+  for(int out_oidx=0;out_oidx<out_grid->oSites();out_oidx++){
+    std::vector<int> out_ocoor(ndim);
+    out_grid->oCoorFromOindex(out_ocoor, out_oidx);
+
+    std::vector<SobjOut*> ptrs(out_nsimd);      
+
+    std::vector<int> lcoor(out_grid->Nd());
+      
+    for(int lane=0; lane < out_nsimd; lane++){
+      for(int mu=0;mu<ndim;mu++)
+	lcoor[mu] = out_ocoor[mu] + out_grid->_rdimensions[mu]*out_icoor[lane][mu];
+	
+      int llex; Lexicographic::IndexFromCoor(lcoor, llex, out_grid->_ldimensions);
+      ptrs[lane] = &in_slex_conv[llex];
+    }
+    merge(out._odata[out_oidx], ptrs, 0);
+  }
+}
+
+
+  
+
  
 }
 #endif

lib/pugixml/pugixml.h

@@ -17,7 +17,7 @@
 #endif
 
 // Include user configuration file (this can define various configuration macros)
-#include <pugixml/pugiconfig.hpp>
+#include "pugiconfig.hpp"
 
 #ifndef HEADER_PUGIXML_HPP
 #define HEADER_PUGIXML_HPP

lib/qcd/QCD.h

@@ -55,10 +55,19 @@ namespace QCD {
     //////////////////////////////////////////////////////////////////////////////
     // QCD iMatrix types
     // Index conventions:                            Lorentz x Spin x Colour
+    // note: static const int or constexpr will work for type deductions
+    //       with the intel compiler (up to version 17)
     //////////////////////////////////////////////////////////////////////////////
-    static const int ColourIndex = 2;
+    #define ColourIndex  2
-    static const int SpinIndex   = 1;
+    #define SpinIndex    1
-    static const int LorentzIndex= 0;
+    #define LorentzIndex 0
+
+  
+    // Also should make these a named enum type
+    static const int DaggerNo=0;
+    static const int DaggerYes=1;
+    static const int InverseNo=0;
+    static const int InverseYes=1;
 
     // Useful traits is this a spin index
     //typename std::enable_if<matchGridTensorIndex<iVector<vtype,Ns>,SpinorIndex>::value,iVector<vtype,Ns> >::type *SFINAE;
@@ -484,16 +493,27 @@ namespace QCD {
 }   //namespace QCD
 } // Grid
 
-#include <qcd/utils/SpaceTimeGrid.h>
+
-#include <qcd/spin/Dirac.h>
+#include <Grid/qcd/utils/SpaceTimeGrid.h>
-#include <qcd/spin/TwoSpinor.h>
+#include <Grid/qcd/spin/Dirac.h>
-#include <qcd/utils/LinalgUtils.h>
+#include <Grid/qcd/spin/TwoSpinor.h>
-#include <qcd/utils/CovariantCshift.h>
+#include <Grid/qcd/utils/LinalgUtils.h>
-#include <qcd/utils/SUn.h>
+#include <Grid/qcd/utils/CovariantCshift.h>
-#include <qcd/action/Actions.h>
+
-#include <qcd/hmc/integrators/Integrator.h>
+// Include representations 	
-#include <qcd/hmc/integrators/Integrator_algorithm.h>
+#include <Grid/qcd/utils/SUn.h>
-#include <qcd/hmc/HMC.h>
+#include <Grid/qcd/utils/SUnAdjoint.h>
+#include <Grid/qcd/utils/SUnTwoIndex.h>
+#include <Grid/qcd/representations/hmc_types.h>
+
+#include <Grid/qcd/action/Actions.h>
+
+#include <Grid/qcd/smearing/Smearing.h>
+
+#include <Grid/qcd/hmc/integrators/Integrator.h>
+#include <Grid/qcd/hmc/integrators/Integrator_algorithm.h>
+#include <Grid/qcd/hmc/HMC.h>
+
 
 
 #endif

lib/qcd/action/ActionBase.h

@@ -23,7 +23,8 @@ Author: neo <cossu@post.kek.jp>
 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
+See the full license in the file "LICENSE" in the top level distribution
+directory
 *************************************************************************************/
 /*  END LEGAL */
 #ifndef QCD_ACTION_BASE
@@ -33,54 +34,120 @@ namespace QCD{
 
 template <class GaugeField>
 class Action {
-
  public:
+  bool is_smeared = false;
   // Boundary conditions? // Heatbath?
-  virtual void  refresh(const GaugeField &U, GridParallelRNG& pRNG) = 0;// refresh pseudofermions
+  virtual void refresh(const GaugeField& U,
+                       GridParallelRNG& pRNG) = 0;  // refresh pseudofermions
   virtual RealD S(const GaugeField& U) = 0;         // evaluate the action
-  virtual void  deriv(const GaugeField &U,GaugeField & dSdU )     = 0;  // evaluate the action derivative
+  virtual void deriv(const GaugeField& U,
+                     GaugeField& dSdU) = 0;  // evaluate the action derivative
   virtual ~Action(){};
 };
 
-// Could derive PseudoFermion action with a PF field, FermionField, and a Grid; implement refresh
+// Indexing of tuple types
+template <class T, class Tuple>
+struct Index;
+
+template <class T, class... Types>
+struct Index<T, std::tuple<T, Types...>> {
+  static const std::size_t value = 0;
+};
+
+template <class T, class U, class... Types>
+struct Index<T, std::tuple<U, Types...>> {
+  static const std::size_t value = 1 + Index<T, std::tuple<Types...>>::value;
+};
+
 /*
-template<class GaugeField, class FermionField>
+template <class GaugeField>
-class PseudoFermionAction : public Action<GaugeField> {
+struct ActionLevel {
  public:
-  FermionField Phi;
+  typedef Action<GaugeField>*
-  GridParallelRNG &pRNG;
+      ActPtr;  // now force the same colours as the rest of the code
-  GridBase &Grid;
 
-  PseudoFermionAction(GridBase &_Grid,GridParallelRNG &_pRNG) : Grid(_Grid), Phi(&_Grid), pRNG(_pRNG) {
+  //Add supported representations here
-  };
 
-  virtual void refresh(const GaugeField &gauge) {
-    gaussian(Phi,pRNG);
-  };
 
-};
+  unsigned int multiplier;
-*/
-
-template<class GaugeField> struct ActionLevel{
-public:
-   
-  typedef Action<GaugeField>*  ActPtr; // now force the same colours as the rest of the code
-
-  int multiplier;
 
   std::vector<ActPtr> actions;
 
-  ActionLevel(int mul = 1) : multiplier(mul) {
+  ActionLevel(unsigned int mul = 1) : actions(0), multiplier(mul) {
-    assert (mul > 0);
+    assert(mul >= 1);
   };
 
-  void push_back(ActPtr ptr){
+  void push_back(ActPtr ptr) { actions.push_back(ptr); }
-    actions.push_back(ptr);
+};
+*/
+
+template <class GaugeField, class Repr = NoHirep >
+struct ActionLevel {
+ public:
+  unsigned int multiplier; 
+
+  // Fundamental repr actions separated because of the smearing
+  typedef Action<GaugeField>* ActPtr;
+
+  // construct a tuple of vectors of the actions for the corresponding higher
+  // representation fields
+  typedef typename AccessTypes<Action, Repr>::VectorCollection action_collection;
+  action_collection actions_hirep;
+  typedef typename  AccessTypes<Action, Repr>::FieldTypeCollection action_hirep_types;
+
+  std::vector<ActPtr>& actions;
+
+  // Temporary conversion between ActionLevel and ActionLevelHirep
+  //ActionLevelHirep(ActionLevel<GaugeField>& AL ):actions(AL.actions), multiplier(AL.multiplier){}
+
+  ActionLevel(unsigned int mul = 1) : actions(std::get<0>(actions_hirep)), multiplier(mul) {
+    // initialize the hirep vectors to zero.
+    //apply(this->resize, actions_hirep, 0); //need a working resize
+    assert(mul >= 1);
+  };
+
+  //void push_back(ActPtr ptr) { actions.push_back(ptr); }
+
+
+
+  template < class Field >
+  void push_back(Action<Field>* ptr) {
+    // insert only in the correct vector
+    std::get< Index < Field, action_hirep_types>::value >(actions_hirep).push_back(ptr);
+  };
+
+ 
+
+  template < class ActPtr>
+  static void resize(ActPtr ap, unsigned int n){
+    ap->resize(n);
+
   }
+
+  //template <std::size_t I>
+  //auto getRepresentation(Repr& R)->decltype(std::get<I>(R).U)  {return std::get<I>(R).U;}
+
+  // Loop on tuple for a callable function
+  template <std::size_t I = 1, typename Callable, typename ...Args>
+  inline typename std::enable_if<I == std::tuple_size<action_collection>::value, void>::type apply(
+      Callable, Repr& R,Args&...) const {}
+
+  template <std::size_t I = 1, typename Callable, typename ...Args>
+  inline typename std::enable_if<I < std::tuple_size<action_collection>::value, void>::type apply(
+      Callable fn, Repr& R, Args&... arguments) const {
+    fn(std::get<I>(actions_hirep), std::get<I>(R.rep), arguments...);
+    apply<I + 1>(fn, R, arguments...);
+  }  
+
 };
 
-template<class GaugeField> using ActionSet = std::vector<ActionLevel< GaugeField > >;
 
+//template <class GaugeField>
+//using ActionSet = std::vector<ActionLevel<GaugeField> >;
 
-}}
+template <class GaugeField, class R>
+using ActionSet = std::vector<ActionLevel<GaugeField, R> >;
+
+}
+}
 #endif

lib/qcd/action/Actions.h

@@ -40,25 +40,25 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 ////////////////////////////////////////////
 // Abstract base interface
 ////////////////////////////////////////////
-#include <qcd/action/ActionBase.h>
+#include <Grid/qcd/action/ActionBase.h>
-#include <qcd/action/ActionParams.h>
+#include <Grid/qcd/action/ActionParams.h>
 
 ////////////////////////////////////////////
 // Utility functions
 ////////////////////////////////////////////
-#include <qcd/action/gauge/GaugeImpl.h>
+#include <Grid/qcd/action/gauge/GaugeImpl.h>
-#include <qcd/utils/WilsonLoops.h>
+#include <Grid/qcd/utils/WilsonLoops.h>
 
-#include <qcd/action/fermion/WilsonCompressor.h>     //used by all wilson type fermions
+#include <Grid/qcd/action/fermion/WilsonCompressor.h>     //used by all wilson type fermions
-#include <qcd/action/fermion/FermionOperatorImpl.h>
+#include <Grid/qcd/action/fermion/FermionOperatorImpl.h>
-#include <qcd/action/fermion/FermionOperator.h>
+#include <Grid/qcd/action/fermion/FermionOperator.h>
-#include <qcd/action/fermion/WilsonKernels.h>        //used by all wilson type fermions
+#include <Grid/qcd/action/fermion/WilsonKernels.h>        //used by all wilson type fermions
 
 ////////////////////////////////////////////
 // Gauge Actions
 ////////////////////////////////////////////
-#include <qcd/action/gauge/WilsonGaugeAction.h>
+#include <Grid/qcd/action/gauge/WilsonGaugeAction.h>
-#include <qcd/action/gauge/PlaqPlusRectangleAction.h>
+#include <Grid/qcd/action/gauge/PlaqPlusRectangleAction.h>
 
 namespace Grid {
 namespace QCD {
@@ -107,41 +107,64 @@ typedef SymanzikGaugeAction<ConjugateGimplD>        ConjugateSymanzikGaugeAction
 // for EVERY .cc file. This define centralises the list and restores global push of impl cases
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 
-#define FermOpTemplateInstantiate(A) \
+
+#define FermOp4dVecTemplateInstantiate(A) \
   template class A<WilsonImplF>;		\
   template class A<WilsonImplD>;		\
+  template class A<ZWilsonImplF>;		\
+  template class A<ZWilsonImplD>;		\
   template class A<GparityWilsonImplF>;		\
   template class A<GparityWilsonImplD>;		
 
+#define AdjointFermOpTemplateInstantiate(A) \
+  template class A<WilsonAdjImplF>; \
+  template class A<WilsonAdjImplD>; 
+
+#define TwoIndexFermOpTemplateInstantiate(A) \
+  template class A<WilsonTwoIndexSymmetricImplF>; \
+  template class A<WilsonTwoIndexSymmetricImplD>; 
+
+#define FermOp5dVecTemplateInstantiate(A) \
+  template class A<DomainWallVec5dImplF>;	\
+  template class A<DomainWallVec5dImplD>;	\
+  template class A<ZDomainWallVec5dImplF>;	\
+  template class A<ZDomainWallVec5dImplD>;	
+
+#define FermOpTemplateInstantiate(A) \
+ FermOp4dVecTemplateInstantiate(A) \
+ FermOp5dVecTemplateInstantiate(A) 
+
+
 #define GparityFermOpTemplateInstantiate(A) 
 
 ////////////////////////////////////////////
 // Fermion operators / actions
 ////////////////////////////////////////////
 
-#include <qcd/action/fermion/WilsonFermion.h>       // 4d wilson like
+#include <Grid/qcd/action/fermion/WilsonFermion.h>       // 4d wilson like
-#include <qcd/action/fermion/WilsonTMFermion.h>       // 4d wilson like
+#include <Grid/qcd/action/fermion/WilsonTMFermion.h>       // 4d wilson like
-#include <qcd/action/fermion/WilsonFermion5D.h>     // 5d base used by all 5d overlap types
+#include <Grid/qcd/action/fermion/WilsonFermion5D.h>     // 5d base used by all 5d overlap types
 
-//#include <qcd/action/fermion/CloverFermion.h>
+//#include <Grid/qcd/action/fermion/CloverFermion.h>
 
-#include <qcd/action/fermion/CayleyFermion5D.h>     // Cayley types
+#include <Grid/qcd/action/fermion/CayleyFermion5D.h>     // Cayley types
-#include <qcd/action/fermion/DomainWallFermion.h>
+#include <Grid/qcd/action/fermion/DomainWallFermion.h>
-#include <qcd/action/fermion/DomainWallFermion.h>
+#include <Grid/qcd/action/fermion/DomainWallFermion.h>
-#include <qcd/action/fermion/MobiusFermion.h>
+#include <Grid/qcd/action/fermion/MobiusFermion.h>
-#include <qcd/action/fermion/ScaledShamirFermion.h>
+#include <Grid/qcd/action/fermion/ZMobiusFermion.h>
-#include <qcd/action/fermion/MobiusZolotarevFermion.h>
+#include <Grid/qcd/action/fermion/ScaledShamirFermion.h>
-#include <qcd/action/fermion/ShamirZolotarevFermion.h>
+#include <Grid/qcd/action/fermion/MobiusZolotarevFermion.h>
-#include <qcd/action/fermion/OverlapWilsonCayleyTanhFermion.h>
+#include <Grid/qcd/action/fermion/ShamirZolotarevFermion.h>
-#include <qcd/action/fermion/OverlapWilsonCayleyZolotarevFermion.h>
+#include <Grid/qcd/action/fermion/OverlapWilsonCayleyTanhFermion.h>
+#include <Grid/qcd/action/fermion/OverlapWilsonCayleyZolotarevFermion.h>
 
-#include <qcd/action/fermion/ContinuedFractionFermion5D.h>               // Continued fraction
+#include <Grid/qcd/action/fermion/ContinuedFractionFermion5D.h>               // Continued fraction
-#include <qcd/action/fermion/OverlapWilsonContfracTanhFermion.h>
+#include <Grid/qcd/action/fermion/OverlapWilsonContfracTanhFermion.h>
-#include <qcd/action/fermion/OverlapWilsonContfracZolotarevFermion.h>
+#include <Grid/qcd/action/fermion/OverlapWilsonContfracZolotarevFermion.h>
 
-#include <qcd/action/fermion/PartialFractionFermion5D.h>                 // Partial fraction
+#include <Grid/qcd/action/fermion/PartialFractionFermion5D.h>                 // Partial fraction
-#include <qcd/action/fermion/OverlapWilsonPartialFractionTanhFermion.h>
+#include <Grid/qcd/action/fermion/OverlapWilsonPartialFractionTanhFermion.h>
-#include <qcd/action/fermion/OverlapWilsonPartialFractionZolotarevFermion.h>
+#include <Grid/qcd/action/fermion/OverlapWilsonPartialFractionZolotarevFermion.h>
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // More maintainable to maintain the following typedef list centrally, as more "impl" targets
@@ -157,6 +180,14 @@ typedef WilsonFermion<WilsonImplR> WilsonFermionR;
 typedef WilsonFermion<WilsonImplF> WilsonFermionF;
 typedef WilsonFermion<WilsonImplD> WilsonFermionD;
 
+typedef WilsonFermion<WilsonAdjImplR> WilsonAdjFermionR;
+typedef WilsonFermion<WilsonAdjImplF> WilsonAdjFermionF;
+typedef WilsonFermion<WilsonAdjImplD> WilsonAdjFermionD;
+
+typedef WilsonFermion<WilsonTwoIndexSymmetricImplR> WilsonTwoIndexSymmetricFermionR;
+typedef WilsonFermion<WilsonTwoIndexSymmetricImplF> WilsonTwoIndexSymmetricFermionF;
+typedef WilsonFermion<WilsonTwoIndexSymmetricImplD> WilsonTwoIndexSymmetricFermionD;
+
 typedef WilsonTMFermion<WilsonImplR> WilsonTMFermionR;
 typedef WilsonTMFermion<WilsonImplF> WilsonTMFermionF;
 typedef WilsonTMFermion<WilsonImplD> WilsonTMFermionD;
@@ -167,6 +198,11 @@ typedef DomainWallFermion<WilsonImplD> DomainWallFermionD;
 typedef MobiusFermion<WilsonImplR> MobiusFermionR;
 typedef MobiusFermion<WilsonImplF> MobiusFermionF;
 typedef MobiusFermion<WilsonImplD> MobiusFermionD;
+
+typedef ZMobiusFermion<ZWilsonImplR> ZMobiusFermionR;
+typedef ZMobiusFermion<ZWilsonImplF> ZMobiusFermionF;
+typedef ZMobiusFermion<ZWilsonImplD> ZMobiusFermionD;
+
 typedef ScaledShamirFermion<WilsonImplR> ScaledShamirFermionR;
 typedef ScaledShamirFermion<WilsonImplF> ScaledShamirFermionF;
 typedef ScaledShamirFermion<WilsonImplD> ScaledShamirFermionD;
@@ -222,21 +258,21 @@ typedef MobiusFermion<GparityWilsonImplD> GparityMobiusFermionD;
 ///////////////////////////////////////////////////////////////////////////////
 // G5 herm -- this has to live in QCD since dirac matrix is not in the broader sector of code
 ///////////////////////////////////////////////////////////////////////////////
-#include <qcd/action/fermion/g5HermitianLinop.h>
+#include <Grid/qcd/action/fermion/g5HermitianLinop.h>
 
 ////////////////////////////////////////
 // Pseudo fermion combinations for HMC
 ////////////////////////////////////////
-#include <qcd/action/pseudofermion/EvenOddSchurDifferentiable.h>
+#include <Grid/qcd/action/pseudofermion/EvenOddSchurDifferentiable.h>
 
-#include <qcd/action/pseudofermion/TwoFlavour.h>
+#include <Grid/qcd/action/pseudofermion/TwoFlavour.h>
-#include <qcd/action/pseudofermion/TwoFlavourRatio.h>
+#include <Grid/qcd/action/pseudofermion/TwoFlavourRatio.h>
-#include <qcd/action/pseudofermion/TwoFlavourEvenOdd.h>
+#include <Grid/qcd/action/pseudofermion/TwoFlavourEvenOdd.h>
-#include <qcd/action/pseudofermion/TwoFlavourEvenOddRatio.h>
+#include <Grid/qcd/action/pseudofermion/TwoFlavourEvenOddRatio.h>
 
-#include <qcd/action/pseudofermion/OneFlavourRational.h>
+#include <Grid/qcd/action/pseudofermion/OneFlavourRational.h>
-#include <qcd/action/pseudofermion/OneFlavourRationalRatio.h>
+#include <Grid/qcd/action/pseudofermion/OneFlavourRationalRatio.h>
-#include <qcd/action/pseudofermion/OneFlavourEvenOddRational.h>
+#include <Grid/qcd/action/pseudofermion/OneFlavourEvenOddRational.h>
-#include <qcd/action/pseudofermion/OneFlavourEvenOddRationalRatio.h>
+#include <Grid/qcd/action/pseudofermion/OneFlavourEvenOddRationalRatio.h>
 
 #endif

lib/qcd/action/fermion/CayleyFermion5D.cc

@@ -28,7 +28,10 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+
 #include <Grid.h>
+
+
 namespace Grid {
 namespace QCD {
 
@@ -45,48 +48,107 @@ namespace QCD {
 		   FourDimGrid,
  	 	   FourDimRedBlackGrid,_M5,p),
    mass(_mass)
- {
+ { }
- }
 
+template<class Impl>  
+void CayleyFermion5D<Impl>::M5D   (const FermionField &psi, FermionField &chi)
+{
+  int Ls=this->Ls;
+  std::vector<Coeff_t> diag (Ls,1.0);
+  std::vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1]=mass;
+  std::vector<Coeff_t> lower(Ls,-1.0); lower[0]   =mass;
+  M5D(psi,chi,chi,lower,diag,upper);
+}
 template<class Impl>
 void CayleyFermion5D<Impl>::Meooe5D    (const FermionField &psi, FermionField &Din)
 {
-    // Assemble Din
   int Ls=this->Ls;
+  std::vector<Coeff_t> diag = bs;
+  std::vector<Coeff_t> upper= cs;
+  std::vector<Coeff_t> lower= cs; 
+  upper[Ls-1]=-mass*upper[Ls-1];
+  lower[0]   =-mass*lower[0];
+  M5D(psi,psi,Din,lower,diag,upper);
+}
+template<class Impl> void CayleyFermion5D<Impl>::Meo5D     (const FermionField &psi, FermionField &chi)
+{
+  int Ls=this->Ls;
+  std::vector<Coeff_t> diag = beo;
+  std::vector<Coeff_t> upper(Ls);
+  std::vector<Coeff_t> lower(Ls);
+  for(int i=0;i<Ls;i++) {
+    upper[i]=-ceo[i];
+    lower[i]=-ceo[i];
+  }
+  upper[Ls-1]=-mass*upper[Ls-1];
+  lower[0]   =-mass*lower[0];
+  M5D(psi,psi,chi,lower,diag,upper);
+}
+template<class Impl>
+void CayleyFermion5D<Impl>::Mooee       (const FermionField &psi, FermionField &chi)
+{
+  int Ls=this->Ls;
+  std::vector<Coeff_t> diag = bee;
+  std::vector<Coeff_t> upper(Ls);
+  std::vector<Coeff_t> lower(Ls);
+  for(int i=0;i<Ls;i++) {
+    upper[i]=-cee[i];
+    lower[i]=-cee[i];
+  }
+  upper[Ls-1]=-mass*upper[Ls-1];
+  lower[0]   =-mass*lower[0];
+  M5D(psi,psi,chi,lower,diag,upper);
+}
+
+template<class Impl>
+void CayleyFermion5D<Impl>::MooeeDag    (const FermionField &psi, FermionField &chi)
+{
+  int Ls=this->Ls;
+  std::vector<Coeff_t> diag = bee;
+  std::vector<Coeff_t> upper(Ls);
+  std::vector<Coeff_t> lower(Ls);
+
   for (int s=0;s<Ls;s++){
+    // Assemble the 5d matrix
     if ( s==0 ) {
-	//	Din = bs psi[s] + cs[s] psi[s+1}
+      upper[s] = -cee[s+1] ;
-	axpby_ssp_pminus(Din,bs[s],psi,cs[s],psi,s,s+1);
+      lower[s] = mass*cee[Ls-1];
-	//      Din+= -mass*cs[s] psi[s+1}
-	axpby_ssp_pplus (Din,1.0,Din,-mass*cs[s],psi,s,Ls-1);
     } else if ( s==(Ls-1)) { 
-	axpby_ssp_pminus(Din,bs[s],psi,-mass*cs[s],psi,s,0);
+      upper[s] = mass*cee[0];
-	axpby_ssp_pplus (Din,1.0,Din,cs[s],psi,s,s-1);
+      lower[s] = -cee[s-1];
     } else {
-	axpby_ssp_pminus(Din,bs[s],psi,cs[s],psi,s,s+1);
+      upper[s]=-cee[s+1];
-	axpby_ssp_pplus(Din,1.0,Din,cs[s],psi,s,s-1);
+      lower[s]=-cee[s-1];
     }
   }
+
+  M5Ddag(psi,psi,chi,lower,diag,upper);
 }
+
+template<class Impl>
+void CayleyFermion5D<Impl>::M5Ddag (const FermionField &psi, FermionField &chi)
+{
+  int Ls=this->Ls;
+  std::vector<Coeff_t> diag(Ls,1.0);
+  std::vector<Coeff_t> upper(Ls,-1.0);
+  std::vector<Coeff_t> lower(Ls,-1.0);
+  upper[Ls-1]=-mass*upper[Ls-1];
+  lower[0]   =-mass*lower[0];
+  M5Ddag(psi,chi,chi,lower,diag,upper);
+}
+
 template<class Impl>
 void CayleyFermion5D<Impl>::MeooeDag5D    (const FermionField &psi, FermionField &Din)
 {
   int Ls=this->Ls;
-    for(int s=0;s<Ls;s++){
+  std::vector<Coeff_t> diag =bs;
-      if ( s==0 ) {
+  std::vector<Coeff_t> upper=cs;
-	axpby_ssp_pplus (Din,bs[s],psi,cs[s+1],psi,s,s+1);
+  std::vector<Coeff_t> lower=cs;
-	axpby_ssp_pminus(Din,1.0,Din,-mass*cs[Ls-1],psi,s,Ls-1);
+  upper[Ls-1]=-mass*upper[Ls-1];
-      } else if ( s==(Ls-1)) { 
+  lower[0]   =-mass*lower[0];
-	axpby_ssp_pplus (Din,bs[s],psi,-mass*cs[0],psi,s,0);
+  M5Ddag(psi,psi,Din,lower,diag,upper);
-	axpby_ssp_pminus(Din,1.0,Din,cs[s-1],psi,s,s-1);
-      } else {
-	axpby_ssp_pplus (Din,bs[s],psi,cs[s+1],psi,s,s+1);
-	axpby_ssp_pminus(Din,1.0,Din,cs[s-1],psi,s,s-1);
-      }
-    }
 }
 
-  // override multiply
 template<class Impl>
 RealD CayleyFermion5D<Impl>::M    (const FermionField &psi, FermionField &chi)
 {
@@ -95,42 +157,14 @@ namespace QCD {
   FermionField Din(psi._grid);
   
   // Assemble Din
-    /*
-    for(int s=0;s<Ls;s++){
-      if ( s==0 ) {
-	//	Din = bs psi[s] + cs[s] psi[s+1}
-	axpby_ssp_pminus(Din,bs[s],psi,cs[s],psi,s,s+1);
-	//      Din+= -mass*cs[s] psi[s+1}
-	axpby_ssp_pplus (Din,1.0,Din,-mass*cs[s],psi,s,Ls-1);
-      } else if ( s==(Ls-1)) { 
-	axpby_ssp_pminus(Din,bs[s],psi,-mass*cs[s],psi,s,0);
-	axpby_ssp_pplus (Din,1.0,Din,cs[s],psi,s,s-1);
-      } else {
-	axpby_ssp_pminus(Din,bs[s],psi,cs[s],psi,s,s+1);
-	axpby_ssp_pplus(Din,1.0,Din,cs[s],psi,s,s-1);
-      }
-    }
-    */
   Meooe5D(psi,Din);
   
   this->DW(Din,chi,DaggerNo);
   // ((b D_W + D_w hop terms +1) on s-diag
   axpby(chi,1.0,1.0,chi,psi); 
   
-    // Call Mooee??
+  M5D(psi,chi);
-    for(int s=0;s<Ls;s++){
+  return(norm2(chi));
-      if ( s==0 ){
-	axpby_ssp_pminus(chi,1.0,chi,-1.0,psi,s,s+1);
-	axpby_ssp_pplus (chi,1.0,chi,mass,psi,s,Ls-1);
-      } else if ( s==(Ls-1)) {
-	axpby_ssp_pminus(chi,1.0,chi,mass,psi,s,0);
-	axpby_ssp_pplus (chi,1.0,chi,-1.0,psi,s,s-1);
-      } else {
-	axpby_ssp_pminus(chi,1.0,chi,-1.0,psi,s,s+1);
-	axpby_ssp_pplus (chi,1.0,chi,-1.0,psi,s,s-1);
-      }
-    }
-    return norm2(chi);
 }
 
 template<class Impl>
@@ -146,39 +180,7 @@ namespace QCD {
   
   MeooeDag5D(Din,chi);
   
-    int Ls=this->Ls;
+  M5Ddag(psi,chi);
-    for(int s=0;s<Ls;s++){
-
-      // Collect the terms in DW
-      //	Chi = bs Din[s] + cs[s] Din[s+1}
-      //    Chi+= -mass*cs[s] psi[s+1}
-      /*
-      if ( s==0 ) {
-	axpby_ssp_pplus (chi,bs[s],Din,cs[s+1],Din,s,s+1);
-	axpby_ssp_pminus(chi,1.0,chi,-mass*cs[Ls-1],Din,s,Ls-1);
-      } else if ( s==(Ls-1)) { 
-	axpby_ssp_pplus (chi,bs[s],Din,-mass*cs[0],Din,s,0);
-	axpby_ssp_pminus(chi,1.0,chi,cs[s-1],Din,s,s-1);
-      } else {
-	axpby_ssp_pplus (chi,bs[s],Din,cs[s+1],Din,s,s+1);
-	axpby_ssp_pminus(chi,1.0,chi,cs[s-1],Din,s,s-1);
-      }
-      */
-
-      // FIXME just call MooeeDag??
-
-      // Collect the terms indept of DW
-      if ( s==0 ){
-	axpby_ssp_pplus (chi,1.0,chi,-1.0,psi,s,s+1);
-	axpby_ssp_pminus(chi,1.0,chi,mass,psi,s,Ls-1);
-      } else if ( s==(Ls-1)) {
-	axpby_ssp_pplus (chi,1.0,chi,mass,psi,s,0);
-	axpby_ssp_pminus(chi,1.0,chi,-1.0,psi,s,s-1);
-      } else {
-	axpby_ssp_pplus(chi,1.0,chi,-1.0,psi,s,s+1);
-	axpby_ssp_pminus(chi,1.0,chi,-1.0,psi,s,s-1);
-      }
-    }
   // ((b D_W + D_w hop terms +1) on s-diag
   axpby (chi,1.0,1.0,chi,psi); 
   return norm2(chi);
@@ -189,30 +191,10 @@ namespace QCD {
 void CayleyFermion5D<Impl>::Meooe       (const FermionField &psi, FermionField &chi)
 {
   int Ls=this->Ls;
-
   FermionField tmp(psi._grid);
-    // Assemble the 5d matrix
-    Meooe5D(psi,tmp); 
-#if 0
-    std::cout << "Meooe Test replacement norm2 tmp = " <<norm2(tmp)<<std::endl;
-    for(int s=0;s<Ls;s++){
-      if ( s==0 ) {
-	//	tmp = bs psi[s] + cs[s] psi[s+1}
-	//      tmp+= -mass*cs[s] psi[s+1}
-	axpby_ssp_pminus(tmp,beo[s],psi,-ceo[s],psi ,s, s+1);
-	axpby_ssp_pplus(tmp,1.0,tmp,mass*ceo[s],psi,s,Ls-1);
-      } else if ( s==(Ls-1)) { 
-	axpby_ssp_pminus(tmp,beo[s],psi,mass*ceo[s],psi,s,0);
-	axpby_ssp_pplus(tmp,1.0,tmp,-ceo[s],psi,s,s-1);
-      } else {
-	axpby_ssp_pminus(tmp,beo[s],psi,-ceo[s],psi,s,s+1);
-	axpby_ssp_pplus (tmp,1.0,tmp,-ceo[s],psi,s,s-1);
-      }
-    }
-    std::cout << "Meooe Test replacement norm2 tmp old = " <<norm2(tmp)<<std::endl;
-#endif
 
-    // Apply 4d dslash
+  Meooe5D(psi,tmp); 
+
   if ( psi.checkerboard == Odd ) {
     this->DhopEO(tmp,chi,DaggerNo);
   } else {
@@ -230,140 +212,16 @@ namespace QCD {
   } else {
     this->DhopOE(psi,tmp,DaggerYes);
   }
-
   MeooeDag5D(tmp,chi); 
-#if 0
-    std::cout << "Meooe Test replacement norm2 chi new = " <<norm2(chi)<<std::endl;
-    // Assemble the 5d matrix
-    int Ls=this->Ls;
-    for(int s=0;s<Ls;s++){
-      if ( s==0 ) {
-	axpby_ssp_pplus(chi,beo[s],tmp,   -ceo[s+1]  ,tmp,s,s+1);
-	axpby_ssp_pminus(chi,   1.0,chi,mass*ceo[Ls-1],tmp,s,Ls-1);
-      } else if ( s==(Ls-1)) { 
-	axpby_ssp_pplus(chi,beo[s],tmp,mass*ceo[0],tmp,s,0);
-	axpby_ssp_pminus(chi,1.0,chi,-ceo[s-1],tmp,s,s-1);
-      } else {
-	axpby_ssp_pplus(chi,beo[s],tmp,-ceo[s+1],tmp,s,s+1);
-	axpby_ssp_pminus(chi,1.0   ,chi,-ceo[s-1],tmp,s,s-1);
-      }
-    }
-    std::cout << "Meooe Test replacement norm2 chi old = " <<norm2(chi)<<std::endl;
-#endif
-
-  }
-
- template<class Impl>
-  void CayleyFermion5D<Impl>::Mooee       (const FermionField &psi, FermionField &chi)
-  {
-    int Ls=this->Ls;
-    for (int s=0;s<Ls;s++){
-      if ( s==0 ) {
-	axpby_ssp_pminus(chi,bee[s],psi ,-cee[s],psi,s,s+1);
-	axpby_ssp_pplus (chi,1.0,chi,mass*cee[s],psi,s,Ls-1);
-      } else if ( s==(Ls-1)) { 
-	axpby_ssp_pminus(chi,bee[s],psi,mass*cee[s],psi,s,0);
-	axpby_ssp_pplus (chi,1.0,chi,-cee[s],psi,s,s-1);
-      } else {
-	axpby_ssp_pminus(chi,bee[s],psi,-cee[s],psi,s,s+1);
-	axpby_ssp_pplus (chi,1.0,chi,-cee[s],psi,s,s-1);
-      }
-    }
 }
 
 template<class Impl>
 void  CayleyFermion5D<Impl>::Mdir (const FermionField &psi, FermionField &chi,int dir,int disp){
-    int Ls=this->Ls;
   FermionField tmp(psi._grid);
-    // Assemble the 5d matrix
+  Meo5D(psi,tmp);
-    for(int s=0;s<Ls;s++){
-      if ( s==0 ) {
-	//	tmp = bs psi[s] + cs[s] psi[s+1}
-	//      tmp+= -mass*cs[s] psi[s+1}
-	axpby_ssp_pminus(tmp,beo[s],psi,-ceo[s],psi ,s, s+1);
-	axpby_ssp_pplus(tmp,1.0,tmp,mass*ceo[s],psi,s,Ls-1);
-      } else if ( s==(Ls-1)) { 
-	axpby_ssp_pminus(tmp,beo[s],psi,mass*ceo[s],psi,s,0);
-	axpby_ssp_pplus(tmp,1.0,tmp,-ceo[s],psi,s,s-1);
-      } else {
-	axpby_ssp_pminus(tmp,beo[s],psi,-ceo[s],psi,s,s+1);
-	axpby_ssp_pplus (tmp,1.0,tmp,-ceo[s],psi,s,s-1);
-      }
-    }
   // Apply 4d dslash fragment
   this->DhopDir(tmp,chi,dir,disp);
 }
-
- template<class Impl>
-  void CayleyFermion5D<Impl>::MooeeDag    (const FermionField &psi, FermionField &chi)
-  {
-    int Ls=this->Ls;
-    for (int s=0;s<Ls;s++){
-      // Assemble the 5d matrix
-      if ( s==0 ) {
-	axpby_ssp_pplus(chi,bee[s],psi,-cee[s+1]  ,psi,s,s+1);
-	axpby_ssp_pminus(chi,1.0,chi,mass*cee[Ls-1],psi,s,Ls-1);
-      } else if ( s==(Ls-1)) { 
-	axpby_ssp_pplus(chi,bee[s],psi,mass*cee[0],psi,s,0);
-	axpby_ssp_pminus(chi,1.0,chi,-cee[s-1],psi,s,s-1);
-      } else {
-	axpby_ssp_pplus(chi,bee[s],psi,-cee[s+1],psi,s,s+1);
-	axpby_ssp_pminus(chi,1.0   ,chi,-cee[s-1],psi,s,s-1);
-      }
-    }
-  }
-
- template<class Impl>
-  void CayleyFermion5D<Impl>::MooeeInv    (const FermionField &psi, FermionField &chi)
-  {
-    int Ls=this->Ls;
-    // Apply (L^{\prime})^{-1}
-    axpby_ssp (chi,1.0,psi,     0.0,psi,0,0);      // chi[0]=psi[0]
-    for (int s=1;s<Ls;s++){
-      axpby_ssp_pplus(chi,1.0,psi,-lee[s-1],chi,s,s-1);// recursion Psi[s] -lee P_+ chi[s-1]
-    }
-    // L_m^{-1} 
-    for (int s=0;s<Ls-1;s++){ // Chi[ee] = 1 - sum[s<Ls-1] -leem[s]P_- chi
-      axpby_ssp_pminus(chi,1.0,chi,-leem[s],chi,Ls-1,s);
-    }
-    // U_m^{-1} D^{-1}
-    for (int s=0;s<Ls-1;s++){
-      // Chi[s] + 1/d chi[s] 
-      axpby_ssp_pplus(chi,1.0/dee[s],chi,-ueem[s]/dee[Ls-1],chi,s,Ls-1);
-    }	
-    axpby_ssp(chi,1.0/dee[Ls-1],chi,0.0,chi,Ls-1,Ls-1); // Modest avoidable 
-    
-    // Apply U^{-1}
-    for (int s=Ls-2;s>=0;s--){
-      axpby_ssp_pminus (chi,1.0,chi,-uee[s],chi,s,s+1);  // chi[Ls]
-    }
-  }
-
- template<class Impl>
-  void CayleyFermion5D<Impl>::MooeeInvDag (const FermionField &psi, FermionField &chi)
-  {
-    int Ls=this->Ls;
-    // Apply (U^{\prime})^{-dagger}
-    axpby_ssp (chi,1.0,psi,     0.0,psi,0,0);      // chi[0]=psi[0]
-    for (int s=1;s<Ls;s++){
-      axpby_ssp_pminus(chi,1.0,psi,-uee[s-1],chi,s,s-1);
-    }
-    // U_m^{-\dagger} 
-    for (int s=0;s<Ls-1;s++){
-      axpby_ssp_pplus(chi,1.0,chi,-ueem[s],chi,Ls-1,s);
-    }
-    // L_m^{-\dagger} D^{-dagger}
-    for (int s=0;s<Ls-1;s++){
-      axpby_ssp_pminus(chi,1.0/dee[s],chi,-leem[s]/dee[Ls-1],chi,s,Ls-1);
-    }	
-    axpby_ssp(chi,1.0/dee[Ls-1],chi,0.0,chi,Ls-1,Ls-1); // Modest avoidable 
-    
-    // Apply L^{-dagger}
-    for (int s=Ls-2;s>=0;s--){
-      axpby_ssp_pplus (chi,1.0,chi,-lee[s],chi,s,s+1);  // chi[Ls]
-    }
-  }
-
 // force terms; five routines; default to Dhop on diagonal
 template<class Impl>
 void CayleyFermion5D<Impl>::MDeriv  (GaugeField &mat,const FermionField &U,const FermionField &V,int dag)
@@ -415,12 +273,21 @@ namespace QCD {
 template<class Impl>
 void CayleyFermion5D<Impl>::SetCoefficientsTanh(Approx::zolotarev_data *zdata,RealD b,RealD c)
 {
-    SetCoefficientsZolotarev(1.0,zdata,b,c);
+  std::vector<Coeff_t> gamma(this->Ls);
-
+  for(int s=0;s<this->Ls;s++) gamma[s] = zdata->gamma[s];
+  SetCoefficientsInternal(1.0,gamma,b,c);
 }
 //Zolo
 template<class Impl>
 void CayleyFermion5D<Impl>::SetCoefficientsZolotarev(RealD zolo_hi,Approx::zolotarev_data *zdata,RealD b,RealD c)
+{
+  std::vector<Coeff_t> gamma(this->Ls);
+  for(int s=0;s<this->Ls;s++) gamma[s] = zdata->gamma[s];
+  SetCoefficientsInternal(zolo_hi,gamma,b,c);
+}
+//Zolo
+template<class Impl>
+void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,std::vector<Coeff_t> & gamma,RealD b,RealD c)
 {
   int Ls=this->Ls;
 
@@ -458,7 +325,7 @@ namespace QCD {
   double bmc = b-c;
   for(int i=0; i < Ls; i++){
     as[i] = 1.0;
-      omega[i] = ((double)zdata->gamma[i])*zolo_hi; //NB reciprocal relative to Chroma NEF code
+    omega[i] = gamma[i]*zolo_hi; //NB reciprocal relative to Chroma NEF code
     bs[i] = 0.5*(bpc/omega[i] + bmc);
     cs[i] = 0.5*(bpc/omega[i] - bmc);
   }
@@ -520,12 +387,14 @@ namespace QCD {
   }
 	
   { 
-      double delta_d=mass*cee[Ls-1];
+    Coeff_t delta_d=mass*cee[Ls-1];
     for(int j=0;j<Ls-1;j++) delta_d *= cee[j]/bee[j];
     dee[Ls-1] += delta_d;
   }  
 }
 
+
+
   FermOpTemplateInstantiate(CayleyFermion5D);
   GparityFermOpTemplateInstantiate(CayleyFermion5D);
 

lib/qcd/action/fermion/CayleyFermion5D.h

@@ -51,6 +51,29 @@ namespace Grid {
       virtual void   MooeeDag    (const FermionField &in, FermionField &out);
       virtual void   MooeeInv    (const FermionField &in, FermionField &out);
       virtual void   MooeeInvDag (const FermionField &in, FermionField &out);
+      virtual void   Meo5D (const FermionField &psi, FermionField &chi);
+
+      virtual void   M5D   (const FermionField &psi, FermionField &chi);
+      virtual void   M5Ddag(const FermionField &psi, FermionField &chi);
+
+      /////////////////////////////////////////////////////
+      // Instantiate different versions depending on Impl
+      /////////////////////////////////////////////////////
+      void M5D(const FermionField &psi,
+	       const FermionField &phi, 
+	       FermionField &chi,
+	       std::vector<Coeff_t> &lower,
+	       std::vector<Coeff_t> &diag,
+	       std::vector<Coeff_t> &upper);
+
+      void M5Ddag(const FermionField &psi,
+		  const FermionField &phi, 
+		  FermionField &chi,
+		  std::vector<Coeff_t> &lower,
+		  std::vector<Coeff_t> &diag,
+		  std::vector<Coeff_t> &upper);
+      void MooeeInternal(const FermionField &in, FermionField &out,int dag,int inv);
+
       virtual void   Instantiatable(void)=0;
 
       // force terms; five routines; default to Dhop on diagonal
@@ -68,23 +91,23 @@ namespace Grid {
       RealD mass;
 
       // Cayley form Moebius (tanh and zolotarev)
-      std::vector<RealD> omega; 
+      std::vector<Coeff_t> omega; 
-      std::vector<RealD> bs;    // S dependent coeffs
+      std::vector<Coeff_t> bs;    // S dependent coeffs
-      std::vector<RealD> cs;    
+      std::vector<Coeff_t> cs;    
-      std::vector<RealD> as;    
+      std::vector<Coeff_t> as;    
       // For preconditioning Cayley form
-      std::vector<RealD> bee;    
+      std::vector<Coeff_t> bee;    
-      std::vector<RealD> cee;    
+      std::vector<Coeff_t> cee;    
-      std::vector<RealD> aee;    
+      std::vector<Coeff_t> aee;    
-      std::vector<RealD> beo;    
+      std::vector<Coeff_t> beo;    
-      std::vector<RealD> ceo;    
+      std::vector<Coeff_t> ceo;    
-      std::vector<RealD> aeo;    
+      std::vector<Coeff_t> aeo;    
       // LDU factorisation of the eeoo matrix
-      std::vector<RealD> lee;    
+      std::vector<Coeff_t> lee;    
-      std::vector<RealD> leem;    
+      std::vector<Coeff_t> leem;    
-      std::vector<RealD> uee;    
+      std::vector<Coeff_t> uee;    
-      std::vector<RealD> ueem;    
+      std::vector<Coeff_t> ueem;    
-      std::vector<RealD> dee;    
+      std::vector<Coeff_t> dee;    
 
       // Constructors
       CayleyFermion5D(GaugeField &_Umu,
@@ -97,9 +120,20 @@ namespace Grid {
     protected:
       void SetCoefficientsZolotarev(RealD zolohi,Approx::zolotarev_data *zdata,RealD b,RealD c);
       void SetCoefficientsTanh(Approx::zolotarev_data *zdata,RealD b,RealD c);
+      void SetCoefficientsInternal(RealD zolo_hi,std::vector<Coeff_t> & gamma,RealD b,RealD c);
     };
 
   }
 }
+#define INSTANTIATE_DPERP(A)\
+template void CayleyFermion5D< A >::M5D(const FermionField &psi,const FermionField &phi,FermionField &chi,\
+					std::vector<Coeff_t> &lower,std::vector<Coeff_t> &diag,std::vector<Coeff_t> &upper); \
+template void CayleyFermion5D< A >::M5Ddag(const FermionField &psi,const FermionField &phi,FermionField &chi,\
+					   std::vector<Coeff_t> &lower,std::vector<Coeff_t> &diag,std::vector<Coeff_t> &upper); \
+template void CayleyFermion5D< A >::MooeeInv    (const FermionField &psi, FermionField &chi); \
+template void CayleyFermion5D< A >::MooeeInvDag (const FermionField &psi, FermionField &chi);
+
+#define CAYLEY_DPERP_CACHE
+#undef  CAYLEY_DPERP_LINALG
 
 #endif

lib/qcd/action/fermion/CayleyFermion5Dcache.cc

@@ -0,0 +1,211 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/qcd/action/fermion/CayleyFermion5D.cc
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <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.h>
+
+
+namespace Grid {
+namespace QCD {
+
+  // FIXME -- make a version of these routines with site loop outermost for cache reuse.
+
+  // Pminus fowards
+  // Pplus  backwards..
+template<class Impl>  
+void CayleyFermion5D<Impl>::M5D(const FermionField &psi,
+				const FermionField &phi, 
+				FermionField &chi,
+				std::vector<Coeff_t> &lower,
+				std::vector<Coeff_t> &diag,
+				std::vector<Coeff_t> &upper)
+{
+  int Ls =this->Ls;
+  GridBase *grid=psi._grid;
+  assert(phi.checkerboard == psi.checkerboard);
+  chi.checkerboard=psi.checkerboard;
+PARALLEL_FOR_LOOP
+  for(int ss=0;ss<grid->oSites();ss+=Ls){ // adds Ls
+    for(int s=0;s<Ls;s++){
+      auto tmp = psi._odata[0];
+      if ( s==0 ) {
+ 	                            spProj5m(tmp,psi._odata[ss+s+1]);
+	chi[ss+s]=diag[s]*phi[ss+s]+upper[s]*tmp;
+
+	                    spProj5p(tmp,psi._odata[ss+Ls-1]);
+	chi[ss+s]=chi[ss+s]+lower[s]*tmp;
+      } else if ( s==(Ls-1)) {
+	                            spProj5m(tmp,psi._odata[ss+0]);
+	chi[ss+s]=diag[s]*phi[ss+s]+upper[s]*tmp;
+
+ 	                    spProj5p(tmp,psi._odata[ss+s-1]);
+	chi[ss+s]=chi[ss+s]+lower[s]*tmp;
+      } else { 
+	                            spProj5m(tmp,psi._odata[ss+s+1]);
+	chi[ss+s]=diag[s]*phi[ss+s]+upper[s]*tmp;
+
+	                    spProj5p(tmp,psi._odata[ss+s-1]);
+	chi[ss+s]=chi[ss+s]+lower[s]*tmp;
+      }
+    }
+  }
+}
+
+template<class Impl>  
+void CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi,
+				   const FermionField &phi, 
+				   FermionField &chi,
+				   std::vector<Coeff_t> &lower,
+				   std::vector<Coeff_t> &diag,
+				   std::vector<Coeff_t> &upper)
+{
+  int Ls =this->Ls;
+  GridBase *grid=psi._grid;
+  assert(phi.checkerboard == psi.checkerboard);
+  chi.checkerboard=psi.checkerboard;
+
+PARALLEL_FOR_LOOP
+  for(int ss=0;ss<grid->oSites();ss+=Ls){ // adds Ls
+    auto tmp = psi._odata[0];
+    for(int s=0;s<Ls;s++){
+      if ( s==0 ) {
+	spProj5p(tmp,psi._odata[ss+s+1]);
+	chi[ss+s]=diag[s]*phi[ss+s]+upper[s]*tmp;
+
+	spProj5m(tmp,psi._odata[ss+Ls-1]);
+	chi[ss+s]=chi[ss+s]+lower[s]*tmp;
+      } else if ( s==(Ls-1)) {
+	spProj5p(tmp,psi._odata[ss+0]);
+	chi[ss+s]=diag[s]*phi[ss+s]+upper[s]*tmp;
+
+	spProj5m(tmp,psi._odata[ss+s-1]);
+	chi[ss+s]=chi[ss+s]+lower[s]*tmp;
+      } else { 
+	spProj5p(tmp,psi._odata[ss+s+1]);
+	chi[ss+s]=diag[s]*phi[ss+s]+upper[s]*tmp;
+
+	spProj5m(tmp,psi._odata[ss+s-1]);
+	chi[ss+s]=chi[ss+s]+lower[s]*tmp;
+      }
+    }
+  }
+}
+
+template<class Impl>
+void CayleyFermion5D<Impl>::MooeeInv    (const FermionField &psi, FermionField &chi)
+{
+  GridBase *grid=psi._grid;
+  int Ls=this->Ls;
+
+  chi.checkerboard=psi.checkerboard;
+
+PARALLEL_FOR_LOOP
+  for(int ss=0;ss<grid->oSites();ss+=Ls){ // adds Ls
+    auto tmp = psi._odata[0];
+
+    // Apply (L^{\prime})^{-1}
+    chi[ss]=psi[ss]; // chi[0]=psi[0]
+    for(int s=1;s<Ls;s++){
+                            spProj5p(tmp,chi[ss+s-1]);  
+      chi[ss+s] = psi[ss+s]-lee[s-1]*tmp;
+    }
+    // L_m^{-1} 
+    for (int s=0;s<Ls-1;s++){ // Chi[ee] = 1 - sum[s<Ls-1] -leem[s]P_- chi
+                                   spProj5m(tmp,chi[ss+s]);    
+      chi[ss+Ls-1] = chi[ss+Ls-1] - leem[s]*tmp;
+    }
+    // U_m^{-1} D^{-1}
+    for (int s=0;s<Ls-1;s++){
+      // Chi[s] + 1/d chi[s] 
+                                                spProj5p(tmp,chi[ss+Ls-1]); 
+      chi[ss+s] = (1.0/dee[s])*chi[ss+s]-(ueem[s]/dee[Ls-1])*tmp;
+    }	
+    chi[ss+Ls-1]= (1.0/dee[Ls-1])*chi[ss+Ls-1];
+      
+    // Apply U^{-1}
+    for (int s=Ls-2;s>=0;s--){
+                            spProj5m(tmp,chi[ss+s+1]);  
+      chi[ss+s] = chi[ss+s] - uee[s]*tmp;
+    }
+  }
+}
+
+template<class Impl>
+void CayleyFermion5D<Impl>::MooeeInvDag (const FermionField &psi, FermionField &chi)
+{
+  GridBase *grid=psi._grid;
+  int Ls=this->Ls;
+
+  assert(psi.checkerboard == psi.checkerboard);
+  chi.checkerboard=psi.checkerboard;
+
+
+PARALLEL_FOR_LOOP
+  for(int ss=0;ss<grid->oSites();ss+=Ls){ // adds Ls
+
+    auto tmp = psi._odata[0];
+
+    // Apply (U^{\prime})^{-dagger}
+    chi[ss]=psi[ss];
+    for (int s=1;s<Ls;s++){
+                            spProj5m(tmp,chi[ss+s-1]);
+      chi[ss+s] = psi[ss+s]-uee[s-1]*tmp;
+    }
+    // U_m^{-\dagger} 
+    for (int s=0;s<Ls-1;s++){
+                                   spProj5p(tmp,chi[ss+s]);
+      chi[ss+Ls-1] = chi[ss+Ls-1] - ueem[s]*tmp;
+    }
+
+    // L_m^{-\dagger} D^{-dagger}
+    for (int s=0;s<Ls-1;s++){
+      spProj5m(tmp,chi[ss+Ls-1]);
+      chi[ss+s] = (1.0/dee[s])*chi[ss+s]-(leem[s]/dee[Ls-1])*tmp;
+    }	
+    chi[ss+Ls-1]= (1.0/dee[Ls-1])*chi[ss+Ls-1];
+  
+    // Apply L^{-dagger}
+    for (int s=Ls-2;s>=0;s--){
+      spProj5p(tmp,chi[ss+s+1]);
+      chi[ss+s] = chi[ss+s] - lee[s]*tmp;
+    }
+  }
+}
+
+#ifdef CAYLEY_DPERP_CACHE
+  INSTANTIATE_DPERP(WilsonImplF);
+  INSTANTIATE_DPERP(WilsonImplD);
+  INSTANTIATE_DPERP(GparityWilsonImplF);
+  INSTANTIATE_DPERP(GparityWilsonImplD);
+  INSTANTIATE_DPERP(ZWilsonImplF);
+  INSTANTIATE_DPERP(ZWilsonImplD);
+#endif
+
+}}

lib/qcd/action/fermion/CayleyFermion5Ddense.cc

@@ -0,0 +1,133 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/qcd/action/fermion/CayleyFermion5D.cc
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <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/Eigen/Dense>
+#include <Grid.h>
+
+
+namespace Grid {
+namespace QCD {
+  /*
+   * Dense matrix versions of routines
+   */
+
+  /*
+template<class Impl>
+void CayleyFermion5D<Impl>::MooeeInvDag (const FermionField &psi, FermionField &chi)
+{
+  this->MooeeInternal(psi,chi,DaggerYes,InverseYes);
+}
+  
+template<class Impl>
+void CayleyFermion5D<Impl>::MooeeInv(const FermionField &psi, FermionField &chi)
+{
+  this->MooeeInternal(psi,chi,DaggerNo,InverseYes);
+}
+  */
+template<class Impl>
+void CayleyFermion5D<Impl>::MooeeInternal(const FermionField &psi, FermionField &chi,int dag, int inv)
+{
+  int Ls=this->Ls;
+  int LLs = psi._grid->_rdimensions[0];
+  int vol = psi._grid->oSites()/LLs;
+  
+  chi.checkerboard=psi.checkerboard;
+  
+  assert(Ls==LLs);
+  
+  Eigen::MatrixXd Pplus  = Eigen::MatrixXd::Zero(Ls,Ls);
+  Eigen::MatrixXd Pminus = Eigen::MatrixXd::Zero(Ls,Ls);
+  
+  for(int s=0;s<Ls;s++){
+    Pplus(s,s) = bee[s];
+    Pminus(s,s)= bee[s];
+  }
+  
+  for(int s=0;s<Ls-1;s++){
+    Pminus(s,s+1) = -cee[s];
+  }
+  
+  for(int s=0;s<Ls-1;s++){
+    Pplus(s+1,s) = -cee[s+1];
+  }
+  Pplus (0,Ls-1) = mass*cee[0];
+  Pminus(Ls-1,0) = mass*cee[Ls-1];
+  
+  Eigen::MatrixXd PplusMat ;
+  Eigen::MatrixXd PminusMat;
+  
+  if ( inv ) {
+    PplusMat =Pplus.inverse();
+    PminusMat=Pminus.inverse();
+  } else { 
+    PplusMat =Pplus;
+    PminusMat=Pminus;
+  }
+  
+  if(dag){
+    PplusMat.adjointInPlace();
+    PminusMat.adjointInPlace();
+  }
+
+  // For the non-vectorised s-direction this is simple
+  
+  for(auto site=0;site<vol;site++){
+    
+    SiteSpinor     SiteChi;
+    SiteHalfSpinor SitePplus;
+    SiteHalfSpinor SitePminus;
+    
+    for(int s1=0;s1<Ls;s1++){
+      SiteChi =zero;
+      for(int s2=0;s2<Ls;s2++){
+	int lex2 = s2+Ls*site;
+	
+	if ( PplusMat(s1,s2) != 0.0 ) {
+	  spProj5p(SitePplus,psi[lex2]);
+	  accumRecon5p(SiteChi,PplusMat (s1,s2)*SitePplus);
+	}
+	
+	if ( PminusMat(s1,s2) != 0.0 ) {
+	  spProj5m(SitePminus,psi[lex2]);
+	  accumRecon5m(SiteChi,PminusMat(s1,s2)*SitePminus);
+	}
+      }
+      chi[s1+Ls*site] = SiteChi*0.5;
+    }
+  }
+}
+
+template void CayleyFermion5D<GparityWilsonImplF>::MooeeInternal(const FermionField &psi, FermionField &chi,int dag, int inv);
+template void CayleyFermion5D<GparityWilsonImplD>::MooeeInternal(const FermionField &psi, FermionField &chi,int dag, int inv);
+template void CayleyFermion5D<WilsonImplF>::MooeeInternal(const FermionField &psi, FermionField &chi,int dag, int inv);
+template void CayleyFermion5D<WilsonImplD>::MooeeInternal(const FermionField &psi, FermionField &chi,int dag, int inv);
+
+}}

lib/qcd/action/fermion/CayleyFermion5Dssp.cc

@@ -0,0 +1,149 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/qcd/action/fermion/CayleyFermion5D.cc
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <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.h>
+
+
+namespace Grid {
+namespace QCD {
+
+  // FIXME -- make a version of these routines with site loop outermost for cache reuse.
+
+  // Pminus fowards
+  // Pplus  backwards
+template<class Impl>  
+void CayleyFermion5D<Impl>::M5D(const FermionField &psi,
+				const FermionField &phi, 
+				FermionField &chi,
+				std::vector<Coeff_t> &lower,
+				std::vector<Coeff_t> &diag,
+				std::vector<Coeff_t> &upper)
+{
+  int Ls=this->Ls;
+  for(int s=0;s<Ls;s++){
+    if ( s==0 ) {
+      axpby_ssp_pminus(chi,diag[s],phi,upper[s],psi,s,s+1);
+      axpby_ssp_pplus (chi,1.0,chi,lower[s],psi,s,Ls-1);
+    } else if ( s==(Ls-1)) { 
+      axpby_ssp_pminus(chi,diag[s],phi,upper[s],psi,s,0);
+      axpby_ssp_pplus (chi,1.0,chi,lower[s],psi,s,s-1);
+    } else {
+      axpby_ssp_pminus(chi,diag[s],phi,upper[s],psi,s,s+1);
+      axpby_ssp_pplus(chi,1.0,chi,lower[s],psi,s,s-1);
+    }
+  }
+}
+template<class Impl>  
+void CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi,
+				   const FermionField &phi, 
+				   FermionField &chi,
+				   std::vector<Coeff_t> &lower,
+				   std::vector<Coeff_t> &diag,
+				   std::vector<Coeff_t> &upper)
+{
+  int Ls=this->Ls;
+  for(int s=0;s<Ls;s++){
+    if ( s==0 ) {
+      axpby_ssp_pplus (chi,diag[s],phi,upper[s],psi,s,s+1);
+      axpby_ssp_pminus(chi,1.0,chi,lower[s],psi,s,Ls-1);
+    } else if ( s==(Ls-1)) { 
+      axpby_ssp_pplus (chi,diag[s],phi,upper[s],psi,s,0);
+      axpby_ssp_pminus(chi,1.0,chi,lower[s],psi,s,s-1);
+    } else {
+      axpby_ssp_pplus (chi,diag[s],phi,upper[s],psi,s,s+1);
+      axpby_ssp_pminus(chi,1.0,chi,lower[s],psi,s,s-1);
+    }
+  }
+}
+
+template<class Impl>
+void CayleyFermion5D<Impl>::MooeeInv    (const FermionField &psi, FermionField &chi)
+{
+  chi.checkerboard=psi.checkerboard;
+  int Ls=this->Ls;
+  // Apply (L^{\prime})^{-1}
+  axpby_ssp (chi,1.0,psi,     0.0,psi,0,0);      // chi[0]=psi[0]
+  for (int s=1;s<Ls;s++){
+    axpby_ssp_pplus(chi,1.0,psi,-lee[s-1],chi,s,s-1);// recursion Psi[s] -lee P_+ chi[s-1]
+  }
+  // L_m^{-1} 
+  for (int s=0;s<Ls-1;s++){ // Chi[ee] = 1 - sum[s<Ls-1] -leem[s]P_- chi
+    axpby_ssp_pminus(chi,1.0,chi,-leem[s],chi,Ls-1,s);
+  }
+  // U_m^{-1} D^{-1}
+  for (int s=0;s<Ls-1;s++){
+    // Chi[s] + 1/d chi[s] 
+    axpby_ssp_pplus(chi,1.0/dee[s],chi,-ueem[s]/dee[Ls-1],chi,s,Ls-1);
+  }	
+  axpby_ssp(chi,1.0/dee[Ls-1],chi,0.0,chi,Ls-1,Ls-1); // Modest avoidable 
+  
+  // Apply U^{-1}
+  for (int s=Ls-2;s>=0;s--){
+    axpby_ssp_pminus (chi,1.0,chi,-uee[s],chi,s,s+1);  // chi[Ls]
+  }
+}
+
+template<class Impl>
+void CayleyFermion5D<Impl>::MooeeInvDag (const FermionField &psi, FermionField &chi)
+{
+  chi.checkerboard=psi.checkerboard;
+  int Ls=this->Ls;
+  // Apply (U^{\prime})^{-dagger}
+  axpby_ssp (chi,1.0,psi,     0.0,psi,0,0);      // chi[0]=psi[0]
+  for (int s=1;s<Ls;s++){
+    axpby_ssp_pminus(chi,1.0,psi,-uee[s-1],chi,s,s-1);
+  }
+  // U_m^{-\dagger} 
+  for (int s=0;s<Ls-1;s++){
+    axpby_ssp_pplus(chi,1.0,chi,-ueem[s],chi,Ls-1,s);
+  }
+  // L_m^{-\dagger} D^{-dagger}
+  for (int s=0;s<Ls-1;s++){
+    axpby_ssp_pminus(chi,1.0/dee[s],chi,-leem[s]/dee[Ls-1],chi,s,Ls-1);
+  }	
+  axpby_ssp(chi,1.0/dee[Ls-1],chi,0.0,chi,Ls-1,Ls-1); // Modest avoidable 
+  
+  // Apply L^{-dagger}
+  for (int s=Ls-2;s>=0;s--){
+    axpby_ssp_pplus (chi,1.0,chi,-lee[s],chi,s,s+1);  // chi[Ls]
+  }
+}
+
+
+#ifdef CAYLEY_DPERP_LINALG
+  INSTANTIATE(WilsonImplF);
+  INSTANTIATE(WilsonImplD);
+  INSTANTIATE(GparityWilsonImplF);
+  INSTANTIATE(GparityWilsonImplD);
+#endif
+
+}
+}

lib/qcd/action/fermion/CayleyFermion5Dvec.cc

@@ -0,0 +1,309 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/qcd/action/fermion/CayleyFermion5D.cc
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <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/Eigen/Dense>
+#include <Grid.h>
+
+
+namespace Grid {
+namespace QCD {
+  /*
+   * Dense matrix versions of routines
+   */
+template<class Impl>
+void CayleyFermion5D<Impl>::MooeeInvDag (const FermionField &psi, FermionField &chi)
+{
+  this->MooeeInternal(psi,chi,DaggerYes,InverseYes);
+}
+  
+template<class Impl>
+void CayleyFermion5D<Impl>::MooeeInv(const FermionField &psi, FermionField &chi)
+{
+  this->MooeeInternal(psi,chi,DaggerNo,InverseYes);
+}
+template<class Impl>  
+void CayleyFermion5D<Impl>::M5D(const FermionField &psi,
+				const FermionField &phi, 
+				FermionField &chi,
+				std::vector<Coeff_t> &lower,
+				std::vector<Coeff_t> &diag,
+				std::vector<Coeff_t> &upper)
+{
+  GridBase *grid=psi._grid;
+  int Ls   = this->Ls;
+  int LLs  = grid->_rdimensions[0];
+  int nsimd= Simd::Nsimd();
+
+  Vector<iSinglet<Simd> > u(LLs);
+  Vector<iSinglet<Simd> > l(LLs);
+  Vector<iSinglet<Simd> > d(LLs);
+
+  assert(Ls/LLs==nsimd);
+  assert(phi.checkerboard == psi.checkerboard);
+
+  chi.checkerboard=psi.checkerboard;
+
+  // just directly address via type pun
+  typedef typename Simd::scalar_type scalar_type;
+  scalar_type * u_p = (scalar_type *)&u[0];
+  scalar_type * l_p = (scalar_type *)&l[0];
+  scalar_type * d_p = (scalar_type *)&d[0];
+
+  for(int o=0;o<LLs;o++){ // outer
+  for(int i=0;i<nsimd;i++){ //inner
+    int s  = o+i*LLs;
+    int ss = o*nsimd+i;
+    u_p[ss] = upper[s];
+    l_p[ss] = lower[s];
+    d_p[ss] = diag[s];
+  }}
+
+PARALLEL_FOR_LOOP
+  for(int ss=0;ss<grid->oSites();ss+=LLs){ // adds LLs
+
+    alignas(64) SiteHalfSpinor hp;
+    alignas(64) SiteHalfSpinor hm;
+    alignas(64) SiteSpinor fp;
+    alignas(64) SiteSpinor fm;
+
+    for(int v=0;v<LLs;v++){
+
+      int vp=(v+1)%LLs;
+      int vm=(v+LLs-1)%LLs;
+
+      spProj5m(hp,psi[ss+vp]);
+      spProj5p(hm,psi[ss+vm]);
+      
+      if ( vp<=v ) rotate(hp,hp,1);
+      if ( vm>=v ) rotate(hm,hm,nsimd-1);
+
+      hp=hp*0.5;
+      hm=hm*0.5;
+      spRecon5m(fp,hp);
+      spRecon5p(fm,hm);
+
+      chi[ss+v] = d[v]*phi[ss+v]+u[v]*fp;
+      chi[ss+v] = chi[ss+v]     +l[v]*fm;
+
+    }
+  }
+}
+
+template<class Impl>  
+void CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi,
+				   const FermionField &phi, 
+				   FermionField &chi,
+				   std::vector<Coeff_t> &lower,
+				   std::vector<Coeff_t> &diag,
+				   std::vector<Coeff_t> &upper)
+{
+  GridBase *grid=psi._grid;
+  int Ls   = this->Ls;
+  int LLs  = grid->_rdimensions[0];
+  int nsimd= Simd::Nsimd();
+
+  Vector<iSinglet<Simd> > u(LLs);
+  Vector<iSinglet<Simd> > l(LLs);
+  Vector<iSinglet<Simd> > d(LLs);
+
+  assert(Ls/LLs==nsimd);
+  assert(phi.checkerboard == psi.checkerboard);
+
+  chi.checkerboard=psi.checkerboard;
+
+  // just directly address via type pun
+  typedef typename Simd::scalar_type scalar_type;
+  scalar_type * u_p = (scalar_type *)&u[0];
+  scalar_type * l_p = (scalar_type *)&l[0];
+  scalar_type * d_p = (scalar_type *)&d[0];
+
+  for(int o=0;o<LLs;o++){ // outer
+  for(int i=0;i<nsimd;i++){ //inner
+    int s  = o+i*LLs;
+    int ss = o*nsimd+i;
+    u_p[ss] = upper[s];
+    l_p[ss] = lower[s];
+    d_p[ss] = diag[s];
+  }}
+
+PARALLEL_FOR_LOOP
+  for(int ss=0;ss<grid->oSites();ss+=LLs){ // adds LLs
+
+    alignas(64) SiteHalfSpinor hp;
+    alignas(64) SiteHalfSpinor hm;
+    alignas(64) SiteSpinor fp;
+    alignas(64) SiteSpinor fm;
+
+    for(int v=0;v<LLs;v++){
+
+      int vp=(v+1)%LLs;
+      int vm=(v+LLs-1)%LLs;
+
+      spProj5p(hp,psi[ss+vp]);
+      spProj5m(hm,psi[ss+vm]);
+
+      if ( vp<=v ) rotate(hp,hp,1);
+      if ( vm>=v ) rotate(hm,hm,nsimd-1);
+      
+      hp=hp*0.5;
+      hm=hm*0.5;
+      spRecon5p(fp,hp);
+      spRecon5m(fm,hm);
+
+      chi[ss+v] = d[v]*phi[ss+v]+u[v]*fp;
+      chi[ss+v] = chi[ss+v]     +l[v]*fm;
+
+    }
+  }
+}
+
+template<class Impl>
+void CayleyFermion5D<Impl>::MooeeInternal(const FermionField &psi, FermionField &chi,int dag, int inv)
+{
+  int Ls=this->Ls;
+  int LLs = psi._grid->_rdimensions[0];
+  int vol = psi._grid->oSites()/LLs;
+
+  chi.checkerboard=psi.checkerboard;
+  
+  Eigen::MatrixXcd Pplus  = Eigen::MatrixXcd::Zero(Ls,Ls);
+  Eigen::MatrixXcd Pminus = Eigen::MatrixXcd::Zero(Ls,Ls);
+  
+  for(int s=0;s<Ls;s++){
+    Pplus(s,s) = bee[s];
+    Pminus(s,s)= bee[s];
+  }
+  
+  for(int s=0;s<Ls-1;s++){
+    Pminus(s,s+1) = -cee[s];
+  }
+  
+  for(int s=0;s<Ls-1;s++){
+    Pplus(s+1,s) = -cee[s+1];
+  }
+  Pplus (0,Ls-1) = mass*cee[0];
+  Pminus(Ls-1,0) = mass*cee[Ls-1];
+  
+  Eigen::MatrixXcd PplusMat ;
+  Eigen::MatrixXcd PminusMat;
+  
+  if ( inv ) {
+    PplusMat =Pplus.inverse();
+    PminusMat=Pminus.inverse();
+  } else { 
+    PplusMat =Pplus;
+    PminusMat=Pminus;
+  }
+  
+  if(dag){
+    PplusMat.adjointInPlace();
+    PminusMat.adjointInPlace();
+  }
+  
+  typedef typename SiteHalfSpinor::scalar_type scalar_type;
+  const int Nsimd=Simd::Nsimd();
+  Vector<iSinglet<Simd> > Matp(Ls*LLs);
+  Vector<iSinglet<Simd> > Matm(Ls*LLs);
+
+  for(int s2=0;s2<Ls;s2++){
+  for(int s1=0;s1<LLs;s1++){
+    int istride = LLs;
+    int ostride = 1;
+      Simd Vp;
+      Simd Vm;
+      scalar_type *sp = (scalar_type *)&Vp;
+      scalar_type *sm = (scalar_type *)&Vm;
+      for(int l=0;l<Nsimd;l++){
+	sp[l] = PplusMat (l*istride+s1*ostride ,s2);
+	sm[l] = PminusMat(l*istride+s1*ostride,s2);
+      }
+      Matp[LLs*s2+s1] = Vp;
+      Matm[LLs*s2+s1] = Vm;
+    }
+  }
+  
+  // Dynamic allocate on stack to get per thread without serialised heap acces
+PARALLEL_FOR_LOOP
+  for(auto site=0;site<vol;site++){
+    
+    //    SiteHalfSpinor *SitePplus =(SiteHalfSpinor *) alloca(LLs*sizeof(SiteHalfSpinor));
+    //    SiteHalfSpinor *SitePminus=(SiteHalfSpinor *) alloca(LLs*sizeof(SiteHalfSpinor));
+    //    SiteSpinor     *SiteChi   =(SiteSpinor *)     alloca(LLs*sizeof(SiteSpinor));
+
+    Vector<SiteHalfSpinor> SitePplus(LLs);
+    Vector<SiteHalfSpinor> SitePminus(LLs);
+    Vector<SiteHalfSpinor> SiteChiP(LLs);
+    Vector<SiteHalfSpinor> SiteChiM(LLs);
+    Vector<SiteSpinor>     SiteChi(LLs);
+
+    SiteHalfSpinor BcastP;
+    SiteHalfSpinor BcastM;
+
+    for(int s=0;s<LLs;s++){
+      int lex = s+LLs*site;
+      spProj5p(SitePplus[s] ,psi[lex]);
+      spProj5m(SitePminus[s],psi[lex]);
+      SiteChiP[s]=zero;
+      SiteChiM[s]=zero;
+    }
+      
+    int s=0;
+    for(int  l=0; l<Simd::Nsimd();l++){ // simd lane
+      for(int s2=0;s2<LLs;s2++){ // Column loop of right hand side
+	vbroadcast(BcastP,SitePplus [s2],l);
+	vbroadcast(BcastM,SitePminus[s2],l);
+	for(int s1=0;s1<LLs;s1++){ // Column loop of reduction variables
+	  SiteChiP[s1]=SiteChiP[s1]+Matp[LLs*s+s1]*BcastP;
+	  SiteChiM[s1]=SiteChiM[s1]+Matm[LLs*s+s1]*BcastM;
+	}
+      s++;
+    }}
+
+    for(int s=0;s<LLs;s++){
+      int lex = s+LLs*site;
+      spRecon5p(SiteChi[s],SiteChiP[s]);
+      accumRecon5m(SiteChi[s],SiteChiM[s]);
+      chi[lex] = SiteChi[s]*0.5;
+    }
+  }
+}
+
+INSTANTIATE_DPERP(DomainWallVec5dImplD);
+INSTANTIATE_DPERP(DomainWallVec5dImplF);
+INSTANTIATE_DPERP(ZDomainWallVec5dImplD);
+INSTANTIATE_DPERP(ZDomainWallVec5dImplF);
+
+template void CayleyFermion5D<DomainWallVec5dImplF>::MooeeInternal(const FermionField &psi, FermionField &chi,int dag, int inv);
+template void CayleyFermion5D<DomainWallVec5dImplD>::MooeeInternal(const FermionField &psi, FermionField &chi,int dag, int inv);
+template void CayleyFermion5D<ZDomainWallVec5dImplF>::MooeeInternal(const FermionField &psi, FermionField &chi,int dag, int inv);
+template void CayleyFermion5D<ZDomainWallVec5dImplD>::MooeeInternal(const FermionField &psi, FermionField &chi,int dag, int inv);
+
+}}

lib/qcd/action/fermion/DomainWallFermion.h

@@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef  GRID_QCD_DOMAIN_WALL_FERMION_H
 #define  GRID_QCD_DOMAIN_WALL_FERMION_H
 
-#include <Grid.h>
+#include <Grid/Grid.h>
 
 namespace Grid {
 
@@ -63,7 +63,7 @@ namespace Grid {
 	Approx::zolotarev_data *zdata = Approx::higham(eps,this->Ls);// eps is ignored for higham
 	assert(zdata->n==this->Ls);
 	
-	std::cout<<GridLogMessage << "DomainWallFermion with Ls="<<this->Ls<<std::endl;
+	//	std::cout<<GridLogMessage << "DomainWallFermion with Ls="<<this->Ls<<std::endl;
 	// Call base setter
 	this->SetCoefficientsTanh(zdata,1.0,0.0);
 

lib/qcd/action/fermion/FermionOperatorImpl.h

@@ -25,14 +25,14 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 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
+See the full license in the file "LICENSE" in the top level distribution
+directory
 *************************************************************************************/
 /*  END LEGAL */
 #ifndef GRID_QCD_FERMION_OPERATOR_IMPL_H
 #define GRID_QCD_FERMION_OPERATOR_IMPL_H
 
 namespace Grid {
-
 namespace QCD {
 
 
@@ -75,7 +75,7 @@ namespace Grid {
   //
   //
   // template<class Impl>
-    // class MyOp : pubic<Impl> { 
+  // class MyOp : public<Impl> { 
   // public:
   //
   //    INHERIT_ALL_IMPL_TYPES(Impl);
@@ -100,26 +100,35 @@ namespace Grid {
   typedef typename Impl::SiteHalfSpinor       SiteHalfSpinor;		\
   typedef typename Impl::Compressor               Compressor;		\
   typedef typename Impl::StencilImpl             StencilImpl;		\
-    typedef typename Impl::ImplParams ImplParams;
+  typedef typename Impl::ImplParams ImplParams;				\
+  typedef typename Impl::Coeff_t       Coeff_t;
   
 #define INHERIT_IMPL_TYPES(Base) \
   INHERIT_GIMPL_TYPES(Base)	 \
   INHERIT_FIMPL_TYPES(Base)
   
-    ///////
+  /////////////////////////////////////////////////////////////////////////////
   // Single flavour four spinors with colour index
-    ///////
+  /////////////////////////////////////////////////////////////////////////////
-    template<class S,int Nrepresentation=Nc>
+  template <class S, class Representation = FundamentalRepresentation,class _Coeff_t = RealD >
-    class WilsonImpl :  public PeriodicGaugeImpl< GaugeImplTypes< S,Nrepresentation> > { 
+  class WilsonImpl : public PeriodicGaugeImpl<GaugeImplTypes<S, Representation::Dimension > > {
+
     public:
 
-      typedef PeriodicGaugeImpl< GaugeImplTypes< S,Nrepresentation> > Gimpl;
+    static const int Dimension = Representation::Dimension;
+    typedef PeriodicGaugeImpl<GaugeImplTypes<S, Dimension > > Gimpl;
+      
+    //Necessary?
+    constexpr bool is_fundamental() const{return Dimension == Nc ? 1 : 0;}
+    
+    const bool LsVectorised=false;
+    typedef _Coeff_t Coeff_t;
 
     INHERIT_GIMPL_TYPES(Gimpl);
       
-      template<typename vtype> using iImplSpinor             = iScalar<iVector<iVector<vtype, Nrepresentation>, Ns> >;
+    template <typename vtype> using iImplSpinor            = iScalar<iVector<iVector<vtype, Dimension>, Ns> >;
-      template<typename vtype> using iImplHalfSpinor         = iScalar<iVector<iVector<vtype, Nrepresentation>, Nhs> >;
+    template <typename vtype> using iImplHalfSpinor        = iScalar<iVector<iVector<vtype, Dimension>, Nhs> >;
-      template<typename vtype> using iImplDoubledGaugeField  = iVector<iScalar<iMatrix<vtype, Nrepresentation> >, Nds >;
+    template <typename vtype> using iImplDoubledGaugeField = iVector<iScalar<iMatrix<vtype, Dimension> >, Nds>;
     
     typedef iImplSpinor<Simd>            SiteSpinor;
     typedef iImplHalfSpinor<Simd>        SiteHalfSpinor;
@@ -138,7 +147,12 @@ namespace Grid {
       
     bool overlapCommsCompute(void) { return Params.overlapCommsCompute; };
       
-      inline void multLink(SiteHalfSpinor &phi,const SiteDoubledGaugeField &U,const SiteHalfSpinor &chi,int mu,StencilEntry *SE,StencilImpl &St){
+    inline void multLink(SiteHalfSpinor &phi,
+			 const SiteDoubledGaugeField &U,
+			 const SiteHalfSpinor &chi,
+			 int mu,
+			 StencilEntry *SE,
+			 StencilImpl &St) {
       mult(&phi(), &U(mu), &chi());
     }
       
@@ -146,8 +160,10 @@ namespace Grid {
     inline void loadLinkElement(Simd &reg, ref &memory) {
       reg = memory;
     }
-      inline void DoubleStore(GridBase *GaugeGrid,DoubledGaugeField &Uds,const GaugeField &Umu)
+      
-      {
+    inline void DoubleStore(GridBase *GaugeGrid,
+			    DoubledGaugeField &Uds,
+			    const GaugeField &Umu) {
       conformable(Uds._grid, GaugeGrid);
       conformable(Umu._grid, GaugeGrid);
       GaugeLinkField U(GaugeGrid);
@@ -168,9 +184,9 @@ namespace Grid {
     inline void InsertForce5D(GaugeField &mat, FermionField &Btilde, FermionField &Atilde,int mu){
       
       int Ls=Btilde._grid->_fdimensions[0];
-
       GaugeLinkField tmp(mat._grid);
       tmp = zero;
+      
       PARALLEL_FOR_LOOP
       for(int sss=0;sss<tmp._grid->oSites();sss++){
 	int sU=sss;
@@ -182,18 +198,19 @@ PARALLEL_FOR_LOOP
       PokeIndex<LorentzIndex>(mat,tmp,mu);
       
     }
-
   };
 
-
+  ////////////////////////////////////////////////////////////////////////////////////
-
-    ///////
   // Single flavour four spinors with colour index, 5d redblack
-    ///////
+  ////////////////////////////////////////////////////////////////////////////////////
-    template<class S,int Nrepresentation=Nc>
+
-    class DomainWallRedBlack5dImpl :  public PeriodicGaugeImpl< GaugeImplTypes< S,Nrepresentation> > { 
+template<class S,int Nrepresentation=Nc,class _Coeff_t = RealD>
+class DomainWallVec5dImpl :  public PeriodicGaugeImpl< GaugeImplTypes< S,Nrepresentation> > { 
   public:
       
+  static const int Dimension = Nrepresentation;
+  const bool LsVectorised=true;
+  typedef _Coeff_t Coeff_t;      
   typedef PeriodicGaugeImpl<GaugeImplTypes<S, Nrepresentation> > Gimpl;
   
   INHERIT_GIMPL_TYPES(Gimpl);
@@ -221,7 +238,7 @@ PARALLEL_FOR_LOOP
   
   ImplParams Params;
   
-      DomainWallRedBlack5dImpl(const ImplParams &p= ImplParams()) : Params(p) {}; 
+  DomainWallVec5dImpl(const ImplParams &p = ImplParams()) : Params(p){};
       
   bool overlapCommsCompute(void) { return false; };
       
@@ -229,8 +246,10 @@ PARALLEL_FOR_LOOP
   inline void loadLinkElement(Simd &reg, ref &memory) {
     vsplat(reg, memory);
   }
-      inline void multLink(SiteHalfSpinor &phi,const SiteDoubledGaugeField &U,const SiteHalfSpinor &chi,int mu,StencilEntry *SE,StencilImpl &St)
+
-      {
+  inline void multLink(SiteHalfSpinor &phi, const SiteDoubledGaugeField &U,
+		       const SiteHalfSpinor &chi, int mu, StencilEntry *SE,
+		       StencilImpl &St) {
     SiteGaugeLink UU;
     for (int i = 0; i < Nrepresentation; i++) {
       for (int j = 0; j < Nrepresentation; j++) {
@@ -265,28 +284,32 @@ PARALLEL_FOR_LOOP
       
       pokeLocalSite(ScalarUds, Uds, lcoor);
     }
-
   }
       
-      inline void InsertForce4D(GaugeField &mat, FermionField &Btilde, FermionField &A,int mu){
+  inline void InsertForce4D(GaugeField &mat, FermionField &Btilde,FermionField &A, int mu) 
+  {
     assert(0);
   }
       
-      inline void InsertForce5D(GaugeField &mat, FermionField &Btilde, FermionField &Atilde,int mu){
+  inline void InsertForce5D(GaugeField &mat, FermionField &Btilde,FermionField &Atilde, int mu) 
+  {
 	assert(0);
   }
-
 };
     
-
     ////////////////////////////////////////////////////////////////////////////////////////
     // Flavour doubled spinors; is Gparity the only? what about C*?
     ////////////////////////////////////////////////////////////////////////////////////////
     
-    template<class S,int Nrepresentation>
+template <class S, int Nrepresentation,class _Coeff_t = RealD>
 class GparityWilsonImpl : public ConjugateGaugeImpl<GaugeImplTypes<S, Nrepresentation> > {
  public:
 
+ static const int Dimension = Nrepresentation;
+
+ const bool LsVectorised=false;
+
+ typedef _Coeff_t Coeff_t;
  typedef ConjugateGaugeImpl< GaugeImplTypes<S,Nrepresentation> > Gimpl;
  
  INHERIT_GIMPL_TYPES(Gimpl);
@@ -313,8 +336,11 @@ PARALLEL_FOR_LOOP
 
  bool overlapCommsCompute(void) { return Params.overlapCommsCompute; };
 
-      // provide the multiply by link that is differentiated between Gparity (with flavour index) and non-Gparity
+ // provide the multiply by link that is differentiated between Gparity (with
-      inline void multLink(SiteHalfSpinor &phi,const SiteDoubledGaugeField &U,const SiteHalfSpinor &chi,int mu,StencilEntry *SE,StencilImpl &St){
+ // flavour index) and non-Gparity
+ inline void multLink(SiteHalfSpinor &phi, const SiteDoubledGaugeField &U,
+		      const SiteHalfSpinor &chi, int mu, StencilEntry *SE,
+		      StencilImpl &St) {
 
   typedef SiteHalfSpinor vobj;
    typedef typename SiteHalfSpinor::scalar_object sobj;
@@ -384,7 +410,6 @@ PARALLEL_FOR_LOOP
 
  inline void DoubleStore(GridBase *GaugeGrid,DoubledGaugeField &Uds,const GaugeField &Umu)
  {
-	
    conformable(Uds._grid,GaugeGrid);
    conformable(Umu._grid,GaugeGrid);
    
@@ -394,7 +419,6 @@ PARALLEL_FOR_LOOP
    
    Lattice<iScalar<vInteger> > coor(GaugeGrid);
 	
-	
    for(int mu=0;mu<Nd;mu++){
 	  
      LatticeCoordinate(coor,mu);
@@ -408,7 +432,6 @@ PARALLEL_FOR_LOOP
        Uconj = where(coor==neglink,-Uconj,Uconj);
      }
 	  
-	  
 PARALLEL_FOR_LOOP
      for(auto ss=U.begin();ss<U.end();ss++){
        Uds[ss](0)(mu) = U[ss]();
@@ -441,6 +464,7 @@ PARALLEL_FOR_LOOP
    }
  }
       
+      
  inline void InsertForce4D(GaugeField &mat, FermionField &Btilde, FermionField &A, int mu) {
 
    // DhopDir provides U or Uconj depending on coor/flavour.
@@ -454,6 +478,7 @@ PARALLEL_FOR_LOOP
    PokeIndex<LorentzIndex>(mat, link, mu);
    return;
  }
+      
  inline void InsertForce5D(GaugeField &mat, FermionField &Btilde, FermionField &Atilde, int mu) {
 
    int Ls = Btilde._grid->_fdimensions[0];
@@ -471,20 +496,37 @@ PARALLEL_FOR_LOOP
    PokeIndex<LorentzIndex>(mat, tmp, mu);
    return;
  }
+
 };
 
-    typedef WilsonImpl<vComplex ,Nc> WilsonImplR; // Real.. whichever prec
+ typedef WilsonImpl<vComplex,  FundamentalRepresentation > WilsonImplR;   // Real.. whichever prec
-    typedef WilsonImpl<vComplexF,Nc> WilsonImplF; // Float
+ typedef WilsonImpl<vComplexF, FundamentalRepresentation > WilsonImplF;  // Float
-    typedef WilsonImpl<vComplexD,Nc> WilsonImplD; // Double
+ typedef WilsonImpl<vComplexD, FundamentalRepresentation > WilsonImplD;  // Double
 
-    typedef DomainWallRedBlack5dImpl<vComplex ,Nc> DomainWallRedBlack5dImplR; // Real.. whichever prec
+ typedef WilsonImpl<vComplex,  FundamentalRepresentation, ComplexD > ZWilsonImplR; // Real.. whichever prec
-    typedef DomainWallRedBlack5dImpl<vComplexF,Nc> DomainWallRedBlack5dImplF; // Float
+ typedef WilsonImpl<vComplexF, FundamentalRepresentation, ComplexD > ZWilsonImplF; // Float
-    typedef DomainWallRedBlack5dImpl<vComplexD,Nc> DomainWallRedBlack5dImplD; // Double
+ typedef WilsonImpl<vComplexD, FundamentalRepresentation, ComplexD > ZWilsonImplD; // Double
+ 
+ typedef WilsonImpl<vComplex,  AdjointRepresentation > WilsonAdjImplR;   // Real.. whichever prec
+ typedef WilsonImpl<vComplexF, AdjointRepresentation > WilsonAdjImplF;  // Float
+ typedef WilsonImpl<vComplexD, AdjointRepresentation > WilsonAdjImplD;  // Double
+ 
+ typedef WilsonImpl<vComplex,  TwoIndexSymmetricRepresentation > WilsonTwoIndexSymmetricImplR;   // Real.. whichever prec
+ typedef WilsonImpl<vComplexF, TwoIndexSymmetricRepresentation > WilsonTwoIndexSymmetricImplF;  // Float
+ typedef WilsonImpl<vComplexD, TwoIndexSymmetricRepresentation > WilsonTwoIndexSymmetricImplD;  // Double
+ 
+ typedef DomainWallVec5dImpl<vComplex ,Nc> DomainWallVec5dImplR; // Real.. whichever prec
+ typedef DomainWallVec5dImpl<vComplexF,Nc> DomainWallVec5dImplF; // Float
+ typedef DomainWallVec5dImpl<vComplexD,Nc> DomainWallVec5dImplD; // Double
+ 
+ typedef DomainWallVec5dImpl<vComplex ,Nc,ComplexD> ZDomainWallVec5dImplR; // Real.. whichever prec
+ typedef DomainWallVec5dImpl<vComplexF,Nc,ComplexD> ZDomainWallVec5dImplF; // Float
+ typedef DomainWallVec5dImpl<vComplexD,Nc,ComplexD> ZDomainWallVec5dImplD; // Double
  
  typedef GparityWilsonImpl<vComplex , Nc> GparityWilsonImplR;  // Real.. whichever prec
  typedef GparityWilsonImpl<vComplexF, Nc> GparityWilsonImplF;  // Float
  typedef GparityWilsonImpl<vComplexD, Nc> GparityWilsonImplD;  // Double
 
-  }
+}}
-}
+
 #endif

lib/qcd/action/fermion/MobiusFermion.h

@@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef  GRID_QCD_MOBIUS_FERMION_H
 #define  GRID_QCD_MOBIUS_FERMION_H
 
-#include <Grid.h>
+#include <Grid/Grid.h>
 
 namespace Grid {
 

lib/qcd/action/fermion/MobiusZolotarevFermion.h

@@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef  GRID_QCD_MOBIUS_ZOLOTAREV_FERMION_H
 #define  GRID_QCD_MOBIUS_ZOLOTAREV_FERMION_H
 
-#include <Grid.h>
+#include <Grid/Grid.h>
 
 namespace Grid {
 

lib/qcd/action/fermion/OverlapWilsonCayleyTanhFermion.h

@@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef OVERLAP_WILSON_CAYLEY_TANH_FERMION_H
 #define OVERLAP_WILSON_CAYLEY_TANH_FERMION_H
 
-#include <Grid.h>
+#include <Grid/Grid.h>
 
 namespace Grid {
 

lib/qcd/action/fermion/OverlapWilsonCayleyZolotarevFermion.h

@@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef  OVERLAP_WILSON_CAYLEY_ZOLOTAREV_FERMION_H
 #define  OVERLAP_WILSON_CAYLEY_ZOLOTAREV_FERMION_H
 
-#include <Grid.h>
+#include <Grid/Grid.h>
 
 namespace Grid {
 

lib/qcd/action/fermion/OverlapWilsonContfracTanhFermion.h

@@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef OVERLAP_WILSON_CONTFRAC_TANH_FERMION_H
 #define OVERLAP_WILSON_CONTFRAC_TANH_FERMION_H
 
-#include <Grid.h>
+#include <Grid/Grid.h>
 
 namespace Grid {
 

lib/qcd/action/fermion/OverlapWilsonContfracZolotarevFermion.h

@@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef OVERLAP_WILSON_CONTFRAC_ZOLOTAREV_FERMION_H
 #define OVERLAP_WILSON_CONTFRAC_ZOLOTAREV_FERMION_H
 
-#include <Grid.h>
+#include <Grid/Grid.h>
 
 namespace Grid {
 

lib/qcd/action/fermion/OverlapWilsonPartialFractionTanhFermion.h

@@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef OVERLAP_WILSON_PARTFRAC_TANH_FERMION_H
 #define OVERLAP_WILSON_PARTFRAC_TANH_FERMION_H
 
-#include <Grid.h>
+#include <Grid/Grid.h>
 
 namespace Grid {
 

lib/qcd/action/fermion/OverlapWilsonPartialFractionZolotarevFermion.h

@@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef OVERLAP_WILSON_PARTFRAC_ZOLOTAREV_FERMION_H
 #define OVERLAP_WILSON_PARTFRAC_ZOLOTAREV_FERMION_H
 
-#include <Grid.h>
+#include <Grid/Grid.h>
 
 namespace Grid {
 

lib/qcd/action/fermion/ScaledShamirFermion.h

@@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef  GRID_QCD_SCALED_SHAMIR_FERMION_H
 #define  GRID_QCD_SCALED_SHAMIR_FERMION_H
 
-#include <Grid.h>
+#include <Grid/Grid.h>
 
 namespace Grid {
 

lib/qcd/action/fermion/ShamirZolotarevFermion.h

@@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef  GRID_QCD_SHAMIR_ZOLOTAREV_FERMION_H
 #define  GRID_QCD_SHAMIR_ZOLOTAREV_FERMION_H
 
-#include <Grid.h>
+#include <Grid/Grid.h>
 
 namespace Grid {
 

lib/qcd/action/fermion/WilsonFermion.cc

@@ -25,7 +25,8 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 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
+See the full license in the file "LICENSE" in the top level distribution
+directory
 *************************************************************************************/
 /*  END LEGAL */
 #include <Grid.h>
@@ -33,8 +34,10 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 namespace Grid {
 namespace QCD {
 
-  const std::vector<int> WilsonFermionStatic::directions   ({0,1,2,3, 0, 1, 2, 3});
+const std::vector<int> WilsonFermionStatic::directions({0, 1, 2, 3, 0, 1, 2,
-  const std::vector<int> WilsonFermionStatic::displacements({1,1,1,1,-1,-1,-1,-1});
+                                                        3});
+const std::vector<int> WilsonFermionStatic::displacements({1, 1, 1, 1, -1, -1,
+                                                           -1, -1});
 int WilsonFermionStatic::HandOptDslash;
 
 /////////////////////////////////
@@ -42,28 +45,29 @@ namespace QCD {
 /////////////////////////////////
 
 template <class Impl>
-  WilsonFermion<Impl>::WilsonFermion(GaugeField &_Umu,
+WilsonFermion<Impl>::WilsonFermion(GaugeField &_Umu, GridCartesian &Fgrid,
-				     GridCartesian         &Fgrid,
+                                   GridRedBlackCartesian &Hgrid, RealD _mass,
-				     GridRedBlackCartesian &Hgrid, 
+                                   const ImplParams &p)
-				     RealD _mass,const ImplParams &p) :
+    : Kernels(p),
-        Kernels(p),
       _grid(&Fgrid),
       _cbgrid(&Hgrid),
       Stencil(&Fgrid, npoint, Even, directions, displacements),
-	StencilEven(&Hgrid,npoint,Even,directions,displacements), // source is Even
+      StencilEven(&Hgrid, npoint, Even, directions,
-	StencilOdd (&Hgrid,npoint,Odd ,directions,displacements), // source is Odd
+                  displacements),  // source is Even
+      StencilOdd(&Hgrid, npoint, Odd, directions,
+                 displacements),  // source is Odd
       mass(_mass),
+      Lebesgue(_grid),
+      LebesgueEvenOdd(_cbgrid),
       Umu(&Fgrid),
       UmuEven(&Hgrid),
-	UmuOdd (&Hgrid) 
+      UmuOdd(&Hgrid) {
-  {
   // Allocate the required comms buffer
   ImportGauge(_Umu);
 }
 
 template <class Impl>
-  void WilsonFermion<Impl>::ImportGauge(const GaugeField &_Umu)
+void WilsonFermion<Impl>::ImportGauge(const GaugeField &_Umu) {
-  {
   GaugeField HUmu(_Umu._grid);
   HUmu = _Umu * (-0.5);
   Impl::DoubleStore(GaugeGrid(), Umu, HUmu);
@@ -76,24 +80,21 @@ namespace QCD {
 /////////////////////////////
 
 template <class Impl>
-  RealD WilsonFermion<Impl>::M(const FermionField &in, FermionField &out) 
+RealD WilsonFermion<Impl>::M(const FermionField &in, FermionField &out) {
-  {
   out.checkerboard = in.checkerboard;
   Dhop(in, out, DaggerNo);
   return axpy_norm(out, 4 + mass, in, out);
 }
 
 template <class Impl>
-  RealD WilsonFermion<Impl>::Mdag(const FermionField &in, FermionField &out) 
+RealD WilsonFermion<Impl>::Mdag(const FermionField &in, FermionField &out) {
-  {
   out.checkerboard = in.checkerboard;
   Dhop(in, out, DaggerYes);
   return axpy_norm(out, 4 + mass, in, out);
 }
 
 template <class Impl>
-  void WilsonFermion<Impl>::Meooe(const FermionField &in, FermionField &out) 
+void WilsonFermion<Impl>::Meooe(const FermionField &in, FermionField &out) {
-  {
   if (in.checkerboard == Odd) {
     DhopEO(in, out, DaggerNo);
   } else {
@@ -101,8 +102,7 @@ namespace QCD {
   }
 }
 template <class Impl>
-  void WilsonFermion<Impl>::MeooeDag(const FermionField &in, FermionField &out) 
+void WilsonFermion<Impl>::MeooeDag(const FermionField &in, FermionField &out) {
-  {
   if (in.checkerboard == Odd) {
     DhopEO(in, out, DaggerYes);
   } else {
@@ -130,7 +130,8 @@ namespace QCD {
 }
 
 template <class Impl>
-  void WilsonFermion<Impl>::MooeeInvDag(const FermionField &in, FermionField &out) {
+void WilsonFermion<Impl>::MooeeInvDag(const FermionField &in,
+                                      FermionField &out) {
   out.checkerboard = in.checkerboard;
   MooeeInv(in, out);
 }
@@ -140,12 +141,9 @@ namespace QCD {
 ///////////////////////////////////
 
 template <class Impl>
-  void WilsonFermion<Impl>::DerivInternal(StencilImpl & st,
+void WilsonFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGaugeField &U,
-					  DoubledGaugeField & U,
+                                        GaugeField &mat, const FermionField &A,
-					  GaugeField &mat,
-					  const FermionField &A,
                                         const FermionField &B, int dag) {
-	
   assert((dag == DaggerNo) || (dag == DaggerYes));
 
   Compressor compressor(dag);
@@ -157,7 +155,6 @@ namespace QCD {
   st.HaloExchange(B, compressor);
 
   for (int mu = 0; mu < Nd; mu++) {
-      
     ////////////////////////////////////////////////////////////////////////
     // Flip gamma (1+g)<->(1-g) if dag
     ////////////////////////////////////////////////////////////////////////
@@ -169,20 +166,20 @@ namespace QCD {
     ////////////////////////
     PARALLEL_FOR_LOOP
     for (int sss = 0; sss < B._grid->oSites(); sss++) {
-	  Kernels::DiracOptDhopDir(st,U,st.comm_buf,sss,sss,B,Btilde,mu,gamma);
+      Kernels::DiracOptDhopDir(st, U, st.CommBuf(), sss, sss, B, Btilde, mu,
+                               gamma);
     }
 
     //////////////////////////////////////////////////
     // spin trace outer product
     //////////////////////////////////////////////////
     Impl::InsertForce4D(mat, Btilde, Atilde, mu);
-
   }
 }
 
 template <class Impl>
-  void WilsonFermion<Impl>::DhopDeriv(GaugeField &mat,const FermionField &U,const FermionField &V,int dag)
+void WilsonFermion<Impl>::DhopDeriv(GaugeField &mat, const FermionField &U,
-  {
+                                    const FermionField &V, int dag) {
   conformable(U._grid, _grid);
   conformable(U._grid, V._grid);
   conformable(U._grid, mat._grid);
@@ -193,8 +190,8 @@ PARALLEL_FOR_LOOP
 }
 
 template <class Impl>
-  void WilsonFermion<Impl>::DhopDerivOE(GaugeField &mat,const FermionField &U,const FermionField &V,int dag)
+void WilsonFermion<Impl>::DhopDerivOE(GaugeField &mat, const FermionField &U,
-  {
+                                      const FermionField &V, int dag) {
   conformable(U._grid, _cbgrid);
   conformable(U._grid, V._grid);
   conformable(U._grid, mat._grid);
@@ -207,8 +204,8 @@ PARALLEL_FOR_LOOP
 }
 
 template <class Impl>
-  void WilsonFermion<Impl>::DhopDerivEO(GaugeField &mat,const FermionField &U,const FermionField &V,int dag)
+void WilsonFermion<Impl>::DhopDerivEO(GaugeField &mat, const FermionField &U,
-  {
+                                      const FermionField &V, int dag) {
   conformable(U._grid, _cbgrid);
   conformable(U._grid, V._grid);
   conformable(U._grid, mat._grid);
@@ -220,118 +217,99 @@ PARALLEL_FOR_LOOP
   DerivInternal(StencilOdd, UmuEven, mat, U, V, dag);
 }
 
-
 template <class Impl>
-  void WilsonFermion<Impl>::Dhop(const FermionField &in, FermionField &out,int dag) {
+void WilsonFermion<Impl>::Dhop(const FermionField &in, FermionField &out,
+                               int dag) {
   conformable(in._grid, _grid);  // verifies full grid
   conformable(in._grid, out._grid);
 
   out.checkerboard = in.checkerboard;
 
-    DhopInternal(Stencil,Umu,in,out,dag);
+  DhopInternal(Stencil, Lebesgue, Umu, in, out, dag);
 }
 
 template <class Impl>
-  void WilsonFermion<Impl>::DhopOE(const FermionField &in, FermionField &out,int dag) {
+void WilsonFermion<Impl>::DhopOE(const FermionField &in, FermionField &out,
+                                 int dag) {
   conformable(in._grid, _cbgrid);    // verifies half grid
   conformable(in._grid, out._grid);  // drops the cb check
 
   assert(in.checkerboard == Even);
   out.checkerboard = Odd;
 
-    DhopInternal(StencilEven,UmuOdd,in,out,dag);
+  DhopInternal(StencilEven, LebesgueEvenOdd, UmuOdd, in, out, dag);
 }
 
 template <class Impl>
-  void WilsonFermion<Impl>::DhopEO(const FermionField &in, FermionField &out,int dag) {
+void WilsonFermion<Impl>::DhopEO(const FermionField &in, FermionField &out,
+                                 int dag) {
   conformable(in._grid, _cbgrid);    // verifies half grid
   conformable(in._grid, out._grid);  // drops the cb check
 
   assert(in.checkerboard == Odd);
   out.checkerboard = Even;
 
-    DhopInternal(StencilOdd,UmuEven,in,out,dag);
+  DhopInternal(StencilOdd, LebesgueEvenOdd, UmuEven, in, out, dag);
 }
 
 template <class Impl>
-  void WilsonFermion<Impl>::Mdir (const FermionField &in, FermionField &out,int dir,int disp) {
+void WilsonFermion<Impl>::Mdir(const FermionField &in, FermionField &out,
+                               int dir, int disp) {
   DhopDir(in, out, dir, disp);
 }
 
-
 template <class Impl>
-  void WilsonFermion<Impl>::DhopDir(const FermionField &in, FermionField &out,int dir,int disp){
+void WilsonFermion<Impl>::DhopDir(const FermionField &in, FermionField &out,
-    
+                                  int dir, int disp) {
   int skip = (disp == 1) ? 0 : 1;
   int dirdisp = dir + skip * 4;
   int gamma = dir + (1 - skip) * 4;
 
   DhopDirDisp(in, out, dirdisp, gamma, DaggerNo);
-    
 };
 
 template <class Impl>
-  void WilsonFermion<Impl>::DhopDirDisp(const FermionField &in, FermionField &out,int dirdisp,int gamma,int dag) {
+void WilsonFermion<Impl>::DhopDirDisp(const FermionField &in, FermionField &out,
-    
+                                      int dirdisp, int gamma, int dag) {
   Compressor compressor(dag);
 
   Stencil.HaloExchange(in, compressor);
 
   PARALLEL_FOR_LOOP
   for (int sss = 0; sss < in._grid->oSites(); sss++) {
-	Kernels::DiracOptDhopDir(Stencil,Umu,Stencil.comm_buf,sss,sss,in,out,dirdisp,gamma);
+    Kernels::DiracOptDhopDir(Stencil, Umu, Stencil.CommBuf(), sss, sss, in, out,
+                             dirdisp, gamma);
   }
-    
 };
 
 template <class Impl>
-  void WilsonFermion<Impl>::DhopInternal(StencilImpl & st,DoubledGaugeField & U,
+void WilsonFermion<Impl>::DhopInternal(StencilImpl &st, LebesgueOrder &lo,
-					 const FermionField &in, FermionField &out,int dag) 
+                                       DoubledGaugeField &U,
-  {
+                                       const FermionField &in,
-    DhopInternalCommsThenCompute(st,U,in,out,dag);
+                                       FermionField &out, int dag) {
-  }
-  template<class Impl>
-  void WilsonFermion<Impl>::DhopInternalCommsThenCompute(StencilImpl & st,DoubledGaugeField & U,
-							 const FermionField &in, FermionField &out,int dag) {
-
   assert((dag == DaggerNo) || (dag == DaggerYes));
 
   Compressor compressor(dag);
   st.HaloExchange(in, compressor);
 
   if (dag == DaggerYes) {
-      if( HandOptDslash ) {
     PARALLEL_FOR_LOOP
     for (int sss = 0; sss < in._grid->oSites(); sss++) {
-	  Kernels::DiracOptHandDhopSiteDag(st,U,st.comm_buf,sss,sss,in,out);
+      Kernels::DiracOptDhopSiteDag(st, lo, U, st.CommBuf(), sss, sss, 1, 1, in,
+                                   out);
     }
   } else {
     PARALLEL_FOR_LOOP
     for (int sss = 0; sss < in._grid->oSites(); sss++) {
-	  Kernels::DiracOptDhopSiteDag(st,U,st.comm_buf,sss,sss,in,out);
+      Kernels::DiracOptDhopSite(st, lo, U, st.CommBuf(), sss, sss, 1, 1, in,
-	}
+                                out);
-      }
-    } else {
-      if( HandOptDslash ) {
-PARALLEL_FOR_LOOP
-        for(int sss=0;sss<in._grid->oSites();sss++){
-	  Kernels::DiracOptHandDhopSite(st,U,st.comm_buf,sss,sss,in,out);
-	}
-      } else { 
-PARALLEL_FOR_LOOP
-        for(int sss=0;sss<in._grid->oSites();sss++){
-	  Kernels::DiracOptDhopSite(st,U,st.comm_buf,sss,sss,in,out);
-	}
     }
   }
 };
 
- 
 FermOpTemplateInstantiate(WilsonFermion);
+AdjointFermOpTemplateInstantiate(WilsonFermion);
+TwoIndexFermOpTemplateInstantiate(WilsonFermion);
 GparityFermOpTemplateInstantiate(WilsonFermion);
-
+}
-
+}
-}}
-
-
-

lib/qcd/action/fermion/WilsonFermion.h

@@ -24,7 +24,8 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 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
+See the full license in the file "LICENSE" in the top level distribution
+directory
 *************************************************************************************/
 /*  END LEGAL */
 #ifndef GRID_QCD_WILSON_FERMION_H
@@ -44,8 +45,7 @@ namespace Grid {
 };
 
 template <class Impl>
-    class WilsonFermion : public WilsonKernels<Impl>, public WilsonFermionStatic
+class WilsonFermion : public WilsonKernels<Impl>, public WilsonFermionStatic {
-    {
  public:
   INHERIT_IMPL_TYPES(Impl);
   typedef WilsonKernels<Impl> Kernels;
@@ -82,10 +82,12 @@ namespace Grid {
   // Derivative interface
   ////////////////////////
   // Interface calls an internal routine
-      void DhopDeriv(GaugeField &mat,const FermionField &U,const FermionField &V,int dag);
+  void DhopDeriv(GaugeField &mat, const FermionField &U, const FermionField &V,
-      void DhopDerivOE(GaugeField &mat,const FermionField &U,const FermionField &V,int dag);
+                 int dag);
-      void DhopDerivEO(GaugeField &mat,const FermionField &U,const FermionField &V,int dag);
+  void DhopDerivOE(GaugeField &mat, const FermionField &U,
-
+                   const FermionField &V, int dag);
+  void DhopDerivEO(GaugeField &mat, const FermionField &U,
+                   const FermionField &V, int dag);
 
   ///////////////////////////////////////////////////////////////
   // non-hermitian hopping term; half cb or both
@@ -99,31 +101,22 @@ namespace Grid {
   ///////////////////////////////////////////////////////////////
   void Mdir(const FermionField &in, FermionField &out, int dir, int disp);
   void DhopDir(const FermionField &in, FermionField &out, int dir, int disp);
-      void DhopDirDisp(const FermionField &in, FermionField &out,int dirdisp,int gamma,int dag) ;
+  void DhopDirDisp(const FermionField &in, FermionField &out, int dirdisp,
+                   int gamma, int dag);
 
   ///////////////////////////////////////////////////////////////
   // Extra methods added by derived
   ///////////////////////////////////////////////////////////////
-      void DerivInternal(StencilImpl & st,
+  void DerivInternal(StencilImpl &st, DoubledGaugeField &U, GaugeField &mat,
-			 DoubledGaugeField & U,
+                     const FermionField &A, const FermionField &B, int dag);
-			 GaugeField &mat,
-			 const FermionField &A,
-			 const FermionField &B,
-			 int dag);
 
-      void DhopInternal(StencilImpl & st,DoubledGaugeField & U,
+  void DhopInternal(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
-			const FermionField &in, FermionField &out,int dag) ;
-
-      void DhopInternalCommsThenCompute(StencilImpl & st,DoubledGaugeField & U,
                     const FermionField &in, FermionField &out, int dag);
 
   // Constructor
-      WilsonFermion(GaugeField &_Umu,
+  WilsonFermion(GaugeField &_Umu, GridCartesian &Fgrid,
-		    GridCartesian         &Fgrid,
+                GridRedBlackCartesian &Hgrid, RealD _mass,
-		    GridRedBlackCartesian &Hgrid, 
+                const ImplParams &p = ImplParams());
-		    RealD _mass,
-		    const ImplParams &p= ImplParams()
-		    ) ;
 
   // DoubleStore impl dependent
   void ImportGauge(const GaugeField &_Umu);
@@ -134,7 +127,6 @@ namespace Grid {
 
   //    protected:
  public:
-
   RealD mass;
 
   GridBase *_grid;
@@ -150,11 +142,14 @@ namespace Grid {
   DoubledGaugeField UmuEven;
   DoubledGaugeField UmuOdd;
 
+  LebesgueOrder Lebesgue;
+  LebesgueOrder LebesgueEvenOdd;
 };
 
 typedef WilsonFermion<WilsonImplF> WilsonFermionF;
 typedef WilsonFermion<WilsonImplD> WilsonFermionD;
 
+
 }
 }
 #endif

lib/qcd/action/fermion/WilsonFermion5D.cc

@@ -1,4 +1,3 @@
-
 /*************************************************************************************
 
     Grid physics library, www.github.com/paboyle/Grid 
@@ -39,8 +38,6 @@ namespace QCD {
 // S-direction is INNERMOST and takes no part in the parity.
 const std::vector<int> WilsonFermion5DStatic::directions   ({1,2,3,4, 1, 2, 3, 4});
 const std::vector<int> WilsonFermion5DStatic::displacements({1,1,1,1,-1,-1,-1,-1});
-int WilsonFermion5DStatic::HandOptDslash;
-int WilsonFermion5DStatic::AsmOptDslash;
 
   // 5d lattice for DWF.
 template<class Impl>
@@ -65,6 +62,41 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
   Lebesgue(_FourDimGrid),
   LebesgueEvenOdd(_FourDimRedBlackGrid)
 {
+  if (Impl::LsVectorised) { 
+
+    int nsimd = Simd::Nsimd();
+    
+    // some assertions
+    assert(FiveDimGrid._ndimension==5);
+    assert(FiveDimRedBlackGrid._ndimension==5);
+    assert(FiveDimRedBlackGrid._checker_dim==1); // Don't checker the s direction
+    assert(FourDimGrid._ndimension==4);
+
+    // Dimension zero of the five-d is the Ls direction
+    Ls=FiveDimGrid._fdimensions[0];
+    assert(FiveDimGrid._processors[0]         ==1);
+    assert(FiveDimGrid._simd_layout[0]        ==nsimd);
+
+    assert(FiveDimRedBlackGrid._fdimensions[0]==Ls);
+    assert(FiveDimRedBlackGrid._processors[0] ==1);
+    assert(FiveDimRedBlackGrid._simd_layout[0]==nsimd);
+
+    // Other dimensions must match the decomposition of the four-D fields 
+    for(int d=0;d<4;d++){
+      assert(FiveDimRedBlackGrid._fdimensions[d+1]==FourDimGrid._fdimensions[d]);
+      assert(FiveDimRedBlackGrid._processors[d+1] ==FourDimGrid._processors[d]);
+      
+      assert(FourDimGrid._simd_layout[d]=1);
+      assert(FourDimRedBlackGrid._simd_layout[d]=1);
+      assert(FiveDimRedBlackGrid._simd_layout[d+1]==1);
+
+      assert(FiveDimGrid._fdimensions[d+1]        ==FourDimGrid._fdimensions[d]);
+      assert(FiveDimGrid._processors[d+1]         ==FourDimGrid._processors[d]);
+      assert(FiveDimGrid._simd_layout[d+1]        ==FourDimGrid._simd_layout[d]);
+    }
+
+  } else {
+
     // some assertions
     assert(FiveDimGrid._ndimension==5);
     assert(FourDimGrid._ndimension==4);
@@ -95,37 +127,18 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
       assert(FiveDimGrid._processors[d+1]         ==FourDimGrid._processors[d]);
       assert(FiveDimGrid._simd_layout[d+1]        ==FourDimGrid._simd_layout[d]);
     }
+  }
     
   // Allocate the required comms buffer
   ImportGauge(_Umu);
-  alltime=0;
-  commtime=0;
-  jointime=0;
-  dslashtime=0;
-  dslash1time=0;
 }
-
+  /*
 template<class Impl>
 WilsonFermion5D<Impl>::WilsonFermion5D(int simd,GaugeField &_Umu,
                GridCartesian         &FiveDimGrid,
                GridRedBlackCartesian &FiveDimRedBlackGrid,
                GridCartesian         &FourDimGrid,
-				       GridRedBlackCartesian &FourDimRedBlackGrid,
                RealD _M5,const ImplParams &p) :
-  Kernels(p),
-  _FiveDimGrid        (&FiveDimGrid),
-  _FiveDimRedBlackGrid(&FiveDimRedBlackGrid),
-  _FourDimGrid        (&FourDimGrid),
-  _FourDimRedBlackGrid(&FourDimRedBlackGrid),
-  Stencil    (_FiveDimGrid,npoint,Even,directions,displacements),
-  StencilEven(_FiveDimRedBlackGrid,npoint,Even,directions,displacements), // source is Even
-  StencilOdd (_FiveDimRedBlackGrid,npoint,Odd ,directions,displacements), // source is Odd
-  M5(_M5),
-  Umu(_FourDimGrid),
-  UmuEven(_FourDimRedBlackGrid),
-  UmuOdd (_FourDimRedBlackGrid),
-  Lebesgue(_FourDimGrid),
-  LebesgueEvenOdd(_FourDimRedBlackGrid)
 {
   int nsimd = Simd::Nsimd();
 
@@ -134,7 +147,6 @@ WilsonFermion5D<Impl>::WilsonFermion5D(int simd, GaugeField &_Umu,
   assert(FiveDimRedBlackGrid._ndimension==5);
   assert(FiveDimRedBlackGrid._checker_dim==0); // Checkerboard the s-direction
   assert(FourDimGrid._ndimension==4);
-  assert(FourDimRedBlackGrid._ndimension==4);
 
   // Dimension zero of the five-d is the Ls direction
   Ls=FiveDimGrid._fdimensions[0];
@@ -147,15 +159,10 @@ WilsonFermion5D<Impl>::WilsonFermion5D(int simd, GaugeField &_Umu,
 
   // Other dimensions must match the decomposition of the four-D fields 
   for(int d=0;d<4;d++){
-    assert(FourDimRedBlackGrid._fdimensions[d]  ==FourDimGrid._fdimensions[d]);
     assert(FiveDimRedBlackGrid._fdimensions[d+1]==FourDimGrid._fdimensions[d]);
-
-    assert(FourDimRedBlackGrid._processors[d]   ==FourDimGrid._processors[d]);
     assert(FiveDimRedBlackGrid._processors[d+1] ==FourDimGrid._processors[d]);
 
     assert(FourDimGrid._simd_layout[d]=1);
-    assert(FourDimRedBlackGrid._simd_layout[d]  ==1);
-    assert(FourDimRedBlackGrid._simd_layout[d]  ==1);
     assert(FiveDimRedBlackGrid._simd_layout[d+1]==1);
 
     assert(FiveDimGrid._fdimensions[d+1]        ==FourDimGrid._fdimensions[d]);
@@ -163,8 +170,68 @@ WilsonFermion5D<Impl>::WilsonFermion5D(int simd, GaugeField &_Umu,
     assert(FiveDimGrid._simd_layout[d+1]        ==FourDimGrid._simd_layout[d]);
   }
 
-  // Allocate the required comms buffer
+  {
-  ImportGauge(_Umu);
+  }
+}  
+  */
+     
+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;
+
+  if ( DhopCalls > 0 ) {
+    std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D Number of Dhop Calls     : " << DhopCalls   << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D Total Communication time : " << DhopCommTime<< " us" << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D CommTime/Calls           : " << DhopCommTime / DhopCalls << " us" << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D Total Compute time       : " << DhopComputeTime << " us" << std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D ComputeTime/Calls        : " << DhopComputeTime / DhopCalls << " us" << std::endl;
+
+    RealD mflops = 1344*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;
+
+   }
+
+  if ( DerivCalls > 0 ) {
+    std::cout << GridLogMessage << "#### Deriv calls report "<< std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D Number of Deriv Calls    : " <<DerivCalls <<std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D Total Communication time : " <<DerivCommTime <<" us"<<std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D CommTime/Calls           : " <<DerivCommTime/DerivCalls<<" us" <<std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D Total Compute time       : " <<DerivComputeTime <<" us"<<std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D ComputeTime/Calls        : " <<DerivComputeTime/DerivCalls<<" us" <<std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D Total Dhop Compute time  : " <<DerivDhopComputeTime <<" us"<<std::endl;
+    std::cout << GridLogMessage << "WilsonFermion5D Dhop ComputeTime/Calls   : " <<DerivDhopComputeTime/DerivCalls<<" us" <<std::endl;
+    
+    RealD mflops = 144*volume*DerivCalls/DerivDhopComputeTime;
+    std::cout << GridLogMessage << "Average mflops/s per call                : " << mflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per node       : " << mflops/NP << std::endl;
+  }
+
+  if (DerivCalls > 0 || DhopCalls > 0){
+    std::cout << GridLogMessage << "WilsonFermion5D Stencil"<<std::endl;  Stencil.Report();
+    std::cout << GridLogMessage << "WilsonFermion5D StencilEven"<<std::endl;  StencilEven.Report();
+    std::cout << GridLogMessage << "WilsonFermion5D StencilOdd"<<std::endl;  StencilOdd.Report();
+  }
+}
+
+template<class Impl>
+void WilsonFermion5D<Impl>::ZeroCounters(void) {
+  DhopCalls       = 0;
+  DhopCommTime    = 0;
+  DhopComputeTime = 0;
+
+  DerivCalls       = 0;
+  DerivCommTime    = 0;
+  DerivComputeTime = 0;
+  DerivDhopComputeTime = 0;
+
+  Stencil.ZeroCounters();
+  StencilEven.ZeroCounters();
+  StencilOdd.ZeroCounters();
 }
 
 
@@ -201,7 +268,7 @@ PARALLEL_FOR_LOOP
     for(int s=0;s<Ls;s++){
       int sU=ss;
       int sF = s+Ls*sU; 
-      Kernels::DiracOptDhopDir(Stencil,Umu,Stencil.comm_buf,sF,sU,in,out,dirdisp,gamma);
+      Kernels::DiracOptDhopDir(Stencil,Umu,Stencil.CommBuf(),sF,sU,in,out,dirdisp,gamma);
     }
   }
 };
@@ -214,6 +281,7 @@ void WilsonFermion5D<Impl>::DerivInternal(StencilImpl & st,
             const FermionField &B,
             int dag)
 {
+  DerivCalls++;
   assert((dag==DaggerNo) ||(dag==DaggerYes));
 
   conformable(st._grid,A._grid);
@@ -224,12 +292,14 @@ void WilsonFermion5D<Impl>::DerivInternal(StencilImpl & st,
   FermionField Btilde(B._grid);
   FermionField Atilde(B._grid);
 
+  DerivCommTime-=usecond();
   st.HaloExchange(B,compressor);
+  DerivCommTime+=usecond();
 
   Atilde=A;
 
+  DerivComputeTime-=usecond();
   for (int mu = 0; mu < Nd; mu++) {
-      
     ////////////////////////////////////////////////////////////////////////
     // Flip gamma if dag
     ////////////////////////////////////////////////////////////////////////
@@ -240,6 +310,7 @@ void WilsonFermion5D<Impl>::DerivInternal(StencilImpl & st,
     // Call the single hop
     ////////////////////////
 
+    DerivDhopComputeTime -= usecond();
     PARALLEL_FOR_LOOP
     for (int sss = 0; sss < U._grid->oSites(); sss++) {
       for (int s = 0; s < Ls; s++) {
@@ -249,19 +320,17 @@ PARALLEL_FOR_LOOP
         assert(sF < B._grid->oSites());
         assert(sU < U._grid->oSites());
 
-	Kernels::DiracOptDhopDir(st,U,st.comm_buf,sF,sU,B,Btilde,mu,gamma);
+        Kernels::DiracOptDhopDir(st, U, st.CommBuf(), sF, sU, B, Btilde, mu, gamma);
 
         ////////////////////////////
         // spin trace outer product
         ////////////////////////////
-
       }
-
     }
-
+    DerivDhopComputeTime += usecond();
     Impl::InsertForce5D(mat, Btilde, Atilde, mu);
-
   }
+  DerivComputeTime += usecond();
 }
 
 template<class Impl>
@@ -297,30 +366,6 @@ void WilsonFermion5D<Impl>::DhopDerivEO(GaugeField &mat,
 }
 
 
-template<class Impl>
-void WilsonFermion5D<Impl>::Report(void)
-{
-  std::cout<<GridLogMessage << "******************** WilsonFermion"<<std::endl;
-  std::cout<<GridLogMessage << "Wilson5d      time "<<alltime <<" us"<<std::endl;
-  std::cout<<GridLogMessage << "HaloBegin     time "<<commtime <<" us"<<std::endl;
-  std::cout<<GridLogMessage << "Dslash        time "<<dslashtime<<" us"<<std::endl;
-  std::cout<<GridLogMessage << "Dslash1       time "<<dslash1time<<" us"<<std::endl;
-  std::cout<<GridLogMessage << "HaloComplete  time "<<jointime<<" us"<<std::endl;
-  std::cout<<GridLogMessage << "******************** Stencil"<<std::endl;
-  std::cout<<GridLogMessage << "Stencil all gather      time "<<Stencil.halogtime<<" us"<<std::endl;
-  std::cout<<GridLogMessage << "Stencil nosplice gather time "<<Stencil.nosplicetime<<" us"<<std::endl;
-  std::cout<<GridLogMessage << "Stencil splice   gather time "<<Stencil.splicetime<<" us"<<std::endl;
-  std::cout<<GridLogMessage << "********************"<<std::endl;
-  std::cout<<GridLogMessage << "Stencil gather        "<<Stencil.gathertime<<" us"<<std::endl;
-  std::cout<<GridLogMessage << "Stencil gather simd   "<<Stencil.gathermtime<<" us"<<std::endl;
-  std::cout<<GridLogMessage << "Stencil merge  simd   "<<Stencil.mergetime<<" us"<<std::endl;
-  std::cout<<GridLogMessage << "Stencil spin   simd   "<<Stencil.spintime<<" us"<<std::endl;
-  std::cout<<GridLogMessage << "********************"<<std::endl;
-  std::cout<<GridLogMessage << "Stencil MB/s          "<<(double)Stencil.comms_bytes/Stencil.commtime<<std::endl;
-  std::cout<<GridLogMessage << "Stencil comm     time "<<Stencil.commtime<<" us"<<std::endl;
-  std::cout<<GridLogMessage << "Stencil join     time "<<Stencil.jointime<<" us"<<std::endl;
-  std::cout<<GridLogMessage << "********************"<<std::endl;
-}
 template<class Impl>
 void WilsonFermion5D<Impl>::DhopDerivOE(GaugeField &mat,
 					const FermionField &A,
@@ -342,96 +387,62 @@ template<class Impl>
 void WilsonFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOrder &lo,
 					 DoubledGaugeField & U,
 					 const FermionField &in, FermionField &out,int dag)
-{
-    DhopInternalCommsThenCompute(st,lo,U,in,out,dag);
-}
-
-template<class Impl>
-void WilsonFermion5D<Impl>::DhopInternalCommsThenCompute(StencilImpl & st, LebesgueOrder &lo,
-					 DoubledGaugeField & U,
-					 const FermionField &in, FermionField &out,int dag)
 {
   //  assert((dag==DaggerNo) ||(dag==DaggerYes));
-  alltime-=usecond();
   Compressor compressor(dag);
 
-  // Assume balanced KMP_AFFINITY; this is forced in GridThread.h
   int LLs = in._grid->_rdimensions[0];
   
-  commtime -=usecond();
+  DhopCommTime-=usecond();
-  //  auto handle = st.HaloExchangeBegin(in,compressor);
-  //  st.HaloExchangeComplete(handle);
   st.HaloExchange(in,compressor);
-  commtime +=usecond();
+  DhopCommTime+=usecond();
-
-  jointime -=usecond();
-  jointime +=usecond();
   
+  DhopComputeTime-=usecond();
   // Dhop takes the 4d grid from U, and makes a 5d index for fermion
-  // Not loop ordering and data layout.
-  // Designed to create 
-  // - per thread reuse in L1 cache for U
-  // - 8 linear access unit stride streams per thread for Fermion for hw prefetchable.
-  dslashtime -=usecond();
   if (dag == DaggerYes) {
-    if( this->HandOptDslash ) {
     PARALLEL_FOR_LOOP
     for (int ss = 0; ss < U._grid->oSites(); ss++) {
-	for(int s=0;s<LLs;s++){
       int sU = ss;
-	  int sF = s+LLs*sU;
+      int sF = LLs * sU;
-	  Kernels::DiracOptHandDhopSiteDag(st,U,st.comm_buf,sF,sU,in,out);
+      Kernels::DiracOptDhopSiteDag(st, lo, U, st.CommBuf(), sF, sU, LLs, 1, in, out);
     }
-      }
+#ifdef AVX512
-    } else { 
+  } else if (stat.is_init() ) {
-PARALLEL_FOR_LOOP
-      for(int ss=0;ss<U._grid->oSites();ss++){
-	for(int s=0;s<LLs;s++){
-	  int sU=ss;
-	  int sF = s+LLs*sU;
-	  Kernels::DiracOptDhopSiteDag(st,U,st.comm_buf,sF,sU,in,out);
-	}
-      }
-    }
-  } else {
-    if( this->AsmOptDslash ) {
 
-PARALLEL_FOR_LOOP
+    int nthreads;
+    stat.start();
+#pragma omp parallel
+    {
+#pragma omp master
+    nthreads = omp_get_num_threads();
+    int mythread = omp_get_thread_num();
+    stat.enter(mythread);
+#pragma omp for nowait
     for(int ss=0;ss<U._grid->oSites();ss++) {
-	for(int s=0;s<LLs;s++){
       int sU=ss;
-	  int sF = s+LLs*sU;
+      int sF=LLs*sU;
-	  Kernels::DiracOptAsmDhopSite(st,U,st.comm_buf,sF,sU,in,out);
+      Kernels::DiracOptDhopSite(st,lo,U,st.CommBuf(),sF,sU,LLs,1,in,out);
-	}
-      }
-    } else if( this->HandOptDslash ) {
-PARALLEL_FOR_LOOP     
-      for(int ss=0;ss<U._grid->oSites();ss++){
-	for(int s=0;s<LLs;s++){
-	  int sU=ss;
-	  int sF = s+LLs*sU;
-	  Kernels::DiracOptHandDhopSite(st,U,st.comm_buf,sF,sU,in,out);
     }
+    stat.exit(mythread);
     }
+    stat.accum(nthreads);
+#endif
   } else {
     PARALLEL_FOR_LOOP
     for (int ss = 0; ss < U._grid->oSites(); ss++) {
-	for(int s=0;s<LLs;s++){
       int sU = ss;
-	  int sF = s+LLs*sU; 
+      int sF = LLs * sU;
-	  Kernels::DiracOptDhopSite(st,U,st.comm_buf,sF,sU,in,out);
+      Kernels::DiracOptDhopSite(st,lo,U,st.CommBuf(),sF,sU,LLs,1,in,out);
     }
   }
-    }
+  DhopComputeTime+=usecond();
-  }
-  dslashtime +=usecond();
-  alltime+=usecond();
 }
 
 
 template<class Impl>
 void WilsonFermion5D<Impl>::DhopOE(const FermionField &in, FermionField &out,int dag)
 {
+  DhopCalls++;
   conformable(in._grid,FermionRedBlackGrid());    // verifies half grid
   conformable(in._grid,out._grid); // drops the cb check
 
@@ -443,6 +454,7 @@ void WilsonFermion5D<Impl>::DhopOE(const FermionField &in, FermionField &out,int
 template<class Impl>
 void WilsonFermion5D<Impl>::DhopEO(const FermionField &in, FermionField &out,int dag)
 {
+  DhopCalls++;
   conformable(in._grid,FermionRedBlackGrid());    // verifies half grid
   conformable(in._grid,out._grid); // drops the cb check
 
@@ -454,6 +466,7 @@ void WilsonFermion5D<Impl>::DhopEO(const FermionField &in, FermionField &out,int
 template<class Impl>
 void WilsonFermion5D<Impl>::Dhop(const FermionField &in, FermionField &out,int dag)
 {
+  DhopCalls+=2;
   conformable(in._grid,FermionGrid()); // verifies full grid
   conformable(in._grid,out._grid);
 
@@ -471,8 +484,6 @@ void WilsonFermion5D<Impl>::DW(const FermionField &in, FermionField &out,int dag
 
 FermOpTemplateInstantiate(WilsonFermion5D);
 GparityFermOpTemplateInstantiate(WilsonFermion5D);
-template class WilsonFermion5D<DomainWallRedBlack5dImplF>;		
-template class WilsonFermion5D<DomainWallRedBlack5dImplD>;
   
 }}
 

lib/qcd/action/fermion/WilsonFermion5D.h

@@ -31,8 +31,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #ifndef  GRID_QCD_WILSON_FERMION_5D_H
 #define  GRID_QCD_WILSON_FERMION_5D_H
 
-namespace Grid {
+#include <Grid/Stat.h>
 
+namespace Grid {
 namespace QCD {
 
   ////////////////////////////////////////////////////////////////////////////////
@@ -49,8 +50,6 @@ namespace Grid {
   class WilsonFermion5DStatic { 
   public:
     // S-direction is INNERMOST and takes no part in the parity.
-      static int AsmOptDslash; // these are a temporary hack
-      static int HandOptDslash; // these are a temporary hack
     static const std::vector<int> directions;
     static const std::vector<int> displacements;
     const int npoint = 8;
@@ -62,11 +61,19 @@ namespace Grid {
   public:
     INHERIT_IMPL_TYPES(Impl);
     typedef WilsonKernels<Impl> Kernels;
-     double alltime;
+    PmuStat stat;
-     double jointime;
+    
-     double commtime;
+    void Report(void);
-     double dslashtime;
+    void ZeroCounters(void);
-     double dslash1time;
+    double DhopCalls;
+    double DhopCommTime;
+    double DhopComputeTime;
+    
+    double DerivCalls;
+    double DerivCommTime;
+    double DerivComputeTime;
+    double DerivDhopComputeTime;
+    
     ///////////////////////////////////////////////////////////////
     // Implement the abstract base
     ///////////////////////////////////////////////////////////////
@@ -122,13 +129,6 @@ namespace Grid {
 		      FermionField &out,
 		      int dag);
     
-      void DhopInternalCommsThenCompute(StencilImpl & st,
-			LebesgueOrder &lo,
-			DoubledGaugeField &U,
-			const FermionField &in, 
-			FermionField &out,
-			int dag);
-
     // Constructors
     WilsonFermion5D(GaugeField &_Umu,
 		    GridCartesian         &FiveDimGrid,
@@ -138,18 +138,18 @@ namespace Grid {
 		    double _M5,const ImplParams &p= ImplParams());
     
     // Constructors
+    /*
       WilsonFermion5D(int simd, 
       GaugeField &_Umu,
       GridCartesian         &FiveDimGrid,
       GridRedBlackCartesian &FiveDimRedBlackGrid,
       GridCartesian         &FourDimGrid,
-		      GridRedBlackCartesian &FourDimRedBlackGrid,
       double _M5,const ImplParams &p= ImplParams());
+    */
     
     // DoubleStore
     void ImportGauge(const GaugeField &_Umu);
     
-      void Report(void);
     ///////////////////////////////////////////////////////////////
     // Data members require to support the functionality
     ///////////////////////////////////////////////////////////////
@@ -181,7 +181,7 @@ namespace Grid {
     std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> >  comm_buf;
     
   };
-  }
+
-}
+}}
 
 #endif