No compile on comms == none fix

Think this is nearly there

.travis.yml

@@ -9,68 +9,6 @@ matrix:
     - os:        osx
       osx_image: xcode8.3
       compiler: clang
-    - compiler: gcc
-      dist: trusty
-      sudo: required
-      addons:
-        apt:
-          sources:
-            - ubuntu-toolchain-r-test
-          packages:
-            - g++-4.9
-            - libmpfr-dev
-            - libgmp-dev
-            - libmpc-dev
-            - libopenmpi-dev
-            - openmpi-bin
-            - binutils-dev
-      env: VERSION=-4.9
-    - compiler: gcc
-      dist: trusty
-      sudo: required
-      addons:
-        apt:
-          sources:
-            - ubuntu-toolchain-r-test
-          packages:
-            - g++-5
-            - libmpfr-dev
-            - libgmp-dev
-            - libmpc-dev
-            - libopenmpi-dev
-            - openmpi-bin
-            - binutils-dev
-      env: VERSION=-5
-    - compiler: clang
-      dist: trusty
-      addons:
-        apt:
-          sources:
-            - ubuntu-toolchain-r-test
-          packages:
-            - g++-4.8
-            - libmpfr-dev
-            - libgmp-dev
-            - libmpc-dev
-            - libopenmpi-dev
-            - openmpi-bin
-            - binutils-dev
-      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
-      dist: trusty
-      addons:
-        apt:
-          sources:
-            - ubuntu-toolchain-r-test
-          packages:
-            - g++-4.8
-            - libmpfr-dev
-            - libgmp-dev
-            - libmpc-dev
-            - libopenmpi-dev
-            - openmpi-bin
-            - binutils-dev
-      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`
@@ -106,9 +44,3 @@ script:
     - make -j4
     - ./benchmarks/Benchmark_dwf --threads 1 --debug-signals
     - make check
-    - echo make clean
-    - if [[ "$TRAVIS_OS_NAME" == "linux" ]] && [[ "$CC" == "clang" ]]; then ../configure --enable-precision=single --enable-simd=SSE4 --enable-comms=mpi-auto ; fi
-    - if [[ "$TRAVIS_OS_NAME" == "linux" ]] && [[ "$CC" == "clang" ]]; then make -j4; fi
-    - if [[ "$TRAVIS_OS_NAME" == "linux" ]] && [[ "$CC" == "clang" ]]; then mpirun.openmpi -n 2 ./benchmarks/Benchmark_dwf --threads 1 --mpi 2.1.1.1; fi
-
-

README.md

@@ -1,27 +1,44 @@
-# Grid
-<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>
+# Grid [![Teamcity status](http://ci.cliath.ph.ed.ac.uk/app/rest/builds/aggregated/strob:(buildType:(affectedProject(id:Grid)),branch:name:develop)/statusIcon.svg)](http://ci.cliath.ph.ed.ac.uk/project.html?projectId=Grid&tab=projectOverview) [![Travis status](https://travis-ci.org/paboyle/Grid.svg?branch=develop)](https://travis-ci.org/paboyle/Grid)
 
 **Data parallel C++ mathematical object library.**
 
 License: GPL v2.
 
-Last update Nov 2016.
+Last update June 2017.
 
 _Please do not send pull requests to the `master` branch which is reserved for releases._
 
+
+
+### 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
+array indices to both MPI tasks and SIMD processing elements.
+
+* Identically shaped arrays then be processed with perfect data parallelisation.
+* Such identically shaped arrays are called conformable arrays.
+
+The transformation is based on the observation that Cartesian array processing involves
+identical processing to be performed on different regions of the Cartesian array.
+
+The library will both geometrically decompose into MPI tasks and across SIMD lanes.
+Local vector loops are parallelised with OpenMP pragmas.
+
+Data parallel array operations can then be specified with a SINGLE data parallel paradigm, but
+optimally use MPI, OpenMP and SIMD parallelism under the hood. This is a significant simplification
+for most programmers.
+
+The layout transformations are parametrised by the SIMD vector length. This adapts according to the architecture.
+Presently SSE4, ARM NEON (128 bits) AVX, AVX2, QPX (256 bits), IMCI and AVX512 (512 bits) targets are supported.
+
+These are presented as `vRealF`, `vRealD`, `vComplexF`, and `vComplexD` internal vector data types. 
+The corresponding scalar types are named `RealF`, `RealD`, `ComplexF` and `ComplexD`.
+
+MPI, OpenMP, and SIMD parallelism are present in the library.
+Please see [this paper](https://arxiv.org/abs/1512.03487) for more detail.
+
+
 ### Compilers
 
 Intel ICPC v16.0.3 and later
@@ -56,35 +73,25 @@ When you file an issue, please go though the following checklist:
 6. Attach the output of `make V=1`.
 7. Describe the issue and any previous attempt to solve it. If relevant, show how to reproduce the issue using a minimal working example.
 
+### Required libraries
+Grid requires:
 
+[GMP](https://gmplib.org/), 
 
-### 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
-array indices to both MPI tasks and SIMD processing elements.
+[MPFR](http://www.mpfr.org/) 
 
-* Identically shaped arrays then be processed with perfect data parallelisation.
-* Such identically shaped arrays are called conformable arrays.
+Bootstrapping grid downloads and uses for internal dense matrix (non-QCD operations) the Eigen library.
 
-The transformation is based on the observation that Cartesian array processing involves
-identical processing to be performed on different regions of the Cartesian array.
+Grid optionally uses:
 
-The library will both geometrically decompose into MPI tasks and across SIMD lanes.
-Local vector loops are parallelised with OpenMP pragmas.
+[HDF5](https://support.hdfgroup.org/HDF5/)  
 
-Data parallel array operations can then be specified with a SINGLE data parallel paradigm, but
-optimally use MPI, OpenMP and SIMD parallelism under the hood. This is a significant simplification
-for most programmers.
+[LIME](http://usqcd-software.github.io/c-lime/) for ILDG and SciDAC file format support. 
 
-The layout transformations are parametrised by the SIMD vector length. This adapts according to the architecture.
-Presently SSE4 (128 bit) AVX, AVX2, QPX (256 bit), IMCI, and AVX512 (512 bit) targets are supported (ARM NEON on the way).
+[FFTW](http://www.fftw.org) either generic version or via the Intel MKL library.
 
-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`.
+LAPACK either generic version or Intel MKL library.
 
-MPI, OpenMP, and SIMD parallelism are present in the library.
-Please see https://arxiv.org/abs/1512.03487 for more detail.
 
 ### Quick start
 First, start by cloning the repository:
@@ -155,7 +162,6 @@ The following options can be use with the `--enable-comms=` option to target dif
 | `none`         | no communications                                             |
 | `mpi[-auto]`   | MPI communications                                            |
 | `mpi3[-auto]`  | MPI communications using MPI 3 shared memory                  |
-| `mpi3l[-auto]` | MPI communications using MPI 3 shared memory and leader model |
 | `shmem `       | Cray SHMEM communications                                     |
 
 For the MPI interfaces the optional `-auto` suffix instructs the `configure` scripts to determine all the necessary compilation and linking flags. This is done by extracting the informations from the MPI wrapper specified in the environment variable `MPICXX` (if not specified `configure` will scan though a list of default names). The `-auto` suffix is not supported by the Cray environment wrapper scripts. Use the standard versions instead.  
@@ -173,7 +179,8 @@ The following options can be use with the `--enable-simd=` option to target diff
 | `AVXFMA4`   | AVX (256 bit) + FMA4                   |
 | `AVX2`      | AVX 2 (256 bit)                        |
 | `AVX512`    | AVX 512 bit                            |
-| `QPX`       | QPX (256 bit)                          |
+| `NEONv8`    | [ARM NEON](http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.den0024a/ch07s03.html) (128 bit)                     |
+| `QPX`       | IBM QPX (256 bit)                      |
 
 Alternatively, some CPU codenames can be directly used:
 
@@ -196,20 +203,204 @@ The following configuration is recommended for the Intel Knights Landing platfor
 ../configure --enable-precision=double\
              --enable-simd=KNL        \
              --enable-comms=mpi-auto  \
-             --with-gmp=<path>        \
-             --with-mpfr=<path>       \
              --enable-mkl             \
              CXX=icpc MPICXX=mpiicpc
 ```
+The MKL flag enables use of BLAS and FFTW from the Intel Math Kernels Library.
 
-where `<path>` is the UNIX prefix where GMP and MPFR are installed. If you are working on a Cray machine that does not use the `mpiicpc` wrapper, please use:
+If you are working on a Cray machine that does not use the `mpiicpc` wrapper, please use:
 
 ``` bash
 ../configure --enable-precision=double\
              --enable-simd=KNL        \
              --enable-comms=mpi       \
-             --with-gmp=<path>        \
-             --with-mpfr=<path>       \
              --enable-mkl             \
              CXX=CC CC=cc
 ```
+
+If gmp and mpfr are NOT in standard places (/usr/) these flags may be needed:
+``` bash
+               --with-gmp=<path>        \
+               --with-mpfr=<path>       \
+```
+where `<path>` is the UNIX prefix where GMP and MPFR are installed. 
+
+Knight's Landing with Intel Omnipath adapters with two adapters per node 
+presently performs better with use of more than one rank per node, using shared memory 
+for interior communication. This is the mpi3 communications implementation. 
+We recommend four ranks per node for best performance, but optimum is local volume dependent.
+
+``` bash
+../configure --enable-precision=double\
+             --enable-simd=KNL        \
+             --enable-comms=mpi3-auto \
+             --enable-mkl             \
+             CC=icpc MPICXX=mpiicpc 
+```
+
+### Build setup for Intel Haswell Xeon platform
+
+The following configuration is recommended for the Intel Haswell platform:
+
+``` bash
+../configure --enable-precision=double\
+             --enable-simd=AVX2       \
+             --enable-comms=mpi3-auto \
+             --enable-mkl             \
+             CXX=icpc MPICXX=mpiicpc
+```
+The MKL flag enables use of BLAS and FFTW from the Intel Math Kernels Library.
+
+If gmp and mpfr are NOT in standard places (/usr/) these flags may be needed:
+``` bash
+               --with-gmp=<path>        \
+               --with-mpfr=<path>       \
+```
+where `<path>` is the UNIX prefix where GMP and MPFR are installed. 
+
+If you are working on a Cray machine that does not use the `mpiicpc` wrapper, please use:
+
+``` bash
+../configure --enable-precision=double\
+             --enable-simd=AVX2       \
+             --enable-comms=mpi3      \
+             --enable-mkl             \
+             CXX=CC CC=cc
+```
+Since Dual socket nodes are commonplace, we recommend MPI-3 as the default with the use of 
+one rank per socket. If using the Intel MPI library, threads should be pinned to NUMA domains using
+```
+        export I_MPI_PIN=1
+```
+This is the default.
+
+### Build setup for Intel Skylake Xeon platform
+
+The following configuration is recommended for the Intel Skylake platform:
+
+``` bash
+../configure --enable-precision=double\
+             --enable-simd=AVX512     \
+             --enable-comms=mpi3      \
+             --enable-mkl             \
+             CXX=mpiicpc
+```
+The MKL flag enables use of BLAS and FFTW from the Intel Math Kernels Library.
+
+If gmp and mpfr are NOT in standard places (/usr/) these flags may be needed:
+``` bash
+               --with-gmp=<path>        \
+               --with-mpfr=<path>       \
+```
+where `<path>` is the UNIX prefix where GMP and MPFR are installed. 
+
+If you are working on a Cray machine that does not use the `mpiicpc` wrapper, please use:
+
+``` bash
+../configure --enable-precision=double\
+             --enable-simd=AVX512     \
+             --enable-comms=mpi3      \
+             --enable-mkl             \
+             CXX=CC CC=cc
+```
+Since Dual socket nodes are commonplace, we recommend MPI-3 as the default with the use of 
+one rank per socket. If using the Intel MPI library, threads should be pinned to NUMA domains using
+``` 
+        export I_MPI_PIN=1
+```
+This is the default. 
+
+#### Expected Skylake Gold 6148 dual socket (single prec, single node 20+20 cores) performance using NUMA MPI mapping): 
+
+mpirun -n 2 benchmarks/Benchmark_dwf --grid 16.16.16.16 --mpi 2.1.1.1 --cacheblocking 2.2.2.2 --dslash-asm --shm 1024 --threads 18 
+
+TBA
+
+
+### Build setup for AMD EPYC / RYZEN
+
+The AMD EPYC is a multichip module comprising 32 cores spread over four distinct chips each with 8 cores.
+So, even with a single socket node there is a quad-chip module. Dual socket nodes with 64 cores total
+are common. Each chip within the module exposes a separate NUMA domain.
+There are four NUMA domains per socket and we recommend one MPI rank per NUMA domain.
+MPI-3 is recommended with the use of four ranks per socket,
+and 8 threads per rank. 
+
+The following configuration is recommended for the AMD EPYC platform.
+
+``` bash
+../configure --enable-precision=double\
+             --enable-simd=AVX2       \
+             --enable-comms=mpi3 \
+             CXX=mpicxx 
+```
+
+If gmp and mpfr are NOT in standard places (/usr/) these flags may be needed:
+``` bash
+               --with-gmp=<path>        \
+               --with-mpfr=<path>       \
+```
+where `<path>` is the UNIX prefix where GMP and MPFR are installed. 
+
+Using MPICH and g++ v4.9.2, best performance can be obtained using explicit GOMP_CPU_AFFINITY flags for each MPI rank.
+This can be done by invoking MPI on a wrapper script omp_bind.sh to handle this. 
+
+It is recommended to run 8 MPI ranks on a single dual socket AMD EPYC, with 8 threads per rank using MPI3 and
+shared memory to communicate within this node:
+
+mpirun -np 8 ./omp_bind.sh ./Benchmark_dwf --mpi 2.2.2.1 --dslash-unroll --threads 8 --grid 16.16.16.16 --cacheblocking 4.4.4.4 
+
+Where omp_bind.sh does the following:
+```
+#!/bin/bash
+
+numanode=` expr $PMI_RANK % 8 `
+basecore=`expr $numanode \* 16`
+core0=`expr $basecore + 0 `
+core1=`expr $basecore + 2 `
+core2=`expr $basecore + 4 `
+core3=`expr $basecore + 6 `
+core4=`expr $basecore + 8 `
+core5=`expr $basecore + 10 `
+core6=`expr $basecore + 12 `
+core7=`expr $basecore + 14 `
+
+export GOMP_CPU_AFFINITY="$core0 $core1 $core2 $core3 $core4 $core5 $core6 $core7"
+echo GOMP_CUP_AFFINITY $GOMP_CPU_AFFINITY
+
+$@
+```
+
+Performance:
+
+#### Expected AMD EPYC 7601 dual socket (single prec, single node 32+32 cores) performance using NUMA MPI mapping): 
+
+mpirun  -np 8 ./omp_bind.sh ./Benchmark_dwf --threads 8 --mpi 2.2.2.1 --dslash-unroll --grid 16.16.16.16 --cacheblocking 4.4.4.4
+
+TBA
+
+### Build setup for BlueGene/Q
+
+To be written...
+
+### Build setup for ARM Neon
+
+To be written...
+
+### Build setup for laptops, other compilers, non-cluster builds
+
+Many versions of g++ and clang++ work with Grid, and involve merely replacing CXX (and MPICXX),
+and omit the enable-mkl flag. 
+
+Single node builds are enabled with 
+```
+            --enable-comms=none
+```
+
+FFTW support that is not in the default search path may then enabled with
+```
+    --with-fftw=<installpath>
+```
+
+BLAS will not be compiled in by default, and Lanczos will default to Eigen diagonalisation.
+

TODO

@@ -2,18 +2,20 @@ TODO:
 ---------------
 
 Large item work list:
-1)- MultiRHS with spread out extra dim -- Go through filesystem with SciDAC I/O
 
+1)- BG/Q port and check ; Andrew says ok.
 2)- Christoph's local basis expansion Lanczos
-3)- BG/Q port and check
-4)- Precision conversion and sort out localConvert      <-- partial
-  - Consistent linear solver flop count/rate -- PARTIAL, time but no flop/s yet
-5)- Physical propagator interface
-6)- Conserved currents
-7)- Multigrid Wilson and DWF, compare to other Multigrid implementations
-8)- HDCR resume
+--
+3a)- RNG I/O in ILDG/SciDAC (minor)
+3b)- Precision conversion and sort out localConvert      <-- partial/easy
+3c)- Consistent linear solver flop count/rate -- PARTIAL, time but no flop/s yet
+4)- Physical propagator interface
+5)- Conserved currents
+6)- Multigrid Wilson and DWF, compare to other Multigrid implementations
+7)- HDCR resume
 
 Recent DONE 
+-- MultiRHS with spread out extra dim -- Go through filesystem with SciDAC I/O ; <-- DONE ; bmark cori
 -- Lanczos Remove DenseVector, DenseMatrix; Use Eigen instead. <-- DONE
 -- GaugeFix into central location                      <-- DONE
 -- Scidac and Ildg metadata handling                   <-- DONE

benchmarks/Benchmark_ITT.cc

@@ -0,0 +1,800 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./benchmarks/Benchmark_memory_bandwidth.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;
+
+typedef WilsonFermion5D<DomainWallVec5dImplR> WilsonFermion5DR;
+typedef WilsonFermion5D<DomainWallVec5dImplF> WilsonFermion5DF;
+typedef WilsonFermion5D<DomainWallVec5dImplD> WilsonFermion5DD;
+
+
+std::vector<int> L_list;
+std::vector<int> Ls_list;
+std::vector<double> mflop_list;
+
+double mflop_ref;
+double mflop_ref_err;
+
+int NN_global;
+
+struct time_statistics{
+  double mean;
+  double err;
+  double min;
+  double max;
+
+  void statistics(std::vector<double> v){
+      double sum = std::accumulate(v.begin(), v.end(), 0.0);
+      mean = sum / v.size();
+
+      std::vector<double> diff(v.size());
+      std::transform(v.begin(), v.end(), diff.begin(), [=](double x) { return x - mean; });
+      double sq_sum = std::inner_product(diff.begin(), diff.end(), diff.begin(), 0.0);
+      err = std::sqrt(sq_sum / (v.size()*(v.size() - 1)));
+
+      auto result = std::minmax_element(v.begin(), v.end());
+      min = *result.first;
+      max = *result.second;
+}
+};
+
+void comms_header(){
+  std::cout <<GridLogMessage << " L  "<<"\t"<<" Ls  "<<"\t"
+            <<std::setw(11)<<"bytes"<<"MB/s uni (err/min/max)"<<"\t\t"<<"MB/s bidi (err/min/max)"<<std::endl;
+};
+
+Gamma::Algebra Gmu [] = {
+  Gamma::Algebra::GammaX,
+  Gamma::Algebra::GammaY,
+  Gamma::Algebra::GammaZ,
+  Gamma::Algebra::GammaT
+};
+struct controls {
+  int Opt;
+  int CommsOverlap;
+  Grid::CartesianCommunicator::CommunicatorPolicy_t CommsAsynch;
+  //  int HugePages;
+};
+
+class Benchmark {
+public:
+  static void Decomposition (void ) {
+
+    int threads = GridThread::GetThreads();
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << "= Grid is setup to use "<<threads<<" threads"<<std::endl;
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage<<"Grid Default Decomposition patterns\n";
+    std::cout<<GridLogMessage<<"\tOpenMP threads : "<<GridThread::GetThreads()<<std::endl;
+    std::cout<<GridLogMessage<<"\tMPI tasks      : "<<GridCmdVectorIntToString(GridDefaultMpi())<<std::endl;
+    std::cout<<GridLogMessage<<"\tvReal          : "<<sizeof(vReal )*8    <<"bits ; " <<GridCmdVectorIntToString(GridDefaultSimd(4,vReal::Nsimd()))<<std::endl;
+    std::cout<<GridLogMessage<<"\tvRealF         : "<<sizeof(vRealF)*8    <<"bits ; " <<GridCmdVectorIntToString(GridDefaultSimd(4,vRealF::Nsimd()))<<std::endl;
+    std::cout<<GridLogMessage<<"\tvRealD         : "<<sizeof(vRealD)*8    <<"bits ; " <<GridCmdVectorIntToString(GridDefaultSimd(4,vRealD::Nsimd()))<<std::endl;
+    std::cout<<GridLogMessage<<"\tvComplex       : "<<sizeof(vComplex )*8 <<"bits ; " <<GridCmdVectorIntToString(GridDefaultSimd(4,vComplex::Nsimd()))<<std::endl;
+    std::cout<<GridLogMessage<<"\tvComplexF      : "<<sizeof(vComplexF)*8 <<"bits ; " <<GridCmdVectorIntToString(GridDefaultSimd(4,vComplexF::Nsimd()))<<std::endl;
+    std::cout<<GridLogMessage<<"\tvComplexD      : "<<sizeof(vComplexD)*8 <<"bits ; " <<GridCmdVectorIntToString(GridDefaultSimd(4,vComplexD::Nsimd()))<<std::endl;
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+
+  }
+
+  static void Comms(void)
+  {
+    int Nloop=200;
+    int nmu=0;
+    int maxlat=32;
+
+    std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplexD::Nsimd());
+    std::vector<int> mpi_layout  = GridDefaultMpi();
+
+    for(int mu=0;mu<Nd;mu++) if (mpi_layout[mu]>1) nmu++;
+
+    std::vector<double> t_time(Nloop);
+    time_statistics timestat;
+
+    std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << "= Benchmarking threaded STENCIL halo exchange in "<<nmu<<" dimensions"<<std::endl;
+    std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
+    comms_header();
+
+    for(int lat=4;lat<=maxlat;lat+=4){
+      for(int Ls=8;Ls<=8;Ls*=2){
+
+	std::vector<int> latt_size  ({lat*mpi_layout[0],
+	      lat*mpi_layout[1],
+	      lat*mpi_layout[2],
+	      lat*mpi_layout[3]});
+
+	GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
+	RealD Nrank = Grid._Nprocessors;
+	RealD Nnode = Grid.NodeCount();
+	RealD ppn = Nrank/Nnode;
+
+	std::vector<HalfSpinColourVectorD *> xbuf(8);
+	std::vector<HalfSpinColourVectorD *> rbuf(8);
+	Grid.ShmBufferFreeAll();
+	for(int d=0;d<8;d++){
+	  xbuf[d] = (HalfSpinColourVectorD *)Grid.ShmBufferMalloc(lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
+	  rbuf[d] = (HalfSpinColourVectorD *)Grid.ShmBufferMalloc(lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
+	  bzero((void *)xbuf[d],lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
+	  bzero((void *)rbuf[d],lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
+	}
+
+	int bytes=lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD);
+	int ncomm;
+	double dbytes;
+	std::vector<double> times(Nloop);
+	for(int i=0;i<Nloop;i++){
+
+	  double start=usecond();
+
+	  dbytes=0;
+	  ncomm=0;
+
+	  parallel_for(int dir=0;dir<8;dir++){
+
+	    double tbytes;
+	    int mu =dir % 4;
+
+	    if (mpi_layout[mu]>1 ) {
+	        
+	      int xmit_to_rank;
+	      int recv_from_rank;
+	      if ( dir == mu ) { 
+		int comm_proc=1;
+		Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
+	      } else { 
+		int comm_proc = mpi_layout[mu]-1;
+		Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
+	      }
+	      tbytes= Grid.StencilSendToRecvFrom((void *)&xbuf[dir][0], xmit_to_rank,
+						 (void *)&rbuf[dir][0], recv_from_rank,
+						 bytes,dir);
+	  
+#ifdef GRID_OMP
+#pragma omp atomic
+#endif
+	      ncomm++;
+
+#ifdef GRID_OMP
+#pragma omp atomic
+#endif
+	      dbytes+=tbytes;
+	    }
+	  }
+	  Grid.Barrier();
+	  double stop=usecond();
+	  t_time[i] = stop-start; // microseconds
+	}
+
+	timestat.statistics(t_time);
+	//	for(int i=0;i<t_time.size();i++){
+	//	  std::cout << i<<" "<<t_time[i]<<std::endl;
+	//	}
+
+	dbytes=dbytes*ppn;
+	double xbytes    = dbytes*0.5;
+	double rbytes    = dbytes*0.5;
+	double bidibytes = dbytes;
+
+	std::cout<<GridLogMessage << std::setw(4) << lat<<"\t"<<Ls<<"\t"
+		 <<std::setw(11) << bytes<< std::fixed << std::setprecision(1) << std::setw(7)
+		 <<std::right<< xbytes/timestat.mean<<"  "<< xbytes*timestat.err/(timestat.mean*timestat.mean)<< " "
+		 <<xbytes/timestat.max <<" "<< xbytes/timestat.min  
+		 << "\t\t"<<std::setw(7)<< bidibytes/timestat.mean<< "  " << bidibytes*timestat.err/(timestat.mean*timestat.mean) << " "
+		 << bidibytes/timestat.max << " " << bidibytes/timestat.min << std::endl;
+
+ 
+	
+	    }
+    }    
+
+    return;
+  }
+
+  static void Memory(void)
+  {
+    const int Nvec=8;
+    typedef Lattice< iVector< vReal,Nvec> > LatticeVec;
+    typedef iVector<vReal,Nvec> Vec;
+
+    std::vector<int> simd_layout = GridDefaultSimd(Nd,vReal::Nsimd());
+    std::vector<int> mpi_layout  = GridDefaultMpi();
+
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << "= Benchmarking a*x + y bandwidth"<<std::endl;
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << "  L  "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s"<<"\t\t"<<"Gflop/s"<<"\t\t seconds"<< "\t\tGB/s / node"<<std::endl;
+    std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
+  
+    uint64_t NP;
+    uint64_t NN;
+
+
+  uint64_t lmax=48;
+#define NLOOP (100*lmax*lmax*lmax*lmax/lat/lat/lat/lat)
+
+    GridSerialRNG          sRNG;      sRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9}));
+    for(int lat=8;lat<=lmax;lat+=4){
+
+      std::vector<int> latt_size  ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
+      int64_t vol= latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
+      GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
+
+      NP= Grid.RankCount();
+      NN =Grid.NodeCount();
+
+      Vec rn ; random(sRNG,rn);
+
+      LatticeVec z(&Grid); z=rn;
+      LatticeVec x(&Grid); x=rn;
+      LatticeVec y(&Grid); y=rn;
+      double a=2.0;
+
+      uint64_t Nloop=NLOOP;
+
+      double start=usecond();
+      for(int i=0;i<Nloop;i++){
+	z=a*x-y;
+        x._odata[0]=z._odata[0]; // force serial dependency to prevent optimise away
+        y._odata[4]=z._odata[4];
+      }
+      double stop=usecond();
+      double time = (stop-start)/Nloop*1000;
+     
+      double flops=vol*Nvec*2;// mul,add
+      double bytes=3.0*vol*Nvec*sizeof(Real);
+      std::cout<<GridLogMessage<<std::setprecision(3) 
+	       << lat<<"\t\t"<<bytes<<"   \t\t"<<bytes/time<<"\t\t"<<flops/time<<"\t\t"<<(stop-start)/1000./1000.
+	       << "\t\t"<< bytes/time/NN <<std::endl;
+
+    }
+  };
+
+  static double DWF5(int Ls,int L)
+  {
+    RealD mass=0.1;
+    RealD M5  =1.8;
+
+    double mflops;
+    double mflops_best = 0;
+    double mflops_worst= 0;
+    std::vector<double> mflops_all;
+
+    ///////////////////////////////////////////////////////
+    // Set/Get the layout & grid size
+    ///////////////////////////////////////////////////////
+    int threads = GridThread::GetThreads();
+    std::vector<int> mpi = GridDefaultMpi(); assert(mpi.size()==4);
+    std::vector<int> local({L,L,L,L});
+
+    GridCartesian         * TmpGrid   = SpaceTimeGrid::makeFourDimGrid(std::vector<int>({64,64,64,64}), 
+								       GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
+    uint64_t NP = TmpGrid->RankCount();
+    uint64_t NN = TmpGrid->NodeCount();
+    NN_global=NN;
+    uint64_t SHM=NP/NN;
+
+    std::vector<int> internal;
+    if      ( SHM == 1 )   internal = std::vector<int>({1,1,1,1});
+    else if ( SHM == 2 )   internal = std::vector<int>({2,1,1,1});
+    else if ( SHM == 4 )   internal = std::vector<int>({2,2,1,1});
+    else if ( SHM == 8 )   internal = std::vector<int>({2,2,2,1});
+    else assert(0);
+
+    std::vector<int> nodes({mpi[0]/internal[0],mpi[1]/internal[1],mpi[2]/internal[2],mpi[3]/internal[3]});
+    std::vector<int> latt4({local[0]*nodes[0],local[1]*nodes[1],local[2]*nodes[2],local[3]*nodes[3]});
+
+    ///////// Welcome message ////////////
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << "Benchmark DWF Ls vec on "<<L<<"^4 local volume "<<std::endl;
+    std::cout<<GridLogMessage << "* Global volume  : "<<GridCmdVectorIntToString(latt4)<<std::endl;
+    std::cout<<GridLogMessage << "* Ls             : "<<Ls<<std::endl;
+    std::cout<<GridLogMessage << "* MPI ranks      : "<<GridCmdVectorIntToString(mpi)<<std::endl;
+    std::cout<<GridLogMessage << "* Intranode      : "<<GridCmdVectorIntToString(internal)<<std::endl;
+    std::cout<<GridLogMessage << "* nodes          : "<<GridCmdVectorIntToString(nodes)<<std::endl;
+    std::cout<<GridLogMessage << "* Using "<<threads<<" threads"<<std::endl;
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+
+    ///////// Lattice Init ////////////
+    GridCartesian         * UGrid    = SpaceTimeGrid::makeFourDimGrid(latt4, GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
+    GridRedBlackCartesian * UrbGrid  = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+    GridCartesian         * sUGrid   = SpaceTimeGrid::makeFourDimDWFGrid(latt4,GridDefaultMpi());
+    GridRedBlackCartesian * sUrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(sUGrid);
+    GridCartesian         * sFGrid   = SpaceTimeGrid::makeFiveDimDWFGrid(Ls,UGrid);
+    GridRedBlackCartesian * sFrbGrid = SpaceTimeGrid::makeFiveDimDWFRedBlackGrid(Ls,UGrid);
+
+    ///////// RNG Init ////////////
+    std::vector<int> seeds4({1,2,3,4});
+    std::vector<int> seeds5({5,6,7,8});
+    GridParallelRNG          RNG4(UGrid);  RNG4.SeedFixedIntegers(seeds4);
+    GridParallelRNG          RNG5(sFGrid);  RNG5.SeedFixedIntegers(seeds5);
+    std::cout << GridLogMessage << "Initialised RNGs" << std::endl;
+
+    ///////// Source preparation ////////////
+    LatticeFermion src   (sFGrid); random(RNG5,src);
+    LatticeFermion tmp   (sFGrid);
+
+    RealD N2 = 1.0/::sqrt(norm2(src));
+    src = src*N2;
+    
+    LatticeGaugeField Umu(UGrid);  SU3::HotConfiguration(RNG4,Umu); 
+
+    WilsonFermion5DR sDw(Umu,*sFGrid,*sFrbGrid,*sUGrid,*sUrbGrid,M5);
+    LatticeFermion src_e (sFrbGrid);
+    LatticeFermion src_o (sFrbGrid);
+    LatticeFermion r_e   (sFrbGrid);
+    LatticeFermion r_o   (sFrbGrid);
+    LatticeFermion r_eo  (sFGrid);
+    LatticeFermion err   (sFGrid);
+    {
+
+      pickCheckerboard(Even,src_e,src);
+      pickCheckerboard(Odd,src_o,src);
+
+#if defined(AVX512) 
+      const int num_cases = 6;
+      std::string fmt("A/S ; A/O ; U/S ; U/O ; G/S ; G/O ");
+#else
+      const int num_cases = 4;
+      std::string fmt("U/S ; U/O ; G/S ; G/O ");
+#endif
+      controls Cases [] = {
+#ifdef AVX512
+	{ QCD::WilsonKernelsStatic::OptInlineAsm , QCD::WilsonKernelsStatic::CommsThenCompute ,CartesianCommunicator::CommunicatorPolicySequential  },
+	{ QCD::WilsonKernelsStatic::OptInlineAsm , QCD::WilsonKernelsStatic::CommsAndCompute  ,CartesianCommunicator::CommunicatorPolicySequential  },
+#endif
+	{ QCD::WilsonKernelsStatic::OptHandUnroll, QCD::WilsonKernelsStatic::CommsThenCompute ,CartesianCommunicator::CommunicatorPolicySequential  },
+	{ QCD::WilsonKernelsStatic::OptHandUnroll, QCD::WilsonKernelsStatic::CommsAndCompute  ,CartesianCommunicator::CommunicatorPolicySequential  },
+	{ QCD::WilsonKernelsStatic::OptGeneric   , QCD::WilsonKernelsStatic::CommsThenCompute ,CartesianCommunicator::CommunicatorPolicySequential  },
+	{ QCD::WilsonKernelsStatic::OptGeneric   , QCD::WilsonKernelsStatic::CommsAndCompute  ,CartesianCommunicator::CommunicatorPolicySequential  }
+      }; 
+
+      for(int c=0;c<num_cases;c++) {
+
+	QCD::WilsonKernelsStatic::Comms = Cases[c].CommsOverlap;
+	QCD::WilsonKernelsStatic::Opt   = Cases[c].Opt;
+	CartesianCommunicator::SetCommunicatorPolicy(Cases[c].CommsAsynch);
+
+	std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+	if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGeneric   ) std::cout << GridLogMessage<< "* Using GENERIC Nc WilsonKernels" <<std::endl;
+	if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptHandUnroll) std::cout << GridLogMessage<< "* Using Nc=3       WilsonKernels" <<std::endl;
+	if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm ) std::cout << GridLogMessage<< "* Using Asm Nc=3   WilsonKernels" <<std::endl;
+	if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute ) std::cout << GridLogMessage<< "* Using Overlapped Comms/Compute" <<std::endl;
+	if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsThenCompute) std::cout << GridLogMessage<< "* Using sequential comms compute" <<std::endl;
+	if ( sizeof(Real)==4 )   std::cout << GridLogMessage<< "* SINGLE precision "<<std::endl;
+	if ( sizeof(Real)==8 )   std::cout << GridLogMessage<< "* DOUBLE precision "<<std::endl;
+	std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+
+	int nwarm = 100;
+	uint64_t ncall = 1000;
+
+	double t0=usecond();
+	sFGrid->Barrier();
+	for(int i=0;i<nwarm;i++){
+	  sDw.DhopEO(src_o,r_e,DaggerNo);
+	}
+	sFGrid->Barrier();
+	double t1=usecond();
+
+	sDw.ZeroCounters();
+	time_statistics timestat;
+	std::vector<double> t_time(ncall);
+	for(uint64_t i=0;i<ncall;i++){
+	  t0=usecond();
+	  sDw.DhopEO(src_o,r_e,DaggerNo);
+	  t1=usecond();
+	  t_time[i] = t1-t0;
+	}
+	sFGrid->Barrier();
+	
+	double volume=Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
+	double flops=(1344.0*volume)/2;
+	double mf_hi, mf_lo, mf_err;
+
+	timestat.statistics(t_time);
+	mf_hi = flops/timestat.min;
+	mf_lo = flops/timestat.max;
+	mf_err= flops/timestat.min * timestat.err/timestat.mean;
+
+	mflops = flops/timestat.mean;
+	mflops_all.push_back(mflops);
+	if ( mflops_best == 0   ) mflops_best = mflops;
+	if ( mflops_worst== 0   ) mflops_worst= mflops;
+	if ( mflops>mflops_best ) mflops_best = mflops;
+	if ( mflops<mflops_worst) mflops_worst= mflops;
+
+	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"sDeo mflop/s =   "<< mflops << " ("<<mf_err<<") " << mf_lo<<"-"<<mf_hi <<std::endl;
+	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"sDeo mflop/s per rank   "<< mflops/NP<<std::endl;
+	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"sDeo mflop/s per node   "<< mflops/NN<<std::endl;
+
+	sDw.Report();
+
+      }
+      double robust = mflops_worst/mflops_best;;
+      std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+      std::cout<<GridLogMessage << L<<"^4 x "<<Ls<< " sDeo Best  mflop/s        =   "<< mflops_best << " ; " << mflops_best/NN<<" per node " <<std::endl;
+      std::cout<<GridLogMessage << L<<"^4 x "<<Ls<< " sDeo Worst mflop/s        =   "<< mflops_worst<< " ; " << mflops_worst/NN<<" per node " <<std::endl;
+
+      std::cout<<GridLogMessage <<std::setprecision(3)<< L<<"^4 x "<<Ls<< " Performance Robustness   =   "<< robust <<std::endl;
+      std::cout<<GridLogMessage <<fmt << std::endl;
+      std::cout<<GridLogMessage;
+
+      for(int i=0;i<mflops_all.size();i++){
+	std::cout<<mflops_all[i]/NN<<" ; " ;
+      }
+      std::cout<<std::endl;
+      std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+
+    }
+    return mflops_best;
+  }
+
+  static double DWF(int Ls,int L, double & robust)
+  {
+    RealD mass=0.1;
+    RealD M5  =1.8;
+
+    double mflops;
+    double mflops_best = 0;
+    double mflops_worst= 0;
+    std::vector<double> mflops_all;
+
+    ///////////////////////////////////////////////////////
+    // Set/Get the layout & grid size
+    ///////////////////////////////////////////////////////
+    int threads = GridThread::GetThreads();
+    std::vector<int> mpi = GridDefaultMpi(); assert(mpi.size()==4);
+    std::vector<int> local({L,L,L,L});
+
+    GridCartesian         * TmpGrid   = SpaceTimeGrid::makeFourDimGrid(std::vector<int>({64,64,64,64}), 
+								       GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
+    uint64_t NP = TmpGrid->RankCount();
+    uint64_t NN = TmpGrid->NodeCount();
+    NN_global=NN;
+    uint64_t SHM=NP/NN;
+
+    std::vector<int> internal;
+    if      ( SHM == 1 )   internal = std::vector<int>({1,1,1,1});
+    else if ( SHM == 2 )   internal = std::vector<int>({2,1,1,1});
+    else if ( SHM == 4 )   internal = std::vector<int>({2,2,1,1});
+    else if ( SHM == 8 )   internal = std::vector<int>({2,2,2,1});
+    else assert(0);
+
+    std::vector<int> nodes({mpi[0]/internal[0],mpi[1]/internal[1],mpi[2]/internal[2],mpi[3]/internal[3]});
+    std::vector<int> latt4({local[0]*nodes[0],local[1]*nodes[1],local[2]*nodes[2],local[3]*nodes[3]});
+
+    ///////// Welcome message ////////////
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << "Benchmark DWF on "<<L<<"^4 local volume "<<std::endl;
+    std::cout<<GridLogMessage << "* Global volume  : "<<GridCmdVectorIntToString(latt4)<<std::endl;
+    std::cout<<GridLogMessage << "* Ls             : "<<Ls<<std::endl;
+    std::cout<<GridLogMessage << "* MPI ranks      : "<<GridCmdVectorIntToString(mpi)<<std::endl;
+    std::cout<<GridLogMessage << "* Intranode      : "<<GridCmdVectorIntToString(internal)<<std::endl;
+    std::cout<<GridLogMessage << "* nodes          : "<<GridCmdVectorIntToString(nodes)<<std::endl;
+    std::cout<<GridLogMessage << "* Using "<<threads<<" threads"<<std::endl;
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+
+
+    ///////// Lattice Init ////////////
+    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);
+
+    
+    ///////// RNG Init ////////////
+    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);
+    std::cout << GridLogMessage << "Initialised RNGs" << std::endl;
+
+    ///////// Source preparation ////////////
+    LatticeFermion src   (FGrid); random(RNG5,src);
+    LatticeFermion ref   (FGrid);
+    LatticeFermion tmp   (FGrid);
+
+    RealD N2 = 1.0/::sqrt(norm2(src));
+    src = src*N2;
+    
+    LatticeGaugeField Umu(UGrid);  SU3::HotConfiguration(RNG4,Umu); 
+
+    DomainWallFermionR Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+
+    ////////////////////////////////////
+    // Naive wilson implementation
+    ////////////////////////////////////
+    {
+      LatticeGaugeField Umu5d(FGrid); 
+      std::vector<LatticeColourMatrix> U(4,FGrid);
+      for(int ss=0;ss<Umu._grid->oSites();ss++){
+	for(int s=0;s<Ls;s++){
+	  Umu5d._odata[Ls*ss+s] = Umu._odata[ss];
+	}
+      }
+      ref = zero;
+      for(int mu=0;mu<Nd;mu++){
+	U[mu] = PeekIndex<LorentzIndex>(Umu5d,mu);
+      }
+      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;
+    }
+
+    LatticeFermion src_e (FrbGrid);
+    LatticeFermion src_o (FrbGrid);
+    LatticeFermion r_e   (FrbGrid);
+    LatticeFermion r_o   (FrbGrid);
+    LatticeFermion r_eo  (FGrid);
+    LatticeFermion err   (FGrid);
+    {
+
+      pickCheckerboard(Even,src_e,src);
+      pickCheckerboard(Odd,src_o,src);
+
+#if defined(AVX512) 
+      const int num_cases = 6;
+      std::string fmt("A/S ; A/O ; U/S ; U/O ; G/S ; G/O ");
+#else
+      const int num_cases = 4;
+      std::string fmt("U/S ; U/O ; G/S ; G/O ");
+#endif
+      controls Cases [] = {
+#ifdef AVX512
+	{ QCD::WilsonKernelsStatic::OptInlineAsm , QCD::WilsonKernelsStatic::CommsThenCompute ,CartesianCommunicator::CommunicatorPolicySequential  },
+	{ QCD::WilsonKernelsStatic::OptInlineAsm , QCD::WilsonKernelsStatic::CommsAndCompute  ,CartesianCommunicator::CommunicatorPolicySequential  },
+#endif
+	{ QCD::WilsonKernelsStatic::OptHandUnroll, QCD::WilsonKernelsStatic::CommsThenCompute ,CartesianCommunicator::CommunicatorPolicySequential  },
+	{ QCD::WilsonKernelsStatic::OptHandUnroll, QCD::WilsonKernelsStatic::CommsAndCompute  ,CartesianCommunicator::CommunicatorPolicySequential  },
+	{ QCD::WilsonKernelsStatic::OptGeneric   , QCD::WilsonKernelsStatic::CommsThenCompute ,CartesianCommunicator::CommunicatorPolicySequential  },
+	{ QCD::WilsonKernelsStatic::OptGeneric   , QCD::WilsonKernelsStatic::CommsAndCompute  ,CartesianCommunicator::CommunicatorPolicySequential  }
+      }; 
+
+      for(int c=0;c<num_cases;c++) {
+
+	QCD::WilsonKernelsStatic::Comms = Cases[c].CommsOverlap;
+	QCD::WilsonKernelsStatic::Opt   = Cases[c].Opt;
+	CartesianCommunicator::SetCommunicatorPolicy(Cases[c].CommsAsynch);
+
+	std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+	if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGeneric   ) std::cout << GridLogMessage<< "* Using GENERIC Nc WilsonKernels" <<std::endl;
+	if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptHandUnroll) std::cout << GridLogMessage<< "* Using Nc=3       WilsonKernels" <<std::endl;
+	if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm ) std::cout << GridLogMessage<< "* Using Asm Nc=3   WilsonKernels" <<std::endl;
+	if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute ) std::cout << GridLogMessage<< "* Using Overlapped Comms/Compute" <<std::endl;
+	if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsThenCompute) std::cout << GridLogMessage<< "* Using sequential comms compute" <<std::endl;
+	if ( sizeof(Real)==4 )   std::cout << GridLogMessage<< "* SINGLE precision "<<std::endl;
+	if ( sizeof(Real)==8 )   std::cout << GridLogMessage<< "* DOUBLE precision "<<std::endl;
+	std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+
+	int nwarm = 200;
+	double t0=usecond();
+	FGrid->Barrier();
+	for(int i=0;i<nwarm;i++){
+	  Dw.DhopEO(src_o,r_e,DaggerNo);
+	}
+	FGrid->Barrier();
+	double t1=usecond();
+	//	uint64_t ncall = (uint64_t) 2.5*1000.0*1000.0*nwarm/(t1-t0);
+	//	if (ncall < 500) ncall = 500;
+	uint64_t ncall = 1000;
+
+	FGrid->Broadcast(0,&ncall,sizeof(ncall));
+
+	//	std::cout << GridLogMessage << " Estimate " << ncall << " calls per second"<<std::endl;
+	Dw.ZeroCounters();
+
+	time_statistics timestat;
+	std::vector<double> t_time(ncall);
+	for(uint64_t i=0;i<ncall;i++){
+	  t0=usecond();
+	  Dw.DhopEO(src_o,r_e,DaggerNo);
+	  t1=usecond();
+	  t_time[i] = t1-t0;
+	}
+	FGrid->Barrier();
+	
+	double volume=Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
+	double flops=(1344.0*volume)/2;
+	double mf_hi, mf_lo, mf_err;
+
+	timestat.statistics(t_time);
+	mf_hi = flops/timestat.min;
+	mf_lo = flops/timestat.max;
+	mf_err= flops/timestat.min * timestat.err/timestat.mean;
+
+	mflops = flops/timestat.mean;
+	mflops_all.push_back(mflops);
+	if ( mflops_best == 0   ) mflops_best = mflops;
+	if ( mflops_worst== 0   ) mflops_worst= mflops;
+	if ( mflops>mflops_best ) mflops_best = mflops;
+	if ( mflops<mflops_worst) mflops_worst= mflops;
+
+	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Deo mflop/s =   "<< mflops << " ("<<mf_err<<") " << mf_lo<<"-"<<mf_hi <<std::endl;
+	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Deo mflop/s per rank   "<< mflops/NP<<std::endl;
+	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Deo mflop/s per node   "<< mflops/NN<<std::endl;
+
+	Dw.Report();
+
+	Dw.DhopEO(src_o,r_e,DaggerNo);
+	Dw.DhopOE(src_e,r_o,DaggerNo);
+	setCheckerboard(r_eo,r_o);
+	setCheckerboard(r_eo,r_e);
+	err = r_eo-ref; 
+	std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
+	assert((norm2(err)<1.0e-4));
+
+      }
+      robust = mflops_worst/mflops_best;
+      std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+      std::cout<<GridLogMessage << L<<"^4 x "<<Ls<< " Deo Best  mflop/s        =   "<< mflops_best << " ; " << mflops_best/NN<<" per node " <<std::endl;
+      std::cout<<GridLogMessage << L<<"^4 x "<<Ls<< " Deo Worst mflop/s        =   "<< mflops_worst<< " ; " << mflops_worst/NN<<" per node " <<std::endl;
+      std::cout<<GridLogMessage << std::fixed<<std::setprecision(3)<< L<<"^4 x "<<Ls<< " Performance Robustness   =   "<< robust  <<std::endl;
+      std::cout<<GridLogMessage <<fmt << std::endl;
+      std::cout<<GridLogMessage ;
+
+      for(int i=0;i<mflops_all.size();i++){
+	std::cout<<mflops_all[i]/NN<<" ; " ;
+      }
+      std::cout<<std::endl;
+      std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+
+    }
+    return mflops_best;
+  }
+
+};
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  CartesianCommunicator::SetCommunicatorPolicy(CartesianCommunicator::CommunicatorPolicySequential);
+#ifdef KNL
+  LebesgueOrder::Block = std::vector<int>({8,2,2,2});
+#else
+  LebesgueOrder::Block = std::vector<int>({2,2,2,2});
+#endif
+  Benchmark::Decomposition();
+
+  int do_memory=1;
+  int do_comms =1;
+  int do_su3   =0;
+  int do_wilson=1;
+  int do_dwf   =1;
+
+  if ( do_su3 ) {
+    // empty for now
+  }
+#if 1
+  int sel=2;
+  std::vector<int> L_list({8,12,16,24});
+#else
+  int sel=1;
+  std::vector<int> L_list({8,12});
+#endif
+  int selm1=sel-1;
+  std::vector<double> robust_list;
+
+  std::vector<double> wilson;
+  std::vector<double> dwf4;
+  std::vector<double> dwf5;
+
+  if ( do_wilson ) {
+    int Ls=1;
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << " Wilson dslash 4D vectorised" <<std::endl;
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    for(int l=0;l<L_list.size();l++){
+      double robust;
+      wilson.push_back(Benchmark::DWF(1,L_list[l],robust));
+    }
+  }
+
+  int Ls=16;
+  if ( do_dwf ) {
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << " Domain wall dslash 4D vectorised" <<std::endl;
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    for(int l=0;l<L_list.size();l++){
+      double robust;
+      double result = Benchmark::DWF(Ls,L_list[l],robust) ;
+      dwf4.push_back(result);
+      robust_list.push_back(robust);
+    }
+  }
+
+  if ( do_dwf ) {
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << " Domain wall dslash 4D vectorised" <<std::endl;
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    for(int l=0;l<L_list.size();l++){
+      dwf5.push_back(Benchmark::DWF5(Ls,L_list[l]));
+    }
+
+  }
+
+  if ( do_dwf ) {
+
+  std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+  std::cout<<GridLogMessage << " Summary table Ls="<<Ls <<std::endl;
+  std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+  std::cout<<GridLogMessage << "L \t\t Wilson \t DWF4 \t DWF5 " <<std::endl;
+  for(int l=0;l<L_list.size();l++){
+    std::cout<<GridLogMessage << L_list[l] <<" \t\t "<< wilson[l]<<" \t "<<dwf4[l]<<" \t "<<dwf5[l] <<std::endl;
+  }
+  std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+  }
+
+  int NN=NN_global;
+  if ( do_memory ) {
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << " Memory benchmark " <<std::endl;
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    Benchmark::Memory();
+  }
+
+  if ( do_comms && (NN>1) ) {
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << " Communications benchmark " <<std::endl;
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    Benchmark::Comms();
+  }
+
+  if ( do_dwf ) {
+  std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+  std::cout<<GridLogMessage << " Per Node Summary table Ls="<<Ls <<std::endl;
+  std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+  std::cout<<GridLogMessage << " L \t\t Wilson\t\t DWF4  \t\t DWF5 " <<std::endl;
+  for(int l=0;l<L_list.size();l++){
+    std::cout<<GridLogMessage << L_list[l] <<" \t\t "<< wilson[l]/NN<<" \t "<<dwf4[l]/NN<<" \t "<<dwf5[l] /NN<<std::endl;
+  }
+  std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+
+  std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+  std::cout<<GridLogMessage << " Comparison point     result: "  << 0.5*(dwf4[sel]+dwf4[selm1])/NN << " Mflop/s per node"<<std::endl;
+  std::cout<<GridLogMessage << " Comparison point is 0.5*("<<dwf4[sel]/NN<<"+"<<dwf4[selm1]/NN << ") "<<std::endl;
+  std::cout<<std::setprecision(3);
+  std::cout<<GridLogMessage << " Comparison point robustness: "  << robust_list[sel] <<std::endl;
+  std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+
+  }
+
+
+  Grid_finalize();
+}

benchmarks/Benchmark_comms.cc

@@ -68,7 +68,7 @@ int main (int argc, char ** argv)
 
   int Nloop=100;
   int nmu=0;
-  int maxlat=24;
+  int maxlat=32;
   for(int mu=0;mu<Nd;mu++) if (mpi_layout[mu]>1) nmu++;
 
   std::cout << GridLogMessage << "Number of iterations to average: "<< Nloop << std::endl;
@@ -80,7 +80,7 @@ int main (int argc, char ** argv)
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
   header();
   for(int lat=4;lat<=maxlat;lat+=4){
-    for(int Ls=8;Ls<=32;Ls*=2){
+    for(int Ls=8;Ls<=8;Ls*=2){
 
       std::vector<int> latt_size  ({lat*mpi_layout[0],
       				    lat*mpi_layout[1],
@@ -92,11 +92,16 @@ int main (int argc, char ** argv)
       RealD Nnode = Grid.NodeCount();
       RealD ppn = Nrank/Nnode;
 
-      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));
+      std::vector<Vector<HalfSpinColourVectorD> > xbuf(8);	
+      std::vector<Vector<HalfSpinColourVectorD> > rbuf(8);
 
       int ncomm;
       int bytes=lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD);
+      for(int mu=0;mu<8;mu++){
+	xbuf[mu].resize(lat*lat*lat*Ls);
+	rbuf[mu].resize(lat*lat*lat*Ls);
+	//	std::cout << " buffers " << std::hex << (uint64_t)&xbuf[mu][0] <<" " << (uint64_t)&rbuf[mu][0] <<std::endl;
+      }
 
       for(int i=0;i<Nloop;i++){
       double start=usecond();
@@ -112,7 +117,6 @@ int main (int argc, char ** argv)
 	    int comm_proc=1;
 	    int xmit_to_rank;
 	    int recv_from_rank;
-	    
 	    Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
 	    Grid.SendToRecvFromBegin(requests,
 				   (void *)&xbuf[mu][0],
@@ -163,7 +167,7 @@ int main (int argc, char ** argv)
   header();
 
   for(int lat=4;lat<=maxlat;lat+=4){
-    for(int Ls=8;Ls<=32;Ls*=2){
+    for(int Ls=8;Ls<=8;Ls*=2){
 
       std::vector<int> latt_size  ({lat,lat,lat,lat});
 
@@ -172,9 +176,14 @@ int main (int argc, char ** argv)
       RealD Nnode = Grid.NodeCount();
       RealD ppn = Nrank/Nnode;
 
-      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));
+      std::vector<Vector<HalfSpinColourVectorD> > xbuf(8);
+      std::vector<Vector<HalfSpinColourVectorD> > rbuf(8);
 
+      for(int mu=0;mu<8;mu++){
+	xbuf[mu].resize(lat*lat*lat*Ls);
+	rbuf[mu].resize(lat*lat*lat*Ls);
+	//	std::cout << " buffers " << std::hex << (uint64_t)&xbuf[mu][0] <<" " << (uint64_t)&rbuf[mu][0] <<std::endl;
+      }
 
       int ncomm;
       int bytes=lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD);
@@ -249,7 +258,7 @@ int main (int argc, char ** argv)
   header();
 
   for(int lat=4;lat<=maxlat;lat+=4){
-    for(int Ls=8;Ls<=32;Ls*=2){
+    for(int Ls=8;Ls<=8;Ls*=2){
 
       std::vector<int> latt_size  ({lat*mpi_layout[0],
       				    lat*mpi_layout[1],
@@ -299,7 +308,7 @@ int main (int argc, char ** argv)
 					      xmit_to_rank,
 					      (void *)&rbuf[mu][0],
 					      recv_from_rank,
-					      bytes);
+					      bytes,mu);
 	
 	    comm_proc = mpi_layout[mu]-1;
 	  
@@ -310,11 +319,11 @@ int main (int argc, char ** argv)
 					      xmit_to_rank,
 					      (void *)&rbuf[mu+4][0],
 					      recv_from_rank,
-					      bytes);
+					      bytes,mu+4);
 	  
 	  }
 	}
-	Grid.StencilSendToRecvFromComplete(requests);
+	Grid.StencilSendToRecvFromComplete(requests,0);
 	Grid.Barrier();
 	double stop=usecond();
 	t_time[i] = stop-start; // microseconds
@@ -346,7 +355,7 @@ int main (int argc, char ** argv)
   header();
 
   for(int lat=4;lat<=maxlat;lat+=4){
-    for(int Ls=8;Ls<=32;Ls*=2){
+    for(int Ls=8;Ls<=8;Ls*=2){
 
       std::vector<int> latt_size  ({lat*mpi_layout[0],
       				    lat*mpi_layout[1],
@@ -393,8 +402,8 @@ int main (int argc, char ** argv)
 					      xmit_to_rank,
 					      (void *)&rbuf[mu][0],
 					      recv_from_rank,
-					      bytes);
-	    Grid.StencilSendToRecvFromComplete(requests);
+					      bytes,mu);
+	    Grid.StencilSendToRecvFromComplete(requests,mu);
 	    requests.resize(0);
 
 	    comm_proc = mpi_layout[mu]-1;
@@ -406,8 +415,8 @@ int main (int argc, char ** argv)
 					      xmit_to_rank,
 					      (void *)&rbuf[mu+4][0],
 					      recv_from_rank,
-					      bytes);
-	    Grid.StencilSendToRecvFromComplete(requests);
+					      bytes,mu+4);
+	    Grid.StencilSendToRecvFromComplete(requests,mu+4);
 	    requests.resize(0);
 	  
 	  }
@@ -436,5 +445,97 @@ int main (int argc, char ** argv)
     }
   }    
 
+
+
+  std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
+  std::cout<<GridLogMessage << "= Benchmarking threaded STENCIL halo exchange in "<<nmu<<" dimensions"<<std::endl;
+  std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
+  header();
+
+  for(int lat=4;lat<=maxlat;lat+=4){
+    for(int Ls=8;Ls<=8;Ls*=2){
+
+      std::vector<int> latt_size  ({lat*mpi_layout[0],
+      				    lat*mpi_layout[1],
+      				    lat*mpi_layout[2],
+      				    lat*mpi_layout[3]});
+
+      GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
+      RealD Nrank = Grid._Nprocessors;
+      RealD Nnode = Grid.NodeCount();
+      RealD ppn = Nrank/Nnode;
+
+      std::vector<HalfSpinColourVectorD *> xbuf(8);
+      std::vector<HalfSpinColourVectorD *> rbuf(8);
+      Grid.ShmBufferFreeAll();
+      for(int d=0;d<8;d++){
+	xbuf[d] = (HalfSpinColourVectorD *)Grid.ShmBufferMalloc(lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
+	rbuf[d] = (HalfSpinColourVectorD *)Grid.ShmBufferMalloc(lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
+	bzero((void *)xbuf[d],lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
+	bzero((void *)rbuf[d],lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
+      }
+
+      int ncomm;
+      int bytes=lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD);
+      double dbytes;
+      for(int i=0;i<Nloop;i++){
+	double start=usecond();
+
+	std::vector<CartesianCommunicator::CommsRequest_t> requests;
+	dbytes=0;
+	ncomm=0;
+
+	parallel_for(int dir=0;dir<8;dir++){
+
+	  double tbytes;
+	  int mu =dir % 4;
+
+	  if (mpi_layout[mu]>1 ) {
+	  
+	    ncomm++;
+	    int xmit_to_rank;
+	    int recv_from_rank;
+	    if ( dir == mu ) { 
+	      int comm_proc=1;
+	      Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
+	    } else { 
+	      int comm_proc = mpi_layout[mu]-1;
+	      Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
+	    }
+
+	    tbytes= Grid.StencilSendToRecvFrom((void *)&xbuf[dir][0], xmit_to_rank,
+					       (void *)&rbuf[dir][0], recv_from_rank, bytes,dir);
+
+#pragma omp atomic
+	    dbytes+=tbytes;
+	  }
+	}
+	Grid.Barrier();
+	double stop=usecond();
+	t_time[i] = stop-start; // microseconds
+      }
+
+      timestat.statistics(t_time);
+
+      dbytes=dbytes*ppn;
+      double xbytes    = dbytes*0.5;
+      double rbytes    = dbytes*0.5;
+      double bidibytes = dbytes;
+
+
+      std::cout<<GridLogMessage << std::setw(4) << lat<<"\t"<<Ls<<"\t"
+               <<std::setw(11) << bytes<< std::fixed << std::setprecision(1) << std::setw(7)
+               <<std::right<< xbytes/timestat.mean<<"  "<< xbytes*timestat.err/(timestat.mean*timestat.mean)<< " "
+               <<xbytes/timestat.max <<" "<< xbytes/timestat.min  
+               << "\t\t"<<std::setw(7)<< bidibytes/timestat.mean<< "  " << bidibytes*timestat.err/(timestat.mean*timestat.mean) << " "
+               << bidibytes/timestat.max << " " << bidibytes/timestat.min << std::endl;
+ 
+    }
+  }    
+
+  std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
+  std::cout<<GridLogMessage << "= All done; Bye Bye"<<std::endl;
+  std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
+
   Grid_finalize();
 }

benchmarks/Benchmark_dwf.cc

@@ -51,7 +51,13 @@ 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=16;
+  int Ls=16;
+  for(int i=0;i<argc;i++)
+    if(std::string(argv[i]) == "-Ls"){
+      std::stringstream ss(argv[i+1]); ss >> Ls;
+    }
+
+
   GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
   GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
   GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
@@ -165,7 +171,7 @@ int main (int argc, char ** argv)
   std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
 
   DomainWallFermionR Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
-  int ncall =1000;
+  int ncall =500;
   if (1) {
     FGrid->Barrier();
     Dw.ZeroCounters();
@@ -303,6 +309,7 @@ int main (int argc, char ** argv)
     }
     assert(sum < 1.0e-4);
 
+    
     if(1){
       std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
       std::cout << GridLogMessage<< "* Benchmarking WilsonFermion5D<DomainWallVec5dImplR>::DhopEO "<<std::endl;
@@ -381,7 +388,22 @@ int main (int argc, char ** argv)
       }
       assert(error<1.0e-4);
     }
+
+  if(0){
+    std::cout << "Single cache warm call to sDw.Dhop " <<std::endl;
+    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();
     }
+  }
+
+  }
+
+
 
   if (1)
   { // Naive wilson dag implementation
@@ -487,9 +509,9 @@ int main (int argc, char ** argv)
   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-4);
-  //assert(norm2(src_o)<1.0e-4);
-
+  assert(norm2(src_e)<1.0e-4);
+  assert(norm2(src_o)<1.0e-4);
   Grid_finalize();
+  exit(0);
 }
 

benchmarks/Benchmark_gparity.cc

@@ -0,0 +1,190 @@
+#include <Grid/Grid.h>
+#include <sstream>
+using namespace std;
+using namespace Grid;
+using namespace Grid::QCD;
+
+template<class d>
+struct scal {
+  d internal;
+};
+
+  Gamma::Algebra Gmu [] = {
+    Gamma::Algebra::GammaX,
+    Gamma::Algebra::GammaY,
+    Gamma::Algebra::GammaZ,
+    Gamma::Algebra::GammaT
+  };
+
+typedef typename GparityDomainWallFermionF::FermionField GparityLatticeFermionF;
+typedef typename GparityDomainWallFermionD::FermionField GparityLatticeFermionD;
+
+
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  int Ls=16;
+  for(int i=0;i<argc;i++)
+    if(std::string(argv[i]) == "-Ls"){
+      std::stringstream ss(argv[i+1]); ss >> Ls;
+    }
+
+
+  int threads = GridThread::GetThreads();
+  std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;
+  std::cout<<GridLogMessage << "Ls = " << Ls << std::endl;
+
+  std::vector<int> latt4 = GridDefaultLatt();
+
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplexF::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});
+  
+  std::cout << GridLogMessage << "Initialising 4d RNG" << std::endl;
+  GridParallelRNG          RNG4(UGrid);  RNG4.SeedFixedIntegers(seeds4);
+  std::cout << GridLogMessage << "Initialising 5d RNG" << std::endl;
+  GridParallelRNG          RNG5(FGrid);  RNG5.SeedFixedIntegers(seeds5);
+  std::cout << GridLogMessage << "Initialised RNGs" << std::endl;
+
+  GparityLatticeFermionF src   (FGrid); random(RNG5,src);
+  RealD N2 = 1.0/::sqrt(norm2(src));
+  src = src*N2;
+
+  GparityLatticeFermionF result(FGrid); result=zero;
+  GparityLatticeFermionF    ref(FGrid);    ref=zero;
+  GparityLatticeFermionF    tmp(FGrid);
+  GparityLatticeFermionF    err(FGrid);
+
+  std::cout << GridLogMessage << "Drawing gauge field" << std::endl;
+  LatticeGaugeFieldF Umu(UGrid); 
+  SU3::HotConfiguration(RNG4,Umu); 
+  std::cout << GridLogMessage << "Random gauge initialised " << std::endl;
+
+  RealD mass=0.1;
+  RealD M5  =1.8;
+
+  RealD NP = UGrid->_Nprocessors;
+  RealD NN = UGrid->NodeCount();
+
+  std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
+  std::cout << GridLogMessage<< "* Kernel options --dslash-generic, --dslash-unroll, --dslash-asm" <<std::endl;
+  std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
+  std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
+  std::cout << GridLogMessage<< "* Benchmarking DomainWallFermion::Dhop                  "<<std::endl;
+  std::cout << GridLogMessage<< "* Vectorising space-time by "<<vComplexF::Nsimd()<<std::endl;
+#ifdef GRID_OMP
+  if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute ) std::cout << GridLogMessage<< "* Using Overlapped Comms/Compute" <<std::endl;
+  if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsThenCompute) std::cout << GridLogMessage<< "* Using sequential comms compute" <<std::endl;
+#endif
+  if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGeneric   ) std::cout << GridLogMessage<< "* Using GENERIC Nc WilsonKernels" <<std::endl;
+  if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptHandUnroll) std::cout << GridLogMessage<< "* Using Nc=3       WilsonKernels" <<std::endl;
+  if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm ) std::cout << GridLogMessage<< "* Using Asm Nc=3   WilsonKernels" <<std::endl;
+  std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
+
+
+
+  std::cout << GridLogMessage<< "* SINGLE/SINGLE"<<std::endl;
+  GparityDomainWallFermionF Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+  int ncall =1000;
+  if (1) {
+    FGrid->Barrier();
+    Dw.ZeroCounters();
+    Dw.Dhop(src,result,0);
+    std::cout<<GridLogMessage<<"Called warmup"<<std::endl;
+    double t0=usecond();
+    for(int i=0;i<ncall;i++){
+      __SSC_START;
+      Dw.Dhop(src,result,0);
+      __SSC_STOP;
+    }
+    double t1=usecond();
+    FGrid->Barrier();
+    
+    double volume=Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
+    double flops=2*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 rank =  "<< flops/(t1-t0)/NP<<std::endl;
+    std::cout<<GridLogMessage << "mflop/s per node =  "<< flops/(t1-t0)/NN<<std::endl;
+    Dw.Report();
+  }
+
+  std::cout << GridLogMessage<< "* SINGLE/HALF"<<std::endl;
+  GparityDomainWallFermionFH DwH(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+  if (1) {
+    FGrid->Barrier();
+    DwH.ZeroCounters();
+    DwH.Dhop(src,result,0);
+    double t0=usecond();
+    for(int i=0;i<ncall;i++){
+      __SSC_START;
+      DwH.Dhop(src,result,0);
+      __SSC_STOP;
+    }
+    double t1=usecond();
+    FGrid->Barrier();
+    
+    double volume=Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
+    double flops=2*1344*volume*ncall;
+
+    std::cout<<GridLogMessage << "Called half prec comms Dw "<<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;
+    std::cout<<GridLogMessage << "mflop/s per node =  "<< flops/(t1-t0)/NN<<std::endl;
+    DwH.Report();
+  }
+
+  GridCartesian         * UGrid_d   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplexD::Nsimd()),GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid_d = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid_d);
+  GridCartesian         * FGrid_d   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid_d);
+  GridRedBlackCartesian * FrbGrid_d = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid_d);
+
+  
+  std::cout << GridLogMessage<< "* DOUBLE/DOUBLE"<<std::endl;
+  GparityLatticeFermionD src_d(FGrid_d);
+  precisionChange(src_d,src);
+
+  LatticeGaugeFieldD Umu_d(UGrid_d); 
+  precisionChange(Umu_d,Umu);
+
+  GparityLatticeFermionD result_d(FGrid_d);
+
+  GparityDomainWallFermionD DwD(Umu_d,*FGrid_d,*FrbGrid_d,*UGrid_d,*UrbGrid_d,mass,M5);
+  if (1) {
+    FGrid_d->Barrier();
+    DwD.ZeroCounters();
+    DwD.Dhop(src_d,result_d,0);
+    std::cout<<GridLogMessage<<"Called warmup"<<std::endl;
+    double t0=usecond();
+    for(int i=0;i<ncall;i++){
+      __SSC_START;
+      DwD.Dhop(src_d,result_d,0);
+      __SSC_STOP;
+    }
+    double t1=usecond();
+    FGrid_d->Barrier();
+    
+    double volume=Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
+    double flops=2*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 rank =  "<< flops/(t1-t0)/NP<<std::endl;
+    std::cout<<GridLogMessage << "mflop/s per node =  "<< flops/(t1-t0)/NN<<std::endl;
+    DwD.Report();
+  }
+
+  Grid_finalize();
+}
+

benchmarks/Benchmark_memory_bandwidth.cc

@@ -55,21 +55,21 @@ int main (int argc, char ** argv)
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
   std::cout<<GridLogMessage << "  L  "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s"<<"\t\t"<<"Gflop/s"<<"\t\t seconds"<<std::endl;
   std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
-  uint64_t lmax=64;
-#define NLOOP (100*lmax*lmax*lmax*lmax/vol)
-  for(int lat=4;lat<=lmax;lat+=4){
+  uint64_t lmax=96;
+#define NLOOP (10*lmax*lmax*lmax*lmax/vol)
+  for(int lat=8;lat<=lmax;lat+=8){
 
       std::vector<int> latt_size  ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
-      int vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
+      int64_t vol= latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
       GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
 
       uint64_t Nloop=NLOOP;
 
-      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9});
+      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9}));
 
-      LatticeVec z(&Grid); //random(pRNG,z);
-      LatticeVec x(&Grid); //random(pRNG,x);
-      LatticeVec y(&Grid); //random(pRNG,y);
+      LatticeVec z(&Grid);// random(pRNG,z);
+      LatticeVec x(&Grid);// random(pRNG,x);
+      LatticeVec y(&Grid);// random(pRNG,y);
       double a=2.0;
 
 
@@ -83,7 +83,7 @@ int main (int argc, char ** argv)
       double time = (stop-start)/Nloop*1000;
       
       double flops=vol*Nvec*2;// mul,add
-      double bytes=3*vol*Nvec*sizeof(Real);
+      double bytes=3.0*vol*Nvec*sizeof(Real);
       std::cout<<GridLogMessage<<std::setprecision(3) << lat<<"\t\t"<<bytes<<"   \t\t"<<bytes/time<<"\t\t"<<flops/time<<"\t\t"<<(stop-start)/1000./1000.<<std::endl;
 
     }
@@ -94,17 +94,17 @@ int main (int argc, char ** argv)
   std::cout<<GridLogMessage << "  L  "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s"<<"\t\t"<<"Gflop/s"<<"\t\t seconds"<<std::endl;
   std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
   
-  for(int lat=4;lat<=lmax;lat+=4){
+  for(int lat=8;lat<=lmax;lat+=8){
 
       std::vector<int> latt_size  ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
-      int vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
+      int64_t vol= latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
       GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
 
-      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9});
+      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9}));
 
-      LatticeVec z(&Grid); //random(pRNG,z);
-      LatticeVec x(&Grid); //random(pRNG,x);
-      LatticeVec y(&Grid); //random(pRNG,y);
+      LatticeVec z(&Grid);// random(pRNG,z);
+      LatticeVec x(&Grid);// random(pRNG,x);
+      LatticeVec y(&Grid);// random(pRNG,y);
       double a=2.0;
 
       uint64_t Nloop=NLOOP;
@@ -119,7 +119,7 @@ int main (int argc, char ** argv)
       double time = (stop-start)/Nloop*1000;
      
       double flops=vol*Nvec*2;// mul,add
-      double bytes=3*vol*Nvec*sizeof(Real);
+      double bytes=3.0*vol*Nvec*sizeof(Real);
       std::cout<<GridLogMessage<<std::setprecision(3) << lat<<"\t\t"<<bytes<<"   \t\t"<<bytes/time<<"\t\t"<<flops/time<<"\t\t"<<(stop-start)/1000./1000.<<std::endl;
 
     }
@@ -129,20 +129,20 @@ int main (int argc, char ** argv)
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
   std::cout<<GridLogMessage << "  L  "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s"<<"\t\t"<<"Gflop/s"<<"\t\t seconds"<<std::endl;
 
-  for(int lat=4;lat<=lmax;lat+=4){
+  for(int lat=8;lat<=lmax;lat+=8){
 
 
       std::vector<int> latt_size  ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
-      int vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
+      int64_t vol= latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
       uint64_t Nloop=NLOOP;
 
       GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
 
-      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9});
+      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9}));
 
-      LatticeVec z(&Grid); //random(pRNG,z);
-      LatticeVec x(&Grid); //random(pRNG,x);
-      LatticeVec y(&Grid); //random(pRNG,y);
+      LatticeVec z(&Grid);// random(pRNG,z);
+      LatticeVec x(&Grid);// random(pRNG,x);
+      LatticeVec y(&Grid);// random(pRNG,y);
       RealD a=2.0;
 
 
@@ -154,7 +154,7 @@ int main (int argc, char ** argv)
       double stop=usecond();
       double time = (stop-start)/Nloop*1000;
       
-      double bytes=2*vol*Nvec*sizeof(Real);
+      double bytes=2.0*vol*Nvec*sizeof(Real);
       double flops=vol*Nvec*1;// mul
       std::cout<<GridLogMessage <<std::setprecision(3) << lat<<"\t\t"<<bytes<<"   \t\t"<<bytes/time<<"\t\t"<<flops/time<<"\t\t"<<(stop-start)/1000./1000.<<std::endl;
 
@@ -166,17 +166,17 @@ int main (int argc, char ** argv)
   std::cout<<GridLogMessage << "  L  "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s"<<"\t\t"<<"Gflop/s"<<"\t\t seconds"<<std::endl;
   std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
 
-  for(int lat=4;lat<=lmax;lat+=4){
+  for(int lat=8;lat<=lmax;lat+=8){
 
       std::vector<int> latt_size  ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
-      int vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
+      int64_t vol= latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
       uint64_t Nloop=NLOOP;
       GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
 
-      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9});
-      LatticeVec z(&Grid); //random(pRNG,z);
-      LatticeVec x(&Grid); //random(pRNG,x);
-      LatticeVec y(&Grid); //random(pRNG,y);
+      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9}));
+      LatticeVec z(&Grid);// random(pRNG,z);
+      LatticeVec x(&Grid);// random(pRNG,x);
+      LatticeVec y(&Grid);// random(pRNG,y);
       RealD a=2.0;
       Real nn;      
       double start=usecond();
@@ -187,7 +187,7 @@ int main (int argc, char ** argv)
       double stop=usecond();
       double time = (stop-start)/Nloop*1000;
       
-      double bytes=vol*Nvec*sizeof(Real);
+      double bytes=1.0*vol*Nvec*sizeof(Real);
       double flops=vol*Nvec*2;// mul,add
       std::cout<<GridLogMessage<<std::setprecision(3) << lat<<"\t\t"<<bytes<<"  \t\t"<<bytes/time<<"\t\t"<<flops/time<< "\t\t"<<(stop-start)/1000./1000.<< "\t\t " <<std::endl;
 

benchmarks/Benchmark_staggered.cc

@@ -40,7 +40,7 @@ int main (int argc, char ** argv)
   std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
   std::vector<int> mpi_layout  = GridDefaultMpi();
   GridCartesian               Grid(latt_size,simd_layout,mpi_layout);
-  GridRedBlackCartesian     RBGrid(latt_size,simd_layout,mpi_layout);
+  GridRedBlackCartesian     RBGrid(&Grid);
 
   int threads = GridThread::GetThreads();
   std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;

benchmarks/Benchmark_su3.cc

@@ -37,12 +37,12 @@ int main (int argc, char ** argv)
   Grid_init(&argc,&argv);
 #define LMAX (64)
 
-  int Nloop=20;
+  int64_t Nloop=20;
 
   std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
   std::vector<int> mpi_layout  = GridDefaultMpi();
 
-  int threads = GridThread::GetThreads();
+  int64_t threads = GridThread::GetThreads();
   std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;
 
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
@@ -54,16 +54,16 @@ int main (int argc, char ** argv)
   for(int lat=2;lat<=LMAX;lat+=2){
 
       std::vector<int> latt_size  ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
-      int vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
+      int64_t vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
       GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
-      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9});
+      GridParallelRNG          pRNG(&Grid);      pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9}));
 
-      LatticeColourMatrix z(&Grid);// random(pRNG,z);
-      LatticeColourMatrix x(&Grid);// random(pRNG,x);
-      LatticeColourMatrix y(&Grid);// random(pRNG,y);
+      LatticeColourMatrix z(&Grid); random(pRNG,z);
+      LatticeColourMatrix x(&Grid); random(pRNG,x);
+      LatticeColourMatrix y(&Grid); random(pRNG,y);
 
       double start=usecond();
-      for(int i=0;i<Nloop;i++){
+      for(int64_t i=0;i<Nloop;i++){
 	x=x*y;
       }
       double stop=usecond();
@@ -86,17 +86,17 @@ int main (int argc, char ** argv)
   for(int lat=2;lat<=LMAX;lat+=2){
 
       std::vector<int> latt_size  ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
-      int vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
+      int64_t vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
 
       GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
-      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9});
+      GridParallelRNG          pRNG(&Grid);      pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9}));
 
-      LatticeColourMatrix z(&Grid); //random(pRNG,z);
-      LatticeColourMatrix x(&Grid); //random(pRNG,x);
-      LatticeColourMatrix y(&Grid); //random(pRNG,y);
+      LatticeColourMatrix z(&Grid); random(pRNG,z);
+      LatticeColourMatrix x(&Grid); random(pRNG,x);
+      LatticeColourMatrix y(&Grid); random(pRNG,y);
 
       double start=usecond();
-      for(int i=0;i<Nloop;i++){
+      for(int64_t i=0;i<Nloop;i++){
 	z=x*y;
       }
       double stop=usecond();
@@ -117,17 +117,17 @@ int main (int argc, char ** argv)
   for(int lat=2;lat<=LMAX;lat+=2){
 
       std::vector<int> latt_size  ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
-      int vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
+      int64_t vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
 
       GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
-      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9});
+      GridParallelRNG          pRNG(&Grid);      pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9}));
 
-      LatticeColourMatrix z(&Grid); //random(pRNG,z);
-      LatticeColourMatrix x(&Grid); //random(pRNG,x);
-      LatticeColourMatrix y(&Grid); //random(pRNG,y);
+      LatticeColourMatrix z(&Grid); random(pRNG,z);
+      LatticeColourMatrix x(&Grid); random(pRNG,x);
+      LatticeColourMatrix y(&Grid); random(pRNG,y);
 
       double start=usecond();
-      for(int i=0;i<Nloop;i++){
+      for(int64_t i=0;i<Nloop;i++){
 	mult(z,x,y);
       }
       double stop=usecond();
@@ -148,17 +148,17 @@ int main (int argc, char ** argv)
   for(int lat=2;lat<=LMAX;lat+=2){
 
       std::vector<int> latt_size  ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
-      int vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
+      int64_t vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
 
       GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
-      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9});
+      GridParallelRNG          pRNG(&Grid);      pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9}));
 
-      LatticeColourMatrix z(&Grid); //random(pRNG,z);
-      LatticeColourMatrix x(&Grid); //random(pRNG,x);
-      LatticeColourMatrix y(&Grid); //random(pRNG,y);
+      LatticeColourMatrix z(&Grid); random(pRNG,z);
+      LatticeColourMatrix x(&Grid); random(pRNG,x);
+      LatticeColourMatrix y(&Grid); random(pRNG,y);
 
       double start=usecond();
-      for(int i=0;i<Nloop;i++){
+      for(int64_t i=0;i<Nloop;i++){
 	mac(z,x,y);
       }
       double stop=usecond();

benchmarks/Benchmark_wilson.cc

@@ -58,7 +58,7 @@ int main (int argc, char ** argv)
   std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
   std::vector<int> mpi_layout  = GridDefaultMpi();
   GridCartesian               Grid(latt_size,simd_layout,mpi_layout);
-  GridRedBlackCartesian     RBGrid(latt_size,simd_layout,mpi_layout);
+  GridRedBlackCartesian     RBGrid(&Grid);
 
   int threads = GridThread::GetThreads();
   std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;

benchmarks/Benchmark_wilson_sweep.cc

@@ -93,7 +93,7 @@ int main (int argc, char ** argv)
 	  std::cout << latt_size.back() << "\t\t";
 
 	  GridCartesian           Grid(latt_size,simd_layout,mpi_layout);
-	  GridRedBlackCartesian RBGrid(latt_size,simd_layout,mpi_layout);
+	  GridRedBlackCartesian RBGrid(&Grid);
 
 	  GridParallelRNG  pRNG(&Grid); pRNG.SeedFixedIntegers(seeds);
 	  LatticeGaugeField Umu(&Grid); random(pRNG,Umu);

configure.ac

@@ -13,6 +13,10 @@ m4_ifdef([AM_SILENT_RULES], [AM_SILENT_RULES([yes])])
 ################ Get git info
 #AC_REVISION([m4_esyscmd_s([./scripts/configure.commit])])
 
+################ Set flags
+# do not move!
+CXXFLAGS="-O3 $CXXFLAGS"
+
 ############### Checks for programs
 AC_PROG_CXX
 AC_PROG_RANLIB
@@ -27,7 +31,6 @@ AX_GXX_VERSION
 AC_DEFINE_UNQUOTED([GXX_VERSION],["$GXX_VERSION"],
       [version of g++ that will compile the code])
 
-CXXFLAGS="-g $CXXFLAGS"
 
 
 ############### Checks for typedefs, structures, and compiler characteristics
@@ -51,9 +54,14 @@ AC_CHECK_HEADERS(malloc/malloc.h)
 AC_CHECK_HEADERS(malloc.h)
 AC_CHECK_HEADERS(endian.h)
 AC_CHECK_HEADERS(execinfo.h)
+AC_CHECK_HEADERS(numaif.h)
 AC_CHECK_DECLS([ntohll],[], [], [[#include <arpa/inet.h>]])
 AC_CHECK_DECLS([be64toh],[], [], [[#include <arpa/inet.h>]])
 
+############## Standard libraries
+AC_CHECK_LIB([m],[cos])
+AC_CHECK_LIB([stdc++],[abort])
+
 ############### GMP and MPFR
 AC_ARG_WITH([gmp],
     [AS_HELP_STRING([--with-gmp=prefix],
@@ -186,9 +194,14 @@ Info at: http://usqcd.jlab.org/usqcd-docs/c-lime/)])
 
 AC_SEARCH_LIBS([crc32], [z],
                [AC_DEFINE([HAVE_ZLIB], [1], [Define to 1 if you have the `LIBZ' library])]
-               [have_zlib=true],
+               [have_zlib=true] [LIBS="${LIBS} -lz"],
 	       [AC_MSG_ERROR(zlib library was not found in your system.)])
 
+AC_SEARCH_LIBS([move_pages], [numa],
+               [AC_DEFINE([HAVE_LIBNUMA], [1], [Define to 1 if you have the `LIBNUMA' library])]
+               [have_libnuma=true] [LIBS="${LIBS} -lnuma"],
+	       [AC_MSG_WARN(libnuma library was not found in your system. Some optimisations will not apply)])
+
 AC_SEARCH_LIBS([H5Fopen], [hdf5_cpp],
                [AC_DEFINE([HAVE_HDF5], [1], [Define to 1 if you have the `HDF5' library])]
                [have_hdf5=true]
@@ -241,6 +254,7 @@ case ${ax_cv_cxx_compiler_vendor} in
         SIMD_FLAGS='';;
       KNL)
         AC_DEFINE([AVX512],[1],[AVX512 intrinsics])
+        AC_DEFINE([KNL],[1],[Knights landing processor])
         SIMD_FLAGS='-march=knl';;
       GEN)
         AC_DEFINE([GEN],[1],[generic vector code])
@@ -248,6 +262,9 @@ case ${ax_cv_cxx_compiler_vendor} in
                            [generic SIMD vector width (in bytes)])
         SIMD_GEN_WIDTH_MSG=" (width= $ac_gen_simd_width)"
         SIMD_FLAGS='';;
+      NEONv8)
+        AC_DEFINE([NEONV8],[1],[ARMv8 NEON])
+        SIMD_FLAGS='-march=armv8-a';;
       QPX|BGQ)
         AC_DEFINE([QPX],[1],[QPX intrinsics for BG/Q])
         SIMD_FLAGS='';;
@@ -276,6 +293,7 @@ case ${ax_cv_cxx_compiler_vendor} in
         SIMD_FLAGS='';;
       KNL)
         AC_DEFINE([AVX512],[1],[AVX512 intrinsics for Knights Landing])
+        AC_DEFINE([KNL],[1],[Knights landing processor])
         SIMD_FLAGS='-xmic-avx512';;
       GEN)
         AC_DEFINE([GEN],[1],[generic vector code])
@@ -313,8 +331,41 @@ case ${ac_PRECISION} in
      double)
        AC_DEFINE([GRID_DEFAULT_PRECISION_DOUBLE],[1],[GRID_DEFAULT_PRECISION is DOUBLE] )
      ;;
+     *)
+     AC_MSG_ERROR([${ac_PRECISION} unsupported --enable-precision option]);
+     ;;
 esac
 
+######################  Shared memory allocation technique under MPI3
+AC_ARG_ENABLE([shm],[AC_HELP_STRING([--enable-shm=shmget|shmopen|hugetlbfs],
+              [Select SHM allocation technique])],[ac_SHM=${enable_shm}],[ac_SHM=shmopen])
+
+case ${ac_SHM} in
+
+     shmget)
+     AC_DEFINE([GRID_MPI3_SHMGET],[1],[GRID_MPI3_SHMGET] )
+     ;;
+
+     shmopen)
+     AC_DEFINE([GRID_MPI3_SHMOPEN],[1],[GRID_MPI3_SHMOPEN] )
+     ;;
+
+     hugetlbfs)
+     AC_DEFINE([GRID_MPI3_SHMMMAP],[1],[GRID_MPI3_SHMMMAP] )
+     ;;
+
+     *)
+     AC_MSG_ERROR([${ac_SHM} unsupported --enable-shm option]);
+     ;;
+esac
+
+######################  Shared base path for SHMMMAP
+AC_ARG_ENABLE([shmpath],[AC_HELP_STRING([--enable-shmpath=path],
+              [Select SHM mmap base path for hugetlbfs])],
+	      [ac_SHMPATH=${enable_shmpath}],
+	      [ac_SHMPATH=/var/lib/hugetlbfs/pagesize-2MB/])
+AC_DEFINE_UNQUOTED([GRID_SHM_PATH],["$ac_SHMPATH"],[Path to a hugetlbfs filesystem for MMAPing])
+
 ############### communication type selection
 AC_ARG_ENABLE([comms],[AC_HELP_STRING([--enable-comms=none|mpi|mpi-auto|mpi3|mpi3-auto|shmem],
               [Select communications])],[ac_COMMS=${enable_comms}],[ac_COMMS=none])
@@ -324,14 +375,14 @@ case ${ac_COMMS} in
         AC_DEFINE([GRID_COMMS_NONE],[1],[GRID_COMMS_NONE] )
         comms_type='none'
      ;;
-     mpi3l*)
-       AC_DEFINE([GRID_COMMS_MPI3L],[1],[GRID_COMMS_MPI3L] )
-       comms_type='mpi3l'
-     ;;
      mpi3*)
         AC_DEFINE([GRID_COMMS_MPI3],[1],[GRID_COMMS_MPI3] )
         comms_type='mpi3'
      ;;
+     mpit)
+        AC_DEFINE([GRID_COMMS_MPIT],[1],[GRID_COMMS_MPIT] )
+        comms_type='mpit'
+     ;;
      mpi*)
         AC_DEFINE([GRID_COMMS_MPI],[1],[GRID_COMMS_MPI] )
         comms_type='mpi'
@@ -359,7 +410,7 @@ esac
 AM_CONDITIONAL(BUILD_COMMS_SHMEM, [ test "${comms_type}X" == "shmemX" ])
 AM_CONDITIONAL(BUILD_COMMS_MPI,   [ test "${comms_type}X" == "mpiX" ])
 AM_CONDITIONAL(BUILD_COMMS_MPI3,  [ test "${comms_type}X" == "mpi3X" ] )
-AM_CONDITIONAL(BUILD_COMMS_MPI3L, [ test "${comms_type}X" == "mpi3lX" ] )
+AM_CONDITIONAL(BUILD_COMMS_MPIT,  [ test "${comms_type}X" == "mpitX" ] )
 AM_CONDITIONAL(BUILD_COMMS_NONE,  [ test "${comms_type}X" == "noneX" ])
 
 ############### RNG selection
@@ -464,6 +515,8 @@ compiler version            : ${ax_cv_gxx_version}
 SIMD                        : ${ac_SIMD}${SIMD_GEN_WIDTH_MSG}
 Threading                   : ${ac_openmp}
 Communications type         : ${comms_type}
+Shared memory allocator     : ${ac_SHM}
+Shared memory mmap path     : ${ac_SHMPATH}
 Default precision           : ${ac_PRECISION}
 Software FP16 conversion    : ${ac_SFW_FP16}
 RNG choice                  : ${ac_RNG}
@@ -497,6 +550,7 @@ AC_CONFIG_FILES(tests/forces/Makefile)
 AC_CONFIG_FILES(tests/hadrons/Makefile)
 AC_CONFIG_FILES(tests/hmc/Makefile)
 AC_CONFIG_FILES(tests/solver/Makefile)
+AC_CONFIG_FILES(tests/lanczos/Makefile)
 AC_CONFIG_FILES(tests/smearing/Makefile)
 AC_CONFIG_FILES(tests/qdpxx/Makefile)
 AC_CONFIG_FILES(tests/testu01/Makefile)

extras/Hadrons/Environment.cc

@@ -41,9 +41,10 @@ using namespace Hadrons;
 // constructor /////////////////////////////////////////////////////////////////
 Environment::Environment(void)
 {
-    nd_ = GridDefaultLatt().size();
+    dim_ = GridDefaultLatt();
+    nd_  = dim_.size();
     grid4d_.reset(SpaceTimeGrid::makeFourDimGrid(
-        GridDefaultLatt(), GridDefaultSimd(nd_, vComplex::Nsimd()),
+        dim_, GridDefaultSimd(nd_, vComplex::Nsimd()),
         GridDefaultMpi()));
     gridRb4d_.reset(SpaceTimeGrid::makeFourDimRedBlackGrid(grid4d_.get()));
     auto loc = getGrid()->LocalDimensions();
@@ -132,6 +133,16 @@ unsigned int Environment::getNd(void) const
     return nd_;
 }
 
+std::vector<int> Environment::getDim(void) const
+{
+    return dim_;
+}
+
+int Environment::getDim(const unsigned int mu) const
+{
+    return dim_[mu];
+}
+
 // random number generator /////////////////////////////////////////////////////
 void Environment::setSeed(const std::vector<int> &seed)
 {
@@ -271,6 +282,21 @@ std::string Environment::getModuleType(const std::string name) const
     return getModuleType(getModuleAddress(name));
 }
 
+std::string Environment::getModuleNamespace(const unsigned int address) const
+{
+    std::string type = getModuleType(address), ns;
+    
+    auto pos2 = type.rfind("::");
+    auto pos1 = type.rfind("::", pos2 - 2);
+    
+    return type.substr(pos1 + 2, pos2 - pos1 - 2);
+}
+
+std::string Environment::getModuleNamespace(const std::string name) const
+{
+    return getModuleNamespace(getModuleAddress(name));
+}
+
 bool Environment::hasModule(const unsigned int address) const
 {
     return (address < module_.size());
@@ -491,9 +517,16 @@ std::string Environment::getObjectName(const unsigned int address) const
 std::string Environment::getObjectType(const unsigned int address) const
 {
     if (hasRegisteredObject(address))
+    {
+        if (object_[address].type)
         {
             return typeName(object_[address].type);
         }
+        else
+        {
+            return "<no type>";
+        }
+    }
     else if (hasObject(address))
     {
         HADRON_ERROR("object with address " + std::to_string(address)
@@ -532,6 +565,23 @@ Environment::Size Environment::getObjectSize(const std::string name) const
     return getObjectSize(getObjectAddress(name));
 }
 
+unsigned int Environment::getObjectModule(const unsigned int address) const
+{
+    if (hasObject(address))
+    {
+        return object_[address].module;
+    }
+    else
+    {
+        HADRON_ERROR("no object with address " + std::to_string(address));
+    }
+}
+
+unsigned int Environment::getObjectModule(const std::string name) const
+{
+    return getObjectModule(getObjectAddress(name));
+}
+
 unsigned int Environment::getObjectLs(const unsigned int address) const
 {
     if (hasRegisteredObject(address))

extras/Hadrons/Environment.hpp

@@ -106,6 +106,8 @@ public:
     void                    createGrid(const unsigned int Ls);
     GridCartesian *         getGrid(const unsigned int Ls = 1) const;
     GridRedBlackCartesian * getRbGrid(const unsigned int Ls = 1) const;
+    std::vector<int>        getDim(void) const;
+    int                     getDim(const unsigned int mu) const;
     unsigned int            getNd(void) const;
     // random number generator
     void                    setSeed(const std::vector<int> &seed);
@@ -131,6 +133,8 @@ public:
     std::string             getModuleName(const unsigned int address) const;
     std::string             getModuleType(const unsigned int address) const;
     std::string             getModuleType(const std::string name) const;
+    std::string             getModuleNamespace(const unsigned int address) const;
+    std::string             getModuleNamespace(const std::string name) const;
     bool                    hasModule(const unsigned int address) const;
     bool                    hasModule(const std::string name) const;
     Graph<unsigned int>     makeModuleGraph(void) const;
@@ -171,6 +175,8 @@ public:
     std::string             getObjectType(const std::string name) const;
     Size                    getObjectSize(const unsigned int address) const;
     Size                    getObjectSize(const std::string name) const;
+    unsigned int            getObjectModule(const unsigned int address) const;
+    unsigned int            getObjectModule(const std::string name) const;
     unsigned int            getObjectLs(const unsigned int address) const;
     unsigned int            getObjectLs(const std::string name) const;
     bool                    hasObject(const unsigned int address) const;
@@ -181,6 +187,10 @@ public:
     bool                    hasCreatedObject(const std::string name) const;
     bool                    isObject5d(const unsigned int address) const;
     bool                    isObject5d(const std::string name) const;
+    template <typename T>
+    bool                    isObjectOfType(const unsigned int address) const;
+    template <typename T>
+    bool                    isObjectOfType(const std::string name) const;
     Environment::Size       getTotalSize(void) const;
     void                    addOwnership(const unsigned int owner,
                                          const unsigned int property);
@@ -197,6 +207,7 @@ private:
     bool                                   dryRun_{false};
     unsigned int                           traj_, locVol_;
     // grids
+    std::vector<int>                       dim_;
     GridPt                                 grid4d_;
     std::map<unsigned int, GridPt>         grid5d_;
     GridRbPt                               gridRb4d_;
@@ -343,7 +354,7 @@ T * Environment::getObject(const unsigned int address) const
         else
         {
             HADRON_ERROR("object with address " + std::to_string(address) +
-                         " does not have type '" + typeid(T).name() +
+                         " does not have type '" + typeName(&typeid(T)) +
                          "' (has type '" + getObjectType(address) + "')");
         }
     }
@@ -380,6 +391,37 @@ T * Environment::createLattice(const std::string name)
     return createLattice<T>(getObjectAddress(name));
 }
 
+template <typename T>
+bool Environment::isObjectOfType(const unsigned int address) const
+{
+    if (hasRegisteredObject(address))
+    {
+        if (auto h = dynamic_cast<Holder<T> *>(object_[address].data.get()))
+        {
+            return true;
+        }
+        else
+        {
+            return false;
+        }
+    }
+    else if (hasObject(address))
+    {
+        HADRON_ERROR("object with address " + std::to_string(address) +
+                     " exists but is not registered");
+    }
+    else
+    {
+        HADRON_ERROR("no object with address " + std::to_string(address));
+    }
+}
+
+template <typename T>
+bool Environment::isObjectOfType(const std::string name) const
+{
+    return isObjectOfType<T>(getObjectAddress(name));
+}
+
 END_HADRONS_NAMESPACE
 
 #endif // Hadrons_Environment_hpp_

extras/Hadrons/Global.hpp

@@ -51,23 +51,43 @@ using Grid::operator<<;
  * error with GCC 5 (clang & GCC 6 compile fine without it).
  */
 
-// FIXME: find a way to do that in a more general fashion
 #ifndef FIMPL
 #define FIMPL WilsonImplR
 #endif
+#ifndef SIMPL
+#define SIMPL ScalarImplCR
+#endif
 
 BEGIN_HADRONS_NAMESPACE
 
 // type aliases
-#define TYPE_ALIASES(FImpl, suffix)\
+#define FERM_TYPE_ALIASES(FImpl, suffix)\
 typedef FermionOperator<FImpl>                       FMat##suffix;             \
 typedef typename FImpl::FermionField                 FermionField##suffix;     \
 typedef typename FImpl::PropagatorField              PropagatorField##suffix;  \
 typedef typename FImpl::SitePropagator               SitePropagator##suffix;   \
-typedef typename FImpl::DoubledGaugeField            DoubledGaugeField##suffix;\
-typedef std::function<void(FermionField##suffix &,                             \
+typedef std::vector<typename FImpl::SitePropagator::scalar_object>             \
+                                                     SlicedPropagator##suffix;
+
+#define GAUGE_TYPE_ALIASES(FImpl, suffix)\
+typedef typename FImpl::DoubledGaugeField DoubledGaugeField##suffix;
+
+#define SCALAR_TYPE_ALIASES(SImpl, suffix)\
+typedef typename SImpl::Field ScalarField##suffix;\
+typedef typename SImpl::Field PropagatorField##suffix;
+
+#define SOLVER_TYPE_ALIASES(FImpl, suffix)\
+typedef std::function<void(FermionField##suffix &,\
                       const FermionField##suffix &)> SolverFn##suffix;
 
+#define SINK_TYPE_ALIASES(suffix)\
+typedef std::function<SlicedPropagator##suffix(const PropagatorField##suffix &)> SinkFn##suffix;
+
+#define FGS_TYPE_ALIASES(FImpl, suffix)\
+FERM_TYPE_ALIASES(FImpl, suffix)\
+GAUGE_TYPE_ALIASES(FImpl, suffix)\
+SOLVER_TYPE_ALIASES(FImpl, suffix)
+
 // logger
 class HadronsLogger: public Logger
 {

extras/Hadrons/Modules.hpp

@@ -1,31 +1,3 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid 
-
-Source file: extras/Hadrons/Modules.hpp
-
-Copyright (C) 2015
-Copyright (C) 2016
-
-Author: Antonin Portelli <antonin.portelli@me.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/Hadrons/Modules/MAction/DWF.hpp>
 #include <Grid/Hadrons/Modules/MAction/Wilson.hpp>
 #include <Grid/Hadrons/Modules/MContraction/Baryon.hpp>
@@ -36,13 +8,18 @@ See the full license in the file "LICENSE" in the top level distribution directo
 #include <Grid/Hadrons/Modules/MContraction/WeakHamiltonianEye.hpp>
 #include <Grid/Hadrons/Modules/MContraction/WeakHamiltonianNonEye.hpp>
 #include <Grid/Hadrons/Modules/MContraction/WeakNeutral4ptDisc.hpp>
+#include <Grid/Hadrons/Modules/MFermion/GaugeProp.hpp>
 #include <Grid/Hadrons/Modules/MGauge/Load.hpp>
 #include <Grid/Hadrons/Modules/MGauge/Random.hpp>
+#include <Grid/Hadrons/Modules/MGauge/StochEm.hpp>
 #include <Grid/Hadrons/Modules/MGauge/Unit.hpp>
 #include <Grid/Hadrons/Modules/MLoop/NoiseLoop.hpp>
+#include <Grid/Hadrons/Modules/MScalar/ChargedProp.hpp>
+#include <Grid/Hadrons/Modules/MScalar/FreeProp.hpp>
+#include <Grid/Hadrons/Modules/MScalar/Scalar.hpp>
+#include <Grid/Hadrons/Modules/MSink/Point.hpp>
 #include <Grid/Hadrons/Modules/MSolver/RBPrecCG.hpp>
 #include <Grid/Hadrons/Modules/MSource/Point.hpp>
 #include <Grid/Hadrons/Modules/MSource/SeqGamma.hpp>
 #include <Grid/Hadrons/Modules/MSource/Wall.hpp>
 #include <Grid/Hadrons/Modules/MSource/Z2.hpp>
-#include <Grid/Hadrons/Modules/Quark.hpp>

extras/Hadrons/Modules/MAction/DWF.hpp

@@ -27,8 +27,8 @@ See the full license in the file "LICENSE" in the top level distribution directo
 *************************************************************************************/
 /*  END LEGAL */
 
-#ifndef Hadrons_DWF_hpp_
-#define Hadrons_DWF_hpp_
+#ifndef Hadrons_MAction_DWF_hpp_
+#define Hadrons_MAction_DWF_hpp_
 
 #include <Grid/Hadrons/Global.hpp>
 #include <Grid/Hadrons/Module.hpp>
@@ -56,7 +56,7 @@ template <typename FImpl>
 class TDWF: public Module<DWFPar>
 {
 public:
-    TYPE_ALIASES(FImpl,);
+    FGS_TYPE_ALIASES(FImpl,);
 public:
     // constructor
     TDWF(const std::string name);
@@ -137,4 +137,4 @@ END_MODULE_NAMESPACE
 
 END_HADRONS_NAMESPACE
 
-#endif // Hadrons_DWF_hpp_
+#endif // Hadrons_MAction_DWF_hpp_

extras/Hadrons/Modules/MAction/Wilson.hpp

@@ -27,8 +27,8 @@ See the full license in the file "LICENSE" in the top level distribution directo
 *************************************************************************************/
 /*  END LEGAL */
 
-#ifndef Hadrons_Wilson_hpp_
-#define Hadrons_Wilson_hpp_
+#ifndef Hadrons_MAction_Wilson_hpp_
+#define Hadrons_MAction_Wilson_hpp_
 
 #include <Grid/Hadrons/Global.hpp>
 #include <Grid/Hadrons/Module.hpp>
@@ -54,7 +54,7 @@ template <typename FImpl>
 class TWilson: public Module<WilsonPar>
 {
 public:
-    TYPE_ALIASES(FImpl,);
+    FGS_TYPE_ALIASES(FImpl,);
 public:
     // constructor
     TWilson(const std::string name);

extras/Hadrons/Modules/MContraction/Baryon.hpp

@@ -27,8 +27,8 @@ See the full license in the file "LICENSE" in the top level distribution directo
 *************************************************************************************/
 /*  END LEGAL */
 
-#ifndef Hadrons_Baryon_hpp_
-#define Hadrons_Baryon_hpp_
+#ifndef Hadrons_MContraction_Baryon_hpp_
+#define Hadrons_MContraction_Baryon_hpp_
 
 #include <Grid/Hadrons/Global.hpp>
 #include <Grid/Hadrons/Module.hpp>
@@ -55,9 +55,9 @@ template <typename FImpl1, typename FImpl2, typename FImpl3>
 class TBaryon: public Module<BaryonPar>
 {
 public:
-    TYPE_ALIASES(FImpl1, 1);
-    TYPE_ALIASES(FImpl2, 2);
-    TYPE_ALIASES(FImpl3, 3);
+    FERM_TYPE_ALIASES(FImpl1, 1);
+    FERM_TYPE_ALIASES(FImpl2, 2);
+    FERM_TYPE_ALIASES(FImpl3, 3);
     class Result: Serializable
     {
     public:
@@ -121,11 +121,11 @@ void TBaryon<FImpl1, FImpl2, FImpl3>::execute(void)
     
     // FIXME: do contractions
     
-    write(writer, "meson", result);
+    // write(writer, "meson", result);
 }
 
 END_MODULE_NAMESPACE
 
 END_HADRONS_NAMESPACE
 
-#endif // Hadrons_Baryon_hpp_
+#endif // Hadrons_MContraction_Baryon_hpp_

extras/Hadrons/Modules/MContraction/DiscLoop.hpp

@@ -26,8 +26,8 @@ See the full license in the file "LICENSE" in the top level distribution directo
 *************************************************************************************/
 /*  END LEGAL */
 
-#ifndef Hadrons_DiscLoop_hpp_
-#define Hadrons_DiscLoop_hpp_
+#ifndef Hadrons_MContraction_DiscLoop_hpp_
+#define Hadrons_MContraction_DiscLoop_hpp_
 
 #include <Grid/Hadrons/Global.hpp>
 #include <Grid/Hadrons/Module.hpp>
@@ -52,7 +52,7 @@ public:
 template <typename FImpl>
 class TDiscLoop: public Module<DiscLoopPar>
 {
-    TYPE_ALIASES(FImpl,);
+    FERM_TYPE_ALIASES(FImpl,);
     class Result: Serializable
     {
     public:
@@ -141,4 +141,4 @@ END_MODULE_NAMESPACE
 
 END_HADRONS_NAMESPACE
 
-#endif // Hadrons_DiscLoop_hpp_
+#endif // Hadrons_MContraction_DiscLoop_hpp_

extras/Hadrons/Modules/MContraction/Gamma3pt.hpp

@@ -26,8 +26,8 @@ See the full license in the file "LICENSE" in the top level distribution directo
 *************************************************************************************/
 /*  END LEGAL */
 
-#ifndef Hadrons_Gamma3pt_hpp_
-#define Hadrons_Gamma3pt_hpp_
+#ifndef Hadrons_MContraction_Gamma3pt_hpp_
+#define Hadrons_MContraction_Gamma3pt_hpp_
 
 #include <Grid/Hadrons/Global.hpp>
 #include <Grid/Hadrons/Module.hpp>
@@ -72,9 +72,9 @@ public:
 template <typename FImpl1, typename FImpl2, typename FImpl3>
 class TGamma3pt: public Module<Gamma3ptPar>
 {
-    TYPE_ALIASES(FImpl1, 1);
-    TYPE_ALIASES(FImpl2, 2);
-    TYPE_ALIASES(FImpl3, 3);
+    FERM_TYPE_ALIASES(FImpl1, 1);
+    FERM_TYPE_ALIASES(FImpl2, 2);
+    FERM_TYPE_ALIASES(FImpl3, 3);
     class Result: Serializable
     {
     public:
@@ -167,4 +167,4 @@ END_MODULE_NAMESPACE
 
 END_HADRONS_NAMESPACE
 
-#endif // Hadrons_Gamma3pt_hpp_
+#endif // Hadrons_MContraction_Gamma3pt_hpp_

extras/Hadrons/Modules/MContraction/Meson.hpp

@@ -29,8 +29,8 @@ See the full license in the file "LICENSE" in the top level distribution directo
 *************************************************************************************/
 /*  END LEGAL */
 
-#ifndef Hadrons_Meson_hpp_
-#define Hadrons_Meson_hpp_
+#ifndef Hadrons_MContraction_Meson_hpp_
+#define Hadrons_MContraction_Meson_hpp_
 
 #include <Grid/Hadrons/Global.hpp>
 #include <Grid/Hadrons/Module.hpp>
@@ -69,7 +69,7 @@ public:
                                     std::string, q1,
                                     std::string, q2,
                                     std::string, gammas,
-                                    std::string, mom,
+                                    std::string, sink,
                                     std::string, output);
 };
 
@@ -77,8 +77,10 @@ template <typename FImpl1, typename FImpl2>
 class TMeson: public Module<MesonPar>
 {
 public:
-    TYPE_ALIASES(FImpl1, 1);
-    TYPE_ALIASES(FImpl2, 2);
+    FERM_TYPE_ALIASES(FImpl1, 1);
+    FERM_TYPE_ALIASES(FImpl2, 2);
+    FERM_TYPE_ALIASES(ScalarImplCR, Scalar);
+    SINK_TYPE_ALIASES(Scalar);
     class Result: Serializable
     {
     public:
@@ -115,7 +117,7 @@ TMeson<FImpl1, FImpl2>::TMeson(const std::string name)
 template <typename FImpl1, typename FImpl2>
 std::vector<std::string> TMeson<FImpl1, FImpl2>::getInput(void)
 {
-    std::vector<std::string> input = {par().q1, par().q2};
+    std::vector<std::string> input = {par().q1, par().q2, par().sink};
     
     return input;
 }
@@ -154,6 +156,9 @@ void TMeson<FImpl1, FImpl2>::parseGammaString(std::vector<GammaPair> &gammaList)
 
 
 // execution ///////////////////////////////////////////////////////////////////
+#define mesonConnected(q1, q2, gSnk, gSrc) \
+(g5*(gSnk))*(q1)*(adj(gSrc)*g5)*adj(q2)
+
 template <typename FImpl1, typename FImpl2>
 void TMeson<FImpl1, FImpl2>::execute(void)
 {
@@ -162,44 +167,73 @@ void TMeson<FImpl1, FImpl2>::execute(void)
                  << std::endl;
     
     CorrWriter             writer(par().output);
-    PropagatorField1       &q1 = *env().template getObject<PropagatorField1>(par().q1);
-    PropagatorField2       &q2 = *env().template getObject<PropagatorField2>(par().q2);
-    LatticeComplex         c(env().getGrid());
-    Gamma                  g5(Gamma::Algebra::Gamma5);
-    std::vector<GammaPair> gammaList;
     std::vector<TComplex>  buf;
     std::vector<Result>    result;
-    std::vector<Real>      p;
-
-    p  = strToVec<Real>(par().mom);
-    LatticeComplex         ph(env().getGrid()), coor(env().getGrid());
-    Complex                i(0.0,1.0);
-    ph = zero;
-    for(unsigned int mu = 0; mu < env().getNd(); mu++)
-    {
-        LatticeCoordinate(coor, mu);
-        ph = ph + p[mu]*coor*((1./(env().getGrid()->_fdimensions[mu])));
-    }
-    ph = exp((Real)(2*M_PI)*i*ph);
+    Gamma                  g5(Gamma::Algebra::Gamma5);
+    std::vector<GammaPair> gammaList;
+    int                    nt = env().getDim(Tp);
     
     parseGammaString(gammaList);
-
     result.resize(gammaList.size());
     for (unsigned int i = 0; i < result.size(); ++i)
+    {
+        result[i].gamma_snk = gammaList[i].first;
+        result[i].gamma_src = gammaList[i].second;
+        result[i].corr.resize(nt);
+    }
+    if (env().template isObjectOfType<SlicedPropagator1>(par().q1) and
+        env().template isObjectOfType<SlicedPropagator2>(par().q2))
+    {
+        SlicedPropagator1 &q1 = *env().template getObject<SlicedPropagator1>(par().q1);
+        SlicedPropagator2 &q2 = *env().template getObject<SlicedPropagator2>(par().q2);
+        
+        LOG(Message) << "(propagator already sinked)" << std::endl;
+        for (unsigned int i = 0; i < result.size(); ++i)
         {
             Gamma gSnk(gammaList[i].first);
             Gamma gSrc(gammaList[i].second);
-        c = trace((g5*gSnk)*q1*(adj(gSrc)*g5)*adj(q2))*ph;
-        sliceSum(c, buf, Tp);
             
-        result[i].gamma_snk = gammaList[i].first;
-        result[i].gamma_src = gammaList[i].second;
-        result[i].corr.resize(buf.size());
+            for (unsigned int t = 0; t < buf.size(); ++t)
+            {
+                result[i].corr[t] = TensorRemove(trace(mesonConnected(q1[t], q2[t], gSnk, gSrc)));
+            }
+        }
+    }
+    else
+    {
+        PropagatorField1 &q1   = *env().template getObject<PropagatorField1>(par().q1);
+        PropagatorField2 &q2   = *env().template getObject<PropagatorField2>(par().q2);
+        LatticeComplex   c(env().getGrid());
+        
+        LOG(Message) << "(using sink '" << par().sink << "')" << std::endl;
+        for (unsigned int i = 0; i < result.size(); ++i)
+        {
+            Gamma       gSnk(gammaList[i].first);
+            Gamma       gSrc(gammaList[i].second);
+            std::string ns;
+                
+            ns = env().getModuleNamespace(env().getObjectModule(par().sink));
+            if (ns == "MSource")
+            {
+                PropagatorField1 &sink =
+                    *env().template getObject<PropagatorField1>(par().sink);
+                
+                c = trace(mesonConnected(q1, q2, gSnk, gSrc)*sink);
+                sliceSum(c, buf, Tp);
+            }
+            else if (ns == "MSink")
+            {
+                SinkFnScalar &sink = *env().template getObject<SinkFnScalar>(par().sink);
+                
+                c   = trace(mesonConnected(q1, q2, gSnk, gSrc));
+                buf = sink(c);
+            }
             for (unsigned int t = 0; t < buf.size(); ++t)
             {
                 result[i].corr[t] = TensorRemove(buf[t]);
             }
         }
+    }
     write(writer, "meson", result);
 }
 
@@ -207,4 +241,4 @@ END_MODULE_NAMESPACE
 
 END_HADRONS_NAMESPACE
 
-#endif // Hadrons_Meson_hpp_
+#endif // Hadrons_MContraction_Meson_hpp_

extras/Hadrons/Modules/MContraction/WeakHamiltonian.hpp

@@ -26,8 +26,8 @@ See the full license in the file "LICENSE" in the top level distribution directo
 *************************************************************************************/
 /*  END LEGAL */
 
-#ifndef Hadrons_WeakHamiltonian_hpp_
-#define Hadrons_WeakHamiltonian_hpp_
+#ifndef Hadrons_MContraction_WeakHamiltonian_hpp_
+#define Hadrons_MContraction_WeakHamiltonian_hpp_
 
 #include <Grid/Hadrons/Global.hpp>
 #include <Grid/Hadrons/Module.hpp>
@@ -83,7 +83,7 @@ public:
 class T##modname: public Module<WeakHamiltonianPar>\
 {\
 public:\
-    TYPE_ALIASES(FIMPL,)\
+    FERM_TYPE_ALIASES(FIMPL,)\
     class Result: Serializable\
     {\
     public:\
@@ -111,4 +111,4 @@ END_MODULE_NAMESPACE
 
 END_HADRONS_NAMESPACE
 
-#endif // Hadrons_WeakHamiltonian_hpp_
+#endif // Hadrons_MContraction_WeakHamiltonian_hpp_

extras/Hadrons/Modules/MContraction/WeakHamiltonianEye.hpp

@@ -26,8 +26,8 @@ See the full license in the file "LICENSE" in the top level distribution directo
 *************************************************************************************/
 /*  END LEGAL */
 
-#ifndef Hadrons_WeakHamiltonianEye_hpp_
-#define Hadrons_WeakHamiltonianEye_hpp_
+#ifndef Hadrons_MContraction_WeakHamiltonianEye_hpp_
+#define Hadrons_MContraction_WeakHamiltonianEye_hpp_
 
 #include <Grid/Hadrons/Modules/MContraction/WeakHamiltonian.hpp>
 
@@ -55,4 +55,4 @@ END_MODULE_NAMESPACE
 
 END_HADRONS_NAMESPACE
 
-#endif // Hadrons_WeakHamiltonianEye_hpp_
+#endif // Hadrons_MContraction_WeakHamiltonianEye_hpp_

extras/Hadrons/Modules/MContraction/WeakHamiltonianNonEye.hpp

@@ -26,8 +26,8 @@ See the full license in the file "LICENSE" in the top level distribution directo
 *************************************************************************************/
 /*  END LEGAL */
 
-#ifndef Hadrons_WeakHamiltonianNonEye_hpp_
-#define Hadrons_WeakHamiltonianNonEye_hpp_
+#ifndef Hadrons_MContraction_WeakHamiltonianNonEye_hpp_
+#define Hadrons_MContraction_WeakHamiltonianNonEye_hpp_
 
 #include <Grid/Hadrons/Modules/MContraction/WeakHamiltonian.hpp>
 
@@ -54,4 +54,4 @@ END_MODULE_NAMESPACE
 
 END_HADRONS_NAMESPACE
 
-#endif // Hadrons_WeakHamiltonianNonEye_hpp_
+#endif // Hadrons_MContraction_WeakHamiltonianNonEye_hpp_

extras/Hadrons/Modules/MContraction/WeakNeutral4ptDisc.hpp

@@ -26,8 +26,8 @@ See the full license in the file "LICENSE" in the top level distribution directo
 *************************************************************************************/
 /*  END LEGAL */
 
-#ifndef Hadrons_WeakNeutral4ptDisc_hpp_
-#define Hadrons_WeakNeutral4ptDisc_hpp_
+#ifndef Hadrons_MContraction_WeakNeutral4ptDisc_hpp_
+#define Hadrons_MContraction_WeakNeutral4ptDisc_hpp_
 
 #include <Grid/Hadrons/Modules/MContraction/WeakHamiltonian.hpp>
 
@@ -56,4 +56,4 @@ END_MODULE_NAMESPACE
 
 END_HADRONS_NAMESPACE
 
-#endif // Hadrons_WeakNeutral4ptDisc_hpp_
+#endif // Hadrons_MContraction_WeakNeutral4ptDisc_hpp_

extras/Hadrons/Modules/MFermion/GaugeProp.hpp

@@ -1,34 +1,5 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid 
-
-Source file: extras/Hadrons/Modules/Quark.hpp
-
-Copyright (C) 2015
-Copyright (C) 2016
-
-Author: Antonin Portelli <antonin.portelli@me.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 */
-
-#ifndef Hadrons_Quark_hpp_
-#define Hadrons_Quark_hpp_
+#ifndef Hadrons_MFermion_GaugeProp_hpp_
+#define Hadrons_MFermion_GaugeProp_hpp_
 
 #include <Grid/Hadrons/Global.hpp>
 #include <Grid/Hadrons/Module.hpp>
@@ -37,27 +8,29 @@ See the full license in the file "LICENSE" in the top level distribution directo
 BEGIN_HADRONS_NAMESPACE
 
 /******************************************************************************
- *                               TQuark                                       *
+ *                                GaugeProp                                   *
  ******************************************************************************/
-class QuarkPar: Serializable
+BEGIN_MODULE_NAMESPACE(MFermion)
+
+class GaugePropPar: Serializable
 {
 public:
-    GRID_SERIALIZABLE_CLASS_MEMBERS(QuarkPar,
+    GRID_SERIALIZABLE_CLASS_MEMBERS(GaugePropPar,
                                     std::string, source,
                                     std::string, solver);
 };
 
 template <typename FImpl>
-class TQuark: public Module<QuarkPar>
+class TGaugeProp: public Module<GaugePropPar>
 {
 public:
-    TYPE_ALIASES(FImpl,);
+    FGS_TYPE_ALIASES(FImpl,);
 public:
     // constructor
-    TQuark(const std::string name);
+    TGaugeProp(const std::string name);
     // destructor
-    virtual ~TQuark(void) = default;
-    // dependencies/products
+    virtual ~TGaugeProp(void) = default;
+    // dependency relation
     virtual std::vector<std::string> getInput(void);
     virtual std::vector<std::string> getOutput(void);
     // setup
@@ -69,20 +42,20 @@ private:
     SolverFn     *solver_{nullptr};
 };
 
-MODULE_REGISTER(Quark, TQuark<FIMPL>);
+MODULE_REGISTER_NS(GaugeProp, TGaugeProp<FIMPL>, MFermion);
 
 /******************************************************************************
- *                          TQuark implementation                             *
+ *                      TGaugeProp implementation                             *
  ******************************************************************************/
 // constructor /////////////////////////////////////////////////////////////////
 template <typename FImpl>
-TQuark<FImpl>::TQuark(const std::string name)
-: Module(name)
+TGaugeProp<FImpl>::TGaugeProp(const std::string name)
+: Module<GaugePropPar>(name)
 {}
 
 // dependencies/products ///////////////////////////////////////////////////////
 template <typename FImpl>
-std::vector<std::string> TQuark<FImpl>::getInput(void)
+std::vector<std::string> TGaugeProp<FImpl>::getInput(void)
 {
     std::vector<std::string> in = {par().source, par().solver};
     
@@ -90,7 +63,7 @@ std::vector<std::string> TQuark<FImpl>::getInput(void)
 }
 
 template <typename FImpl>
-std::vector<std::string> TQuark<FImpl>::getOutput(void)
+std::vector<std::string> TGaugeProp<FImpl>::getOutput(void)
 {
     std::vector<std::string> out = {getName(), getName() + "_5d"};
     
@@ -99,7 +72,7 @@ std::vector<std::string> TQuark<FImpl>::getOutput(void)
 
 // setup ///////////////////////////////////////////////////////////////////////
 template <typename FImpl>
-void TQuark<FImpl>::setup(void)
+void TGaugeProp<FImpl>::setup(void)
 {
     Ls_ = env().getObjectLs(par().solver);
     env().template registerLattice<PropagatorField>(getName());
@@ -111,7 +84,7 @@ void TQuark<FImpl>::setup(void)
 
 // execution ///////////////////////////////////////////////////////////////////
 template <typename FImpl>
-void TQuark<FImpl>::execute(void)
+void TGaugeProp<FImpl>::execute(void)
 {
     LOG(Message) << "Computing quark propagator '" << getName() << "'"
     << std::endl;
@@ -180,6 +153,8 @@ void TQuark<FImpl>::execute(void)
     }
 }
 
+END_MODULE_NAMESPACE
+
 END_HADRONS_NAMESPACE
 
-#endif // Hadrons_Quark_hpp_
+#endif // Hadrons_MFermion_GaugeProp_hpp_

extras/Hadrons/Modules/MGauge/Load.hpp

@@ -27,8 +27,8 @@ See the full license in the file "LICENSE" in the top level distribution directo
 *************************************************************************************/
 /*  END LEGAL */
 
-#ifndef Hadrons_Load_hpp_
-#define Hadrons_Load_hpp_
+#ifndef Hadrons_MGauge_Load_hpp_
+#define Hadrons_MGauge_Load_hpp_
 
 #include <Grid/Hadrons/Global.hpp>
 #include <Grid/Hadrons/Module.hpp>
@@ -70,4 +70,4 @@ END_MODULE_NAMESPACE
 
 END_HADRONS_NAMESPACE
 
-#endif // Hadrons_Load_hpp_
+#endif // Hadrons_MGauge_Load_hpp_

extras/Hadrons/Modules/MGauge/Random.hpp

@@ -27,8 +27,8 @@ See the full license in the file "LICENSE" in the top level distribution directo
 *************************************************************************************/
 /*  END LEGAL */
 
-#ifndef Hadrons_Random_hpp_
-#define Hadrons_Random_hpp_
+#ifndef Hadrons_MGauge_Random_hpp_
+#define Hadrons_MGauge_Random_hpp_
 
 #include <Grid/Hadrons/Global.hpp>
 #include <Grid/Hadrons/Module.hpp>
@@ -63,4 +63,4 @@ END_MODULE_NAMESPACE
 
 END_HADRONS_NAMESPACE
 
-#endif // Hadrons_Random_hpp_
+#endif // Hadrons_MGauge_Random_hpp_

extras/Hadrons/Modules/MGauge/StochEm.cc

@@ -0,0 +1,88 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MGauge/StochEm.cc
+
+Copyright (C) 2015
+Copyright (C) 2016
+
+
+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/Hadrons/Modules/MGauge/StochEm.hpp>
+
+using namespace Grid;
+using namespace Hadrons;
+using namespace MGauge;
+
+/******************************************************************************
+*                  TStochEm implementation                             *
+******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+TStochEm::TStochEm(const std::string name)
+: Module<StochEmPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+std::vector<std::string> TStochEm::getInput(void)
+{
+    std::vector<std::string> in;
+    
+    return in;
+}
+
+std::vector<std::string> TStochEm::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+void TStochEm::setup(void)
+{
+    if (!env().hasRegisteredObject("_" + getName() + "_weight"))
+    {
+        env().registerLattice<EmComp>("_" + getName() + "_weight");
+    }
+    env().registerLattice<EmField>(getName());
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+void TStochEm::execute(void)
+{
+    PhotonR photon(par().gauge, par().zmScheme);
+    EmField &a = *env().createLattice<EmField>(getName());
+    EmComp  *w;
+    
+    if (!env().hasCreatedObject("_" + getName() + "_weight"))
+    {
+        LOG(Message) << "Caching stochatic EM potential weight (gauge: "
+                     << par().gauge << ", zero-mode scheme: "
+                     << par().zmScheme << ")..." << std::endl;
+        w = env().createLattice<EmComp>("_" + getName() + "_weight");
+        photon.StochasticWeight(*w);
+    }
+    else
+    {
+        w = env().getObject<EmComp>("_" + getName() + "_weight");
+    }
+    LOG(Message) << "Generating stochatic EM potential..." << std::endl;
+    photon.StochasticField(a, *env().get4dRng(), *w);
+}

extras/Hadrons/Modules/MGauge/StochEm.hpp

@@ -0,0 +1,75 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MGauge/StochEm.hpp
+
+Copyright (C) 2015
+Copyright (C) 2016
+
+
+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 Hadrons_MGauge_StochEm_hpp_
+#define Hadrons_MGauge_StochEm_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                         StochEm                                 *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MGauge)
+
+class StochEmPar: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(StochEmPar,
+                                    PhotonR::Gauge,    gauge,
+                                    PhotonR::ZmScheme, zmScheme);
+};
+
+class TStochEm: public Module<StochEmPar>
+{
+public:
+    typedef PhotonR::GaugeField     EmField;
+    typedef PhotonR::GaugeLinkField EmComp;
+public:
+    // constructor
+    TStochEm(const std::string name);
+    // destructor
+    virtual ~TStochEm(void) = default;
+    // dependency relation
+    virtual std::vector<std::string> getInput(void);
+    virtual std::vector<std::string> getOutput(void);
+    // setup
+    virtual void setup(void);
+    // execution
+    virtual void execute(void);
+};
+
+MODULE_REGISTER_NS(StochEm, TStochEm, MGauge);
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_MGauge_StochEm_hpp_

extras/Hadrons/Modules/MGauge/Unit.hpp

@@ -27,8 +27,8 @@ See the full license in the file "LICENSE" in the top level distribution directo
 *************************************************************************************/
 /*  END LEGAL */
 
-#ifndef Hadrons_Unit_hpp_
-#define Hadrons_Unit_hpp_
+#ifndef Hadrons_MGauge_Unit_hpp_
+#define Hadrons_MGauge_Unit_hpp_
 
 #include <Grid/Hadrons/Global.hpp>
 #include <Grid/Hadrons/Module.hpp>
@@ -63,4 +63,4 @@ END_MODULE_NAMESPACE
 
 END_HADRONS_NAMESPACE
 
-#endif // Hadrons_Unit_hpp_
+#endif // Hadrons_MGauge_Unit_hpp_

extras/Hadrons/Modules/MLoop/NoiseLoop.hpp

@@ -26,8 +26,8 @@ See the full license in the file "LICENSE" in the top level distribution directo
 *************************************************************************************/
 /*  END LEGAL */
 
-#ifndef Hadrons_NoiseLoop_hpp_
-#define Hadrons_NoiseLoop_hpp_
+#ifndef Hadrons_MLoop_NoiseLoop_hpp_
+#define Hadrons_MLoop_NoiseLoop_hpp_
 
 #include <Grid/Hadrons/Global.hpp>
 #include <Grid/Hadrons/Module.hpp>
@@ -65,7 +65,7 @@ template <typename FImpl>
 class TNoiseLoop: public Module<NoiseLoopPar>
 {
 public:
-    TYPE_ALIASES(FImpl,);
+    FERM_TYPE_ALIASES(FImpl,);
 public:
     // constructor
     TNoiseLoop(const std::string name);
@@ -129,4 +129,4 @@ END_MODULE_NAMESPACE
 
 END_HADRONS_NAMESPACE
 
-#endif // Hadrons_NoiseLoop_hpp_
+#endif // Hadrons_MLoop_NoiseLoop_hpp_

extras/Hadrons/Modules/MScalar/ChargedProp.cc

@@ -0,0 +1,226 @@
+#include <Grid/Hadrons/Modules/MScalar/ChargedProp.hpp>
+#include <Grid/Hadrons/Modules/MScalar/Scalar.hpp>
+
+using namespace Grid;
+using namespace Hadrons;
+using namespace MScalar;
+
+/******************************************************************************
+*                     TChargedProp implementation                             *
+******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+TChargedProp::TChargedProp(const std::string name)
+: Module<ChargedPropPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+std::vector<std::string> TChargedProp::getInput(void)
+{
+    std::vector<std::string> in = {par().source, par().emField};
+    
+    return in;
+}
+
+std::vector<std::string> TChargedProp::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+void TChargedProp::setup(void)
+{
+    freeMomPropName_ = FREEMOMPROP(par().mass);
+    phaseName_.clear();
+    for (unsigned int mu = 0; mu < env().getNd(); ++mu)
+    {
+        phaseName_.push_back("_shiftphase_" + std::to_string(mu));
+    }
+    GFSrcName_ = "_" + getName() + "_DinvSrc";
+    if (!env().hasRegisteredObject(freeMomPropName_))
+    {
+        env().registerLattice<ScalarField>(freeMomPropName_);
+    }
+    if (!env().hasRegisteredObject(phaseName_[0]))
+    {
+        for (unsigned int mu = 0; mu < env().getNd(); ++mu)
+        {
+            env().registerLattice<ScalarField>(phaseName_[mu]);
+        }
+    }
+    if (!env().hasRegisteredObject(GFSrcName_))
+    {
+        env().registerLattice<ScalarField>(GFSrcName_);
+    }
+    env().registerLattice<ScalarField>(getName());
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+void TChargedProp::execute(void)
+{
+    // CACHING ANALYTIC EXPRESSIONS
+    ScalarField &source = *env().getObject<ScalarField>(par().source);
+    Complex     ci(0.0,1.0);
+    FFT         fft(env().getGrid());
+    
+    // cache free scalar propagator
+    if (!env().hasCreatedObject(freeMomPropName_))
+    {
+        LOG(Message) << "Caching momentum space free scalar propagator"
+                     << " (mass= " << par().mass << ")..." << std::endl;
+        freeMomProp_ = env().createLattice<ScalarField>(freeMomPropName_);
+        SIMPL::MomentumSpacePropagator(*freeMomProp_, par().mass);
+    }
+    else
+    {
+        freeMomProp_ = env().getObject<ScalarField>(freeMomPropName_);
+    }
+    // cache G*F*src
+    if (!env().hasCreatedObject(GFSrcName_))
+        
+    {
+        GFSrc_ = env().createLattice<ScalarField>(GFSrcName_);
+        fft.FFT_all_dim(*GFSrc_, source, FFT::forward);
+        *GFSrc_ = (*freeMomProp_)*(*GFSrc_);
+    }
+    else
+    {
+        GFSrc_ = env().getObject<ScalarField>(GFSrcName_);
+    }
+    // cache phases
+    if (!env().hasCreatedObject(phaseName_[0]))
+    {
+        std::vector<int> &l = env().getGrid()->_fdimensions;
+        
+        LOG(Message) << "Caching shift phases..." << std::endl;
+        for (unsigned int mu = 0; mu < env().getNd(); ++mu)
+        {
+            Real    twoPiL = M_PI*2./l[mu];
+            
+            phase_.push_back(env().createLattice<ScalarField>(phaseName_[mu]));
+            LatticeCoordinate(*(phase_[mu]), mu);
+            *(phase_[mu]) = exp(ci*twoPiL*(*(phase_[mu])));
+        }
+    }
+    else
+    {
+        for (unsigned int mu = 0; mu < env().getNd(); ++mu)
+        {
+            phase_.push_back(env().getObject<ScalarField>(phaseName_[mu]));
+        }
+    }
+
+    // PROPAGATOR CALCULATION
+    LOG(Message) << "Computing charged scalar propagator"
+                 << " (mass= " << par().mass
+                 << ", charge= " << par().charge << ")..." << std::endl;
+    
+    ScalarField &prop   = *env().createLattice<ScalarField>(getName());
+    ScalarField buf(env().getGrid());
+    ScalarField &GFSrc = *GFSrc_, &G = *freeMomProp_;
+    double      q = par().charge;
+    
+    // G*F*Src
+    prop = GFSrc;
+
+    // - q*G*momD1*G*F*Src (momD1 = F*D1*Finv)
+    buf = GFSrc;
+    momD1(buf, fft);
+    buf = G*buf;
+    prop = prop - q*buf;
+
+    // + q^2*G*momD1*G*momD1*G*F*Src (here buf = G*momD1*G*F*Src)
+    momD1(buf, fft);
+    prop = prop + q*q*G*buf;
+
+    // - q^2*G*momD2*G*F*Src (momD2 = F*D2*Finv)
+    buf = GFSrc;
+    momD2(buf, fft);
+    prop = prop - q*q*G*buf;
+
+    // final FT
+    fft.FFT_all_dim(prop, prop, FFT::backward);
+    
+    // OUTPUT IF NECESSARY
+    if (!par().output.empty())
+    {
+        std::string           filename = par().output + "." +
+                                         std::to_string(env().getTrajectory());
+        
+        LOG(Message) << "Saving zero-momentum projection to '"
+                     << filename << "'..." << std::endl;
+        
+        CorrWriter            writer(filename);
+        std::vector<TComplex> vecBuf;
+        std::vector<Complex>  result;
+        
+        sliceSum(prop, vecBuf, Tp);
+        result.resize(vecBuf.size());
+        for (unsigned int t = 0; t < vecBuf.size(); ++t)
+        {
+            result[t] = TensorRemove(vecBuf[t]);
+        }
+        write(writer, "charge", q);
+        write(writer, "prop", result);
+    }
+}
+
+void TChargedProp::momD1(ScalarField &s, FFT &fft)
+{
+    EmField     &A = *env().getObject<EmField>(par().emField);
+    ScalarField buf(env().getGrid()), result(env().getGrid()),
+                Amu(env().getGrid());
+    Complex     ci(0.0,1.0);
+
+    result = zero;
+
+    for (unsigned int mu = 0; mu < env().getNd(); ++mu)
+    {
+        Amu = peekLorentz(A, mu);
+        buf = (*phase_[mu])*s;
+        fft.FFT_all_dim(buf, buf, FFT::backward);
+        buf = Amu*buf;
+        fft.FFT_all_dim(buf, buf, FFT::forward);
+        result = result - ci*buf;
+    }
+    fft.FFT_all_dim(s, s, FFT::backward);
+    for (unsigned int mu = 0; mu < env().getNd(); ++mu)
+    {
+        Amu = peekLorentz(A, mu);
+        buf = Amu*s;
+        fft.FFT_all_dim(buf, buf, FFT::forward);
+        result = result + ci*adj(*phase_[mu])*buf;
+    }
+
+    s = result;
+}
+
+void TChargedProp::momD2(ScalarField &s, FFT &fft)
+{
+    EmField     &A = *env().getObject<EmField>(par().emField);
+    ScalarField buf(env().getGrid()), result(env().getGrid()),
+                Amu(env().getGrid());
+
+    result = zero;
+    
+    for (unsigned int mu = 0; mu < env().getNd(); ++mu)
+    {
+        Amu = peekLorentz(A, mu);
+        buf = (*phase_[mu])*s;
+        fft.FFT_all_dim(buf, buf, FFT::backward);
+        buf = Amu*Amu*buf;
+        fft.FFT_all_dim(buf, buf, FFT::forward);
+        result = result + .5*buf;
+    }
+    fft.FFT_all_dim(s, s, FFT::backward);
+    for (unsigned int mu = 0; mu < env().getNd(); ++mu)
+    {
+        Amu = peekLorentz(A, mu);        
+        buf = Amu*Amu*s;
+        fft.FFT_all_dim(buf, buf, FFT::forward);
+        result = result + .5*adj(*phase_[mu])*buf;
+    }
+
+    s = result;
+}

extras/Hadrons/Modules/MScalar/ChargedProp.hpp

@@ -0,0 +1,61 @@
+#ifndef Hadrons_MScalar_ChargedProp_hpp_
+#define Hadrons_MScalar_ChargedProp_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                       Charged scalar propagator                            *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MScalar)
+
+class ChargedPropPar: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(ChargedPropPar,
+                                    std::string, emField,
+                                    std::string, source,
+                                    double,      mass,
+                                    double,      charge,
+                                    std::string, output);
+};
+
+class TChargedProp: public Module<ChargedPropPar>
+{
+public:
+    SCALAR_TYPE_ALIASES(SIMPL,);
+    typedef PhotonR::GaugeField     EmField;
+    typedef PhotonR::GaugeLinkField EmComp;
+public:
+    // constructor
+    TChargedProp(const std::string name);
+    // destructor
+    virtual ~TChargedProp(void) = default;
+    // dependency relation
+    virtual std::vector<std::string> getInput(void);
+    virtual std::vector<std::string> getOutput(void);
+    // setup
+    virtual void setup(void);
+    // execution
+    virtual void execute(void);
+private:
+    void momD1(ScalarField &s, FFT &fft);
+    void momD2(ScalarField &s, FFT &fft);
+private:
+    std::string                freeMomPropName_, GFSrcName_;
+    std::vector<std::string>   phaseName_;
+    ScalarField                *freeMomProp_, *GFSrc_;
+    std::vector<ScalarField *> phase_;
+    EmField                    *A;
+};
+
+MODULE_REGISTER_NS(ChargedProp, TChargedProp, MScalar);
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_MScalar_ChargedProp_hpp_

extras/Hadrons/Modules/MScalar/FreeProp.cc

@@ -0,0 +1,79 @@
+#include <Grid/Hadrons/Modules/MScalar/FreeProp.hpp>
+#include <Grid/Hadrons/Modules/MScalar/Scalar.hpp>
+
+using namespace Grid;
+using namespace Hadrons;
+using namespace MScalar;
+
+/******************************************************************************
+*                        TFreeProp implementation                             *
+******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+TFreeProp::TFreeProp(const std::string name)
+: Module<FreePropPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+std::vector<std::string> TFreeProp::getInput(void)
+{
+    std::vector<std::string> in = {par().source};
+    
+    return in;
+}
+
+std::vector<std::string> TFreeProp::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+void TFreeProp::setup(void)
+{
+    freeMomPropName_ = FREEMOMPROP(par().mass);
+    
+    if (!env().hasRegisteredObject(freeMomPropName_))
+    {
+        env().registerLattice<ScalarField>(freeMomPropName_);
+    }
+    env().registerLattice<ScalarField>(getName());
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+void TFreeProp::execute(void)
+{
+    ScalarField &prop   = *env().createLattice<ScalarField>(getName());
+    ScalarField &source = *env().getObject<ScalarField>(par().source);
+    ScalarField *freeMomProp;
+
+    if (!env().hasCreatedObject(freeMomPropName_))
+    {
+        LOG(Message) << "Caching momentum space free scalar propagator"
+                     << " (mass= " << par().mass << ")..." << std::endl;
+        freeMomProp = env().createLattice<ScalarField>(freeMomPropName_);
+        SIMPL::MomentumSpacePropagator(*freeMomProp, par().mass);
+    }
+    else
+    {
+        freeMomProp = env().getObject<ScalarField>(freeMomPropName_);
+    }
+    LOG(Message) << "Computing free scalar propagator..." << std::endl;
+    SIMPL::FreePropagator(source, prop, *freeMomProp);
+    
+    if (!par().output.empty())
+    {
+        TextWriter            writer(par().output + "." +
+                                     std::to_string(env().getTrajectory()));
+        std::vector<TComplex> buf;
+        std::vector<Complex>  result;
+        
+        sliceSum(prop, buf, Tp);
+        result.resize(buf.size());
+        for (unsigned int t = 0; t < buf.size(); ++t)
+        {
+            result[t] = TensorRemove(buf[t]);
+        }
+        write(writer, "prop", result);
+    }
+}

extras/Hadrons/Modules/MScalar/FreeProp.hpp

@@ -0,0 +1,50 @@
+#ifndef Hadrons_MScalar_FreeProp_hpp_
+#define Hadrons_MScalar_FreeProp_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                               FreeProp                                     *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MScalar)
+
+class FreePropPar: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(FreePropPar,
+                                    std::string, source,
+                                    double,      mass,
+                                    std::string, output);
+};
+
+class TFreeProp: public Module<FreePropPar>
+{
+public:
+    SCALAR_TYPE_ALIASES(SIMPL,);
+public:
+    // constructor
+    TFreeProp(const std::string name);
+    // destructor
+    virtual ~TFreeProp(void) = default;
+    // dependency relation
+    virtual std::vector<std::string> getInput(void);
+    virtual std::vector<std::string> getOutput(void);
+    // setup
+    virtual void setup(void);
+    // execution
+    virtual void execute(void);
+private:
+    std::string freeMomPropName_;
+};
+
+MODULE_REGISTER_NS(FreeProp, TFreeProp, MScalar);
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_MScalar_FreeProp_hpp_

extras/Hadrons/Modules/MScalar/Scalar.hpp

@@ -0,0 +1,6 @@
+#ifndef Hadrons_Scalar_hpp_
+#define Hadrons_Scalar_hpp_
+
+#define FREEMOMPROP(m) "_scalar_mom_prop_" + std::to_string(m)
+
+#endif // Hadrons_Scalar_hpp_

extras/Hadrons/Modules/MSink/Point.hpp

@@ -0,0 +1,114 @@
+#ifndef Hadrons_MSink_Point_hpp_
+#define Hadrons_MSink_Point_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                                   Point                                    *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MSink)
+
+class PointPar: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(PointPar,
+                                    std::string, mom);
+};
+
+template <typename FImpl>
+class TPoint: public Module<PointPar>
+{
+public:
+    FERM_TYPE_ALIASES(FImpl,);
+    SINK_TYPE_ALIASES();
+public:
+    // constructor
+    TPoint(const std::string name);
+    // destructor
+    virtual ~TPoint(void) = default;
+    // dependency relation
+    virtual std::vector<std::string> getInput(void);
+    virtual std::vector<std::string> getOutput(void);
+    // setup
+    virtual void setup(void);
+    // execution
+    virtual void execute(void);
+};
+
+MODULE_REGISTER_NS(Point,       TPoint<FIMPL>,        MSink);
+MODULE_REGISTER_NS(ScalarPoint, TPoint<ScalarImplCR>, MSink);
+
+/******************************************************************************
+ *                          TPoint implementation                             *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+template <typename FImpl>
+TPoint<FImpl>::TPoint(const std::string name)
+: Module<PointPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+template <typename FImpl>
+std::vector<std::string> TPoint<FImpl>::getInput(void)
+{
+    std::vector<std::string> in;
+    
+    return in;
+}
+
+template <typename FImpl>
+std::vector<std::string> TPoint<FImpl>::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TPoint<FImpl>::setup(void)
+{
+    unsigned int size;
+    
+    size = env().template lattice4dSize<LatticeComplex>();
+    env().registerObject(getName(), size);
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TPoint<FImpl>::execute(void)
+{
+    std::vector<Real> p = strToVec<Real>(par().mom);
+    LatticeComplex    ph(env().getGrid()), coor(env().getGrid());
+    Complex           i(0.0,1.0);
+    
+    LOG(Message) << "Setting up point sink function for momentum ["
+                 << par().mom << "]" << std::endl;
+    ph = zero;
+    for(unsigned int mu = 0; mu < env().getNd(); mu++)
+    {
+        LatticeCoordinate(coor, mu);
+        ph = ph + (p[mu]/env().getGrid()->_fdimensions[mu])*coor;
+    }
+    ph = exp((Real)(2*M_PI)*i*ph);
+    auto sink = [ph](const PropagatorField &field)
+    {
+        SlicedPropagator res;
+        PropagatorField  tmp = ph*field;
+        
+        sliceSum(tmp, res, Tp);
+        
+        return res;
+    };
+    env().setObject(getName(), new SinkFn(sink));
+}
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_MSink_Point_hpp_

extras/Hadrons/Modules/MSolver/RBPrecCG.hpp

@@ -27,8 +27,8 @@ See the full license in the file "LICENSE" in the top level distribution directo
 *************************************************************************************/
 /*  END LEGAL */
 
-#ifndef Hadrons_RBPrecCG_hpp_
-#define Hadrons_RBPrecCG_hpp_
+#ifndef Hadrons_MSolver_RBPrecCG_hpp_
+#define Hadrons_MSolver_RBPrecCG_hpp_
 
 #include <Grid/Hadrons/Global.hpp>
 #include <Grid/Hadrons/Module.hpp>
@@ -53,7 +53,7 @@ template <typename FImpl>
 class TRBPrecCG: public Module<RBPrecCGPar>
 {
 public:
-    TYPE_ALIASES(FImpl,);
+    FGS_TYPE_ALIASES(FImpl,);
 public:
     // constructor
     TRBPrecCG(const std::string name);
@@ -129,4 +129,4 @@ END_MODULE_NAMESPACE
 
 END_HADRONS_NAMESPACE
 
-#endif // Hadrons_RBPrecCG_hpp_
+#endif // Hadrons_MSolver_RBPrecCG_hpp_

extras/Hadrons/Modules/MSource/Point.hpp

@@ -27,8 +27,8 @@ See the full license in the file "LICENSE" in the top level distribution directo
 *************************************************************************************/
 /*  END LEGAL */
 
-#ifndef Hadrons_Point_hpp_
-#define Hadrons_Point_hpp_
+#ifndef Hadrons_MSource_Point_hpp_
+#define Hadrons_MSource_Point_hpp_
 
 #include <Grid/Hadrons/Global.hpp>
 #include <Grid/Hadrons/Module.hpp>
@@ -63,7 +63,7 @@ template <typename FImpl>
 class TPoint: public Module<PointPar>
 {
 public:
-    TYPE_ALIASES(FImpl,);
+    FERM_TYPE_ALIASES(FImpl,);
 public:
     // constructor
     TPoint(const std::string name);
@@ -79,6 +79,7 @@ public:
 };
 
 MODULE_REGISTER_NS(Point,       TPoint<FIMPL>,        MSource);
+MODULE_REGISTER_NS(ScalarPoint, TPoint<ScalarImplCR>, MSource);
 
 /******************************************************************************
  *                       TPoint template implementation                       *
@@ -132,4 +133,4 @@ END_MODULE_NAMESPACE
 
 END_HADRONS_NAMESPACE
 
-#endif // Hadrons_Point_hpp_
+#endif // Hadrons_MSource_Point_hpp_

extras/Hadrons/Modules/MSource/SeqGamma.hpp

@@ -28,8 +28,8 @@ See the full license in the file "LICENSE" in the top level distribution directo
 *************************************************************************************/
 /*  END LEGAL */
 
-#ifndef Hadrons_SeqGamma_hpp_
-#define Hadrons_SeqGamma_hpp_
+#ifndef Hadrons_MSource_SeqGamma_hpp_
+#define Hadrons_MSource_SeqGamma_hpp_
 
 #include <Grid/Hadrons/Global.hpp>
 #include <Grid/Hadrons/Module.hpp>
@@ -72,7 +72,7 @@ template <typename FImpl>
 class TSeqGamma: public Module<SeqGammaPar>
 {
 public:
-    TYPE_ALIASES(FImpl,);
+    FGS_TYPE_ALIASES(FImpl,);
 public:
     // constructor
     TSeqGamma(const std::string name);
@@ -161,4 +161,4 @@ END_MODULE_NAMESPACE
 
 END_HADRONS_NAMESPACE
 
-#endif // Hadrons_SeqGamma_hpp_
+#endif // Hadrons_MSource_SeqGamma_hpp_

extras/Hadrons/Modules/MSource/Wall.hpp

@@ -26,8 +26,8 @@ See the full license in the file "LICENSE" in the top level distribution directo
 *************************************************************************************/
 /*  END LEGAL */
 
-#ifndef Hadrons_WallSource_hpp_
-#define Hadrons_WallSource_hpp_
+#ifndef Hadrons_MSource_WallSource_hpp_
+#define Hadrons_MSource_WallSource_hpp_
 
 #include <Grid/Hadrons/Global.hpp>
 #include <Grid/Hadrons/Module.hpp>
@@ -64,7 +64,7 @@ template <typename FImpl>
 class TWall: public Module<WallPar>
 {
 public:
-    TYPE_ALIASES(FImpl,);
+    FERM_TYPE_ALIASES(FImpl,);
 public:
     // constructor
     TWall(const std::string name);
@@ -144,4 +144,4 @@ END_MODULE_NAMESPACE
 
 END_HADRONS_NAMESPACE
 
-#endif // Hadrons_WallSource_hpp_
+#endif // Hadrons_MSource_WallSource_hpp_

extras/Hadrons/Modules/MSource/Z2.hpp

@@ -27,8 +27,8 @@ See the full license in the file "LICENSE" in the top level distribution directo
 *************************************************************************************/
 /*  END LEGAL */
 
-#ifndef Hadrons_Z2_hpp_
-#define Hadrons_Z2_hpp_
+#ifndef Hadrons_MSource_Z2_hpp_
+#define Hadrons_MSource_Z2_hpp_
 
 #include <Grid/Hadrons/Global.hpp>
 #include <Grid/Hadrons/Module.hpp>
@@ -67,7 +67,7 @@ template <typename FImpl>
 class TZ2: public Module<Z2Par>
 {
 public:
-    TYPE_ALIASES(FImpl,);
+    FERM_TYPE_ALIASES(FImpl,);
 public:
     // constructor
     TZ2(const std::string name);
@@ -83,6 +83,7 @@ public:
 };
 
 MODULE_REGISTER_NS(Z2,       TZ2<FIMPL>,        MSource);
+MODULE_REGISTER_NS(ScalarZ2, TZ2<ScalarImplCR>, MSource);
 
 /******************************************************************************
  *                       TZ2 template implementation                          *
@@ -148,4 +149,4 @@ END_MODULE_NAMESPACE
 
 END_HADRONS_NAMESPACE
 
-#endif // Hadrons_Z2_hpp_
+#endif // Hadrons_MSource_Z2_hpp_

extras/Hadrons/Modules/templates/Module_in_NS.hpp.template

@@ -1,5 +1,5 @@
-#ifndef Hadrons____FILEBASENAME____hpp_
-#define Hadrons____FILEBASENAME____hpp_
+#ifndef Hadrons____NAMESPACE_______FILEBASENAME____hpp_
+#define Hadrons____NAMESPACE_______FILEBASENAME____hpp_
 
 #include <Grid/Hadrons/Global.hpp>
 #include <Grid/Hadrons/Module.hpp>
@@ -41,4 +41,4 @@ END_MODULE_NAMESPACE
 
 END_HADRONS_NAMESPACE
 
-#endif // Hadrons____FILEBASENAME____hpp_
+#endif // Hadrons____NAMESPACE_______FILEBASENAME____hpp_

extras/Hadrons/Modules/templates/Module_tmp_in_NS.hpp.template

@@ -1,5 +1,5 @@
-#ifndef Hadrons____FILEBASENAME____hpp_
-#define Hadrons____FILEBASENAME____hpp_
+#ifndef Hadrons____NAMESPACE_______FILEBASENAME____hpp_
+#define Hadrons____NAMESPACE_______FILEBASENAME____hpp_
 
 #include <Grid/Hadrons/Global.hpp>
 #include <Grid/Hadrons/Module.hpp>
@@ -82,4 +82,4 @@ END_MODULE_NAMESPACE
 
 END_HADRONS_NAMESPACE
 
-#endif // Hadrons____FILEBASENAME____hpp_
+#endif // Hadrons____NAMESPACE_______FILEBASENAME____hpp_

extras/Hadrons/modules.inc

@@ -4,7 +4,10 @@ modules_cc =\
   Modules/MContraction/WeakNeutral4ptDisc.cc \
   Modules/MGauge/Load.cc \
   Modules/MGauge/Random.cc \
-  Modules/MGauge/Unit.cc
+  Modules/MGauge/StochEm.cc \
+  Modules/MGauge/Unit.cc \
+  Modules/MScalar/ChargedProp.cc \
+  Modules/MScalar/FreeProp.cc
 
 modules_hpp =\
   Modules/MAction/DWF.hpp \
@@ -17,14 +20,19 @@ modules_hpp =\
   Modules/MContraction/WeakHamiltonianEye.hpp \
   Modules/MContraction/WeakHamiltonianNonEye.hpp \
   Modules/MContraction/WeakNeutral4ptDisc.hpp \
+  Modules/MFermion/GaugeProp.hpp \
   Modules/MGauge/Load.hpp \
   Modules/MGauge/Random.hpp \
+  Modules/MGauge/StochEm.hpp \
   Modules/MGauge/Unit.hpp \
   Modules/MLoop/NoiseLoop.hpp \
+  Modules/MScalar/ChargedProp.hpp \
+  Modules/MScalar/FreeProp.hpp \
+  Modules/MScalar/Scalar.hpp \
+  Modules/MSink/Point.hpp \
   Modules/MSolver/RBPrecCG.hpp \
   Modules/MSource/Point.hpp \
   Modules/MSource/SeqGamma.hpp \
   Modules/MSource/Wall.hpp \
-  Modules/MSource/Z2.hpp \
-  Modules/Quark.hpp
+  Modules/MSource/Z2.hpp
 

extras/qed-fvol/Global.cc

@@ -0,0 +1,11 @@
+#include <qed-fvol/Global.hpp>
+
+using namespace Grid;
+using namespace QCD;
+using namespace QedFVol;
+
+QedFVolLogger QedFVol::QedFVolLogError(1,"Error");
+QedFVolLogger QedFVol::QedFVolLogWarning(1,"Warning");
+QedFVolLogger QedFVol::QedFVolLogMessage(1,"Message");
+QedFVolLogger QedFVol::QedFVolLogIterative(1,"Iterative");
+QedFVolLogger QedFVol::QedFVolLogDebug(1,"Debug");

extras/qed-fvol/Global.hpp

@@ -0,0 +1,42 @@
+#ifndef QedFVol_Global_hpp_
+#define QedFVol_Global_hpp_
+
+#include <Grid/Grid.h>
+
+#define BEGIN_QEDFVOL_NAMESPACE \
+namespace Grid {\
+using namespace QCD;\
+namespace QedFVol {\
+using Grid::operator<<;
+#define END_QEDFVOL_NAMESPACE }}
+
+/* the 'using Grid::operator<<;' statement prevents a very nasty compilation
+ * error with GCC (clang compiles fine without it).
+ */
+
+BEGIN_QEDFVOL_NAMESPACE
+
+class QedFVolLogger: public Logger
+{
+public:
+    QedFVolLogger(int on, std::string nm): Logger("QedFVol", on, nm,
+                                                  GridLogColours, "BLACK"){};
+};
+
+#define LOG(channel) std::cout << QedFVolLog##channel
+#define QEDFVOL_ERROR(msg)\
+LOG(Error) << msg << " (" << __FUNCTION__ << " at " << __FILE__ << ":"\
+           << __LINE__ << ")" << std::endl;\
+abort();
+
+#define DEBUG_VAR(var) LOG(Debug) << #var << "= " << (var) << std::endl;
+
+extern QedFVolLogger QedFVolLogError;
+extern QedFVolLogger QedFVolLogWarning;
+extern QedFVolLogger QedFVolLogMessage;
+extern QedFVolLogger QedFVolLogIterative;
+extern QedFVolLogger QedFVolLogDebug;
+
+END_QEDFVOL_NAMESPACE
+
+#endif // QedFVol_Global_hpp_

extras/qed-fvol/Makefile.am

@@ -0,0 +1,9 @@
+AM_CXXFLAGS += -I$(top_srcdir)/extras
+
+bin_PROGRAMS = qed-fvol
+
+qed_fvol_SOURCES =   \
+    qed-fvol.cc      \
+    Global.cc
+
+qed_fvol_LDADD   = -lGrid

extras/qed-fvol/WilsonLoops.h

@@ -0,0 +1,265 @@
+#ifndef QEDFVOL_WILSONLOOPS_H
+#define QEDFVOL_WILSONLOOPS_H
+
+#include <Global.hpp>
+
+BEGIN_QEDFVOL_NAMESPACE
+
+template <class Gimpl> class NewWilsonLoops : public Gimpl {
+public:
+  INHERIT_GIMPL_TYPES(Gimpl);
+
+  typedef typename Gimpl::GaugeLinkField GaugeMat;
+  typedef typename Gimpl::GaugeField GaugeLorentz;
+
+  //////////////////////////////////////////////////
+  // directed plaquette oriented in mu,nu plane
+  //////////////////////////////////////////////////
+  static void dirPlaquette(GaugeMat &plaq, const std::vector<GaugeMat> &U,
+                           const int mu, const int nu) {
+    // Annoyingly, must use either scope resolution to find dependent base
+    // class,
+    // or this-> ; there is no "this" in a static method. This forces explicit
+    // Gimpl scope
+    // resolution throughout the usage in this file, and rather defeats the
+    // purpose of deriving
+    // from Gimpl.
+    plaq = Gimpl::CovShiftBackward(
+        U[mu], mu, Gimpl::CovShiftBackward(
+                       U[nu], nu, Gimpl::CovShiftForward(U[mu], mu, U[nu])));
+  }
+  //////////////////////////////////////////////////
+  // trace of directed plaquette oriented in mu,nu plane
+  //////////////////////////////////////////////////
+  static void traceDirPlaquette(LatticeComplex &plaq,
+                                const std::vector<GaugeMat> &U, const int mu,
+                                const int nu) {
+    GaugeMat sp(U[0]._grid);
+    dirPlaquette(sp, U, mu, nu);
+    plaq = trace(sp);
+  }
+  //////////////////////////////////////////////////
+  // sum over all planes of plaquette
+  //////////////////////////////////////////////////
+  static void sitePlaquette(LatticeComplex &Plaq,
+                            const std::vector<GaugeMat> &U) {
+    LatticeComplex sitePlaq(U[0]._grid);
+    Plaq = zero;
+    for (int mu = 1; mu < U[0]._grid->_ndimension; mu++) {
+      for (int nu = 0; nu < mu; nu++) {
+        traceDirPlaquette(sitePlaq, U, mu, nu);
+        Plaq = Plaq + sitePlaq;
+      }
+    }
+  }
+  //////////////////////////////////////////////////
+  // sum over all x,y,z,t and over all planes of plaquette
+  //////////////////////////////////////////////////
+  static Real sumPlaquette(const GaugeLorentz &Umu) {
+    std::vector<GaugeMat> U(4, Umu._grid);
+
+    for (int mu = 0; mu < Umu._grid->_ndimension; mu++) {
+      U[mu] = PeekIndex<LorentzIndex>(Umu, mu);
+    }
+
+    LatticeComplex Plaq(Umu._grid);
+
+    sitePlaquette(Plaq, U);
+
+    TComplex Tp = sum(Plaq);
+    Complex p = TensorRemove(Tp);
+    return p.real();
+  }
+  //////////////////////////////////////////////////
+  // average over all x,y,z,t and over all planes of plaquette
+  //////////////////////////////////////////////////
+  static Real avgPlaquette(const GaugeLorentz &Umu) {
+    int ndim = Umu._grid->_ndimension;
+    Real sumplaq = sumPlaquette(Umu);
+    Real vol = Umu._grid->gSites();
+    Real faces = (1.0 * ndim * (ndim - 1)) / 2.0;
+    return sumplaq / vol / faces / Nc; // Nc dependent... FIXME
+  }
+
+  //////////////////////////////////////////////////
+  // Wilson loop of size (R1, R2), oriented in mu,nu plane
+  //////////////////////////////////////////////////
+  static void wilsonLoop(GaugeMat &wl, const std::vector<GaugeMat> &U,
+                           const int Rmu, const int Rnu,
+                           const int mu, const int nu) {
+    wl = U[nu];
+
+    for(int i = 0; i < Rnu-1; i++){
+      wl = Gimpl::CovShiftForward(U[nu], nu, wl);
+    }
+
+    for(int i = 0; i < Rmu; i++){
+      wl = Gimpl::CovShiftForward(U[mu], mu, wl);
+    }
+
+    for(int i = 0; i < Rnu; i++){
+      wl = Gimpl::CovShiftBackward(U[nu], nu, wl);
+    }
+
+    for(int i = 0; i < Rmu; i++){
+      wl = Gimpl::CovShiftBackward(U[mu], mu, wl);
+    }
+  }
+  //////////////////////////////////////////////////
+  // trace of Wilson Loop oriented in mu,nu plane
+  //////////////////////////////////////////////////
+  static void traceWilsonLoop(LatticeComplex &wl,
+                                const std::vector<GaugeMat> &U,
+                                const int Rmu, const int Rnu,
+                                const int mu, const int nu) {
+    GaugeMat sp(U[0]._grid);
+    wilsonLoop(sp, U, Rmu, Rnu, mu, nu);
+    wl = trace(sp);
+  }
+  //////////////////////////////////////////////////
+  // sum over all planes of Wilson loop
+  //////////////////////////////////////////////////
+  static void siteWilsonLoop(LatticeComplex &Wl,
+                            const std::vector<GaugeMat> &U,
+                            const int R1, const int R2) {
+    LatticeComplex siteWl(U[0]._grid);
+    Wl = zero;
+    for (int mu = 1; mu < U[0]._grid->_ndimension; mu++) {
+      for (int nu = 0; nu < mu; nu++) {
+        traceWilsonLoop(siteWl, U, R1, R2, mu, nu);
+        Wl = Wl + siteWl;
+        traceWilsonLoop(siteWl, U, R2, R1, mu, nu);
+        Wl = Wl + siteWl;
+      }
+    }
+  }
+  //////////////////////////////////////////////////
+  // sum over planes of Wilson loop with length R1
+  // in the time direction
+  //////////////////////////////////////////////////
+  static void siteTimelikeWilsonLoop(LatticeComplex &Wl,
+                            const std::vector<GaugeMat> &U,
+                            const int R1, const int R2) {
+    LatticeComplex siteWl(U[0]._grid);
+
+    int ndim = U[0]._grid->_ndimension;
+
+    Wl = zero;
+    for (int nu = 0; nu < ndim - 1; nu++) {
+      traceWilsonLoop(siteWl, U, R1, R2, ndim-1, nu);
+      Wl = Wl + siteWl;
+    }
+  }
+  //////////////////////////////////////////////////
+  // sum Wilson loop over all planes orthogonal to the time direction
+  //////////////////////////////////////////////////
+  static void siteSpatialWilsonLoop(LatticeComplex &Wl,
+                            const std::vector<GaugeMat> &U,
+                            const int R1, const int R2) {
+    LatticeComplex siteWl(U[0]._grid);
+
+    Wl = zero;
+    for (int mu = 1; mu < U[0]._grid->_ndimension - 1; mu++) {
+      for (int nu = 0; nu < mu; nu++) {
+        traceWilsonLoop(siteWl, U, R1, R2, mu, nu);
+        Wl = Wl + siteWl;
+        traceWilsonLoop(siteWl, U, R2, R1, mu, nu);
+        Wl = Wl + siteWl;
+      }
+    }
+  }
+  //////////////////////////////////////////////////
+  // sum over all x,y,z,t and over all planes of Wilson loop
+  //////////////////////////////////////////////////
+  static Real sumWilsonLoop(const GaugeLorentz &Umu,
+                            const int R1, const int R2) {
+    std::vector<GaugeMat> U(4, Umu._grid);
+
+    for (int mu = 0; mu < Umu._grid->_ndimension; mu++) {
+      U[mu] = PeekIndex<LorentzIndex>(Umu, mu);
+    }
+
+    LatticeComplex Wl(Umu._grid);
+
+    siteWilsonLoop(Wl, U, R1, R2);
+
+    TComplex Tp = sum(Wl);
+    Complex p = TensorRemove(Tp);
+    return p.real();
+  }
+  //////////////////////////////////////////////////
+  // sum over all x,y,z,t and over all planes of timelike Wilson loop
+  //////////////////////////////////////////////////
+  static Real sumTimelikeWilsonLoop(const GaugeLorentz &Umu,
+                            const int R1, const int R2) {
+    std::vector<GaugeMat> U(4, Umu._grid);
+
+    for (int mu = 0; mu < Umu._grid->_ndimension; mu++) {
+      U[mu] = PeekIndex<LorentzIndex>(Umu, mu);
+    }
+
+    LatticeComplex Wl(Umu._grid);
+
+    siteTimelikeWilsonLoop(Wl, U, R1, R2);
+
+    TComplex Tp = sum(Wl);
+    Complex p = TensorRemove(Tp);
+    return p.real();
+  }
+  //////////////////////////////////////////////////
+  // sum over all x,y,z,t and over all planes of spatial Wilson loop
+  //////////////////////////////////////////////////
+  static Real sumSpatialWilsonLoop(const GaugeLorentz &Umu,
+                            const int R1, const int R2) {
+    std::vector<GaugeMat> U(4, Umu._grid);
+
+    for (int mu = 0; mu < Umu._grid->_ndimension; mu++) {
+      U[mu] = PeekIndex<LorentzIndex>(Umu, mu);
+    }
+
+    LatticeComplex Wl(Umu._grid);
+
+    siteSpatialWilsonLoop(Wl, U, R1, R2);
+
+    TComplex Tp = sum(Wl);
+    Complex p = TensorRemove(Tp);
+    return p.real();
+  }
+  //////////////////////////////////////////////////
+  // average over all x,y,z,t and over all planes of Wilson loop
+  //////////////////////////////////////////////////
+  static Real avgWilsonLoop(const GaugeLorentz &Umu,
+                            const int R1, const int R2) {
+    int ndim = Umu._grid->_ndimension;
+    Real sumWl = sumWilsonLoop(Umu, R1, R2);
+    Real vol = Umu._grid->gSites();
+    Real faces = 1.0 * ndim * (ndim - 1);
+    return sumWl / vol / faces / Nc; // Nc dependent... FIXME
+  }
+  //////////////////////////////////////////////////
+  // average over all x,y,z,t and over all planes of timelike Wilson loop
+  //////////////////////////////////////////////////
+  static Real avgTimelikeWilsonLoop(const GaugeLorentz &Umu,
+                            const int R1, const int R2) {
+    int ndim = Umu._grid->_ndimension;
+    Real sumWl = sumTimelikeWilsonLoop(Umu, R1, R2);
+    Real vol = Umu._grid->gSites();
+    Real faces = 1.0 * (ndim - 1);
+    return sumWl / vol / faces / Nc; // Nc dependent... FIXME
+  }
+  //////////////////////////////////////////////////
+  // average over all x,y,z,t and over all planes of spatial Wilson loop
+  //////////////////////////////////////////////////
+  static Real avgSpatialWilsonLoop(const GaugeLorentz &Umu,
+                            const int R1, const int R2) {
+    int ndim = Umu._grid->_ndimension;
+    Real sumWl = sumSpatialWilsonLoop(Umu, R1, R2);
+    Real vol = Umu._grid->gSites();
+    Real faces = 1.0 * (ndim - 1) * (ndim - 2);
+    return sumWl / vol / faces / Nc; // Nc dependent... FIXME
+  }
+};
+
+END_QEDFVOL_NAMESPACE
+
+#endif // QEDFVOL_WILSONLOOPS_H

extras/qed-fvol/qed-fvol.cc

@@ -0,0 +1,88 @@
+#include <Global.hpp>
+#include <WilsonLoops.h>
+
+using namespace Grid;
+using namespace QCD;
+using namespace QedFVol;
+
+typedef PeriodicGaugeImpl<QedGimplR>    QedPeriodicGimplR;
+typedef PhotonR::GaugeField             EmField;
+typedef PhotonR::GaugeLinkField         EmComp;
+
+const int NCONFIGS = 10;
+const int NWILSON = 10;
+
+int main(int argc, char *argv[])
+{
+    // parse command line
+    std::string parameterFileName;
+    
+    if (argc < 2)
+    {
+        std::cerr << "usage: " << argv[0] << " <parameter file> [Grid options]";
+        std::cerr << std::endl;
+        std::exit(EXIT_FAILURE);
+    }
+    parameterFileName = argv[1];
+    
+    // initialization
+    Grid_init(&argc, &argv);
+    QedFVolLogError.Active(GridLogError.isActive());
+    QedFVolLogWarning.Active(GridLogWarning.isActive());
+    QedFVolLogMessage.Active(GridLogMessage.isActive());
+    QedFVolLogIterative.Active(GridLogIterative.isActive());
+    QedFVolLogDebug.Active(GridLogDebug.isActive());
+    LOG(Message) << "Grid initialized" << std::endl;
+    
+    // QED stuff
+    std::vector<int> latt_size   = GridDefaultLatt();
+    std::vector<int> simd_layout = GridDefaultSimd(4, vComplex::Nsimd());
+    std::vector<int> mpi_layout  = GridDefaultMpi();
+    GridCartesian    grid(latt_size,simd_layout,mpi_layout);
+    GridParallelRNG  pRNG(&grid);
+    PhotonR          photon(PhotonR::Gauge::feynman,
+                            PhotonR::ZmScheme::qedL);
+    EmField          a(&grid);
+    EmField          expA(&grid);
+
+    Complex imag_unit(0, 1);
+
+    Real wlA;
+    std::vector<Real> logWlAvg(NWILSON, 0.0), logWlTime(NWILSON, 0.0), logWlSpace(NWILSON, 0.0);
+
+    pRNG.SeedRandomDevice();
+
+    LOG(Message) << "Wilson loop calculation beginning" << std::endl;
+    for(int ic = 0; ic < NCONFIGS; ic++){
+        LOG(Message) << "Configuration " << ic <<std::endl;
+        photon.StochasticField(a, pRNG);
+
+        // Exponentiate photon field
+        expA = exp(imag_unit*a);
+
+        // Calculate Wilson loops
+        for(int iw=1; iw<=NWILSON; iw++){
+            wlA = NewWilsonLoops<QedPeriodicGimplR>::avgWilsonLoop(expA, iw, iw) * 3;
+            logWlAvg[iw-1] -= 2*log(wlA);
+            wlA = NewWilsonLoops<QedPeriodicGimplR>::avgTimelikeWilsonLoop(expA, iw, iw) * 3;
+            logWlTime[iw-1] -= 2*log(wlA);
+            wlA = NewWilsonLoops<QedPeriodicGimplR>::avgSpatialWilsonLoop(expA, iw, iw) * 3;
+            logWlSpace[iw-1] -= 2*log(wlA);
+        }
+    }
+    LOG(Message) << "Wilson loop calculation completed" << std::endl;
+    
+    // Calculate Wilson loops
+    for(int iw=1; iw<=10; iw++){
+        LOG(Message) << iw << 'x' << iw << " Wilson loop" << std::endl;
+        LOG(Message) << "-2log(W) average: " << logWlAvg[iw-1]/NCONFIGS << std::endl;
+        LOG(Message) << "-2log(W) timelike: " << logWlTime[iw-1]/NCONFIGS << std::endl;
+        LOG(Message) << "-2log(W) spatial: " << logWlSpace[iw-1]/NCONFIGS << std::endl;
+    }
+
+    // epilogue
+    LOG(Message) << "Grid is finalizing now" << std::endl;
+    Grid_finalize();
+    
+    return EXIT_SUCCESS;
+}

lib/Makefile.am

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

lib/algorithms/Algorithms.h

@@ -37,6 +37,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/algorithms/approx/Chebyshev.h>
 #include <Grid/algorithms/approx/Remez.h>
 #include <Grid/algorithms/approx/MultiShiftFunction.h>
+#include <Grid/algorithms/approx/Forecast.h>
 
 #include <Grid/algorithms/iterative/ConjugateGradient.h>
 #include <Grid/algorithms/iterative/ConjugateResidual.h>
@@ -44,30 +45,16 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/algorithms/iterative/SchurRedBlack.h>
 #include <Grid/algorithms/iterative/ConjugateGradientMultiShift.h>
 #include <Grid/algorithms/iterative/ConjugateGradientMixedPrec.h>
-
-// Lanczos support
-//#include <Grid/algorithms/iterative/MatrixUtils.h>
+#include <Grid/algorithms/iterative/BlockConjugateGradient.h>
+#include <Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h>
 #include <Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h>
 #include <Grid/algorithms/CoarsenedMatrix.h>
 #include <Grid/algorithms/FFT.h>
 
-// Eigen/lanczos
 // EigCg
-// MCR
 // Pcg
-// Multishift CG
 // Hdcg
 // GCR
 // etc..
 
-// integrator/Leapfrog
-// integrator/Omelyan
-// integrator/ForceGradient
-
-// montecarlo/hmc
-// montecarlo/rhmc
-// montecarlo/metropolis
-// etc...
-
-
 #endif

lib/algorithms/CoarsenedMatrix.h

@@ -103,29 +103,32 @@ namespace Grid {
     GridBase *CoarseGrid;
     GridBase *FineGrid;
     std::vector<Lattice<Fobj> > subspace;
+    int checkerboard;
 
-    Aggregation(GridBase *_CoarseGrid,GridBase *_FineGrid) : 
+  Aggregation(GridBase *_CoarseGrid,GridBase *_FineGrid,int _checkerboard) : 
     CoarseGrid(_CoarseGrid),
       FineGrid(_FineGrid),
-      subspace(nbasis,_FineGrid)
+      subspace(nbasis,_FineGrid),
+      checkerboard(_checkerboard)
 	{
 	};
   
     void Orthogonalise(void){
       CoarseScalar InnerProd(CoarseGrid); 
+      std::cout << GridLogMessage <<" Gramm-Schmidt pass 1"<<std::endl;
       blockOrthogonalise(InnerProd,subspace);
+      std::cout << GridLogMessage <<" Gramm-Schmidt pass 2"<<std::endl;
+      blockOrthogonalise(InnerProd,subspace);
+      //      std::cout << GridLogMessage <<" Gramm-Schmidt checking orthogonality"<<std::endl;
+      //      CheckOrthogonal();
     } 
     void CheckOrthogonal(void){
       CoarseVector iProj(CoarseGrid); 
       CoarseVector eProj(CoarseGrid); 
-      Lattice<CComplex> pokey(CoarseGrid);
-
-      
       for(int i=0;i<nbasis;i++){
 	blockProject(iProj,subspace[i],subspace);
-
 	eProj=zero; 
-	for(int ss=0;ss<CoarseGrid->oSites();ss++){
+	parallel_for(int ss=0;ss<CoarseGrid->oSites();ss++){
 	  eProj._odata[ss](i)=CComplex(1.0);
 	}
 	eProj=eProj - iProj;
@@ -137,6 +140,7 @@ namespace Grid {
       blockProject(CoarseVec,FineVec,subspace);
     }
     void PromoteFromSubspace(const CoarseVector &CoarseVec,FineField &FineVec){
+      FineVec.checkerboard = subspace[0].checkerboard;
       blockPromote(CoarseVec,FineVec,subspace);
     }
     void CreateSubspaceRandom(GridParallelRNG &RNG){
@@ -147,6 +151,7 @@ namespace Grid {
       Orthogonalise();
     }
 
+    /*
     virtual void CreateSubspaceLanczos(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,int nn=nbasis) 
     {
       // Run a Lanczos with sloppy convergence
@@ -195,7 +200,7 @@ namespace Grid {
 	  std::cout << GridLogMessage <<"subspace["<<b<<"] = "<<norm2(subspace[b])<<std::endl;
 	}
     }
-
+    */
     virtual void CreateSubspace(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,int nn=nbasis) {
 
       RealD scale;

lib/algorithms/FFT.h

@@ -230,6 +230,7 @@ namespace Grid {
       // Barrel shift and collect global pencil
       std::vector<int> lcoor(Nd), gcoor(Nd);
       result = source;
+      int pc = processor_coor[dim];
       for(int p=0;p<processors[dim];p++) {
         PARALLEL_REGION
         {
@@ -240,7 +241,8 @@ namespace Grid {
           for(int idx=0;idx<sgrid->lSites();idx++) {
             sgrid->LocalIndexToLocalCoor(idx,cbuf);
             peekLocalSite(s,result,cbuf);
-            cbuf[dim]+=p*L;
+	    cbuf[dim]+=((pc+p) % processors[dim])*L;
+	    //            cbuf[dim]+=p*L;
             pokeLocalSite(s,pgbuf,cbuf);
           }
         }
@@ -278,7 +280,6 @@ namespace Grid {
       flops+= flops_call*NN;
       
       // writing out result
-      int pc = processor_coor[dim];
       PARALLEL_REGION
       {
         std::vector<int> clbuf(Nd), cgbuf(Nd);

lib/algorithms/LinearOperator.h

@@ -162,15 +162,10 @@ namespace Grid {
 	_Mat.M(in,out);
       }
       void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
-	ComplexD dot;
-
 	_Mat.M(in,out);
 	
-	dot= innerProduct(in,out);
-	n1=real(dot);
-
-	dot = innerProduct(out,out);
-	n2=real(dot);
+	ComplexD dot= innerProduct(in,out); n1=real(dot);
+	n2=norm2(out);
       }
       void HermOp(const Field &in, Field &out){
 	_Mat.M(in,out);
@@ -192,10 +187,10 @@ namespace Grid {
 	ni=Mpc(in,tmp);
 	no=MpcDag(tmp,out);
       }
-      void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
+      virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
 	MpcDagMpc(in,out,n1,n2);
       }
-      void HermOp(const Field &in, Field &out){
+      virtual void HermOp(const Field &in, Field &out){
 	RealD n1,n2;
 	HermOpAndNorm(in,out,n1,n2);
       }
@@ -212,7 +207,6 @@ namespace Grid {
       void OpDir  (const Field &in, Field &out,int dir,int disp) {
 	assert(0);
       }
-
     };
     template<class Matrix,class Field>
       class SchurDiagMooeeOperator :  public SchurOperatorBase<Field> {
@@ -270,7 +264,6 @@ namespace Grid {
 	return axpy_norm(out,-1.0,tmp,in);
       }
     };
-
     template<class Matrix,class Field>
       class SchurDiagTwoOperator :  public SchurOperatorBase<Field> {
     protected:
@@ -299,6 +292,45 @@ namespace Grid {
 	return axpy_norm(out,-1.0,tmp,in);
       }
     };
+    ///////////////////////////////////////////////////////////////////////////////////////////////////
+    // Left  handed Moo^-1 ; (Moo - Moe Mee^-1 Meo) psi = eta  -->  ( 1 - Moo^-1 Moe Mee^-1 Meo ) psi = Moo^-1 eta
+    // Right handed Moo^-1 ; (Moo - Moe Mee^-1 Meo) Moo^-1 Moo psi = eta  -->  ( 1 - Moe Mee^-1 Meo ) Moo^-1 phi=eta ; psi = Moo^-1 phi
+    ///////////////////////////////////////////////////////////////////////////////////////////////////
+    template<class Matrix,class Field> using SchurDiagOneRH = SchurDiagTwoOperator<Matrix,Field> ;
+    template<class Matrix,class Field> using SchurDiagOneLH = SchurDiagOneOperator<Matrix,Field> ;
+    ///////////////////////////////////////////////////////////////////////////////////////////////////
+    //  Staggered use
+    ///////////////////////////////////////////////////////////////////////////////////////////////////
+    template<class Matrix,class Field>
+      class SchurStaggeredOperator :  public SchurOperatorBase<Field> {
+    protected:
+      Matrix &_Mat;
+    public:
+      SchurStaggeredOperator (Matrix &Mat): _Mat(Mat){};
+      virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
+	n2 = Mpc(in,out);
+	ComplexD dot= innerProduct(in,out);
+	n1 = real(dot);
+      }
+      virtual void HermOp(const Field &in, Field &out){
+	Mpc(in,out);
+      }
+      virtual  RealD Mpc      (const Field &in, Field &out) {
+	Field tmp(in._grid);
+	_Mat.Meooe(in,tmp);
+	_Mat.MooeeInv(tmp,out);
+	_Mat.Meooe(out,tmp);
+	_Mat.Mooee(in,out);
+        return axpy_norm(out,-1.0,tmp,out);
+      }
+      virtual  RealD MpcDag   (const Field &in, Field &out){
+	return Mpc(in,out);
+      }
+      virtual void MpcDagMpc(const Field &in, Field &out,RealD &ni,RealD &no) {
+	assert(0);// Never need with staggered
+      }
+    };
+    template<class Matrix,class Field> using SchurStagOperator = SchurStaggeredOperator<Matrix,Field>;
 
 
     /////////////////////////////////////////////////////////////
@@ -314,6 +346,14 @@ namespace Grid {
       virtual void operator() (const Field &in, Field &out) = 0;
     };
 
+    template<class Field> class IdentityLinearFunction : public LinearFunction<Field> {
+    public:
+      void operator() (const Field &in, Field &out){
+	out = in;
+      };
+    };
+
+
     /////////////////////////////////////////////////////////////
     // Base classes for Multishift solvers for operators
     /////////////////////////////////////////////////////////////
@@ -336,6 +376,64 @@ namespace Grid {
      };
     */
 
+  ////////////////////////////////////////////////////////////////////////////////////////////
+  // Hermitian operator Linear function and operator function
+  ////////////////////////////////////////////////////////////////////////////////////////////
+    template<class Field>
+      class HermOpOperatorFunction : public OperatorFunction<Field> {
+      void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
+	Linop.HermOp(in,out);
+      };
+    };
+
+    template<typename Field>
+      class PlainHermOp : public LinearFunction<Field> {
+    public:
+      LinearOperatorBase<Field> &_Linop;
+      
+      PlainHermOp(LinearOperatorBase<Field>& linop) : _Linop(linop) 
+      {}
+      
+      void operator()(const Field& in, Field& out) {
+	_Linop.HermOp(in,out);
+      }
+    };
+
+    template<typename Field>
+    class FunctionHermOp : public LinearFunction<Field> {
+    public:
+      OperatorFunction<Field>   & _poly;
+      LinearOperatorBase<Field> &_Linop;
+      
+      FunctionHermOp(OperatorFunction<Field> & poly,LinearOperatorBase<Field>& linop) 
+	: _poly(poly), _Linop(linop) {};
+      
+      void operator()(const Field& in, Field& out) {
+	_poly(_Linop,in,out);
+      }
+    };
+
+  template<class Field>
+  class Polynomial : public OperatorFunction<Field> {
+  private:
+    std::vector<RealD> Coeffs;
+  public:
+    Polynomial(std::vector<RealD> &_Coeffs) : Coeffs(_Coeffs) { };
+
+    // Implement the required interface
+    void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
+
+      Field AtoN(in._grid);
+      Field Mtmp(in._grid);
+      AtoN = in;
+      out = AtoN*Coeffs[0];
+      for(int n=1;n<Coeffs.size();n++){
+	Mtmp = AtoN;
+	Linop.HermOp(Mtmp,AtoN);
+	out=out+AtoN*Coeffs[n];
+      }
+    };
+  };
 
 }
 

lib/algorithms/approx/Chebyshev.h

@@ -8,6 +8,7 @@
 
 Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Christoph Lehner <clehner@bnl.gov>
 
     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
@@ -33,41 +34,12 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 
 namespace Grid {
 
-  ////////////////////////////////////////////////////////////////////////////////////////////
-  // Simple general polynomial with user supplied coefficients
-  ////////////////////////////////////////////////////////////////////////////////////////////
-  template<class Field>
-  class HermOpOperatorFunction : public OperatorFunction<Field> {
-    void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
-      Linop.HermOp(in,out);
-    };
-  };
-
-  template<class Field>
-  class Polynomial : public OperatorFunction<Field> {
-  private:
-    std::vector<RealD> Coeffs;
-  public:
-    Polynomial(std::vector<RealD> &_Coeffs) : Coeffs(_Coeffs) { };
-
-    // Implement the required interface
-    void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
-
-      Field AtoN(in._grid);
-      Field Mtmp(in._grid);
-      AtoN = in;
-      out = AtoN*Coeffs[0];
-//            std::cout <<"Poly in " <<norm2(in)<<" size "<< Coeffs.size()<<std::endl;
-//            std::cout <<"Coeffs[0]= "<<Coeffs[0]<< " 0 " <<norm2(out)<<std::endl;
-      for(int n=1;n<Coeffs.size();n++){
-	Mtmp = AtoN;
-	Linop.HermOp(Mtmp,AtoN);
-	out=out+AtoN*Coeffs[n];
-//            std::cout <<"Coeffs "<<n<<"= "<< Coeffs[n]<< " 0 " <<std::endl;
-//		std::cout << n<<" " <<norm2(out)<<std::endl;
-      }
-    };
-  };
+struct ChebyParams : Serializable {
+  GRID_SERIALIZABLE_CLASS_MEMBERS(ChebyParams,
+				  RealD, alpha,  
+				  RealD, beta,   
+				  int, Npoly);
+};
 
   ////////////////////////////////////////////////////////////////////////////////////////////
   // Generic Chebyshev approximations
@@ -83,7 +55,9 @@ namespace Grid {
   public:
     void csv(std::ostream &out){
       RealD diff = hi-lo;
-      for (RealD x=lo-0.2*diff; x<hi+0.2*diff; x+=(hi-lo)/1000) {
+      RealD delta = (hi-lo)*1.0e-9;
+      for (RealD x=lo; x<hi; x+=delta) {
+	delta*=1.1;
 	RealD f = approx(x);
 	out<< x<<" "<<f<<std::endl;
       }
@@ -99,6 +73,7 @@ namespace Grid {
     };
 
     Chebyshev(){};
+    Chebyshev(ChebyParams p){ Init(p.alpha,p.beta,p.Npoly);};
     Chebyshev(RealD _lo,RealD _hi,int _order, RealD (* func)(RealD) ) {Init(_lo,_hi,_order,func);};
     Chebyshev(RealD _lo,RealD _hi,int _order) {Init(_lo,_hi,_order);};
 
@@ -193,6 +168,47 @@ namespace Grid {
       return sum;
     };
 
+    RealD approxD(RealD x)
+    {
+      RealD Un;
+      RealD Unm;
+      RealD Unp;
+      
+      RealD y=( x-0.5*(hi+lo))/(0.5*(hi-lo));
+      
+      RealD U0=1;
+      RealD U1=2*y;
+      
+      RealD sum;
+      sum = Coeffs[1]*U0;
+      sum+= Coeffs[2]*U1*2.0;
+      
+      Un =U1;
+      Unm=U0;
+      for(int i=2;i<order-1;i++){
+	Unp=2*y*Un-Unm;
+	Unm=Un;
+	Un =Unp;
+	sum+= Un*Coeffs[i+1]*(i+1.0);
+      }
+      return sum/(0.5*(hi-lo));
+    };
+    
+    RealD approxInv(RealD z, RealD x0, int maxiter, RealD resid) {
+      RealD x = x0;
+      RealD eps;
+      
+      int i;
+      for (i=0;i<maxiter;i++) {
+	eps = approx(x) - z;
+	if (fabs(eps / z) < resid)
+	  return x;
+	x = x - eps / approxD(x);
+      }
+      
+      return std::numeric_limits<double>::quiet_NaN();
+    }
+    
     // Implement the required interface
     void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
 

lib/algorithms/approx/Forecast.h

@@ -0,0 +1,152 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/algorithms/approx/Forecast.h
+
+Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: David Murphy <dmurphy@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 INCLUDED_FORECAST_H
+#define INCLUDED_FORECAST_H
+
+namespace Grid {
+
+  // Abstract base class.
+  // Takes a matrix (Mat), a source (phi), and a vector of Fields (chi)
+  // and returns a forecasted solution to the system D*psi = phi (psi).
+  template<class Matrix, class Field>
+  class Forecast
+  {
+    public:
+      virtual Field operator()(Matrix &Mat, const Field& phi, const std::vector<Field>& chi) = 0;
+  };
+
+  // Implementation of Brower et al.'s chronological inverter (arXiv:hep-lat/9509012),
+  // used to forecast solutions across poles of the EOFA heatbath.
+  //
+  // Modified from CPS (cps_pp/src/util/dirac_op/d_op_base/comsrc/minresext.C)
+  template<class Matrix, class Field>
+  class ChronoForecast : public Forecast<Matrix,Field>
+  {
+    public:
+      Field operator()(Matrix &Mat, const Field& phi, const std::vector<Field>& prev_solns)
+      {
+        int degree = prev_solns.size();
+        Field chi(phi); // forecasted solution
+
+        // Trivial cases
+        if(degree == 0){ chi = zero; return chi; }
+        else if(degree == 1){ return prev_solns[0]; }
+
+        RealD dot;
+        ComplexD xp;
+        Field r(phi); // residual
+        Field Mv(phi);
+        std::vector<Field> v(prev_solns); // orthonormalized previous solutions
+        std::vector<Field> MdagMv(degree,phi);
+
+        // Array to hold the matrix elements
+        std::vector<std::vector<ComplexD>> G(degree, std::vector<ComplexD>(degree));
+
+        // Solution and source vectors
+        std::vector<ComplexD> a(degree);
+        std::vector<ComplexD> b(degree);
+
+        // Orthonormalize the vector basis
+        for(int i=0; i<degree; i++){
+          v[i] *= 1.0/std::sqrt(norm2(v[i]));
+          for(int j=i+1; j<degree; j++){ v[j] -= innerProduct(v[i],v[j]) * v[i]; }
+        }
+
+        // Perform sparse matrix multiplication and construct rhs
+        for(int i=0; i<degree; i++){
+          b[i] = innerProduct(v[i],phi);
+          Mat.M(v[i],Mv);
+          Mat.Mdag(Mv,MdagMv[i]);
+          G[i][i] = innerProduct(v[i],MdagMv[i]);
+        }
+
+        // Construct the matrix
+        for(int j=0; j<degree; j++){
+        for(int k=j+1; k<degree; k++){
+          G[j][k] = innerProduct(v[j],MdagMv[k]);
+          G[k][j] = std::conj(G[j][k]);
+        }}
+
+        // Gauss-Jordan elimination with partial pivoting
+        for(int i=0; i<degree; i++){
+
+          // Perform partial pivoting
+          int k = i;
+          for(int j=i+1; j<degree; j++){ if(std::abs(G[j][j]) > std::abs(G[k][k])){ k = j; } }
+          if(k != i){
+            xp = b[k];
+            b[k] = b[i];
+            b[i] = xp;
+            for(int j=0; j<degree; j++){
+              xp = G[k][j];
+              G[k][j] = G[i][j];
+              G[i][j] = xp;
+            }
+          }
+
+          // Convert matrix to upper triangular form
+          for(int j=i+1; j<degree; j++){
+            xp = G[j][i]/G[i][i];
+            b[j] -= xp * b[i];
+            for(int k=0; k<degree; k++){ G[j][k] -= xp*G[i][k]; }
+          }
+        }
+
+        // Use Gaussian elimination to solve equations and calculate initial guess
+        chi = zero;
+        r = phi;
+        for(int i=degree-1; i>=0; i--){
+          a[i] = 0.0;
+          for(int j=i+1; j<degree; j++){ a[i] += G[i][j] * a[j]; }
+          a[i] = (b[i]-a[i])/G[i][i];
+          chi += a[i]*v[i];
+          r -= a[i]*MdagMv[i];
+        }
+
+        RealD true_r(0.0);
+        ComplexD tmp;
+        for(int i=0; i<degree; i++){
+          tmp = -b[i];
+          for(int j=0; j<degree; j++){ tmp += G[i][j]*a[j]; }
+          tmp = std::conj(tmp)*tmp;
+          true_r += std::sqrt(tmp.real());
+        }
+
+        RealD error = std::sqrt(norm2(r)/norm2(phi));
+        std::cout << GridLogMessage << "ChronoForecast: |res|/|src| = " << error << std::endl;
+
+        return chi;
+      };
+  };
+
+}
+
+#endif

lib/algorithms/iterative/BlockConjugateGradient.h

@@ -87,15 +87,22 @@ void ThinQRfact (Eigen::MatrixXcd &m_rr,
   ////////////////////////////////////////////////////////////////////////////////////////////////////
   sliceInnerProductMatrix(m_rr,R,R,Orthog);
 
-  ////////////////////////////////////////////////////////////////////////////////////////////////////
-  // Cholesky from Eigen
-  // There exists a ldlt that is documented as more stable
-  ////////////////////////////////////////////////////////////////////////////////////////////////////
-  Eigen::MatrixXcd L    = m_rr.llt().matrixL(); 
+  // Force manifest hermitian to avoid rounding related
+  m_rr = 0.5*(m_rr+m_rr.adjoint());
 
+#if 0
+  std::cout << " Calling Cholesky  ldlt on m_rr "  << m_rr <<std::endl;
+  Eigen::MatrixXcd L_ldlt = m_rr.ldlt().matrixL(); 
+  std::cout << " Called Cholesky  ldlt on m_rr "  << L_ldlt <<std::endl;
+  auto  D_ldlt = m_rr.ldlt().vectorD(); 
+  std::cout << " Called Cholesky  ldlt on m_rr "  << D_ldlt <<std::endl;
+#endif
+
+  //  std::cout << " Calling Cholesky  llt on m_rr "  <<std::endl;
+  Eigen::MatrixXcd L    = m_rr.llt().matrixL(); 
+  //  std::cout << " Called Cholesky  llt on m_rr "  << L <<std::endl;
   C    = L.adjoint();
   Cinv = C.inverse();
-
   ////////////////////////////////////////////////////////////////////////////////////////////////////
   // Q = R C^{-1}
   //
@@ -103,7 +110,6 @@ void ThinQRfact (Eigen::MatrixXcd &m_rr,
   //
   // NB maddMatrix conventions are Right multiplication X[j] a[j,i] already
   ////////////////////////////////////////////////////////////////////////////////////////////////////
-  // FIXME:: make a sliceMulMatrix to avoid zero vector
   sliceMulMatrix(Q,Cinv,R,Orthog);
 }
 ////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -199,7 +205,12 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
 
   Linop.HermOp(X, AD);
   tmp = B - AD;  
+  //std::cout << GridLogMessage << " initial tmp " << norm2(tmp)<< std::endl;
   ThinQRfact (m_rr, m_C, m_Cinv, Q, tmp);
+  //std::cout << GridLogMessage << " initial Q " << norm2(Q)<< std::endl;
+  //std::cout << GridLogMessage << " m_rr " << m_rr<<std::endl;
+  //std::cout << GridLogMessage << " m_C " << m_C<<std::endl;
+  //std::cout << GridLogMessage << " m_Cinv " << m_Cinv<<std::endl;
   D=Q;
 
   std::cout << GridLogMessage<<"BlockCGrQ computed initial residual and QR fact " <<std::endl;
@@ -221,12 +232,14 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
     MatrixTimer.Start();
     Linop.HermOp(D, Z);      
     MatrixTimer.Stop();
+    //std::cout << GridLogMessage << " norm2 Z " <<norm2(Z)<<std::endl;
 
     //4. M  = [D^dag Z]^{-1}
     sliceInnerTimer.Start();
     sliceInnerProductMatrix(m_DZ,D,Z,Orthog);
     sliceInnerTimer.Stop();
     m_M       = m_DZ.inverse();
+    //std::cout << GridLogMessage << " m_DZ " <<m_DZ<<std::endl;
     
     //5. X  = X + D MC
     m_tmp     = m_M * m_C;

lib/algorithms/iterative/ConjugateGradient.h

@@ -52,8 +52,8 @@ class ConjugateGradient : public OperatorFunction<Field> {
         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);
 
@@ -78,12 +78,12 @@ class ConjugateGradient : public OperatorFunction<Field> {
     cp = a;
     ssq = norm2(src);
 
-    std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient: guess " << guess << std::endl;
-    std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient:   src " << ssq << std::endl;
-    std::cout << GridLogIterative << std::setprecision(4) << "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;
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient: guess " << guess << std::endl;
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient:   src " << ssq << std::endl;
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient:    mp " << d << std::endl;
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient:   mmp " << b << std::endl;
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient:  cp,r " << cp << std::endl;
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient:     p " << a << std::endl;
 
     RealD rsq = Tolerance * Tolerance * ssq;
 
@@ -92,7 +92,7 @@ class ConjugateGradient : public OperatorFunction<Field> {
       return;
     }
 
-    std::cout << GridLogIterative << std::setprecision(4)
+    std::cout << GridLogIterative << std::setprecision(8)
               << "ConjugateGradient: k=0 residual " << cp << " target " << rsq << std::endl;
 
     GridStopWatch LinalgTimer;

lib/algorithms/iterative/ConjugateGradientReliableUpdate.h

@@ -0,0 +1,256 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/ConjugateGradientReliableUpdate.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_RELIABLE_UPDATE_H
+#define GRID_CONJUGATE_GRADIENT_RELIABLE_UPDATE_H
+
+namespace Grid {
+
+  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 ConjugateGradientReliableUpdate : public LinearFunction<FieldD> {
+  public:
+    bool ErrorOnNoConverge;  // throw an assert when the CG fails to converge.
+    // Defaults true.
+    RealD Tolerance;
+    Integer MaxIterations;
+    Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
+    Integer ReliableUpdatesPerformed;
+
+    bool DoFinalCleanup; //Final DP cleanup, defaults to true
+    Integer IterationsToCleanup; //Final DP cleanup step iterations
+    
+    LinearOperatorBase<FieldF> &Linop_f;
+    LinearOperatorBase<FieldD> &Linop_d;
+    GridBase* SinglePrecGrid;
+    RealD Delta; //reliable update parameter
+
+    //Optional ability to switch to a different linear operator once the tolerance reaches a certain point. Useful for single/half -> single/single
+    LinearOperatorBase<FieldF> *Linop_fallback;
+    RealD fallback_transition_tol;
+
+    
+    ConjugateGradientReliableUpdate(RealD tol, Integer maxit, RealD _delta, GridBase* _sp_grid, LinearOperatorBase<FieldF> &_Linop_f, LinearOperatorBase<FieldD> &_Linop_d, bool err_on_no_conv = true)
+      : Tolerance(tol),
+        MaxIterations(maxit),
+	Delta(_delta),
+	Linop_f(_Linop_f),
+	Linop_d(_Linop_d),
+	SinglePrecGrid(_sp_grid),
+        ErrorOnNoConverge(err_on_no_conv),
+	DoFinalCleanup(true),
+	Linop_fallback(NULL)
+    {};
+
+    void setFallbackLinop(LinearOperatorBase<FieldF> &_Linop_fallback, const RealD _fallback_transition_tol){
+      Linop_fallback = &_Linop_fallback;
+      fallback_transition_tol = _fallback_transition_tol;      
+    }
+    
+    void operator()(const FieldD &src, FieldD &psi) {
+      LinearOperatorBase<FieldF> *Linop_f_use = &Linop_f;
+      bool using_fallback = false;
+      
+      psi.checkerboard = src.checkerboard;
+      conformable(psi, src);
+
+      RealD cp, c, a, d, b, ssq, qq, b_pred;
+
+      FieldD p(src);
+      FieldD mmp(src);
+      FieldD r(src);
+
+      // Initial residual computation & set up
+      RealD guess = norm2(psi);
+      assert(std::isnan(guess) == 0);
+    
+      Linop_d.HermOpAndNorm(psi, mmp, d, b);
+    
+      r = src - mmp;
+      p = r;
+
+      a = norm2(p);
+      cp = a;
+      ssq = norm2(src);
+
+      std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate: guess " << guess << std::endl;
+      std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate:   src " << ssq << std::endl;
+      std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate:    mp " << d << std::endl;
+      std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate:   mmp " << b << std::endl;
+      std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate:  cp,r " << cp << std::endl;
+      std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate:     p " << a << std::endl;
+
+      RealD rsq = Tolerance * Tolerance * ssq;
+
+      // Check if guess is really REALLY good :)
+      if (cp <= rsq) {
+	std::cout << GridLogMessage << "ConjugateGradientReliableUpdate guess was REALLY good\n";
+	std::cout << GridLogMessage << "\tComputed residual " << sqrt(cp / ssq)<<std::endl;
+	return;
+      }
+
+      //Single prec initialization
+      FieldF r_f(SinglePrecGrid);
+      r_f.checkerboard = r.checkerboard;
+      precisionChange(r_f, r);
+
+      FieldF psi_f(r_f);
+      psi_f = zero;
+
+      FieldF p_f(r_f);
+      FieldF mmp_f(r_f);
+
+      RealD MaxResidSinceLastRelUp = cp; //initial residual    
+    
+      std::cout << GridLogIterative << std::setprecision(4)
+		<< "ConjugateGradient: k=0 residual " << cp << " target " << rsq << std::endl;
+
+      GridStopWatch LinalgTimer;
+      GridStopWatch MatrixTimer;
+      GridStopWatch SolverTimer;
+
+      SolverTimer.Start();
+      int k = 0;
+      int l = 0;
+    
+      for (k = 1; k <= MaxIterations; k++) {
+	c = cp;
+
+	MatrixTimer.Start();
+	Linop_f_use->HermOpAndNorm(p_f, mmp_f, d, qq);
+	MatrixTimer.Stop();
+
+	LinalgTimer.Start();
+
+	a = c / d;
+	b_pred = a * (a * qq - d) / c;
+
+	cp = axpy_norm(r_f, -a, mmp_f, r_f);
+	b = cp / c;
+
+	// Fuse these loops ; should be really easy
+	psi_f = a * p_f + psi_f;
+	//p_f = p_f * b + r_f;
+
+	LinalgTimer.Stop();
+
+	std::cout << GridLogIterative << "ConjugateGradientReliableUpdate: Iteration " << k
+		  << " residual " << cp << " target " << rsq << std::endl;
+	std::cout << GridLogDebug << "a = "<< a << " b_pred = "<< b_pred << "  b = "<< b << std::endl;
+	std::cout << GridLogDebug << "qq = "<< qq << " d = "<< d << "  c = "<< c << std::endl;
+
+	if(cp > MaxResidSinceLastRelUp){
+	  std::cout << GridLogIterative << "ConjugateGradientReliableUpdate: updating MaxResidSinceLastRelUp : " << MaxResidSinceLastRelUp << " -> " << cp << std::endl;
+	  MaxResidSinceLastRelUp = cp;
+	}
+	  
+	// Stopping condition
+	if (cp <= rsq) {
+	  //Although not written in the paper, I assume that I have to add on the final solution
+	  precisionChange(mmp, psi_f);
+	  psi = psi + mmp;
+	
+	
+	  SolverTimer.Stop();
+	  Linop_d.HermOpAndNorm(psi, mmp, d, qq);
+	  p = mmp - src;
+
+	  RealD srcnorm = sqrt(norm2(src));
+	  RealD resnorm = sqrt(norm2(p));
+	  RealD true_residual = resnorm / srcnorm;
+
+	  std::cout << GridLogMessage << "ConjugateGradientReliableUpdate Converged on iteration " << k << " after " << l << " reliable updates" << std::endl;
+	  std::cout << GridLogMessage << "\tComputed residual " << sqrt(cp / ssq)<<std::endl;
+	  std::cout << GridLogMessage << "\tTrue residual " << true_residual<<std::endl;
+	  std::cout << GridLogMessage << "\tTarget " << Tolerance << std::endl;
+
+	  std::cout << GridLogMessage << "Time breakdown "<<std::endl;
+	  std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed() <<std::endl;
+	  std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
+	  std::cout << GridLogMessage << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
+
+	  IterationsToComplete = k;	
+	  ReliableUpdatesPerformed = l;
+	  
+	  if(DoFinalCleanup){
+	    //Do a final CG to cleanup
+	    std::cout << GridLogMessage << "ConjugateGradientReliableUpdate performing final cleanup.\n";
+	    ConjugateGradient<FieldD> CG(Tolerance,MaxIterations);
+	    CG.ErrorOnNoConverge = ErrorOnNoConverge;
+	    CG(Linop_d,src,psi);
+	    IterationsToCleanup = CG.IterationsToComplete;
+	  }
+	  else if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
+
+	  std::cout << GridLogMessage << "ConjugateGradientReliableUpdate complete.\n";
+	  return;
+	}
+	else if(cp < Delta * MaxResidSinceLastRelUp) { //reliable update
+	  std::cout << GridLogMessage << "ConjugateGradientReliableUpdate "
+		    << cp << "(residual) < " << Delta << "(Delta) * " << MaxResidSinceLastRelUp << "(MaxResidSinceLastRelUp) on iteration " << k << " : performing reliable update\n";
+	  precisionChange(mmp, psi_f);
+	  psi = psi + mmp;
+
+	  Linop_d.HermOpAndNorm(psi, mmp, d, qq);
+	  r = src - mmp;
+
+	  psi_f = zero;
+	  precisionChange(r_f, r);
+	  cp = norm2(r);
+	  MaxResidSinceLastRelUp = cp;
+
+	  b = cp/c;
+	  
+	  std::cout << GridLogMessage << "ConjugateGradientReliableUpdate new residual " << cp << std::endl;
+	  
+	  l = l+1;
+	}
+
+	p_f = p_f * b + r_f; //update search vector after reliable update appears to help convergence
+
+	if(!using_fallback && Linop_fallback != NULL && cp < fallback_transition_tol){
+	  std::cout << GridLogMessage << "ConjugateGradientReliableUpdate switching to fallback linear operator on iteration " << k << " at residual " << cp << std::endl;
+	  Linop_f_use = Linop_fallback;
+	  using_fallback = true;
+	}
+
+	
+      }
+      std::cout << GridLogMessage << "ConjugateGradientReliableUpdate did NOT converge"
+		<< std::endl;
+      
+      if (ErrorOnNoConverge) assert(0);
+      IterationsToComplete = k;
+      ReliableUpdatesPerformed = l;      
+    }    
+  };
+
+
+};
+
+
+
+#endif

lib/algorithms/iterative/ImplicitlyRestartedLanczos.h

@@ -7,8 +7,9 @@
     Copyright (C) 2015
 
 Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Chulwoo Jung
-Author: Guido Cossu
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Chulwoo Jung <chulwoo@bnl.gov>
+Author: Christoph Lehner <clehner@bnl.gov>
 
     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
@@ -27,108 +28,269 @@ Author: Guido Cossu
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#ifndef GRID_IRL_H
-#define GRID_IRL_H
+#ifndef GRID_BIRL_H
+#define GRID_BIRL_H
 
 #include <string.h> //memset
+//#include <zlib.h>
+#include <sys/stat.h>
 
 namespace Grid { 
 
-  enum IRLdiagonalisation { 
-    IRLdiagonaliseWithDSTEGR,
-    IRLdiagonaliseWithQR,
-    IRLdiagonaliseWithEigen
-  };
-
-////////////////////////////////////////////////////////////////////////////////
-// Helper class for sorting the evalues AND evectors by Field
-// Use pointer swizzle on vectors
-////////////////////////////////////////////////////////////////////////////////
+  ////////////////////////////////////////////////////////
+  // Move following 100 LOC to lattice/Lattice_basis.h
+  ////////////////////////////////////////////////////////
 template<class Field>
-class SortEigen {
- private:
-  static bool less_lmd(RealD left,RealD right){
-    return left > right;
-  }  
-  static bool less_pair(std::pair<RealD,Field const*>& left,
-                        std::pair<RealD,Field const*>& right){
-    return left.first > (right.first);
+void basisOrthogonalize(std::vector<Field> &basis,Field &w,int k) 
+{
+  for(int j=0; j<k; ++j){
+    auto ip = innerProduct(basis[j],w);
+    w = w - ip*basis[j];
   }
+}
 
- public:
-  void push(std::vector<RealD>& lmd,std::vector<Field>& evec,int N) {
+template<class Field>
+void basisRotate(std::vector<Field> &basis,Eigen::MatrixXd& Qt,int j0, int j1, int k0,int k1,int Nm) 
+{
+  typedef typename Field::vector_object vobj;
+  GridBase* grid = basis[0]._grid;
       
-    ////////////////////////////////////////////////////////////////////////
-    // PAB: FIXME: VERY VERY VERY wasteful: takes a copy of the entire vector set.
-    //    : The vector reorder should be done by pointer swizzle somehow
-    ////////////////////////////////////////////////////////////////////////
-    std::vector<Field> cpy(lmd.size(),evec[0]._grid);
-    for(int i=0;i<lmd.size();i++) cpy[i] = evec[i];
+  parallel_region
+  {
+    std::vector < vobj > B(Nm); // Thread private
         
-    std::vector<std::pair<RealD, Field const*> > emod(lmd.size());    
+    parallel_for_internal(int ss=0;ss < grid->oSites();ss++){
+      for(int j=j0; j<j1; ++j) B[j]=0.;
       
-    for(int i=0;i<lmd.size();++i)  emod[i] = std::pair<RealD,Field const*>(lmd[i],&cpy[i]);
-
-    partial_sort(emod.begin(),emod.begin()+N,emod.end(),less_pair);
-
-    typename std::vector<std::pair<RealD, Field const*> >::iterator it = emod.begin();
-    for(int i=0;i<N;++i){
-      lmd[i]=it->first;
-      evec[i]=*(it->second);
-      ++it;
+      for(int j=j0; j<j1; ++j){
+	for(int k=k0; k<k1; ++k){
+	  B[j] +=Qt(j,k) * basis[k]._odata[ss];
 	}
       }
-  void push(std::vector<RealD>& lmd,int N) {
-    std::partial_sort(lmd.begin(),lmd.begin()+N,lmd.end(),less_lmd);
+      for(int j=j0; j<j1; ++j){
+	  basis[j]._odata[ss] = B[j];
       }
-  bool saturated(RealD lmd, RealD thrs) {
-    return fabs(lmd) > fabs(thrs);
     }
-};
+  }
+}
+
+// Extract a single rotated vector
+template<class Field>
+void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,int j, int k0,int k1,int Nm) 
+{
+  typedef typename Field::vector_object vobj;
+  GridBase* grid = basis[0]._grid;
+
+  result.checkerboard = basis[0].checkerboard;
+  parallel_for(int ss=0;ss < grid->oSites();ss++){
+    vobj B = zero;
+    for(int k=k0; k<k1; ++k){
+      B +=Qt(j,k) * basis[k]._odata[ss];
+    }
+    result._odata[ss] = B;
+  }
+}
+
+template<class Field>
+void basisReorderInPlace(std::vector<Field> &_v,std::vector<RealD>& sort_vals, std::vector<int>& idx) 
+{
+  int vlen = idx.size();
+
+  assert(vlen>=1);
+  assert(vlen<=sort_vals.size());
+  assert(vlen<=_v.size());
+
+  for (size_t i=0;i<vlen;i++) {
+
+    if (idx[i] != i) {
+
+      //////////////////////////////////////
+      // idx[i] is a table of desired sources giving a permutation.
+      // Swap v[i] with v[idx[i]].
+      // Find  j>i for which _vnew[j] = _vold[i],
+      // track the move idx[j] => idx[i]
+      // track the move idx[i] => i
+      //////////////////////////////////////
+      size_t j;
+      for (j=i;j<idx.size();j++)
+	if (idx[j]==i)
+	  break;
+
+      assert(idx[i] > i);     assert(j!=idx.size());      assert(idx[j]==i);
+
+      std::swap(_v[i]._odata,_v[idx[i]]._odata); // should use vector move constructor, no data copy
+      std::swap(sort_vals[i],sort_vals[idx[i]]);
+
+      idx[j] = idx[i];
+      idx[i] = i;
+    }
+  }
+}
+
+inline std::vector<int> basisSortGetIndex(std::vector<RealD>& sort_vals) 
+{
+  std::vector<int> idx(sort_vals.size());
+  std::iota(idx.begin(), idx.end(), 0);
+
+  // sort indexes based on comparing values in v
+  std::sort(idx.begin(), idx.end(), [&sort_vals](int i1, int i2) {
+    return ::fabs(sort_vals[i1]) < ::fabs(sort_vals[i2]);
+  });
+  return idx;
+}
+
+template<class Field>
+void basisSortInPlace(std::vector<Field> & _v,std::vector<RealD>& sort_vals, bool reverse) 
+{
+  std::vector<int> idx = basisSortGetIndex(sort_vals);
+  if (reverse)
+    std::reverse(idx.begin(), idx.end());
+  
+  basisReorderInPlace(_v,sort_vals,idx);
+}
+
+// PAB: faster to compute the inner products first then fuse loops.
+// If performance critical can improve.
+template<class Field>
+void basisDeflate(const std::vector<Field> &_v,const std::vector<RealD>& eval,const Field& src_orig,Field& result) {
+  result = zero;
+  assert(_v.size()==eval.size());
+  int N = (int)_v.size();
+  for (int i=0;i<N;i++) {
+    Field& tmp = _v[i];
+    axpy(result,TensorRemove(innerProduct(tmp,src_orig)) / eval[i],tmp,result);
+  }
+}
 
 /////////////////////////////////////////////////////////////
 // Implicitly restarted lanczos
 /////////////////////////////////////////////////////////////
+template<class Field> class ImplicitlyRestartedLanczosTester 
+{
+ public:
+  virtual int TestConvergence(int j,RealD resid,Field &evec, RealD &eval,RealD evalMaxApprox);
+  virtual int ReconstructEval(int j,RealD resid,Field &evec, RealD &eval,RealD evalMaxApprox);
+};
+
+enum IRLdiagonalisation { 
+  IRLdiagonaliseWithDSTEGR,
+  IRLdiagonaliseWithQR,
+  IRLdiagonaliseWithEigen
+};
+
+template<class Field> class ImplicitlyRestartedLanczosHermOpTester  : public ImplicitlyRestartedLanczosTester<Field>
+{
+ public:
+  LinearFunction<Field>       &_HermOpTest;
+  ImplicitlyRestartedLanczosHermOpTester(LinearFunction<Field> &HermOpTest) : _HermOpTest(HermOpTest)  {  };
+  int ReconstructEval(int j,RealD resid,Field &B, RealD &eval,RealD evalMaxApprox)
+  {
+    return TestConvergence(j,resid,B,eval,evalMaxApprox);
+  }
+  int TestConvergence(int j,RealD eresid,Field &B, RealD &eval,RealD evalMaxApprox)
+  {
+    Field v(B);
+    RealD eval_poly = eval;
+    // Apply operator
+    _HermOpTest(B,v);
+
+    RealD vnum = real(innerProduct(B,v)); // HermOp.
+    RealD vden = norm2(B);
+    RealD vv0  = norm2(v);
+    eval   = vnum/vden;
+    v -= eval*B;
+
+    RealD vv = norm2(v) / ::pow(evalMaxApprox,2.0);
+
+    std::cout.precision(13);
+    std::cout<<GridLogIRL  << "[" << std::setw(3)<<j<<"] "
+	     <<"eval = "<<std::setw(25)<< eval << " (" << eval_poly << ")"
+	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
+	     <<std::endl;
+
+    int conv=0;
+    if( (vv<eresid*eresid) ) conv = 1;
+
+    return conv;
+  }
+};
+
 template<class Field> 
 class ImplicitlyRestartedLanczos {
-
-private:       
-
-  int MaxIter;   // Max iterations
+ private:
+  const RealD small = 1.0e-8;
+  int MaxIter;
+  int MinRestart; // Minimum number of restarts; only check for convergence after
   int Nstop;   // Number of evecs checked for convergence
   int Nk;      // Number of converged sought
+  //  int Np;      // Np -- Number of spare vecs in krylov space //  == Nm - Nk
   int Nm;      // Nm -- total number of vectors
-  RealD eresid;
   IRLdiagonalisation diagonalisation;
-  ////////////////////////////////////
-  // Embedded objects
-  ////////////////////////////////////
-           SortEigen<Field> _sort;
-  LinearOperatorBase<Field> &_Linop;
-    OperatorFunction<Field> &_poly;
+  int orth_period;
     
+  RealD OrthoTime;
+  RealD eresid, betastp;
+  ////////////////////////////////
+  // Embedded objects
+  ////////////////////////////////
+  LinearFunction<Field>       &_HermOp;
+  LinearFunction<Field>       &_HermOpTest;
+  ImplicitlyRestartedLanczosTester<Field> &_Tester;
+  // Default tester provided (we need a ref to something in default case)
+  ImplicitlyRestartedLanczosHermOpTester<Field> SimpleTester;
   /////////////////////////
   // Constructor
   /////////////////////////
+  
 public:       
- ImplicitlyRestartedLanczos(LinearOperatorBase<Field> &Linop, // op
-			    OperatorFunction<Field> & poly,   // polynomial
-			    int _Nstop, // really sought vecs
+  //////////////////////////////////////////////////////////////////
+  // PAB:
+  //////////////////////////////////////////////////////////////////
+  // Too many options  & knobs. Do we really need orth_period
+  // What is the theoretical basis & guarantees of betastp ?
+  // Nstop=Nk viable?
+  // MinRestart avoidable with new convergence test?
+  // Could cut to HermOp, HermOpTest, Tester, Nk, Nm, resid, maxiter (+diagonalisation)
+  // HermOpTest could be eliminated if we dropped the Power method for max eval.
+  // -- also: The eval, eval2, eval2_copy stuff is still unnecessarily unclear
+  //////////////////////////////////////////////////////////////////
+ ImplicitlyRestartedLanczos(LinearFunction<Field> & HermOp,
+			    LinearFunction<Field> & HermOpTest,
+			    ImplicitlyRestartedLanczosTester<Field> & Tester,
+			    int _Nstop, // sought vecs
 			    int _Nk, // sought vecs
-			    int _Nm,    // total vecs
-			    RealD _eresid, // resid in lmd deficit 
+			    int _Nm, // spare vecs
+			    RealD _eresid, // resid in lmdue deficit 
 			    int _MaxIter, // Max iterations
-			    IRLdiagonalisation _diagonalisation= IRLdiagonaliseWithEigen ) :
-    _Linop(Linop),    _poly(poly),
-      Nstop(_Nstop), Nk(_Nk), Nm(_Nm),
-      eresid(_eresid),  MaxIter(_MaxIter),
-      diagonalisation(_diagonalisation)
-      { };
+			    RealD _betastp=0.0, // if beta(k) < betastp: converged
+			    int _MinRestart=1, int _orth_period = 1,
+			    IRLdiagonalisation _diagonalisation= IRLdiagonaliseWithEigen) :
+    SimpleTester(HermOpTest), _HermOp(HermOp),      _HermOpTest(HermOpTest), _Tester(Tester),
+    Nstop(_Nstop)  ,      Nk(_Nk),      Nm(_Nm),
+    eresid(_eresid),      betastp(_betastp),
+    MaxIter(_MaxIter)  ,      MinRestart(_MinRestart),
+    orth_period(_orth_period), diagonalisation(_diagonalisation)  { };
+
+    ImplicitlyRestartedLanczos(LinearFunction<Field> & HermOp,
+			       LinearFunction<Field> & HermOpTest,
+			       int _Nstop, // sought vecs
+			       int _Nk, // sought vecs
+			       int _Nm, // spare vecs
+			       RealD _eresid, // resid in lmdue deficit 
+			       int _MaxIter, // Max iterations
+			       RealD _betastp=0.0, // if beta(k) < betastp: converged
+			       int _MinRestart=1, int _orth_period = 1,
+			       IRLdiagonalisation _diagonalisation= IRLdiagonaliseWithEigen) :
+    SimpleTester(HermOpTest),  _HermOp(HermOp),      _HermOpTest(HermOpTest), _Tester(SimpleTester),
+    Nstop(_Nstop)  ,      Nk(_Nk),      Nm(_Nm),
+    eresid(_eresid),      betastp(_betastp),
+    MaxIter(_MaxIter)  ,      MinRestart(_MinRestart),
+    orth_period(_orth_period), diagonalisation(_diagonalisation)  { };
 
   ////////////////////////////////
   // Helpers
   ////////////////////////////////
-  static RealD normalise(Field& v) 
+  template<typename T>  static RealD normalise(T& v) 
   {
     RealD nn = norm2(v);
     nn = sqrt(nn);
@@ -136,16 +298,12 @@ public:
     return nn;
   }
 
-  void orthogonalize(Field& w, std::vector<Field>& evec, int k)
+  void orthogonalize(Field& w, std::vector<Field>& evec,int k)
   {
-    typedef typename Field::scalar_type MyComplex;
-    MyComplex ip;
-    
-    for(int j=0; j<k; ++j){
-      ip = innerProduct(evec[j],w); 
-      w = w - ip * evec[j];
-    }
+    OrthoTime-=usecond()/1e6;
+    basisOrthogonalize(evec,w,k);
     normalise(w);
+    OrthoTime+=usecond()/1e6;
   }
 
 /* Rudy Arthur's thesis pp.137
@@ -165,184 +323,234 @@ repeat
   →AVK =VKHK +fKe†K † Extend to an M = K + P step factorization AVM = VMHM + fMeM
 until convergence
 */
-  void calc(std::vector<RealD>& eval,  std::vector<Field>& evec, const Field& src, int& Nconv)
+  void calc(std::vector<RealD>& eval, std::vector<Field>& evec,  const Field& src, int& Nconv, bool reverse=true)
   {
+    GridBase *grid = src._grid;
+    assert(grid == evec[0]._grid);
     
-    GridBase *grid = evec[0]._grid;
-    assert(grid == src._grid);
-    
-    std::cout << GridLogMessage <<"**************************************************************************"<< std::endl;
-    std::cout << GridLogMessage <<" ImplicitlyRestartedLanczos::calc() starting iteration 0 /  "<< MaxIter<< std::endl;
-    std::cout << GridLogMessage <<"**************************************************************************"<< std::endl;
-    std::cout << GridLogMessage <<" -- seek   Nk    = " << Nk    <<" vectors"<< std::endl;
-    std::cout << GridLogMessage <<" -- accept Nstop = " << Nstop <<" vectors"<< std::endl;
-    std::cout << GridLogMessage <<" -- total  Nm    = " << Nm    <<" vectors"<< std::endl;
-    std::cout << GridLogMessage <<" -- size of eval = " << eval.size() << std::endl;
-    std::cout << GridLogMessage <<" -- size of evec = " << evec.size() << std::endl;
+    GridLogIRL.TimingMode(1);
+    std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
+    std::cout << GridLogIRL <<" ImplicitlyRestartedLanczos::calc() starting iteration 0 /  "<< MaxIter<< std::endl;
+    std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
+    std::cout << GridLogIRL <<" -- seek   Nk    = " << Nk    <<" vectors"<< std::endl;
+    std::cout << GridLogIRL <<" -- accept Nstop = " << Nstop <<" vectors"<< std::endl;
+    std::cout << GridLogIRL <<" -- total  Nm    = " << Nm    <<" vectors"<< std::endl;
+    std::cout << GridLogIRL <<" -- size of eval = " << eval.size() << std::endl;
+    std::cout << GridLogIRL <<" -- size of evec = " << evec.size() << std::endl;
     if ( diagonalisation == IRLdiagonaliseWithDSTEGR ) {
-      std::cout << GridLogMessage << "Diagonalisation is DSTEGR "<<std::endl;
+      std::cout << GridLogIRL << "Diagonalisation is DSTEGR "<<std::endl;
     } else if ( diagonalisation == IRLdiagonaliseWithQR ) { 
-      std::cout << GridLogMessage << "Diagonalisation is QR "<<std::endl;
+      std::cout << GridLogIRL << "Diagonalisation is QR "<<std::endl;
     }  else if ( diagonalisation == IRLdiagonaliseWithEigen ) { 
-      std::cout << GridLogMessage << "Diagonalisation is Eigen "<<std::endl;
+      std::cout << GridLogIRL << "Diagonalisation is Eigen "<<std::endl;
     }
-    std::cout << GridLogMessage <<"**************************************************************************"<< std::endl;
+    std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
 	
-    assert(Nm == evec.size() && Nm == eval.size());
+    assert(Nm <= evec.size() && Nm <= eval.size());
+    
+    // quickly get an idea of the largest eigenvalue to more properly normalize the residuum
+    RealD evalMaxApprox = 0.0;
+    {
+      auto src_n = src;
+      auto tmp = src;
+      const int _MAX_ITER_IRL_MEVAPP_ = 50;
+      for (int i=0;i<_MAX_ITER_IRL_MEVAPP_;i++) {
+	_HermOpTest(src_n,tmp);
+	RealD vnum = real(innerProduct(src_n,tmp)); // HermOp.
+	RealD vden = norm2(src_n);
+	RealD na = vnum/vden;
+	if (fabs(evalMaxApprox/na - 1.0) < 0.05)
+	  i=_MAX_ITER_IRL_MEVAPP_;
+	evalMaxApprox = na;
+	std::cout << GridLogIRL << " Approximation of largest eigenvalue: " << evalMaxApprox << std::endl;
+	src_n = tmp;
+      }
+    }
 	
     std::vector<RealD> lme(Nm);  
     std::vector<RealD> lme2(Nm);
     std::vector<RealD> eval2(Nm);
-
+    std::vector<RealD> eval2_copy(Nm);
     Eigen::MatrixXd Qt = Eigen::MatrixXd::Zero(Nm,Nm);
 
-    std::vector<int>   Iconv(Nm);
-    std::vector<Field>  B(Nm,grid); // waste of space replicating
-    
     Field f(grid);
     Field v(grid);
-    
     int k1 = 1;
     int k2 = Nk;
+    RealD beta_k;
 
     Nconv = 0;
   
-    RealD beta_k;
-  
     // Set initial vector
     evec[0] = src;
-    std::cout << GridLogMessage <<"norm2(src)= " << norm2(src)<<std::endl;
-    
     normalise(evec[0]);
-    std::cout << GridLogMessage <<"norm2(evec[0])= " << norm2(evec[0]) <<std::endl;
 	
     // Initial Nk steps
+    OrthoTime=0.;
     for(int k=0; k<Nk; ++k) step(eval,lme,evec,f,Nm,k);
+    std::cout<<GridLogIRL <<"Initial "<< Nk <<"steps done "<<std::endl;
+    std::cout<<GridLogIRL <<"Initial steps:OrthoTime "<<OrthoTime<< "seconds"<<std::endl;
 
+    //////////////////////////////////
     // Restarting loop begins
+    //////////////////////////////////
     int iter;
     for(iter = 0; iter<MaxIter; ++iter){
       
+      OrthoTime=0.;
+
       std::cout<< GridLogMessage <<" **********************"<< std::endl;
       std::cout<< GridLogMessage <<" Restart iteration = "<< iter << std::endl;
       std::cout<< GridLogMessage <<" **********************"<< std::endl;
 
+      std::cout<<GridLogIRL <<" running "<<Nm-Nk <<" steps: "<<std::endl;
       for(int k=Nk; k<Nm; ++k) step(eval,lme,evec,f,Nm,k);
-      
       f *= lme[Nm-1];
 
+      std::cout<<GridLogIRL <<" "<<Nm-Nk <<" steps done "<<std::endl;
+      std::cout<<GridLogIRL <<"Initial steps:OrthoTime "<<OrthoTime<< "seconds"<<std::endl;
+	  
+      //////////////////////////////////
       // getting eigenvalues
+      //////////////////////////////////
       for(int k=0; k<Nm; ++k){
 	eval2[k] = eval[k+k1-1];
 	lme2[k] = lme[k+k1-1];
       }
       Qt = Eigen::MatrixXd::Identity(Nm,Nm);
       diagonalize(eval2,lme2,Nm,Nm,Qt,grid);
+      std::cout<<GridLogIRL <<" diagonalized "<<std::endl;
 
+      //////////////////////////////////
       // sorting
-      _sort.push(eval2,Nm);
+      //////////////////////////////////
+      eval2_copy = eval2;
+      std::partial_sort(eval2.begin(),eval2.begin()+Nm,eval2.end(),std::greater<RealD>());
+      std::cout<<GridLogIRL <<" evals sorted "<<std::endl;
+      const int chunk=8;
+      for(int io=0; io<k2;io+=chunk){
+	std::cout<<GridLogIRL << "eval "<< std::setw(3) << io ;
+	for(int ii=0;ii<chunk;ii++){
+	  if ( (io+ii)<k2 )
+	    std::cout<< " "<< std::setw(12)<< eval2[io+ii];
+	}
+	std::cout << std::endl;
+      }
 
+      //////////////////////////////////
       // Implicitly shifted QR transformations
+      //////////////////////////////////
       Qt = Eigen::MatrixXd::Identity(Nm,Nm);
       for(int ip=k2; ip<Nm; ++ip){ 
-	// Eigen replacement for qr_decomp ???
-	qr_decomp(eval,lme,Nm,Nm,Qt,eval2[ip],k1,Nm);
+	QR_decomp(eval,lme,Nm,Nm,Qt,eval2[ip],k1,Nm);
       }
+      std::cout<<GridLogIRL <<"QR decomposed "<<std::endl;
 
-      for(int i=0; i<(Nk+1); ++i) B[i] = 0.0;
+      assert(k2<Nm);      assert(k2<Nm);      assert(k1>0);
 
-      for(int j=k1-1; j<k2+1; ++j){
-	for(int k=0; k<Nm; ++k){
-	  B[j].checkerboard = evec[k].checkerboard;
-	  B[j] += Qt(j,k) * evec[k];
-	}
-      }
-      for(int j=k1-1; j<k2+1; ++j) evec[j] = B[j];
+      basisRotate(evec,Qt,k1-1,k2+1,0,Nm,Nm); /// big constraint on the basis
+      std::cout<<GridLogIRL <<"basisRotated  by Qt"<<std::endl;
       
+      ////////////////////////////////////////////////////
       // Compressed vector f and beta(k2)
+      ////////////////////////////////////////////////////
       f *= Qt(k2-1,Nm-1);
       f += lme[k2-1] * evec[k2];
       beta_k = norm2(f);
       beta_k = sqrt(beta_k);
-      std::cout<< GridLogMessage<<" beta(k) = "<<beta_k<<std::endl;
+      std::cout<<GridLogIRL<<" beta(k) = "<<beta_k<<std::endl;
 	  
       RealD betar = 1.0/beta_k;
       evec[k2] = betar * f;
       lme[k2-1] = beta_k;
 	  
+      ////////////////////////////////////////////////////
       // Convergence test
+      ////////////////////////////////////////////////////
       for(int k=0; k<Nm; ++k){    
 	eval2[k] = eval[k];
 	lme2[k] = lme[k];
       }
       Qt = Eigen::MatrixXd::Identity(Nm,Nm);
       diagonalize(eval2,lme2,Nk,Nm,Qt,grid);
-      
-      for(int k = 0; k<Nk; ++k) B[k]=0.0;
-      
-      for(int j = 0; j<Nk; ++j){
-	for(int k = 0; k<Nk; ++k){
-	  B[j].checkerboard = evec[k].checkerboard;
-	  B[j] += Qt(j,k) * evec[k];
-	}
-      }
+      std::cout<<GridLogIRL <<" Diagonalized "<<std::endl;
 	  
       Nconv = 0;
-      for(int i=0; i<Nk; ++i){
+      if (iter >= MinRestart) {
 
-	_Linop.HermOp(B[i],v);
+	std::cout << GridLogIRL << "Test convergence: rotate subset of vectors to test convergence " << std::endl;
 
-	RealD vnum = real(innerProduct(B[i],v)); // HermOp.
-	RealD vden = norm2(B[i]);
-	eval2[i] = vnum/vden;
-	v -= eval2[i]*B[i];
-	RealD vv = norm2(v);
+	Field B(grid); B.checkerboard = evec[0].checkerboard;
 
-	std::cout.precision(13);
-	std::cout << GridLogMessage << "[" << std::setw(3)<< std::setiosflags(std::ios_base::right) <<i<<"] ";
-	std::cout << "eval = "<<std::setw(25)<< std::setiosflags(std::ios_base::left)<< eval2[i];
-	std::cout << " |H B[i] - eval[i]B[i]|^2 "<< std::setw(25)<< std::setiosflags(std::ios_base::right)<< vv<< std::endl;
-	
-	// change the criteria as evals are supposed to be sorted, all evals smaller(larger) than Nstop should have converged
-	if((vv<eresid*eresid) && (i == Nconv) ){
-	  Iconv[Nconv] = i;
-	  ++Nconv;
+	//  power of two search pattern;  not every evalue in eval2 is assessed.
+	for(int jj = 1; jj<=Nstop; jj*=2){
+	  int j = Nstop-jj;
+	  RealD e = eval2_copy[j]; // Discard the evalue
+	  basisRotateJ(B,evec,Qt,j,0,Nk,Nm);	    
+	  if( _Tester.TestConvergence(j,eresid,B,e,evalMaxApprox) ) {
+	    if ( j > Nconv ) {
+	      Nconv=j+1;
+	      jj=Nstop; // Terminate the scan
 	    }
-	
-      }  // i-loop end
-      
-      std::cout<< GridLogMessage <<" #modes converged: "<<Nconv<<std::endl;
-
-      if( Nconv>=Nstop ){
+	  }
+	}
+	// Do evec[0] for good measure
+	{ 
+	  int j=0;
+	  RealD e = eval2_copy[0]; 
+	  basisRotateJ(B,evec,Qt,j,0,Nk,Nm);	    
+	  _Tester.TestConvergence(j,eresid,B,e,evalMaxApprox);
+	}
+	// test if we converged, if so, terminate
+	std::cout<<GridLogIRL<<" #modes converged: >= "<<Nconv<<"/"<<Nstop<<std::endl;
+	//	if( Nconv>=Nstop || beta_k < betastp){
+	if( Nconv>=Nstop){
 	  goto converged;
 	}
-    } // end of iter loop
 	  
-    std::cout << GridLogMessage <<"**************************************************************************"<< std::endl;
-    std::cout<< GridLogError    <<" ImplicitlyRestartedLanczos::calc() NOT converged.";
-    std::cout << GridLogMessage <<"**************************************************************************"<< std::endl;
+      } else {
+	std::cout << GridLogIRL << "iter < MinRestart: do not yet test for convergence\n";
+      } // end of iter loop
+    }
+
+    std::cout<<GridLogError<<"\n NOT converged.\n";
     abort();
 	
   converged:
-    // Sorting
-    eval.resize(Nconv);
-    evec.resize(Nconv,grid);
-    for(int i=0; i<Nconv; ++i){
-      eval[i] = eval2[Iconv[i]];
-      evec[i] = B[Iconv[i]];
+    {
+      Field B(grid); B.checkerboard = evec[0].checkerboard;
+      basisRotate(evec,Qt,0,Nk,0,Nk,Nm);	    
+      std::cout << GridLogIRL << " Rotated basis"<<std::endl;
+      Nconv=0;
+      //////////////////////////////////////////////////////////////////////
+      // Full final convergence test; unconditionally applied
+      //////////////////////////////////////////////////////////////////////
+      for(int j = 0; j<=Nk; j++){
+	B=evec[j];
+	if( _Tester.ReconstructEval(j,eresid,B,eval2[j],evalMaxApprox) ) {
+	  Nconv++;
 	}
-    _sort.push(eval,evec,Nconv);
-    
-    std::cout << GridLogMessage <<"**************************************************************************"<< std::endl;
-    std::cout << GridLogMessage << "ImplicitlyRestartedLanczos CONVERGED ; Summary :\n";
-    std::cout << GridLogMessage <<"**************************************************************************"<< std::endl;
-    std::cout << GridLogMessage << " -- Iterations  = "<< iter   << "\n";
-    std::cout << GridLogMessage << " -- beta(k)     = "<< beta_k << "\n";
-    std::cout << GridLogMessage << " -- Nconv       = "<< Nconv  << "\n";
-    std::cout << GridLogMessage <<"**************************************************************************"<< std::endl;
       }
 
-private:
+      if ( Nconv < Nstop )
+	std::cout << GridLogIRL << "Nconv ("<<Nconv<<") < Nstop ("<<Nstop<<")"<<std::endl;
+
+      eval=eval2;
+
+      basisSortInPlace(evec,eval,reverse);
+      
+    }
+       
+    std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
+    std::cout << GridLogIRL << "ImplicitlyRestartedLanczos CONVERGED ; Summary :\n";
+    std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
+    std::cout << GridLogIRL << " -- Iterations  = "<< iter   << "\n";
+    std::cout << GridLogIRL << " -- beta(k)     = "<< beta_k << "\n";
+    std::cout << GridLogIRL << " -- Nconv       = "<< Nconv  << "\n";
+    std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
+  }
+
+ private:
 /* Saad PP. 195
 1. Choose an initial vector v1 of 2-norm unity. Set β1 ≡ 0, v0 ≡ 0
 2. For k = 1,2,...,m Do:
@@ -361,14 +569,18 @@ private:
     const RealD tiny = 1.0e-20;
     assert( k< Nm );
 
-    _poly(_Linop,evec[k],w);      // 3. wk:=Avk−βkv_{k−1}
+    GridStopWatch gsw_op,gsw_o;
+
+    Field& evec_k = evec[k];
+
+    _HermOp(evec_k,w);    std::cout<<GridLogIRL << "Poly(HermOp)" <<std::endl;
 
     if(k>0) w -= lme[k-1] * evec[k-1];
 
-    ComplexD zalph = innerProduct(evec[k],w); // 4. αk:=(wk,vk)
+    ComplexD zalph = innerProduct(evec_k,w); // 4. αk:=(wk,vk)
     RealD     alph = real(zalph);
 
-    w = w - alph * evec[k];// 5. wk:=wk−αkvk
+    w = w - alph * evec_k;// 5. wk:=wk−αkvk
 
     RealD beta = normalise(w); // 6. βk+1 := ∥wk∥2. If βk+1 = 0 then Stop
     // 7. vk+1 := wk/βk+1
@@ -376,10 +588,16 @@ private:
     lmd[k] = alph;
     lme[k] = beta;
 
-    if ( k > 0 ) orthogonalize(w,evec,k); // orthonormalise
-    if ( k < Nm-1) evec[k+1] = w;
+    if (k>0 && k % orth_period == 0) {
+      orthogonalize(w,evec,k); // orthonormalise
+      std::cout<<GridLogIRL << "Orthogonalised " <<std::endl;
+    }
 
-    if ( beta < tiny ) std::cout << GridLogMessage << " beta is tiny "<<beta<<std::endl;
+    if(k < Nm-1) evec[k+1] = w;
+
+    std::cout<<GridLogIRL << "alpha[" << k << "] = " << zalph << " beta[" << k << "] = "<<beta<<std::endl;
+    if ( beta < tiny ) 
+      std::cout<<GridLogIRL << " beta is tiny "<<beta<<std::endl;
   }
 
   void diagonalize_Eigen(std::vector<RealD>& lmd, std::vector<RealD>& lme, 
@@ -404,11 +622,11 @@ private:
       }
     }
   }
-  ///////////////////////////////////////////////////////////////////////////
-  // File could end here if settle on Eigen ???
-  ///////////////////////////////////////////////////////////////////////////
 
-  void qr_decomp(std::vector<RealD>& lmd,   // Nm 
+  ///////////////////////////////////////////////////////////////////////////
+  // File could end here if settle on Eigen ??? !!!
+  ///////////////////////////////////////////////////////////////////////////
+  void QR_decomp(std::vector<RealD>& lmd,   // Nm 
 		 std::vector<RealD>& lme,   // Nm 
 		 int Nk, int Nm,            // Nk, Nm
 		 Eigen::MatrixXd& Qt,       // Nm x Nm matrix
@@ -575,17 +793,17 @@ void diagonalize_lapack(std::vector<RealD>& lmd,
 #endif
 }
 
-  void diagonalize_QR(std::vector<RealD>& lmd, std::vector<RealD>& lme, 
+void diagonalize_QR(std::vector<RealD>& lmd, std::vector<RealD>& lme, 
 		    int Nk, int Nm,   
 		    Eigen::MatrixXd & Qt,
 		    GridBase *grid)
-  {
-    int Niter = 100*Nm;
+{
+  int QRiter = 100*Nm;
   int kmin = 1;
   int kmax = Nk;
   
   // (this should be more sophisticated)
-    for(int iter=0; iter<Niter; ++iter){
+  for(int iter=0; iter<QRiter; ++iter){
     
     // determination of 2x2 leading submatrix
     RealD dsub = lmd[kmax-1]-lmd[kmax-2];
@@ -594,7 +812,7 @@ void diagonalize_lapack(std::vector<RealD>& lmd,
     // (Dsh: shift)
     
     // transformation
-      qr_decomp(lmd,lme,Nk,Nm,Qt,Dsh,kmin,kmax); // Nk, Nm
+    QR_decomp(lmd,lme,Nk,Nm,Qt,Dsh,kmin,kmax); // Nk, Nm
     
     // Convergence criterion (redef of kmin and kamx)
     for(int j=kmax-1; j>= kmin; --j){
@@ -604,7 +822,7 @@ void diagonalize_lapack(std::vector<RealD>& lmd,
 	goto continued;
       }
     }
-      Niter = iter;
+    QRiter = iter;
     return;
     
   continued:
@@ -616,10 +834,9 @@ void diagonalize_lapack(std::vector<RealD>& lmd,
       }
     }
   }
-    std::cout << GridLogError << "[QL method] Error - Too many iteration: "<<Niter<<"\n";
+  std::cout << GridLogError << "[QL method] Error - Too many iteration: "<<QRiter<<"\n";
   abort();
-  }
-
- };
+}
+};
 }
 #endif

lib/algorithms/iterative/LocalCoherenceLanczos.h

@@ -0,0 +1,352 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/LocalCoherenceLanczos.h
+
+    Copyright (C) 2015
+
+Author: Christoph Lehner <clehner@bnl.gov>
+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 */
+#ifndef GRID_LOCAL_COHERENCE_IRL_H
+#define GRID_LOCAL_COHERENCE_IRL_H
+namespace Grid { 
+struct LanczosParams : Serializable {
+ public:
+  GRID_SERIALIZABLE_CLASS_MEMBERS(LanczosParams,
+				  ChebyParams, Cheby,/*Chebyshev*/
+				  int, Nstop,    /*Vecs in Lanczos must converge Nstop < Nk < Nm*/
+				  int, Nk,       /*Vecs in Lanczos seek converge*/
+				  int, Nm,       /*Total vecs in Lanczos include restart*/
+				  RealD, resid,  /*residual*/
+ 				  int, MaxIt, 
+				  RealD, betastp,  /* ? */
+				  int, MinRes);    // Must restart
+};
+
+struct LocalCoherenceLanczosParams : Serializable {
+ public:
+  GRID_SERIALIZABLE_CLASS_MEMBERS(LocalCoherenceLanczosParams,
+				  bool, doFine,
+				  bool, doFineRead,
+				  bool, doCoarse,
+	       			  bool, doCoarseRead,
+				  LanczosParams, FineParams,
+				  LanczosParams, CoarseParams,
+				  ChebyParams,   Smoother,
+				  RealD        , coarse_relax_tol,
+				  std::vector<int>, blockSize,
+				  std::string, config,
+				  std::vector < std::complex<double>  >, omega,
+				  RealD, mass,
+				  RealD, M5);
+};
+
+// Duplicate functionality; ProjectedFunctionHermOp could be used with the trivial function
+template<class Fobj,class CComplex,int nbasis>
+class ProjectedHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > {
+public:
+  typedef iVector<CComplex,nbasis >           CoarseSiteVector;
+  typedef Lattice<CoarseSiteVector>           CoarseField;
+  typedef Lattice<CComplex>   CoarseScalar; // used for inner products on fine field
+  typedef Lattice<Fobj>          FineField;
+
+  LinearOperatorBase<FineField> &_Linop;
+  Aggregation<Fobj,CComplex,nbasis> &_Aggregate;
+
+  ProjectedHermOp(LinearOperatorBase<FineField>& linop,  Aggregation<Fobj,CComplex,nbasis> &aggregate) : 
+    _Linop(linop),
+    _Aggregate(aggregate)  {  };
+
+  void operator()(const CoarseField& in, CoarseField& out) {
+
+    GridBase *FineGrid = _Aggregate.FineGrid;
+    FineField fin(FineGrid);
+    FineField fout(FineGrid);
+
+    _Aggregate.PromoteFromSubspace(in,fin);    std::cout<<GridLogIRL<<"ProjectedHermop : Promote to fine"<<std::endl;
+    _Linop.HermOp(fin,fout);                   std::cout<<GridLogIRL<<"ProjectedHermop : HermOp (fine) "<<std::endl;
+    _Aggregate.ProjectToSubspace(out,fout);    std::cout<<GridLogIRL<<"ProjectedHermop : Project to coarse "<<std::endl;
+  }
+};
+
+template<class Fobj,class CComplex,int nbasis>
+class ProjectedFunctionHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > {
+public:
+  typedef iVector<CComplex,nbasis >           CoarseSiteVector;
+  typedef Lattice<CoarseSiteVector>           CoarseField;
+  typedef Lattice<CComplex>   CoarseScalar; // used for inner products on fine field
+  typedef Lattice<Fobj>          FineField;
+
+
+  OperatorFunction<FineField>   & _poly;
+  LinearOperatorBase<FineField> &_Linop;
+  Aggregation<Fobj,CComplex,nbasis> &_Aggregate;
+
+  ProjectedFunctionHermOp(OperatorFunction<FineField> & poly,LinearOperatorBase<FineField>& linop, 
+			  Aggregation<Fobj,CComplex,nbasis> &aggregate) : 
+    _poly(poly),
+    _Linop(linop),
+    _Aggregate(aggregate)  {  };
+
+  void operator()(const CoarseField& in, CoarseField& out) {
+
+    GridBase *FineGrid = _Aggregate.FineGrid;
+
+    FineField fin(FineGrid) ;fin.checkerboard  =_Aggregate.checkerboard;
+    FineField fout(FineGrid);fout.checkerboard =_Aggregate.checkerboard;
+    
+    _Aggregate.PromoteFromSubspace(in,fin);    std::cout<<GridLogIRL<<"ProjectedFunctionHermop : Promote to fine"<<std::endl;
+    _poly(_Linop,fin,fout);                    std::cout<<GridLogIRL<<"ProjectedFunctionHermop : Poly "<<std::endl;
+    _Aggregate.ProjectToSubspace(out,fout);    std::cout<<GridLogIRL<<"ProjectedFunctionHermop : Project to coarse "<<std::endl;
+  }
+};
+
+template<class Fobj,class CComplex,int nbasis>
+class ImplicitlyRestartedLanczosSmoothedTester  : public ImplicitlyRestartedLanczosTester<Lattice<iVector<CComplex,nbasis > > >
+{
+ public:
+  typedef iVector<CComplex,nbasis >           CoarseSiteVector;
+  typedef Lattice<CoarseSiteVector>           CoarseField;
+  typedef Lattice<CComplex>   CoarseScalar; // used for inner products on fine field
+  typedef Lattice<Fobj>          FineField;
+
+  LinearFunction<CoarseField> & _Poly;
+  OperatorFunction<FineField>   & _smoother;
+  LinearOperatorBase<FineField> &_Linop;
+  Aggregation<Fobj,CComplex,nbasis> &_Aggregate;
+  RealD                             _coarse_relax_tol;
+  ImplicitlyRestartedLanczosSmoothedTester(LinearFunction<CoarseField>   &Poly,
+					   OperatorFunction<FineField>   &smoother,
+					   LinearOperatorBase<FineField> &Linop,
+					   Aggregation<Fobj,CComplex,nbasis> &Aggregate,
+					   RealD coarse_relax_tol=5.0e3) 
+    : _smoother(smoother), _Linop(Linop),_Aggregate(Aggregate), _Poly(Poly), _coarse_relax_tol(coarse_relax_tol)  {    };
+
+  int TestConvergence(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox)
+  {
+    CoarseField v(B);
+    RealD eval_poly = eval;
+    // Apply operator
+    _Poly(B,v);
+
+    RealD vnum = real(innerProduct(B,v)); // HermOp.
+    RealD vden = norm2(B);
+    RealD vv0  = norm2(v);
+    eval   = vnum/vden;
+    v -= eval*B;
+
+    RealD vv = norm2(v) / ::pow(evalMaxApprox,2.0);
+
+    std::cout.precision(13);
+    std::cout<<GridLogIRL  << "[" << std::setw(3)<<j<<"] "
+	     <<"eval = "<<std::setw(25)<< eval << " (" << eval_poly << ")"
+	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
+	     <<std::endl;
+
+    int conv=0;
+    if( (vv<eresid*eresid) ) conv = 1;
+    return conv;
+  }
+  int ReconstructEval(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox)
+  {
+    GridBase *FineGrid = _Aggregate.FineGrid;
+
+    int checkerboard   = _Aggregate.checkerboard;
+
+    FineField fB(FineGrid);fB.checkerboard =checkerboard;
+    FineField fv(FineGrid);fv.checkerboard =checkerboard;
+
+    _Aggregate.PromoteFromSubspace(B,fv);
+    _smoother(_Linop,fv,fB); 
+
+    RealD eval_poly = eval;
+    _Linop.HermOp(fB,fv);
+
+    RealD vnum = real(innerProduct(fB,fv)); // HermOp.
+    RealD vden = norm2(fB);
+    RealD vv0  = norm2(fv);
+    eval   = vnum/vden;
+    fv -= eval*fB;
+    RealD vv = norm2(fv) / ::pow(evalMaxApprox,2.0);
+
+    std::cout.precision(13);
+    std::cout<<GridLogIRL  << "[" << std::setw(3)<<j<<"] "
+	     <<"eval = "<<std::setw(25)<< eval << " (" << eval_poly << ")"
+	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
+	     <<std::endl;
+    if ( j > nbasis ) eresid = eresid*_coarse_relax_tol;
+    if( (vv<eresid*eresid) ) return 1;
+    return 0;
+  }
+};
+
+////////////////////////////////////////////
+// Make serializable Lanczos params
+////////////////////////////////////////////
+template<class Fobj,class CComplex,int nbasis>
+class LocalCoherenceLanczos 
+{
+public:
+  typedef iVector<CComplex,nbasis >           CoarseSiteVector;
+  typedef Lattice<CComplex>                   CoarseScalar; // used for inner products on fine field
+  typedef Lattice<CoarseSiteVector>           CoarseField;
+  typedef Lattice<Fobj>                       FineField;
+
+protected:
+  GridBase *_CoarseGrid;
+  GridBase *_FineGrid;
+  int _checkerboard;
+  LinearOperatorBase<FineField>                 & _FineOp;
+  
+  // FIXME replace Aggregation with vector of fine; the code reuse is too small for
+  // the hassle and complexity of cross coupling.
+  Aggregation<Fobj,CComplex,nbasis>               _Aggregate;  
+  std::vector<RealD>                              evals_fine;
+  std::vector<RealD>                              evals_coarse; 
+  std::vector<CoarseField>                        evec_coarse;
+public:
+  LocalCoherenceLanczos(GridBase *FineGrid,
+		GridBase *CoarseGrid,
+		LinearOperatorBase<FineField> &FineOp,
+		int checkerboard) :
+    _CoarseGrid(CoarseGrid),
+    _FineGrid(FineGrid),
+    _Aggregate(CoarseGrid,FineGrid,checkerboard),
+    _FineOp(FineOp),
+    _checkerboard(checkerboard)
+  {
+    evals_fine.resize(0);
+    evals_coarse.resize(0);
+  };
+  void Orthogonalise(void ) { _Aggregate.Orthogonalise(); }
+
+  template<typename T>  static RealD normalise(T& v) 
+  {
+    RealD nn = norm2(v);
+    nn = ::sqrt(nn);
+    v = v * (1.0/nn);
+    return nn;
+  }
+
+  void fakeFine(void)
+  {
+    int Nk = nbasis;
+    _Aggregate.subspace.resize(Nk,_FineGrid);
+    _Aggregate.subspace[0]=1.0;
+    _Aggregate.subspace[0].checkerboard=_checkerboard;
+    normalise(_Aggregate.subspace[0]);
+    PlainHermOp<FineField>    Op(_FineOp);
+    for(int k=1;k<Nk;k++){
+      _Aggregate.subspace[k].checkerboard=_checkerboard;
+      Op(_Aggregate.subspace[k-1],_Aggregate.subspace[k]);
+      normalise(_Aggregate.subspace[k]);
+    }
+  }
+
+  void testFine(RealD resid) 
+  {
+    assert(evals_fine.size() == nbasis);
+    assert(_Aggregate.subspace.size() == nbasis);
+    PlainHermOp<FineField>    Op(_FineOp);
+    ImplicitlyRestartedLanczosHermOpTester<FineField> SimpleTester(Op);
+    for(int k=0;k<nbasis;k++){
+      assert(SimpleTester.ReconstructEval(k,resid,_Aggregate.subspace[k],evals_fine[k],1.0)==1);
+    }
+  }
+
+  void testCoarse(RealD resid,ChebyParams cheby_smooth,RealD relax) 
+  {
+    assert(evals_fine.size() == nbasis);
+    assert(_Aggregate.subspace.size() == nbasis);
+    //////////////////////////////////////////////////////////////////////////////////////////////////
+    // create a smoother and see if we can get a cheap convergence test and smooth inside the IRL
+    //////////////////////////////////////////////////////////////////////////////////////////////////
+    Chebyshev<FineField>                          ChebySmooth(cheby_smooth);
+    ProjectedFunctionHermOp<Fobj,CComplex,nbasis> ChebyOp (ChebySmooth,_FineOp,_Aggregate);
+    ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,_Aggregate,relax);
+
+    for(int k=0;k<evec_coarse.size();k++){
+      if ( k < nbasis ) { 
+	assert(ChebySmoothTester.ReconstructEval(k,resid,evec_coarse[k],evals_coarse[k],1.0)==1);
+      } else { 
+	assert(ChebySmoothTester.ReconstructEval(k,resid*relax,evec_coarse[k],evals_coarse[k],1.0)==1);
+      }
+    }
+  }
+
+  void calcFine(ChebyParams cheby_parms,int Nstop,int Nk,int Nm,RealD resid, 
+		RealD MaxIt, RealD betastp, int MinRes)
+  {
+    assert(nbasis<=Nm);
+    Chebyshev<FineField>      Cheby(cheby_parms);
+    FunctionHermOp<FineField> ChebyOp(Cheby,_FineOp);
+    PlainHermOp<FineField>    Op(_FineOp);
+
+    evals_fine.resize(Nm);
+    _Aggregate.subspace.resize(Nm,_FineGrid);
+
+    ImplicitlyRestartedLanczos<FineField> IRL(ChebyOp,Op,Nstop,Nk,Nm,resid,MaxIt,betastp,MinRes);
+
+    FineField src(_FineGrid); src=1.0; src.checkerboard = _checkerboard;
+
+    int Nconv;
+    IRL.calc(evals_fine,_Aggregate.subspace,src,Nconv,false);
+    
+    // Shrink down to number saved
+    assert(Nstop>=nbasis);
+    assert(Nconv>=nbasis);
+    evals_fine.resize(nbasis);
+    _Aggregate.subspace.resize(nbasis,_FineGrid);
+  }
+  void calcCoarse(ChebyParams cheby_op,ChebyParams cheby_smooth,RealD relax,
+		  int Nstop, int Nk, int Nm,RealD resid, 
+		  RealD MaxIt, RealD betastp, int MinRes)
+  {
+    Chebyshev<FineField>                          Cheby(cheby_op);
+    ProjectedHermOp<Fobj,CComplex,nbasis>         Op(_FineOp,_Aggregate);
+    ProjectedFunctionHermOp<Fobj,CComplex,nbasis> ChebyOp (Cheby,_FineOp,_Aggregate);
+    //////////////////////////////////////////////////////////////////////////////////////////////////
+    // create a smoother and see if we can get a cheap convergence test and smooth inside the IRL
+    //////////////////////////////////////////////////////////////////////////////////////////////////
+
+    Chebyshev<FineField>                                           ChebySmooth(cheby_smooth);
+    ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,_Aggregate,relax);
+
+    evals_coarse.resize(Nm);
+    evec_coarse.resize(Nm,_CoarseGrid);
+
+    CoarseField src(_CoarseGrid);     src=1.0; 
+
+    ImplicitlyRestartedLanczos<CoarseField> IRL(ChebyOp,ChebyOp,ChebySmoothTester,Nstop,Nk,Nm,resid,MaxIt,betastp,MinRes);
+    int Nconv=0;
+    IRL.calc(evals_coarse,evec_coarse,src,Nconv,false);
+    assert(Nconv>=Nstop);
+    evals_coarse.resize(Nstop);
+    evec_coarse.resize (Nstop,_CoarseGrid);
+    for (int i=0;i<Nstop;i++){
+      std::cout << i << " Coarse eval = " << evals_coarse[i]  << std::endl;
+    }
+  }
+};
+
+}
+#endif

lib/algorithms/iterative/SchurRedBlack.h

@@ -53,16 +53,119 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
    *     M psi = eta
    ***********************
    *Odd
-   * i)   (D_oo)^{\dag} D_oo psi_o = (D_oo)^dag L^{-1}  eta_o
+   * i)                 D_oo psi_o =  L^{-1}  eta_o
    *                        eta_o' = (D_oo)^dag (eta_o - Moe Mee^{-1} eta_e)
+   *
+   * Wilson:
+   *      (D_oo)^{\dag} D_oo psi_o = (D_oo)^dag L^{-1}  eta_o
+   * Stag:
+   *      D_oo psi_o = L^{-1}  eta =    (eta_o - Moe Mee^{-1} eta_e)
+   *
+   * L^-1 eta_o= (1              0 ) (e
+   *             (-MoeMee^{-1}   1 )   
+   *
    *Even
    * ii)  Mee psi_e + Meo psi_o = src_e
    *
    *   => sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
    *
+   * 
+   * TODO: Other options:
+   * 
+   * a) change checkerboards for Schur e<->o
+   *
+   * Left precon by Moo^-1
+   * b) Doo^{dag} M_oo^-dag Moo^-1 Doo psi_0 =  (D_oo)^dag M_oo^-dag Moo^-1 L^{-1}  eta_o
+   *                              eta_o'     = (D_oo)^dag  M_oo^-dag Moo^-1 (eta_o - Moe Mee^{-1} eta_e)
+   *
+   * Right precon by Moo^-1
+   * c) M_oo^-dag Doo^{dag} Doo Moo^-1 phi_0 = M_oo^-dag (D_oo)^dag L^{-1}  eta_o
+   *                              eta_o'     = M_oo^-dag (D_oo)^dag (eta_o - Moe Mee^{-1} eta_e)
+   *                              psi_o = M_oo^-1 phi_o
+   * TODO: Deflation 
    */
 namespace Grid {
 
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Take a matrix and form a Red Black solver calling a Herm solver
+  // Use of RB info prevents making SchurRedBlackSolve conform to standard interface
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////
+
+  template<class Field> class SchurRedBlackStaggeredSolve {
+  private:
+    OperatorFunction<Field> & _HermitianRBSolver;
+    int CBfactorise;
+  public:
+
+    /////////////////////////////////////////////////////
+    // Wrap the usual normal equations Schur trick
+    /////////////////////////////////////////////////////
+  SchurRedBlackStaggeredSolve(OperatorFunction<Field> &HermitianRBSolver)  :
+     _HermitianRBSolver(HermitianRBSolver) 
+    { 
+      CBfactorise=0;
+    };
+
+    template<class Matrix>
+      void operator() (Matrix & _Matrix,const Field &in, Field &out){
+
+      // FIXME CGdiagonalMee not implemented virtual function
+      // FIXME use CBfactorise to control schur decomp
+      GridBase *grid = _Matrix.RedBlackGrid();
+      GridBase *fgrid= _Matrix.Grid();
+
+      SchurStaggeredOperator<Matrix,Field> _HermOpEO(_Matrix);
+ 
+      Field src_e(grid);
+      Field src_o(grid);
+      Field sol_e(grid);
+      Field sol_o(grid);
+      Field   tmp(grid);
+      Field  Mtmp(grid);
+      Field resid(fgrid);
+
+      pickCheckerboard(Even,src_e,in);
+      pickCheckerboard(Odd ,src_o,in);
+      pickCheckerboard(Even,sol_e,out);
+      pickCheckerboard(Odd ,sol_o,out);
+    
+      /////////////////////////////////////////////////////
+      // src_o = (source_o - Moe MeeInv source_e)
+      /////////////////////////////////////////////////////
+      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.checkerboard ==Even);
+      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.checkerboard ==Odd);     
+      tmp=src_o-Mtmp;                  assert(  tmp.checkerboard ==Odd);     
+
+      src_o = tmp;     assert(src_o.checkerboard ==Odd);
+      //  _Matrix.Mooee(tmp,src_o); // Extra factor of "m" in source
+
+      //////////////////////////////////////////////////////////////
+      // Call the red-black solver
+      //////////////////////////////////////////////////////////////
+      std::cout<<GridLogMessage << "SchurRedBlackStaggeredSolver calling the Mpc solver" <<std::endl;
+      _HermitianRBSolver(_HermOpEO,src_o,sol_o);  assert(sol_o.checkerboard==Odd);
+
+      ///////////////////////////////////////////////////
+      // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
+      ///////////////////////////////////////////////////
+      _Matrix.Meooe(sol_o,tmp);        assert(  tmp.checkerboard   ==Even);
+      src_e = src_e-tmp;               assert(  src_e.checkerboard ==Even);
+      _Matrix.MooeeInv(src_e,sol_e);   assert(  sol_e.checkerboard ==Even);
+     
+      setCheckerboard(out,sol_e); assert(  sol_e.checkerboard ==Even);
+      setCheckerboard(out,sol_o); assert(  sol_o.checkerboard ==Odd );
+
+      // Verify the unprec residual
+      _Matrix.M(out,resid); 
+      resid = resid-in;
+      RealD ns = norm2(in);
+      RealD nr = norm2(resid);
+
+      std::cout<<GridLogMessage << "SchurRedBlackStaggered solver true unprec resid "<< std::sqrt(nr/ns) <<" nr "<< nr <<" ns "<<ns << std::endl;
+    }     
+  };
+  template<class Field> using SchurRedBlackStagSolve = SchurRedBlackStaggeredSolve<Field>;
+
   ///////////////////////////////////////////////////////////////////////////////////////////////////////
   // Take a matrix and form a Red Black solver calling a Herm solver
   // Use of RB info prevents making SchurRedBlackSolve conform to standard interface
@@ -76,12 +179,10 @@ namespace Grid {
     /////////////////////////////////////////////////////
     // Wrap the usual normal equations Schur trick
     /////////////////////////////////////////////////////
-  SchurRedBlackDiagMooeeSolve(OperatorFunction<Field> &HermitianRBSolver)  :
-     _HermitianRBSolver(HermitianRBSolver) 
+  SchurRedBlackDiagMooeeSolve(OperatorFunction<Field> &HermitianRBSolver,int cb=0)  :  _HermitianRBSolver(HermitianRBSolver) 
   { 
-      CBfactorise=0;
+    CBfactorise=cb;
   };
-
     template<class Matrix>
       void operator() (Matrix & _Matrix,const Field &in, Field &out){
 
@@ -141,5 +242,166 @@ namespace Grid {
     }     
   };
 
+
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Take a matrix and form a Red Black solver calling a Herm solver
+  // Use of RB info prevents making SchurRedBlackSolve conform to standard interface
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////
+  template<class Field> class SchurRedBlackDiagTwoSolve {
+  private:
+    OperatorFunction<Field> & _HermitianRBSolver;
+    int CBfactorise;
+  public:
+
+    /////////////////////////////////////////////////////
+    // Wrap the usual normal equations Schur trick
+    /////////////////////////////////////////////////////
+  SchurRedBlackDiagTwoSolve(OperatorFunction<Field> &HermitianRBSolver)  :
+     _HermitianRBSolver(HermitianRBSolver) 
+    { 
+      CBfactorise=0;
+    };
+
+    template<class Matrix>
+      void operator() (Matrix & _Matrix,const Field &in, Field &out){
+
+      // FIXME CGdiagonalMee not implemented virtual function
+      // FIXME use CBfactorise to control schur decomp
+      GridBase *grid = _Matrix.RedBlackGrid();
+      GridBase *fgrid= _Matrix.Grid();
+
+      SchurDiagTwoOperator<Matrix,Field> _HermOpEO(_Matrix);
+ 
+      Field src_e(grid);
+      Field src_o(grid);
+      Field sol_e(grid);
+      Field sol_o(grid);
+      Field   tmp(grid);
+      Field  Mtmp(grid);
+      Field resid(fgrid);
+
+      pickCheckerboard(Even,src_e,in);
+      pickCheckerboard(Odd ,src_o,in);
+      pickCheckerboard(Even,sol_e,out);
+      pickCheckerboard(Odd ,sol_o,out);
+    
+      /////////////////////////////////////////////////////
+      // src_o = Mdag * (source_o - Moe MeeInv source_e)
+      /////////////////////////////////////////////////////
+      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.checkerboard ==Even);
+      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.checkerboard ==Odd);     
+      tmp=src_o-Mtmp;                  assert(  tmp.checkerboard ==Odd);     
+
+      // get the right MpcDag
+      _HermOpEO.MpcDag(tmp,src_o);     assert(src_o.checkerboard ==Odd);       
+
+      //////////////////////////////////////////////////////////////
+      // Call the red-black solver
+      //////////////////////////////////////////////////////////////
+      std::cout<<GridLogMessage << "SchurRedBlack solver calling the MpcDagMp solver" <<std::endl;
+//      _HermitianRBSolver(_HermOpEO,src_o,sol_o);  assert(sol_o.checkerboard==Odd);
+      _HermitianRBSolver(_HermOpEO,src_o,tmp);  assert(tmp.checkerboard==Odd);
+      _Matrix.MooeeInv(tmp,sol_o);        assert(  sol_o.checkerboard   ==Odd);
+
+      ///////////////////////////////////////////////////
+      // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
+      ///////////////////////////////////////////////////
+      _Matrix.Meooe(sol_o,tmp);        assert(  tmp.checkerboard   ==Even);
+      src_e = src_e-tmp;               assert(  src_e.checkerboard ==Even);
+      _Matrix.MooeeInv(src_e,sol_e);   assert(  sol_e.checkerboard ==Even);
+     
+      setCheckerboard(out,sol_e); assert(  sol_e.checkerboard ==Even);
+      setCheckerboard(out,sol_o); assert(  sol_o.checkerboard ==Odd );
+
+      // Verify the unprec residual
+      _Matrix.M(out,resid); 
+      resid = resid-in;
+      RealD ns = norm2(in);
+      RealD nr = norm2(resid);
+
+      std::cout<<GridLogMessage << "SchurRedBlackDiagTwo solver true unprec resid "<< std::sqrt(nr/ns) <<" nr "<< nr <<" ns "<<ns << std::endl;
+    }     
+  };
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Take a matrix and form a Red Black solver calling a Herm solver
+  // Use of RB info prevents making SchurRedBlackSolve conform to standard interface
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////
+  template<class Field> class SchurRedBlackDiagTwoMixed {
+  private:
+    LinearFunction<Field> & _HermitianRBSolver;
+    int CBfactorise;
+  public:
+
+    /////////////////////////////////////////////////////
+    // Wrap the usual normal equations Schur trick
+    /////////////////////////////////////////////////////
+  SchurRedBlackDiagTwoMixed(LinearFunction<Field> &HermitianRBSolver)  :
+     _HermitianRBSolver(HermitianRBSolver) 
+    { 
+      CBfactorise=0;
+    };
+
+    template<class Matrix>
+      void operator() (Matrix & _Matrix,const Field &in, Field &out){
+
+      // FIXME CGdiagonalMee not implemented virtual function
+      // FIXME use CBfactorise to control schur decomp
+      GridBase *grid = _Matrix.RedBlackGrid();
+      GridBase *fgrid= _Matrix.Grid();
+
+      SchurDiagTwoOperator<Matrix,Field> _HermOpEO(_Matrix);
+ 
+      Field src_e(grid);
+      Field src_o(grid);
+      Field sol_e(grid);
+      Field sol_o(grid);
+      Field   tmp(grid);
+      Field  Mtmp(grid);
+      Field resid(fgrid);
+
+      pickCheckerboard(Even,src_e,in);
+      pickCheckerboard(Odd ,src_o,in);
+      pickCheckerboard(Even,sol_e,out);
+      pickCheckerboard(Odd ,sol_o,out);
+    
+      /////////////////////////////////////////////////////
+      // src_o = Mdag * (source_o - Moe MeeInv source_e)
+      /////////////////////////////////////////////////////
+      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.checkerboard ==Even);
+      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.checkerboard ==Odd);     
+      tmp=src_o-Mtmp;                  assert(  tmp.checkerboard ==Odd);     
+
+      // get the right MpcDag
+      _HermOpEO.MpcDag(tmp,src_o);     assert(src_o.checkerboard ==Odd);       
+
+      //////////////////////////////////////////////////////////////
+      // Call the red-black solver
+      //////////////////////////////////////////////////////////////
+      std::cout<<GridLogMessage << "SchurRedBlack solver calling the MpcDagMp solver" <<std::endl;
+//      _HermitianRBSolver(_HermOpEO,src_o,sol_o);  assert(sol_o.checkerboard==Odd);
+//      _HermitianRBSolver(_HermOpEO,src_o,tmp);  assert(tmp.checkerboard==Odd);
+      _HermitianRBSolver(src_o,tmp);  assert(tmp.checkerboard==Odd);
+      _Matrix.MooeeInv(tmp,sol_o);        assert(  sol_o.checkerboard   ==Odd);
+
+      ///////////////////////////////////////////////////
+      // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
+      ///////////////////////////////////////////////////
+      _Matrix.Meooe(sol_o,tmp);        assert(  tmp.checkerboard   ==Even);
+      src_e = src_e-tmp;               assert(  src_e.checkerboard ==Even);
+      _Matrix.MooeeInv(src_e,sol_e);   assert(  sol_e.checkerboard ==Even);
+     
+      setCheckerboard(out,sol_e); assert(  sol_e.checkerboard ==Even);
+      setCheckerboard(out,sol_o); assert(  sol_o.checkerboard ==Odd );
+
+      // Verify the unprec residual
+      _Matrix.M(out,resid); 
+      resid = resid-in;
+      RealD ns = norm2(in);
+      RealD nr = norm2(resid);
+
+      std::cout<<GridLogMessage << "SchurRedBlackDiagTwo solver true unprec resid "<< std::sqrt(nr/ns) <<" nr "<< nr <<" ns "<<ns << std::endl;
+    }     
+  };
+
 }
 #endif

lib/allocator/AlignedAllocator.cc

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

lib/allocator/AlignedAllocator.h

@@ -64,6 +64,8 @@ namespace Grid {
 
   };
 
+  void check_huge_pages(void *Buf,uint64_t BYTES);
+
 ////////////////////////////////////////////////////////////////////
 // A lattice of something, but assume the something is SIMDized.
 ////////////////////////////////////////////////////////////////////
@@ -92,18 +94,34 @@ public:
     size_type bytes = __n*sizeof(_Tp);
 
     _Tp *ptr = (_Tp *) PointerCache::Lookup(bytes);
+    //    if ( ptr != NULL ) 
+    //      std::cout << "alignedAllocator "<<__n << " cache hit "<< std::hex << ptr <<std::dec <<std::endl;
 
+    //////////////////
+    // Hack 2MB align; could make option probably doesn't need configurability
+    //////////////////
+//define GRID_ALLOC_ALIGN (128)
+#define GRID_ALLOC_ALIGN (2*1024*1024)
 #ifdef HAVE_MM_MALLOC_H
-    if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) _mm_malloc(bytes,128);
+    if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) _mm_malloc(bytes,GRID_ALLOC_ALIGN);
 #else
-    if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) memalign(128,bytes);
+    if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) memalign(GRID_ALLOC_ALIGN,bytes);
 #endif
-
+    //    std::cout << "alignedAllocator " << std::hex << ptr <<std::dec <<std::endl;
+    // First touch optimise in threaded loop
+    uint8_t *cp = (uint8_t *)ptr;
+#ifdef GRID_OMP
+#pragma omp parallel for
+#endif
+    for(size_type n=0;n<bytes;n+=4096){
+      cp[n]=0;
+    }
     return ptr;
   }
 
   void deallocate(pointer __p, size_type __n) { 
     size_type bytes = __n * sizeof(_Tp);
+
     pointer __freeme = (pointer)PointerCache::Insert((void *)__p,bytes);
 
 #ifdef HAVE_MM_MALLOC_H
@@ -182,10 +200,19 @@ public:
   pointer allocate(size_type __n, const void* _p= 0) 
   {
 #ifdef HAVE_MM_MALLOC_H
-    _Tp * ptr = (_Tp *) _mm_malloc(__n*sizeof(_Tp),128);
+    _Tp * ptr = (_Tp *) _mm_malloc(__n*sizeof(_Tp),GRID_ALLOC_ALIGN);
 #else
-    _Tp * ptr = (_Tp *) memalign(128,__n*sizeof(_Tp));
+    _Tp * ptr = (_Tp *) memalign(GRID_ALLOC_ALIGN,__n*sizeof(_Tp));
 #endif
+    size_type bytes = __n*sizeof(_Tp);
+    uint8_t *cp = (uint8_t *)ptr;
+    if ( ptr ) { 
+    // One touch per 4k page, static OMP loop to catch same loop order
+#pragma omp parallel for schedule(static)
+      for(size_type n=0;n<bytes;n+=4096){
+	cp[n]=0;
+      }
+    }
     return ptr;
   }
   void deallocate(pointer __p, size_type) { 

lib/cartesian/Cartesian_base.h

@@ -44,11 +44,20 @@ namespace Grid{
   class GridBase : public CartesianCommunicator , public GridThread {
 
 public:
-
+    int dummy;
     // Give Lattice access
     template<class object> friend class Lattice;
 
     GridBase(const std::vector<int> & processor_grid) : CartesianCommunicator(processor_grid) {};
+    GridBase(const std::vector<int> & processor_grid,
+	     const CartesianCommunicator &parent,
+	     int &split_rank) 
+      : CartesianCommunicator(processor_grid,parent,split_rank) {};
+    GridBase(const std::vector<int> & processor_grid,
+	     const CartesianCommunicator &parent) 
+      : CartesianCommunicator(processor_grid,parent,dummy) {};
+
+    virtual ~GridBase() = default;
 
     // Physics Grid information.
     std::vector<int> _simd_layout;// Which dimensions get relayed out over simd lanes.
@@ -185,17 +194,18 @@ public:
     ////////////////////////////////////////////////////////////////
 
     void show_decomposition(){
-      std::cout << GridLogMessage << "Full Dimensions    : " << _fdimensions << std::endl;
-      std::cout << GridLogMessage << "Global Dimensions  : " << _gdimensions << std::endl;
-      std::cout << GridLogMessage << "Local Dimensions   : " << _ldimensions << std::endl;
-      std::cout << GridLogMessage << "Reduced Dimensions : " << _rdimensions << std::endl;
-      std::cout << GridLogMessage << "Outer strides      : " << _ostride << std::endl;
-      std::cout << GridLogMessage << "Inner strides      : " << _istride << std::endl;
-      std::cout << GridLogMessage << "iSites             : " << _isites << std::endl;
-      std::cout << GridLogMessage << "oSites             : " << _osites << std::endl;
-      std::cout << GridLogMessage << "lSites             : " << lSites() << std::endl;        
-      std::cout << GridLogMessage << "gSites             : " << gSites() << std::endl;
-      std::cout << GridLogMessage << "Nd                 : " << _ndimension << std::endl;             
+      std::cout << GridLogMessage << "\tFull Dimensions    : " << _fdimensions << std::endl;
+      std::cout << GridLogMessage << "\tSIMD layout        : " << _simd_layout << std::endl;
+      std::cout << GridLogMessage << "\tGlobal Dimensions  : " << _gdimensions << std::endl;
+      std::cout << GridLogMessage << "\tLocal Dimensions   : " << _ldimensions << std::endl;
+      std::cout << GridLogMessage << "\tReduced Dimensions : " << _rdimensions << std::endl;
+      std::cout << GridLogMessage << "\tOuter strides      : " << _ostride << std::endl;
+      std::cout << GridLogMessage << "\tInner strides      : " << _istride << std::endl;
+      std::cout << GridLogMessage << "\tiSites             : " << _isites << std::endl;
+      std::cout << GridLogMessage << "\toSites             : " << _osites << std::endl;
+      std::cout << GridLogMessage << "\tlSites             : " << lSites() << std::endl;        
+      std::cout << GridLogMessage << "\tgSites             : " << gSites() << std::endl;
+      std::cout << GridLogMessage << "\tNd                 : " << _ndimension << std::endl;             
     } 
 
     ////////////////////////////////////////////////////////////////
@@ -209,9 +219,6 @@ public:
       assert(lidx<lSites());
       Lexicographic::CoorFromIndex(lcoor,lidx,_ldimensions);
     }
-
-
-
     void GlobalCoorToGlobalIndex(const std::vector<int> & gcoor,int & gidx){
       gidx=0;
       int mult=1;

lib/cartesian/Cartesian_full.h

@@ -38,7 +38,7 @@ namespace Grid{
 class GridCartesian: public GridBase {
 
 public:
-
+    int dummy;
     virtual int  CheckerBoardFromOindexTable (int Oindex) {
       return 0;
     }
@@ -61,10 +61,38 @@ public:
     virtual int CheckerBoardShift(int source_cb,int dim,int shift, int osite){
       return shift;
     }
+    /////////////////////////////////////////////////////////////////////////
+    // Constructor takes a parent grid and possibly subdivides communicator.
+    /////////////////////////////////////////////////////////////////////////
     GridCartesian(const std::vector<int> &dimensions,
 		  const std::vector<int> &simd_layout,
-		  const std::vector<int> &processor_grid
-		  ) : GridBase(processor_grid)
+		  const std::vector<int> &processor_grid,
+		  const GridCartesian &parent) : GridBase(processor_grid,parent,dummy)
+    {
+      Init(dimensions,simd_layout,processor_grid);
+    }
+    GridCartesian(const std::vector<int> &dimensions,
+		  const std::vector<int> &simd_layout,
+		  const std::vector<int> &processor_grid,
+		  const GridCartesian &parent,int &split_rank) : GridBase(processor_grid,parent,split_rank)
+    {
+      Init(dimensions,simd_layout,processor_grid);
+    }
+    /////////////////////////////////////////////////////////////////////////
+    // Construct from comm world
+    /////////////////////////////////////////////////////////////////////////
+    GridCartesian(const std::vector<int> &dimensions,
+		  const std::vector<int> &simd_layout,
+		  const std::vector<int> &processor_grid) : GridBase(processor_grid)
+    {
+      Init(dimensions,simd_layout,processor_grid);
+    }
+
+    virtual ~GridCartesian() = default;
+
+    void Init(const std::vector<int> &dimensions,
+	      const std::vector<int> &simd_layout,
+	      const std::vector<int> &processor_grid)
     {
       ///////////////////////
       // Grid information
@@ -84,30 +112,36 @@ public:
 
       _fsites = _gsites = _osites = _isites = 1;
 
-        for(int d=0;d<_ndimension;d++){
+      for (int d = 0; d < _ndimension; d++)
+      {
         _fdimensions[d] = dimensions[d];   // Global dimensions
         _gdimensions[d] = _fdimensions[d]; // Global dimensions
         _simd_layout[d] = simd_layout[d];
         _fsites = _fsites * _fdimensions[d];
         _gsites = _gsites * _gdimensions[d];
 
-	  //FIXME check for exact division
-
         // Use a reduced simd grid
-	  _ldimensions[d]= _gdimensions[d]/_processors[d];  //local dimensions
-	  _rdimensions[d]= _ldimensions[d]/_simd_layout[d]; //overdecomposition
-	  _lstart[d]     = _processor_coor[d]*_ldimensions[d];
-	  _lend[d]       = _processor_coor[d]*_ldimensions[d]+_ldimensions[d]-1;
+        _ldimensions[d] = _gdimensions[d] / _processors[d]; //local dimensions
+        assert(_ldimensions[d] * _processors[d] == _gdimensions[d]);
+
+        _rdimensions[d] = _ldimensions[d] / _simd_layout[d]; //overdecomposition
+        assert(_rdimensions[d] * _simd_layout[d] == _ldimensions[d]);
+
+        _lstart[d] = _processor_coor[d] * _ldimensions[d];
+        _lend[d] = _processor_coor[d] * _ldimensions[d] + _ldimensions[d] - 1;
         _osites *= _rdimensions[d];
         _isites *= _simd_layout[d];
 
         // Addressing support
-	  if ( d==0 ) {
+        if (d == 0)
+        {
           _ostride[d] = 1;
           _istride[d] = 1;
-	  } else {
-	    _ostride[d] = _ostride[d-1]*_rdimensions[d-1];
-	    _istride[d] = _istride[d-1]*_simd_layout[d-1];
+        }
+        else
+        {
+          _ostride[d] = _ostride[d - 1] * _rdimensions[d - 1];
+          _istride[d] = _istride[d - 1] * _simd_layout[d - 1];
         }
       }
 
@@ -118,21 +152,20 @@ public:
       _slice_stride.resize(_ndimension);
       _slice_nblock.resize(_ndimension);
 
-        int block =1;
-        int nblock=1;
-        for(int d=0;d<_ndimension;d++) nblock*=_rdimensions[d];
+      int block = 1;
+      int nblock = 1;
+      for (int d = 0; d < _ndimension; d++)
+        nblock *= _rdimensions[d];
 
-        for(int d=0;d<_ndimension;d++){
-            nblock/=_rdimensions[d];
-            _slice_block[d] =block;
-            _slice_stride[d]=_ostride[d]*_rdimensions[d];
-            _slice_nblock[d]=nblock;
-            block = block*_rdimensions[d];
+      for (int d = 0; d < _ndimension; d++)
+      {
+        nblock /= _rdimensions[d];
+        _slice_block[d] = block;
+        _slice_stride[d] = _ostride[d] * _rdimensions[d];
+        _slice_nblock[d] = nblock;
+        block = block * _rdimensions[d];
       }
-
     };
 };
-
-
 }
 #endif

lib/cartesian/Cartesian_red_black.h

@@ -112,24 +112,59 @@ public:
       }
     };
 
-    GridRedBlackCartesian(const GridBase *base) : GridRedBlackCartesian(base->_fdimensions,base->_simd_layout,base->_processors)  {};
+    ////////////////////////////////////////////////////////////
+    // Create Redblack from original grid; require full grid pointer ?
+    ////////////////////////////////////////////////////////////
+    GridRedBlackCartesian(const GridBase *base) : GridBase(base->_processors,*base)
+    {
+      int dims = base->_ndimension;
+      std::vector<int> checker_dim_mask(dims,1);
+      int checker_dim = 0;
+      Init(base->_fdimensions,base->_simd_layout,base->_processors,checker_dim_mask,checker_dim);
+    };
 
-    GridRedBlackCartesian(const std::vector<int> &dimensions,
+    ////////////////////////////////////////////////////////////
+    // Create redblack from original grid, with non-trivial checker dim mask
+    ////////////////////////////////////////////////////////////
+    GridRedBlackCartesian(const GridBase *base,
+			  const std::vector<int> &checker_dim_mask,
+			  int checker_dim
+			  ) :  GridBase(base->_processors,*base) 
+    {
+      Init(base->_fdimensions,base->_simd_layout,base->_processors,checker_dim_mask,checker_dim)  ;
+    }
+
+    virtual ~GridRedBlackCartesian() = default;
+#if 0
+    ////////////////////////////////////////////////////////////
+    // Create redblack grid ;; deprecate these. Should not
+    // need direct creation of redblack without a full grid to base on
+    ////////////////////////////////////////////////////////////
+    GridRedBlackCartesian(const GridBase *base,
+			  const std::vector<int> &dimensions,
 			  const std::vector<int> &simd_layout,
 			  const std::vector<int> &processor_grid,
 			  const std::vector<int> &checker_dim_mask,
 			  int checker_dim
-			  ) :  GridBase(processor_grid) 
+			  ) :  GridBase(processor_grid,*base) 
     {
       Init(dimensions,simd_layout,processor_grid,checker_dim_mask,checker_dim);
     }
-    GridRedBlackCartesian(const std::vector<int> &dimensions,
+
+    ////////////////////////////////////////////////////////////
+    // Create redblack grid
+    ////////////////////////////////////////////////////////////
+    GridRedBlackCartesian(const GridBase *base,
+			  const std::vector<int> &dimensions,
 			  const std::vector<int> &simd_layout,
-			  const std::vector<int> &processor_grid) : GridBase(processor_grid) 
+			  const std::vector<int> &processor_grid) : GridBase(processor_grid,*base) 
     {
       std::vector<int> checker_dim_mask(dimensions.size(),1);
-      Init(dimensions,simd_layout,processor_grid,checker_dim_mask,0);
+      int checker_dim = 0;
+      Init(dimensions,simd_layout,processor_grid,checker_dim_mask,checker_dim);
     }
+#endif
+
     void Init(const std::vector<int> &dimensions,
               const std::vector<int> &simd_layout,
               const std::vector<int> &processor_grid,
@@ -140,11 +175,11 @@ public:
       // Grid information
       ///////////////////////
       _checker_dim = checker_dim;
-      assert(checker_dim_mask[checker_dim]==1);
+      assert(checker_dim_mask[checker_dim] == 1);
       _ndimension = dimensions.size();
-      assert(checker_dim_mask.size()==_ndimension);
-      assert(processor_grid.size()==_ndimension);
-      assert(simd_layout.size()==_ndimension);
+      assert(checker_dim_mask.size() == _ndimension);
+      assert(processor_grid.size() == _ndimension);
+      assert(simd_layout.size() == _ndimension);
 
       _fdimensions.resize(_ndimension);
       _gdimensions.resize(_ndimension);
@@ -159,31 +194,39 @@ public:
 
       _fsites = _gsites = _osites = _isites = 1;
 
-      _checker_dim_mask=checker_dim_mask;
+      _checker_dim_mask = checker_dim_mask;
 
-      for(int d=0;d<_ndimension;d++){
+      for (int d = 0; d < _ndimension; d++)
+      {
         _fdimensions[d] = dimensions[d];
         _gdimensions[d] = _fdimensions[d];
         _fsites = _fsites * _fdimensions[d];
         _gsites = _gsites * _gdimensions[d];
 
-	if (d==_checker_dim) {
-	  _gdimensions[d] = _gdimensions[d]/2; // Remove a checkerboard
+        if (d == _checker_dim)
+        {
+          assert((_gdimensions[d] & 0x1) == 0);
+          _gdimensions[d] = _gdimensions[d] / 2; // Remove a checkerboard
+	  _gsites /= 2;
         }
-	_ldimensions[d] = _gdimensions[d]/_processors[d];
-	_lstart[d]     = _processor_coor[d]*_ldimensions[d];
-	_lend[d]       = _processor_coor[d]*_ldimensions[d]+_ldimensions[d]-1;
+        _ldimensions[d] = _gdimensions[d] / _processors[d];
+        assert(_ldimensions[d] * _processors[d] == _gdimensions[d]);
+        _lstart[d] = _processor_coor[d] * _ldimensions[d];
+        _lend[d] = _processor_coor[d] * _ldimensions[d] + _ldimensions[d] - 1;
 
         // Use a reduced simd grid
         _simd_layout[d] = simd_layout[d];
-	_rdimensions[d]= _ldimensions[d]/_simd_layout[d];
-	assert(_rdimensions[d]>0);
+        _rdimensions[d] = _ldimensions[d] / _simd_layout[d]; // this is not checking if this is integer
+        assert(_rdimensions[d] * _simd_layout[d] == _ldimensions[d]);
+        assert(_rdimensions[d] > 0);
 
         // all elements of a simd vector must have same checkerboard.
         // If Ls vectorised, this must still be the case; e.g. dwf rb5d
-	if ( _simd_layout[d]>1 ) {
-	  if ( checker_dim_mask[d] ) { 
-	    assert( (_rdimensions[d]&0x1) == 0 );
+        if (_simd_layout[d] > 1)
+        {
+          if (checker_dim_mask[d])
+          {
+            assert((_rdimensions[d] & 0x1) == 0);
           }
         }
 
@@ -191,15 +234,16 @@ public:
         _isites *= _simd_layout[d];
 
         // Addressing support
-	if ( d==0 ) {
+        if (d == 0)
+        {
           _ostride[d] = 1;
           _istride[d] = 1;
-	} else {
-	  _ostride[d] = _ostride[d-1]*_rdimensions[d-1];
-	  _istride[d] = _istride[d-1]*_simd_layout[d-1];
         }
-
-
+        else
+        {
+          _ostride[d] = _ostride[d - 1] * _rdimensions[d - 1];
+          _istride[d] = _istride[d - 1] * _simd_layout[d - 1];
+        }
       }
 
       ////////////////////////////////////////////////////////////////////////////////////////////
@@ -209,40 +253,48 @@ public:
       _slice_stride.resize(_ndimension);
       _slice_nblock.resize(_ndimension);
 
-      int block =1;
-      int nblock=1;
-      for(int d=0;d<_ndimension;d++) nblock*=_rdimensions[d];
+      int block = 1;
+      int nblock = 1;
+      for (int d = 0; d < _ndimension; d++)
+        nblock *= _rdimensions[d];
 
-      for(int d=0;d<_ndimension;d++){
-	nblock/=_rdimensions[d];
-	_slice_block[d] =block;
-	_slice_stride[d]=_ostride[d]*_rdimensions[d];
-	_slice_nblock[d]=nblock;
-	block = block*_rdimensions[d];
+      for (int d = 0; d < _ndimension; d++)
+      {
+        nblock /= _rdimensions[d];
+        _slice_block[d] = block;
+        _slice_stride[d] = _ostride[d] * _rdimensions[d];
+        _slice_nblock[d] = nblock;
+        block = block * _rdimensions[d];
       }
 
       ////////////////////////////////////////////////
       // Create a checkerboard lookup table
       ////////////////////////////////////////////////
       int rvol = 1;
-      for(int d=0;d<_ndimension;d++){
-	rvol=rvol * _rdimensions[d];
+      for (int d = 0; d < _ndimension; d++)
+      {
+        rvol = rvol * _rdimensions[d];
       }
       _checker_board.resize(rvol);
-      for(int osite=0;osite<_osites;osite++){
-	_checker_board[osite] = CheckerBoardFromOindex (osite);
+      for (int osite = 0; osite < _osites; osite++)
+      {
+        _checker_board[osite] = CheckerBoardFromOindex(osite);
       }
-      
     };
-protected:
+
+  protected:
     virtual int oIndex(std::vector<int> &coor)
     {
-      int idx=0;
-      for(int d=0;d<_ndimension;d++) {
-	if( d==_checker_dim ) {
-	  idx+=_ostride[d]*((coor[d]/2)%_rdimensions[d]);
-	} else {
-	  idx+=_ostride[d]*(coor[d]%_rdimensions[d]);
+      int idx = 0;
+      for (int d = 0; d < _ndimension; d++)
+      {
+        if (d == _checker_dim)
+        {
+          idx += _ostride[d] * ((coor[d] / 2) % _rdimensions[d]);
+        }
+        else
+        {
+          idx += _ostride[d] * (coor[d] % _rdimensions[d]);
         }
       }
       return idx;
@@ -250,17 +302,20 @@ protected:
 
     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]);
+      int idx = 0;
+      for (int d = 0; d < _ndimension; d++)
+      {
+        if (d == _checker_dim)
+        {
+          idx += _istride[d] * (lcoor[d] / (2 * _rdimensions[d]));
+        }
+        else
+        {
+          idx += _istride[d] * (lcoor[d] / _rdimensions[d]);
         }
       }
       return idx;
     }
 };
-
 }
 #endif

lib/communicator/Communicator_base.cc

@@ -26,6 +26,10 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     *************************************************************************************/
     /*  END LEGAL */
 #include <Grid/GridCore.h>
+#include <fcntl.h>
+#include <unistd.h>
+#include <limits.h>
+#include <sys/mman.h>
 
 namespace Grid {
 
@@ -33,8 +37,11 @@ namespace Grid {
 // Info that is setup once and indept of cartesian layout
 ///////////////////////////////////////////////////////////////
 void *              CartesianCommunicator::ShmCommBuf;
-uint64_t            CartesianCommunicator::MAX_MPI_SHM_BYTES   = 128*1024*1024; 
-CartesianCommunicator::CommunicatorPolicy_t  CartesianCommunicator::CommunicatorPolicy= CartesianCommunicator::CommunicatorPolicyConcurrent;
+uint64_t            CartesianCommunicator::MAX_MPI_SHM_BYTES   = 1024LL*1024LL*1024LL; 
+CartesianCommunicator::CommunicatorPolicy_t  
+CartesianCommunicator::CommunicatorPolicy= CartesianCommunicator::CommunicatorPolicyConcurrent;
+int CartesianCommunicator::nCommThreads = -1;
+int CartesianCommunicator::Hugepages = 0;
 
 /////////////////////////////////
 // Alloc, free shmem region
@@ -89,25 +96,175 @@ void CartesianCommunicator::GlobalSumVector(ComplexD *c,int N)
   GlobalSumVector((double *)c,2*N);
 }
 
-#if !defined( GRID_COMMS_MPI3) && !defined (GRID_COMMS_MPI3L)
+
+#if defined( GRID_COMMS_MPI) || defined (GRID_COMMS_MPIT) || defined (GRID_COMMS_MPI3)
+
+CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent,int &srank) 
+{
+  _ndimension = processors.size();
+  assert(_ndimension = parent._ndimension);
+  
+  //////////////////////////////////////////////////////////////////////////////////////////////////////
+  // split the communicator
+  //////////////////////////////////////////////////////////////////////////////////////////////////////
+  int Nparent;
+  MPI_Comm_size(parent.communicator,&Nparent);
+
+  int childsize=1;
+  for(int d=0;d<processors.size();d++) {
+    childsize *= processors[d];
+  }
+  int Nchild = Nparent/childsize;
+  assert (childsize * Nchild == Nparent);
+
+  std::vector<int> ccoor(_ndimension); // coor within subcommunicator
+  std::vector<int> scoor(_ndimension); // coor of split within parent
+  std::vector<int> ssize(_ndimension); // coor of split within parent
+
+  for(int d=0;d<_ndimension;d++){
+    ccoor[d] = parent._processor_coor[d] % processors[d];
+    scoor[d] = parent._processor_coor[d] / processors[d];
+    ssize[d] = parent._processors[d]     / processors[d];
+  }
+  int crank;  // rank within subcomm ; srank is rank of subcomm within blocks of subcomms
+  // Mpi uses the reverse Lexico convention to us
+  Lexicographic::IndexFromCoorReversed(ccoor,crank,processors);
+  Lexicographic::IndexFromCoorReversed(scoor,srank,ssize);
+
+  MPI_Comm comm_split;
+  if ( Nchild > 1 ) { 
+
+    /*
+    std::cout << GridLogMessage<<"Child communicator of "<< std::hex << parent.communicator << std::dec<<std::endl;
+    std::cout << GridLogMessage<<" parent grid["<< parent._ndimension<<"]    ";
+    for(int d=0;d<parent._processors.size();d++)  std::cout << parent._processors[d] << " ";
+    std::cout<<std::endl;
+
+    std::cout << GridLogMessage<<" child grid["<< _ndimension <<"]    ";
+    for(int d=0;d<processors.size();d++)  std::cout << processors[d] << " ";
+    std::cout<<std::endl;
+
+    std::cout << GridLogMessage<<" old rank "<< parent._processor<<" coor ["<< _ndimension <<"]    ";
+    for(int d=0;d<processors.size();d++)  std::cout << parent._processor_coor[d] << " ";
+    std::cout<<std::endl;
+
+    std::cout << GridLogMessage<<" new rank "<< crank<<" coor ["<< _ndimension <<"]    ";
+    for(int d=0;d<processors.size();d++)  std::cout << ccoor[d] << " ";
+    std::cout<<std::endl;
+
+    std::cout << GridLogMessage<<" new coor ["<< _ndimension <<"]    ";
+    for(int d=0;d<processors.size();d++)  std::cout << parent._processor_coor[d] << " ";
+    std::cout<<std::endl;
+    */
+
+    int ierr= MPI_Comm_split(parent.communicator,srank,crank,&comm_split);
+    assert(ierr==0);
+    //////////////////////////////////////////////////////////////////////////////////////////////////////
+    // Declare victory
+    //////////////////////////////////////////////////////////////////////////////////////////////////////
+    /*
+    std::cout << GridLogMessage<<"Divided communicator "<< parent._Nprocessors<<" into "
+	      << Nchild <<" communicators with " << childsize << " ranks"<<std::endl;
+    */
+  } else {
+    comm_split=parent.communicator;
+    srank = 0;
+  }
+
+  //////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Set up from the new split communicator
+  //////////////////////////////////////////////////////////////////////////////////////////////////////
+  InitFromMPICommunicator(processors,comm_split);
+}
+//////////////////////////////////////////////////////////////////////////////////////////////////////
+// Take an MPI_Comm and self assemble
+//////////////////////////////////////////////////////////////////////////////////////////////////////
+void CartesianCommunicator::InitFromMPICommunicator(const std::vector<int> &processors, MPI_Comm communicator_base)
+{
+  _ndimension = processors.size();
+  _processor_coor.resize(_ndimension);
+
+  /////////////////////////////////
+  // Count the requested nodes
+  /////////////////////////////////
+  _Nprocessors=1;
+  _processors = processors;
+  for(int i=0;i<_ndimension;i++){
+    _Nprocessors*=_processors[i];
+  }
+
+  std::vector<int> periodic(_ndimension,1);
+  MPI_Cart_create(communicator_base, _ndimension,&_processors[0],&periodic[0],0,&communicator);
+  MPI_Comm_rank(communicator,&_processor);
+  MPI_Cart_coords(communicator,_processor,_ndimension,&_processor_coor[0]);
+
+  if ( communicator_base != communicator_world ) {
+    std::cout << "Cartesian communicator created with a non-world communicator"<<std::endl;
+    
+    std::cout << " new communicator rank "<<_processor<< " coor ["<<_ndimension<<"] ";
+    for(int d=0;d<_processors.size();d++){
+      std::cout << _processor_coor[d]<<" ";
+    }
+    std::cout << std::endl;
+  }
+
+  int Size;
+  MPI_Comm_size(communicator,&Size);
+
+#ifdef GRID_COMMS_MPIT
+  communicator_halo.resize (2*_ndimension);
+  for(int i=0;i<_ndimension*2;i++){
+    MPI_Comm_dup(communicator,&communicator_halo[i]);
+  }
+#endif
+  
+  assert(Size==_Nprocessors);
+}
+
+CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors) 
+{
+  InitFromMPICommunicator(processors,communicator_world);
+}
+
+#endif
+
+#if !defined( GRID_COMMS_MPI3) 
 
 int                      CartesianCommunicator::NodeCount(void)    { return ProcessorCount();};
 int                      CartesianCommunicator::RankCount(void)    { return ProcessorCount();};
-
+#endif
+#if !defined( GRID_COMMS_MPI3) && !defined (GRID_COMMS_MPIT)
+double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
+						     int xmit_to_rank,
+						     void *recv,
+						     int recv_from_rank,
+						     int bytes, int dir)
+{
+  std::vector<CommsRequest_t> list;
+  // Discard the "dir"
+  SendToRecvFromBegin   (list,xmit,xmit_to_rank,recv,recv_from_rank,bytes);
+  SendToRecvFromComplete(list);
+  return 2.0*bytes;
+}
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 							 void *xmit,
 							 int xmit_to_rank,
 							 void *recv,
 							 int recv_from_rank,
-						       int bytes)
+							 int bytes, int dir)
 {
+  // Discard the "dir"
   SendToRecvFromBegin(list,xmit,xmit_to_rank,recv,recv_from_rank,bytes);
   return 2.0*bytes;
 }
-void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall)
+void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
 {
   SendToRecvFromComplete(waitall);
 }
+#endif
+
+#if !defined( GRID_COMMS_MPI3) 
+
 void CartesianCommunicator::StencilBarrier(void){};
 
 commVector<uint8_t> CartesianCommunicator::ShmBufStorageVector;
@@ -121,8 +278,30 @@ void *CartesianCommunicator::ShmBufferTranslate(int rank,void * local_p) {
   return NULL;
 }
 void CartesianCommunicator::ShmInitGeneric(void){
+#if 1
+  int mmap_flag =0;
+#ifdef MAP_ANONYMOUS
+  mmap_flag = mmap_flag| MAP_SHARED | MAP_ANONYMOUS;
+#endif
+#ifdef MAP_ANON
+  mmap_flag = mmap_flag| MAP_SHARED | MAP_ANON;
+#endif
+#ifdef MAP_HUGETLB
+  if ( Hugepages ) mmap_flag |= MAP_HUGETLB;
+#endif
+  ShmCommBuf =(void *) mmap(NULL, MAX_MPI_SHM_BYTES, PROT_READ | PROT_WRITE, mmap_flag, -1, 0); 
+  if (ShmCommBuf == (void *)MAP_FAILED) {
+    perror("mmap failed ");
+    exit(EXIT_FAILURE);  
+  }
+#ifdef MADV_HUGEPAGE
+  if (!Hugepages ) madvise(ShmCommBuf,MAX_MPI_SHM_BYTES,MADV_HUGEPAGE);
+#endif
+#else 
   ShmBufStorageVector.resize(MAX_MPI_SHM_BYTES);
   ShmCommBuf=(void *)&ShmBufStorageVector[0];
+#endif
+  bzero(ShmCommBuf,MAX_MPI_SHM_BYTES);
 }
 
 #endif

lib/communicator/Communicator_base.h

@@ -38,7 +38,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifdef GRID_COMMS_MPI3
 #include <mpi.h>
 #endif
-#ifdef GRID_COMMS_MPI3L
+#ifdef GRID_COMMS_MPIT
 #include <mpi.h>
 #endif
 #ifdef GRID_COMMS_SHMEM
@@ -50,12 +50,24 @@ namespace Grid {
 class CartesianCommunicator {
   public:    
 
-  // 65536 ranks per node adequate for now
+
+  ////////////////////////////////////////////
+  // Isend/Irecv/Wait, or Sendrecv blocking
+  ////////////////////////////////////////////
+  enum CommunicatorPolicy_t { CommunicatorPolicyConcurrent, CommunicatorPolicySequential };
+  static CommunicatorPolicy_t CommunicatorPolicy;
+  static void SetCommunicatorPolicy(CommunicatorPolicy_t policy ) { CommunicatorPolicy = policy; }
+
+  ///////////////////////////////////////////
+  // Up to 65536 ranks per node adequate for now
   // 128MB shared memory for comms enought for 48^4 local vol comms
   // Give external control (command line override?) of this
-
+  ///////////////////////////////////////////
   static const int MAXLOG2RANKSPERNODE = 16;            
   static uint64_t  MAX_MPI_SHM_BYTES;
+  static int       nCommThreads;
+  // use explicit huge pages
+  static int       Hugepages;
 
   // Communicator should know nothing of the physics grid, only processor grid.
   int              _Nprocessors;     // How many in all
@@ -64,14 +76,19 @@ class CartesianCommunicator {
   std::vector<int> _processor_coor;  // linear processor coordinate
   unsigned long _ndimension;
 
-#if defined (GRID_COMMS_MPI) || defined (GRID_COMMS_MPI3) || defined (GRID_COMMS_MPI3L)
+#if defined (GRID_COMMS_MPI) || defined (GRID_COMMS_MPI3) || defined (GRID_COMMS_MPIT)
   static MPI_Comm communicator_world;
+
   MPI_Comm              communicator;
+  std::vector<MPI_Comm> communicator_halo;
+
   typedef MPI_Request CommsRequest_t;
+
 #else 
   typedef int CommsRequest_t;
 #endif
 
+
   ////////////////////////////////////////////////////////////////////
   // Helper functionality for SHM Windows common to all other impls
   ////////////////////////////////////////////////////////////////////
@@ -117,10 +134,6 @@ class CartesianCommunicator {
   /////////////////////////////////
   static void * ShmCommBuf;
 
-  // Isend/Irecv/Wait, or Sendrecv blocking
-  enum CommunicatorPolicy_t { CommunicatorPolicyConcurrent, CommunicatorPolicySequential };
-  static CommunicatorPolicy_t CommunicatorPolicy;
-  static void SetCommunicatorPolicy(CommunicatorPolicy_t policy ) { CommunicatorPolicy = policy; }
   
   size_t heap_top;
   size_t heap_bytes;
@@ -137,9 +150,22 @@ class CartesianCommunicator {
   static void Init(int *argc, char ***argv);
 
   ////////////////////////////////////////////////
-  // Constructor of any given grid
+  // Constructors to sub-divide a parent communicator
+  // and default to comm world
   ////////////////////////////////////////////////
+  CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent,int &srank);
   CartesianCommunicator(const std::vector<int> &pdimensions_in);
+  virtual ~CartesianCommunicator();
+
+ private:
+#if defined (GRID_COMMS_MPI) || defined (GRID_COMMS_MPIT) 
+  ////////////////////////////////////////////////
+  // Private initialise from an MPI communicator
+  // Can use after an MPI_Comm_split, but hidden from user so private
+  ////////////////////////////////////////////////
+  void InitFromMPICommunicator(const std::vector<int> &processors, MPI_Comm communicator_base);
+#endif
+ public:
   
   ////////////////////////////////////////////////////////////////////////////////////////
   // Wraps MPI_Cart routines, or implements equivalent on other impls
@@ -211,14 +237,21 @@ class CartesianCommunicator {
   
   void SendToRecvFromComplete(std::vector<CommsRequest_t> &waitall);
 
+  double StencilSendToRecvFrom(void *xmit,
+			       int xmit_to_rank,
+			       void *recv,
+			       int recv_from_rank,
+			       int bytes,int dir);
+
   double StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 				    void *xmit,
 				    int xmit_to_rank,
 				    void *recv,
 				    int recv_from_rank,
-				  int bytes);
+				    int bytes,int dir);
   
-  void StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall);
+  
+  void StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int i);
   void StencilBarrier(void);
 
   ////////////////////////////////////////////////////////////
@@ -231,6 +264,27 @@ class CartesianCommunicator {
   ////////////////////////////////////////////////////////////
   void Broadcast(int root,void* data, int bytes);
 
+  ////////////////////////////////////////////////////////////
+  // All2All down one dimension
+  ////////////////////////////////////////////////////////////
+  template<class T> void AllToAll(int dim,std::vector<T> &in, std::vector<T> &out){
+    assert(dim>=0);
+    assert(dim<_ndimension);
+    int numnode = _processors[dim];
+    //    std::cerr << " AllToAll in.size()  "<<in.size()<<std::endl;
+    //    std::cerr << " AllToAll out.size() "<<out.size()<<std::endl;
+    assert(in.size()==out.size());
+    uint64_t bytes=sizeof(T);
+    uint64_t words=in.size()/numnode;
+
+    assert(numnode * words == in.size());
+    assert(words < (1ULL<<32));
+
+    AllToAll(dim,(void *)&in[0],(void *)&out[0],words,bytes);
+  }
+  void AllToAll(int dim  ,void *in,void *out,uint64_t words,uint64_t bytes);
+  void AllToAll(void  *in,void *out,uint64_t words         ,uint64_t bytes);
+  
   template<class obj> void Broadcast(int root,obj &data)
     {
       Broadcast(root,(void *)&data,sizeof(data));

lib/communicator/Communicator_mpi.cc

@@ -53,28 +53,14 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
   ShmInitGeneric();
 }
 
-CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
+CartesianCommunicator::~CartesianCommunicator()
 {
-  _ndimension = processors.size();
-  std::vector<int> periodic(_ndimension,1);
-
-  _Nprocessors=1;
-  _processors = processors;
-  _processor_coor.resize(_ndimension);
-  
-  MPI_Cart_create(communicator_world, _ndimension,&_processors[0],&periodic[0],1,&communicator);
-  MPI_Comm_rank(communicator,&_processor);
-  MPI_Cart_coords(communicator,_processor,_ndimension,&_processor_coor[0]);
-
-  for(int i=0;i<_ndimension;i++){
-    _Nprocessors*=_processors[i];
-  }
-  
-  int Size; 
-  MPI_Comm_size(communicator,&Size);
-  
-  assert(Size==_Nprocessors);
+  int MPI_is_finalised;
+  MPI_Finalized(&MPI_is_finalised);
+  if (communicator && MPI_is_finalised)
+    MPI_Comm_free(&communicator);
 }
+
 void CartesianCommunicator::GlobalSum(uint32_t &u){
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_SUM,communicator);
   assert(ierr==0);
@@ -210,6 +196,36 @@ void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
 		     root,
 		     communicator);
   assert(ierr==0);
+}
+void CartesianCommunicator::AllToAll(int dim,void  *in,void *out,uint64_t words,uint64_t bytes)
+{
+  std::vector<int> row(_ndimension,1);
+  assert(dim>=0 && dim<_ndimension);
+
+  //  Split the communicator
+  row[dim] = _processors[dim];
+
+  int me;
+  CartesianCommunicator Comm(row,*this,me);
+  Comm.AllToAll(in,out,words,bytes);
+}
+void CartesianCommunicator::AllToAll(void  *in,void *out,uint64_t words,uint64_t bytes)
+{
+  // MPI is a pain and uses "int" arguments
+  // 64*64*64*128*16 == 500Million elements of data.
+  // When 24*4 bytes multiples get 50x 10^9 >>> 2x10^9 Y2K bug.
+  // (Turns up on 32^3 x 64 Gparity too)
+  MPI_Datatype object;
+  int iwords; 
+  int ibytes;
+  iwords = words;
+  ibytes = bytes;
+  assert(words == iwords); // safe to cast to int ?
+  assert(bytes == ibytes); // safe to cast to int ?
+  MPI_Type_contiguous(ibytes,MPI_BYTE,&object);
+  MPI_Type_commit(&object);
+  MPI_Alltoall(in,iwords,object,out,iwords,object,communicator);
+  MPI_Type_free(&object);
 }
   ///////////////////////////////////////////////////////
   // Should only be used prior to Grid Init finished.
@@ -230,5 +246,7 @@ void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
   assert(ierr==0);
 }
 
+
+
 }
 

lib/communicator/Communicator_mpi3.cc

@@ -37,11 +37,12 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <sys/ipc.h>
 #include <sys/shm.h>
 #include <sys/mman.h>
-//#include <zlib.h>
-#ifndef SHM_HUGETLB
-#define SHM_HUGETLB 04000
+#include <zlib.h>
+#ifdef HAVE_NUMAIF_H
+#include <numaif.h>
 #endif
 
+
 namespace Grid {
 
 ///////////////////////////////////////////////////////////////////////////////////////////////////
@@ -197,7 +198,46 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
   ShmCommBuf = 0;
   ShmCommBufs.resize(ShmSize);
 
-#if 1
+  ////////////////////////////////////////////////////////////////////////////////////////////
+  // Hugetlbf and others map filesystems as mappable huge pages
+  ////////////////////////////////////////////////////////////////////////////////////////////
+#ifdef GRID_MPI3_SHMMMAP
+  char shm_name [NAME_MAX];
+  for(int r=0;r<ShmSize;r++){
+    
+    size_t size = CartesianCommunicator::MAX_MPI_SHM_BYTES;
+    sprintf(shm_name,GRID_SHM_PATH "/Grid_mpi3_shm_%d_%d",GroupRank,r);
+    //sprintf(shm_name,"/var/lib/hugetlbfs/group/wheel/pagesize-2MB/" "Grid_mpi3_shm_%d_%d",GroupRank,r);
+    //    printf("Opening file %s \n",shm_name);
+    int fd=open(shm_name,O_RDWR|O_CREAT,0666);
+    if ( fd == -1) { 
+      printf("open %s failed\n",shm_name);
+      perror("open hugetlbfs");
+      exit(0);
+    }
+    int mmap_flag = MAP_SHARED ;
+#ifdef MAP_POPULATE    
+    mmap_flag|=MAP_POPULATE;
+#endif
+#ifdef MAP_HUGETLB
+    if ( Hugepages ) mmap_flag |= MAP_HUGETLB;
+#endif
+    void *ptr = (void *) mmap(NULL, MAX_MPI_SHM_BYTES, PROT_READ | PROT_WRITE, mmap_flag,fd, 0); 
+    if ( ptr == (void *)MAP_FAILED ) {    
+      printf("mmap %s failed\n",shm_name);
+      perror("failed mmap");      assert(0);    
+    }
+    assert(((uint64_t)ptr&0x3F)==0);
+    ShmCommBufs[r] =ptr;
+    
+  }
+#endif
+  ////////////////////////////////////////////////////////////////////////////////////////////
+  // POSIX SHMOPEN ; as far as I know Linux does not allow EXPLICIT HugePages with this case
+  // tmpfs (Larry Meadows says) does not support explicit huge page, and this is used for 
+  // the posix shm virtual file system
+  ////////////////////////////////////////////////////////////////////////////////////////////
+#ifdef GRID_MPI3_SHMOPEN
   char shm_name [NAME_MAX];
   if ( ShmRank == 0 ) {
     for(int r=0;r<ShmSize;r++){
@@ -211,9 +251,37 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
       if ( fd < 0 ) {	perror("failed shm_open");	assert(0);      }
       ftruncate(fd, size);
       
-      void * ptr =  mmap(NULL,size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
-      if ( ptr == MAP_FAILED ) {       perror("failed mmap");      assert(0);    }
+      int mmap_flag = MAP_SHARED;
+#ifdef MAP_POPULATE 
+      mmap_flag |= MAP_POPULATE;
+#endif
+#ifdef MAP_HUGETLB
+      if (Hugepages) mmap_flag |= MAP_HUGETLB;
+#endif
+      void * ptr =  mmap(NULL,size, PROT_READ | PROT_WRITE, mmap_flag, fd, 0);
+
+      if ( ptr == (void * )MAP_FAILED ) {       perror("failed mmap");      assert(0);    }
       assert(((uint64_t)ptr&0x3F)==0);
+
+// Experiments; Experiments; Try to force numa domain on the shm segment if we have numaif.h
+#if 0
+//#ifdef HAVE_NUMAIF_H
+	int status;
+	int flags=MPOL_MF_MOVE;
+#ifdef KNL
+	int nodes=1; // numa domain == MCDRAM
+	// Find out if in SNC2,SNC4 mode ?
+#else
+	int nodes=r; // numa domain == MPI ID
+#endif
+	unsigned long count=1;
+	for(uint64_t page=0;page<size;page+=4096){
+	  void *pages = (void *) ( page + (uint64_t)ptr );
+	  uint64_t *cow_it = (uint64_t *)pages;	*cow_it = 1;
+	  ierr= move_pages(0,count, &pages,&nodes,&status,flags);
+	  if (ierr && (page==0)) perror("numa relocate command failed");
+	}
+#endif
 	ShmCommBufs[r] =ptr;
       
     }
@@ -236,23 +304,34 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
       ShmCommBufs[r] =ptr;
     }
   }
-
-#else
+#endif
+  ////////////////////////////////////////////////////////////////////////////////////////////
+  // SHMGET SHMAT and SHM_HUGETLB flag
+  ////////////////////////////////////////////////////////////////////////////////////////////
+#ifdef GRID_MPI3_SHMGET
   std::vector<int> shmids(ShmSize);
 
   if ( ShmRank == 0 ) {
     for(int r=0;r<ShmSize;r++){
       size_t size = CartesianCommunicator::MAX_MPI_SHM_BYTES;
-      key_t key   = 0x4545 + r;
-      if ((shmids[r]= shmget(key,size, SHM_HUGETLB | IPC_CREAT | SHM_R | SHM_W)) < 0) {
+      key_t key   = IPC_PRIVATE;
+      int flags = IPC_CREAT | SHM_R | SHM_W;
+#ifdef SHM_HUGETLB
+      if (Hugepages) flags|=SHM_HUGETLB;
+#endif
+      if ((shmids[r]= shmget(key,size, flags)) ==-1) {
 	int errsv = errno;
 	printf("Errno %d\n",errsv);
+	printf("key   %d\n",key);
+	printf("size  %lld\n",size);
+	printf("flags %d\n",flags);
 	perror("shmget");
 	exit(1);
-      }
+      } else { 
 	printf("shmid: 0x%x\n", shmids[r]);
       }
     }
+  }
   MPI_Barrier(ShmComm);
   MPI_Bcast(&shmids[0],ShmSize*sizeof(int),MPI_BYTE,0,ShmComm);
   MPI_Barrier(ShmComm);
@@ -371,12 +450,27 @@ void  CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &c
   assert(lr!=-1);
   Lexicographic::CoorFromIndex(coor,lr,_processors);
 }
+
+//////////////////////////////////
+// Try to subdivide communicator
+//////////////////////////////////
+CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent) 
+  : CartesianCommunicator(processors) 
+{
+  std::cout << "Attempts to split MPI3 communicators will fail until implemented" <<std::endl;
+}
 CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
 { 
   int ierr;
   communicator=communicator_world;
+
   _ndimension = processors.size();
 
+  communicator_halo.resize (2*_ndimension);
+  for(int i=0;i<_ndimension*2;i++){
+    MPI_Comm_dup(communicator,&communicator_halo[i]);
+  }
+
   ////////////////////////////////////////////////////////////////
   // Assert power of two shm_size.
   ////////////////////////////////////////////////////////////////
@@ -599,13 +693,28 @@ void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &lis
   }
 }
 
+double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
+						     int dest,
+						     void *recv,
+						     int from,
+						     int bytes,int dir)
+{
+  std::vector<CommsRequest_t> list;
+  double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,recv,from,bytes,dir);
+  StencilSendToRecvFromComplete(list,dir);
+  return offbytes;
+}
+
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 							 void *xmit,
 							 int dest,
 							 void *recv,
 							 int from,
-						       int bytes)
+							 int bytes,int dir)
 {
+  int ncomm  =communicator_halo.size(); 
+  int commdir=dir%ncomm;
+
   MPI_Request xrq;
   MPI_Request rrq;
 
@@ -624,26 +733,26 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
   gfrom = MPI_UNDEFINED;
 #endif
   if ( gfrom ==MPI_UNDEFINED) {
-    ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
+    ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator_halo[commdir],&rrq);
     assert(ierr==0);
     list.push_back(rrq);
     off_node_bytes+=bytes;
   }
 
   if ( gdest == MPI_UNDEFINED ) {
-    ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
+    ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator_halo[commdir],&xrq);
     assert(ierr==0);
     list.push_back(xrq);
     off_node_bytes+=bytes;
   }
 
   if ( CommunicatorPolicy == CommunicatorPolicySequential ) { 
-    this->StencilSendToRecvFromComplete(list);
+    this->StencilSendToRecvFromComplete(list,dir);
   }
 
   return off_node_bytes;
 }
-void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall)
+void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
 {
   SendToRecvFromComplete(waitall);
 }

lib/communicator/Communicator_mpit.cc

@@ -0,0 +1,268 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/communicator/Communicator_mpi.cc
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#include <Grid/GridCore.h>
+#include <Grid/GridQCDcore.h>
+#include <Grid/qcd/action/ActionCore.h>
+#include <mpi.h>
+
+namespace Grid {
+
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+// Info that is setup once and indept of cartesian layout
+///////////////////////////////////////////////////////////////////////////////////////////////////
+MPI_Comm CartesianCommunicator::communicator_world;
+
+// Should error check all MPI calls.
+void CartesianCommunicator::Init(int *argc, char ***argv) {
+  int flag;
+  int provided;
+  MPI_Initialized(&flag); // needed to coexist with other libs apparently
+  if ( !flag ) {
+    MPI_Init_thread(argc,argv,MPI_THREAD_MULTIPLE,&provided);
+    if ( provided != MPI_THREAD_MULTIPLE ) {
+      QCD::WilsonKernelsStatic::Comms = QCD::WilsonKernelsStatic::CommsThenCompute;
+    }
+  }
+  MPI_Comm_dup (MPI_COMM_WORLD,&communicator_world);
+  ShmInitGeneric();
+}
+
+CartesianCommunicator::~CartesianCommunicator()
+{
+  if (communicator && !MPI::Is_finalized())
+    MPI_Comm_free(&communicator);
+}
+
+
+void CartesianCommunicator::GlobalSum(uint32_t &u){
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_SUM,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalSum(uint64_t &u){
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_SUM,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalXOR(uint32_t &u){
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_BXOR,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalXOR(uint64_t &u){
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_BXOR,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalSum(float &f){
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalSumVector(float *f,int N)
+{
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,f,N,MPI_FLOAT,MPI_SUM,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalSum(double &d)
+{
+  int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_SUM,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalSumVector(double *d,int N)
+{
+  int ierr = MPI_Allreduce(MPI_IN_PLACE,d,N,MPI_DOUBLE,MPI_SUM,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
+{
+  int ierr=MPI_Cart_shift(communicator,dim,shift,&source,&dest);
+  assert(ierr==0);
+}
+int CartesianCommunicator::RankFromProcessorCoor(std::vector<int> &coor)
+{
+  int rank;
+  int ierr=MPI_Cart_rank  (communicator, &coor[0], &rank);
+  assert(ierr==0);
+  return rank;
+}
+void  CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor)
+{
+  coor.resize(_ndimension);
+  int ierr=MPI_Cart_coords  (communicator, rank, _ndimension,&coor[0]);
+  assert(ierr==0);
+}
+
+// Basic Halo comms primitive
+void CartesianCommunicator::SendToRecvFrom(void *xmit,
+					   int dest,
+					   void *recv,
+					   int from,
+					   int bytes)
+{
+  std::vector<CommsRequest_t> reqs(0);
+  SendToRecvFromBegin(reqs,xmit,dest,recv,from,bytes);
+  SendToRecvFromComplete(reqs);
+}
+
+void CartesianCommunicator::SendRecvPacket(void *xmit,
+					   void *recv,
+					   int sender,
+					   int receiver,
+					   int bytes)
+{
+  MPI_Status stat;
+  assert(sender != receiver);
+  int tag = sender;
+  if ( _processor == sender ) {
+    MPI_Send(xmit, bytes, MPI_CHAR,receiver,tag,communicator);
+  }
+  if ( _processor == receiver ) { 
+    MPI_Recv(recv, bytes, MPI_CHAR,sender,tag,communicator,&stat);
+  }
+}
+
+// Basic Halo comms primitive
+void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+						void *xmit,
+						int dest,
+						void *recv,
+						int from,
+						int bytes)
+{
+  int myrank = _processor;
+  int ierr;
+  if ( CommunicatorPolicy == CommunicatorPolicyConcurrent ) { 
+    MPI_Request xrq;
+    MPI_Request rrq;
+
+    ierr =MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
+    ierr|=MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
+    
+    assert(ierr==0);
+    list.push_back(xrq);
+    list.push_back(rrq);
+  } else { 
+    // Give the CPU to MPI immediately; can use threads to overlap optionally
+    ierr=MPI_Sendrecv(xmit,bytes,MPI_CHAR,dest,myrank,
+		      recv,bytes,MPI_CHAR,from, from,
+		      communicator,MPI_STATUS_IGNORE);
+    assert(ierr==0);
+  }
+}
+void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
+{
+  if ( CommunicatorPolicy == CommunicatorPolicyConcurrent ) { 
+    int nreq=list.size();
+    std::vector<MPI_Status> status(nreq);
+    int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
+    assert(ierr==0);
+  }
+}
+
+void CartesianCommunicator::Barrier(void)
+{
+  int ierr = MPI_Barrier(communicator);
+  assert(ierr==0);
+}
+
+void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
+{
+  int ierr=MPI_Bcast(data,
+		     bytes,
+		     MPI_BYTE,
+		     root,
+		     communicator);
+  assert(ierr==0);
+}
+  ///////////////////////////////////////////////////////
+  // Should only be used prior to Grid Init finished.
+  // Check for this?
+  ///////////////////////////////////////////////////////
+int CartesianCommunicator::RankWorld(void){ 
+  int r; 
+  MPI_Comm_rank(communicator_world,&r);
+  return r;
+}
+void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
+{
+  int ierr= MPI_Bcast(data,
+		      bytes,
+		      MPI_BYTE,
+		      root,
+		      communicator_world);
+  assert(ierr==0);
+}
+
+double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+							 void *xmit,
+							 int xmit_to_rank,
+							 void *recv,
+							 int recv_from_rank,
+							 int bytes,int dir)
+{
+  int myrank = _processor;
+  int ierr;
+  int ncomm  =communicator_halo.size(); 
+  int commdir=dir%ncomm;
+  
+  //  std::cout << " sending on communicator "<<dir<<" " <<communicator_halo[dir]<<std::endl;
+  // Give the CPU to MPI immediately; can use threads to overlap optionally
+  MPI_Request req[2];
+  MPI_Irecv(recv,bytes,MPI_CHAR,recv_from_rank,recv_from_rank, communicator_halo[commdir],&req[1]);
+  MPI_Isend(xmit,bytes,MPI_CHAR,xmit_to_rank  ,myrank        , communicator_halo[commdir],&req[0]);
+
+  list.push_back(req[0]);
+  list.push_back(req[1]);
+  return 2.0*bytes;
+}
+void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
+{ 
+  int nreq=waitall.size();
+  MPI_Waitall(nreq, &waitall[0], MPI_STATUSES_IGNORE);
+};
+double CartesianCommunicator::StencilSendToRecvFrom(void *xmit,
+						    int xmit_to_rank,
+						    void *recv,
+						    int recv_from_rank,
+						    int bytes,int dir)
+{
+  int myrank = _processor;
+  int ierr;
+  //  std::cout << " sending on communicator "<<dir<<" " <<communicator_halo.size()<< <std::endl;
+
+  int ncomm  =communicator_halo.size(); 
+  int commdir=dir%ncomm;
+  // Give the CPU to MPI immediately; can use threads to overlap optionally
+  MPI_Request req[2];
+  MPI_Irecv(recv,bytes,MPI_CHAR,recv_from_rank,recv_from_rank, communicator_halo[commdir],&req[1]);
+  MPI_Isend(xmit,bytes,MPI_CHAR,xmit_to_rank  ,myrank        , communicator_halo[commdir],&req[0]);
+  MPI_Waitall(2, req, MPI_STATUSES_IGNORE);
+  return 2.0*bytes;
+}
+
+
+
+}
+

lib/communicator/Communicator_none.cc

@@ -38,6 +38,9 @@ void CartesianCommunicator::Init(int *argc, char *** arv)
   ShmInitGeneric();
 }
 
+CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent,int &srank) 
+  : CartesianCommunicator(processors) { srank=0;}
+
 CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
 {
   _processors = processors;
@@ -53,6 +56,8 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
   }
 }
 
+CartesianCommunicator::~CartesianCommunicator(){}
+
 void CartesianCommunicator::GlobalSum(float &){}
 void CartesianCommunicator::GlobalSumVector(float *,int N){}
 void CartesianCommunicator::GlobalSum(double &){}
@@ -95,6 +100,14 @@ void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &
 {
   assert(0);
 }
+void CartesianCommunicator::AllToAll(int dim,void  *in,void *out,uint64_t words,uint64_t bytes)
+{
+  bcopy(in,out,bytes*words);
+}
+void CartesianCommunicator::AllToAll(void  *in,void *out,uint64_t words,uint64_t bytes)
+{
+  bcopy(in,out,bytes*words);
+}
 
 int  CartesianCommunicator::RankWorld(void){return 0;}
 void CartesianCommunicator::Barrier(void){}

lib/communicator/Communicator_shmem.cc

@@ -75,6 +75,11 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
   ShmInitGeneric();
 }
 
+CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent) 
+  : CartesianCommunicator(processors) 
+{
+  std::cout << "Attempts to split SHMEM communicators will fail " <<std::endl;
+}
 CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
 {
   _ndimension = processors.size();

lib/cshift/Cshift.h

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

lib/lattice/Lattice_reduction.h

@@ -369,6 +369,7 @@ static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice
   }
 };
 
+/*
 inline GridBase         *makeSubSliceGrid(const GridBase *BlockSolverGrid,int Orthog)
 {
   int NN    = BlockSolverGrid->_ndimension;
@@ -387,6 +388,7 @@ inline GridBase         *makeSubSliceGrid(const GridBase *BlockSolverGrid,int Or
   }
   return (GridBase *)new GridCartesian(latt_phys,simd_phys,mpi_phys); 
 }
+*/
 
 template<class vobj>
 static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,const Lattice<vobj> &Y,int Orthog,RealD scale=1.0) 
@@ -398,14 +400,15 @@ static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice
   int Nblock = X._grid->GlobalDimensions()[Orthog];
 
   GridBase *FullGrid  = X._grid;
-  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
+  //  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
 
-  Lattice<vobj> Xslice(SliceGrid);
-  Lattice<vobj> Rslice(SliceGrid);
+  //  Lattice<vobj> Xslice(SliceGrid);
+  //  Lattice<vobj> Rslice(SliceGrid);
 
   assert( FullGrid->_simd_layout[Orthog]==1);
   int nh =  FullGrid->_ndimension;
-  int nl = SliceGrid->_ndimension;
+  //  int nl = SliceGrid->_ndimension;
+  int nl = nh-1;
 
   //FIXME package in a convenient iterator
   //Should loop over a plane orthogonal to direction "Orthog"
@@ -448,14 +451,14 @@ static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<
   int Nblock = X._grid->GlobalDimensions()[Orthog];
 
   GridBase *FullGrid  = X._grid;
-  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
-
-  Lattice<vobj> Xslice(SliceGrid);
-  Lattice<vobj> Rslice(SliceGrid);
+  //  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
+  //  Lattice<vobj> Xslice(SliceGrid);
+  //  Lattice<vobj> Rslice(SliceGrid);
 
   assert( FullGrid->_simd_layout[Orthog]==1);
   int nh =  FullGrid->_ndimension;
-  int nl = SliceGrid->_ndimension;
+  //  int nl = SliceGrid->_ndimension;
+  int nl=1;
 
   //FIXME package in a convenient iterator
   //Should loop over a plane orthogonal to direction "Orthog"
@@ -498,18 +501,19 @@ static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj>
   typedef typename vobj::vector_type vector_type;
   
   GridBase *FullGrid  = lhs._grid;
-  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
+  //  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
   
   int Nblock = FullGrid->GlobalDimensions()[Orthog];
   
-  Lattice<vobj> Lslice(SliceGrid);
-  Lattice<vobj> Rslice(SliceGrid);
+  //  Lattice<vobj> Lslice(SliceGrid);
+  //  Lattice<vobj> Rslice(SliceGrid);
   
   mat = Eigen::MatrixXcd::Zero(Nblock,Nblock);
 
   assert( FullGrid->_simd_layout[Orthog]==1);
   int nh =  FullGrid->_ndimension;
-  int nl = SliceGrid->_ndimension;
+  //  int nl = SliceGrid->_ndimension;
+  int nl = nh-1;
 
   //FIXME package in a convenient iterator
   //Should loop over a plane orthogonal to direction "Orthog"
@@ -540,7 +544,7 @@ static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj>
       for(int i=0;i<Nblock;i++){
       for(int j=0;j<Nblock;j++){
 	auto tmp = innerProduct(Left[i],Right[j]);
-	vector_typeD rtmp = TensorRemove(tmp);
+	auto rtmp = TensorRemove(tmp);
 	mat_thread(i,j) += Reduce(rtmp);
       }}
     }}
@@ -549,6 +553,14 @@ static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj>
       mat += mat_thread;
     }  
   }
+
+  for(int i=0;i<Nblock;i++){
+  for(int j=0;j<Nblock;j++){
+    ComplexD sum = mat(i,j);
+    FullGrid->GlobalSum(sum);
+    mat(i,j)=sum;
+  }}
+
   return;
 }
 

lib/lattice/Lattice_transfer.h

@@ -109,8 +109,8 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
 
   coarseData=zero;
 
-  // Loop with a cache friendly loop ordering
-  for(int sf=0;sf<fine->oSites();sf++){
+  // Loop over coars parallel, and then loop over fine associated with coarse.
+  parallel_for(int sf=0;sf<fine->oSites();sf++){
 
     int sc;
     std::vector<int> coor_c(_ndimension);
@@ -119,6 +119,7 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
     for(int d=0;d<_ndimension;d++) coor_c[d]=coor_f[d]/block_r[d];
     Lexicographic::IndexFromCoor(coor_c,sc,coarse->_rdimensions);
 
+PARALLEL_CRITICAL
     for(int i=0;i<nbasis;i++) {
 
       coarseData._odata[sc](i)=coarseData._odata[sc](i)
@@ -139,6 +140,7 @@ inline void blockZAXPY(Lattice<vobj> &fineZ,
   GridBase * coarse= coarseA._grid;
 
   fineZ.checkerboard=fineX.checkerboard;
+  assert(fineX.checkerboard==fineY.checkerboard);
   subdivides(coarse,fine); // require they map
   conformable(fineX,fineY);
   conformable(fineX,fineZ);
@@ -180,9 +182,10 @@ template<class vobj,class CComplex>
   GridBase *coarse(CoarseInner._grid);
   GridBase *fine  (fineX._grid);
 
-  Lattice<dotp> fine_inner(fine);
+  Lattice<dotp> fine_inner(fine); fine_inner.checkerboard = fineX.checkerboard;
   Lattice<dotp> coarse_inner(coarse);
 
+  // Precision promotion?
   fine_inner = localInnerProduct(fineX,fineY);
   blockSum(coarse_inner,fine_inner);
   parallel_for(int ss=0;ss<coarse->oSites();ss++){
@@ -193,7 +196,7 @@ template<class vobj,class CComplex>
 inline void blockNormalise(Lattice<CComplex> &ip,Lattice<vobj> &fineX)
 {
   GridBase *coarse = ip._grid;
-  Lattice<vobj> zz(fineX._grid); zz=zero;
+  Lattice<vobj> zz(fineX._grid); zz=zero; zz.checkerboard=fineX.checkerboard;
   blockInnerProduct(ip,fineX,fineX);
   ip = pow(ip,-0.5);
   blockZAXPY(fineX,ip,fineX,zz);
@@ -216,20 +219,26 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
     block_r[d] = fine->_rdimensions[d] / coarse->_rdimensions[d];
   }
 
+  // Turn this around to loop threaded over sc and interior loop 
+  // over sf would thread better
   coarseData=zero;
-  for(int sf=0;sf<fine->oSites();sf++){
+  parallel_region {
 
     int sc;
     std::vector<int> coor_c(_ndimension);
     std::vector<int> coor_f(_ndimension);
 
+    parallel_for_internal(int sf=0;sf<fine->oSites();sf++){
+    
       Lexicographic::CoorFromIndex(coor_f,sf,fine->_rdimensions);
       for(int d=0;d<_ndimension;d++) coor_c[d]=coor_f[d]/block_r[d];
       Lexicographic::IndexFromCoor(coor_c,sc,coarse->_rdimensions);
       
+PARALLEL_CRITICAL
       coarseData._odata[sc]=coarseData._odata[sc]+fineData._odata[sf];
 
     }
+  }
   return;
 }
 
@@ -238,7 +247,7 @@ inline void blockPick(GridBase *coarse,const Lattice<vobj> &unpicked,Lattice<vob
 {
   GridBase * fine = unpicked._grid;
 
-  Lattice<vobj> zz(fine);
+  Lattice<vobj> zz(fine); zz.checkerboard = unpicked.checkerboard;
   Lattice<iScalar<vInteger> > fcoor(fine);
 
   zz = zero;
@@ -303,12 +312,13 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
   }
 
   // Loop with a cache friendly loop ordering
-  for(int sf=0;sf<fine->oSites();sf++){
-
+  parallel_region {
     int sc;
     std::vector<int> coor_c(_ndimension);
     std::vector<int> coor_f(_ndimension);
 
+    parallel_for_internal(int sf=0;sf<fine->oSites();sf++){
+
       Lexicographic::CoorFromIndex(coor_f,sf,fine->_rdimensions);
       for(int d=0;d<_ndimension;d++) coor_c[d]=coor_f[d]/block_r[d];
       Lexicographic::IndexFromCoor(coor_c,sc,coarse->_rdimensions);
@@ -316,7 +326,7 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
       for(int i=0;i<nbasis;i++) {
 	if(i==0) fineData._odata[sf]=coarseData._odata[sc](i) * Basis[i]._odata[sf];
 	else     fineData._odata[sf]=fineData._odata[sf]+coarseData._odata[sc](i)*Basis[i]._odata[sf];
-
+      }
     }
   }
   return;
@@ -685,5 +695,314 @@ void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in){
   }
 }
 
+////////////////////////////////////////////////////////////////////////////////
+// Communicate between grids
+////////////////////////////////////////////////////////////////////////////////
+//
+// All to all plan
+//
+// Subvolume on fine grid is v.    Vectors a,b,c,d 
+//
+///////////////////////////////////////////////////////////////////////////////////////////////////////////
+// SIMPLEST CASE:
+///////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Mesh of nodes (2) ; subdivide to  1 subdivisions
+//
+// Lex ord:   
+//          N0 va0 vb0  N1 va1 vb1 
+//
+// For each dimension do an all to all
+//
+// full AllToAll(0)
+//          N0 va0 va1    N1 vb0 vb1
+//
+// REARRANGE
+//          N0 va01       N1 vb01
+//
+// Must also rearrange data to get into the NEW lex order of grid at each stage. Some kind of "insert/extract".
+// NB: Easiest to programme if keep in lex order.
+//
+///////////////////////////////////////////////////////////////////////////////////////////////////////////
+// SIMPLE CASE:
+///////////////////////////////////////////////////////////////////////////////////////////////////////////
+//
+// Mesh of nodes (2x2) ; subdivide to  1x1 subdivisions
+//
+// Lex ord:   
+//          N0 va0 vb0 vc0 vd0       N1 va1 vb1 vc1 vd1  
+//          N2 va2 vb2 vc2 vd2       N3 va3 vb3 vc3 vd3 
+//
+// Ratio = full[dim] / split[dim]
+//
+// For each dimension do an all to all; get Nvec -> Nvec / ratio
+//                                          Ldim -> Ldim * ratio
+//                                          LocalVol -> LocalVol * ratio
+// full AllToAll(0)
+//          N0 va0 vb0 va1 vb1       N1 vc0 vd0 vc1 vd1   
+//          N2 va2 vb2 va3 vb3       N3 vc2 vd2 vc3 vd3 
+//
+// REARRANGE
+//          N0 va01 vb01      N1 vc01 vd01
+//          N2 va23 vb23      N3 vc23 vd23
+//
+// full AllToAll(1)           // Not what is wanted. FIXME
+//          N0 va01 va23      N1 vc01 vc23 
+//          N2 vb01 vb23      N3 vd01 vd23
+// 
+// REARRANGE
+//          N0 va0123      N1 vc0123
+//          N2 vb0123      N3 vd0123
+//
+// Must also rearrange data to get into the NEW lex order of grid at each stage. Some kind of "insert/extract".
+// NB: Easiest to programme if keep in lex order.
+//
+/////////////////////////////////////////////////////////
+
+template<class Vobj>
+void Grid_split(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
+{
+  typedef typename Vobj::scalar_object Sobj;
+
+  int full_vecs   = full.size();
+
+  assert(full_vecs>=1);
+
+  GridBase * full_grid = full[0]._grid;
+  GridBase *split_grid = split._grid;
+
+  int       ndim  = full_grid->_ndimension;
+  int  full_nproc = full_grid->_Nprocessors;
+  int split_nproc =split_grid->_Nprocessors;
+
+  ////////////////////////////////
+  // Checkerboard management
+  ////////////////////////////////
+  int cb = full[0].checkerboard;
+  split.checkerboard = cb;
+
+  //////////////////////////////
+  // Checks
+  //////////////////////////////
+  assert(full_grid->_ndimension==split_grid->_ndimension);
+  for(int n=0;n<full_vecs;n++){
+    assert(full[n].checkerboard == cb);
+    for(int d=0;d<ndim;d++){
+      assert(full[n]._grid->_gdimensions[d]==split._grid->_gdimensions[d]);
+      assert(full[n]._grid->_fdimensions[d]==split._grid->_fdimensions[d]);
+    }
+  }
+
+  int   nvector   =full_nproc/split_nproc; 
+  assert(nvector*split_nproc==full_nproc);
+  assert(nvector == full_vecs);
+
+  std::vector<int> ratio(ndim);
+  for(int d=0;d<ndim;d++){
+    ratio[d] = full_grid->_processors[d]/ split_grid->_processors[d];
+  }
+
+  uint64_t lsites = full_grid->lSites();
+  uint64_t     sz = lsites * nvector;
+  std::vector<Sobj> tmpdata(sz);
+  std::vector<Sobj> alldata(sz);
+  std::vector<Sobj> scalardata(lsites); 
+
+  for(int v=0;v<nvector;v++){
+    unvectorizeToLexOrdArray(scalardata,full[v]);    
+    parallel_for(int site=0;site<lsites;site++){
+      alldata[v*lsites+site] = scalardata[site];
+    }
+  }
+
+  int nvec = nvector; // Counts down to 1 as we collapse dims
+  std::vector<int> ldims = full_grid->_ldimensions;
+  std::vector<int> lcoor(ndim);
+
+  for(int d=ndim-1;d>=0;d--){
+
+    if ( ratio[d] != 1 ) {
+
+      full_grid ->AllToAll(d,alldata,tmpdata);
+      //      std::cout << GridLogMessage << "Grid_split: dim " <<d<<" ratio "<<ratio[d]<<" nvec "<<nvec<<" procs "<<split_grid->_processors[d]<<std::endl;
+      //      for(int v=0;v<nvec;v++){
+      //	std::cout << "Grid_split: alldata["<<v<<"] " << alldata[v] <<std::endl;
+      //	std::cout << "Grid_split: tmpdata["<<v<<"] " << tmpdata[v] <<std::endl;
+      //      }
+      //////////////////////////////////////////
+      //Local volume for this dimension is expanded by ratio of processor extents
+      // Number of vectors is decreased by same factor
+      // Rearrange to lexico for bigger volume
+      //////////////////////////////////////////
+      nvec    /= ratio[d];
+
+      auto rdims = ldims; rdims[d]  *=   ratio[d];
+      auto rsites= lsites*ratio[d];
+      for(int v=0;v<nvec;v++){
+
+	// For loop over each site within old subvol
+	for(int lsite=0;lsite<lsites;lsite++){
+
+	  Lexicographic::CoorFromIndex(lcoor, lsite, ldims);	  
+
+	  for(int r=0;r<ratio[d];r++){ // ratio*nvec terms
+
+	    auto rcoor = lcoor;	    rcoor[d]  += r*ldims[d];
+
+	    int rsite; Lexicographic::IndexFromCoor(rcoor, rsite, rdims);	  
+	    rsite += v * rsites;
+
+	    int rmul=nvec*lsites;
+	    int vmul=     lsites;
+	    alldata[rsite] = tmpdata[lsite+r*rmul+v*vmul];
+	    //	    if ( lsite==0 ) {
+	    //	      std::cout << "Grid_split: grow alldata["<<rsite<<"] " << alldata[rsite] << " <- tmpdata["<< lsite+r*rmul+v*vmul<<"] "<<tmpdata[lsite+r*rmul+v*vmul]  <<std::endl;
+	    //	    }	      
+	  }
+	}
+      }
+      ldims[d]*= ratio[d];
+      lsites  *= ratio[d];
+
+      if ( split_grid->_processors[d] > 1 ) {
+	tmpdata = alldata;
+	split_grid->AllToAll(d,tmpdata,alldata);
+      }
+    }
+  }
+  vectorizeFromLexOrdArray(alldata,split);    
+}
+
+template<class Vobj>
+void Grid_split(Lattice<Vobj> &full,Lattice<Vobj>   & split)
+{
+  int nvector = full._grid->_Nprocessors / split._grid->_Nprocessors;
+  std::vector<Lattice<Vobj> > full_v(nvector,full._grid);
+  for(int n=0;n<nvector;n++){
+    full_v[n] = full;
+  }
+  Grid_split(full_v,split);
+}
+
+template<class Vobj>
+void Grid_unsplit(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
+{
+  typedef typename Vobj::scalar_object Sobj;
+
+  int full_vecs   = full.size();
+
+  assert(full_vecs>=1);
+
+  GridBase * full_grid = full[0]._grid;
+  GridBase *split_grid = split._grid;
+
+  int       ndim  = full_grid->_ndimension;
+  int  full_nproc = full_grid->_Nprocessors;
+  int split_nproc =split_grid->_Nprocessors;
+
+  ////////////////////////////////
+  // Checkerboard management
+  ////////////////////////////////
+  int cb = full[0].checkerboard;
+  split.checkerboard = cb;
+
+  //////////////////////////////
+  // Checks
+  //////////////////////////////
+  assert(full_grid->_ndimension==split_grid->_ndimension);
+  for(int n=0;n<full_vecs;n++){
+    assert(full[n].checkerboard == cb);
+    for(int d=0;d<ndim;d++){
+      assert(full[n]._grid->_gdimensions[d]==split._grid->_gdimensions[d]);
+      assert(full[n]._grid->_fdimensions[d]==split._grid->_fdimensions[d]);
+    }
+  }
+
+  int   nvector   =full_nproc/split_nproc; 
+  assert(nvector*split_nproc==full_nproc);
+  assert(nvector == full_vecs);
+
+  std::vector<int> ratio(ndim);
+  for(int d=0;d<ndim;d++){
+    ratio[d] = full_grid->_processors[d]/ split_grid->_processors[d];
+  }
+
+  uint64_t lsites = full_grid->lSites();
+  uint64_t     sz = lsites * nvector;
+  std::vector<Sobj> tmpdata(sz);
+  std::vector<Sobj> alldata(sz);
+  std::vector<Sobj> scalardata(lsites); 
+
+  unvectorizeToLexOrdArray(alldata,split);    
+
+  /////////////////////////////////////////////////////////////////
+  // Start from split grid and work towards full grid
+  /////////////////////////////////////////////////////////////////
+  std::vector<int> lcoor(ndim);
+  std::vector<int> rcoor(ndim);
+
+  int nvec = 1;
+  lsites = split_grid->lSites();
+  std::vector<int> ldims = split_grid->_ldimensions;
+
+  //  for(int d=ndim-1;d>=0;d--){
+  for(int d=0;d<ndim;d++){
+
+    if ( ratio[d] != 1 ) {
+
+
+      if ( split_grid->_processors[d] > 1 ) {
+	tmpdata = alldata;
+	split_grid->AllToAll(d,tmpdata,alldata);
+      }
+
+      //////////////////////////////////////////
+      //Local volume for this dimension is expanded by ratio of processor extents
+      // Number of vectors is decreased by same factor
+      // Rearrange to lexico for bigger volume
+      //////////////////////////////////////////
+      auto rsites= lsites/ratio[d];
+      auto rdims = ldims; rdims[d]/=ratio[d];
+
+      for(int v=0;v<nvec;v++){
+
+	// rsite, rcoor --> smaller local volume
+	// lsite, lcoor --> bigger original (single node?) volume
+	// For loop over each site within smaller subvol
+	for(int rsite=0;rsite<rsites;rsite++){
+
+	  Lexicographic::CoorFromIndex(rcoor, rsite, rdims);	  
+	  int lsite;
+
+	  for(int r=0;r<ratio[d];r++){ 
+
+	    lcoor = rcoor; lcoor[d] += r*rdims[d];
+	    Lexicographic::IndexFromCoor(lcoor, lsite, ldims); lsite += v * lsites;
+
+	    int rmul=nvec*rsites;
+	    int vmul=     rsites;
+	    tmpdata[rsite+r*rmul+v*vmul]=alldata[lsite];
+
+	  }
+	}
+      }
+      nvec   *= ratio[d];
+      ldims[d]=rdims[d];
+      lsites  =rsites;
+
+      full_grid ->AllToAll(d,tmpdata,alldata);
+    }
+  }
+
+  lsites = full_grid->lSites();
+  for(int v=0;v<nvector;v++){
+    assert(v<full.size());
+    parallel_for(int site=0;site<lsites;site++){
+      scalardata[site] = alldata[v*lsites+site];
+    }
+    vectorizeFromLexOrdArray(scalardata,full[v]);    
+  }
+}
+
+ 
 }
 #endif

lib/log/Log.cc

@@ -50,7 +50,7 @@ namespace Grid {
     return (status==0) ? res.get() : name ;
   }
   
-GridStopWatch Logger::StopWatch;
+GridStopWatch Logger::GlobalStopWatch;
 int Logger::timestamp;
 std::ostream Logger::devnull(0);
 
@@ -59,13 +59,15 @@ void GridLogTimestamp(int on){
 }
 
 Colours GridLogColours(0);
-GridLogger GridLogError(1, "Error", GridLogColours, "RED");
+GridLogger GridLogIRL    (1, "IRL"   , GridLogColours, "NORMAL");
+GridLogger GridLogSolver (1, "Solver", GridLogColours, "NORMAL");
+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 GridLogDebug  (1, "Debug", GridLogColours, "PURPLE");
 GridLogger GridLogPerformance(1, "Performance", GridLogColours, "GREEN");
-GridLogger GridLogIterative(1, "Iterative", GridLogColours, "BLUE");
-GridLogger GridLogIntegrator(1, "Integrator", GridLogColours, "BLUE");
+GridLogger GridLogIterative  (1, "Iterative", GridLogColours, "BLUE");
+GridLogger GridLogIntegrator (1, "Integrator", GridLogColours, "BLUE");
 
 void GridLogConfigure(std::vector<std::string> &logstreams) {
   GridLogError.Active(0);
@@ -95,7 +97,7 @@ void GridLogConfigure(std::vector<std::string> &logstreams) {
 ////////////////////////////////////////////////////////////
 void Grid_quiesce_nodes(void) {
   int me = 0;
-#if defined(GRID_COMMS_MPI) || defined(GRID_COMMS_MPI3) || defined(GRID_COMMS_MPI3L)
+#if defined(GRID_COMMS_MPI) || defined(GRID_COMMS_MPI3) || defined(GRID_COMMS_MPIT)
   MPI_Comm_rank(MPI_COMM_WORLD, &me);
 #endif
 #ifdef GRID_COMMS_SHMEM

lib/log/Log.h

@@ -85,12 +85,15 @@ class Logger {
 protected:
   Colours &Painter;
   int active;
+  int timing_mode;
   static int timestamp;
   std::string name, topName;
   std::string COLOUR;
 
 public:
-  static GridStopWatch StopWatch;
+  static GridStopWatch GlobalStopWatch;
+  GridStopWatch         LocalStopWatch;
+  GridStopWatch *StopWatch;
   static std::ostream devnull;
 
   std::string background() {return Painter.colour["NORMAL"];}
@@ -101,22 +104,38 @@ public:
     name(nm),
     topName(topNm),
     Painter(col_class),
-    COLOUR(col) {} ;
+    timing_mode(0),
+    COLOUR(col) 
+    {
+      StopWatch = & GlobalStopWatch;
+    };
   
   void Active(int on) {active = on;};
   int  isActive(void) {return active;};
   static void Timestamp(int on) {timestamp = on;};
+  void Reset(void) { 
+    StopWatch->Reset(); 
+    StopWatch->Start(); 
+  }
+  void TimingMode(int on) { 
+    timing_mode = on; 
+    if(on) { 
+      StopWatch = &LocalStopWatch;
+      Reset(); 
+    }
+  }
 
   friend std::ostream& operator<< (std::ostream& stream, Logger& log){
 
     if ( log.active ) {
-      stream << log.background()<< std::setw(8) << std::left << log.topName << log.background()<< " : ";
-      stream << log.colour() << std::setw(10) << std::left << log.name << log.background() << " : ";
+      stream << log.background()<<  std::left << log.topName << log.background()<< " : ";
+      stream << log.colour() <<  std::left << log.name << log.background() << " : ";
       if ( log.timestamp ) {
-	StopWatch.Stop();
-	GridTime now = StopWatch.Elapsed();
-	StopWatch.Start();
-	stream << log.evidence()<< now << log.background() << " : " ;
+	log.StopWatch->Stop();
+	GridTime now = log.StopWatch->Elapsed();
+	if ( log.timing_mode==1 ) log.StopWatch->Reset();
+	log.StopWatch->Start();
+	stream << log.evidence()<< std::setw(6)<<now << log.background() << " : " ;
       }
       stream << log.colour();
       return stream;
@@ -135,6 +154,8 @@ public:
 
 void GridLogConfigure(std::vector<std::string> &logstreams);
 
+extern GridLogger GridLogIRL;
+extern GridLogger GridLogSolver;
 extern GridLogger GridLogError;
 extern GridLogger GridLogWarning;
 extern GridLogger GridLogMessage;

lib/parallelIO/BinaryIO.h

@@ -29,7 +29,7 @@
 #ifndef GRID_BINARY_IO_H
 #define GRID_BINARY_IO_H
 
-#if defined(GRID_COMMS_MPI) || defined(GRID_COMMS_MPI3) 
+#if defined(GRID_COMMS_MPI) || defined(GRID_COMMS_MPI3) || defined(GRID_COMMS_MPIT) 
 #define USE_MPI_IO
 #else
 #undef  USE_MPI_IO
@@ -99,34 +99,38 @@ class BinaryIO {
     NerscChecksum(grid,scalardata,nersc_csum);
   }
 
-  template<class fobj> static inline void NerscChecksum(GridBase *grid,std::vector<fobj> &fbuf,uint32_t &nersc_csum)
+  template <class fobj>
+  static inline void NerscChecksum(GridBase *grid, std::vector<fobj> &fbuf, uint32_t &nersc_csum)
   {
-    const uint64_t size32 = sizeof(fobj)/sizeof(uint32_t);
+    const uint64_t size32 = sizeof(fobj) / sizeof(uint32_t);
 
-
-    uint64_t lsites              =grid->lSites();
-    if (fbuf.size()==1) {
-      lsites=1;
-    }
-
-#pragma omp parallel
+    uint64_t lsites = grid->lSites();
+    if (fbuf.size() == 1)
     {
-      uint32_t nersc_csum_thr=0;
+      lsites = 1;
+    }
 
-#pragma omp for
-      for(uint64_t local_site=0;local_site<lsites;local_site++){
-	uint32_t * site_buf = (uint32_t *)&fbuf[local_site];
-	for(uint64_t j=0;j<size32;j++){
-	  nersc_csum_thr=nersc_csum_thr+site_buf[j];
+    #pragma omp parallel
+    {
+      uint32_t nersc_csum_thr = 0;
+
+      #pragma omp for
+      for (uint64_t local_site = 0; local_site < lsites; local_site++)
+      {
+        uint32_t *site_buf = (uint32_t *)&fbuf[local_site];
+        for (uint64_t j = 0; j < size32; j++)
+        {
+          nersc_csum_thr = nersc_csum_thr + site_buf[j];
         }
       }
 
-#pragma omp critical
+      #pragma omp critical
       {
         nersc_csum += nersc_csum_thr;
       }
     }
   }
+
   template<class fobj> static inline void ScidacChecksum(GridBase *grid,std::vector<fobj> &fbuf,uint32_t &scidac_csuma,uint32_t &scidac_csumb)
   {
     const uint64_t size32 = sizeof(fobj)/sizeof(uint32_t);
@@ -257,7 +261,7 @@ class BinaryIO {
 			      GridBase *grid,
 			      std::vector<fobj> &iodata,
 			      std::string file,
-			      int offset,
+			      Integer offset,
 			      const std::string &format, int control,
 			      uint32_t &nersc_csum,
 			      uint32_t &scidac_csuma,
@@ -352,7 +356,7 @@ class BinaryIO {
 
       if ( (control & BINARYIO_LEXICOGRAPHIC) && (nrank > 1) ) {
 #ifdef USE_MPI_IO
-	std::cout<< GridLogMessage<< "MPI read I/O "<< file<< std::endl;
+	std::cout<< GridLogMessage<<"IOobject: MPI read I/O "<< file<< std::endl;
 	ierr=MPI_File_open(grid->communicator,(char *) file.c_str(), MPI_MODE_RDONLY, MPI_INFO_NULL, &fh);    assert(ierr==0);
 	ierr=MPI_File_set_view(fh, disp, mpiObject, fileArray, "native", MPI_INFO_NULL);    assert(ierr==0);
 	ierr=MPI_File_read_all(fh, &iodata[0], 1, localArray, &status);    assert(ierr==0);
@@ -363,16 +367,20 @@ class BinaryIO {
 	assert(0);
 #endif
       } else {
-	std::cout<< GridLogMessage<< "C++ read I/O "<< file<<" : "
-		 << iodata.size()*sizeof(fobj)<<" bytes"<<std::endl;
+	std::cout << GridLogMessage <<"IOobject: C++ read I/O " << file << " : "
+                  << iodata.size() * sizeof(fobj) << " bytes" << std::endl;
         std::ifstream fin;
-	fin.open(file,std::ios::binary|std::ios::in);
-	if ( control & BINARYIO_MASTER_APPEND )  {
-	  fin.seekg(-sizeof(fobj),fin.end);
-	} else { 
-	  fin.seekg(offset+myrank*lsites*sizeof(fobj));
+        fin.open(file, std::ios::binary | std::ios::in);
+        if (control & BINARYIO_MASTER_APPEND)
+        {
+          fin.seekg(-sizeof(fobj), fin.end);
         }
-	fin.read((char *)&iodata[0],iodata.size()*sizeof(fobj));assert( fin.fail()==0);
+        else
+        {
+          fin.seekg(offset + myrank * lsites * sizeof(fobj));
+        }
+        fin.read((char *)&iodata[0], iodata.size() * sizeof(fobj));
+        assert(fin.fail() == 0);
         fin.close();
       }
       timer.Stop();
@@ -405,10 +413,30 @@ class BinaryIO {
       timer.Start();
       if ( (control & BINARYIO_LEXICOGRAPHIC) && (nrank > 1) ) {
 #ifdef USE_MPI_IO
-	std::cout<< GridLogMessage<< "MPI write I/O "<< file<< std::endl;
-	ierr=MPI_File_open(grid->communicator,(char *) file.c_str(), MPI_MODE_RDWR|MPI_MODE_CREATE,MPI_INFO_NULL, &fh); assert(ierr==0);
-	ierr=MPI_File_set_view(fh, disp, mpiObject, fileArray, "native", MPI_INFO_NULL);                        assert(ierr==0);
-	ierr=MPI_File_write_all(fh, &iodata[0], 1, localArray, &status);                                        assert(ierr==0);
+        std::cout << GridLogMessage <<"IOobject: MPI write I/O " << file << std::endl;
+        ierr = MPI_File_open(grid->communicator, (char *)file.c_str(), MPI_MODE_RDWR | MPI_MODE_CREATE, MPI_INFO_NULL, &fh);
+	//        std::cout << GridLogMessage << "Checking for errors" << std::endl;
+        if (ierr != MPI_SUCCESS)
+        {
+          char error_string[BUFSIZ];
+          int length_of_error_string, error_class;
+
+          MPI_Error_class(ierr, &error_class);
+          MPI_Error_string(error_class, error_string, &length_of_error_string);
+          fprintf(stderr, "%3d: %s\n", myrank, error_string);
+          MPI_Error_string(ierr, error_string, &length_of_error_string);
+          fprintf(stderr, "%3d: %s\n", myrank, error_string);
+          MPI_Abort(MPI_COMM_WORLD, 1); //assert(ierr == 0);
+        }
+
+        std::cout << GridLogDebug << "MPI read I/O set view " << file << std::endl;
+        ierr = MPI_File_set_view(fh, disp, mpiObject, fileArray, "native", MPI_INFO_NULL);
+        assert(ierr == 0);
+
+        std::cout << GridLogDebug << "MPI read I/O write all " << file << std::endl;
+        ierr = MPI_File_write_all(fh, &iodata[0], 1, localArray, &status);
+        assert(ierr == 0);
+
         MPI_File_close(&fh);
         MPI_Type_free(&fileArray);
         MPI_Type_free(&localArray);
@@ -416,15 +444,51 @@ class BinaryIO {
 	assert(0);
 #endif
       } else { 
-	std::ofstream fout; fout.open(file,std::ios::binary|std::ios::out|std::ios::in);
-	std::cout<< GridLogMessage<< "C++ write I/O "<< file<<" : "
-		 << iodata.size()*sizeof(fobj)<<" bytes"<<std::endl;
-	if ( control & BINARYIO_MASTER_APPEND )  {
-	  fout.seekp(0,fout.end);
-	} else {
-	  fout.seekp(offset+myrank*lsites*sizeof(fobj));
+
+        std::cout << GridLogMessage << "IOobject: C++ write I/O " << file << " : "
+                  << iodata.size() * sizeof(fobj) << " bytes" << std::endl;
+        
+	std::ofstream fout; 
+	fout.exceptions ( std::fstream::failbit | std::fstream::badbit );
+	try {
+	  fout.open(file,std::ios::binary|std::ios::out|std::ios::in);
+	} catch (const std::fstream::failure& exc) {
+	  std::cout << GridLogError << "Error in opening the file " << file << " for output" <<std::endl;
+	  std::cout << GridLogError << "Exception description: " << exc.what() << std::endl;
+	  std::cout << GridLogError << "Probable cause: wrong path, inaccessible location "<< std::endl;
+#ifdef USE_MPI_IO
+	  MPI_Abort(MPI_COMM_WORLD,1);
+#else
+	  exit(1);
+#endif
+	}
+	
+	if ( control & BINARYIO_MASTER_APPEND )  {
+	  try {
+	    fout.seekp(0,fout.end);
+	  } catch (const std::fstream::failure& exc) {
+	    std::cout << "Exception in seeking file end " << file << std::endl;
+	  }
+	} else {
+	  try { 
+	    fout.seekp(offset+myrank*lsites*sizeof(fobj));
+	  } catch (const std::fstream::failure& exc) {
+	    std::cout << "Exception in seeking file " << file <<" offset "<< offset << std::endl;
+	  }
+	}
+
+	try {
+	  fout.write((char *)&iodata[0],iodata.size()*sizeof(fobj));//assert( fout.fail()==0);
+	}
+	catch (const std::fstream::failure& exc) {
+	  std::cout << "Exception in writing file " << file << std::endl;
+	  std::cout << GridLogError << "Exception description: "<< exc.what() << std::endl;
+#ifdef USE_MPI_IO
+	  MPI_Abort(MPI_COMM_WORLD,1);
+#else
+	  exit(1);
+#endif
 	}
-	fout.write((char *)&iodata[0],iodata.size()*sizeof(fobj));assert( fout.fail()==0);
 	fout.close();
       }
       timer.Stop();
@@ -442,12 +506,15 @@ class BinaryIO {
     //////////////////////////////////////////////////////////////////////////////
     // Safety check
     //////////////////////////////////////////////////////////////////////////////
+    // if the data size is 1 we do not want to sum over the MPI ranks
+    if (iodata.size() != 1){
       grid->Barrier();
       grid->GlobalSum(nersc_csum);
       grid->GlobalXOR(scidac_csuma);
       grid->GlobalXOR(scidac_csumb);
       grid->Barrier();
     }
+  }
 
   /////////////////////////////////////////////////////////////////////////////
   // Read a Lattice of object
@@ -456,7 +523,7 @@ class BinaryIO {
   static inline void readLatticeObject(Lattice<vobj> &Umu,
 				       std::string file,
 				       munger munge,
-				       int offset,
+				       Integer offset,
 				       const std::string &format,
 				       uint32_t &nersc_csum,
 				       uint32_t &scidac_csuma,
@@ -493,7 +560,7 @@ class BinaryIO {
     static inline void writeLatticeObject(Lattice<vobj> &Umu,
 					  std::string file,
 					  munger munge,
-					  int offset,
+					  Integer offset,
 					  const std::string &format,
 					  uint32_t &nersc_csum,
 					  uint32_t &scidac_csuma,
@@ -530,7 +597,7 @@ class BinaryIO {
   static inline void readRNG(GridSerialRNG &serial,
 			     GridParallelRNG &parallel,
 			     std::string file,
-			     int offset,
+			     Integer offset,
 			     uint32_t &nersc_csum,
 			     uint32_t &scidac_csuma,
 			     uint32_t &scidac_csumb)
@@ -546,9 +613,9 @@ class BinaryIO {
     int gsites = grid->gSites();
     int lsites = grid->lSites();
 
-    uint32_t nersc_csum_tmp;
-    uint32_t scidac_csuma_tmp;
-    uint32_t scidac_csumb_tmp;
+    uint32_t nersc_csum_tmp   = 0;
+    uint32_t scidac_csuma_tmp = 0;
+    uint32_t scidac_csumb_tmp = 0;
 
     GridStopWatch timer;
 
@@ -592,7 +659,7 @@ class BinaryIO {
   static inline void writeRNG(GridSerialRNG &serial,
 			      GridParallelRNG &parallel,
 			      std::string file,
-			      int offset,
+			      Integer offset,
 			      uint32_t &nersc_csum,
 			      uint32_t &scidac_csuma,
 			      uint32_t &scidac_csumb)

lib/parallelIO/IldgIO.h

@@ -84,10 +84,6 @@ namespace QCD {
    stream << "GRID_";
    stream << ScidacWordMnemonic<stype>();
 
-   //   std::cout << " Lorentz N/S/V/M : " << _LorentzN<<" "<<_LorentzScalar<<"/"<<_LorentzVector<<"/"<<_LorentzMatrix<<std::endl;
-   //   std::cout << " Spin    N/S/V/M : " << _SpinN   <<" "<<_SpinScalar   <<"/"<<_SpinVector   <<"/"<<_SpinMatrix<<std::endl;
-   //   std::cout << " Colour  N/S/V/M : " << _ColourN <<" "<<_ColourScalar <<"/"<<_ColourVector <<"/"<<_ColourMatrix<<std::endl;
-
    if ( _LorentzVector )   stream << "_LorentzVector"<<_LorentzN;
    if ( _LorentzMatrix )   stream << "_LorentzMatrix"<<_LorentzN;
 
@@ -151,7 +147,7 @@ namespace QCD {
 
    _scidacRecord = sr;
 
-   std::cout << GridLogMessage << "Build SciDAC datatype " <<sr.datatype<<std::endl;
+   //   std::cout << GridLogMessage << "Build SciDAC datatype " <<sr.datatype<<std::endl;
  }
  
  ///////////////////////////////////////////////////////
@@ -182,7 +178,7 @@ class GridLimeReader : public BinaryIO {
    /////////////////////////////////////////////
    // Open the file
    /////////////////////////////////////////////
-   void open(std::string &_filename) 
+   void open(const std::string &_filename) 
    {
      filename= _filename;
      File = fopen(filename.c_str(), "r");
@@ -210,24 +206,38 @@ class GridLimeReader : public BinaryIO {
 
     while ( limeReaderNextRecord(LimeR) == LIME_SUCCESS ) { 
 
-      std::cout << GridLogMessage << limeReaderType(LimeR) <<std::endl;
+      uint64_t file_bytes =limeReaderBytes(LimeR);
 
-      if ( strncmp(limeReaderType(LimeR), record_name.c_str(),strlen(record_name.c_str()) )  ) {
+      //      std::cout << GridLogMessage << limeReaderType(LimeR) << " "<< file_bytes <<" bytes "<<std::endl;
+      //      std::cout << GridLogMessage<< " readLimeObject seeking "<<  record_name <<" found record :" <<limeReaderType(LimeR) <<std::endl;
 
+      if ( !strncmp(limeReaderType(LimeR), record_name.c_str(),strlen(record_name.c_str()) )  ) {
 
-	off_t offset= ftell(File);
+	//	std::cout << GridLogMessage<< " readLimeLatticeBinaryObject matches ! " <<std::endl;
+
+	uint64_t PayloadSize = sizeof(sobj) * field._grid->_gsites;
+
+	//	std::cout << "R sizeof(sobj)= " <<sizeof(sobj)<<std::endl;
+	//	std::cout << "R Gsites " <<field._grid->_gsites<<std::endl;
+	//	std::cout << "R Payload expected " <<PayloadSize<<std::endl;
+	//	std::cout << "R file size " <<file_bytes <<std::endl;
+
+	assert(PayloadSize == file_bytes);// Must match or user error
+
+	uint64_t offset= ftello(File);
+	//	std::cout << " ReadLatticeObject from offset "<<offset << std::endl;
 	BinarySimpleMunger<sobj,sobj> munge;
-	BinaryIO::readLatticeObject< sobj, sobj >(field, filename, munge, offset, format,nersc_csum,scidac_csuma,scidac_csumb);
+	BinaryIO::readLatticeObject< vobj, sobj >(field, filename, munge, offset, format,nersc_csum,scidac_csuma,scidac_csumb);
 
 	/////////////////////////////////////////////
 	// Insist checksum is next record
 	/////////////////////////////////////////////
-	readLimeObject(scidacChecksum_,std::string("scidacChecksum"),record_name);
+	readLimeObject(scidacChecksum_,std::string("scidacChecksum"),std::string(SCIDAC_CHECKSUM));
 
 	/////////////////////////////////////////////
 	// Verify checksums
 	/////////////////////////////////////////////
-	scidacChecksumVerify(scidacChecksum_,scidac_csuma,scidac_csumb);
+	assert(scidacChecksumVerify(scidacChecksum_,scidac_csuma,scidac_csumb)==1);
 	return;
       }
     }
@@ -242,11 +252,16 @@ class GridLimeReader : public BinaryIO {
     // should this be a do while; can we miss a first record??
     while ( limeReaderNextRecord(LimeR) == LIME_SUCCESS ) { 
 
+      //      std::cout << GridLogMessage<< " readLimeObject seeking "<< record_name <<" found record :" <<limeReaderType(LimeR) <<std::endl;
       uint64_t nbytes = limeReaderBytes(LimeR);//size of this record (configuration)
 
-      if ( strncmp(limeReaderType(LimeR), record_name.c_str(),strlen(record_name.c_str()) )  ) {
+      if ( !strncmp(limeReaderType(LimeR), record_name.c_str(),strlen(record_name.c_str()) )  ) {
+
+	//	std::cout << GridLogMessage<< " readLimeObject matches ! " << record_name <<std::endl;
 	std::vector<char> xmlc(nbytes+1,'\0');
 	limeReaderReadData((void *)&xmlc[0], &nbytes, LimeR);    
+	//	std::cout << GridLogMessage<< " readLimeObject matches XML " << &xmlc[0] <<std::endl;
+
 	XmlReader RD(&xmlc[0],"");
 	read(RD,object_name,object);
 	return;
@@ -261,13 +276,14 @@ class GridLimeWriter : public BinaryIO {
  public:
    ///////////////////////////////////////////////////
    // FIXME: format for RNG? Now just binary out instead
+   // FIXME: collective calls or not ?
+   //      : must know if I am the I/O boss
    ///////////////////////////////////////////////////
-
    FILE       *File;
    LimeWriter *LimeW;
    std::string filename;
 
-   void open(std::string &_filename) { 
+   void open(const std::string &_filename) { 
      filename= _filename;
      File = fopen(filename.c_str(), "w");
      LimeW = limeCreateWriter(File); assert(LimeW != NULL );
@@ -302,14 +318,18 @@ class GridLimeWriter : public BinaryIO {
       write(WR,object_name,object);
       xmlstring = WR.XmlString();
     }
+    //    std::cout << "WriteLimeObject" << record_name <<std::endl;
     uint64_t nbytes = xmlstring.size();
+    //    std::cout << " xmlstring "<< nbytes<< " " << xmlstring <<std::endl;
     int err;
-    LimeRecordHeader *h = limeCreateHeader(MB, ME,(char *)record_name.c_str(), nbytes); assert(h!= NULL);
+    LimeRecordHeader *h = limeCreateHeader(MB, ME,const_cast<char *>(record_name.c_str()), nbytes); 
+    assert(h!= NULL);
 
     err=limeWriteRecordHeader(h, LimeW);                    assert(err>=0);
     err=limeWriteRecordData(&xmlstring[0], &nbytes, LimeW); assert(err>=0);
     err=limeWriterCloseRecord(LimeW);                       assert(err>=0);
     limeDestroyHeader(h);
+    //    std::cout << " File offset is now"<<ftello(File) << std::endl;
   }
   ////////////////////////////////////////////
   // Write a generic lattice field and csum
@@ -326,6 +346,10 @@ class GridLimeWriter : public BinaryIO {
     uint64_t PayloadSize = sizeof(sobj) * field._grid->_gsites;
     createLimeRecordHeader(record_name, 0, 0, PayloadSize);
 
+    //    std::cout << "W sizeof(sobj)"      <<sizeof(sobj)<<std::endl;
+    //    std::cout << "W Gsites "           <<field._grid->_gsites<<std::endl;
+    //    std::cout << "W Payload expected " <<PayloadSize<<std::endl;
+
     ////////////////////////////////////////////////////////////////////
     // NB: FILE and iostream are jointly writing disjoint sequences in the
     // the same file through different file handles (integer units).
@@ -333,17 +357,20 @@ class GridLimeWriter : public BinaryIO {
     // These are both buffered, so why I think this code is right is as follows.
     //
     // i)  write record header to FILE *File, telegraphing the size. 
-    // ii) ftell reads the offset from FILE *File .
+    // ii) ftello reads the offset from FILE *File .
     // iii) iostream / MPI Open independently seek this offset. Write sequence direct to disk.
     //      Closes iostream and flushes.
     // iv) fseek on FILE * to end of this disjoint section.
     //  v) Continue writing scidac record.
     ////////////////////////////////////////////////////////////////////
-    off_t offset = ftell(File);
+    uint64_t offset = ftello(File);
+    //    std::cout << " Writing to offset "<<offset << std::endl;
     std::string format = getFormatString<vobj>();
     BinarySimpleMunger<sobj,sobj> munge;
     BinaryIO::writeLatticeObject<vobj,sobj>(field, filename, munge, offset, format,nersc_csum,scidac_csuma,scidac_csumb);
+    //    fseek(File,0,SEEK_END);    offset = ftello(File);std::cout << " offset now "<<offset << std::endl;
     err=limeWriterCloseRecord(LimeW);  assert(err>=0);
+
     ////////////////////////////////////////
     // Write checksum element, propagaing forward from the BinaryIO
     // Always pair a checksum with a binary object, and close message
@@ -353,8 +380,8 @@ class GridLimeWriter : public BinaryIO {
     std::stringstream streamb; streamb << std::hex << scidac_csumb;
     checksum.suma= streama.str();
     checksum.sumb= streamb.str();
-    std::cout << GridLogMessage<<" writing scidac checksums "<<std::hex<<scidac_csuma<<"/"<<scidac_csumb<<std::dec<<std::endl;
-    writeLimeObject(0,1,checksum,std::string("scidacChecksum"    ),std::string(SCIDAC_CHECKSUM));
+    //    std::cout << GridLogMessage<<" writing scidac checksums "<<std::hex<<scidac_csuma<<"/"<<scidac_csumb<<std::dec<<std::endl;
+    writeLimeObject(0,1,checksum,std::string("scidacChecksum"),std::string(SCIDAC_CHECKSUM));
   }
 };
 
@@ -374,8 +401,6 @@ class ScidacWriter : public GridLimeWriter {
   template <class vobj, class userRecord>
   void writeScidacFieldRecord(Lattice<vobj> &field,userRecord _userRecord) 
   {
-    typedef typename vobj::scalar_object sobj;
-    uint64_t nbytes;
     GridBase * grid = field._grid;
 
     ////////////////////////////////////////
@@ -397,6 +422,66 @@ class ScidacWriter : public GridLimeWriter {
   }
 };
 
+
+class ScidacReader : public GridLimeReader {
+ public:
+
+   template<class SerialisableUserFile>
+   void readScidacFileRecord(GridBase *grid,SerialisableUserFile &_userFile)
+   {
+     scidacFile    _scidacFile(grid);
+     readLimeObject(_scidacFile,_scidacFile.SerialisableClassName(),std::string(SCIDAC_PRIVATE_FILE_XML));
+     readLimeObject(_userFile,_userFile.SerialisableClassName(),std::string(SCIDAC_FILE_XML));
+   }
+  ////////////////////////////////////////////////
+  // Write generic lattice field in scidac format
+  ////////////////////////////////////////////////
+  template <class vobj, class userRecord>
+  void readScidacFieldRecord(Lattice<vobj> &field,userRecord &_userRecord) 
+  {
+    typedef typename vobj::scalar_object sobj;
+    GridBase * grid = field._grid;
+
+    ////////////////////////////////////////
+    // fill the Grid header
+    ////////////////////////////////////////
+    FieldMetaData header;
+    scidacRecord  _scidacRecord;
+    scidacFile    _scidacFile;
+
+    //////////////////////////////////////////////
+    // Fill the Lime file record by record
+    //////////////////////////////////////////////
+    readLimeObject(header ,std::string("FieldMetaData"),std::string(GRID_FORMAT)); // Open message 
+    readLimeObject(_userRecord,_userRecord.SerialisableClassName(),std::string(SCIDAC_RECORD_XML));
+    readLimeObject(_scidacRecord,_scidacRecord.SerialisableClassName(),std::string(SCIDAC_PRIVATE_RECORD_XML));
+    readLimeLatticeBinaryObject(field,std::string(ILDG_BINARY_DATA));
+  }
+  void skipPastBinaryRecord(void) {
+    std::string rec_name(ILDG_BINARY_DATA);
+    while ( limeReaderNextRecord(LimeR) == LIME_SUCCESS ) { 
+      if ( !strncmp(limeReaderType(LimeR), rec_name.c_str(),strlen(rec_name.c_str()) )  ) {
+	skipPastObjectRecord(std::string(SCIDAC_CHECKSUM));
+	return;
+      }
+    }    
+  }
+  void skipPastObjectRecord(std::string rec_name) {
+    while ( limeReaderNextRecord(LimeR) == LIME_SUCCESS ) { 
+      if ( !strncmp(limeReaderType(LimeR), rec_name.c_str(),strlen(rec_name.c_str()) )  ) {
+	return;
+      }
+    }
+  }
+  void skipScidacFieldRecord() {
+    skipPastObjectRecord(std::string(GRID_FORMAT));
+    skipPastObjectRecord(std::string(SCIDAC_RECORD_XML));
+    skipPastObjectRecord(std::string(SCIDAC_PRIVATE_RECORD_XML));
+    skipPastBinaryRecord();
+  }
+};
+
+
 class IldgWriter : public ScidacWriter {
  public:
 
@@ -425,8 +510,6 @@ class IldgWriter : public ScidacWriter {
     typedef iLorentzColourMatrix<vsimd> vobj;
     typedef typename vobj::scalar_object sobj;
 
-    uint64_t nbytes;
-
     ////////////////////////////////////////
     // fill the Grid header
     ////////////////////////////////////////
@@ -557,7 +640,7 @@ class IldgReader : public GridLimeReader {
 	// Copy out the string
 	std::vector<char> xmlc(nbytes+1,'\0');
 	limeReaderReadData((void *)&xmlc[0], &nbytes, LimeR);    
-	std::cout << GridLogMessage<< "Non binary record :" <<limeReaderType(LimeR) <<std::endl; //<<"\n"<<(&xmlc[0])<<std::endl;
+	//	std::cout << GridLogMessage<< "Non binary record :" <<limeReaderType(LimeR) <<std::endl; //<<"\n"<<(&xmlc[0])<<std::endl;
 
 	//////////////////////////////////
 	// ILDG format record
@@ -601,7 +684,7 @@ class IldgReader : public GridLimeReader {
 	  std::string xmls(&xmlc[0]);
 	  // is it a USQCD info field
 	  if ( xmls.find(std::string("usqcdInfo")) != std::string::npos ) { 
-	    std::cout << GridLogMessage<<"...found a usqcdInfo field"<<std::endl;
+	    //	    std::cout << GridLogMessage<<"...found a usqcdInfo field"<<std::endl;
 	    XmlReader RD(&xmlc[0],"");
 	    read(RD,"usqcdInfo",usqcdInfo_);
 	    found_usqcdInfo = 1;
@@ -619,8 +702,7 @@ class IldgReader : public GridLimeReader {
 	// Binary data
 	/////////////////////////////////
 	std::cout << GridLogMessage << "ILDG Binary record found : "  ILDG_BINARY_DATA << std::endl;
-	off_t offset= ftell(File);
-
+	uint64_t offset= ftello(File);
 	if ( format == std::string("IEEE64BIG") ) {
 	  GaugeSimpleMunger<dobj, sobj> munge;
 	  BinaryIO::readLatticeObject< vobj, dobj >(Umu, filename, munge, offset, format,nersc_csum,scidac_csuma,scidac_csumb);

lib/parallelIO/IldgIOtypes.h

@@ -64,6 +64,11 @@ namespace Grid {
 // file compatability, so should be correct to assume the undocumented but defacto file structure.
 /////////////////////////////////////////////////////////////////////////////////
 
+struct emptyUserRecord : Serializable { 
+  GRID_SERIALIZABLE_CLASS_MEMBERS(emptyUserRecord,int,dummy);
+  emptyUserRecord() { dummy=0; };
+};
+
 ////////////////////////
 // Scidac private file xml
 // <?xml version="1.0" encoding="UTF-8"?><scidacFile><version>1.1</version><spacetime>4</spacetime><dims>16 16 16 32 </dims><volfmt>0</volfmt></scidacFile>

lib/parallelIO/MetaData.h

@@ -85,6 +85,9 @@ namespace Grid {
 	nd=4;
 	dimension.resize(4);
 	boundary.resize(4);
+	scidac_checksuma=0;
+	scidac_checksumb=0;
+	checksum=0;
       }
     };
 
@@ -104,6 +107,7 @@ namespace Grid {
       header.nd = nd;
       header.dimension.resize(nd);
       header.boundary.resize(nd);
+      header.data_start = 0;
       for(int d=0;d<nd;d++) {
 	header.dimension[d] = grid->_fdimensions[d];
       }

lib/perfmon/PerfCount.cc

@@ -40,7 +40,7 @@ const PerformanceCounter::PerformanceCounterConfig PerformanceCounter::Performan
   { PERF_TYPE_HARDWARE, PERF_COUNT_HW_CPU_CYCLES          ,  "CPUCYCLES.........." , INSTRUCTIONS},
   { PERF_TYPE_HARDWARE, PERF_COUNT_HW_INSTRUCTIONS        ,  "INSTRUCTIONS......." , CPUCYCLES   },
     // 4
-#ifdef AVX512
+#ifdef KNL
     { PERF_TYPE_RAW, RawConfig(0x40,0x04), "ALL_LOADS..........", CPUCYCLES    },
     { PERF_TYPE_RAW, RawConfig(0x01,0x04), "L1_MISS_LOADS......", L1D_READ_ACCESS  },
     { PERF_TYPE_RAW, RawConfig(0x40,0x04), "ALL_LOADS..........", L1D_READ_ACCESS    },

lib/qcd/action/fermion/AbstractEOFAFermion.h

@@ -0,0 +1,100 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/AbstractEOFAFermion.h
+
+Copyright (C) 2017
+
+Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: David Murphy <dmurphy@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_QCD_ABSTRACT_EOFA_FERMION_H
+#define  GRID_QCD_ABSTRACT_EOFA_FERMION_H
+
+#include <Grid/qcd/action/fermion/CayleyFermion5D.h>
+
+namespace Grid {
+namespace QCD {
+
+  // DJM: Abstract base class for EOFA fermion types.
+  // Defines layout of additional EOFA-specific parameters and operators.
+  // Use to construct EOFA pseudofermion actions that are agnostic to
+  // Shamir / Mobius / etc., and ensure that no one can construct EOFA
+  // pseudofermion action with non-EOFA fermion type.
+  template<class Impl>
+  class AbstractEOFAFermion : public CayleyFermion5D<Impl> {
+    public:
+      INHERIT_IMPL_TYPES(Impl);
+
+    public:
+      // Fermion operator: D(mq1) + shift*\gamma_{5}*R_{5}*\Delta_{\pm}(mq2,mq3)*P_{\pm}
+      RealD mq1;
+      RealD mq2;
+      RealD mq3;
+      RealD shift;
+      int pm;
+
+      RealD alpha; // Mobius scale
+      RealD k;     // EOFA normalization constant
+
+      virtual void Instantiatable(void) = 0;
+
+      // EOFA-specific operations
+      // Force user to implement in derived classes
+      virtual void  Omega    (const FermionField& in, FermionField& out, int sign, int dag) = 0;
+      virtual void  Dtilde   (const FermionField& in, FermionField& out) = 0;
+      virtual void  DtildeInv(const FermionField& in, FermionField& out) = 0;
+
+      // Implement derivatives in base class:
+      // for EOFA both DWF and Mobius just need d(Dw)/dU
+      virtual void MDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag){
+        this->DhopDeriv(mat, U, V, dag);
+      };
+      virtual void MoeDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag){
+        this->DhopDerivOE(mat, U, V, dag);
+      };
+      virtual void MeoDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag){
+        this->DhopDerivEO(mat, U, V, dag);
+      };
+
+      // Recompute 5D coefficients for different value of shift constant
+      // (needed for heatbath loop over poles)
+      virtual void RefreshShiftCoefficients(RealD new_shift) = 0;
+
+      // Constructors
+      AbstractEOFAFermion(GaugeField& _Umu, GridCartesian& FiveDimGrid, GridRedBlackCartesian& FiveDimRedBlackGrid,
+        GridCartesian& FourDimGrid, GridRedBlackCartesian& FourDimRedBlackGrid,
+        RealD _mq1, RealD _mq2, RealD _mq3, RealD _shift, int _pm,
+        RealD _M5, RealD _b, RealD _c, const ImplParams& p=ImplParams())
+        : CayleyFermion5D<Impl>(_Umu, FiveDimGrid, FiveDimRedBlackGrid, FourDimGrid, FourDimRedBlackGrid,
+          _mq1, _M5, p), mq1(_mq1), mq2(_mq2), mq3(_mq3), shift(_shift), pm(_pm)
+      {
+        int Ls = this->Ls;
+        this->alpha = _b + _c;
+        this->k = this->alpha * (_mq3-_mq2) * std::pow(this->alpha+1.0,2*Ls) /
+                    ( std::pow(this->alpha+1.0,Ls) + _mq2*std::pow(this->alpha-1.0,Ls) ) /
+                    ( std::pow(this->alpha+1.0,Ls) + _mq3*std::pow(this->alpha-1.0,Ls) );
+      };
+  };
+}}
+
+#endif

lib/qcd/action/fermion/CayleyFermion5D.cc

@@ -77,7 +77,6 @@ void CayleyFermion5D<Impl>::DminusDag(const FermionField &psi, FermionField &chi
   }
 }
 
-
 template<class Impl> void CayleyFermion5D<Impl>::CayleyReport(void)
 {
   this->Report();
@@ -119,7 +118,6 @@ template<class Impl> void CayleyFermion5D<Impl>::CayleyZeroCounters(void)
   MooeeInvTime=0;
 }
 
-
 template<class Impl>  
 void CayleyFermion5D<Impl>::M5D   (const FermionField &psi, FermionField &chi)
 {
@@ -414,7 +412,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,std::vector<Co
   for(int i=0; i < Ls; i++){
     as[i] = 1.0;
     omega[i] = gamma[i]*zolo_hi; //NB reciprocal relative to Chroma NEF code
-    //    assert(fabs(omega[i])>0.0);
+    assert(omega[i]!=Coeff_t(0.0));
     bs[i] = 0.5*(bpc/omega[i] + bmc);
     cs[i] = 0.5*(bpc/omega[i] - bmc);
   }
@@ -429,7 +427,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,std::vector<Co
   
   for(int i=0;i<Ls;i++){
     bee[i]=as[i]*(bs[i]*(4.0-this->M5) +1.0);     
-    //    assert(fabs(bee[i])>0.0);
+    assert(bee[i]!=Coeff_t(0.0));
     cee[i]=as[i]*(1.0-cs[i]*(4.0-this->M5));
     beo[i]=as[i]*bs[i];
     ceo[i]=-as[i]*cs[i];
@@ -456,10 +454,16 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,std::vector<Co
     
     if ( i < Ls-1 ) {
 
+      assert(bee[i]!=Coeff_t(0.0));
+      assert(bee[0]!=Coeff_t(0.0));
+      
       lee[i] =-cee[i+1]/bee[i]; // sub-diag entry on the ith column
       
       leem[i]=mass*cee[Ls-1]/bee[0];
-      for(int j=0;j<i;j++)  leem[i]*= aee[j]/bee[j+1];
+      for(int j=0;j<i;j++) {
+	assert(bee[j+1]!=Coeff_t(0.0));
+	leem[i]*= aee[j]/bee[j+1];
+      }
       
       uee[i] =-aee[i]/bee[i];   // up-diag entry on the ith row
       
@@ -478,7 +482,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,std::vector<Co
   { 
     Coeff_t delta_d=mass*cee[Ls-1];
     for(int j=0;j<Ls-1;j++) {
-      //      assert(fabs(bee[j])>0.0);
+      assert(bee[j] != Coeff_t(0.0));
       delta_d *= cee[j]/bee[j];
     }
     dee[Ls-1] += delta_d;

lib/qcd/action/fermion/CayleyFermion5D.h

@@ -179,9 +179,9 @@ namespace Grid {
      double MooeeInvTime;
 
     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);
+      virtual void SetCoefficientsZolotarev(RealD zolohi,Approx::zolotarev_data *zdata,RealD b,RealD c);
+      virtual void SetCoefficientsTanh(Approx::zolotarev_data *zdata,RealD b,RealD c);
+      virtual void SetCoefficientsInternal(RealD zolo_hi,std::vector<Coeff_t> & gamma,RealD b,RealD c);
     };
 
   }

lib/qcd/action/fermion/DomainWallEOFAFermion.cc

@@ -0,0 +1,438 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/DomainWallEOFAFermion.cc
+
+Copyright (C) 2017
+
+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>
+Author: David Murphy <dmurphy@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 */
+
+#include <Grid/Grid_Eigen_Dense.h>
+#include <Grid/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/DomainWallEOFAFermion.h>
+
+namespace Grid {
+namespace QCD {
+
+    template<class Impl>
+    DomainWallEOFAFermion<Impl>::DomainWallEOFAFermion(
+      GaugeField            &_Umu,
+      GridCartesian         &FiveDimGrid,
+      GridRedBlackCartesian &FiveDimRedBlackGrid,
+      GridCartesian         &FourDimGrid,
+      GridRedBlackCartesian &FourDimRedBlackGrid,
+      RealD _mq1, RealD _mq2, RealD _mq3,
+      RealD _shift, int _pm, RealD _M5, const ImplParams &p) :
+    AbstractEOFAFermion<Impl>(_Umu, FiveDimGrid, FiveDimRedBlackGrid,
+        FourDimGrid, FourDimRedBlackGrid, _mq1, _mq2, _mq3,
+        _shift, _pm, _M5, 1.0, 0.0, p)
+    {
+        RealD eps = 1.0;
+        Approx::zolotarev_data *zdata = Approx::higham(eps,this->Ls);
+        assert(zdata->n == this->Ls);
+
+        std::cout << GridLogMessage << "DomainWallEOFAFermion with Ls=" << this->Ls << std::endl;
+        this->SetCoefficientsTanh(zdata, 1.0, 0.0);
+
+        Approx::zolotarev_free(zdata);
+    }
+
+    /***************************************************************
+     * Additional EOFA operators only called outside the inverter.
+     * Since speed is not essential, simple axpby-style
+     * implementations should be fine.
+     ***************************************************************/
+    template<class Impl>
+    void DomainWallEOFAFermion<Impl>::Omega(const FermionField& psi, FermionField& Din, int sign, int dag)
+    {
+        int Ls = this->Ls;
+
+        Din = zero;
+        if((sign == 1) && (dag == 0)){ axpby_ssp(Din, 0.0, psi, 1.0, psi, Ls-1, 0); }
+        else if((sign == -1) && (dag == 0)){ axpby_ssp(Din, 0.0, psi, 1.0, psi, 0, 0); }
+        else if((sign == 1 ) && (dag == 1)){ axpby_ssp(Din, 0.0, psi, 1.0, psi, 0, Ls-1); }
+        else if((sign == -1) && (dag == 1)){ axpby_ssp(Din, 0.0, psi, 1.0, psi, 0, 0); }
+    }
+
+    // This is just the identity for DWF
+    template<class Impl>
+    void DomainWallEOFAFermion<Impl>::Dtilde(const FermionField& psi, FermionField& chi){ chi = psi; }
+
+    // This is just the identity for DWF
+    template<class Impl>
+    void DomainWallEOFAFermion<Impl>::DtildeInv(const FermionField& psi, FermionField& chi){ chi = psi; }
+
+    /*****************************************************************************************************/
+
+    template<class Impl>
+    RealD DomainWallEOFAFermion<Impl>::M(const FermionField& psi, FermionField& chi)
+    {
+        int Ls = this->Ls;
+
+        FermionField Din(psi._grid);
+
+        this->Meooe5D(psi, Din);
+        this->DW(Din, chi, DaggerNo);
+        axpby(chi, 1.0, 1.0, chi, psi);
+        this->M5D(psi, chi);
+        return(norm2(chi));
+    }
+
+    template<class Impl>
+    RealD DomainWallEOFAFermion<Impl>::Mdag(const FermionField& psi, FermionField& chi)
+    {
+        int Ls = this->Ls;
+
+        FermionField Din(psi._grid);
+
+        this->DW(psi, Din, DaggerYes);
+        this->MeooeDag5D(Din, chi);
+        this->M5Ddag(psi, chi);
+        axpby(chi, 1.0, 1.0, chi, psi);
+        return(norm2(chi));
+    }
+
+    /********************************************************************
+     * Performance critical fermion operators called inside the inverter
+     ********************************************************************/
+
+    template<class Impl>
+    void DomainWallEOFAFermion<Impl>::M5D(const FermionField& psi, FermionField& chi)
+    {
+        int   Ls    = this->Ls;
+        int   pm    = this->pm;
+        RealD shift = this->shift;
+        RealD mq1   = this->mq1;
+        RealD mq2   = this->mq2;
+        RealD mq3   = this->mq3;
+
+        // coefficients for shift operator ( = shift*\gamma_{5}*R_{5}*\Delta_{\pm}(mq2,mq3)*P_{\pm} )
+        Coeff_t shiftp(0.0), shiftm(0.0);
+        if(shift != 0.0){
+          if(pm == 1){ shiftp = shift*(mq3-mq2); }
+          else{ shiftm = -shift*(mq3-mq2); }
+        }
+
+        std::vector<Coeff_t> diag(Ls,1.0);
+        std::vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1] = mq1 + shiftm;
+        std::vector<Coeff_t> lower(Ls,-1.0); lower[0]    = mq1 + shiftp;
+
+        #if(0)
+            std::cout << GridLogMessage << "DomainWallEOFAFermion::M5D(FF&,FF&):" << std::endl;
+            for(int i=0; i<diag.size(); ++i){
+                std::cout << GridLogMessage << "diag[" << i << "] =" << diag[i] << std::endl;
+            }
+            for(int i=0; i<upper.size(); ++i){
+                std::cout << GridLogMessage << "upper[" << i << "] =" << upper[i] << std::endl;
+            }
+            for(int i=0; i<lower.size(); ++i){
+                std::cout << GridLogMessage << "lower[" << i << "] =" << lower[i] << std::endl;
+            }
+        #endif
+
+        this->M5D(psi, chi, chi, lower, diag, upper);
+    }
+
+    template<class Impl>
+    void DomainWallEOFAFermion<Impl>::M5Ddag(const FermionField& psi, FermionField& chi)
+    {
+        int   Ls    = this->Ls;
+        int   pm    = this->pm;
+        RealD shift = this->shift;
+        RealD mq1   = this->mq1;
+        RealD mq2   = this->mq2;
+        RealD mq3   = this->mq3;
+
+        // coefficients for shift operator ( = shift*\gamma_{5}*R_{5}*\Delta_{\pm}(mq2,mq3)*P_{\pm} )
+        Coeff_t shiftp(0.0), shiftm(0.0);
+        if(shift != 0.0){
+          if(pm == 1){ shiftp = shift*(mq3-mq2); }
+          else{ shiftm = -shift*(mq3-mq2); }
+        }
+
+        std::vector<Coeff_t> diag(Ls,1.0);
+        std::vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1] = mq1 + shiftp;
+        std::vector<Coeff_t> lower(Ls,-1.0); lower[0]    = mq1 + shiftm;
+
+        #if(0)
+            std::cout << GridLogMessage << "DomainWallEOFAFermion::M5Ddag(FF&,FF&):" << std::endl;
+            for(int i=0; i<diag.size(); ++i){
+                std::cout << GridLogMessage << "diag[" << i << "] =" << diag[i] << std::endl;
+            }
+            for(int i=0; i<upper.size(); ++i){
+                std::cout << GridLogMessage << "upper[" << i << "] =" << upper[i] << std::endl;
+            }
+            for(int i=0; i<lower.size(); ++i){
+                std::cout << GridLogMessage << "lower[" << i << "] =" << lower[i] << std::endl;
+            }
+        #endif
+
+        this->M5Ddag(psi, chi, chi, lower, diag, upper);
+    }
+
+    // half checkerboard operations
+    template<class Impl>
+    void DomainWallEOFAFermion<Impl>::Mooee(const FermionField& psi, FermionField& chi)
+    {
+        int Ls = this->Ls;
+
+        std::vector<Coeff_t> diag = this->bee;
+        std::vector<Coeff_t> upper(Ls);
+        std::vector<Coeff_t> lower(Ls);
+
+        for(int s=0; s<Ls; s++){
+          upper[s] = -this->cee[s];
+          lower[s] = -this->cee[s];
+        }
+        upper[Ls-1] = this->dm;
+        lower[0]    = this->dp;
+
+        this->M5D(psi, psi, chi, lower, diag, upper);
+    }
+
+    template<class Impl>
+    void DomainWallEOFAFermion<Impl>::MooeeDag(const FermionField& psi, FermionField& chi)
+    {
+        int Ls = this->Ls;
+
+        std::vector<Coeff_t> diag = this->bee;
+        std::vector<Coeff_t> upper(Ls);
+        std::vector<Coeff_t> lower(Ls);
+
+        for(int s=0; s<Ls; s++){
+          upper[s] = -this->cee[s];
+          lower[s] = -this->cee[s];
+        }
+        upper[Ls-1] = this->dp;
+        lower[0]    = this->dm;
+
+        this->M5Ddag(psi, psi, chi, lower, diag, upper);
+    }
+
+    /****************************************************************************************/
+
+    //Zolo
+    template<class Impl>
+    void DomainWallEOFAFermion<Impl>::SetCoefficientsInternal(RealD zolo_hi, std::vector<Coeff_t>& gamma, RealD b, RealD c)
+    {
+        int   Ls    = this->Ls;
+        int   pm    = this->pm;
+        RealD mq1   = this->mq1;
+        RealD mq2   = this->mq2;
+        RealD mq3   = this->mq3;
+        RealD shift = this->shift;
+
+        ////////////////////////////////////////////////////////
+        // Constants for the preconditioned matrix Cayley form
+        ////////////////////////////////////////////////////////
+        this->bs.resize(Ls);
+        this->cs.resize(Ls);
+        this->aee.resize(Ls);
+        this->aeo.resize(Ls);
+        this->bee.resize(Ls);
+        this->beo.resize(Ls);
+        this->cee.resize(Ls);
+        this->ceo.resize(Ls);
+
+        for(int i=0; i<Ls; ++i){
+          this->bee[i] = 4.0 - this->M5 + 1.0;
+          this->cee[i] = 1.0;
+        }
+
+        for(int i=0; i<Ls; ++i){
+          this->aee[i] = this->cee[i];
+          this->bs[i] = this->beo[i] = 1.0;
+          this->cs[i] = this->ceo[i] = 0.0;
+        }
+
+        //////////////////////////////////////////
+        // EOFA shift terms
+        //////////////////////////////////////////
+        if(pm == 1){
+          this->dp = mq1*this->cee[0] + shift*(mq3-mq2);
+          this->dm = mq1*this->cee[Ls-1];
+        } else if(this->pm == -1) {
+          this->dp = mq1*this->cee[0];
+          this->dm = mq1*this->cee[Ls-1] - shift*(mq3-mq2);
+        } else {
+          this->dp = mq1*this->cee[0];
+          this->dm = mq1*this->cee[Ls-1];
+        }
+
+        //////////////////////////////////////////
+        // LDU decomposition of eeoo
+        //////////////////////////////////////////
+        this->dee.resize(Ls+1);
+        this->lee.resize(Ls);
+        this->leem.resize(Ls);
+        this->uee.resize(Ls);
+        this->ueem.resize(Ls);
+
+        for(int i=0; i<Ls; ++i){
+
+          if(i < Ls-1){
+
+            this->lee[i] = -this->cee[i+1]/this->bee[i]; // sub-diag entry on the ith column
+
+            this->leem[i] = this->dm/this->bee[i];
+            for(int j=0; j<i; j++){ this->leem[i] *= this->aee[j]/this->bee[j]; }
+
+            this->dee[i] = this->bee[i];
+
+            this->uee[i] = -this->aee[i]/this->bee[i];   // up-diag entry on the ith row
+
+            this->ueem[i] = this->dp / this->bee[0];
+            for(int j=1; j<=i; j++){ this->ueem[i] *= this->cee[j]/this->bee[j]; }
+
+          } else {
+
+            this->lee[i]  = 0.0;
+            this->leem[i] = 0.0;
+            this->uee[i]  = 0.0;
+            this->ueem[i] = 0.0;
+
+          }
+        }
+
+        {
+          Coeff_t delta_d = 1.0 / this->bee[0];
+          for(int j=1; j<Ls-1; j++){ delta_d *= this->cee[j] / this->bee[j]; }
+          this->dee[Ls-1] = this->bee[Ls-1] + this->cee[0] * this->dm * delta_d;
+          this->dee[Ls] = this->bee[Ls-1] + this->cee[Ls-1] * this->dp * delta_d;
+        }
+
+        int inv = 1;
+        this->MooeeInternalCompute(0, inv, this->MatpInv, this->MatmInv);
+        this->MooeeInternalCompute(1, inv, this->MatpInvDag, this->MatmInvDag);
+    }
+
+    // Recompute Cayley-form coefficients for different shift
+    template<class Impl>
+    void DomainWallEOFAFermion<Impl>::RefreshShiftCoefficients(RealD new_shift)
+    {
+        this->shift = new_shift;
+        Approx::zolotarev_data *zdata = Approx::higham(1.0, this->Ls);
+        this->SetCoefficientsTanh(zdata, 1.0, 0.0);
+    }
+
+    template<class Impl>
+    void DomainWallEOFAFermion<Impl>::MooeeInternalCompute(int dag, int inv,
+        Vector<iSinglet<Simd> >& Matp, Vector<iSinglet<Simd> >& Matm)
+    {
+        int Ls = this->Ls;
+
+        GridBase* grid = this->FermionRedBlackGrid();
+        int LLs = grid->_rdimensions[0];
+
+        if(LLs == Ls){ return; } // Not vectorised in 5th direction
+
+        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)  = this->bee[s];
+            Pminus(s,s) = this->bee[s];
+        }
+
+        for(int s=0; s<Ls-1; s++){
+            Pminus(s,s+1) = -this->cee[s];
+        }
+
+        for(int s=0; s<Ls-1; s++){
+            Pplus(s+1,s) = -this->cee[s+1];
+        }
+
+        Pplus (0,Ls-1) = this->dp;
+        Pminus(Ls-1,0) = this->dm;
+
+        Eigen::MatrixXcd PplusMat ;
+        Eigen::MatrixXcd PminusMat;
+
+        #if(0)
+            std::cout << GridLogMessage << "Pplus:" << std::endl;
+            for(int s=0; s<Ls; ++s){
+                for(int ss=0; ss<Ls; ++ss){
+                    std::cout << Pplus(s,ss) << "\t";
+                }
+                std::cout << std::endl;
+            }
+            std::cout << GridLogMessage << "Pminus:" << std::endl;
+            for(int s=0; s<Ls; ++s){
+                for(int ss=0; ss<Ls; ++ss){
+                    std::cout << Pminus(s,ss) << "\t";
+                }
+                std::cout << std::endl;
+            }
+        #endif
+
+        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();
+        Matp.resize(Ls*LLs);
+        Matm.resize(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++){
+                if(switcheroo<Coeff_t>::iscomplex()) {
+                    sp[l] = PplusMat (l*istride+s1*ostride,s2);
+                    sm[l] = PminusMat(l*istride+s1*ostride,s2);
+                } else {
+                    // if real
+                    scalar_type tmp;
+                    tmp = PplusMat (l*istride+s1*ostride,s2);
+                    sp[l] = scalar_type(tmp.real(),tmp.real());
+                    tmp = PminusMat(l*istride+s1*ostride,s2);
+                    sm[l] = scalar_type(tmp.real(),tmp.real());
+                }
+            }
+            Matp[LLs*s2+s1] = Vp;
+            Matm[LLs*s2+s1] = Vm;
+        }}
+    }
+
+    FermOpTemplateInstantiate(DomainWallEOFAFermion);
+    GparityFermOpTemplateInstantiate(DomainWallEOFAFermion);
+
+}}

lib/qcd/action/fermion/DomainWallEOFAFermion.h

@@ -0,0 +1,115 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/DomainWallEOFAFermion.h
+
+Copyright (C) 2017
+
+Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: David Murphy <dmurphy@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_QCD_DOMAIN_WALL_EOFA_FERMION_H
+#define  GRID_QCD_DOMAIN_WALL_EOFA_FERMION_H
+
+#include <Grid/qcd/action/fermion/AbstractEOFAFermion.h>
+
+namespace Grid {
+namespace QCD {
+
+  template<class Impl>
+  class DomainWallEOFAFermion : public AbstractEOFAFermion<Impl>
+  {
+    public:
+      INHERIT_IMPL_TYPES(Impl);
+
+    public:
+      // Modified (0,Ls-1) and (Ls-1,0) elements of Mooee
+      // for red-black preconditioned Shamir EOFA
+      Coeff_t dm;
+      Coeff_t dp;
+
+      virtual void Instantiatable(void) {};
+
+      // EOFA-specific operations
+      virtual void  Omega      (const FermionField& in, FermionField& out, int sign, int dag);
+      virtual void  Dtilde     (const FermionField& in, FermionField& out);
+      virtual void  DtildeInv  (const FermionField& in, FermionField& out);
+
+      // override multiply
+      virtual RealD M          (const FermionField& in, FermionField& out);
+      virtual RealD Mdag       (const FermionField& in, FermionField& out);
+
+      // half checkerboard operations
+      virtual void  Mooee      (const FermionField& in, FermionField& out);
+      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   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);
+
+      void MooeeInternalCompute(int dag, int inv, Vector<iSinglet<Simd>>& Matp, Vector<iSinglet<Simd>>& Matm);
+
+      void MooeeInternalAsm(const FermionField& in, FermionField& out, int LLs, int site,
+        Vector<iSinglet<Simd>>& Matp, Vector<iSinglet<Simd>>& Matm);
+
+      void MooeeInternalZAsm(const FermionField& in, FermionField& out, int LLs, int site,
+        Vector<iSinglet<Simd>>& Matp, Vector<iSinglet<Simd>>& Matm);
+
+      virtual void RefreshShiftCoefficients(RealD new_shift);
+
+      // Constructors
+      DomainWallEOFAFermion(GaugeField& _Umu, GridCartesian& FiveDimGrid, GridRedBlackCartesian& FiveDimRedBlackGrid,
+        GridCartesian& FourDimGrid, GridRedBlackCartesian& FourDimRedBlackGrid,
+        RealD _mq1, RealD _mq2, RealD _mq3, RealD _shift, int pm,
+        RealD _M5, const ImplParams& p=ImplParams());
+
+    protected:
+      void SetCoefficientsInternal(RealD zolo_hi, std::vector<Coeff_t>& gamma, RealD b, RealD c);
+  };
+}}
+
+#define INSTANTIATE_DPERP_DWF_EOFA(A)\
+template void DomainWallEOFAFermion<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 DomainWallEOFAFermion<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 DomainWallEOFAFermion<A>::MooeeInv(const FermionField& psi, FermionField& chi); \
+template void DomainWallEOFAFermion<A>::MooeeInvDag(const FermionField& psi, FermionField& chi);
+
+#undef  DOMAIN_WALL_EOFA_DPERP_DENSE
+#define DOMAIN_WALL_EOFA_DPERP_CACHE
+#undef  DOMAIN_WALL_EOFA_DPERP_LINALG
+#define DOMAIN_WALL_EOFA_DPERP_VEC
+
+#endif

lib/qcd/action/fermion/DomainWallEOFAFermioncache.cc

@@ -0,0 +1,248 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/DomainWallEOFAFermioncache.cc
+
+Copyright (C) 2017
+
+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>
+Author: David Murphy <dmurphy@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 */
+
+#include <Grid/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/DomainWallEOFAFermion.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 DomainWallEOFAFermion<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;
+        // Flops = 6.0*(Nc*Ns) *Ls*vol
+        this->M5Dcalls++;
+        this->M5Dtime -= usecond();
+
+        parallel_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;
+                }
+            }
+        }
+
+        this->M5Dtime += usecond();
+    }
+
+    template<class Impl>
+    void DomainWallEOFAFermion<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;
+
+        // Flops = 6.0*(Nc*Ns) *Ls*vol
+        this->M5Dcalls++;
+        this->M5Dtime -= usecond();
+
+        parallel_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;
+                }
+            }
+        }
+
+        this->M5Dtime += usecond();
+    }
+
+    template<class Impl>
+    void DomainWallEOFAFermion<Impl>::MooeeInv(const FermionField& psi, FermionField& chi)
+    {
+        GridBase* grid = psi._grid;
+        int Ls = this->Ls;
+
+        chi.checkerboard = psi.checkerboard;
+
+        this->MooeeInvCalls++;
+        this->MooeeInvTime -= usecond();
+
+        parallel_for(int ss=0; ss<grid->oSites(); ss+=Ls){ // adds Ls
+
+            auto tmp1 = psi._odata[0];
+            auto tmp2 = psi._odata[0];
+
+            // flops = 12*2*Ls + 12*2*Ls + 3*12*Ls + 12*2*Ls  = 12*Ls * (9) = 108*Ls flops
+            // Apply (L^{\prime})^{-1}
+            chi[ss] = psi[ss]; // chi[0]=psi[0]
+            for(int s=1; s<Ls; s++){
+                spProj5p(tmp1, chi[ss+s-1]);
+                chi[ss+s] = psi[ss+s] - this->lee[s-1]*tmp1;
+            }
+
+            // L_m^{-1}
+            for(int s=0; s<Ls-1; s++){ // Chi[ee] = 1 - sum[s<Ls-1] -leem[s]P_- chi
+                spProj5m(tmp1, chi[ss+s]);
+                chi[ss+Ls-1] = chi[ss+Ls-1] - this->leem[s]*tmp1;
+            }
+
+            // U_m^{-1} D^{-1}
+            for(int s=0; s<Ls-1; s++){ // Chi[s] + 1/d chi[s]
+                spProj5p(tmp1, chi[ss+Ls-1]);
+                chi[ss+s] = (1.0/this->dee[s])*chi[ss+s] - (this->ueem[s]/this->dee[Ls])*tmp1;
+            }
+            spProj5m(tmp2, chi[ss+Ls-1]);
+            chi[ss+Ls-1] = (1.0/this->dee[Ls])*tmp1 + (1.0/this->dee[Ls-1])*tmp2;
+
+            // Apply U^{-1}
+            for(int s=Ls-2; s>=0; s--){
+                spProj5m(tmp1, chi[ss+s+1]);
+                chi[ss+s] = chi[ss+s] - this->uee[s]*tmp1;
+            }
+        }
+
+        this->MooeeInvTime += usecond();
+    }
+
+    template<class Impl>
+    void DomainWallEOFAFermion<Impl>::MooeeInvDag(const FermionField& psi, FermionField& chi)
+    {
+        GridBase* grid = psi._grid;
+        int Ls = this->Ls;
+
+        assert(psi.checkerboard == psi.checkerboard);
+        chi.checkerboard = psi.checkerboard;
+
+        std::vector<Coeff_t> ueec(Ls);
+        std::vector<Coeff_t> deec(Ls+1);
+        std::vector<Coeff_t> leec(Ls);
+        std::vector<Coeff_t> ueemc(Ls);
+        std::vector<Coeff_t> leemc(Ls);
+
+        for(int s=0; s<ueec.size(); s++){
+            ueec[s]  = conjugate(this->uee[s]);
+            deec[s]  = conjugate(this->dee[s]);
+            leec[s]  = conjugate(this->lee[s]);
+            ueemc[s] = conjugate(this->ueem[s]);
+            leemc[s] = conjugate(this->leem[s]);
+        }
+        deec[Ls] = conjugate(this->dee[Ls]);
+
+        this->MooeeInvCalls++;
+        this->MooeeInvTime -= usecond();
+
+        parallel_for(int ss=0; ss<grid->oSites(); ss+=Ls){ // adds Ls
+
+            auto tmp1 = psi._odata[0];
+            auto tmp2 = psi._odata[0];
+
+            // Apply (U^{\prime})^{-dagger}
+            chi[ss] = psi[ss];
+            for(int s=1; s<Ls; s++){
+                spProj5m(tmp1, chi[ss+s-1]);
+                chi[ss+s] = psi[ss+s] - ueec[s-1]*tmp1;
+            }
+
+            // U_m^{-\dagger}
+            for(int s=0; s<Ls-1; s++){
+                spProj5p(tmp1, chi[ss+s]);
+                chi[ss+Ls-1] = chi[ss+Ls-1] - ueemc[s]*tmp1;
+            }
+
+            // L_m^{-\dagger} D^{-dagger}
+            for(int s=0; s<Ls-1; s++){
+                spProj5m(tmp1, chi[ss+Ls-1]);
+                chi[ss+s] = (1.0/deec[s])*chi[ss+s] - (leemc[s]/deec[Ls-1])*tmp1;
+            }
+            spProj5p(tmp2, chi[ss+Ls-1]);
+            chi[ss+Ls-1] = (1.0/deec[Ls-1])*tmp1 + (1.0/deec[Ls])*tmp2;
+
+            // Apply L^{-dagger}
+            for(int s=Ls-2; s>=0; s--){
+                spProj5p(tmp1, chi[ss+s+1]);
+                chi[ss+s] = chi[ss+s] - leec[s]*tmp1;
+            }
+        }
+
+        this->MooeeInvTime += usecond();
+    }
+
+    #ifdef DOMAIN_WALL_EOFA_DPERP_CACHE
+
+        INSTANTIATE_DPERP_DWF_EOFA(WilsonImplF);
+        INSTANTIATE_DPERP_DWF_EOFA(WilsonImplD);
+        INSTANTIATE_DPERP_DWF_EOFA(GparityWilsonImplF);
+        INSTANTIATE_DPERP_DWF_EOFA(GparityWilsonImplD);
+        INSTANTIATE_DPERP_DWF_EOFA(ZWilsonImplF);
+        INSTANTIATE_DPERP_DWF_EOFA(ZWilsonImplD);
+
+        INSTANTIATE_DPERP_DWF_EOFA(WilsonImplFH);
+        INSTANTIATE_DPERP_DWF_EOFA(WilsonImplDF);
+        INSTANTIATE_DPERP_DWF_EOFA(GparityWilsonImplFH);
+        INSTANTIATE_DPERP_DWF_EOFA(GparityWilsonImplDF);
+        INSTANTIATE_DPERP_DWF_EOFA(ZWilsonImplFH);
+        INSTANTIATE_DPERP_DWF_EOFA(ZWilsonImplDF);
+
+    #endif
+
+}}

lib/qcd/action/fermion/DomainWallEOFAFermiondense.cc

@@ -0,0 +1,159 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/DomainWallEOFAFermiondense.cc
+
+Copyright (C) 2017
+
+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>
+Author: David Murphy <dmurphy@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 */
+
+#include <Grid/Grid_Eigen_Dense.h>
+#include <Grid/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/DomainWallEOFAFermion.h>
+
+namespace Grid {
+namespace QCD {
+
+    /*
+    * Dense matrix versions of routines
+    */
+    template<class Impl>
+    void DomainWallEOFAFermion<Impl>::MooeeInvDag(const FermionField& psi, FermionField& chi)
+    {
+        this->MooeeInternal(psi, chi, DaggerYes, InverseYes);
+    }
+
+    template<class Impl>
+    void DomainWallEOFAFermion<Impl>::MooeeInv(const FermionField& psi, FermionField& chi)
+    {
+        this->MooeeInternal(psi, chi, DaggerNo, InverseYes);
+    }
+
+    template<class Impl>
+    void DomainWallEOFAFermion<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)  = this->bee[s];
+            Pminus(s,s) = this->bee[s];
+        }
+
+        for(int s=0; s<Ls-1; s++){
+            Pminus(s,s+1) = -this->cee[s];
+        }
+
+        for(int s=0; s<Ls-1; s++){
+            Pplus(s+1,s) = -this->cee[s+1];
+        }
+
+        Pplus (0,Ls-1) = this->dp;
+        Pminus(Ls-1,0) = this->dm;
+
+        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;
+            }
+        }
+    }
+
+    #ifdef DOMAIN_WALL_EOFA_DPERP_DENSE
+
+        INSTANTIATE_DPERP_DWF_EOFA(GparityWilsonImplF);
+        INSTANTIATE_DPERP_DWF_EOFA(GparityWilsonImplD);
+        INSTANTIATE_DPERP_DWF_EOFA(WilsonImplF);
+        INSTANTIATE_DPERP_DWF_EOFA(WilsonImplD);
+        INSTANTIATE_DPERP_DWF_EOFA(ZWilsonImplF);
+        INSTANTIATE_DPERP_DWF_EOFA(ZWilsonImplD);
+
+        template void DomainWallEOFAFermion<GparityWilsonImplF>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
+        template void DomainWallEOFAFermion<GparityWilsonImplD>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
+        template void DomainWallEOFAFermion<WilsonImplF>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
+        template void DomainWallEOFAFermion<WilsonImplD>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
+        template void DomainWallEOFAFermion<ZWilsonImplF>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
+        template void DomainWallEOFAFermion<ZWilsonImplD>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
+
+        INSTANTIATE_DPERP_DWF_EOFA(GparityWilsonImplFH);
+        INSTANTIATE_DPERP_DWF_EOFA(GparityWilsonImplDF);
+        INSTANTIATE_DPERP_DWF_EOFA(WilsonImplFH);
+        INSTANTIATE_DPERP_DWF_EOFA(WilsonImplDF);
+        INSTANTIATE_DPERP_DWF_EOFA(ZWilsonImplFH);
+        INSTANTIATE_DPERP_DWF_EOFA(ZWilsonImplDF);
+
+        template void DomainWallEOFAFermion<GparityWilsonImplFH>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
+        template void DomainWallEOFAFermion<GparityWilsonImplDF>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
+        template void DomainWallEOFAFermion<WilsonImplFH>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
+        template void DomainWallEOFAFermion<WilsonImplDF>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
+        template void DomainWallEOFAFermion<ZWilsonImplFH>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
+        template void DomainWallEOFAFermion<ZWilsonImplDF>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
+
+    #endif
+
+}}

lib/qcd/action/fermion/DomainWallEOFAFermionssp.cc

@@ -0,0 +1,168 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/DomainWallEOFAFermionssp.cc
+
+Copyright (C) 2017
+
+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>
+Author: David Murphy <dmurphy@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 */
+
+#include <Grid/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/DomainWallEOFAFermion.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 DomainWallEOFAFermion<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)
+    {
+        Coeff_t one(1.0);
+        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, one, 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, one, 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, one, chi, lower[s], psi, s, s-1);
+            }
+        }
+    }
+
+    template<class Impl>
+    void DomainWallEOFAFermion<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)
+    {
+        Coeff_t one(1.0);
+        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, one, 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, one, 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, one, chi, lower[s], psi, s, s-1);
+            }
+        }
+    }
+
+    template<class Impl>
+    void DomainWallEOFAFermion<Impl>::MooeeInv(const FermionField& psi, FermionField& chi)
+    {
+        Coeff_t one(1.0);
+        Coeff_t czero(0.0);
+        chi.checkerboard = psi.checkerboard;
+        int Ls = this->Ls;
+
+        FermionField tmp(psi._grid);
+
+        // Apply (L^{\prime})^{-1}
+        axpby_ssp(chi, one, psi, czero, psi, 0, 0);      // chi[0]=psi[0]
+        for(int s=1; s<Ls; s++){
+            axpby_ssp_pplus(chi, one, psi, -this->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, one, chi, -this->leem[s], chi, Ls-1, s);
+        }
+
+        // U_m^{-1} D^{-1}
+        for(int s=0; s<Ls-1; s++){
+            axpby_ssp_pplus(chi, one/this->dee[s], chi, -this->ueem[s]/this->dee[Ls], chi, s, Ls-1);
+        }
+        axpby_ssp_pminus(tmp, czero, chi, one/this->dee[Ls-1], chi, Ls-1, Ls-1);
+        axpby_ssp_pplus(chi, one, tmp, one/this->dee[Ls], chi, Ls-1, Ls-1);
+
+        // Apply U^{-1}
+        for(int s=Ls-2; s>=0; s--){
+            axpby_ssp_pminus(chi, one, chi, -this->uee[s], chi, s, s+1);  // chi[Ls]
+        }
+    }
+
+    template<class Impl>
+    void DomainWallEOFAFermion<Impl>::MooeeInvDag(const FermionField& psi, FermionField& chi)
+    {
+        Coeff_t one(1.0);
+        Coeff_t czero(0.0);
+        chi.checkerboard = psi.checkerboard;
+        int Ls = this->Ls;
+
+        FermionField tmp(psi._grid);
+
+        // Apply (U^{\prime})^{-dagger}
+        axpby_ssp(chi, one, psi, czero, psi, 0, 0);      // chi[0]=psi[0]
+        for(int s=1; s<Ls; s++){
+            axpby_ssp_pminus(chi, one, psi, -conjugate(this->uee[s-1]), chi, s, s-1);
+        }
+
+        // U_m^{-\dagger}
+        for(int s=0; s<Ls-1; s++){
+            axpby_ssp_pplus(chi, one, chi, -conjugate(this->ueem[s]), chi, Ls-1, s);
+        }
+
+        // L_m^{-\dagger} D^{-dagger}
+        for(int s=0; s<Ls-1; s++){
+            axpby_ssp_pminus(chi, one/conjugate(this->dee[s]), chi, -conjugate(this->leem[s]/this->dee[Ls-1]), chi, s, Ls-1);
+        }
+        axpby_ssp_pminus(tmp, czero, chi, one/conjugate(this->dee[Ls-1]), chi, Ls-1, Ls-1);
+        axpby_ssp_pplus(chi, one, tmp, one/conjugate(this->dee[Ls]), chi, Ls-1, Ls-1);
+
+        // Apply L^{-dagger}
+        for(int s=Ls-2; s>=0; s--){
+            axpby_ssp_pplus(chi, one, chi, -conjugate(this->lee[s]), chi, s, s+1);  // chi[Ls]
+        }
+    }
+
+    #ifdef DOMAIN_WALL_EOFA_DPERP_LINALG
+
+        INSTANTIATE_DPERP_DWF_EOFA(WilsonImplF);
+        INSTANTIATE_DPERP_DWF_EOFA(WilsonImplD);
+        INSTANTIATE_DPERP_DWF_EOFA(GparityWilsonImplF);
+        INSTANTIATE_DPERP_DWF_EOFA(GparityWilsonImplD);
+        INSTANTIATE_DPERP_DWF_EOFA(ZWilsonImplF);
+        INSTANTIATE_DPERP_DWF_EOFA(ZWilsonImplD);
+
+        INSTANTIATE_DPERP_DWF_EOFA(WilsonImplFH);
+        INSTANTIATE_DPERP_DWF_EOFA(WilsonImplDF);
+        INSTANTIATE_DPERP_DWF_EOFA(GparityWilsonImplFH);
+        INSTANTIATE_DPERP_DWF_EOFA(GparityWilsonImplDF);
+        INSTANTIATE_DPERP_DWF_EOFA(ZWilsonImplFH);
+        INSTANTIATE_DPERP_DWF_EOFA(ZWilsonImplDF);
+
+    #endif
+
+}}

lib/qcd/action/fermion/DomainWallEOFAFermionvec.cc

@@ -0,0 +1,605 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/qcd/action/fermion/DomainWallEOFAFermionvec.cc
+
+Copyright (C) 2017
+
+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>
+Author: David Murphy <dmurphy@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 */
+
+#include <Grid/qcd/action/fermion/FermionCore.h>
+#include <Grid/qcd/action/fermion/DomainWallEOFAFermion.h>
+
+namespace Grid {
+namespace QCD {
+
+    /*
+    * Dense matrix versions of routines
+    */
+    template<class Impl>
+    void DomainWallEOFAFermion<Impl>::MooeeInvDag(const FermionField& psi, FermionField& chi)
+    {
+        this->MooeeInternal(psi, chi, DaggerYes, InverseYes);
+    }
+
+    template<class Impl>
+    void DomainWallEOFAFermion<Impl>::MooeeInv(const FermionField& psi, FermionField& chi)
+    {
+        this->MooeeInternal(psi, chi, DaggerNo, InverseYes);
+    }
+
+    template<class Impl>
+    void DomainWallEOFAFermion<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];
+        const 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];
+        }}
+
+        this->M5Dcalls++;
+        this->M5Dtime -= usecond();
+
+        assert(Nc == 3);
+
+        parallel_for(int ss=0; ss<grid->oSites(); ss+=LLs){ // adds LLs
+
+            #if 0
+
+                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 = 0.5*hp;
+                    hm = 0.5*hm;
+
+                    spRecon5m(fp, hp);
+                    spRecon5p(fm, hm);
+
+                    chi[ss+v] = d[v]*phi[ss+v];
+                    chi[ss+v] = chi[ss+v] + u[v]*fp;
+                    chi[ss+v] = chi[ss+v] + l[v]*fm;
+
+                }
+
+            #else
+
+                for(int v=0; v<LLs; v++){
+
+                    vprefetch(psi[ss+v+LLs]);
+
+                    int vp = (v==LLs-1) ? 0     : v+1;
+                    int vm = (v==0)     ? LLs-1 : v-1;
+
+                    Simd hp_00 = psi[ss+vp]()(2)(0);
+                    Simd hp_01 = psi[ss+vp]()(2)(1);
+                    Simd hp_02 = psi[ss+vp]()(2)(2);
+                    Simd hp_10 = psi[ss+vp]()(3)(0);
+                    Simd hp_11 = psi[ss+vp]()(3)(1);
+                    Simd hp_12 = psi[ss+vp]()(3)(2);
+
+                    Simd hm_00 = psi[ss+vm]()(0)(0);
+                    Simd hm_01 = psi[ss+vm]()(0)(1);
+                    Simd hm_02 = psi[ss+vm]()(0)(2);
+                    Simd hm_10 = psi[ss+vm]()(1)(0);
+                    Simd hm_11 = psi[ss+vm]()(1)(1);
+                    Simd hm_12 = psi[ss+vm]()(1)(2);
+
+                    if(vp <= v){
+                        hp_00.v = Optimization::Rotate::tRotate<2>(hp_00.v);
+                        hp_01.v = Optimization::Rotate::tRotate<2>(hp_01.v);
+                        hp_02.v = Optimization::Rotate::tRotate<2>(hp_02.v);
+                        hp_10.v = Optimization::Rotate::tRotate<2>(hp_10.v);
+                        hp_11.v = Optimization::Rotate::tRotate<2>(hp_11.v);
+                        hp_12.v = Optimization::Rotate::tRotate<2>(hp_12.v);
+                    }
+
+                    if(vm >= v){
+                        hm_00.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_00.v);
+                        hm_01.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_01.v);
+                        hm_02.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_02.v);
+                        hm_10.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_10.v);
+                        hm_11.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_11.v);
+                        hm_12.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_12.v);
+                    }
+
+                    // Can force these to real arithmetic and save 2x.
+                    Simd p_00 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_00);
+                    Simd p_01 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_01);
+                    Simd p_02 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_02);
+                    Simd p_10 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_10);
+                    Simd p_11 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_11);
+                    Simd p_12 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_12);
+                    Simd p_20 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_00);
+                    Simd p_21 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_01);
+                    Simd p_22 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_02);
+                    Simd p_30 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_10);
+                    Simd p_31 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_11);
+                    Simd p_32 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_12);
+
+                    vstream(chi[ss+v]()(0)(0), p_00);
+                    vstream(chi[ss+v]()(0)(1), p_01);
+                    vstream(chi[ss+v]()(0)(2), p_02);
+                    vstream(chi[ss+v]()(1)(0), p_10);
+                    vstream(chi[ss+v]()(1)(1), p_11);
+                    vstream(chi[ss+v]()(1)(2), p_12);
+                    vstream(chi[ss+v]()(2)(0), p_20);
+                    vstream(chi[ss+v]()(2)(1), p_21);
+                    vstream(chi[ss+v]()(2)(2), p_22);
+                    vstream(chi[ss+v]()(3)(0), p_30);
+                    vstream(chi[ss+v]()(3)(1), p_31);
+                    vstream(chi[ss+v]()(3)(2), p_32);
+                }
+
+            #endif
+        }
+
+        this->M5Dtime += usecond();
+    }
+
+    template<class Impl>
+    void DomainWallEOFAFermion<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];
+        }}
+
+        this->M5Dcalls++;
+        this->M5Dtime -= usecond();
+
+        parallel_for(int ss=0; ss<grid->oSites(); ss+=LLs){ // adds LLs
+
+        #if 0
+
+            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;
+            }
+
+        #else
+
+            for(int v=0; v<LLs; v++){
+
+                vprefetch(psi[ss+v+LLs]);
+
+                int vp = (v == LLs-1) ? 0     : v+1;
+                int vm = (v == 0    ) ? LLs-1 : v-1;
+
+                Simd hp_00 = psi[ss+vp]()(0)(0);
+                Simd hp_01 = psi[ss+vp]()(0)(1);
+                Simd hp_02 = psi[ss+vp]()(0)(2);
+                Simd hp_10 = psi[ss+vp]()(1)(0);
+                Simd hp_11 = psi[ss+vp]()(1)(1);
+                Simd hp_12 = psi[ss+vp]()(1)(2);
+
+                Simd hm_00 = psi[ss+vm]()(2)(0);
+                Simd hm_01 = psi[ss+vm]()(2)(1);
+                Simd hm_02 = psi[ss+vm]()(2)(2);
+                Simd hm_10 = psi[ss+vm]()(3)(0);
+                Simd hm_11 = psi[ss+vm]()(3)(1);
+                Simd hm_12 = psi[ss+vm]()(3)(2);
+
+                if (vp <= v){
+                    hp_00.v = Optimization::Rotate::tRotate<2>(hp_00.v);
+                    hp_01.v = Optimization::Rotate::tRotate<2>(hp_01.v);
+                    hp_02.v = Optimization::Rotate::tRotate<2>(hp_02.v);
+                    hp_10.v = Optimization::Rotate::tRotate<2>(hp_10.v);
+                    hp_11.v = Optimization::Rotate::tRotate<2>(hp_11.v);
+                    hp_12.v = Optimization::Rotate::tRotate<2>(hp_12.v);
+                }
+
+                if(vm >= v){
+                    hm_00.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_00.v);
+                    hm_01.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_01.v);
+                    hm_02.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_02.v);
+                    hm_10.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_10.v);
+                    hm_11.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_11.v);
+                    hm_12.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_12.v);
+                }
+
+                Simd p_00 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_00);
+                Simd p_01 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_01);
+                Simd p_02 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_02);
+                Simd p_10 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_10);
+                Simd p_11 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_11);
+                Simd p_12 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_12);
+                Simd p_20 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_00);
+                Simd p_21 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_01);
+                Simd p_22 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_02);
+                Simd p_30 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_10);
+                Simd p_31 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_11);
+                Simd p_32 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_12);
+
+                vstream(chi[ss+v]()(0)(0), p_00);
+                vstream(chi[ss+v]()(0)(1), p_01);
+                vstream(chi[ss+v]()(0)(2), p_02);
+                vstream(chi[ss+v]()(1)(0), p_10);
+                vstream(chi[ss+v]()(1)(1), p_11);
+                vstream(chi[ss+v]()(1)(2), p_12);
+                vstream(chi[ss+v]()(2)(0), p_20);
+                vstream(chi[ss+v]()(2)(1), p_21);
+                vstream(chi[ss+v]()(2)(2), p_22);
+                vstream(chi[ss+v]()(3)(0), p_30);
+                vstream(chi[ss+v]()(3)(1), p_31);
+                vstream(chi[ss+v]()(3)(2), p_32);
+            }
+        #endif
+
+        }
+
+        this->M5Dtime += usecond();
+    }
+
+    #ifdef AVX512
+        #include<simd/Intel512common.h>
+        #include<simd/Intel512avx.h>
+        #include<simd/Intel512single.h>
+    #endif
+
+    template<class Impl>
+    void DomainWallEOFAFermion<Impl>::MooeeInternalAsm(const FermionField& psi, FermionField& chi,
+        int LLs, int site, Vector<iSinglet<Simd> >& Matp, Vector<iSinglet<Simd> >& Matm)
+    {
+        #ifndef AVX512
+        {
+            SiteHalfSpinor BcastP;
+            SiteHalfSpinor BcastM;
+            SiteHalfSpinor SiteChiP;
+            SiteHalfSpinor SiteChiM;
+
+            // Ls*Ls * 2 * 12 * vol flops
+            for(int s1=0; s1<LLs; s1++){
+
+                for(int s2=0; s2<LLs; s2++){
+                for(int l=0; l < Simd::Nsimd(); l++){ // simd lane
+
+                    int s = s2 + l*LLs;
+                    int lex = s2 + LLs*site;
+
+                    if( s2==0 && l==0 ){
+                        SiteChiP=zero;
+                        SiteChiM=zero;
+                    }
+
+                    for(int sp=0; sp<2;  sp++){
+                    for(int co=0; co<Nc; co++){
+                        vbroadcast(BcastP()(sp)(co), psi[lex]()(sp)(co), l);
+                    }}
+
+                    for(int sp=0; sp<2;  sp++){
+                    for(int co=0; co<Nc; co++){
+                        vbroadcast(BcastM()(sp)(co), psi[lex]()(sp+2)(co), l);
+                    }}
+
+                    for(int sp=0; sp<2;  sp++){
+                    for(int co=0; co<Nc; co++){
+                        SiteChiP()(sp)(co) = real_madd(Matp[LLs*s+s1]()()(), BcastP()(sp)(co), SiteChiP()(sp)(co)); // 1100 us.
+                        SiteChiM()(sp)(co) = real_madd(Matm[LLs*s+s1]()()(), BcastM()(sp)(co), SiteChiM()(sp)(co)); // each found by commenting out
+                    }}
+                }}
+
+                {
+                    int lex = s1 + LLs*site;
+                    for(int sp=0; sp<2;  sp++){
+                    for(int co=0; co<Nc; co++){
+                        vstream(chi[lex]()(sp)(co),   SiteChiP()(sp)(co));
+                        vstream(chi[lex]()(sp+2)(co), SiteChiM()(sp)(co));
+                    }}
+                }
+            }
+
+        }
+        #else
+        {
+            // pointers
+            //  MASK_REGS;
+            #define Chi_00 %%zmm1
+            #define Chi_01 %%zmm2
+            #define Chi_02 %%zmm3
+            #define Chi_10 %%zmm4
+            #define Chi_11 %%zmm5
+            #define Chi_12 %%zmm6
+            #define Chi_20 %%zmm7
+            #define Chi_21 %%zmm8
+            #define Chi_22 %%zmm9
+            #define Chi_30 %%zmm10
+            #define Chi_31 %%zmm11
+            #define Chi_32 %%zmm12
+
+            #define BCAST0  %%zmm13
+            #define BCAST1  %%zmm14
+            #define BCAST2  %%zmm15
+            #define BCAST3  %%zmm16
+            #define BCAST4  %%zmm17
+            #define BCAST5  %%zmm18
+            #define BCAST6  %%zmm19
+            #define BCAST7  %%zmm20
+            #define BCAST8  %%zmm21
+            #define BCAST9  %%zmm22
+            #define BCAST10 %%zmm23
+            #define BCAST11 %%zmm24
+
+            int incr = LLs*LLs*sizeof(iSinglet<Simd>);
+            for(int s1=0; s1<LLs; s1++){
+
+                for(int s2=0; s2<LLs; s2++){
+
+                    int lex = s2 + LLs*site;
+                    uint64_t a0 = (uint64_t) &Matp[LLs*s2+s1]; // should be cacheable
+                    uint64_t a1 = (uint64_t) &Matm[LLs*s2+s1];
+                    uint64_t a2 = (uint64_t) &psi[lex];
+
+                    for(int l=0; l<Simd::Nsimd(); l++){ // simd lane
+                        if((s2+l)==0) {
+                            asm(
+                                    VPREFETCH1(0,%2)              VPREFETCH1(0,%1)
+                                    VPREFETCH1(12,%2)  	          VPREFETCH1(13,%2)
+                                    VPREFETCH1(14,%2)  	          VPREFETCH1(15,%2)
+                                    VBCASTCDUP(0,%2,BCAST0)
+                                    VBCASTCDUP(1,%2,BCAST1)
+                                    VBCASTCDUP(2,%2,BCAST2)
+                                    VBCASTCDUP(3,%2,BCAST3)
+                                    VBCASTCDUP(4,%2,BCAST4)       VMULMEM(0,%0,BCAST0,Chi_00)
+                                    VBCASTCDUP(5,%2,BCAST5)       VMULMEM(0,%0,BCAST1,Chi_01)
+                                    VBCASTCDUP(6,%2,BCAST6)       VMULMEM(0,%0,BCAST2,Chi_02)
+                                    VBCASTCDUP(7,%2,BCAST7)       VMULMEM(0,%0,BCAST3,Chi_10)
+                                    VBCASTCDUP(8,%2,BCAST8)       VMULMEM(0,%0,BCAST4,Chi_11)
+                                    VBCASTCDUP(9,%2,BCAST9)       VMULMEM(0,%0,BCAST5,Chi_12)
+                                    VBCASTCDUP(10,%2,BCAST10)     VMULMEM(0,%1,BCAST6,Chi_20)
+                                    VBCASTCDUP(11,%2,BCAST11)     VMULMEM(0,%1,BCAST7,Chi_21)
+                                    VMULMEM(0,%1,BCAST8,Chi_22)
+                                    VMULMEM(0,%1,BCAST9,Chi_30)
+                                    VMULMEM(0,%1,BCAST10,Chi_31)
+                                    VMULMEM(0,%1,BCAST11,Chi_32)
+                                    : : "r" (a0), "r" (a1), "r" (a2)                            );
+                        } else {
+                            asm(
+                                    VBCASTCDUP(0,%2,BCAST0)   VMADDMEM(0,%0,BCAST0,Chi_00)
+                                    VBCASTCDUP(1,%2,BCAST1)   VMADDMEM(0,%0,BCAST1,Chi_01)
+                                    VBCASTCDUP(2,%2,BCAST2)   VMADDMEM(0,%0,BCAST2,Chi_02)
+                                    VBCASTCDUP(3,%2,BCAST3)   VMADDMEM(0,%0,BCAST3,Chi_10)
+                                    VBCASTCDUP(4,%2,BCAST4)   VMADDMEM(0,%0,BCAST4,Chi_11)
+                                    VBCASTCDUP(5,%2,BCAST5)   VMADDMEM(0,%0,BCAST5,Chi_12)
+                                    VBCASTCDUP(6,%2,BCAST6)   VMADDMEM(0,%1,BCAST6,Chi_20)
+                                    VBCASTCDUP(7,%2,BCAST7)   VMADDMEM(0,%1,BCAST7,Chi_21)
+                                    VBCASTCDUP(8,%2,BCAST8)   VMADDMEM(0,%1,BCAST8,Chi_22)
+                                    VBCASTCDUP(9,%2,BCAST9)   VMADDMEM(0,%1,BCAST9,Chi_30)
+                                    VBCASTCDUP(10,%2,BCAST10) VMADDMEM(0,%1,BCAST10,Chi_31)
+                                    VBCASTCDUP(11,%2,BCAST11) VMADDMEM(0,%1,BCAST11,Chi_32)
+                                    : : "r" (a0), "r" (a1), "r" (a2)                            );
+                        }
+                        a0 = a0 + incr;
+                        a1 = a1 + incr;
+                        a2 = a2 + sizeof(Simd::scalar_type);
+                    }
+                }
+
+                {
+                  int lexa = s1+LLs*site;
+                  asm (
+                     VSTORE(0,%0,Chi_00) VSTORE(1 ,%0,Chi_01)  VSTORE(2 ,%0,Chi_02)
+                     VSTORE(3,%0,Chi_10) VSTORE(4 ,%0,Chi_11)  VSTORE(5 ,%0,Chi_12)
+                     VSTORE(6,%0,Chi_20) VSTORE(7 ,%0,Chi_21)  VSTORE(8 ,%0,Chi_22)
+                     VSTORE(9,%0,Chi_30) VSTORE(10,%0,Chi_31)  VSTORE(11,%0,Chi_32)
+                     : : "r" ((uint64_t)&chi[lexa]) : "memory" );
+
+                }
+            }
+        }
+
+        #undef Chi_00
+        #undef Chi_01
+        #undef Chi_02
+        #undef Chi_10
+        #undef Chi_11
+        #undef Chi_12
+        #undef Chi_20
+        #undef Chi_21
+        #undef Chi_22
+        #undef Chi_30
+        #undef Chi_31
+        #undef Chi_32
+
+        #undef BCAST0
+        #undef BCAST1
+        #undef BCAST2
+        #undef BCAST3
+        #undef BCAST4
+        #undef BCAST5
+        #undef BCAST6
+        #undef BCAST7
+        #undef BCAST8
+        #undef BCAST9
+        #undef BCAST10
+        #undef BCAST11
+        #endif
+    };
+
+    // Z-mobius version
+    template<class Impl>
+    void DomainWallEOFAFermion<Impl>::MooeeInternalZAsm(const FermionField& psi, FermionField& chi,
+        int LLs, int site, Vector<iSinglet<Simd> >& Matp, Vector<iSinglet<Simd> >& Matm)
+    {
+        std::cout << "Error: zMobius not implemented for EOFA" << std::endl;
+        exit(-1);
+    };
+
+    template<class Impl>
+    void DomainWallEOFAFermion<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;
+
+        Vector<iSinglet<Simd> > Matp;
+        Vector<iSinglet<Simd> > Matm;
+        Vector<iSinglet<Simd> > *_Matp;
+        Vector<iSinglet<Simd> > *_Matm;
+
+        //  MooeeInternalCompute(dag,inv,Matp,Matm);
+        if(inv && dag){
+            _Matp = &this->MatpInvDag;
+            _Matm = &this->MatmInvDag;
+        }
+
+        if(inv && (!dag)){
+            _Matp = &this->MatpInv;
+            _Matm = &this->MatmInv;
+        }
+
+        if(!inv){
+            MooeeInternalCompute(dag, inv, Matp, Matm);
+            _Matp = &Matp;
+            _Matm = &Matm;
+        }
+
+        assert(_Matp->size() == Ls*LLs);
+
+        this->MooeeInvCalls++;
+        this->MooeeInvTime -= usecond();
+
+        if(switcheroo<Coeff_t>::iscomplex()){
+            parallel_for(auto site=0; site<vol; site++){
+                MooeeInternalZAsm(psi, chi, LLs, site, *_Matp, *_Matm);
+            }
+        } else {
+            parallel_for(auto site=0; site<vol; site++){
+                MooeeInternalAsm(psi, chi, LLs, site, *_Matp, *_Matm);
+            }
+        }
+
+        this->MooeeInvTime += usecond();
+    }
+
+    #ifdef DOMAIN_WALL_EOFA_DPERP_VEC
+
+        INSTANTIATE_DPERP_DWF_EOFA(DomainWallVec5dImplD);
+        INSTANTIATE_DPERP_DWF_EOFA(DomainWallVec5dImplF);
+        INSTANTIATE_DPERP_DWF_EOFA(ZDomainWallVec5dImplD);
+        INSTANTIATE_DPERP_DWF_EOFA(ZDomainWallVec5dImplF);
+
+        INSTANTIATE_DPERP_DWF_EOFA(DomainWallVec5dImplDF);
+        INSTANTIATE_DPERP_DWF_EOFA(DomainWallVec5dImplFH);
+        INSTANTIATE_DPERP_DWF_EOFA(ZDomainWallVec5dImplDF);
+        INSTANTIATE_DPERP_DWF_EOFA(ZDomainWallVec5dImplFH);
+
+        template void DomainWallEOFAFermion<DomainWallVec5dImplF>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
+        template void DomainWallEOFAFermion<DomainWallVec5dImplD>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
+        template void DomainWallEOFAFermion<ZDomainWallVec5dImplF>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
+        template void DomainWallEOFAFermion<ZDomainWallVec5dImplD>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
+
+        template void DomainWallEOFAFermion<DomainWallVec5dImplFH>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
+        template void DomainWallEOFAFermion<DomainWallVec5dImplDF>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
+        template void DomainWallEOFAFermion<ZDomainWallVec5dImplFH>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
+        template void DomainWallEOFAFermion<ZDomainWallVec5dImplDF>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
+
+    #endif
+
+}}

lib/qcd/action/fermion/Fermion.h

@@ -38,6 +38,8 @@ Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
 // - ContinuedFractionFermion5D.cc
 // - WilsonFermion.cc
 // - WilsonKernels.cc
+// - DomainWallEOFAFermion.cc
+// - MobiusEOFAFermion.cc
 //
 // The explicit instantiation is only avoidable if we move this source to headers and end up with include/parse/recompile
 // for EVERY .cc file. This define centralises the list and restores global push of impl cases
@@ -55,8 +57,9 @@ Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
 #include <Grid/qcd/action/fermion/ImprovedStaggeredFermion5D.h>
 #include <Grid/qcd/action/fermion/CayleyFermion5D.h>     // Cayley types
 #include <Grid/qcd/action/fermion/DomainWallFermion.h>
-#include <Grid/qcd/action/fermion/DomainWallFermion.h>
+#include <Grid/qcd/action/fermion/DomainWallEOFAFermion.h>
 #include <Grid/qcd/action/fermion/MobiusFermion.h>
+#include <Grid/qcd/action/fermion/MobiusEOFAFermion.h>
 #include <Grid/qcd/action/fermion/ZMobiusFermion.h>
 #include <Grid/qcd/action/fermion/SchurDiagTwoKappa.h>
 #include <Grid/qcd/action/fermion/ScaledShamirFermion.h>
@@ -113,6 +116,14 @@ typedef DomainWallFermion<WilsonImplRL> DomainWallFermionRL;
 typedef DomainWallFermion<WilsonImplFH> DomainWallFermionFH;
 typedef DomainWallFermion<WilsonImplDF> DomainWallFermionDF;
 
+typedef DomainWallEOFAFermion<WilsonImplR> DomainWallEOFAFermionR;
+typedef DomainWallEOFAFermion<WilsonImplF> DomainWallEOFAFermionF;
+typedef DomainWallEOFAFermion<WilsonImplD> DomainWallEOFAFermionD;
+
+typedef DomainWallEOFAFermion<WilsonImplRL> DomainWallEOFAFermionRL;
+typedef DomainWallEOFAFermion<WilsonImplFH> DomainWallEOFAFermionFH;
+typedef DomainWallEOFAFermion<WilsonImplDF> DomainWallEOFAFermionDF;
+
 typedef MobiusFermion<WilsonImplR> MobiusFermionR;
 typedef MobiusFermion<WilsonImplF> MobiusFermionF;
 typedef MobiusFermion<WilsonImplD> MobiusFermionD;
@@ -121,6 +132,14 @@ typedef MobiusFermion<WilsonImplRL> MobiusFermionRL;
 typedef MobiusFermion<WilsonImplFH> MobiusFermionFH;
 typedef MobiusFermion<WilsonImplDF> MobiusFermionDF;
 
+typedef MobiusEOFAFermion<WilsonImplR> MobiusEOFAFermionR;
+typedef MobiusEOFAFermion<WilsonImplF> MobiusEOFAFermionF;
+typedef MobiusEOFAFermion<WilsonImplD> MobiusEOFAFermionD;
+
+typedef MobiusEOFAFermion<WilsonImplRL> MobiusEOFAFermionRL;
+typedef MobiusEOFAFermion<WilsonImplFH> MobiusEOFAFermionFH;
+typedef MobiusEOFAFermion<WilsonImplDF> MobiusEOFAFermionDF;
+
 typedef ZMobiusFermion<ZWilsonImplR> ZMobiusFermionR;
 typedef ZMobiusFermion<ZWilsonImplF> ZMobiusFermionF;
 typedef ZMobiusFermion<ZWilsonImplD> ZMobiusFermionD;
@@ -138,6 +157,14 @@ typedef DomainWallFermion<DomainWallVec5dImplRL> DomainWallFermionVec5dRL;
 typedef DomainWallFermion<DomainWallVec5dImplFH> DomainWallFermionVec5dFH;
 typedef DomainWallFermion<DomainWallVec5dImplDF> DomainWallFermionVec5dDF;
 
+typedef DomainWallEOFAFermion<DomainWallVec5dImplR> DomainWallEOFAFermionVec5dR;
+typedef DomainWallEOFAFermion<DomainWallVec5dImplF> DomainWallEOFAFermionVec5dF;
+typedef DomainWallEOFAFermion<DomainWallVec5dImplD> DomainWallEOFAFermionVec5dD;
+
+typedef DomainWallEOFAFermion<DomainWallVec5dImplRL> DomainWallEOFAFermionVec5dRL;
+typedef DomainWallEOFAFermion<DomainWallVec5dImplFH> DomainWallEOFAFermionVec5dFH;
+typedef DomainWallEOFAFermion<DomainWallVec5dImplDF> DomainWallEOFAFermionVec5dDF;
+
 typedef MobiusFermion<DomainWallVec5dImplR> MobiusFermionVec5dR;
 typedef MobiusFermion<DomainWallVec5dImplF> MobiusFermionVec5dF;
 typedef MobiusFermion<DomainWallVec5dImplD> MobiusFermionVec5dD;
@@ -146,6 +173,14 @@ typedef MobiusFermion<DomainWallVec5dImplRL> MobiusFermionVec5dRL;
 typedef MobiusFermion<DomainWallVec5dImplFH> MobiusFermionVec5dFH;
 typedef MobiusFermion<DomainWallVec5dImplDF> MobiusFermionVec5dDF;
 
+typedef MobiusEOFAFermion<DomainWallVec5dImplR> MobiusEOFAFermionVec5dR;
+typedef MobiusEOFAFermion<DomainWallVec5dImplF> MobiusEOFAFermionVec5dF;
+typedef MobiusEOFAFermion<DomainWallVec5dImplD> MobiusEOFAFermionVec5dD;
+
+typedef MobiusEOFAFermion<DomainWallVec5dImplRL> MobiusEOFAFermionVec5dRL;
+typedef MobiusEOFAFermion<DomainWallVec5dImplFH> MobiusEOFAFermionVec5dFH;
+typedef MobiusEOFAFermion<DomainWallVec5dImplDF> MobiusEOFAFermionVec5dDF;
+
 typedef ZMobiusFermion<ZDomainWallVec5dImplR> ZMobiusFermionVec5dR;
 typedef ZMobiusFermion<ZDomainWallVec5dImplF> ZMobiusFermionVec5dF;
 typedef ZMobiusFermion<ZDomainWallVec5dImplD> ZMobiusFermionVec5dD;
@@ -206,6 +241,14 @@ typedef DomainWallFermion<GparityWilsonImplRL> GparityDomainWallFermionRL;
 typedef DomainWallFermion<GparityWilsonImplFH> GparityDomainWallFermionFH;
 typedef DomainWallFermion<GparityWilsonImplDF> GparityDomainWallFermionDF;
 
+typedef DomainWallEOFAFermion<GparityWilsonImplR> GparityDomainWallEOFAFermionR;
+typedef DomainWallEOFAFermion<GparityWilsonImplF> GparityDomainWallEOFAFermionF;
+typedef DomainWallEOFAFermion<GparityWilsonImplD> GparityDomainWallEOFAFermionD;
+
+typedef DomainWallEOFAFermion<GparityWilsonImplRL> GparityDomainWallEOFAFermionRL;
+typedef DomainWallEOFAFermion<GparityWilsonImplFH> GparityDomainWallEOFAFermionFH;
+typedef DomainWallEOFAFermion<GparityWilsonImplDF> GparityDomainWallEOFAFermionDF;
+
 typedef WilsonTMFermion<GparityWilsonImplR> GparityWilsonTMFermionR;
 typedef WilsonTMFermion<GparityWilsonImplF> GparityWilsonTMFermionF;
 typedef WilsonTMFermion<GparityWilsonImplD> GparityWilsonTMFermionD;
@@ -222,6 +265,14 @@ typedef MobiusFermion<GparityWilsonImplRL> GparityMobiusFermionRL;
 typedef MobiusFermion<GparityWilsonImplFH> GparityMobiusFermionFH;
 typedef MobiusFermion<GparityWilsonImplDF> GparityMobiusFermionDF;
 
+typedef MobiusEOFAFermion<GparityWilsonImplR> GparityMobiusEOFAFermionR;
+typedef MobiusEOFAFermion<GparityWilsonImplF> GparityMobiusEOFAFermionF;
+typedef MobiusEOFAFermion<GparityWilsonImplD> GparityMobiusEOFAFermionD;
+
+typedef MobiusEOFAFermion<GparityWilsonImplRL> GparityMobiusEOFAFermionRL;
+typedef MobiusEOFAFermion<GparityWilsonImplFH> GparityMobiusEOFAFermionFH;
+typedef MobiusEOFAFermion<GparityWilsonImplDF> GparityMobiusEOFAFermionDF;
+
 typedef ImprovedStaggeredFermion<StaggeredImplR> ImprovedStaggeredFermionR;
 typedef ImprovedStaggeredFermion<StaggeredImplF> ImprovedStaggeredFermionF;
 typedef ImprovedStaggeredFermion<StaggeredImplD> ImprovedStaggeredFermionD;
@@ -237,4 +288,11 @@ typedef ImprovedStaggeredFermion5D<StaggeredVec5dImplD> ImprovedStaggeredFermion
 
   }}
 
+////////////////////
+// Scalar QED actions
+// TODO: this needs to move to another header after rename to Fermion.h
+////////////////////
+#include <Grid/qcd/action/scalar/Scalar.h>
+#include <Grid/qcd/action/gauge/Photon.h>
+
 #endif

lib/qcd/action/fermion/FermionOperatorImpl.h

@@ -538,6 +538,12 @@ class GparityWilsonImpl : public ConjugateGaugeImpl<GaugeImplTypes<S, Nrepresent
    
  }
 
+
+ template <class ref>
+ inline void loadLinkElement(Simd &reg, ref &memory) {
+   reg = memory;
+ }
+
  inline void DoubleStore(GridBase *GaugeGrid,DoubledGaugeField &Uds,const GaugeField &Umu)
  {
    conformable(Uds._grid,GaugeGrid);