Made checkerboard choice in staggered preconditioned solvers switchable

Add support for DWF with the exact one flavor algorithm

.gitignore

@@ -92,6 +92,7 @@ build*/*
 #####################
 *.xcodeproj/*
 build.sh
+.vscode
 
 # Eigen source #
 ################
@@ -106,6 +107,10 @@ lib/fftw/*
 m4/lt*
 m4/libtool.m4
 
+# github pages #
+################
+gh-pages/
+
 # Buck files #
 ##############
 .buck*
@@ -116,4 +121,5 @@ make-bin-BUCK.sh
 # generated sources #
 #####################
 lib/qcd/spin/gamma-gen/*.h
-lib/qcd/spin/gamma-gen/*.cc
+lib/qcd/spin/gamma-gen/*.cc
+

.travis.yml

@@ -9,62 +9,6 @@ matrix:
     - os:        osx
       osx_image: xcode8.3
       compiler: clang
-    - compiler: gcc
-      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
-      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
-      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
-      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`
@@ -78,6 +22,10 @@ install:
     - export CC=$CC$VERSION
     - export CXX=$CXX$VERSION
     - echo $PATH
+    - which autoconf
+    - autoconf  --version
+    - which automake
+    - automake  --version
     - which $CC
     - $CC  --version
     - which $CXX
@@ -95,9 +43,4 @@ script:
     - ../configure --enable-precision=double --enable-simd=SSE4 --enable-comms=none
     - make -j4
     - ./benchmarks/Benchmark_dwf --threads 1 --debug-signals
-    - echo make clean
-    - if [[ "$TRAVIS_OS_NAME" == "linux" ]]; then ../configure --enable-precision=single --enable-simd=SSE4 --enable-comms=mpi-auto CXXFLAGS='-DMPI_UINT32_T=MPI_UNSIGNED -DMPI_UINT64_T=MPI_UNSIGNED_LONG'; fi
-    - if [[ "$TRAVIS_OS_NAME" == "linux" ]]; then make -j4; fi
-    - if [[ "$TRAVIS_OS_NAME" == "linux" ]]; then mpirun.openmpi -n 2 ./benchmarks/Benchmark_dwf --threads 1 --mpi 2.1.1.1; fi
-
-
+    - make check

Makefile.am

@@ -3,10 +3,15 @@ SUBDIRS = lib benchmarks tests extras
 
 include $(top_srcdir)/doxygen.inc
 
-tests: all
-	$(MAKE) -C tests tests
+bin_SCRIPTS=grid-config
 
-.PHONY: tests doxygen-run doxygen-doc $(DX_PS_GOAL) $(DX_PDF_GOAL)
+
+.PHONY: bench check tests doxygen-run doxygen-doc $(DX_PS_GOAL) $(DX_PDF_GOAL)
+
+tests-local: all
+bench-local: all
+check-local: all
 
 AM_CXXFLAGS += -I$(top_builddir)/include
+
 ACLOCAL_AMFLAGS = -I m4

README.md

@@ -1,41 +1,13 @@
-# 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._
 
-### Bug report
-
-_To help us tracking and solving more efficiently issues with Grid, please report problems using the issue system of GitHub rather than sending emails to Grid developers._
-
-When you file an issue, please go though the following checklist:
-
-1. Check that the code is pointing to the `HEAD` of `develop` or any commit in `master` which is tagged with a version number. 
-2. Give a description of the target platform (CPU, network, compiler). Please give the full CPU part description, using for example `cat /proc/cpuinfo | grep 'model name' | uniq` (Linux) or `sysctl machdep.cpu.brand_string` (macOS) and the full output the `--version` option of your compiler.
-3. Give the exact `configure` command used.
-4. Attach `config.log`.
-5. Attach `config.summary`.
-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.
-
 
 
 ### Description
@@ -58,13 +30,68 @@ optimally use MPI, OpenMP and SIMD parallelism under the hood. This is a signifi
 for most programmers.
 
 The layout transformations are parametrised by the SIMD vector length. This adapts according to the architecture.
-Presently SSE4 (128 bit) AVX, AVX2, QPX (256 bit), IMCI, and AVX512 (512 bit) targets are supported (ARM NEON on the way).
+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. These may be useful in themselves for other programmers.
+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 https://arxiv.org/abs/1512.03487 for more detail.
+Please see [this paper](https://arxiv.org/abs/1512.03487) for more detail.
+
+
+### Compilers
+
+Intel ICPC v16.0.3 and later
+
+Clang v3.5 and later (need 3.8 and later for OpenMP)
+
+GCC   v4.9.x (recommended)
+
+GCC   v6.3 and later
+
+### Important: 
+
+Some versions of GCC appear to have a bug under high optimisation (-O2, -O3).
+
+The safety of these compiler versions cannot be guaranteed at this time. Follow Issue 100 for details and updates.
+
+GCC   v5.x
+
+GCC   v6.1, v6.2
+
+### Bug report
+
+_To help us tracking and solving more efficiently issues with Grid, please report problems using the issue system of GitHub rather than sending emails to Grid developers._
+
+When you file an issue, please go though the following checklist:
+
+1. Check that the code is pointing to the `HEAD` of `develop` or any commit in `master` which is tagged with a version number. 
+2. Give a description of the target platform (CPU, network, compiler). Please give the full CPU part description, using for example `cat /proc/cpuinfo | grep 'model name' | uniq` (Linux) or `sysctl machdep.cpu.brand_string` (macOS) and the full output the `--version` option of your compiler.
+3. Give the exact `configure` command used.
+4. Attach `config.log`.
+5. Attach `grid.config.summary`.
+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/), 
+
+[MPFR](http://www.mpfr.org/) 
+
+Bootstrapping grid downloads and uses for internal dense matrix (non-QCD operations) the Eigen library.
+
+Grid optionally uses:
+
+[HDF5](https://support.hdfgroup.org/HDF5/)  
+
+[LIME](http://usqcd-software.github.io/c-lime/) for ILDG and SciDAC file format support. 
+
+[FFTW](http://www.fftw.org) either generic version or via the Intel MKL library.
+
+LAPACK either generic version or Intel MKL library.
+
 
 ### Quick start
 First, start by cloning the repository:
@@ -95,10 +122,10 @@ install Grid. Other options are detailed in the next section, you can also use `
 `CXX`, `CXXFLAGS`, `LDFLAGS`, ... environment variables can be modified to
 customise the build.
 
-Finally, you can build and install Grid:
+Finally, you can build, check, and install Grid:
 
 ``` bash
-make; make install
+make; make check; make install
 ```
 
 To minimise the build time, only the tests at the root of the `tests` directory are built by default. If you want to build tests in the sub-directory `<subdir>` you can execute:
@@ -121,7 +148,7 @@ If you want to build all the tests at once just use `make tests`.
 - `--enable-gen-simd-width=<size>`: select the size (in bytes) of the generic SIMD vector type (default: 32 bytes).
 - `--enable-precision={single|double}`: set the default precision (default: `double`).
 - `--enable-precision=<comm>`: Use `<comm>` for message passing (default: `none`). A list of possible SIMD targets is detailed in a section below.
-- `--enable-rng={ranlux48|mt19937}`: choose the RNG (default: `ranlux48 `).
+- `--enable-rng={sitmo|ranlux48|mt19937}`: choose the RNG (default: `sitmo `).
 - `--disable-timers`: disable system dependent high-resolution timers.
 - `--enable-chroma`: enable Chroma regression tests.
 - `--enable-doxygen-doc`: enable the Doxygen documentation generation (build with `make doxygen-doc`)
@@ -135,7 +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.  
@@ -153,13 +179,13 @@ 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:
 
 | `<code>`    | Description                            |
 | ----------- | -------------------------------------- |
-| `KNC`       | [Intel Xeon Phi codename Knights Corner](http://ark.intel.com/products/codename/57721/Knights-Corner) |
 | `KNL`       | [Intel Xeon Phi codename Knights Landing](http://ark.intel.com/products/codename/48999/Knights-Landing) |
 | `BGQ`       | Blue Gene/Q                            |
 
@@ -176,21 +202,205 @@ The following configuration is recommended for the Intel Knights Landing platfor
 ``` bash
 ../configure --enable-precision=double\
              --enable-simd=KNL        \
-             --enable-comms=mpi-auto \
-             --with-gmp=<path>        \
-             --with-mpfr=<path>       \
+             --enable-comms=mpi-auto  \
              --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

@@ -1,23 +1,32 @@
 TODO:
 ---------------
 
-Peter's work list:
-2)- Precision conversion and sort out localConvert      <-- 
-3)- Remove DenseVector, DenseMatrix; Use Eigen instead. <-- started 
-4)- Binary I/O speed up & x-strips
--- Profile CG, BlockCG, etc... Flop count/rate -- PARTIAL, time but no flop/s yet
--- Physical propagator interface
--- Conserved currents
--- GaugeFix into central location
--- Multigrid Wilson and DWF, compare to other Multigrid implementations
--- HDCR resume
+Large item work list:
+
+1)- BG/Q port and check ; Andrew says ok.
+2)- Christoph's local basis expansion Lanczos
+--
+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
+-- Binary I/O MPI2 IO                                  <-- DONE
+-- Binary I/O speed up & x-strips                      <-- DONE
 -- Cut down the exterior overhead                      <-- DONE
 -- Interior legs from SHM comms                        <-- DONE
 -- Half-precision comms                                <-- DONE
--- Merge high precision reduction into develop        
--- multiRHS DWF; benchmark on Cori/BNL for comms elimination
+-- Merge high precision reduction into develop         <-- DONE
+-- BlockCG, BCGrQ                                      <-- DONE
+-- multiRHS DWF; benchmark on Cori/BNL for comms elimination <-- DONE
    -- slice* linalg routines for multiRHS, BlockCG    
 
 -----

VERSION

@@ -1,6 +1,5 @@
-Version : 0.6.0
+Version : 0.7.0
 
-- AVX512, AVX2, AVX, SSE good
-- Clang 3.5 and above, ICPC v16 and above, GCC 4.9 and above
-- MPI and MPI3
-- HiRep, Smearing, Generic gauge group
+- Clang 3.5 and above, ICPC v16 and above, GCC 6.3 and above recommended
+- MPI and MPI3 comms optimisations for KNL and OPA finished
+- Half precision comms

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

@@ -31,6 +31,32 @@ using namespace std;
 using namespace Grid;
 using namespace Grid::QCD;
 
+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 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;
+};
+
 int main (int argc, char ** argv)
 {
   Grid_init(&argc,&argv);
@@ -40,17 +66,21 @@ int main (int argc, char ** argv)
   int threads = GridThread::GetThreads();
   std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;
 
-  int Nloop=10;
+  int Nloop=100;
   int nmu=0;
+  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;
+  std::vector<double> t_time(Nloop);
+  time_statistics timestat;
+
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
   std::cout<<GridLogMessage << "= Benchmarking concurrent halo exchange in "<<nmu<<" dimensions"<<std::endl;
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
-  std::cout<<GridLogMessage << "  L  "<<"\t\t"<<" Ls  "<<"\t\t"<<"bytes"<<"\t\t"<<"MB/s uni"<<"\t\t"<<"MB/s bidi"<<std::endl;
-  int maxlat=24;
+  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],
@@ -58,15 +88,23 @@ int main (int argc, char ** argv)
       				    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<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;
+      }
 
-      double start=usecond();
       for(int i=0;i<Nloop;i++){
+      double start=usecond();
 
 	std::vector<CartesianCommunicator::CommsRequest_t> requests;
 
@@ -79,7 +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],
@@ -102,18 +139,24 @@ int main (int argc, char ** argv)
 	}
 	Grid.SendToRecvFromComplete(requests);
 	Grid.Barrier();
-
+	double stop=usecond();
+	t_time[i] = stop-start; // microseconds
       }
-      double stop=usecond();
 
-      double dbytes    = bytes;
-      double xbytes    = Nloop*dbytes*2.0*ncomm;
+      timestat.statistics(t_time);
+
+      double dbytes    = bytes*ppn;
+      double xbytes    = dbytes*2.0*ncomm;
       double rbytes    = xbytes;
       double bidibytes = xbytes+rbytes;
 
-      double time = stop-start; // microseconds
+      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 << lat<<"\t\t"<<Ls<<"\t\t"<<bytes<<"\t\t"<<xbytes/time<<"\t\t"<<bidibytes/time<<std::endl;
     }
   }    
 
@@ -121,25 +164,32 @@ int main (int argc, char ** argv)
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
   std::cout<<GridLogMessage << "= Benchmarking sequential halo exchange in "<<nmu<<" dimensions"<<std::endl;
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
-  std::cout<<GridLogMessage << "  L  "<<"\t\t"<<" Ls  "<<"\t\t"<<"bytes"<<"\t\t"<<"MB/s uni"<<"\t\t"<<"MB/s bidi"<<std::endl;
-
+  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});
 
       GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
+      RealD Nrank = Grid._Nprocessors;
+      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);
 
-      double start=usecond();
       for(int i=0;i<Nloop;i++){
+      double start=usecond();
     
 	ncomm=0;
 	for(int mu=0;mu<4;mu++){
@@ -178,30 +228,37 @@ int main (int argc, char ** argv)
 	  }
 	}
 	Grid.Barrier();
+	double stop=usecond();
+	t_time[i] = stop-start; // microseconds
+
       }
 
-      double stop=usecond();
+      timestat.statistics(t_time);
       
-      double dbytes    = bytes;
-      double xbytes    = Nloop*dbytes*2.0*ncomm;
+      double dbytes    = bytes*ppn;
+      double xbytes    = dbytes*2.0*ncomm;
       double rbytes    = xbytes;
       double bidibytes = xbytes+rbytes;
 
-      double time = stop-start;
+    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 << lat<<"\t\t"<<Ls<<"\t\t"<<bytes<<"\t\t"<<xbytes/time<<"\t\t"<<bidibytes/time<<std::endl;
+      
     }
   }  
 
 
-  Nloop=10;
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
   std::cout<<GridLogMessage << "= Benchmarking concurrent STENCIL halo exchange in "<<nmu<<" dimensions"<<std::endl;
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
-  std::cout<<GridLogMessage << "  L  "<<"\t\t"<<" Ls  "<<"\t\t"<<"bytes"<<"\t\t"<<"MB/s uni"<<"\t\t"<<"MB/s bidi"<<std::endl;
+  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],
@@ -209,6 +266,9 @@ int main (int argc, char ** argv)
       				    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);
@@ -216,73 +276,86 @@ int main (int argc, char ** argv)
       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 start=usecond();
+      double dbytes;
       for(int i=0;i<Nloop;i++){
+	double start=usecond();
+
+	dbytes=0;
+	ncomm=0;
 
 	std::vector<CartesianCommunicator::CommsRequest_t> requests;
 
-	ncomm=0;
 	for(int mu=0;mu<4;mu++){
 	
+
 	  if (mpi_layout[mu]>1 ) {
 	  
 	    ncomm++;
 	    int comm_proc=1;
 	    int xmit_to_rank;
 	    int recv_from_rank;
-	    
 	    Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
-	    Grid.StencilSendToRecvFromBegin(requests,
-					    (void *)&xbuf[mu][0],
-					    xmit_to_rank,
-					    (void *)&rbuf[mu][0],
-					    recv_from_rank,
-					    bytes);
+	    dbytes+=
+	      Grid.StencilSendToRecvFromBegin(requests,
+					      (void *)&xbuf[mu][0],
+					      xmit_to_rank,
+					      (void *)&rbuf[mu][0],
+					      recv_from_rank,
+					      bytes,mu);
 	
 	    comm_proc = mpi_layout[mu]-1;
 	  
 	    Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
-	    Grid.StencilSendToRecvFromBegin(requests,
-					    (void *)&xbuf[mu+4][0],
-					    xmit_to_rank,
-					    (void *)&rbuf[mu+4][0],
-					    recv_from_rank,
-					    bytes);
+	    dbytes+=
+	      Grid.StencilSendToRecvFromBegin(requests,
+					      (void *)&xbuf[mu+4][0],
+					      xmit_to_rank,
+					      (void *)&rbuf[mu+4][0],
+					      recv_from_rank,
+					      bytes,mu+4);
 	  
 	  }
 	}
-	Grid.StencilSendToRecvFromComplete(requests);
+	Grid.StencilSendToRecvFromComplete(requests,0);
 	Grid.Barrier();
-
+	double stop=usecond();
+	t_time[i] = stop-start; // microseconds
+	
       }
-      double stop=usecond();
 
-      double dbytes    = bytes;
-      double xbytes    = Nloop*dbytes*2.0*ncomm;
-      double rbytes    = xbytes;
-      double bidibytes = xbytes+rbytes;
+      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;
 
-      double time = stop-start; // microseconds
 
-      std::cout<<GridLogMessage << lat<<"\t\t"<<Ls<<"\t\t"<<bytes<<"\t\t"<<xbytes/time<<"\t\t"<<bidibytes/time<<std::endl;
     }
   }    
 
 
-
-  Nloop=100;
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
   std::cout<<GridLogMessage << "= Benchmarking sequential STENCIL halo exchange in "<<nmu<<" dimensions"<<std::endl;
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
-  std::cout<<GridLogMessage << "  L  "<<"\t\t"<<" Ls  "<<"\t\t"<<"bytes"<<"\t\t"<<"MB/s uni"<<"\t\t"<<"MB/s bidi"<<std::endl;
+  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],
@@ -290,6 +363,9 @@ int main (int argc, char ** argv)
       				    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);
@@ -297,16 +373,18 @@ int main (int argc, char ** argv)
       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 start=usecond();
+      double dbytes;
       for(int i=0;i<Nloop;i++){
+	double start=usecond();
 
 	std::vector<CartesianCommunicator::CommsRequest_t> requests;
-
+	dbytes=0;
 	ncomm=0;
 	for(int mu=0;mu<4;mu++){
 	
@@ -318,44 +396,146 @@ int main (int argc, char ** argv)
 	    int recv_from_rank;
 	    
 	    Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
-	    Grid.StencilSendToRecvFromBegin(requests,
-					    (void *)&xbuf[mu][0],
-					    xmit_to_rank,
-					    (void *)&rbuf[mu][0],
-					    recv_from_rank,
-					    bytes);
-	    Grid.StencilSendToRecvFromComplete(requests);
+	    dbytes+=
+	      Grid.StencilSendToRecvFromBegin(requests,
+					      (void *)&xbuf[mu][0],
+					      xmit_to_rank,
+					      (void *)&rbuf[mu][0],
+					      recv_from_rank,
+					      bytes,mu);
+	    Grid.StencilSendToRecvFromComplete(requests,mu);
 	    requests.resize(0);
 
 	    comm_proc = mpi_layout[mu]-1;
 	  
 	    Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
-	    Grid.StencilSendToRecvFromBegin(requests,
-					    (void *)&xbuf[mu+4][0],
-					    xmit_to_rank,
-					    (void *)&rbuf[mu+4][0],
-					    recv_from_rank,
-					    bytes);
-	    Grid.StencilSendToRecvFromComplete(requests);
+	    dbytes+=
+	      Grid.StencilSendToRecvFromBegin(requests,
+					      (void *)&xbuf[mu+4][0],
+					      xmit_to_rank,
+					      (void *)&rbuf[mu+4][0],
+					      recv_from_rank,
+					      bytes,mu+4);
+	    Grid.StencilSendToRecvFromComplete(requests,mu+4);
 	    requests.resize(0);
 	  
 	  }
 	}
 	Grid.Barrier();
-
+	double stop=usecond();
+	t_time[i] = stop-start; // microseconds
+	
       }
-      double stop=usecond();
 
-      double dbytes    = bytes;
-      double xbytes    = Nloop*dbytes*2.0*ncomm;
-      double rbytes    = xbytes;
-      double bidibytes = xbytes+rbytes;
+      timestat.statistics(t_time);
 
-      double time = stop-start; // microseconds
+      dbytes=dbytes*ppn;
+      double xbytes    = dbytes*0.5;
+      double rbytes    = dbytes*0.5;
+      double bidibytes = dbytes;
 
-      std::cout<<GridLogMessage << lat<<"\t\t"<<Ls<<"\t\t"<<bytes<<"\t\t"<<xbytes/time<<"\t\t"<<bidibytes/time<<std::endl;
+
+      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 << "= 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

@@ -1,28 +1,22 @@
-    /*************************************************************************************
-
+ /*************************************************************************************
     Grid physics library, www.github.com/paboyle/Grid 
-
     Source file: ./benchmarks/Benchmark_dwf.cc
-
     Copyright (C) 2015
 
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: paboyle <paboyle@ph.ed.ac.uk>
+    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 */
@@ -57,7 +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);
@@ -151,6 +151,7 @@ int main (int argc, char ** argv)
   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;
@@ -160,13 +161,17 @@ int main (int argc, char ** argv)
   std::cout << GridLogMessage<< "* Vectorising space-time by "<<vComplex::Nsimd()<<std::endl;
   if ( sizeof(Real)==4 )   std::cout << GridLogMessage<< "* SINGLE precision "<<std::endl;
   if ( sizeof(Real)==8 )   std::cout << GridLogMessage<< "* DOUBLE precision "<<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;
 
   DomainWallFermionR Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
-  int ncall =1000;
+  int ncall =500;
   if (1) {
     FGrid->Barrier();
     Dw.ZeroCounters();
@@ -189,6 +194,7 @@ int main (int argc, char ** argv)
     //    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;
     err = ref-result; 
     std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
 
@@ -225,6 +231,7 @@ int main (int argc, char ** argv)
     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;
     err = ref-result; 
     std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
 
@@ -240,6 +247,10 @@ int main (int argc, char ** argv)
     std::cout << GridLogMessage<< "* Vectorising fifth dimension by "<<vComplex::Nsimd()<<std::endl;
     if ( sizeof(Real)==4 )   std::cout << GridLogMessage<< "* SINGLE precision "<<std::endl;
     if ( sizeof(Real)==8 )   std::cout << GridLogMessage<< "* DOUBLE precision "<<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;
@@ -271,6 +282,7 @@ int main (int argc, char ** argv)
     std::cout<<GridLogMessage << "Called Dw s_inner "<<ncall<<" times in "<<t1-t0<<" us"<<std::endl;
     std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t1-t0)<<std::endl;
     std::cout<<GridLogMessage << "mflop/s per rank =  "<< flops/(t1-t0)/NP<<std::endl;
+    std::cout<<GridLogMessage << "mflop/s per node =  "<< flops/(t1-t0)/NN<<std::endl;
     //    std::cout<<GridLogMessage<< "res norms "<< norm2(result)<<" " <<norm2(sresult)<<std::endl;
     sDw.Report();
     RealD sum=0;
@@ -296,6 +308,7 @@ int main (int argc, char ** argv)
       std::cout<< "sD ERR   \n " << err  <<std::endl;
     }
     assert(sum < 1.0e-4);
+
     
     if(1){
       std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
@@ -303,6 +316,10 @@ int main (int argc, char ** argv)
       std::cout << GridLogMessage<< "* Vectorising fifth dimension by "<<vComplex::Nsimd()<<std::endl;
       if ( sizeof(Real)==4 )   std::cout << GridLogMessage<< "* SINGLE precision "<<std::endl;
       if ( sizeof(Real)==8 )   std::cout << GridLogMessage<< "* DOUBLE precision "<<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) 
@@ -342,6 +359,7 @@ int main (int argc, char ** argv)
 
       std::cout<<GridLogMessage << "sDeo mflop/s =   "<< flops/(t1-t0)<<std::endl;
       std::cout<<GridLogMessage << "sDeo mflop/s per rank   "<< flops/(t1-t0)/NP<<std::endl;
+      std::cout<<GridLogMessage << "sDeo mflop/s per node   "<< flops/(t1-t0)/NN<<std::endl;
       sDw.Report();
 
       sDw.DhopEO(ssrc_o,sr_e,DaggerNo);
@@ -370,8 +388,23 @@ 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
     ref = zero;
@@ -420,14 +453,15 @@ int main (int argc, char ** argv)
 
 
   // S-direction is INNERMOST and takes no part in the parity.
-  static int Opt;  // these are a temporary hack
-  static int Comms;  // these are a temporary hack
-
   std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
   std::cout << GridLogMessage<< "* Benchmarking DomainWallFermionR::DhopEO                "<<std::endl;
   std::cout << GridLogMessage<< "* Vectorising space-time by "<<vComplex::Nsimd()<<std::endl;
   if ( sizeof(Real)==4 )   std::cout << GridLogMessage<< "* SINGLE precision "<<std::endl;
   if ( sizeof(Real)==8 )   std::cout << GridLogMessage<< "* DOUBLE precision "<<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;
@@ -448,6 +482,7 @@ int main (int argc, char ** argv)
 
     std::cout<<GridLogMessage << "Deo mflop/s =   "<< flops/(t1-t0)<<std::endl;
     std::cout<<GridLogMessage << "Deo mflop/s per rank   "<< flops/(t1-t0)/NP<<std::endl;
+    std::cout<<GridLogMessage << "Deo mflop/s per node   "<< flops/(t1-t0)/NN<<std::endl;
     Dw.Report();
   }
   Dw.DhopEO(src_o,r_e,DaggerNo);
@@ -474,8 +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=44;
-#define NLOOP (1*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

@@ -35,13 +35,14 @@ using namespace Grid::QCD;
 int main (int argc, char ** argv)
 {
   Grid_init(&argc,&argv);
+#define LMAX (64)
 
-  int Nloop=1000;
+  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;
@@ -50,19 +51,19 @@ int main (int argc, char ** argv)
   std::cout<<GridLogMessage << "  L  "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s\t\t GFlop/s"<<std::endl;
   std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
 
-  for(int lat=2;lat<=32;lat+=2){
+  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();
@@ -82,20 +83,20 @@ int main (int argc, char ** argv)
   std::cout<<GridLogMessage << "  L  "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s\t\t GFlop/s"<<std::endl;
   std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
 
-  for(int lat=2;lat<=32;lat+=2){
+  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();
@@ -113,20 +114,20 @@ int main (int argc, char ** argv)
   std::cout<<GridLogMessage << "  L  "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s\t\t GFlop/s"<<std::endl;
   std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
 
-  for(int lat=2;lat<=32;lat+=2){
+  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();
@@ -144,20 +145,20 @@ int main (int argc, char ** argv)
   std::cout<<GridLogMessage << "  L  "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s\t\t GFlop/s"<<std::endl;
   std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
 
-  for(int lat=2;lat<=32;lat+=2){
+  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);

benchmarks/Makefile.am

@@ -1,11 +1,7 @@
 include Make.inc
 
-simple: simple_su3_test.o simple_su3_expr.o simple_simd_test.o
-
-EXTRA_LIBRARIES = libsimple_su3_test.a libsimple_su3_expr.a libsimple_simd_test.a
-
-libsimple_su3_test_a_SOURCES = simple_su3_test.cc
-
-libsimple_su3_expr_a_SOURCES = simple_su3_expr.cc
-
-libsimple_simd_test_a_SOURCES = simple_simd_test.cc
+bench-local: all
+	./Benchmark_su3
+	./Benchmark_memory_bandwidth
+	./Benchmark_wilson
+	./Benchmark_dwf --dslash-unroll

bootstrap.sh

@@ -1,11 +1,11 @@
 #!/usr/bin/env bash
 
-EIGEN_URL='http://bitbucket.org/eigen/eigen/get/3.2.9.tar.bz2'
+EIGEN_URL='http://bitbucket.org/eigen/eigen/get/3.3.3.tar.bz2'
 
 echo "-- deploying Eigen source..."
 wget ${EIGEN_URL} --no-check-certificate
 ./scripts/update_eigen.sh `basename ${EIGEN_URL}`
-rm `basename ${EIGEN_URL}`
+#rm `basename ${EIGEN_URL}`
 
 echo '-- generating Make.inc files...'
 ./scripts/filelist

configure.ac

@@ -1,16 +1,23 @@
 AC_PREREQ([2.63])
-AC_INIT([Grid], [0.6.0], [https://github.com/paboyle/Grid], [Grid])
+AC_INIT([Grid], [0.7.0], [https://github.com/paboyle/Grid], [Grid])
 AC_CANONICAL_BUILD
 AC_CANONICAL_HOST
 AC_CANONICAL_TARGET
-AM_INIT_AUTOMAKE(subdir-objects)
+AM_INIT_AUTOMAKE([subdir-objects 1.13])
+AM_EXTRA_RECURSIVE_TARGETS([tests bench])
 AC_CONFIG_MACRO_DIR([m4])
 AC_CONFIG_SRCDIR([lib/Grid.h])
 AC_CONFIG_HEADERS([lib/Config.h],[sed -i 's|PACKAGE_|GRID_|' lib/Config.h])
 m4_ifdef([AM_SILENT_RULES], [AM_SILENT_RULES([yes])])
 
-############### Checks for programs
+################ 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
 
@@ -24,12 +31,14 @@ AX_GXX_VERSION
 AC_DEFINE_UNQUOTED([GXX_VERSION],["$GXX_VERSION"],
       [version of g++ that will compile the code])
 
+
+
 ############### Checks for typedefs, structures, and compiler characteristics
 AC_TYPE_SIZE_T
 AC_TYPE_UINT32_T
 AC_TYPE_UINT64_T
 
-############### OpenMP 
+############### OpenMP
 AC_OPENMP
 ac_openmp=no
 if test "${OPENMP_CXXFLAGS}X" != "X"; then
@@ -45,9 +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],
@@ -60,16 +74,23 @@ AC_ARG_WITH([mpfr],
     [AM_CXXFLAGS="-I$with_mpfr/include $AM_CXXFLAGS"]
     [AM_LDFLAGS="-L$with_mpfr/lib $AM_LDFLAGS"])
 
-############### FFTW3 
-AC_ARG_WITH([fftw],    
+############### FFTW3
+AC_ARG_WITH([fftw],
             [AS_HELP_STRING([--with-fftw=prefix],
             [try this for a non-standard install prefix of the FFTW3 library])],
             [AM_CXXFLAGS="-I$with_fftw/include $AM_CXXFLAGS"]
             [AM_LDFLAGS="-L$with_fftw/lib $AM_LDFLAGS"])
 
-############### lapack 
+############### LIME
+AC_ARG_WITH([lime],
+            [AS_HELP_STRING([--with-lime=prefix],
+            [try this for a non-standard install prefix of the LIME library])],
+            [AM_CXXFLAGS="-I$with_lime/include $AM_CXXFLAGS"]
+            [AM_LDFLAGS="-L$with_lime/lib $AM_LDFLAGS"])
+
+############### lapack
 AC_ARG_ENABLE([lapack],
-    [AC_HELP_STRING([--enable-lapack=yes|no|prefix], [enable LAPACK])], 
+    [AC_HELP_STRING([--enable-lapack=yes|no|prefix], [enable LAPACK])],
     [ac_LAPACK=${enable_lapack}], [ac_LAPACK=no])
 
 case ${ac_LAPACK} in
@@ -85,7 +106,7 @@ esac
 
 ############### FP16 conversions
 AC_ARG_ENABLE([sfw-fp16],
-    [AC_HELP_STRING([--enable-sfw-fp16=yes|no], [enable software fp16 comms])], 
+    [AC_HELP_STRING([--enable-sfw-fp16=yes|no], [enable software fp16 comms])],
     [ac_SFW_FP16=${enable_sfw_fp16}], [ac_SFW_FP16=yes])
 case ${ac_SFW_FP16} in
     yes)
@@ -120,7 +141,7 @@ AC_ARG_WITH([hdf5],
 
 ############### first-touch
 AC_ARG_ENABLE([numa],
-    [AC_HELP_STRING([--enable-numa=yes|no|prefix], [enable first touch numa opt])], 
+    [AC_HELP_STRING([--enable-numa=yes|no|prefix], [enable first touch numa opt])],
     [ac_NUMA=${enable_NUMA}],[ac_NUMA=no])
 
 case ${ac_NUMA} in
@@ -146,8 +167,8 @@ if test "${ac_MKL}x" != "nox"; then
 fi
 
 AC_SEARCH_LIBS([__gmpf_init], [gmp],
-               [AC_SEARCH_LIBS([mpfr_init], [mpfr], 
-                               [AC_DEFINE([HAVE_LIBMPFR], [1], 
+               [AC_SEARCH_LIBS([mpfr_init], [mpfr],
+                               [AC_DEFINE([HAVE_LIBMPFR], [1],
                                           [Define to 1 if you have the `MPFR' library])]
                                [have_mpfr=true], [AC_MSG_ERROR([MPFR library not found])])]
                [AC_DEFINE([HAVE_LIBGMP], [1], [Define to 1 if you have the `GMP' library])]
@@ -156,7 +177,7 @@ AC_SEARCH_LIBS([__gmpf_init], [gmp],
 if test "${ac_LAPACK}x" != "nox"; then
     AC_SEARCH_LIBS([LAPACKE_sbdsdc], [lapack], [],
                    [AC_MSG_ERROR("LAPACK enabled but library not found")])
-fi   
+fi
 
 AC_SEARCH_LIBS([fftw_execute], [fftw3],
                [AC_SEARCH_LIBS([fftwf_execute], [fftw3f], [],
@@ -164,6 +185,23 @@ AC_SEARCH_LIBS([fftw_execute], [fftw3],
                [AC_DEFINE([HAVE_FFTW], [1], [Define to 1 if you have the `FFTW' library])]
                [have_fftw=true])
 
+AC_SEARCH_LIBS([limeCreateReader], [lime],
+               [AC_DEFINE([HAVE_LIME], [1], [Define to 1 if you have the `LIME' library])]
+               [have_lime=true],
+	       [AC_MSG_WARN(C-LIME library was not found in your system.
+In order to use ILGG file format please install or provide the correct path to your installation
+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] [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]
@@ -216,6 +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])
@@ -223,6 +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='';;
@@ -251,6 +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])
@@ -288,8 +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])
@@ -299,14 +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'
@@ -316,7 +392,7 @@ case ${ac_COMMS} in
         comms_type='shmem'
      ;;
      *)
-        AC_MSG_ERROR([${ac_COMMS} unsupported --enable-comms option]); 
+        AC_MSG_ERROR([${ac_COMMS} unsupported --enable-comms option]);
      ;;
 esac
 case ${ac_COMMS} in
@@ -334,7 +410,7 @@ esac
 AM_CONDITIONAL(BUILD_COMMS_SHMEM, [ test "${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
@@ -353,7 +429,7 @@ case ${ac_RNG} in
       AC_DEFINE([RNG_SITMO],[1],[RNG_SITMO] )
      ;;
      *)
-      AC_MSG_ERROR([${ac_RNG} unsupported --enable-rng option]); 
+      AC_MSG_ERROR([${ac_RNG} unsupported --enable-rng option]);
      ;;
 esac
 
@@ -370,7 +446,7 @@ case ${ac_TIMERS} in
       AC_DEFINE([TIMERS_OFF],[1],[TIMERS_OFF] )
      ;;
      *)
-      AC_MSG_ERROR([${ac_TIMERS} unsupported --enable-timers option]); 
+      AC_MSG_ERROR([${ac_TIMERS} unsupported --enable-timers option]);
      ;;
 esac
 
@@ -382,7 +458,7 @@ case ${ac_CHROMA} in
      yes|no)
      ;;
      *)
-       AC_MSG_ERROR([${ac_CHROMA} unsupported --enable-chroma option]); 
+       AC_MSG_ERROR([${ac_CHROMA} unsupported --enable-chroma option]);
      ;;
 esac
 
@@ -403,12 +479,67 @@ DX_INIT_DOXYGEN([$PACKAGE_NAME], [doxygen.cfg])
 
 ############### Ouput
 cwd=`pwd -P`; cd ${srcdir}; abs_srcdir=`pwd -P`; cd ${cwd}
+GRID_CXXFLAGS="$AM_CXXFLAGS $CXXFLAGS"
+GRID_LDFLAGS="$AM_LDFLAGS $LDFLAGS"
+GRID_LIBS=$LIBS
+GRID_SHORT_SHA=`git rev-parse --short HEAD`
+GRID_SHA=`git rev-parse HEAD`
+GRID_BRANCH=`git rev-parse --abbrev-ref HEAD`
 AM_CXXFLAGS="-I${abs_srcdir}/include $AM_CXXFLAGS"
 AM_CFLAGS="-I${abs_srcdir}/include $AM_CFLAGS"
 AM_LDFLAGS="-L${cwd}/lib $AM_LDFLAGS"
 AC_SUBST([AM_CFLAGS])
 AC_SUBST([AM_CXXFLAGS])
 AC_SUBST([AM_LDFLAGS])
+AC_SUBST([GRID_CXXFLAGS])
+AC_SUBST([GRID_LDFLAGS])
+AC_SUBST([GRID_LIBS])
+AC_SUBST([GRID_SHA])
+AC_SUBST([GRID_BRANCH])
+
+git_commit=`cd $srcdir && ./scripts/configure.commit`
+
+echo "~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Summary of configuration for $PACKAGE v$VERSION
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+----- GIT VERSION -------------------------------------
+$git_commit
+----- PLATFORM ----------------------------------------
+architecture (build)        : $build_cpu
+os (build)                  : $build_os
+architecture (target)       : $target_cpu
+os (target)                 : $target_os
+compiler vendor             : ${ax_cv_cxx_compiler_vendor}
+compiler version            : ${ax_cv_gxx_version}
+----- BUILD OPTIONS -----------------------------------
+SIMD                        : ${ac_SIMD}${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                  : ${ac_RNG}
+GMP                         : `if test "x$have_gmp" = xtrue; then echo yes; else echo no; fi`
+LAPACK                      : ${ac_LAPACK}
+FFTW                        : `if test "x$have_fftw" = xtrue; then echo yes; else echo no; fi`
+LIME (ILDG support)         : `if test "x$have_lime" = xtrue; then echo yes; else echo no; fi`
+HDF5                        : `if test "x$have_hdf5" = xtrue; then echo yes; else echo no; fi`
+build DOXYGEN documentation : `if test "$DX_FLAG_doc" = '1'; then echo yes; else echo no; fi`
+----- BUILD FLAGS -------------------------------------
+CXXFLAGS:
+`echo ${AM_CXXFLAGS} ${CXXFLAGS} | tr ' ' '\n' | sed 's/^-/    -/g'`
+LDFLAGS:
+`echo ${AM_LDFLAGS} ${LDFLAGS} | tr ' ' '\n' | sed 's/^-/    -/g'`
+LIBS:
+`echo ${LIBS} | tr ' ' '\n' | sed 's/^-/    -/g'`
+-------------------------------------------------------" > grid.configure.summary
+
+GRID_SUMMARY="`cat grid.configure.summary`"
+AM_SUBST_NOTMAKE([GRID_SUMMARY])
+AC_SUBST([GRID_SUMMARY])
+
+AC_CONFIG_FILES([grid-config], [chmod +x grid-config])
 AC_CONFIG_FILES(Makefile)
 AC_CONFIG_FILES(lib/Makefile)
 AC_CONFIG_FILES(tests/Makefile)
@@ -419,6 +550,8 @@ 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)
 AC_CONFIG_FILES(benchmarks/Makefile)
@@ -426,36 +559,7 @@ AC_CONFIG_FILES(extras/Makefile)
 AC_CONFIG_FILES(extras/Hadrons/Makefile)
 AC_OUTPUT
 
-echo "~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-Summary of configuration for $PACKAGE v$VERSION
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
------ PLATFORM ----------------------------------------
-architecture (build)        : $build_cpu
-os (build)                  : $build_os
-architecture (target)       : $target_cpu
-os (target)                 : $target_os
-compiler vendor             : ${ax_cv_cxx_compiler_vendor}
-compiler version            : ${ax_cv_gxx_version}
------ BUILD OPTIONS -----------------------------------
-SIMD                        : ${ac_SIMD}${SIMD_GEN_WIDTH_MSG}
-Threading                   : ${ac_openmp} 
-Communications type         : ${comms_type}
-Default precision           : ${ac_PRECISION}
-Software FP16 conversion    : ${ac_SFW_FP16}
-RNG choice                  : ${ac_RNG} 
-GMP                         : `if test "x$have_gmp" = xtrue; then echo yes; else echo no; fi`
-LAPACK                      : ${ac_LAPACK}
-FFTW                        : `if test "x$have_fftw" = xtrue; then echo yes; else echo no; fi`
-HDF5                        : `if test "x$have_hdf5" = xtrue; then echo yes; else echo no; fi`
-build DOXYGEN documentation : `if test "$DX_FLAG_doc" = '1'; then echo yes; else echo no; fi`
------ BUILD FLAGS -------------------------------------
-CXXFLAGS:
-`echo ${AM_CXXFLAGS} ${CXXFLAGS} | tr ' ' '\n' | sed 's/^-/    -/g'`
-LDFLAGS:
-`echo ${AM_LDFLAGS} ${LDFLAGS} | tr ' ' '\n' | sed 's/^-/    -/g'`
-LIBS:
-`echo ${LIBS} | tr ' ' '\n' | sed 's/^-/    -/g'`
--------------------------------------------------------" > config.summary
 echo ""
-cat config.summary
+cat grid.configure.summary
 echo ""
+

extras/Hadrons/Application.cc

@@ -162,7 +162,8 @@ void Application::saveParameterFile(const std::string parameterFileName)
 sizeString((size)*locVol_) << " (" << sizeString(size)  << "/site)"
 
 #define DEFINE_MEMPEAK \
-auto memPeak = [this](const std::vector<unsigned int> &program)\
+GeneticScheduler<unsigned int>::ObjFunc memPeak = \
+[this](const std::vector<unsigned int> &program)\
 {\
     unsigned int memPeak;\
     bool         msg;\

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());
@@ -492,7 +518,14 @@ std::string Environment::getObjectType(const unsigned int address) const
 {
     if (hasRegisteredObject(address))
     {
-        return typeName(object_[address].type);
+        if (object_[address].type)
+        {
+            return typeName(object_[address].type);
+        }
+        else
+        {
+            return "<no type>";
+        }
     }
     else if (hasObject(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
 {
@@ -145,6 +165,15 @@ std::string typeName(void)
     return typeName(typeIdPt<T>());
 }
 
+// default writers/readers
+#ifdef HAVE_HDF5
+typedef Hdf5Reader CorrReader;
+typedef Hdf5Writer CorrWriter;
+#else
+typedef XmlReader CorrReader;
+typedef XmlWriter CorrWriter;
+#endif
+
 END_HADRONS_NAMESPACE
 
 #endif // Hadrons_Global_hpp_

extras/Hadrons/Modules.hpp

@@ -1,40 +1,25 @@
-/*************************************************************************************
-
-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>
+#include <Grid/Hadrons/Modules/MContraction/DiscLoop.hpp>
+#include <Grid/Hadrons/Modules/MContraction/Gamma3pt.hpp>
 #include <Grid/Hadrons/Modules/MContraction/Meson.hpp>
+#include <Grid/Hadrons/Modules/MContraction/WeakHamiltonian.hpp>
+#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>
@@ -48,14 +48,15 @@ public:
                                     std::string, gauge,
                                     unsigned int, Ls,
                                     double      , mass,
-                                    double      , M5);
+                                    double      , M5,
+                                    std::string , boundary);
 };
 
 template <typename FImpl>
 class TDWF: public Module<DWFPar>
 {
 public:
-    TYPE_ALIASES(FImpl,);
+    FGS_TYPE_ALIASES(FImpl,);
 public:
     // constructor
     TDWF(const std::string name);
@@ -116,14 +117,19 @@ void TDWF<FImpl>::execute(void)
                  << par().mass << ", M5= " << par().M5 << " and Ls= "
                  << par().Ls << " using gauge field '" << par().gauge << "'"
                  << std::endl;
+    LOG(Message) << "Fermion boundary conditions: " << par().boundary 
+                 << std::endl;
     env().createGrid(par().Ls);
     auto &U      = *env().template getObject<LatticeGaugeField>(par().gauge);
     auto &g4     = *env().getGrid();
     auto &grb4   = *env().getRbGrid();
     auto &g5     = *env().getGrid(par().Ls);
     auto &grb5   = *env().getRbGrid(par().Ls);
+    std::vector<Complex> boundary = strToVec<Complex>(par().boundary);
+    typename DomainWallFermion<FImpl>::ImplParams implParams(boundary);
     FMat *fMatPt = new DomainWallFermion<FImpl>(U, g5, grb5, g4, grb4,
-                                                par().mass, par().M5);
+                                                par().mass, par().M5,
+                                                implParams);
     env().setObject(getName(), fMatPt);
 }
 
@@ -131,4 +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>
@@ -46,14 +46,15 @@ class WilsonPar: Serializable
 public:
     GRID_SERIALIZABLE_CLASS_MEMBERS(WilsonPar,
                                     std::string, gauge,
-                                    double     , mass);
+                                    double     , mass,
+                                    std::string, boundary);
 };
 
 template <typename FImpl>
 class TWilson: public Module<WilsonPar>
 {
 public:
-    TYPE_ALIASES(FImpl,);
+    FGS_TYPE_ALIASES(FImpl,);
 public:
     // constructor
     TWilson(const std::string name);
@@ -112,10 +113,15 @@ void TWilson<FImpl>::execute()
 {
     LOG(Message) << "Setting up TWilson fermion matrix with m= " << par().mass
                  << " using gauge field '" << par().gauge << "'" << std::endl;
+    LOG(Message) << "Fermion boundary conditions: " << par().boundary 
+                 << std::endl;
     auto &U      = *env().template getObject<LatticeGaugeField>(par().gauge);
     auto &grid   = *env().getGrid();
     auto &gridRb = *env().getRbGrid();
-    FMat *fMatPt = new WilsonFermion<FImpl>(U, grid, gridRb, par().mass);
+    std::vector<Complex> boundary = strToVec<Complex>(par().boundary);
+    typename WilsonFermion<FImpl>::ImplParams implParams(boundary);
+    FMat *fMatPt = new WilsonFermion<FImpl>(U, grid, gridRb, par().mass,
+                                            implParams);
     env().setObject(getName(), fMatPt);
 }
 

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:
@@ -112,7 +112,7 @@ void TBaryon<FImpl1, FImpl2, FImpl3>::execute(void)
                  << " quarks '" << par().q1 << "', '" << par().q2 << "', and '"
                  << par().q3 << "'" << std::endl;
     
-    XmlWriter             writer(par().output);
+    CorrWriter             writer(par().output);
     PropagatorField1      &q1 = *env().template getObject<PropagatorField1>(par().q1);
     PropagatorField2      &q2 = *env().template getObject<PropagatorField2>(par().q2);
     PropagatorField3      &q3 = *env().template getObject<PropagatorField3>(par().q2);
@@ -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

@@ -0,0 +1,144 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MContraction/DiscLoop.hpp
+
+Copyright (C) 2017
+
+Author: Andrew Lawson    <andrew.lawson1991@gmail.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_MContraction_DiscLoop_hpp_
+#define Hadrons_MContraction_DiscLoop_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                                DiscLoop                                    *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MContraction)
+
+class DiscLoopPar: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(DiscLoopPar,
+                                    std::string,    q_loop,
+                                    Gamma::Algebra, gamma,
+                                    std::string,    output);
+};
+
+template <typename FImpl>
+class TDiscLoop: public Module<DiscLoopPar>
+{
+    FERM_TYPE_ALIASES(FImpl,);
+    class Result: Serializable
+    {
+    public:
+        GRID_SERIALIZABLE_CLASS_MEMBERS(Result,
+                                        Gamma::Algebra, gamma,
+                                        std::vector<Complex>, corr);
+    };
+public:
+    // constructor
+    TDiscLoop(const std::string name);
+    // destructor
+    virtual ~TDiscLoop(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(DiscLoop, TDiscLoop<FIMPL>, MContraction);
+
+/******************************************************************************
+ *                       TDiscLoop implementation                             *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+template <typename FImpl>
+TDiscLoop<FImpl>::TDiscLoop(const std::string name)
+: Module<DiscLoopPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+template <typename FImpl>
+std::vector<std::string> TDiscLoop<FImpl>::getInput(void)
+{
+    std::vector<std::string> in = {par().q_loop};
+    
+    return in;
+}
+
+template <typename FImpl>
+std::vector<std::string> TDiscLoop<FImpl>::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TDiscLoop<FImpl>::setup(void)
+{
+    
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TDiscLoop<FImpl>::execute(void)
+{
+    LOG(Message) << "Computing disconnected loop contraction '" << getName() 
+                 << "' using '" << par().q_loop << "' with " << par().gamma 
+                 << " insertion." << std::endl;
+
+    CorrWriter            writer(par().output);
+    PropagatorField       &q_loop = *env().template getObject<PropagatorField>(par().q_loop);
+    LatticeComplex        c(env().getGrid());
+    Gamma                 gamma(par().gamma);
+    std::vector<TComplex> buf;
+    Result                result;
+
+    c = trace(gamma*q_loop);
+    sliceSum(c, buf, Tp);
+
+    result.gamma = par().gamma;
+    result.corr.resize(buf.size());
+    for (unsigned int t = 0; t < buf.size(); ++t)
+    {
+        result.corr[t] = TensorRemove(buf[t]);
+    }
+
+    write(writer, "disc", result);
+}
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_MContraction_DiscLoop_hpp_

extras/Hadrons/Modules/MContraction/Gamma3pt.hpp

@@ -0,0 +1,170 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MContraction/Gamma3pt.hpp
+
+Copyright (C) 2017
+
+Author: Andrew Lawson    <andrew.lawson1991@gmail.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_MContraction_Gamma3pt_hpp_
+#define Hadrons_MContraction_Gamma3pt_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/*
+ * 3pt contraction with gamma matrix insertion.
+ *
+ * Schematic:
+ *
+ *             q2           q3
+ *        /----<------*------<----¬
+ *       /          gamma          \
+ *      /                           \
+ *   i *                            * f
+ *      \                          /
+ *       \                        /
+ *        \----------->----------/
+ *                   q1
+ *
+ *      trace(g5*q1*adj(q2)*g5*gamma*q3)
+ */
+
+/******************************************************************************
+ *                               Gamma3pt                                     *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MContraction)
+
+class Gamma3ptPar: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(Gamma3ptPar,
+                                    std::string,    q1,
+                                    std::string,    q2,
+                                    std::string,    q3,
+                                    Gamma::Algebra, gamma,
+                                    std::string,    output);
+};
+
+template <typename FImpl1, typename FImpl2, typename FImpl3>
+class TGamma3pt: public Module<Gamma3ptPar>
+{
+    FERM_TYPE_ALIASES(FImpl1, 1);
+    FERM_TYPE_ALIASES(FImpl2, 2);
+    FERM_TYPE_ALIASES(FImpl3, 3);
+    class Result: Serializable
+    {
+    public:
+        GRID_SERIALIZABLE_CLASS_MEMBERS(Result,
+                                        Gamma::Algebra, gamma,
+                                        std::vector<Complex>, corr);
+    };
+public:
+    // constructor
+    TGamma3pt(const std::string name);
+    // destructor
+    virtual ~TGamma3pt(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(Gamma3pt, ARG(TGamma3pt<FIMPL, FIMPL, FIMPL>), MContraction);
+
+/******************************************************************************
+ *                       TGamma3pt implementation                             *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+template <typename FImpl1, typename FImpl2, typename FImpl3>
+TGamma3pt<FImpl1, FImpl2, FImpl3>::TGamma3pt(const std::string name)
+: Module<Gamma3ptPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+template <typename FImpl1, typename FImpl2, typename FImpl3>
+std::vector<std::string> TGamma3pt<FImpl1, FImpl2, FImpl3>::getInput(void)
+{
+    std::vector<std::string> in = {par().q1, par().q2, par().q3};
+    
+    return in;
+}
+
+template <typename FImpl1, typename FImpl2, typename FImpl3>
+std::vector<std::string> TGamma3pt<FImpl1, FImpl2, FImpl3>::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+template <typename FImpl1, typename FImpl2, typename FImpl3>
+void TGamma3pt<FImpl1, FImpl2, FImpl3>::setup(void)
+{
+    
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+template <typename FImpl1, typename FImpl2, typename FImpl3>
+void TGamma3pt<FImpl1, FImpl2, FImpl3>::execute(void)
+{
+    LOG(Message) << "Computing 3pt contractions '" << getName() << "' using"
+                 << " quarks '" << par().q1 << "', '" << par().q2 << "' and '"
+                 << par().q3 << "', with " << par().gamma << " insertion." 
+                 << std::endl;
+
+    CorrWriter            writer(par().output);
+    PropagatorField1      &q1 = *env().template getObject<PropagatorField1>(par().q1);
+    PropagatorField2      &q2 = *env().template getObject<PropagatorField2>(par().q2);
+    PropagatorField3      &q3 = *env().template getObject<PropagatorField3>(par().q3);
+    LatticeComplex        c(env().getGrid());
+    Gamma                 g5(Gamma::Algebra::Gamma5);
+    Gamma                 gamma(par().gamma);
+    std::vector<TComplex> buf;
+    Result                result;
+
+    c = trace(g5*q1*adj(q2)*(g5*gamma)*q3);
+    sliceSum(c, buf, Tp);
+
+    result.gamma = par().gamma;
+    result.corr.resize(buf.size());
+    for (unsigned int t = 0; t < buf.size(); ++t)
+    {
+        result.corr[t] = TensorRemove(buf[t]);
+    }
+
+    write(writer, "gamma3pt", result);
+}
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_MContraction_Gamma3pt_hpp_

extras/Hadrons/Modules/MContraction/Meson.hpp

@@ -6,8 +6,10 @@ Source file: extras/Hadrons/Modules/MContraction/Meson.hpp
 
 Copyright (C) 2015
 Copyright (C) 2016
+Copyright (C) 2017
 
 Author: Antonin Portelli <antonin.portelli@me.com>
+        Andrew Lawson    <andrew.lawson1991@gmail.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
@@ -27,8 +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>
@@ -36,32 +38,56 @@ See the full license in the file "LICENSE" in the top level distribution directo
 
 BEGIN_HADRONS_NAMESPACE
 
+/*
+ 
+ Meson contractions
+ -----------------------------
+ 
+ * options:
+ - q1: input propagator 1 (string)
+ - q2: input propagator 2 (string)
+ - gammas: gamma products to insert at sink & source, pairs of gamma matrices 
+           (space-separated strings) in angled brackets (i.e. <g_sink g_src>),
+           in a sequence (e.g. "<Gamma5 Gamma5><Gamma5 GammaT>").
+
+           Special values: "all" - perform all possible contractions.
+ - mom: momentum insertion, space-separated float sequence (e.g ".1 .2 1. 0."),
+        given as multiples of (2*pi) / L.
+*/
+
 /******************************************************************************
  *                                TMeson                                       *
  ******************************************************************************/
 BEGIN_MODULE_NAMESPACE(MContraction)
 
+typedef std::pair<Gamma::Algebra, Gamma::Algebra> GammaPair;
+
 class MesonPar: Serializable
 {
 public:
     GRID_SERIALIZABLE_CLASS_MEMBERS(MesonPar,
-                                    std::string,    q1,
-                                    std::string,    q2,
-                                    std::string,    output,
-                                    Gamma::Algebra, gammaSource,
-                                    Gamma::Algebra, gammaSink);
+                                    std::string, q1,
+                                    std::string, q2,
+                                    std::string, gammas,
+                                    std::string, sink,
+                                    std::string, output);
 };
 
 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:
-        GRID_SERIALIZABLE_CLASS_MEMBERS(Result, std::vector<Complex>, corr);
+        GRID_SERIALIZABLE_CLASS_MEMBERS(Result,
+                                        Gamma::Algebra, gamma_snk,
+                                        Gamma::Algebra, gamma_src,
+                                        std::vector<Complex>, corr);
     };
 public:
     // constructor
@@ -71,6 +97,7 @@ public:
     // dependencies/products
     virtual std::vector<std::string> getInput(void);
     virtual std::vector<std::string> getOutput(void);
+    virtual void parseGammaString(std::vector<GammaPair> &gammaList);
     // execution
     virtual void execute(void);
 };
@@ -90,7 +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;
 }
@@ -103,7 +130,35 @@ std::vector<std::string> TMeson<FImpl1, FImpl2>::getOutput(void)
     return output;
 }
 
+template <typename FImpl1, typename FImpl2>
+void TMeson<FImpl1, FImpl2>::parseGammaString(std::vector<GammaPair> &gammaList)
+{
+    gammaList.clear();
+    // Determine gamma matrices to insert at source/sink.
+    if (par().gammas.compare("all") == 0)
+    {
+        // Do all contractions.
+        for (unsigned int i = 1; i < Gamma::nGamma; i += 2)
+        {
+            for (unsigned int j = 1; j < Gamma::nGamma; j += 2)
+            {
+                gammaList.push_back(std::make_pair((Gamma::Algebra)i, 
+                                                   (Gamma::Algebra)j));
+            }
+        }
+    }
+    else
+    {
+        // Parse individual contractions from input string.
+        gammaList = strToVec<GammaPair>(par().gammas);
+    }
+}
+
+
 // 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)
 {
@@ -111,21 +166,73 @@ void TMeson<FImpl1, FImpl2>::execute(void)
                  << " quarks '" << par().q1 << "' and '" << par().q2 << "'"
                  << std::endl;
     
-    XmlWriter             writer(par().output);
-    PropagatorField1      &q1 = *env().template getObject<PropagatorField1>(par().q1);
-    PropagatorField2      &q2 = *env().template getObject<PropagatorField2>(par().q2);
-    LatticeComplex        c(env().getGrid());
-    Gamma                 gSrc(par().gammaSource), gSnk(par().gammaSink);
-    Gamma                 g5(Gamma::Algebra::Gamma5);
-    std::vector<TComplex> buf;
-    Result                result;
+    CorrWriter             writer(par().output);
+    std::vector<TComplex>  buf;
+    std::vector<Result>    result;
+    Gamma                  g5(Gamma::Algebra::Gamma5);
+    std::vector<GammaPair> gammaList;
+    int                    nt = env().getDim(Tp);
     
-    c = trace(gSnk*q1*adj(gSrc)*g5*adj(q2)*g5);
-    sliceSum(c, buf, Tp);
-    result.corr.resize(buf.size());
-    for (unsigned int t = 0; t < buf.size(); ++t)
+    parseGammaString(gammaList);
+    result.resize(gammaList.size());
+    for (unsigned int i = 0; i < result.size(); ++i)
     {
-        result.corr[t] = TensorRemove(buf[t]);
+        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);
+            
+            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);
 }
@@ -134,4 +241,4 @@ END_MODULE_NAMESPACE
 
 END_HADRONS_NAMESPACE
 
-#endif // Hadrons_Meson_hpp_
+#endif // Hadrons_MContraction_Meson_hpp_

extras/Hadrons/Modules/MContraction/WeakHamiltonian.hpp

@@ -0,0 +1,114 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MContraction/WeakHamiltonian.hpp
+
+Copyright (C) 2017
+
+Author: Andrew Lawson    <andrew.lawson1991@gmail.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_MContraction_WeakHamiltonian_hpp_
+#define Hadrons_MContraction_WeakHamiltonian_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                         WeakHamiltonian                                    *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MContraction)
+
+/*******************************************************************************
+ * Utilities for contractions involving the Weak Hamiltonian.
+ ******************************************************************************/
+//// Sum and store correlator.
+#define MAKE_DIAG(exp, buf, res, n)\
+sliceSum(exp, buf, Tp);\
+res.name = (n);\
+res.corr.resize(buf.size());\
+for (unsigned int t = 0; t < buf.size(); ++t)\
+{\
+    res.corr[t] = TensorRemove(buf[t]);\
+}
+
+//// Contraction of mu index: use 'mu' variable in exp.
+#define SUM_MU(buf,exp)\
+buf = zero;\
+for (unsigned int mu = 0; mu < ndim; ++mu)\
+{\
+    buf += exp;\
+}
+
+enum 
+{
+  i_V = 0,
+  i_A = 1,
+  n_i = 2
+};
+
+class WeakHamiltonianPar: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(WeakHamiltonianPar,
+                                    std::string, q1,
+                                    std::string, q2,
+                                    std::string, q3,
+                                    std::string, q4,
+                                    std::string, output);
+};
+
+#define MAKE_WEAK_MODULE(modname)\
+class T##modname: public Module<WeakHamiltonianPar>\
+{\
+public:\
+    FERM_TYPE_ALIASES(FIMPL,)\
+    class Result: Serializable\
+    {\
+    public:\
+        GRID_SERIALIZABLE_CLASS_MEMBERS(Result,\
+                                        std::string, name,\
+                                        std::vector<Complex>, corr);\
+    };\
+public:\
+    /* constructor */ \
+    T##modname(const std::string name);\
+    /* destructor */ \
+    virtual ~T##modname(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);\
+    std::vector<std::string> VA_label = {"V", "A"};\
+};\
+MODULE_REGISTER_NS(modname, T##modname, MContraction);
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_MContraction_WeakHamiltonian_hpp_

extras/Hadrons/Modules/MContraction/WeakHamiltonianEye.cc

@@ -0,0 +1,137 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MContraction/WeakHamiltonianEye.cc
+
+Copyright (C) 2017
+
+Author: Andrew Lawson    <andrew.lawson1991@gmail.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/MContraction/WeakHamiltonianEye.hpp>
+
+using namespace Grid;
+using namespace Hadrons;
+using namespace MContraction;
+
+/*
+ * Weak Hamiltonian current-current contractions, Eye-type.
+ * 
+ * These contractions are generated by the Q1 and Q2 operators in the physical
+ * basis (see e.g. Fig 3 of arXiv:1507.03094).
+ * 
+ * Schematics:        q4                 |                  
+ *                  /-<-¬                |                             
+ *                 /     \               |             q2           q3
+ *                 \     /               |        /----<------*------<----¬                        
+ *            q2    \   /    q3          |       /          /-*-¬          \
+ *       /-----<-----* *-----<----¬      |      /          /     \          \
+ *    i *            H_W           * f   |   i *           \     /  q4      * f
+ *       \                        /      |      \           \->-/          /   
+ *        \                      /       |       \                        /       
+ *         \---------->---------/        |        \----------->----------/        
+ *                   q1                  |                   q1                  
+ *                                       |
+ *                Saucer (S)             |                  Eye (E)
+ * 
+ * S: trace(q3*g5*q1*adj(q2)*g5*gL[mu][p_1]*q4*gL[mu][p_2])
+ * E: trace(q3*g5*q1*adj(q2)*g5*gL[mu][p_1])*trace(q4*gL[mu][p_2])
+ */
+
+/******************************************************************************
+ *                  TWeakHamiltonianEye implementation                        *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+TWeakHamiltonianEye::TWeakHamiltonianEye(const std::string name)
+: Module<WeakHamiltonianPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+std::vector<std::string> TWeakHamiltonianEye::getInput(void)
+{
+    std::vector<std::string> in = {par().q1, par().q2, par().q3, par().q4};
+    
+    return in;
+}
+
+std::vector<std::string> TWeakHamiltonianEye::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+void TWeakHamiltonianEye::setup(void)
+{
+
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+void TWeakHamiltonianEye::execute(void)
+{
+    LOG(Message) << "Computing Weak Hamiltonian (Eye type) contractions '" 
+                 << getName() << "' using quarks '" << par().q1 << "', '" 
+                 << par().q2 << ", '" << par().q3 << "' and '" << par().q4 
+                 << "'." << std::endl;
+
+    CorrWriter             writer(par().output);
+    PropagatorField &q1 = *env().template getObject<PropagatorField>(par().q1);
+    PropagatorField &q2 = *env().template getObject<PropagatorField>(par().q2);
+    PropagatorField &q3 = *env().template getObject<PropagatorField>(par().q3);
+    PropagatorField &q4 = *env().template getObject<PropagatorField>(par().q4);
+    Gamma g5            = Gamma(Gamma::Algebra::Gamma5);
+    LatticeComplex        expbuf(env().getGrid());
+    std::vector<TComplex> corrbuf;
+    std::vector<Result>   result(n_eye_diag);
+    unsigned int ndim   = env().getNd();
+
+    PropagatorField              tmp1(env().getGrid());
+    LatticeComplex               tmp2(env().getGrid());
+    std::vector<PropagatorField> S_body(ndim, tmp1);
+    std::vector<PropagatorField> S_loop(ndim, tmp1);
+    std::vector<LatticeComplex>  E_body(ndim, tmp2);
+    std::vector<LatticeComplex>  E_loop(ndim, tmp2);
+
+    // Setup for S-type contractions.
+    for (int mu = 0; mu < ndim; ++mu)
+    {
+        S_body[mu] = MAKE_SE_BODY(q1, q2, q3, GammaL(Gamma::gmu[mu]));
+        S_loop[mu] = MAKE_SE_LOOP(q4, GammaL(Gamma::gmu[mu]));
+    }
+
+    // Perform S-type contractions.    
+    SUM_MU(expbuf, trace(S_body[mu]*S_loop[mu]))
+    MAKE_DIAG(expbuf, corrbuf, result[S_diag], "HW_S")
+
+    // Recycle sub-expressions for E-type contractions.
+    for (unsigned int mu = 0; mu < ndim; ++mu)
+    {
+        E_body[mu] = trace(S_body[mu]);
+        E_loop[mu] = trace(S_loop[mu]);
+    }
+
+    // Perform E-type contractions.
+    SUM_MU(expbuf, E_body[mu]*E_loop[mu])
+    MAKE_DIAG(expbuf, corrbuf, result[E_diag], "HW_E")
+
+    write(writer, "HW_Eye", result);
+}

extras/Hadrons/Modules/MContraction/WeakHamiltonianEye.hpp

@@ -0,0 +1,58 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MContraction/WeakHamiltonianEye.hpp
+
+Copyright (C) 2017
+
+Author: Andrew Lawson    <andrew.lawson1991@gmail.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_MContraction_WeakHamiltonianEye_hpp_
+#define Hadrons_MContraction_WeakHamiltonianEye_hpp_
+
+#include <Grid/Hadrons/Modules/MContraction/WeakHamiltonian.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                         WeakHamiltonianEye                                 *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MContraction)
+
+enum
+{
+    S_diag = 0,
+    E_diag = 1,
+    n_eye_diag = 2
+};
+
+// Saucer and Eye subdiagram contractions.
+#define MAKE_SE_BODY(Q_1, Q_2, Q_3, gamma) (Q_3*g5*Q_1*adj(Q_2)*g5*gamma)
+#define MAKE_SE_LOOP(Q_loop, gamma) (Q_loop*gamma)
+
+MAKE_WEAK_MODULE(WeakHamiltonianEye)
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_MContraction_WeakHamiltonianEye_hpp_

extras/Hadrons/Modules/MContraction/WeakHamiltonianNonEye.cc

@@ -0,0 +1,139 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MContraction/WeakHamiltonianNonEye.cc
+
+Copyright (C) 2017
+
+Author: Andrew Lawson    <andrew.lawson1991@gmail.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/MContraction/WeakHamiltonianNonEye.hpp>
+
+using namespace Grid;
+using namespace Hadrons;
+using namespace MContraction;
+
+/*
+ * Weak Hamiltonian current-current contractions, Non-Eye-type.
+ * 
+ * These contractions are generated by the Q1 and Q2 operators in the physical
+ * basis (see e.g. Fig 3 of arXiv:1507.03094).
+ * 
+ * Schematic:     
+ *            q2             q3          |           q2              q3
+ *          /--<--¬       /--<--¬        |        /--<--¬         /--<--¬       
+ *         /       \     /       \       |       /       \       /       \      
+ *        /         \   /         \      |      /         \     /         \     
+ *       /           \ /           \     |     /           \   /           \    
+ *    i *             * H_W         *  f |  i *             * * H_W         * f 
+ *      \             *             |    |     \           /   \           /
+ *       \           / \           /     |      \         /     \         /    
+ *        \         /   \         /      |       \       /       \       /  
+ *         \       /     \       /       |        \-->--/         \-->--/      
+ *          \-->--/       \-->--/        |          q1               q4 
+ *            q1             q4          |
+ *                Connected (C)          |                 Wing (W)
+ *
+ * C: trace(q1*adj(q2)*g5*gL[mu]*q3*adj(q4)*g5*gL[mu])
+ * W: trace(q1*adj(q2)*g5*gL[mu])*trace(q3*adj(q4)*g5*gL[mu])
+ * 
+ */
+
+/******************************************************************************
+ *                  TWeakHamiltonianNonEye implementation                     *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+TWeakHamiltonianNonEye::TWeakHamiltonianNonEye(const std::string name)
+: Module<WeakHamiltonianPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+std::vector<std::string> TWeakHamiltonianNonEye::getInput(void)
+{
+    std::vector<std::string> in = {par().q1, par().q2, par().q3, par().q4};
+    
+    return in;
+}
+
+std::vector<std::string> TWeakHamiltonianNonEye::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+void TWeakHamiltonianNonEye::setup(void)
+{
+
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+void TWeakHamiltonianNonEye::execute(void)
+{
+    LOG(Message) << "Computing Weak Hamiltonian (Non-Eye type) contractions '" 
+                 << getName() << "' using quarks '" << par().q1 << "', '" 
+                 << par().q2 << ", '" << par().q3 << "' and '" << par().q4 
+                 << "'." << std::endl;
+    
+    CorrWriter             writer(par().output);
+    PropagatorField &q1 = *env().template getObject<PropagatorField>(par().q1);
+    PropagatorField &q2 = *env().template getObject<PropagatorField>(par().q2);
+    PropagatorField &q3 = *env().template getObject<PropagatorField>(par().q3);
+    PropagatorField &q4 = *env().template getObject<PropagatorField>(par().q4);
+    Gamma g5            = Gamma(Gamma::Algebra::Gamma5);
+    LatticeComplex        expbuf(env().getGrid());
+    std::vector<TComplex> corrbuf;
+    std::vector<Result>   result(n_noneye_diag); 
+    unsigned int ndim   = env().getNd();
+
+    PropagatorField              tmp1(env().getGrid());
+    LatticeComplex               tmp2(env().getGrid());
+    std::vector<PropagatorField> C_i_side_loop(ndim, tmp1);
+    std::vector<PropagatorField> C_f_side_loop(ndim, tmp1);
+    std::vector<LatticeComplex>  W_i_side_loop(ndim, tmp2);
+    std::vector<LatticeComplex>  W_f_side_loop(ndim, tmp2);
+
+    // Setup for C-type contractions.
+    for (int mu = 0; mu < ndim; ++mu)
+    {
+        C_i_side_loop[mu] = MAKE_CW_SUBDIAG(q1, q2, GammaL(Gamma::gmu[mu]));
+        C_f_side_loop[mu] = MAKE_CW_SUBDIAG(q3, q4, GammaL(Gamma::gmu[mu]));
+    }
+
+    // Perform C-type contractions.    
+    SUM_MU(expbuf, trace(C_i_side_loop[mu]*C_f_side_loop[mu]))
+    MAKE_DIAG(expbuf, corrbuf, result[C_diag], "HW_C")
+
+    // Recycle sub-expressions for W-type contractions.
+    for (unsigned int mu = 0; mu < ndim; ++mu)
+    {
+        W_i_side_loop[mu] = trace(C_i_side_loop[mu]);
+        W_f_side_loop[mu] = trace(C_f_side_loop[mu]);
+    }
+
+    // Perform W-type contractions.
+    SUM_MU(expbuf, W_i_side_loop[mu]*W_f_side_loop[mu])
+    MAKE_DIAG(expbuf, corrbuf, result[W_diag], "HW_W")
+
+    write(writer, "HW_NonEye", result);
+}

extras/Hadrons/Modules/MContraction/WeakHamiltonianNonEye.hpp

@@ -0,0 +1,57 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MContraction/WeakHamiltonianNonEye.hpp
+
+Copyright (C) 2017
+
+Author: Andrew Lawson    <andrew.lawson1991@gmail.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_MContraction_WeakHamiltonianNonEye_hpp_
+#define Hadrons_MContraction_WeakHamiltonianNonEye_hpp_
+
+#include <Grid/Hadrons/Modules/MContraction/WeakHamiltonian.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                         WeakHamiltonianNonEye                              *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MContraction)
+
+enum
+{
+    W_diag = 0,
+    C_diag = 1,
+    n_noneye_diag = 2
+};
+
+// Wing and Connected subdiagram contractions
+#define MAKE_CW_SUBDIAG(Q_1, Q_2, gamma) (Q_1*adj(Q_2)*g5*gamma)
+
+MAKE_WEAK_MODULE(WeakHamiltonianNonEye)
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_MContraction_WeakHamiltonianNonEye_hpp_

extras/Hadrons/Modules/MContraction/WeakNeutral4ptDisc.cc

@@ -0,0 +1,135 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MContraction/WeakNeutral4ptDisc.cc
+
+Copyright (C) 2017
+
+Author: Andrew Lawson    <andrew.lawson1991@gmail.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/MContraction/WeakNeutral4ptDisc.hpp>
+
+using namespace Grid;
+using namespace Hadrons;
+using namespace MContraction;
+
+/*
+ * Weak Hamiltonian + current contractions, disconnected topology for neutral 
+ * mesons.
+ * 
+ * These contractions are generated by operators Q_1,...,10 of the dS=1 Weak
+ * Hamiltonian in the physical basis and an additional current J (see e.g. 
+ * Fig 11 of arXiv:1507.03094).
+ * 
+ * Schematic:
+ *                        
+ *           q2          q4             q3
+ *       /--<--¬     /---<--¬       /---<--¬
+ *     /         \ /         \     /        \
+ *  i *           * H_W      |  J *          * f
+ *     \         / \         /     \        /
+ *      \--->---/   \-------/       \------/
+ *          q1 
+ * 
+ * options
+ * - q1: input propagator 1 (string)
+ * - q2: input propagator 2 (string)
+ * - q3: input propagator 3 (string), assumed to be sequential propagator 
+ * - q4: input propagator 4 (string), assumed to be a loop
+ * 
+ * type 1: trace(q1*adj(q2)*g5*gL[mu])*trace(loop*gL[mu])*trace(q3*g5)
+ * type 2: trace(q1*adj(q2)*g5*gL[mu]*loop*gL[mu])*trace(q3*g5)
+ */
+
+/*******************************************************************************
+ *                  TWeakNeutral4ptDisc implementation                         *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+TWeakNeutral4ptDisc::TWeakNeutral4ptDisc(const std::string name)
+: Module<WeakHamiltonianPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+std::vector<std::string> TWeakNeutral4ptDisc::getInput(void)
+{
+    std::vector<std::string> in = {par().q1, par().q2, par().q3, par().q4};
+    
+    return in;
+}
+
+std::vector<std::string> TWeakNeutral4ptDisc::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+void TWeakNeutral4ptDisc::setup(void)
+{
+
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+void TWeakNeutral4ptDisc::execute(void)
+{
+    LOG(Message) << "Computing Weak Hamiltonian neutral disconnected contractions '" 
+                 << getName() << "' using quarks '" << par().q1 << "', '" 
+                 << par().q2 << ", '" << par().q3 << "' and '" << par().q4 
+                 << "'." << std::endl;
+
+    CorrWriter             writer(par().output);
+    PropagatorField &q1 = *env().template getObject<PropagatorField>(par().q1);
+    PropagatorField &q2 = *env().template getObject<PropagatorField>(par().q2);
+    PropagatorField &q3 = *env().template getObject<PropagatorField>(par().q3);
+    PropagatorField &q4 = *env().template getObject<PropagatorField>(par().q4);
+    Gamma g5            = Gamma(Gamma::Algebra::Gamma5);
+    LatticeComplex        expbuf(env().getGrid());
+    std::vector<TComplex> corrbuf;
+    std::vector<Result>   result(n_neut_disc_diag);
+    unsigned int ndim   = env().getNd();
+
+    PropagatorField              tmp(env().getGrid());
+    std::vector<PropagatorField> meson(ndim, tmp);
+    std::vector<PropagatorField> loop(ndim, tmp);
+    LatticeComplex               curr(env().getGrid());
+
+    // Setup for type 1 contractions.
+    for (int mu = 0; mu < ndim; ++mu)
+    {
+        meson[mu] = MAKE_DISC_MESON(q1, q2, GammaL(Gamma::gmu[mu]));
+        loop[mu] = MAKE_DISC_LOOP(q4, GammaL(Gamma::gmu[mu]));
+    }
+    curr = MAKE_DISC_CURR(q3, GammaL(Gamma::Algebra::Gamma5));
+
+    // Perform type 1 contractions.    
+    SUM_MU(expbuf, trace(meson[mu]*loop[mu]))
+    expbuf *= curr;
+    MAKE_DIAG(expbuf, corrbuf, result[neut_disc_1_diag], "HW_disc0_1")
+
+    // Perform type 2 contractions.
+    SUM_MU(expbuf, trace(meson[mu])*trace(loop[mu]))
+    expbuf *= curr;
+    MAKE_DIAG(expbuf, corrbuf, result[neut_disc_2_diag], "HW_disc0_2")
+
+    write(writer, "HW_disc0", result);
+}

extras/Hadrons/Modules/MContraction/WeakNeutral4ptDisc.hpp

@@ -0,0 +1,59 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MContraction/WeakNeutral4ptDisc.hpp
+
+Copyright (C) 2017
+
+Author: Andrew Lawson    <andrew.lawson1991@gmail.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_MContraction_WeakNeutral4ptDisc_hpp_
+#define Hadrons_MContraction_WeakNeutral4ptDisc_hpp_
+
+#include <Grid/Hadrons/Modules/MContraction/WeakHamiltonian.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                         WeakNeutral4ptDisc                                 *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MContraction)
+
+enum
+{
+    neut_disc_1_diag = 0,
+    neut_disc_2_diag = 1,
+    n_neut_disc_diag = 2
+};
+
+// Neutral 4pt disconnected subdiagram contractions.
+#define MAKE_DISC_MESON(Q_1, Q_2, gamma) (Q_1*adj(Q_2)*g5*gamma)
+#define MAKE_DISC_LOOP(Q_LOOP, gamma) (Q_LOOP*gamma)
+#define MAKE_DISC_CURR(Q_c, gamma) (trace(Q_c*gamma))
+
+MAKE_WEAK_MODULE(WeakNeutral4ptDisc)
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#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());
@@ -111,13 +84,13 @@ 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;
+    << std::endl;
     
     FermionField    source(env().getGrid(Ls_)), sol(env().getGrid(Ls_)),
-                    tmp(env().getGrid());
+    tmp(env().getGrid());
     std::string     propName = (Ls_ == 1) ? getName() : (getName() + "_5d");
     PropagatorField &prop    = *env().template createLattice<PropagatorField>(propName);
     PropagatorField &fullSrc = *env().template getObject<PropagatorField>(par().source);
@@ -128,7 +101,7 @@ void TQuark<FImpl>::execute(void)
     }
     
     LOG(Message) << "Inverting using solver '" << par().solver
-                 << "' on source '" << par().source << "'" << std::endl;
+    << "' on source '" << par().source << "'" << std::endl;
     for (unsigned int s = 0; s < Ns; ++s)
     for (unsigned int c = 0; c < Nc; ++c)
     {
@@ -170,16 +143,18 @@ void TQuark<FImpl>::execute(void)
         if (Ls_ > 1)
         {
             PropagatorField &p4d =
-                *env().template getObject<PropagatorField>(getName());
+            *env().template getObject<PropagatorField>(getName());
             
             axpby_ssp_pminus(sol, 0., sol, 1., sol, 0, 0);
-            axpby_ssp_pplus(sol, 0., sol, 1., sol, 0, Ls_-1);
+            axpby_ssp_pplus(sol, 1., sol, 1., sol, 0, Ls_-1);
             ExtractSlice(tmp, sol, 0, 0);
             FermToProp(p4d, tmp, s, c);
         }
     }
 }
 
+END_MODULE_NAMESPACE
+
 END_HADRONS_NAMESPACE
 
-#endif // Hadrons_Quark_hpp_
+#endif // Hadrons_MFermion_GaugeProp_hpp_

extras/Hadrons/Modules/MGauge/Load.cc

@@ -65,7 +65,7 @@ void TLoad::setup(void)
 // execution ///////////////////////////////////////////////////////////////////
 void TLoad::execute(void)
 {
-    NerscField  header;
+    FieldMetaData  header;
     std::string fileName = par().file + "."
                            + std::to_string(env().getTrajectory());
     
@@ -74,5 +74,5 @@ void TLoad::execute(void)
     LatticeGaugeField &U = *env().createLattice<LatticeGaugeField>(getName());
     NerscIO::readConfiguration(U, header, fileName);
     LOG(Message) << "NERSC header:" << std::endl;
-    dump_nersc_header(header, LOG(Message));
+    dump_meta_data(header, LOG(Message));
 }

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

@@ -0,0 +1,132 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MLoop/NoiseLoop.hpp
+
+Copyright (C) 2016
+
+Author: Andrew Lawson <andrew.lawson1991@gmail.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_MLoop_NoiseLoop_hpp_
+#define Hadrons_MLoop_NoiseLoop_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/*
+ 
+ Noise loop propagator
+ -----------------------------
+ * loop_x = q_x * adj(eta_x)
+ 
+ * options:
+ - q = Result of inversion on noise source.
+ - eta = noise source.
+
+ */
+
+
+/******************************************************************************
+ *                         NoiseLoop                                          *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MLoop)
+
+class NoiseLoopPar: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(NoiseLoopPar,
+                                    std::string, q,
+                                    std::string, eta);
+};
+
+template <typename FImpl>
+class TNoiseLoop: public Module<NoiseLoopPar>
+{
+public:
+    FERM_TYPE_ALIASES(FImpl,);
+public:
+    // constructor
+    TNoiseLoop(const std::string name);
+    // destructor
+    virtual ~TNoiseLoop(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(NoiseLoop, TNoiseLoop<FIMPL>, MLoop);
+
+/******************************************************************************
+ *                 TNoiseLoop implementation                                  *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+template <typename FImpl>
+TNoiseLoop<FImpl>::TNoiseLoop(const std::string name)
+: Module<NoiseLoopPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+template <typename FImpl>
+std::vector<std::string> TNoiseLoop<FImpl>::getInput(void)
+{
+    std::vector<std::string> in = {par().q, par().eta};
+    
+    return in;
+}
+
+template <typename FImpl>
+std::vector<std::string> TNoiseLoop<FImpl>::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TNoiseLoop<FImpl>::setup(void)
+{
+    env().template registerLattice<PropagatorField>(getName());
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TNoiseLoop<FImpl>::execute(void)
+{
+    PropagatorField &loop = *env().template createLattice<PropagatorField>(getName());
+    PropagatorField &q    = *env().template getObject<PropagatorField>(par().q);
+    PropagatorField &eta  = *env().template getObject<PropagatorField>(par().eta);
+    loop = q*adj(eta);
+}
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#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);
@@ -78,7 +78,8 @@ public:
     virtual void execute(void);
 };
 
-MODULE_REGISTER_NS(Point, TPoint<FIMPL>, MSource);
+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

@@ -6,6 +6,7 @@ Source file: extras/Hadrons/Modules/MSource/SeqGamma.hpp
 
 Copyright (C) 2015
 Copyright (C) 2016
+Copyright (C) 2017
 
 Author: Antonin Portelli <antonin.portelli@me.com>
 
@@ -27,8 +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>
@@ -71,7 +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);
@@ -149,9 +150,9 @@ void TSeqGamma<FImpl>::execute(void)
     for(unsigned int mu = 0; mu < env().getNd(); mu++)
     {
         LatticeCoordinate(coor, mu);
-        ph = ph + p[mu]*coor;
+        ph = ph + p[mu]*coor*((1./(env().getGrid()->_fdimensions[mu])));
     }
-    ph = exp(i*ph);
+    ph = exp((Real)(2*M_PI)*i*ph);
     LatticeCoordinate(t, Tp);
     src = where((t >= par().tA) and (t <= par().tB), ph*(g*q), 0.*q);
 }
@@ -160,4 +161,4 @@ END_MODULE_NAMESPACE
 
 END_HADRONS_NAMESPACE
 
-#endif // Hadrons_SeqGamma_hpp_
+#endif // Hadrons_MSource_SeqGamma_hpp_

extras/Hadrons/Modules/MSource/Wall.hpp

@@ -0,0 +1,147 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MSource/Wall.hpp
+
+Copyright (C) 2017
+
+Author: Andrew Lawson <andrew.lawson1991@gmail.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_MSource_WallSource_hpp_
+#define Hadrons_MSource_WallSource_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/*
+ 
+ Wall source
+ -----------------------------
+ * src_x = delta(x_3 - tW) * exp(i x.mom)
+ 
+ * options:
+ - tW: source timeslice (integer)
+ - mom: momentum insertion, space-separated float sequence (e.g ".1 .2 1. 0.")
+ 
+ */
+
+/******************************************************************************
+ *                         Wall                                               *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MSource)
+
+class WallPar: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(WallPar,
+                                    unsigned int, tW,
+                                    std::string, mom);
+};
+
+template <typename FImpl>
+class TWall: public Module<WallPar>
+{
+public:
+    FERM_TYPE_ALIASES(FImpl,);
+public:
+    // constructor
+    TWall(const std::string name);
+    // destructor
+    virtual ~TWall(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(Wall, TWall<FIMPL>, MSource);
+
+/******************************************************************************
+ *                 TWall implementation                                       *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+template <typename FImpl>
+TWall<FImpl>::TWall(const std::string name)
+: Module<WallPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+template <typename FImpl>
+std::vector<std::string> TWall<FImpl>::getInput(void)
+{
+    std::vector<std::string> in;
+    
+    return in;
+}
+
+template <typename FImpl>
+std::vector<std::string> TWall<FImpl>::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TWall<FImpl>::setup(void)
+{
+    env().template registerLattice<PropagatorField>(getName());
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TWall<FImpl>::execute(void)
+{    
+    LOG(Message) << "Generating wall source at t = " << par().tW 
+                 << " with momentum " << par().mom << std::endl;
+    
+    PropagatorField &src = *env().template createLattice<PropagatorField>(getName());
+    Lattice<iScalar<vInteger>> t(env().getGrid());
+    LatticeComplex             ph(env().getGrid()), coor(env().getGrid());
+    std::vector<Real>          p;
+    Complex                    i(0.0,1.0);
+    
+    p  = strToVec<Real>(par().mom);
+    ph = zero;
+    for(unsigned int mu = 0; mu < Nd; mu++)
+    {
+        LatticeCoordinate(coor, mu);
+        ph = ph + p[mu]*coor*((1./(env().getGrid()->_fdimensions[mu])));
+    }
+    ph = exp((Real)(2*M_PI)*i*ph);
+    LatticeCoordinate(t, Tp);
+    src = 1.;
+    src = where((t == par().tW), src*ph, 0.*src);
+}
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#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);
@@ -82,7 +82,8 @@ public:
     virtual void execute(void);
 };
 
-MODULE_REGISTER_NS(Z2, TZ2<FIMPL>, MSource);
+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

@@ -1,19 +1,38 @@
 modules_cc =\
+  Modules/MContraction/WeakHamiltonianEye.cc \
+  Modules/MContraction/WeakHamiltonianNonEye.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 \
   Modules/MAction/Wilson.hpp \
   Modules/MContraction/Baryon.hpp \
+  Modules/MContraction/DiscLoop.hpp \
+  Modules/MContraction/Gamma3pt.hpp \
   Modules/MContraction/Meson.hpp \
+  Modules/MContraction/WeakHamiltonian.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/Z2.hpp \
-  Modules/Quark.hpp
+  Modules/MSource/Wall.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;
+}

gcc-bug-report/README

@@ -20,4 +20,17 @@ The simple testcase in this directory is the submitted bug report that encapsula
 problem. The test case works with icpc and with clang++, but fails consistently on g++
 current variants.
 
-Peter
+Peter
+
+
+************
+
+Second GCC bug reported, see Issue 100.
+
+https://wandbox.org/permlink/tzssJza6R9XnqANw
+https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80652
+
+Getting Travis fails under gcc-5 for Test_simd, now that I added more comprehensive testing to the
+CI test suite. The limitations of Travis runtime limits & weak cores are being shown.
+
+Travis uses 5.4.1 for g++-5.

grid-config.in

@@ -0,0 +1,86 @@
+#! /bin/sh
+
+prefix=@prefix@
+exec_prefix=@exec_prefix@
+includedir=@includedir@
+
+usage()
+{
+  cat <<EOF
+Usage: grid-config [OPTION]
+
+Known values for OPTION are:
+
+  --prefix     show Grid installation prefix
+  --cxxflags   print pre-processor and compiler flags
+  --ldflags    print library linking flags
+  --libs       print library linking information
+  --summary    print full build summary
+  --help       display this help and exit
+  --version    output version information
+  --git        print git revision
+
+EOF
+  
+  exit $1
+}
+
+if test $# -eq 0; then
+  usage 1
+fi
+
+cflags=false
+libs=false
+
+while test $# -gt 0; do
+  case "$1" in
+    -*=*) optarg=`echo "$1" | sed 's/[-_a-zA-Z0-9]*=//'` ;;
+    *) optarg= ;;
+  esac
+  
+  case "$1" in
+    --prefix)
+      echo $prefix
+    ;;
+    
+    --version)
+      echo @VERSION@
+      exit 0
+    ;;
+    
+    --git)
+      echo "@GRID_BRANCH@ @GRID_SHA@"
+      exit 0
+    ;;
+    
+    --help)
+      usage 0
+    ;;
+    
+    --cxxflags)
+      echo @GRID_CXXFLAGS@
+    ;;
+    
+    --ldflags)
+      echo @GRID_LDFLAGS@
+    ;;
+    
+    --libs)
+      echo @GRID_LIBS@
+    ;;
+    
+    --summary)
+      echo ""
+      echo "@GRID_SUMMARY@"
+      echo ""
+    ;;
+    
+    *)
+      usage
+      exit 1
+    ;;
+  esac
+  shift
+done
+
+exit 0

lib/DisableWarnings.h

@@ -0,0 +1,37 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/DisableWarnings.h
+
+Copyright (C) 2016
+
+Author: Guido Cossu <guido.cossu@ed.ac.uk>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+/*  END LEGAL */
+
+#ifndef DISABLE_WARNINGS_H
+#define DISABLE_WARNINGS_H
+
+ //disables and intel compiler specific warning (in json.hpp)
+#pragma warning disable 488  
+
+
+#endif

lib/Grid.h

@@ -41,7 +41,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #include <Grid/GridCore.h>
 #include <Grid/GridQCDcore.h>
 #include <Grid/qcd/action/Action.h>
+#include <Grid/qcd/utils/GaugeFix.h>
 #include <Grid/qcd/smearing/Smearing.h>
+#include <Grid/parallelIO/MetaData.h>
 #include <Grid/qcd/hmc/HMC_aggregate.h>
 
 #endif

lib/GridCore.h

@@ -38,28 +38,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_BASE_H
 #define GRID_BASE_H
 
-///////////////////
-// Std C++ dependencies
-///////////////////
-#include <cassert>
-#include <complex>
-#include <vector>
-#include <iostream>
-#include <iomanip>
-#include <random>
-#include <functional>
-#include <stdio.h>
-#include <stdlib.h>
-#include <stdio.h>
-#include <signal.h>
-#include <ctime>
-#include <sys/time.h>
-#include <chrono>
-
-///////////////////
-// Grid headers
-///////////////////
-#include "Config.h"
+#include <Grid/GridStd.h>
 
 #include <Grid/perfmon/Timer.h>
 #include <Grid/perfmon/PerfCount.h>

lib/GridStd.h

@@ -0,0 +1,29 @@
+#ifndef GRID_STD_H
+#define GRID_STD_H
+
+///////////////////
+// Std C++ dependencies
+///////////////////
+#include <cassert>
+#include <complex>
+#include <vector>
+#include <string>
+#include <iostream>
+#include <iomanip>
+#include <random>
+#include <functional>
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdio.h>
+#include <signal.h>
+#include <ctime>
+#include <sys/time.h>
+#include <chrono>
+#include <zlib.h>
+
+///////////////////
+// Grid config
+///////////////////
+#include "Config.h"
+
+#endif /* GRID_STD_H */

lib/Grid_Eigen_Dense.h

@@ -0,0 +1,9 @@
+#pragma once
+#if defined __GNUC__
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wdeprecated-declarations"
+#endif
+#include <Grid/Eigen/Dense>
+#if defined __GNUC__
+#pragma GCC diagnostic pop
+#endif

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

@@ -1,6 +1,6 @@
     /*************************************************************************************
 
-    Grid physics library, www.github.com/paboyle/Grid 
+    Grid physics library, www.github.com/paboyle/Grid
 
     Source file: ./lib/Algorithms.h
 
@@ -37,6 +37,11 @@ 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/densematrix/DenseMatrix.h>
+#include <Grid/algorithms/densematrix/Francis.h>
+#include <Grid/algorithms/densematrix/Householder.h>
 
 #include <Grid/algorithms/iterative/ConjugateGradient.h>
 #include <Grid/algorithms/iterative/ConjugateResidual.h>
@@ -44,30 +49,18 @@ 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/iterative/ImplicitlyRestartedLanczosCJ.h>
+#include <Grid/algorithms/iterative/SimpleLanczos.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/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

@@ -8,6 +8,7 @@
 
 Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
 Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Chulwoo Jung <chulwoo@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
@@ -162,15 +163,10 @@ 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 +188,10 @@ 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 +208,6 @@ 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> {
@@ -235,7 +230,7 @@ namespace Grid {
 	Field tmp(in._grid);
 
 	_Mat.MeooeDag(in,tmp);
-	_Mat.MooeeInvDag(tmp,out);
+        _Mat.MooeeInvDag(tmp,out);
 	_Mat.MeooeDag(out,tmp);
 
 	_Mat.MooeeDag(in,out);
@@ -270,7 +265,6 @@ namespace Grid {
 	return axpy_norm(out,-1.0,tmp,in);
       }
     };
-
     template<class Matrix,class Field>
       class SchurDiagTwoOperator :  public SchurOperatorBase<Field> {
     protected:
@@ -299,6 +293,168 @@ 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>;
+    template<class Matrix,class Field>
+//      class SchurStagOperator :  public LinearOperatorBase<Field> {
+      class SchurStagOperator :  public SchurOperatorBase<Field> {
+    protected:
+      Matrix &_Mat;
+    public:
+      SchurStagOperator (Matrix &Mat): _Mat(Mat){};
+      virtual  RealD Mpc      (const Field &in, Field &out) {
+	Field tmp(in._grid);
+	Field tmp2(in._grid);
+
+	_Mat.Mooee(in,out);
+	_Mat.Mooee(out,tmp);
+
+	_Mat.Meooe(in,out);
+	_Mat.Meooe(out,tmp2);
+
+	return axpy_norm(out,-1.0,tmp2,tmp);
+      }
+      virtual  RealD MpcDag   (const Field &in, Field &out){
+
+	return Mpc(in,out);
+      }
+#if 0
+      virtual void MpcDagMpc(const Field &in, Field &out,RealD &ni,RealD &no) {
+	Field tmp(in._grid);
+	ni=Mpc(in,tmp);
+	no=MpcDag(tmp,out);
+      }
+#endif
+      void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
+	n2 = Mpc(in,out);
+	ComplexD dot = innerProduct(in,out);
+	n1 = real(dot);
+      }
+      void HermOp(const Field &in, Field &out){
+	RealD n1,n2;
+	HermOpAndNorm(in,out,n1,n2);
+      }
+      void Op     (const Field &in, Field &out){
+	Mpc(in,out);
+      }
+      void AdjOp     (const Field &in, Field &out){ 
+	MpcDag(in,out);
+      }
+      // Support for coarsening to a multigrid
+      void OpDiag (const Field &in, Field &out) {
+	assert(0); // must coarsen the unpreconditioned system
+      }
+      void OpDir  (const Field &in, Field &out,int dir,int disp) {
+	assert(0);
+      }
+
+    };
+
+#if 0
+  // This is specific to (Z)mobius fermions
+  template<class Matrix, class Field>
+    class KappaSimilarityTransform {
+  public:
+//    INHERIT_IMPL_TYPES(Matrix);
+    typedef typename Matrix::Coeff_t                     Coeff_t;
+    std::vector<Coeff_t> kappa, kappaDag, kappaInv, kappaInvDag;
+
+    KappaSimilarityTransform (Matrix &zmob) {
+      for (int i=0;i<(int)zmob.bs.size();i++) {
+	Coeff_t k = 1.0 / ( 2.0 * (zmob.bs[i] *(4 - zmob.M5) + 1.0) );
+	kappa.push_back( k );
+	kappaDag.push_back( conj(k) );
+	kappaInv.push_back( 1.0 / k );
+	kappaInvDag.push_back( 1.0 / conj(k) );
+      }
+    }
+
+  template<typename vobj>
+    void sscale(const Lattice<vobj>& in, Lattice<vobj>& out, Coeff_t* s) {
+    GridBase *grid=out._grid;
+    out.checkerboard = in.checkerboard;
+    assert(grid->_simd_layout[0] == 1); // should be fine for ZMobius for now
+    int Ls = grid->_rdimensions[0];
+    parallel_for(int ss=0;ss<grid->oSites();ss++){
+      vobj tmp = s[ss % Ls]*in._odata[ss];
+      vstream(out._odata[ss],tmp);
+    }
+  }
+
+  RealD sscale_norm(const Field& in, Field& out, Coeff_t* s) {
+    sscale(in,out,s);
+    return norm2(out);
+  }
+
+  virtual RealD M       (const Field& in, Field& out) { return sscale_norm(in,out,&kappa[0]);   }
+  virtual RealD MDag    (const Field& in, Field& out) { return sscale_norm(in,out,&kappaDag[0]);}
+  virtual RealD MInv    (const Field& in, Field& out) { return sscale_norm(in,out,&kappaInv[0]);}
+  virtual RealD MInvDag (const Field& in, Field& out) { return sscale_norm(in,out,&kappaInvDag[0]);}
+
+  };
+
+  template<class Matrix,class Field>
+    class SchurDiagTwoKappaOperator :  public SchurOperatorBase<Field> {
+  public:
+    KappaSimilarityTransform<Matrix, Field> _S;
+    SchurDiagTwoOperator<Matrix, Field> _Mat;
+
+    SchurDiagTwoKappaOperator (Matrix &Mat): _S(Mat), _Mat(Mat) {};
+
+    virtual  RealD Mpc      (const Field &in, Field &out) {
+      Field tmp(in._grid);
+
+      _S.MInv(in,out);
+      _Mat.Mpc(out,tmp);
+      return _S.M(tmp,out);
+
+    }
+    virtual  RealD MpcDag   (const Field &in, Field &out){
+      Field tmp(in._grid);
+
+      _S.MDag(in,out);
+      _Mat.MpcDag(out,tmp);
+      return _S.MInvDag(tmp,out);
+    }
+  };
+#endif
 
 
     /////////////////////////////////////////////////////////////

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
@@ -193,12 +194,54 @@ 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) {
 
       GridBase *grid=in._grid;
-//std::cout << "Chevyshef(): in._grid="<<in._grid<<std::endl;
-//<<" Linop.Grid()="<<Linop.Grid()<<"Linop.RedBlackGrid()="<<Linop.RedBlackGrid()<<std::endl;
+
+      // std::cout << "Chevyshef(): in._grid="<<in._grid<<std::endl;
+      //std::cout <<" Linop.Grid()="<<Linop.Grid()<<"Linop.RedBlackGrid()="<<Linop.RedBlackGrid()<<std::endl;
 
       int vol=grid->gSites();
 

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/approx/Remez.h

@@ -16,7 +16,7 @@
 #define INCLUDED_ALG_REMEZ_H
 
 #include <stddef.h>
-#include <Config.h>
+#include <Grid/GridStd.h>
 
 #ifdef HAVE_LIBGMP
 #include "bigfloat.h"

lib/algorithms/iterative/BlockConjugateGradient.h

@@ -33,6 +33,8 @@ directory
 
 namespace Grid {
 
+enum BlockCGtype { BlockCG, BlockCGrQ, CGmultiRHS };
+
 //////////////////////////////////////////////////////////////////////////
 // Block conjugate gradient. Dimension zero should be the block direction
 //////////////////////////////////////////////////////////////////////////
@@ -40,25 +42,286 @@ template <class Field>
 class BlockConjugateGradient : public OperatorFunction<Field> {
  public:
 
+
   typedef typename Field::scalar_type scomplex;
 
-  const int blockDim = 0;
-
+  int blockDim ;
   int Nblock;
+
+  BlockCGtype CGtype;
   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
   
-  BlockConjugateGradient(RealD tol, Integer maxit, bool err_on_no_conv = true)
-    : Tolerance(tol),
-    MaxIterations(maxit),
-    ErrorOnNoConverge(err_on_no_conv){};
+  BlockConjugateGradient(BlockCGtype cgtype,int _Orthog,RealD tol, Integer maxit, bool err_on_no_conv = true)
+    : Tolerance(tol), CGtype(cgtype),   blockDim(_Orthog),  MaxIterations(maxit), ErrorOnNoConverge(err_on_no_conv)
+  {};
 
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// Thin QR factorisation (google it)
+////////////////////////////////////////////////////////////////////////////////////////////////////
+void ThinQRfact (Eigen::MatrixXcd &m_rr,
+		 Eigen::MatrixXcd &C,
+		 Eigen::MatrixXcd &Cinv,
+		 Field & Q,
+		 const Field & R)
+{
+  int Orthog = blockDim; // First dimension is block dim; this is an assumption
+  ////////////////////////////////////////////////////////////////////////////////////////////////////
+  //Dimensions
+  // R_{ferm x Nblock} =  Q_{ferm x Nblock} x  C_{Nblock x Nblock} -> ferm x Nblock
+  //
+  // Rdag R = m_rr = Herm = L L^dag        <-- Cholesky decomposition (LLT routine in Eigen)
+  //
+  //   Q  C = R => Q = R C^{-1}
+  //
+  // Want  Ident = Q^dag Q = C^{-dag} R^dag R C^{-1} = C^{-dag} L L^dag C^{-1} = 1_{Nblock x Nblock} 
+  //
+  // Set C = L^{dag}, and then Q^dag Q = ident 
+  //
+  // Checks:
+  // Cdag C = Rdag R ; passes.
+  // QdagQ  = 1      ; passes
+  ////////////////////////////////////////////////////////////////////////////////////////////////////
+  sliceInnerProductMatrix(m_rr,R,R,Orthog);
+
+  // 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}
+  //
+  // Q_j  = R_i Cinv(i,j) 
+  //
+  // NB maddMatrix conventions are Right multiplication X[j] a[j,i] already
+  ////////////////////////////////////////////////////////////////////////////////////////////////////
+  sliceMulMatrix(Q,Cinv,R,Orthog);
+}
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// Call one of several implementations
+////////////////////////////////////////////////////////////////////////////////////////////////////
 void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi) 
 {
-  int Orthog = 0; // First dimension is block dim
+  if ( CGtype == BlockCGrQ ) {
+    BlockCGrQsolve(Linop,Src,Psi);
+  } else if (CGtype == BlockCG ) {
+    BlockCGsolve(Linop,Src,Psi);
+  } else if (CGtype == CGmultiRHS ) {
+    CGmultiRHSsolve(Linop,Src,Psi);
+  } else {
+    assert(0);
+  }
+}
+
+////////////////////////////////////////////////////////////////////////////
+// BlockCGrQ implementation:
+//--------------------------
+// X is guess/Solution
+// B is RHS
+// Solve A X_i = B_i    ;        i refers to Nblock index
+////////////////////////////////////////////////////////////////////////////
+void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X) 
+{
+  int Orthog = blockDim; // First dimension is block dim; this is an assumption
+  Nblock = B._grid->_fdimensions[Orthog];
+
+  std::cout<<GridLogMessage<<" Block Conjugate Gradient : Orthog "<<Orthog<<" Nblock "<<Nblock<<std::endl;
+
+  X.checkerboard = B.checkerboard;
+  conformable(X, B);
+
+  Field tmp(B);
+  Field Q(B);
+  Field D(B);
+  Field Z(B);
+  Field AD(B);
+
+  Eigen::MatrixXcd m_DZ     = Eigen::MatrixXcd::Identity(Nblock,Nblock);
+  Eigen::MatrixXcd m_M      = Eigen::MatrixXcd::Identity(Nblock,Nblock);
+  Eigen::MatrixXcd m_rr     = Eigen::MatrixXcd::Zero(Nblock,Nblock);
+
+  Eigen::MatrixXcd m_C      = Eigen::MatrixXcd::Zero(Nblock,Nblock);
+  Eigen::MatrixXcd m_Cinv   = Eigen::MatrixXcd::Zero(Nblock,Nblock);
+  Eigen::MatrixXcd m_S      = Eigen::MatrixXcd::Zero(Nblock,Nblock);
+  Eigen::MatrixXcd m_Sinv   = Eigen::MatrixXcd::Zero(Nblock,Nblock);
+
+  Eigen::MatrixXcd m_tmp    = Eigen::MatrixXcd::Identity(Nblock,Nblock);
+  Eigen::MatrixXcd m_tmp1   = Eigen::MatrixXcd::Identity(Nblock,Nblock);
+
+  // Initial residual computation & set up
+  std::vector<RealD> residuals(Nblock);
+  std::vector<RealD> ssq(Nblock);
+
+  sliceNorm(ssq,B,Orthog);
+  RealD sssum=0;
+  for(int b=0;b<Nblock;b++) sssum+=ssq[b];
+
+  sliceNorm(residuals,B,Orthog);
+  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
+
+  sliceNorm(residuals,X,Orthog);
+  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
+
+  /************************************************************************
+   * Block conjugate gradient rQ (Sebastien Birk Thesis, after Dubrulle 2001)
+   ************************************************************************
+   * Dimensions:
+   *
+   *   X,B==(Nferm x Nblock)
+   *   A==(Nferm x Nferm)
+   *  
+   * Nferm = Nspin x Ncolour x Ncomplex x Nlattice_site
+   * 
+   * QC = R = B-AX, D = Q     ; QC => Thin QR factorisation (google it)
+   * for k: 
+   *   Z  = AD
+   *   M  = [D^dag Z]^{-1}
+   *   X  = X + D MC
+   *   QS = Q - ZM
+   *   D  = Q + D S^dag
+   *   C  = S C
+   */
+  ///////////////////////////////////////
+  // Initial block: initial search dir is guess
+  ///////////////////////////////////////
+  std::cout << GridLogMessage<<"BlockCGrQ algorithm initialisation " <<std::endl;
+
+  //1.  QC = R = B-AX, D = Q     ; QC => Thin QR factorisation (google it)
+
+  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;
+
+  ///////////////////////////////////////
+  // Timers
+  ///////////////////////////////////////
+  GridStopWatch sliceInnerTimer;
+  GridStopWatch sliceMaddTimer;
+  GridStopWatch QRTimer;
+  GridStopWatch MatrixTimer;
+  GridStopWatch SolverTimer;
+  SolverTimer.Start();
+
+  int k;
+  for (k = 1; k <= MaxIterations; k++) {
+
+    //3. Z  = AD
+    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;
+    sliceMaddTimer.Start();
+    sliceMaddMatrix(X,m_tmp, D,X,Orthog);     
+    sliceMaddTimer.Stop();
+
+    //6. QS = Q - ZM
+    sliceMaddTimer.Start();
+    sliceMaddMatrix(tmp,m_M,Z,Q,Orthog,-1.0);
+    sliceMaddTimer.Stop();
+    QRTimer.Start();
+    ThinQRfact (m_rr, m_S, m_Sinv, Q, tmp);
+    QRTimer.Stop();
+    
+    //7. D  = Q + D S^dag
+    m_tmp = m_S.adjoint();
+    sliceMaddTimer.Start();
+    sliceMaddMatrix(D,m_tmp,D,Q,Orthog);
+    sliceMaddTimer.Stop();
+
+    //8. C  = S C
+    m_C = m_S*m_C;
+    
+    /*********************
+     * convergence monitor
+     *********************
+     */
+    m_rr = m_C.adjoint() * m_C;
+
+    RealD max_resid=0;
+    RealD rrsum=0;
+    RealD rr;
+
+    for(int b=0;b<Nblock;b++) {
+      rrsum+=real(m_rr(b,b));
+      rr = real(m_rr(b,b))/ssq[b];
+      if ( rr > max_resid ) max_resid = rr;
+    }
+
+    std::cout << GridLogIterative << "\titeration "<<k<<" rr_sum "<<rrsum<<" ssq_sum "<< sssum
+	      <<" ave "<<std::sqrt(rrsum/sssum) << " max "<< max_resid <<std::endl;
+
+    if ( max_resid < Tolerance*Tolerance ) { 
+
+      SolverTimer.Stop();
+
+      std::cout << GridLogMessage<<"BlockCGrQ converged in "<<k<<" iterations"<<std::endl;
+
+      for(int b=0;b<Nblock;b++){
+	std::cout << GridLogMessage<< "\t\tblock "<<b<<" computed resid "
+		  << std::sqrt(real(m_rr(b,b))/ssq[b])<<std::endl;
+      }
+      std::cout << GridLogMessage<<"\tMax residual is "<<std::sqrt(max_resid)<<std::endl;
+
+      Linop.HermOp(X, AD);
+      AD = AD-B;
+      std::cout << GridLogMessage <<"\t True residual is " << std::sqrt(norm2(AD)/norm2(B)) <<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 << "\tInnerProd  " << sliceInnerTimer.Elapsed() <<std::endl;
+      std::cout << GridLogMessage << "\tMaddMatrix " << sliceMaddTimer.Elapsed()  <<std::endl;
+      std::cout << GridLogMessage << "\tThinQRfact " << QRTimer.Elapsed()  <<std::endl;
+	    
+      IterationsToComplete = k;
+      return;
+    }
+
+  }
+  std::cout << GridLogMessage << "BlockConjugateGradient(rQ) did NOT converge" << std::endl;
+
+  if (ErrorOnNoConverge) assert(0);
+  IterationsToComplete = k;
+}
+//////////////////////////////////////////////////////////////////////////
+// Block conjugate gradient; Original O'Leary Dimension zero should be the block direction
+//////////////////////////////////////////////////////////////////////////
+void BlockCGsolve(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi) 
+{
+  int Orthog = blockDim; // First dimension is block dim; this is an assumption
   Nblock = Src._grid->_fdimensions[Orthog];
 
   std::cout<<GridLogMessage<<" Block Conjugate Gradient : Orthog "<<Orthog<<" Nblock "<<Nblock<<std::endl;
@@ -162,8 +425,9 @@ void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
      *********************
      */
     RealD max_resid=0;
+    RealD rr;
     for(int b=0;b<Nblock;b++){
-      RealD rr = real(m_rr(b,b))/ssq[b];
+      rr = real(m_rr(b,b))/ssq[b];
       if ( rr > max_resid ) max_resid = rr;
     }
     
@@ -173,13 +437,14 @@ void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
 
       std::cout << GridLogMessage<<"BlockCG converged in "<<k<<" iterations"<<std::endl;
       for(int b=0;b<Nblock;b++){
-	std::cout << GridLogMessage<< "\t\tblock "<<b<<" resid "<< std::sqrt(real(m_rr(b,b))/ssq[b])<<std::endl;
+	std::cout << GridLogMessage<< "\t\tblock "<<b<<" computed resid "
+		  << std::sqrt(real(m_rr(b,b))/ssq[b])<<std::endl;
       }
       std::cout << GridLogMessage<<"\tMax residual is "<<std::sqrt(max_resid)<<std::endl;
 
       Linop.HermOp(Psi, AP);
       AP = AP-Src;
-      std::cout << GridLogMessage <<"\tTrue residual is " << std::sqrt(norm2(AP)/norm2(Src)) <<std::endl;
+      std::cout << GridLogMessage <<"\t True residual is " << std::sqrt(norm2(AP)/norm2(Src)) <<std::endl;
 
       std::cout << GridLogMessage << "Time Breakdown "<<std::endl;
       std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed()     <<std::endl;
@@ -197,35 +462,13 @@ void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
   if (ErrorOnNoConverge) assert(0);
   IterationsToComplete = k;
 }
-};
-
-
 //////////////////////////////////////////////////////////////////////////
 // multiRHS conjugate gradient. Dimension zero should be the block direction
+// Use this for spread out across nodes
 //////////////////////////////////////////////////////////////////////////
-template <class Field>
-class MultiRHSConjugateGradient : public OperatorFunction<Field> {
- public:
-
-  typedef typename Field::scalar_type scomplex;
-
-  const int blockDim = 0;
-
-  int Nblock;
-  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
-  
-   MultiRHSConjugateGradient(RealD tol, Integer maxit, bool err_on_no_conv = true)
-    : Tolerance(tol),
-    MaxIterations(maxit),
-    ErrorOnNoConverge(err_on_no_conv){};
-
-void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi) 
+void CGmultiRHSsolve(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi) 
 {
-  int Orthog = 0; // First dimension is block dim
+  int Orthog = blockDim; // First dimension is block dim
   Nblock = Src._grid->_fdimensions[Orthog];
 
   std::cout<<GridLogMessage<<"MultiRHS Conjugate Gradient : Orthog "<<Orthog<<" Nblock "<<Nblock<<std::endl;
@@ -285,12 +528,10 @@ void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
     MatrixTimer.Stop();
 
     // Alpha
-    //    sliceInnerProductVectorTest(v_pAp_test,P,AP,Orthog);
     sliceInnerTimer.Start();
     sliceInnerProductVector(v_pAp,P,AP,Orthog);
     sliceInnerTimer.Stop();
     for(int b=0;b<Nblock;b++){
-      //      std::cout << " "<< v_pAp[b]<<" "<< v_pAp_test[b]<<std::endl;
       v_alpha[b] = v_rr[b]/real(v_pAp[b]);
     }
 
@@ -332,7 +573,7 @@ void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
 
       std::cout << GridLogMessage<<"MultiRHS solver converged in " <<k<<" iterations"<<std::endl;
       for(int b=0;b<Nblock;b++){
-	std::cout << GridLogMessage<< "\t\tBlock "<<b<<" resid "<< std::sqrt(v_rr[b]/ssq[b])<<std::endl;
+	std::cout << GridLogMessage<< "\t\tBlock "<<b<<" computed resid "<< std::sqrt(v_rr[b]/ssq[b])<<std::endl;
       }
       std::cout << GridLogMessage<<"\tMax residual is "<<std::sqrt(max_resid)<<std::endl;
 
@@ -358,9 +599,8 @@ void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
   if (ErrorOnNoConverge) assert(0);
   IterationsToComplete = k;
 }
+
 };
 
-
-
 }
 #endif

lib/algorithms/iterative/BlockImplicitlyRestartedLanczos/BlockImplicitlyRestartedLanczos.h

@@ -0,0 +1,753 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/ImplicitlyRestartedLanczos.h
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+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
+    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_BIRL_H
+#define GRID_BIRL_H
+
+#include <string.h> //memset
+
+#include <zlib.h>
+#include <sys/stat.h>
+
+#include <Grid/algorithms/iterative/BlockImplicitlyRestartedLanczos/BlockedGrid.h>
+#include <Grid/algorithms/iterative/BlockImplicitlyRestartedLanczos/FieldBasisVector.h>
+#include <Grid/algorithms/iterative/BlockImplicitlyRestartedLanczos/BlockProjector.h>
+
+namespace Grid { 
+
+/////////////////////////////////////////////////////////////
+// Implicitly restarted lanczos
+/////////////////////////////////////////////////////////////
+
+ template<class Field> 
+ class BlockImplicitlyRestartedLanczos {
+
+    const RealD small = 1.0e-16;
+public:       
+    int lock;
+    int get;
+    int Niter;
+    int converged;
+
+    int Nminres; // 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 kryloc space
+    int Nm;      // Nm -- total number of vectors
+
+    int orth_period;
+
+    RealD OrthoTime;
+
+    RealD eresid, betastp;
+    SortEigen<Field> _sort;
+    LinearFunction<Field> &_HermOp;
+    LinearFunction<Field> &_HermOpTest;
+    /////////////////////////
+    // Constructor
+    /////////////////////////
+
+    BlockImplicitlyRestartedLanczos(
+			       LinearFunction<Field> & HermOp,
+			       LinearFunction<Field> & HermOpTest,
+			       int _Nstop, // sought vecs
+			       int _Nk, // sought vecs
+			       int _Nm, // spare vecs
+			       RealD _eresid, // resid in lmdue deficit 
+			       RealD _betastp, // if beta(k) < betastp: converged
+			       int _Niter, // Max iterations
+			       int _Nminres, int _orth_period = 1) :
+      _HermOp(HermOp),
+      _HermOpTest(HermOpTest),
+      Nstop(_Nstop),
+      Nk(_Nk),
+      Nm(_Nm),
+      eresid(_eresid),
+      betastp(_betastp),
+      Niter(_Niter),
+	Nminres(_Nminres),
+	orth_period(_orth_period)
+    { 
+      Np = Nm-Nk; assert(Np>0);
+    };
+
+    BlockImplicitlyRestartedLanczos(
+			       LinearFunction<Field> & HermOp,
+			       LinearFunction<Field> & HermOpTest,
+			       int _Nk, // sought vecs
+			       int _Nm, // spare vecs
+			       RealD _eresid, // resid in lmdue deficit 
+			       RealD _betastp, // if beta(k) < betastp: converged
+			       int _Niter, // Max iterations
+			       int _Nminres,
+			       int _orth_period = 1) : 
+      _HermOp(HermOp),
+      _HermOpTest(HermOpTest),
+      Nstop(_Nk),
+      Nk(_Nk),
+      Nm(_Nm),
+      eresid(_eresid),
+      betastp(_betastp),
+      Niter(_Niter),
+	Nminres(_Nminres),
+	orth_period(_orth_period)
+    { 
+      Np = Nm-Nk; assert(Np>0);
+    };
+
+
+/* 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:
+3. wk:=Avk−βkv_{k−1}      
+4. αk:=(wk,vk)       // 
+5. wk:=wk−αkvk       // wk orthog vk 
+6. βk+1 := ∥wk∥2. If βk+1 = 0 then Stop
+7. vk+1 := wk/βk+1
+8. EndDo
+ */
+    void step(std::vector<RealD>& lmd,
+	      std::vector<RealD>& lme, 
+	      BasisFieldVector<Field>& evec,
+	      Field& w,int Nm,int k)
+    {
+      assert( k< Nm );
+
+      GridStopWatch gsw_op,gsw_o;
+
+      Field& evec_k = evec[k];
+
+      gsw_op.Start();
+      _HermOp(evec_k,w);
+      gsw_op.Stop();
+
+      if(k>0){
+	w -= lme[k-1] * evec[k-1];
+      }    
+
+      ComplexD zalph = innerProduct(evec_k,w); // 4. αk:=(wk,vk)
+      RealD     alph = real(zalph);
+
+      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
+
+      std::cout<<GridLogMessage << "alpha[" << k << "] = " << zalph << " beta[" << k << "] = "<<beta<<std::endl;
+      const RealD tiny = 1.0e-20;
+      if ( beta < tiny ) { 
+	std::cout<<GridLogMessage << " beta is tiny "<<beta<<std::endl;
+     }
+      lmd[k] = alph;
+      lme[k]  = beta;
+
+      gsw_o.Start();
+      if (k>0 && k % orth_period == 0) { 
+	orthogonalize(w,evec,k); // orthonormalise
+      }
+      gsw_o.Stop();
+
+      if(k < Nm-1) { 
+	evec[k+1] = w;
+      }
+
+      std::cout << GridLogMessage << "Timing: operator=" << gsw_op.Elapsed() <<
+	" orth=" << gsw_o.Elapsed() << std::endl;
+
+    }
+
+    void qr_decomp(std::vector<RealD>& lmd,
+		   std::vector<RealD>& lme,
+		   int Nk,
+		   int Nm,
+		   std::vector<RealD>& Qt,
+		   RealD Dsh, 
+		   int kmin,
+		   int kmax)
+    {
+      int k = kmin-1;
+      RealD x;
+
+      RealD Fden = 1.0/hypot(lmd[k]-Dsh,lme[k]);
+      RealD c = ( lmd[k] -Dsh) *Fden;
+      RealD s = -lme[k] *Fden;
+      
+      RealD tmpa1 = lmd[k];
+      RealD tmpa2 = lmd[k+1];
+      RealD tmpb  = lme[k];
+
+      lmd[k]   = c*c*tmpa1 +s*s*tmpa2 -2.0*c*s*tmpb;
+      lmd[k+1] = s*s*tmpa1 +c*c*tmpa2 +2.0*c*s*tmpb;
+      lme[k]   = c*s*(tmpa1-tmpa2) +(c*c-s*s)*tmpb;
+      x        =-s*lme[k+1];
+      lme[k+1] = c*lme[k+1];
+      
+      for(int i=0; i<Nk; ++i){
+	RealD Qtmp1 = Qt[i+Nm*k  ];
+	RealD Qtmp2 = Qt[i+Nm*(k+1)];
+	Qt[i+Nm*k    ] = c*Qtmp1 - s*Qtmp2;
+	Qt[i+Nm*(k+1)] = s*Qtmp1 + c*Qtmp2; 
+      }
+
+      // Givens transformations
+      for(int k = kmin; k < kmax-1; ++k){
+
+	RealD Fden = 1.0/hypot(x,lme[k-1]);
+	RealD c = lme[k-1]*Fden;
+	RealD s = - x*Fden;
+	
+	RealD tmpa1 = lmd[k];
+	RealD tmpa2 = lmd[k+1];
+	RealD tmpb  = lme[k];
+
+	lmd[k]   = c*c*tmpa1 +s*s*tmpa2 -2.0*c*s*tmpb;
+	lmd[k+1] = s*s*tmpa1 +c*c*tmpa2 +2.0*c*s*tmpb;
+	lme[k]   = c*s*(tmpa1-tmpa2) +(c*c-s*s)*tmpb;
+	lme[k-1] = c*lme[k-1] -s*x;
+
+	if(k != kmax-2){
+	  x = -s*lme[k+1];
+	  lme[k+1] = c*lme[k+1];
+	}
+
+	for(int i=0; i<Nk; ++i){
+	  RealD Qtmp1 = Qt[i+Nm*k    ];
+	  RealD Qtmp2 = Qt[i+Nm*(k+1)];
+	  Qt[i+Nm*k    ] = c*Qtmp1 -s*Qtmp2;
+	  Qt[i+Nm*(k+1)] = s*Qtmp1 +c*Qtmp2;
+	}
+      }
+    }
+
+#ifdef USE_LAPACK_IRL
+#define LAPACK_INT int
+//long long
+    void diagonalize_lapack(std::vector<RealD>& lmd,
+			    std::vector<RealD>& lme, 
+			    int N1,
+			    int N2,
+			    std::vector<RealD>& Qt,
+			    GridBase *grid){
+
+      std::cout << GridLogMessage << "diagonalize_lapack start\n";
+      GridStopWatch gsw;
+
+      const int size = Nm;
+      //  tevals.resize(size);
+      //  tevecs.resize(size);
+      LAPACK_INT NN = N1;
+      std::vector<double> evals_tmp(NN);
+      std::vector<double> evec_tmp(NN*NN);
+      memset(&evec_tmp[0],0,sizeof(double)*NN*NN);
+      //  double AA[NN][NN];
+      std::vector<double> DD(NN);
+      std::vector<double> EE(NN);
+      for (int i = 0; i< NN; i++)
+	for (int j = i - 1; j <= i + 1; j++)
+	  if ( j < NN && j >= 0 ) {
+	    if (i==j) DD[i] = lmd[i];
+	    if (i==j) evals_tmp[i] = lmd[i];
+	    if (j==(i-1)) EE[j] = lme[j];
+	  }
+      LAPACK_INT evals_found;
+      LAPACK_INT lwork = ( (18*NN) > (1+4*NN+NN*NN)? (18*NN):(1+4*NN+NN*NN)) ;
+      LAPACK_INT liwork =  3+NN*10 ;
+      std::vector<LAPACK_INT> iwork(liwork);
+      std::vector<double> work(lwork);
+      std::vector<LAPACK_INT> isuppz(2*NN);
+      char jobz = 'V'; // calculate evals & evecs
+      char range = 'I'; // calculate all evals
+      //    char range = 'A'; // calculate all evals
+      char uplo = 'U'; // refer to upper half of original matrix
+      char compz = 'I'; // Compute eigenvectors of tridiagonal matrix
+      std::vector<int> ifail(NN);
+      LAPACK_INT info;
+      //  int total = QMP_get_number_of_nodes();
+      //  int node = QMP_get_node_number();
+      //  GridBase *grid = evec[0]._grid;
+      int total = grid->_Nprocessors;
+      int node = grid->_processor;
+      int interval = (NN/total)+1;
+      double vl = 0.0, vu = 0.0;
+      LAPACK_INT il = interval*node+1 , iu = interval*(node+1);
+      if (iu > NN)  iu=NN;
+      double tol = 0.0;
+      if (1) {
+	memset(&evals_tmp[0],0,sizeof(double)*NN);
+	if ( il <= NN){
+	  std::cout << GridLogMessage << "dstegr started" << std::endl; 
+	  gsw.Start();
+	  dstegr(&jobz, &range, &NN,
+		 (double*)&DD[0], (double*)&EE[0],
+		 &vl, &vu, &il, &iu, // these four are ignored if second parameteris 'A'
+		 &tol, // tolerance
+		 &evals_found, &evals_tmp[0], (double*)&evec_tmp[0], &NN,
+		 &isuppz[0],
+		 &work[0], &lwork, &iwork[0], &liwork,
+		 &info);
+	  gsw.Stop();
+	  std::cout << GridLogMessage << "dstegr completed in " << gsw.Elapsed() << std::endl;
+	  for (int i = iu-1; i>= il-1; i--){
+	    evals_tmp[i] = evals_tmp[i - (il-1)];
+	    if (il>1) evals_tmp[i-(il-1)]=0.;
+	    for (int j = 0; j< NN; j++){
+	      evec_tmp[i*NN + j] = evec_tmp[(i - (il-1)) * NN + j];
+	      if (il>1) evec_tmp[(i-(il-1)) * NN + j]=0.;
+	    }
+	  }
+	}
+	{
+	  //        QMP_sum_double_array(evals_tmp,NN);
+	  //        QMP_sum_double_array((double *)evec_tmp,NN*NN);
+	  grid->GlobalSumVector(&evals_tmp[0],NN);
+	  grid->GlobalSumVector(&evec_tmp[0],NN*NN);
+	}
+      } 
+      // cheating a bit. It is better to sort instead of just reversing it, but the document of the routine says evals are sorted in increasing order. qr gives evals in decreasing order.
+      for(int i=0;i<NN;i++){
+	for(int j=0;j<NN;j++)
+	  Qt[(NN-1-i)*N2+j]=evec_tmp[i*NN + j];
+	lmd [NN-1-i]=evals_tmp[i];
+      }
+
+      std::cout << GridLogMessage << "diagonalize_lapack complete\n";
+    }
+#undef LAPACK_INT 
+#endif
+
+
+    void diagonalize(std::vector<RealD>& lmd,
+		     std::vector<RealD>& lme, 
+		     int N2,
+		     int N1,
+		     std::vector<RealD>& Qt,
+		     GridBase *grid)
+    {
+
+#ifdef USE_LAPACK_IRL
+    const int check_lapack=0; // just use lapack if 0, check against lapack if 1
+
+    if(!check_lapack)
+	return diagonalize_lapack(lmd,lme,N2,N1,Qt,grid);
+
+	std::vector <RealD> lmd2(N1);
+	std::vector <RealD> lme2(N1);
+	std::vector<RealD> Qt2(N1*N1);
+         for(int k=0; k<N1; ++k){
+	    lmd2[k] = lmd[k];
+	    lme2[k] = lme[k];
+	  }
+         for(int k=0; k<N1*N1; ++k)
+	Qt2[k] = Qt[k];
+
+//	diagonalize_lapack(lmd2,lme2,Nm2,Nm,Qt,grid);
+#endif
+
+      int Niter = 10000*N1;
+      int kmin = 1;
+      int kmax = N2;
+      // (this should be more sophisticated)
+
+      for(int iter=0; ; ++iter){
+      if ( (iter+1)%(100*N1)==0) 
+      std::cout<<GridLogMessage << "[QL method] Not converged - iteration "<<iter+1<<"\n";
+
+	// determination of 2x2 leading submatrix
+	RealD dsub = lmd[kmax-1]-lmd[kmax-2];
+	RealD dd = sqrt(dsub*dsub + 4.0*lme[kmax-2]*lme[kmax-2]);
+	RealD Dsh = 0.5*(lmd[kmax-2]+lmd[kmax-1] +dd*(dsub/fabs(dsub)));
+	// (Dsh: shift)
+	
+	// transformation
+	qr_decomp(lmd,lme,N2,N1,Qt,Dsh,kmin,kmax);
+	
+	// Convergence criterion (redef of kmin and kamx)
+	for(int j=kmax-1; j>= kmin; --j){
+	  RealD dds = fabs(lmd[j-1])+fabs(lmd[j]);
+	  if(fabs(lme[j-1])+dds > dds){
+	    kmax = j+1;
+	    goto continued;
+	  }
+	}
+	Niter = iter;
+#ifdef USE_LAPACK_IRL
+    if(check_lapack){
+	const double SMALL=1e-8;
+	diagonalize_lapack(lmd2,lme2,N2,N1,Qt2,grid);
+	std::vector <RealD> lmd3(N2);
+         for(int k=0; k<N2; ++k) lmd3[k]=lmd[k];
+        _sort.push(lmd3,N2);
+        _sort.push(lmd2,N2);
+         for(int k=0; k<N2; ++k){
+	    if (fabs(lmd2[k] - lmd3[k]) >SMALL)  std::cout<<GridLogMessage <<"lmd(qr) lmd(lapack) "<< k << ": " << lmd2[k] <<" "<< lmd3[k] <<std::endl;
+//	    if (fabs(lme2[k] - lme[k]) >SMALL)  std::cout<<GridLogMessage <<"lme(qr)-lme(lapack) "<< k << ": " << lme2[k] - lme[k] <<std::endl;
+	  }
+         for(int k=0; k<N1*N1; ++k){
+//	    if (fabs(Qt2[k] - Qt[k]) >SMALL)  std::cout<<GridLogMessage <<"Qt(qr)-Qt(lapack) "<< k << ": " << Qt2[k] - Qt[k] <<std::endl;
+	}
+    }
+#endif
+	return;
+
+      continued:
+	for(int j=0; j<kmax-1; ++j){
+	  RealD dds = fabs(lmd[j])+fabs(lmd[j+1]);
+	  if(fabs(lme[j])+dds > dds){
+	    kmin = j+1;
+	    break;
+	  }
+	}
+      }
+      std::cout<<GridLogMessage << "[QL method] Error - Too many iteration: "<<Niter<<"\n";
+      abort();
+    }
+
+#if 1
+    template<typename T>
+    static RealD normalise(T& v) 
+    {
+      RealD nn = norm2(v);
+      nn = sqrt(nn);
+      v = v * (1.0/nn);
+      return nn;
+    }
+
+    void orthogonalize(Field& w,
+		       BasisFieldVector<Field>& evec,
+		       int k)
+    {
+      double t0=-usecond()/1e6;
+
+      evec.orthogonalize(w,k);
+
+      normalise(w);
+      t0+=usecond()/1e6;
+      OrthoTime +=t0;
+    }
+
+    void setUnit_Qt(int Nm, std::vector<RealD> &Qt) {
+      for(int i=0; i<Qt.size(); ++i) Qt[i] = 0.0;
+      for(int k=0; k<Nm; ++k) Qt[k + k*Nm] = 1.0;
+    }
+
+/* Rudy Arthur's thesis pp.137
+------------------------
+Require: M > K P = M − K †
+Compute the factorization AVM = VM HM + fM eM 
+repeat
+  Q=I
+  for i = 1,...,P do
+    QiRi =HM −θiI Q = QQi
+    H M = Q †i H M Q i
+  end for
+  βK =HM(K+1,K) σK =Q(M,K)
+  r=vK+1βK +rσK
+  VK =VM(1:M)Q(1:M,1:K)
+  HK =HM(1:K,1:K)
+  →AVK =VKHK +fKe†K † Extend to an M = K + P step factorization AVM = VMHM + fMeM
+until convergence
+*/
+
+    void calc(std::vector<RealD>& eval,
+	      BasisFieldVector<Field>& evec,
+	      const Field& src,
+	      int& Nconv,
+	      bool reverse,
+	      int SkipTest)
+      {
+
+	GridBase *grid = evec._v[0]._grid;//evec.get(0 + evec_offset)._grid;
+	assert(grid == src._grid);
+
+	std::cout<<GridLogMessage << " -- Nk = " << Nk << " Np = "<< Np << std::endl;
+	std::cout<<GridLogMessage << " -- Nm = " << Nm << std::endl;
+	std::cout<<GridLogMessage << " -- size of eval   = " << eval.size() << std::endl;
+	std::cout<<GridLogMessage << " -- size of evec  = " << evec.size() << std::endl;
+	
+	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 << GridLogMessage << " 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);
+	std::vector<RealD> Qt(Nm*Nm);
+
+
+	Field f(grid);
+	Field v(grid);
+  
+	int k1 = 1;
+	int k2 = Nk;
+
+	Nconv = 0;
+
+	RealD beta_k;
+  
+	// Set initial vector
+	evec[0] = src;
+	normalise(evec[0]);
+	std:: cout<<GridLogMessage <<"norm2(evec[0])= " << norm2(evec[0])<<std::endl;
+	
+	// Initial Nk steps
+	OrthoTime=0.;
+	double t0=usecond()/1e6;
+	for(int k=0; k<Nk; ++k) step(eval,lme,evec,f,Nm,k);
+	double t1=usecond()/1e6;
+	std::cout<<GridLogMessage <<"IRL::Initial steps: "<<t1-t0<< "seconds"<<std::endl; t0=t1;
+	std::cout<<GridLogMessage <<"IRL::Initial steps:OrthoTime "<<OrthoTime<< "seconds"<<std::endl;
+	t1=usecond()/1e6;
+
+	// Restarting loop begins
+	for(int iter = 0; iter<Niter; ++iter){
+	  
+	  std::cout<<GridLogMessage<<"\n Restart iteration = "<< iter << std::endl;
+	  
+	  // 
+	  // Rudy does a sort first which looks very different. Getting fed up with sorting out the algo defs.
+	  // We loop over 
+	  //
+	  OrthoTime=0.;
+	  for(int k=Nk; k<Nm; ++k) step(eval,lme,evec,f,Nm,k);
+	  t1=usecond()/1e6;
+	  std::cout<<GridLogMessage <<"IRL:: "<<Np <<" steps: "<<t1-t0<< "seconds"<<std::endl; t0=t1;
+	  std::cout<<GridLogMessage <<"IRL::Initial steps:OrthoTime "<<OrthoTime<< "seconds"<<std::endl;
+	  f *= lme[Nm-1];
+	  
+	  t1=usecond()/1e6;
+
+	  
+	  // getting eigenvalues
+	  for(int k=0; k<Nm; ++k){
+	    eval2[k] = eval[k+k1-1];
+	    lme2[k] = lme[k+k1-1];
+	  }
+	  setUnit_Qt(Nm,Qt);
+	  diagonalize(eval2,lme2,Nm,Nm,Qt,grid);
+	  t1=usecond()/1e6;
+	  std::cout<<GridLogMessage <<"IRL:: diagonalize: "<<t1-t0<< "seconds"<<std::endl; t0=t1;
+	  
+	  // sorting
+	  eval2_copy = eval2;
+
+	  _sort.push(eval2,Nm);
+	  t1=usecond()/1e6;
+	  std::cout<<GridLogMessage <<"IRL:: eval sorting: "<<t1-t0<< "seconds"<<std::endl; t0=t1;
+	  
+	  // Implicitly shifted QR transformations
+	  setUnit_Qt(Nm,Qt);
+	  for(int ip=0; ip<k2; ++ip){
+	    std::cout<<GridLogMessage << "eval "<< ip << " "<< eval2[ip] << std::endl;
+	  }
+
+	  for(int ip=k2; ip<Nm; ++ip){ 
+	    std::cout<<GridLogMessage << "qr_decomp "<< ip << " "<< eval2[ip] << std::endl;
+	    qr_decomp(eval,lme,Nm,Nm,Qt,eval2[ip],k1,Nm);
+	    
+	  }
+	  t1=usecond()/1e6;
+	  std::cout<<GridLogMessage <<"IRL::qr_decomp: "<<t1-t0<< "seconds"<<std::endl; t0=t1;
+	  assert(k2<Nm);
+	  
+
+	  assert(k2<Nm);
+	  assert(k1>0);
+	  evec.rotate(Qt,k1-1,k2+1,0,Nm,Nm);
+	  
+	  t1=usecond()/1e6;
+	  std::cout<<GridLogMessage <<"IRL::QR rotation: "<<t1-t0<< "seconds"<<std::endl; t0=t1;
+	  fflush(stdout);
+	  
+	  // Compressed vector f and beta(k2)
+	  f *= Qt[Nm-1+Nm*(k2-1)];
+	  f += lme[k2-1] * evec[k2];
+	  beta_k = norm2(f);
+	  beta_k = sqrt(beta_k);
+	  std::cout<<GridLogMessage<<" beta(k) = "<<beta_k<<std::endl;
+	  
+	  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];
+
+	    std::cout<<GridLogMessage << "eval2[" << k << "] = " << eval2[k] << std::endl;
+	  }
+	  setUnit_Qt(Nm,Qt);
+	  diagonalize(eval2,lme2,Nk,Nm,Qt,grid);
+	  t1=usecond()/1e6;
+	  std::cout<<GridLogMessage <<"IRL::diagonalize: "<<t1-t0<< "seconds"<<std::endl; t0=t1;
+	  
+	  
+	  Nconv = 0;
+	  
+	  if (iter >= Nminres) {
+	    std::cout << GridLogMessage << "Rotation to test convergence " << std::endl;
+	    
+	    Field ev0_orig(grid);
+	    ev0_orig = evec[0];
+	    
+	    evec.rotate(Qt,0,Nk,0,Nk,Nm);
+	    
+	    {
+	      std::cout << GridLogMessage << "Test convergence" << std::endl;
+	      Field B(grid);
+	      
+	      for(int j = 0; j<Nk; j+=SkipTest){
+		B=evec[j];
+		//std::cout << "Checkerboard: " << evec[j].checkerboard << std::endl; 
+		B.checkerboard = evec[0].checkerboard;
+
+		_HermOpTest(B,v);
+		
+		RealD vnum = real(innerProduct(B,v)); // HermOp.
+		RealD vden = norm2(B);
+		RealD vv0 = norm2(v);
+		eval2[j] = vnum/vden;
+		v -= eval2[j]*B;
+		RealD vv = norm2(v) / ::pow(evalMaxApprox,2.0);
+		std::cout.precision(13);
+		std::cout<<GridLogMessage << "[" << std::setw(3)<< std::setiosflags(std::ios_base::right) <<j<<"] "
+			 <<"eval = "<<std::setw(25)<< std::setiosflags(std::ios_base::left)<< eval2[j] << " (" << eval2_copy[j] << ")"
+			 <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25)<< std::setiosflags(std::ios_base::right)<< vv
+			 <<" "<< vnum/(sqrt(vden)*sqrt(vv0))
+			 << " norm(B["<<j<<"])="<< vden <<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) && (j == Nconv) ){
+		  Nconv+=SkipTest;
+		}
+	      }
+	      
+	      // test if we converged, if so, terminate
+	      t1=usecond()/1e6;
+	      std::cout<<GridLogMessage <<"IRL::convergence testing: "<<t1-t0<< "seconds"<<std::endl; t0=t1;
+	      
+	      std::cout<<GridLogMessage<<" #modes converged: "<<Nconv<<std::endl;
+	      
+	      if( Nconv>=Nstop || beta_k < betastp){
+		goto converged;
+	      }
+	      
+	      std::cout << GridLogMessage << "Rotate back" << std::endl;
+	      //B[j] +=Qt[k+_Nm*j] * _v[k]._odata[ss];
+	      {
+		Eigen::MatrixXd qm = Eigen::MatrixXd::Zero(Nk,Nk);
+		for (int k=0;k<Nk;k++)
+		  for (int j=0;j<Nk;j++)
+		    qm(j,k) = Qt[k+Nm*j];
+		GridStopWatch timeInv;
+		timeInv.Start();
+		Eigen::MatrixXd qmI = qm.inverse();
+		timeInv.Stop();
+		std::vector<RealD> QtI(Nm*Nm);
+		for (int k=0;k<Nk;k++)
+		  for (int j=0;j<Nk;j++)
+		    QtI[k+Nm*j] = qmI(j,k);
+		
+		RealD res_check_rotate_inverse = (qm*qmI - Eigen::MatrixXd::Identity(Nk,Nk)).norm(); // sqrt( |X|^2 )
+		assert(res_check_rotate_inverse < 1e-7);
+		evec.rotate(QtI,0,Nk,0,Nk,Nm);
+		
+		axpy(ev0_orig,-1.0,evec[0],ev0_orig);
+		std::cout << GridLogMessage << "Rotation done (in " << timeInv.Elapsed() << " = " << timeInv.useconds() << " us" <<
+		  ", error = " << res_check_rotate_inverse << 
+		  "); | evec[0] - evec[0]_orig | = " << ::sqrt(norm2(ev0_orig)) << std::endl;
+	      }
+	    }
+	  } else {
+	    std::cout << GridLogMessage << "iter < Nminres: do not yet test for convergence\n";
+	  } // end of iter loop
+	}
+
+	std::cout<<GridLogMessage<<"\n NOT converged.\n";
+	abort();
+	
+      converged:
+
+	if (SkipTest == 1) {
+	  eval = eval2;
+	} else {
+
+	  // test quickly
+	  for (int j=0;j<Nstop;j+=SkipTest) {
+	    std::cout<<GridLogMessage << "Eigenvalue[" << j << "] = " << eval2[j] << " (" << eval2_copy[j] << ")" << std::endl;
+	  }
+
+	  eval2_copy.resize(eval2.size());
+	  eval = eval2_copy;
+	}
+
+	evec.sortInPlace(eval,reverse);
+
+	{
+	  
+	 // test
+	 for (int j=0;j<Nstop;j++) {
+	   std::cout<<GridLogMessage << " |e[" << j << "]|^2 = " << norm2(evec[j]) << std::endl;
+	 }
+       }
+       
+       //_sort.push(eval,evec,Nconv);
+       //evec.sort(eval,Nconv);
+       
+       std::cout<<GridLogMessage << "\n Converged\n Summary :\n";
+       std::cout<<GridLogMessage << " -- Iterations  = "<< Nconv  << "\n";
+       std::cout<<GridLogMessage << " -- beta(k)     = "<< beta_k << "\n";
+       std::cout<<GridLogMessage << " -- Nconv       = "<< Nconv  << "\n";
+      }
+#endif
+
+    };
+
+}
+#endif
+

lib/algorithms/iterative/BlockImplicitlyRestartedLanczos/BlockProjector.h

@@ -0,0 +1,143 @@
+namespace Grid { 
+
+/*
+  BlockProjector
+
+  If _HP_BLOCK_PROJECTORS_ is defined, we assume that _evec is a basis that is not
+  fully orthonormalized (to the precision of the coarse field) and we allow for higher-precision
+  coarse field than basis field.
+
+*/
+//#define _HP_BLOCK_PROJECTORS_
+
+template<typename Field>
+class BlockProjector {
+public:
+
+  BasisFieldVector<Field>& _evec;
+  BlockedGrid<Field>& _bgrid;
+
+  BlockProjector(BasisFieldVector<Field>& evec, BlockedGrid<Field>& bgrid) : _evec(evec), _bgrid(bgrid) {
+  }
+
+  void createOrthonormalBasis(RealD thres = 0.0) {
+
+    GridStopWatch sw;
+    sw.Start();
+
+    int cnt = 0;
+
+#pragma omp parallel shared(cnt)
+    {
+      int lcnt = 0;
+
+#pragma omp for
+      for (int b=0;b<_bgrid._o_blocks;b++) {
+	
+	for (int i=0;i<_evec._Nm;i++) {
+	  
+	  auto nrm0 = _bgrid.block_sp(b,_evec._v[i],_evec._v[i]);
+	  
+	  // |i> -= <j|i> |j>
+	  for (int j=0;j<i;j++) {
+	    _bgrid.block_caxpy(b,_evec._v[i],-_bgrid.block_sp(b,_evec._v[j],_evec._v[i]),_evec._v[j],_evec._v[i]);
+	  }
+	  
+	  auto nrm = _bgrid.block_sp(b,_evec._v[i],_evec._v[i]);
+	  
+	  auto eps = nrm/nrm0;
+	  if (Reduce(eps).real() < thres) {
+	    lcnt++;
+	  }
+	  
+	  // TODO: if norm is too small, remove this eigenvector/mark as not needed; in practice: set it to zero norm here and return a mask
+	  // that is then used later to decide not to write certain eigenvectors to disk (add a norm calculation before subtraction step and look at nrm/nrm0 < eps to decide)
+	  _bgrid.block_cscale(b,1.0 / sqrt(nrm),_evec._v[i]);
+	  
+	}
+	
+      }
+
+#pragma omp critical
+      {
+	cnt += lcnt;
+      }
+    }
+    sw.Stop();
+    std::cout << GridLogMessage << "Gram-Schmidt to create blocked basis took " << sw.Elapsed() << " (" << ((RealD)cnt / (RealD)_bgrid._o_blocks / (RealD)_evec._Nm) 
+	      << " below threshold)" << std::endl;
+
+  }
+
+  template<typename CoarseField>
+  void coarseToFine(const CoarseField& in, Field& out) {
+
+    out = zero;
+    out.checkerboard = _evec._v[0].checkerboard;
+
+    int Nbasis = sizeof(in._odata[0]._internal._internal) / sizeof(in._odata[0]._internal._internal[0]);
+    assert(Nbasis == _evec._Nm);
+    
+#pragma omp parallel for
+    for (int b=0;b<_bgrid._o_blocks;b++) {
+      for (int j=0;j<_evec._Nm;j++) {
+	_bgrid.block_caxpy(b,out,in._odata[b]._internal._internal[j],_evec._v[j],out);
+      }
+    }
+
+  }
+
+  template<typename CoarseField>
+  void fineToCoarse(const Field& in, CoarseField& out) {
+
+    out = zero;
+
+    int Nbasis = sizeof(out._odata[0]._internal._internal) / sizeof(out._odata[0]._internal._internal[0]);
+    assert(Nbasis == _evec._Nm);
+
+
+    Field tmp(_bgrid._grid);
+    tmp = in;
+    
+#pragma omp parallel for
+    for (int b=0;b<_bgrid._o_blocks;b++) {
+      for (int j=0;j<_evec._Nm;j++) {
+	// |rhs> -= <j|rhs> |j>
+	auto c = _bgrid.block_sp(b,_evec._v[j],tmp);
+	_bgrid.block_caxpy(b,tmp,-c,_evec._v[j],tmp); // may make this more numerically stable
+	out._odata[b]._internal._internal[j] = c;
+      }
+    }
+
+  }
+
+  template<typename CoarseField>
+    void deflateFine(BasisFieldVector<CoarseField>& _coef,const std::vector<RealD>& eval,int N,const Field& src_orig,Field& result) {
+    result = zero;
+    for (int i=0;i<N;i++) {
+      Field tmp(result._grid);
+      coarseToFine(_coef._v[i],tmp);
+      axpy(result,TensorRemove(innerProduct(tmp,src_orig)) / eval[i],tmp,result);
+    }
+  }
+
+  template<typename CoarseField>
+    void deflateCoarse(BasisFieldVector<CoarseField>& _coef,const std::vector<RealD>& eval,int N,const Field& src_orig,Field& result) {
+    CoarseField src_coarse(_coef._v[0]._grid);
+    CoarseField result_coarse = src_coarse;
+    result_coarse = zero;
+    fineToCoarse(src_orig,src_coarse);
+    for (int i=0;i<N;i++) {
+      axpy(result_coarse,TensorRemove(innerProduct(_coef._v[i],src_coarse)) / eval[i],_coef._v[i],result_coarse);
+    }
+    coarseToFine(result_coarse,result);
+  }
+
+  template<typename CoarseField>
+    void deflate(BasisFieldVector<CoarseField>& _coef,const std::vector<RealD>& eval,int N,const Field& src_orig,Field& result) {
+    // Deflation on coarse Grid is much faster, so use it by default.  Deflation on fine Grid is kept for legacy reasons for now.
+    deflateCoarse(_coef,eval,N,src_orig,result);
+  }
+
+};
+}

lib/algorithms/iterative/BlockImplicitlyRestartedLanczos/BlockedGrid.h

@@ -0,0 +1,401 @@
+namespace Grid {
+
+template<typename Field>
+class BlockedGrid {
+public:
+  GridBase* _grid;
+  typedef typename Field::scalar_type  Coeff_t;
+  typedef typename Field::vector_type vCoeff_t;
+  
+  std::vector<int> _bs; // block size
+  std::vector<int> _nb; // number of blocks
+  std::vector<int> _l;  // local dimensions irrespective of cb
+  std::vector<int> _l_cb;  // local dimensions of checkerboarded vector
+  std::vector<int> _l_cb_o;  // local dimensions of inner checkerboarded vector
+  std::vector<int> _bs_cb; // block size in checkerboarded vector
+  std::vector<int> _nb_o; // number of blocks of simd o-sites
+
+  int _nd, _blocks, _cf_size, _cf_block_size, _cf_o_block_size, _o_blocks, _block_sites;
+  
+  BlockedGrid(GridBase* grid, const std::vector<int>& block_size) :
+    _grid(grid), _bs(block_size), _nd((int)_bs.size()), 
+      _nb(block_size), _l(block_size), _l_cb(block_size), _nb_o(block_size),
+      _l_cb_o(block_size), _bs_cb(block_size) {
+
+    _blocks = 1;
+    _o_blocks = 1;
+    _l = grid->FullDimensions();
+    _l_cb = grid->LocalDimensions();
+    _l_cb_o = grid->_rdimensions;
+
+    _cf_size = 1;
+    _block_sites = 1;
+    for (int i=0;i<_nd;i++) {
+      _l[i] /= grid->_processors[i];
+
+      assert(!(_l[i] % _bs[i])); // lattice must accommodate choice of blocksize
+
+      int r = _l[i] / _l_cb[i];
+      assert(!(_bs[i] % r)); // checkerboarding must accommodate choice of blocksize
+      _bs_cb[i] = _bs[i] / r;
+      _block_sites *= _bs_cb[i];
+      _nb[i] = _l[i] / _bs[i];
+      _nb_o[i] = _nb[i] / _grid->_simd_layout[i];
+      if (_nb[i] % _grid->_simd_layout[i]) { // simd must accommodate choice of blocksize
+	std::cout << GridLogMessage << "Problem: _nb[" << i << "] = " << _nb[i] << " _grid->_simd_layout[" << i << "] = " << _grid->_simd_layout[i] << std::endl;
+	assert(0);
+      }
+      _blocks *= _nb[i];
+      _o_blocks *= _nb_o[i];
+      _cf_size *= _l[i];
+    }
+
+    _cf_size *= 12 / 2;
+    _cf_block_size = _cf_size / _blocks;
+    _cf_o_block_size = _cf_size / _o_blocks;
+
+    std::cout << GridLogMessage << "BlockedGrid:" << std::endl;
+    std::cout << GridLogMessage << " _l     = " << _l << std::endl;
+    std::cout << GridLogMessage << " _l_cb     = " << _l_cb << std::endl;
+    std::cout << GridLogMessage << " _l_cb_o     = " << _l_cb_o << std::endl;
+    std::cout << GridLogMessage << " _bs    = " << _bs << std::endl;
+    std::cout << GridLogMessage << " _bs_cb    = " << _bs_cb << std::endl;
+
+    std::cout << GridLogMessage << " _nb    = " << _nb << std::endl;
+    std::cout << GridLogMessage << " _nb_o    = " << _nb_o << std::endl;
+    std::cout << GridLogMessage << " _blocks = " << _blocks << std::endl;
+    std::cout << GridLogMessage << " _o_blocks = " << _o_blocks << std::endl;
+    std::cout << GridLogMessage << " sizeof(vCoeff_t) = " << sizeof(vCoeff_t) << std::endl;
+    std::cout << GridLogMessage << " _cf_size = " << _cf_size << std::endl;
+    std::cout << GridLogMessage << " _cf_block_size = " << _cf_block_size << std::endl;
+    std::cout << GridLogMessage << " _block_sites = " << _block_sites << std::endl;
+    std::cout << GridLogMessage << " _grid->oSites() = " << _grid->oSites() << std::endl;
+
+    //    _grid->Barrier();
+    //abort();
+  }
+
+    void block_to_coor(int b, std::vector<int>& x0) {
+
+      std::vector<int> bcoor;
+      bcoor.resize(_nd);
+      x0.resize(_nd);
+      assert(b < _o_blocks);
+      Lexicographic::CoorFromIndex(bcoor,b,_nb_o);
+      int i;
+
+      for (i=0;i<_nd;i++) {
+	x0[i] = bcoor[i]*_bs_cb[i];
+      }
+
+      //std::cout << GridLogMessage << "Map block b -> " << x0 << std::endl;
+
+    }
+
+    void block_site_to_o_coor(const std::vector<int>& x0, std::vector<int>& coor, int i) {
+      Lexicographic::CoorFromIndex(coor,i,_bs_cb);
+      for (int j=0;j<_nd;j++)
+	coor[j] += x0[j];
+    }
+
+    int block_site_to_o_site(const std::vector<int>& x0, int i) {
+      std::vector<int> coor;  coor.resize(_nd);
+      block_site_to_o_coor(x0,coor,i);
+      Lexicographic::IndexFromCoor(coor,i,_l_cb_o);
+      return i;
+    }
+
+    vCoeff_t block_sp(int b, const Field& x, const Field& y) {
+
+      std::vector<int> x0;
+      block_to_coor(b,x0);
+
+      vCoeff_t ret = 0.0;
+      for (int i=0;i<_block_sites;i++) { // only odd sites
+	int ss = block_site_to_o_site(x0,i);
+	ret += TensorRemove(innerProduct(x._odata[ss],y._odata[ss]));
+      }
+
+      return ret;
+
+    }
+
+    vCoeff_t block_sp(int b, const Field& x, const std::vector< ComplexD >& y) {
+
+      std::vector<int> x0;
+      block_to_coor(b,x0);
+
+      constexpr int nsimd = sizeof(vCoeff_t) / sizeof(Coeff_t);
+      int lsize = _cf_o_block_size / _block_sites;
+
+      std::vector< ComplexD > ret(nsimd);
+      for (int i=0;i<nsimd;i++)
+	ret[i] = 0.0;
+
+      for (int i=0;i<_block_sites;i++) { // only odd sites
+	int ss = block_site_to_o_site(x0,i);
+
+	int n = lsize / nsimd;
+	for (int l=0;l<n;l++) {
+	  for (int j=0;j<nsimd;j++) {
+	    int t = lsize * i + l*nsimd + j;
+
+	    ret[j] += conjugate(((Coeff_t*)&x._odata[ss]._internal)[l*nsimd + j]) * y[t];
+	  }
+	}
+      }
+
+      vCoeff_t vret;
+      for (int i=0;i<nsimd;i++)
+	((Coeff_t*)&vret)[i] = (Coeff_t)ret[i];
+
+      return vret;
+
+    }
+
+    template<class T>
+      void vcaxpy(iScalar<T>& r,const vCoeff_t& a,const iScalar<T>& x,const iScalar<T>& y) {
+      vcaxpy(r._internal,a,x._internal,y._internal);
+    }
+
+    template<class T,int N>
+      void vcaxpy(iVector<T,N>& r,const vCoeff_t& a,const iVector<T,N>& x,const iVector<T,N>& y) {
+      for (int i=0;i<N;i++)
+	vcaxpy(r._internal[i],a,x._internal[i],y._internal[i]);
+    }
+
+    void vcaxpy(vCoeff_t& r,const vCoeff_t& a,const vCoeff_t& x,const vCoeff_t& y) {
+      r = a*x + y;
+    }
+
+    void block_caxpy(int b, Field& ret, const vCoeff_t& a, const Field& x, const Field& y) {
+
+      std::vector<int> x0;
+      block_to_coor(b,x0);
+
+      for (int i=0;i<_block_sites;i++) { // only odd sites
+	int ss = block_site_to_o_site(x0,i);
+	vcaxpy(ret._odata[ss],a,x._odata[ss],y._odata[ss]);
+      }
+
+    }
+
+    void block_caxpy(int b, std::vector< ComplexD >& ret, const vCoeff_t& a, const Field& x, const std::vector< ComplexD >& y) {
+      std::vector<int> x0;
+      block_to_coor(b,x0);
+
+      constexpr int nsimd = sizeof(vCoeff_t) / sizeof(Coeff_t);
+      int lsize = _cf_o_block_size / _block_sites;
+
+      for (int i=0;i<_block_sites;i++) { // only odd sites
+	int ss = block_site_to_o_site(x0,i);
+
+	int n = lsize / nsimd;
+	for (int l=0;l<n;l++) {
+	  vCoeff_t r = a* ((vCoeff_t*)&x._odata[ss]._internal)[l];
+
+	  for (int j=0;j<nsimd;j++) {
+	    int t = lsize * i + l*nsimd + j;
+	    ret[t] = y[t] + ((Coeff_t*)&r)[j];
+	  }
+	}
+      }
+
+    }
+
+    void block_set(int b, Field& ret, const std::vector< ComplexD >& x) {
+      std::vector<int> x0;
+      block_to_coor(b,x0);
+
+      int lsize = _cf_o_block_size / _block_sites;
+
+      for (int i=0;i<_block_sites;i++) { // only odd sites
+	int ss = block_site_to_o_site(x0,i);
+
+	for (int l=0;l<lsize;l++)
+	  ((Coeff_t*)&ret._odata[ss]._internal)[l] = (Coeff_t)x[lsize * i + l]; // convert precision
+      }
+
+    }
+
+    void block_get(int b, const Field& ret, std::vector< ComplexD >& x) {
+      std::vector<int> x0;
+      block_to_coor(b,x0);
+
+      int lsize = _cf_o_block_size / _block_sites;
+
+      for (int i=0;i<_block_sites;i++) { // only odd sites
+	int ss = block_site_to_o_site(x0,i);
+
+	for (int l=0;l<lsize;l++)
+	  x[lsize * i + l] = (ComplexD)((Coeff_t*)&ret._odata[ss]._internal)[l];
+      }
+
+    }
+
+    template<class T>
+    void vcscale(iScalar<T>& r,const vCoeff_t& a,const iScalar<T>& x) {
+      vcscale(r._internal,a,x._internal);
+    }
+
+    template<class T,int N>
+    void vcscale(iVector<T,N>& r,const vCoeff_t& a,const iVector<T,N>& x) {
+      for (int i=0;i<N;i++)
+	vcscale(r._internal[i],a,x._internal[i]);
+    }
+
+    void vcscale(vCoeff_t& r,const vCoeff_t& a,const vCoeff_t& x) {
+      r = a*x;
+    }
+
+    void block_cscale(int b, const vCoeff_t& a, Field& ret) {
+
+      std::vector<int> x0;
+      block_to_coor(b,x0);
+      
+      for (int i=0;i<_block_sites;i++) { // only odd sites
+	int ss = block_site_to_o_site(x0,i);
+	vcscale(ret._odata[ss],a,ret._odata[ss]);
+      }
+    }
+
+    void getCanonicalBlockOffset(int cb, std::vector<int>& x0) {
+      const int ndim = 5;
+      assert(_nb.size() == ndim);
+      std::vector<int> _nbc = { _nb[1], _nb[2], _nb[3], _nb[4], _nb[0] };
+      std::vector<int> _bsc = { _bs[1], _bs[2], _bs[3], _bs[4], _bs[0] };
+      x0.resize(ndim);
+
+      assert(cb >= 0);
+      assert(cb < _nbc[0]*_nbc[1]*_nbc[2]*_nbc[3]*_nbc[4]);
+
+      Lexicographic::CoorFromIndex(x0,cb,_nbc);
+      int i;
+
+      for (i=0;i<ndim;i++) {
+	x0[i] *= _bsc[i];
+      }
+
+      //if (cb < 2)
+      //	std::cout << GridLogMessage << "Map: " << cb << " To: " << x0 << std::endl;
+    }
+
+    void pokeBlockOfVectorCanonical(int cb,Field& v,const std::vector<float>& buf) {
+      std::vector<int> _bsc = { _bs[1], _bs[2], _bs[3], _bs[4], _bs[0] };
+      std::vector<int> ldim = v._grid->LocalDimensions();
+      std::vector<int> cldim = { ldim[1], ldim[2], ldim[3], ldim[4], ldim[0] };
+      const int _nbsc = _bs_cb[0]*_bs_cb[1]*_bs_cb[2]*_bs_cb[3]*_bs_cb[4];
+      // take canonical block cb of v and put it in canonical ordering in buf
+      std::vector<int> cx0;
+      getCanonicalBlockOffset(cb,cx0);
+
+#pragma omp parallel
+      {
+	std::vector<int> co0,cl0;
+	co0=cx0; cl0=cx0;
+
+#pragma omp for
+	for (int i=0;i<_nbsc;i++) {
+	  Lexicographic::CoorFromIndex(co0,2*i,_bsc); // 2* for eo
+	  for (int j=0;j<(int)_bsc.size();j++)
+	    cl0[j] = cx0[j] + co0[j];
+	  
+	  std::vector<int> l0 = { cl0[4], cl0[0], cl0[1], cl0[2], cl0[3] };
+	  int oi = v._grid->oIndex(l0);
+	  int ii = v._grid->iIndex(l0);
+	  int lti = i;
+
+	  //if (cb < 2 && i<2)
+	  //  std::cout << GridLogMessage << "Map: " << cb << ", " << i << " To: " << cl0 << ", " << cx0 << ", " << oi << ", " << ii << std::endl;
+	  
+	  for (int s=0;s<4;s++)
+	    for (int c=0;c<3;c++) {
+	      Coeff_t& ld = ((Coeff_t*)&v._odata[oi]._internal._internal[s]._internal[c])[ii];
+	      int ti = 12*lti + 3*s + c;
+	      ld = Coeff_t(buf[2*ti+0], buf[2*ti+1]);
+	    }
+	}
+      }
+    }
+
+    void peekBlockOfVectorCanonical(int cb,const Field& v,std::vector<float>& buf) {
+      std::vector<int> _bsc = { _bs[1], _bs[2], _bs[3], _bs[4], _bs[0] };
+      std::vector<int> ldim = v._grid->LocalDimensions();
+      std::vector<int> cldim = { ldim[1], ldim[2], ldim[3], ldim[4], ldim[0] };
+      const int _nbsc = _bs_cb[0]*_bs_cb[1]*_bs_cb[2]*_bs_cb[3]*_bs_cb[4];
+      // take canonical block cb of v and put it in canonical ordering in buf
+      std::vector<int> cx0;
+      getCanonicalBlockOffset(cb,cx0);
+
+      buf.resize(_cf_block_size * 2);
+
+#pragma omp parallel
+      {
+	std::vector<int> co0,cl0;
+	co0=cx0; cl0=cx0;
+
+#pragma omp for
+	for (int i=0;i<_nbsc;i++) {
+	  Lexicographic::CoorFromIndex(co0,2*i,_bsc); // 2* for eo
+	  for (int j=0;j<(int)_bsc.size();j++)
+	    cl0[j] = cx0[j] + co0[j];
+	  
+	  std::vector<int> l0 = { cl0[4], cl0[0], cl0[1], cl0[2], cl0[3] };
+	  int oi = v._grid->oIndex(l0);
+	  int ii = v._grid->iIndex(l0);
+	  int lti = i;
+	  
+	  //if (cb < 2 && i<2)
+	  //  std::cout << GridLogMessage << "Map: " << cb << ", " << i << " To: " << cl0 << ", " << cx0 << ", " << oi << ", " << ii << std::endl;
+
+	  for (int s=0;s<4;s++)
+	    for (int c=0;c<3;c++) {
+	      Coeff_t& ld = ((Coeff_t*)&v._odata[oi]._internal._internal[s]._internal[c])[ii];
+	      int ti = 12*lti + 3*s + c;
+	      buf[2*ti+0] = ld.real();
+	      buf[2*ti+1] = ld.imag();
+	    }
+	}
+      }
+    }
+
+    int globalToLocalCanonicalBlock(int slot,const std::vector<int>& src_nodes,int nb) {
+      // processor coordinate
+      int _nd = (int)src_nodes.size();
+      std::vector<int> _src_nodes = src_nodes;
+      std::vector<int> pco(_nd);
+      Lexicographic::CoorFromIndex(pco,slot,_src_nodes);
+      std::vector<int> cpco = { pco[1], pco[2], pco[3], pco[4], pco[0] };
+
+      // get local block
+      std::vector<int> _nbc = { _nb[1], _nb[2], _nb[3], _nb[4], _nb[0] };
+      assert(_nd == 5);
+      std::vector<int> c_src_local_blocks(_nd);
+      for (int i=0;i<_nd;i++) {
+	assert(_grid->_fdimensions[i] % (src_nodes[i] * _bs[i]) == 0);
+	c_src_local_blocks[(i+4) % 5] = _grid->_fdimensions[i] / src_nodes[i] / _bs[i];
+      }
+      std::vector<int> cbcoor(_nd); // coordinate of block in slot in canonical form
+      Lexicographic::CoorFromIndex(cbcoor,nb,c_src_local_blocks);
+
+      // cpco, cbcoor
+      std::vector<int> clbcoor(_nd);
+      for (int i=0;i<_nd;i++) {
+	int cgcoor = cpco[i] * c_src_local_blocks[i] + cbcoor[i]; // global block coordinate
+	int pcoor = cgcoor / _nbc[i]; // processor coordinate in my Grid
+	int tpcoor = _grid->_processor_coor[(i+1)%5];
+	if (pcoor != tpcoor)
+	  return -1;
+	clbcoor[i] = cgcoor - tpcoor * _nbc[i]; // canonical local block coordinate for canonical dimension i
+      }
+
+      int lnb;
+      Lexicographic::IndexFromCoor(clbcoor,lnb,_nbc);
+      //std::cout << "Mapped slot = " << slot << " nb = " << nb << " to " << lnb << std::endl;
+      return lnb;
+    }
+
+
+ };
+
+}

lib/algorithms/iterative/BlockImplicitlyRestartedLanczos/FieldBasisVector.h

@@ -0,0 +1,163 @@
+namespace Grid { 
+
+template<class Field>
+class BasisFieldVector {
+ public:
+  int _Nm;
+
+  typedef typename Field::scalar_type Coeff_t;
+  typedef typename Field::vector_type vCoeff_t;
+  typedef typename Field::vector_object vobj;
+  typedef typename vobj::scalar_object sobj;
+
+  std::vector<Field> _v; // _Nfull vectors
+
+  void report(int n,GridBase* value) {
+
+    std::cout << GridLogMessage << "BasisFieldVector allocated:\n";
+    std::cout << GridLogMessage << " Delta N = " << n << "\n";
+    std::cout << GridLogMessage << " Size of full vectors (size) = " << 
+      ((double)n*sizeof(vobj)*value->oSites() / 1024./1024./1024.) << " GB\n";
+    std::cout << GridLogMessage << " Size = " << _v.size() << " Capacity = " << _v.capacity() << std::endl;
+
+    value->Barrier();
+
+    if (value->IsBoss()) {
+      system("cat /proc/meminfo");
+    }
+
+    value->Barrier();
+
+  }
+
+  BasisFieldVector(int Nm,GridBase* value) : _Nm(Nm), _v(Nm,value) {
+    report(Nm,value);
+  }
+  
+  ~BasisFieldVector() {
+  }
+
+  Field& operator[](int i) {
+    return _v[i];
+  }
+
+  void orthogonalize(Field& w, int k) {
+    for(int j=0; j<k; ++j){
+      Coeff_t ip = (Coeff_t)innerProduct(_v[j],w);
+      w = w - ip*_v[j];
+    }
+  }
+
+  void rotate(std::vector<RealD>& Qt,int j0, int j1, int k0,int k1,int Nm) {
+    
+    GridBase* grid = _v[0]._grid;
+      
+#pragma omp parallel
+    {
+      std::vector < vobj > B(Nm);
+      
+#pragma omp for
+      for(int ss=0;ss < grid->oSites();ss++){
+	for(int j=j0; j<j1; ++j) B[j]=0.;
+	
+	for(int j=j0; j<j1; ++j){
+	  for(int k=k0; k<k1; ++k){
+	    B[j] +=Qt[k+Nm*j] * _v[k]._odata[ss];
+	  }
+	}
+	for(int j=j0; j<j1; ++j){
+	  _v[j]._odata[ss] = B[j];
+	}
+      }
+    }
+
+  }
+
+  size_t size() const {
+    return _Nm;
+  }
+
+  void resize(int n) {
+    if (n > _Nm)
+      _v.reserve(n);
+    
+    _v.resize(n,_v[0]._grid);
+
+    if (n < _Nm)
+      _v.shrink_to_fit();
+
+    report(n - _Nm,_v[0]._grid);
+
+    _Nm = n;
+  }
+
+  std::vector<int> getIndex(std::vector<RealD>& sort_vals) {
+
+    std::vector<int> idx(sort_vals.size());
+    iota(idx.begin(), idx.end(), 0);
+
+    // sort indexes based on comparing values in v
+    sort(idx.begin(), idx.end(),
+	 [&sort_vals](int i1, int i2) {return ::fabs(sort_vals[i1]) < ::fabs(sort_vals[i2]);});
+
+    return idx;
+  }
+
+  void reorderInPlace(std::vector<RealD>& sort_vals, std::vector<int>& idx) {
+    GridStopWatch gsw;
+    gsw.Start();
+
+    int nswaps = 0;
+    for (size_t i=0;i<idx.size();i++) {
+      if (idx[i] != i) {
+
+	// find proper place (this could be done in logarithmic time, don't bother for now)
+	size_t j;
+	for (j=i;j<idx.size();j++)
+	  if (idx[j]==i)
+	    break;
+	assert(j!=idx.size());
+	
+	Field _t(_v[0]._grid);
+	_t = _v[idx[j]];
+	_v[idx[j]] = _v[idx[i]];
+	_v[idx[i]] = _t;
+
+	RealD _td = sort_vals[idx[j]];
+	sort_vals[idx[j]] = sort_vals[idx[i]];
+	sort_vals[idx[i]] = _td;
+
+	int _tt = idx[i];
+	idx[i] = idx[j];
+	idx[j] = _tt;
+	
+	nswaps++;
+      }
+    }
+
+    // sort values
+    gsw.Stop();
+    std::cout << GridLogMessage << "Sorted eigenspace in place in " << gsw.Elapsed() << " using " << nswaps << " swaps" << std::endl;
+  }
+
+  void sortInPlace(std::vector<RealD>& sort_vals, bool reverse) {
+
+    std::vector<int> idx = getIndex(sort_vals);
+    if (reverse)
+      std::reverse(idx.begin(), idx.end());
+
+    reorderInPlace(sort_vals,idx);
+
+  }
+
+  void deflate(const std::vector<RealD>& eval,const Field& src_orig,Field& result) {
+    result = zero;
+    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);
+    }
+  }
+
+ }; 
+}

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);
 
@@ -123,8 +123,11 @@ class ConjugateGradient : public OperatorFunction<Field> {
       p = p * b + r;
 
       LinalgTimer.Stop();
+
       std::cout << GridLogIterative << "ConjugateGradient: 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;
 
       // Stopping condition
       if (cp <= rsq) {
@@ -132,8 +135,6 @@ class ConjugateGradient : public OperatorFunction<Field> {
         Linop.HermOpAndNorm(psi, mmp, d, qq);
         p = mmp - src;
 
-        RealD mmpnorm = sqrt(norm2(mmp));
-        RealD psinorm = sqrt(norm2(psi));
         RealD srcnorm = sqrt(norm2(src));
         RealD resnorm = sqrt(norm2(p));
         RealD true_residual = resnorm / srcnorm;
@@ -157,8 +158,10 @@ class ConjugateGradient : public OperatorFunction<Field> {
     }
     std::cout << GridLogMessage << "ConjugateGradient did NOT converge"
               << std::endl;
+
     if (ErrorOnNoConverge) assert(0);
     IterationsToComplete = k;
+
   }
 };
 }

lib/algorithms/iterative/ConjugateGradientMixedPrec.h

@@ -60,6 +60,7 @@ namespace Grid {
     }
   
     void operator() (const FieldD &src_d_in, FieldD &sol_d){
+
       TotalInnerIterations = 0;
 	
       GridStopWatch TotalTimer;

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/EigenSort.h

@@ -1,81 +0,0 @@
-    /*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/algorithms/iterative/EigenSort.h
-
-    Copyright (C) 2015
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-    *************************************************************************************/
-    /*  END LEGAL */
-#ifndef GRID_EIGENSORT_H
-#define GRID_EIGENSORT_H
-
-
-namespace Grid {
-    /////////////////////////////////////////////////////////////
-    // Eigen sorter to begin with
-    /////////////////////////////////////////////////////////////
-
-template<class Field>
-class SortEigen {
- private:
-  
-//hacking for testing for now
- 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);
-  }  
-  
-  
- public:
-
-  void push(DenseVector<RealD>& lmd,
-            DenseVector<Field>& evec,int N) {
-    DenseVector<Field> cpy(lmd.size(),evec[0]._grid);
-    for(int i=0;i<lmd.size();i++) cpy[i] = evec[i];
-    
-    DenseVector<std::pair<RealD, Field const*> > emod(lmd.size());    
-    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 DenseVector<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;
-    }
-  }
-  void push(DenseVector<RealD>& lmd,int N) {
-    std::partial_sort(lmd.begin(),lmd.begin()+N,lmd.end(),less_lmd);
-  }
-  bool saturated(RealD lmd, RealD thrs) {
-    return fabs(lmd) > fabs(thrs);
-  }
-};
-
-}
-#endif

lib/algorithms/iterative/SchurRedBlack.h

@@ -7,6 +7,7 @@
     Copyright (C) 2015
 
 Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Chulwoo Jung <chulwoo@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
@@ -53,16 +54,194 @@ 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>;
+  template<class Field> class SchurRedBlackStagSolve {
+  private:
+    OperatorFunction<Field> & _HermitianRBSolver;
+    int CBfactorise;
+  public:
+
+    /////////////////////////////////////////////////////
+    // Wrap the usual normal equations Schur trick
+    /////////////////////////////////////////////////////
+  SchurRedBlackStagSolve(OperatorFunction<Field> &HermitianRBSolver, int cb)  :
+     _HermitianRBSolver(HermitianRBSolver), CBfactorise(cb) {}
+
+    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();
+
+      SchurStagOperator<Matrix,Field> _HermOpEO(_Matrix);
+      int Schur = CBfactorise;
+      int Other = 1 - CBfactorise;
+ 
+      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(Other,src_e,in);
+      pickCheckerboard(Schur ,src_o,in);
+      pickCheckerboard(Other,sol_e,out);
+      pickCheckerboard(Schur ,sol_o,out);
+    
+      /////////////////////////////////////////////////////
+      // src_o = Mdag * (source_o - Moe MeeInv source_e)
+      /////////////////////////////////////////////////////
+      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.checkerboard ==Other);
+      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.checkerboard ==Schur);     
+      tmp=src_o-Mtmp;                  assert(  tmp.checkerboard ==Schur);     
+
+#if 0
+      // get the right MpcDag
+//      _HermOpEO.MpcDag(tmp,src_o);     assert(src_o.checkerboard ==Schur);       
+#else
+      _Matrix.Mooee(tmp,src_o);     assert(src_o.checkerboard ==Schur);
+#endif
+      //////////////////////////////////////////////////////////////
+      // 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==Schur);
+
+      ///////////////////////////////////////////////////
+      // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
+      ///////////////////////////////////////////////////
+      _Matrix.Meooe(sol_o,tmp);        assert(  tmp.checkerboard   ==Other);
+      src_e = src_e-tmp;               assert(  src_e.checkerboard ==Other);
+      _Matrix.MooeeInv(src_e,sol_e);   assert(  sol_e.checkerboard ==Other);
+     
+      setCheckerboard(out,sol_e); assert(  sol_e.checkerboard ==Other);
+      setCheckerboard(out,sol_o); assert(  sol_o.checkerboard ==Schur );
+
+      // Verify the unprec residual
+      _Matrix.M(out,resid); 
+      resid = resid-in;
+      RealD ns = norm2(in);
+      RealD nr = norm2(resid);
+
+      std::cout<<GridLogMessage << "SchurRedBlackStag 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
@@ -76,12 +255,10 @@ namespace Grid {
     /////////////////////////////////////////////////////
     // Wrap the usual normal equations Schur trick
     /////////////////////////////////////////////////////
-  SchurRedBlackDiagMooeeSolve(OperatorFunction<Field> &HermitianRBSolver)  :
-     _HermitianRBSolver(HermitianRBSolver) 
-    { 
-      CBfactorise=0;
-    };
-
+  SchurRedBlackDiagMooeeSolve(OperatorFunction<Field> &HermitianRBSolver,int cb=0)  :  _HermitianRBSolver(HermitianRBSolver) 
+  { 
+    CBfactorise=cb;
+  };
     template<class Matrix>
       void operator() (Matrix & _Matrix,const Field &in, Field &out){
 
@@ -141,5 +318,238 @@ 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;
+    }     
+  };
+
+  template<class Field> class SchurRedBlackStagMixed {
+  private:
+    LinearFunction<Field> & _HermitianRBSolver;
+    int CBfactorise;
+  public:
+
+    /////////////////////////////////////////////////////
+    // Wrap the usual normal equations Schur trick
+    /////////////////////////////////////////////////////
+  SchurRedBlackStagMixed(LinearFunction<Field> &HermitianRBSolver, int cb)  :
+     _HermitianRBSolver(HermitianRBSolver), CBfactorise(cb) {}
+
+    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();
+
+      SchurStagOperator<Matrix,Field> _HermOpEO(_Matrix);
+      int Schur = CBfactorise;
+      int Other = 1 - CBfactorise;
+ 
+      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(Other,src_e,in);
+      pickCheckerboard(Schur ,src_o,in);
+      pickCheckerboard(Other,sol_e,out);
+      pickCheckerboard(Schur ,sol_o,out);
+    
+      /////////////////////////////////////////////////////
+      // src_o = Mdag * (source_o - Moe MeeInv source_e)
+      /////////////////////////////////////////////////////
+      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.checkerboard ==Other);
+      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.checkerboard ==Schur);     
+      tmp=src_o-Mtmp;                  assert(  tmp.checkerboard ==Schur);     
+
+      _Matrix.Mooee(tmp,src_o);     assert(src_o.checkerboard ==Schur);
+      //////////////////////////////////////////////////////////////
+      // 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==Schur);
+      _HermitianRBSolver(src_o,sol_o);  assert(sol_o.checkerboard==Other);
+
+      ///////////////////////////////////////////////////
+      // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
+      ///////////////////////////////////////////////////
+      _Matrix.Meooe(sol_o,tmp);        assert(  tmp.checkerboard   ==Other);
+      src_e = src_e-tmp;               assert(  src_e.checkerboard ==Other);
+      _Matrix.MooeeInv(src_e,sol_e);   assert(  sol_e.checkerboard ==Other);
+     
+      setCheckerboard(out,sol_e); assert(  sol_e.checkerboard ==Other);
+      setCheckerboard(out,sol_o); assert(  sol_o.checkerboard ==Schur );
+
+      // Verify the unprec residual
+      _Matrix.M(out,resid); 
+      resid = resid-in;
+      RealD ns = norm2(in);
+      RealD nr = norm2(resid);
+
+      std::cout<<GridLogMessage << "SchurRedBlackStag solver true unprec resid "<< std::sqrt(nr/ns) <<" nr "<< nr <<" ns "<<ns << std::endl;
+    }     
+  };
+
 }
 #endif

lib/algorithms/iterative/SimpleLanczos.h

@@ -0,0 +1,933 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/ImplicitlyRestartedLanczos.h
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Chulwoo Jung <chulwoo@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
+    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_LANC_H
+#define GRID_LANC_H
+
+#include <string.h>		//memset
+
+#ifdef USE_LAPACK
+#ifdef USE_MKL
+#include<mkl_lapack.h>
+#else
+void LAPACK_dstegr (char *jobz, char *range, int *n, double *d, double *e,
+		    double *vl, double *vu, int *il, int *iu, double *abstol,
+		    int *m, double *w, double *z, int *ldz, int *isuppz,
+		    double *work, int *lwork, int *iwork, int *liwork,
+		    int *info);
+//#include <lapacke/lapacke.h>
+#endif
+#endif
+
+#include <Grid/algorithms/densematrix/DenseMatrix.h>
+//#include <Grid/algorithms/iterative/EigenSort.h>
+
+// eliminate temorary vector in calc()
+#define MEM_SAVE
+
+namespace Grid
+{
+
+  struct Bisection
+  {
+
+#if 0
+    static void get_eig2 (int row_num, std::vector < RealD > &ALPHA,
+			  std::vector < RealD > &BETA,
+			  std::vector < RealD > &eig)
+    {
+      int i, j;
+        std::vector < RealD > evec1 (row_num + 3);
+        std::vector < RealD > evec2 (row_num + 3);
+      RealD eps2;
+        ALPHA[1] = 0.;
+        BETHA[1] = 0.;
+      for (i = 0; i < row_num - 1; i++)
+	{
+	  ALPHA[i + 1] = A[i * (row_num + 1)].real ();
+	  BETHA[i + 2] = A[i * (row_num + 1) + 1].real ();
+	}
+      ALPHA[row_num] = A[(row_num - 1) * (row_num + 1)].real ();
+        bisec (ALPHA, BETHA, row_num, 1, row_num, 1e-10, 1e-10, evec1, eps2);
+        bisec (ALPHA, BETHA, row_num, 1, row_num, 1e-16, 1e-16, evec2, eps2);
+
+      // Do we really need to sort here?
+      int begin = 1;
+      int end = row_num;
+      int swapped = 1;
+      while (swapped)
+	{
+	  swapped = 0;
+	  for (i = begin; i < end; i++)
+	    {
+	      if (mag (evec2[i]) > mag (evec2[i + 1]))
+		{
+		  swap (evec2 + i, evec2 + i + 1);
+		  swapped = 1;
+		}
+	    }
+	  end--;
+	  for (i = end - 1; i >= begin; i--)
+	    {
+	      if (mag (evec2[i]) > mag (evec2[i + 1]))
+		{
+		  swap (evec2 + i, evec2 + i + 1);
+		  swapped = 1;
+		}
+	    }
+	  begin++;
+	}
+
+      for (i = 0; i < row_num; i++)
+	{
+	  for (j = 0; j < row_num; j++)
+	    {
+	      if (i == j)
+		H[i * row_num + j] = evec2[i + 1];
+	      else
+		H[i * row_num + j] = 0.;
+	    }
+	}
+    }
+#endif
+
+    static void bisec (std::vector < RealD > &c,
+		       std::vector < RealD > &b,
+		       int n,
+		       int m1,
+		       int m2,
+		       RealD eps1,
+		       RealD relfeh, std::vector < RealD > &x, RealD & eps2)
+    {
+      std::vector < RealD > wu (n + 2);
+
+      RealD h, q, x1, xu, x0, xmin, xmax;
+      int i, a, k;
+
+      b[1] = 0.0;
+      xmin = c[n] - fabs (b[n]);
+      xmax = c[n] + fabs (b[n]);
+      for (i = 1; i < n; i++)
+	{
+	  h = fabs (b[i]) + fabs (b[i + 1]);
+	  if (c[i] + h > xmax)
+	    xmax = c[i] + h;
+	  if (c[i] - h < xmin)
+	    xmin = c[i] - h;
+	}
+      xmax *= 2.;
+
+      eps2 = relfeh * ((xmin + xmax) > 0.0 ? xmax : -xmin);
+      if (eps1 <= 0.0)
+	eps1 = eps2;
+      eps2 = 0.5 * eps1 + 7.0 * (eps2);
+      x0 = xmax;
+      for (i = m1; i <= m2; i++)
+	{
+	  x[i] = xmax;
+	  wu[i] = xmin;
+	}
+
+      for (k = m2; k >= m1; k--)
+	{
+	  xu = xmin;
+	  i = k;
+	  do
+	    {
+	      if (xu < wu[i])
+		{
+		  xu = wu[i];
+		  i = m1 - 1;
+		}
+	      i--;
+	    }
+	  while (i >= m1);
+	  if (x0 > x[k])
+	    x0 = x[k];
+	  while ((x0 - xu) > 2 * relfeh * (fabs (xu) + fabs (x0)) + eps1)
+	    {
+	      x1 = (xu + x0) / 2;
+
+	      a = 0;
+	      q = 1.0;
+	      for (i = 1; i <= n; i++)
+		{
+		  q =
+		    c[i] - x1 -
+		    ((q != 0.0) ? b[i] * b[i] / q : fabs (b[i]) / relfeh);
+		  if (q < 0)
+		    a++;
+		}
+//      printf("x1=%0.14e a=%d\n",x1,a);
+	      if (a < k)
+		{
+		  if (a < m1)
+		    {
+		      xu = x1;
+		      wu[m1] = x1;
+		    }
+		  else
+		    {
+		      xu = x1;
+		      wu[a + 1] = x1;
+		      if (x[a] > x1)
+			x[a] = x1;
+		    }
+		}
+	      else
+		x0 = x1;
+	    }
+	  printf ("x0=%0.14e xu=%0.14e k=%d\n", x0, xu, k);
+	  x[k] = (x0 + xu) / 2;
+	}
+    }
+  };
+
+/////////////////////////////////////////////////////////////
+// Implicitly restarted lanczos
+/////////////////////////////////////////////////////////////
+
+
+  template < class Field > class SimpleLanczos
+  {
+
+    const RealD small = 1.0e-16;
+  public:
+    int lock;
+    int get;
+    int Niter;
+    int converged;
+
+    int Nstop;			// Number of evecs checked for convergence
+    int Nk;			// Number of converged sought
+    int Np;			// Np -- Number of spare vecs in kryloc space
+    int Nm;			// Nm -- total number of vectors
+
+
+    RealD OrthoTime;
+
+    RealD eresid;
+
+    SortEigen < Field > _sort;
+
+    LinearOperatorBase < Field > &_Linop;
+
+    OperatorFunction < Field > &_poly;
+
+    /////////////////////////
+    // Constructor
+    /////////////////////////
+    void init (void)
+    {
+    };
+    void Abort (int ff, DenseVector < RealD > &evals,
+		DenseVector < DenseVector < RealD > >&evecs);
+
+    SimpleLanczos (LinearOperatorBase < Field > &Linop,	// op
+		   OperatorFunction < Field > &poly,	// polynmial
+		   int _Nstop,	// sought vecs
+		   int _Nk,	// sought vecs
+		   int _Nm,	// spare vecs
+		   RealD _eresid,	// resid in lmdue deficit 
+		   int _Niter):	// Max iterations
+     
+      _Linop (Linop),
+      _poly (poly),
+      Nstop (_Nstop), Nk (_Nk), Nm (_Nm), eresid (_eresid), Niter (_Niter)
+    {
+      Np = Nm - Nk;
+      assert (Np > 0);
+    };
+
+    /////////////////////////
+    // Sanity checked this routine (step) against Saad.
+    /////////////////////////
+    void RitzMatrix (DenseVector < Field > &evec, int k)
+    {
+
+      if (1)
+	return;
+
+      GridBase *grid = evec[0]._grid;
+      Field w (grid);
+      std::cout << GridLogMessage << "RitzMatrix " << std::endl;
+      for (int i = 0; i < k; i++)
+	{
+	  _Linop.HermOp (evec[i], w);
+//      _poly(_Linop,evec[i],w);
+	  std::cout << GridLogMessage << "[" << i << "] ";
+	  for (int j = 0; j < k; j++)
+	    {
+	      ComplexD in = innerProduct (evec[j], w);
+	      if (fabs ((double) i - j) > 1)
+		{
+		  if (abs (in) > 1.0e-9)
+		    {
+		      std::cout << GridLogMessage << "oops" << std::endl;
+		      abort ();
+		    }
+		  else
+		    std::cout << GridLogMessage << " 0 ";
+		}
+	      else
+		{
+		  std::cout << GridLogMessage << " " << in << " ";
+		}
+	    }
+	  std::cout << GridLogMessage << std::endl;
+	}
+    }
+
+    void step (DenseVector < RealD > &lmd,
+	       DenseVector < RealD > &lme,
+	       Field & last, Field & current, Field & next, uint64_t k)
+    {
+      if (lmd.size () <= k)
+	lmd.resize (k + Nm);
+      if (lme.size () <= k)
+	lme.resize (k + Nm);
+
+
+//      _poly(_Linop,current,next );   // 3. wk:=Avk−βkv_{k−1}
+      _Linop.HermOp (current, next);	// 3. wk:=Avk−βkv_{k−1}
+      if (k > 0)
+	{
+	  next -= lme[k - 1] * last;
+	}
+//      std::cout<<GridLogMessage << "<last|next>" << innerProduct(last,next) <<std::endl;
+
+      ComplexD zalph = innerProduct (current, next);	// 4. αk:=(wk,vk)
+      RealD alph = real (zalph);
+
+      next = next - alph * current;	// 5. wk:=wk−αkvk
+//      std::cout<<GridLogMessage << "<current|next>" << innerProduct(current,next) <<std::endl;
+
+      RealD beta = normalise (next);	// 6. βk+1 := ∥wk∥2. If βk+1 = 0 then Stop
+      // 7. vk+1 := wk/βk+1
+//       norm=beta;
+
+      int interval = Nm / 100 + 1;
+      if ((k % interval) == 0)
+	std::
+	  cout << GridLogMessage << k << " : alpha = " << zalph << " beta " <<
+	  beta << std::endl;
+      const RealD tiny = 1.0e-20;
+      if (beta < tiny)
+	{
+	  std::cout << GridLogMessage << " beta is tiny " << beta << std::
+	    endl;
+	}
+      lmd[k] = alph;
+      lme[k] = beta;
+
+    }
+
+    void qr_decomp (DenseVector < RealD > &lmd,
+		    DenseVector < RealD > &lme,
+		    int Nk,
+		    int Nm,
+		    DenseVector < RealD > &Qt, RealD Dsh, int kmin, int kmax)
+    {
+      int k = kmin - 1;
+      RealD x;
+
+      RealD Fden = 1.0 / hypot (lmd[k] - Dsh, lme[k]);
+      RealD c = (lmd[k] - Dsh) * Fden;
+      RealD s = -lme[k] * Fden;
+
+      RealD tmpa1 = lmd[k];
+      RealD tmpa2 = lmd[k + 1];
+      RealD tmpb = lme[k];
+
+      lmd[k] = c * c * tmpa1 + s * s * tmpa2 - 2.0 * c * s * tmpb;
+      lmd[k + 1] = s * s * tmpa1 + c * c * tmpa2 + 2.0 * c * s * tmpb;
+      lme[k] = c * s * (tmpa1 - tmpa2) + (c * c - s * s) * tmpb;
+      x = -s * lme[k + 1];
+      lme[k + 1] = c * lme[k + 1];
+
+      for (int i = 0; i < Nk; ++i)
+	{
+	  RealD Qtmp1 = Qt[i + Nm * k];
+	  RealD Qtmp2 = Qt[i + Nm * (k + 1)];
+	  Qt[i + Nm * k] = c * Qtmp1 - s * Qtmp2;
+	  Qt[i + Nm * (k + 1)] = s * Qtmp1 + c * Qtmp2;
+	}
+
+      // Givens transformations
+      for (int k = kmin; k < kmax - 1; ++k)
+	{
+
+	  RealD Fden = 1.0 / hypot (x, lme[k - 1]);
+	  RealD c = lme[k - 1] * Fden;
+	  RealD s = -x * Fden;
+
+	  RealD tmpa1 = lmd[k];
+	  RealD tmpa2 = lmd[k + 1];
+	  RealD tmpb = lme[k];
+
+	  lmd[k] = c * c * tmpa1 + s * s * tmpa2 - 2.0 * c * s * tmpb;
+	  lmd[k + 1] = s * s * tmpa1 + c * c * tmpa2 + 2.0 * c * s * tmpb;
+	  lme[k] = c * s * (tmpa1 - tmpa2) + (c * c - s * s) * tmpb;
+	  lme[k - 1] = c * lme[k - 1] - s * x;
+
+	  if (k != kmax - 2)
+	    {
+	      x = -s * lme[k + 1];
+	      lme[k + 1] = c * lme[k + 1];
+	    }
+
+	  for (int i = 0; i < Nk; ++i)
+	    {
+	      RealD Qtmp1 = Qt[i + Nm * k];
+	      RealD Qtmp2 = Qt[i + Nm * (k + 1)];
+	      Qt[i + Nm * k] = c * Qtmp1 - s * Qtmp2;
+	      Qt[i + Nm * (k + 1)] = s * Qtmp1 + c * Qtmp2;
+	    }
+	}
+    }
+
+#ifdef USE_LAPACK
+#ifdef USE_MKL
+#define LAPACK_INT MKL_INT
+#else
+#define LAPACK_INT long long
+#endif
+    void diagonalize_lapack (DenseVector < RealD > &lmd, DenseVector < RealD > &lme, int N1,	// all
+			     int N2,	// get
+			     GridBase * grid)
+    {
+      const int size = Nm;
+      LAPACK_INT NN = N1;
+      double evals_tmp[NN];
+      double DD[NN];
+      double EE[NN];
+      for (int i = 0; i < NN; i++)
+	for (int j = i - 1; j <= i + 1; j++)
+	  if (j < NN && j >= 0)
+	    {
+	      if (i == j)
+		DD[i] = lmd[i];
+	      if (i == j)
+		evals_tmp[i] = lmd[i];
+	      if (j == (i - 1))
+		EE[j] = lme[j];
+	    }
+      LAPACK_INT evals_found;
+      LAPACK_INT lwork =
+	((18 * NN) >
+	 (1 + 4 * NN + NN * NN) ? (18 * NN) : (1 + 4 * NN + NN * NN));
+      LAPACK_INT liwork = 3 + NN * 10;
+      LAPACK_INT iwork[liwork];
+      double work[lwork];
+      LAPACK_INT isuppz[2 * NN];
+      char jobz = 'N';		// calculate evals only
+      char range = 'I';		// calculate il-th to iu-th evals
+      //    char range = 'A'; // calculate all evals
+      char uplo = 'U';		// refer to upper half of original matrix
+      char compz = 'I';		// Compute eigenvectors of tridiagonal matrix
+      int ifail[NN];
+      LAPACK_INT info;
+//  int total = QMP_get_number_of_nodes();
+//  int node = QMP_get_node_number();
+//  GridBase *grid = evec[0]._grid;
+      int total = grid->_Nprocessors;
+      int node = grid->_processor;
+      int interval = (NN / total) + 1;
+      double vl = 0.0, vu = 0.0;
+      LAPACK_INT il = interval * node + 1, iu = interval * (node + 1);
+      if (iu > NN)
+	iu = NN;
+      double tol = 0.0;
+      if (1)
+	{
+	  memset (evals_tmp, 0, sizeof (double) * NN);
+	  if (il <= NN)
+	    {
+	      printf ("total=%d node=%d il=%d iu=%d\n", total, node, il, iu);
+#ifdef USE_MKL
+	      dstegr (&jobz, &range, &NN,
+#else
+	      LAPACK_dstegr (&jobz, &range, &NN,
+#endif
+			     (double *) DD, (double *) EE, &vl, &vu, &il, &iu,	// these four are ignored if second parameteris 'A'
+			     &tol,	// tolerance
+			     &evals_found, evals_tmp, (double *) NULL, &NN,
+			     isuppz, work, &lwork, iwork, &liwork, &info);
+	      for (int i = iu - 1; i >= il - 1; i--)
+		{
+		  printf ("node=%d evals_found=%d evals_tmp[%d] = %g\n", node,
+			  evals_found, i - (il - 1), evals_tmp[i - (il - 1)]);
+		  evals_tmp[i] = evals_tmp[i - (il - 1)];
+		  if (il > 1)
+		    evals_tmp[i - (il - 1)] = 0.;
+		}
+	    }
+	  {
+	    grid->GlobalSumVector (evals_tmp, NN);
+	  }
+	}
+// cheating a bit. It is better to sort instead of just reversing it, but the document of the routine says evals are sorted in increasing order. qr gives evals in decreasing order.
+    }
+#undef LAPACK_INT
+#endif
+
+
+    void diagonalize (DenseVector < RealD > &lmd,
+		      DenseVector < RealD > &lme,
+		      int N2, int N1, GridBase * grid)
+    {
+
+#ifdef USE_LAPACK
+      const int check_lapack = 0;	// just use lapack if 0, check against lapack if 1
+
+      if (!check_lapack)
+	return diagonalize_lapack (lmd, lme, N2, N1, grid);
+
+//      diagonalize_lapack(lmd2,lme2,Nm2,Nm,Qt,grid);
+#endif
+    }
+
+#if 1
+    static RealD normalise (Field & v)
+    {
+      RealD nn = norm2 (v);
+      nn = sqrt (nn);
+      v = v * (1.0 / nn);
+      return nn;
+    }
+
+    void orthogonalize (Field & w, DenseVector < Field > &evec, int k)
+    {
+      double t0 = -usecond () / 1e6;
+      typedef typename Field::scalar_type MyComplex;
+      MyComplex ip;
+
+      if (0)
+	{
+	  for (int j = 0; j < k; ++j)
+	    {
+	      normalise (evec[j]);
+	      for (int i = 0; i < j; i++)
+		{
+		  ip = innerProduct (evec[i], evec[j]);	// are the evecs normalised? ; this assumes so.
+		  evec[j] = evec[j] - ip * evec[i];
+		}
+	    }
+	}
+
+      for (int j = 0; j < k; ++j)
+	{
+	  ip = innerProduct (evec[j], w);	// are the evecs normalised? ; this assumes so.
+	  w = w - ip * evec[j];
+	}
+      normalise (w);
+      t0 += usecond () / 1e6;
+      OrthoTime += t0;
+    }
+
+    void setUnit_Qt (int Nm, DenseVector < RealD > &Qt)
+    {
+      for (int i = 0; i < Qt.size (); ++i)
+	Qt[i] = 0.0;
+      for (int k = 0; k < Nm; ++k)
+	Qt[k + k * Nm] = 1.0;
+    }
+
+
+    void calc (DenseVector < RealD > &eval, const Field & src, int &Nconv)
+    {
+
+      GridBase *grid = src._grid;
+//      assert(grid == src._grid);
+
+      std::
+	cout << GridLogMessage << " -- Nk = " << Nk << " Np = " << Np << std::
+	endl;
+      std::cout << GridLogMessage << " -- Nm = " << Nm << std::endl;
+      std::cout << GridLogMessage << " -- size of eval   = " << eval.
+	size () << std::endl;
+
+//      assert(c.size() && Nm == eval.size());
+
+      DenseVector < RealD > lme (Nm);
+      DenseVector < RealD > lmd (Nm);
+
+
+      Field current (grid);
+      Field last (grid);
+      Field next (grid);
+
+      Nconv = 0;
+
+      RealD beta_k;
+
+      // Set initial vector
+      // (uniform vector) Why not src??
+      //      evec[0] = 1.0;
+      current = src;
+      std::cout << GridLogMessage << "norm2(src)= " << norm2 (src) << std::
+	endl;
+      normalise (current);
+      std::
+	cout << GridLogMessage << "norm2(evec[0])= " << norm2 (current) <<
+	std::endl;
+
+      // Initial Nk steps
+      OrthoTime = 0.;
+      double t0 = usecond () / 1e6;
+      RealD norm;		// sqrt norm of last vector
+
+      uint64_t iter = 0;
+
+      bool initted = false;
+      std::vector < RealD > low (Nstop * 10);
+      std::vector < RealD > high (Nstop * 10);
+      RealD cont = 0.;
+      while (1) {
+	  cont = 0.;
+	  std::vector < RealD > lme2 (Nm);
+	  std::vector < RealD > lmd2 (Nm);
+	  for (uint64_t k = 0; k < Nm; ++k, iter++) {
+	      step (lmd, lme, last, current, next, iter);
+	      last = current;
+	      current = next;
+	    }
+	  double t1 = usecond () / 1e6;
+	  std::cout << GridLogMessage << "IRL::Initial steps: " << t1 -
+	    t0 << "seconds" << std::endl;
+	  t0 = t1;
+	  std::
+	    cout << GridLogMessage << "IRL::Initial steps:OrthoTime " <<
+	    OrthoTime << "seconds" << std::endl;
+
+	  // getting eigenvalues
+	  lmd2.resize (iter + 2);
+	  lme2.resize (iter + 2);
+	  for (uint64_t k = 0; k < iter; ++k) {
+	      lmd2[k + 1] = lmd[k];
+	      lme2[k + 2] = lme[k];
+	    }
+	  t1 = usecond () / 1e6;
+	  std::cout << GridLogMessage << "IRL:: copy: " << t1 -
+	    t0 << "seconds" << std::endl;
+	  t0 = t1;
+	  {
+	    int total = grid->_Nprocessors;
+	    int node = grid->_processor;
+	    int interval = (Nstop / total) + 1;
+	    int iu = (iter + 1) - (interval * node + 1);
+	    int il = (iter + 1) - (interval * (node + 1));
+	    std::vector < RealD > eval2 (iter + 3);
+	    RealD eps2;
+	    Bisection::bisec (lmd2, lme2, iter, il, iu, 1e-16, 1e-10, eval2,
+			      eps2);
+//        diagonalize(eval2,lme2,iter,Nk,grid);
+	    RealD diff = 0.;
+	    for (int i = il; i <= iu; i++) {
+		if (initted)
+		  diff =
+		    fabs (eval2[i] - high[iu-i]) / (fabs (eval2[i]) +
+						      fabs (high[iu-i]));
+		if (initted && (diff > eresid))
+		  cont = 1.;
+		if (initted)
+		  printf ("eval[%d]=%0.14e %0.14e, %0.14e\n", i, eval2[i],
+			  high[iu-i], diff);
+		high[iu-i] = eval2[i];
+	      }
+	    il = (interval * node + 1);
+	    iu = (interval * (node + 1));
+	    Bisection::bisec (lmd2, lme2, iter, il, iu, 1e-16, 1e-10, eval2,
+			      eps2);
+	    for (int i = il; i <= iu; i++) {
+		if (initted)
+		  diff =
+		    fabs (eval2[i] - low[i]) / (fabs (eval2[i]) +
+						fabs (low[i]));
+		if (initted && (diff > eresid))
+		  cont = 1.;
+		if (initted)
+		  printf ("eval[%d]=%0.14e %0.14e, %0.14e\n", i, eval2[i],
+			  low[i], diff);
+		low[i] = eval2[i];
+	      }
+	    t1 = usecond () / 1e6;
+	    std::cout << GridLogMessage << "IRL:: diagonalize: " << t1 -
+	      t0 << "seconds" << std::endl;
+	    t0 = t1;
+	  }
+
+	  for (uint64_t k = 0; k < Nk; ++k) {
+//          eval[k] = eval2[k];
+	    }
+	  if (initted)
+	    {
+	      grid->GlobalSumVector (&cont, 1);
+	      if (cont < 1.) return;
+	    }
+	  initted = true;
+	}
+
+    }
+
+
+
+
+
+
+/**
+   There is some matrix Q such that for any vector y
+   Q.e_1 = y and Q is unitary.
+**/
+    template < class T >
+      static T orthQ (DenseMatrix < T > &Q, DenseVector < T > y)
+    {
+      int N = y.size ();	//Matrix Size
+      Fill (Q, 0.0);
+      T tau;
+      for (int i = 0; i < N; i++)
+	{
+	  Q[i][0] = y[i];
+	}
+      T sig = conj (y[0]) * y[0];
+      T tau0 = fabs (sqrt (sig));
+
+      for (int j = 1; j < N; j++)
+	{
+	  sig += conj (y[j]) * y[j];
+	  tau = abs (sqrt (sig));
+
+	  if (abs (tau0) > 0.0)
+	    {
+
+	      T gam = conj ((y[j] / tau) / tau0);
+	      for (int k = 0; k <= j - 1; k++)
+		{
+		  Q[k][j] = -gam * y[k];
+		}
+	      Q[j][j] = tau0 / tau;
+	    }
+	  else
+	    {
+	      Q[j - 1][j] = 1.0;
+	    }
+	  tau0 = tau;
+	}
+      return tau;
+    }
+
+/**
+	There is some matrix Q such that for any vector y
+	Q.e_k = y and Q is unitary.
+**/
+    template < class T >
+      static T orthU (DenseMatrix < T > &Q, DenseVector < T > y)
+    {
+      T tau = orthQ (Q, y);
+      SL (Q);
+      return tau;
+    }
+
+
+/**
+	Wind up with a matrix with the first con rows untouched
+
+say con = 2
+	Q is such that Qdag H Q has {x, x, val, 0, 0, 0, 0, ...} as 1st colum
+	and the matrix is upper hessenberg
+	and with f and Q appropriately modidied with Q is the arnoldi factorization
+
+**/
+
+    template < class T > static void Lock (DenseMatrix < T > &H,	///Hess mtx     
+					   DenseMatrix < T > &Q,	///Lock Transform
+					   T val,	///value to be locked
+					   int con,	///number already locked
+					   RealD small, int dfg, bool herm)
+    {
+      //ForceTridiagonal(H);
+
+      int M = H.dim;
+      DenseVector < T > vec;
+      Resize (vec, M - con);
+
+      DenseMatrix < T > AH;
+      Resize (AH, M - con, M - con);
+      AH = GetSubMtx (H, con, M, con, M);
+
+      DenseMatrix < T > QQ;
+      Resize (QQ, M - con, M - con);
+
+      Unity (Q);
+      Unity (QQ);
+
+      DenseVector < T > evals;
+      Resize (evals, M - con);
+      DenseMatrix < T > evecs;
+      Resize (evecs, M - con, M - con);
+
+      Wilkinson < T > (AH, evals, evecs, small);
+
+      int k = 0;
+      RealD cold = abs (val - evals[k]);
+      for (int i = 1; i < M - con; i++)
+	{
+	  RealD cnew = abs (val - evals[i]);
+	  if (cnew < cold)
+	    {
+	      k = i;
+	      cold = cnew;
+	    }
+	}
+      vec = evecs[k];
+
+      ComplexD tau;
+      orthQ (QQ, vec);
+      //orthQM(QQ,AH,vec);
+
+      AH = Hermitian (QQ) * AH;
+      AH = AH * QQ;
+
+      for (int i = con; i < M; i++)
+	{
+	  for (int j = con; j < M; j++)
+	    {
+	      Q[i][j] = QQ[i - con][j - con];
+	      H[i][j] = AH[i - con][j - con];
+	    }
+	}
+
+      for (int j = M - 1; j > con + 2; j--)
+	{
+
+	  DenseMatrix < T > U;
+	  Resize (U, j - 1 - con, j - 1 - con);
+	  DenseVector < T > z;
+	  Resize (z, j - 1 - con);
+	  T nm = norm (z);
+	  for (int k = con + 0; k < j - 1; k++)
+	    {
+	      z[k - con] = conj (H (j, k + 1));
+	    }
+	  normalise (z);
+
+	  RealD tmp = 0;
+	  for (int i = 0; i < z.size () - 1; i++)
+	    {
+	      tmp = tmp + abs (z[i]);
+	    }
+
+	  if (tmp < small / ((RealD) z.size () - 1.0))
+	    {
+	      continue;
+	    }
+
+	  tau = orthU (U, z);
+
+	  DenseMatrix < T > Hb;
+	  Resize (Hb, j - 1 - con, M);
+
+	  for (int a = 0; a < M; a++)
+	    {
+	      for (int b = 0; b < j - 1 - con; b++)
+		{
+		  T sum = 0;
+		  for (int c = 0; c < j - 1 - con; c++)
+		    {
+		      sum += H[a][con + 1 + c] * U[c][b];
+		    }		//sum += H(a,con+1+c)*U(c,b);}
+		  Hb[b][a] = sum;
+		}
+	    }
+
+	  for (int k = con + 1; k < j; k++)
+	    {
+	      for (int l = 0; l < M; l++)
+		{
+		  H[l][k] = Hb[k - 1 - con][l];
+		}
+	    }			//H(Hb[k-1-con][l] , l,k);}}
+
+	  DenseMatrix < T > Qb;
+	  Resize (Qb, M, M);
+
+	  for (int a = 0; a < M; a++)
+	    {
+	      for (int b = 0; b < j - 1 - con; b++)
+		{
+		  T sum = 0;
+		  for (int c = 0; c < j - 1 - con; c++)
+		    {
+		      sum += Q[a][con + 1 + c] * U[c][b];
+		    }		//sum += Q(a,con+1+c)*U(c,b);}
+		  Qb[b][a] = sum;
+		}
+	    }
+
+	  for (int k = con + 1; k < j; k++)
+	    {
+	      for (int l = 0; l < M; l++)
+		{
+		  Q[l][k] = Qb[k - 1 - con][l];
+		}
+	    }			//Q(Qb[k-1-con][l] , l,k);}}
+
+	  DenseMatrix < T > Hc;
+	  Resize (Hc, M, M);
+
+	  for (int a = 0; a < j - 1 - con; a++)
+	    {
+	      for (int b = 0; b < M; b++)
+		{
+		  T sum = 0;
+		  for (int c = 0; c < j - 1 - con; c++)
+		    {
+		      sum += conj (U[c][a]) * H[con + 1 + c][b];
+		    }		//sum += conj( U(c,a) )*H(con+1+c,b);}
+		  Hc[b][a] = sum;
+		}
+	    }
+
+	  for (int k = 0; k < M; k++)
+	    {
+	      for (int l = con + 1; l < j; l++)
+		{
+		  H[l][k] = Hc[k][l - 1 - con];
+		}
+	    }			//H(Hc[k][l-1-con] , l,k);}}
+
+	}
+    }
+#endif
+
+
+  };
+
+}
+#endif

lib/algorithms/iterative/bisec.c

@@ -0,0 +1,122 @@
+#include <math.h>
+#include <stdlib.h>
+#include <vector>
+
+struct Bisection {
+
+static void get_eig2(int row_num,std::vector<RealD> &ALPHA,std::vector<RealD> &BETA, std::vector<RealD> & eig)
+{
+  int i,j;
+  std::vector<RealD> evec1(row_num+3);
+  std::vector<RealD> evec2(row_num+3);
+  RealD eps2;
+  ALPHA[1]=0.;
+  BETHA[1]=0.;
+  for(i=0;i<row_num-1;i++) {
+    ALPHA[i+1] = A[i*(row_num+1)].real();
+    BETHA[i+2] = A[i*(row_num+1)+1].real();
+  }
+  ALPHA[row_num] = A[(row_num-1)*(row_num+1)].real();
+  bisec(ALPHA,BETHA,row_num,1,row_num,1e-10,1e-10,evec1,eps2);
+  bisec(ALPHA,BETHA,row_num,1,row_num,1e-16,1e-16,evec2,eps2);
+
+  // Do we really need to sort here?
+  int begin=1;
+  int end = row_num;
+  int swapped=1;
+  while(swapped) {
+    swapped=0;
+    for(i=begin;i<end;i++){
+      if(mag(evec2[i])>mag(evec2[i+1]))	{
+	swap(evec2+i,evec2+i+1);
+	swapped=1;
+      }
+    }
+    end--;
+    for(i=end-1;i>=begin;i--){
+      if(mag(evec2[i])>mag(evec2[i+1]))	{
+	swap(evec2+i,evec2+i+1);
+	swapped=1;
+      }
+    }
+    begin++;
+  }
+
+  for(i=0;i<row_num;i++){
+    for(j=0;j<row_num;j++) {
+      if(i==j) H[i*row_num+j]=evec2[i+1];
+      else H[i*row_num+j]=0.;
+    }
+  }
+}
+
+static void bisec(std::vector<RealD> &c,   
+		  std::vector<RealD> &b,
+		  int n,
+		  int m1,
+		  int m2,
+		  RealD eps1,
+		  RealD relfeh,
+		  std::vector<RealD> &x,
+		  RealD &eps2)
+{
+  std::vector<RealD> wu(n+2);
+
+  RealD h,q,x1,xu,x0,xmin,xmax; 
+  int i,a,k;
+
+  b[1]=0.0;
+  xmin=c[n]-fabs(b[n]);
+  xmax=c[n]+fabs(b[n]);
+  for(i=1;i<n;i++){
+    h=fabs(b[i])+fabs(b[i+1]);
+    if(c[i]+h>xmax) xmax= c[i]+h;
+    if(c[i]-h<xmin) xmin= c[i]-h;
+  }
+  xmax *=2.;
+
+  eps2=relfeh*((xmin+xmax)>0.0 ? xmax : -xmin);
+  if(eps1<=0.0) eps1=eps2;
+  eps2=0.5*eps1+7.0*(eps2);
+  x0=xmax;
+  for(i=m1;i<=m2;i++){
+    x[i]=xmax;
+    wu[i]=xmin;
+  }
+
+  for(k=m2;k>=m1;k--){
+    xu=xmin;
+    i=k;
+    do{
+      if(xu<wu[i]){
+	xu=wu[i];
+	i=m1-1;
+      }
+      i--;
+    }while(i>=m1);
+    if(x0>x[k]) x0=x[k];
+    while((x0-xu)>2*relfeh*(fabs(xu)+fabs(x0))+eps1){
+      x1=(xu+x0)/2;
+
+      a=0;
+      q=1.0;
+      for(i=1;i<=n;i++){
+	q=c[i]-x1-((q!=0.0)? b[i]*b[i]/q:fabs(b[i])/relfeh);
+	if(q<0) a++;
+      }
+      //			printf("x1=%e a=%d\n",x1,a);
+      if(a<k){
+	if(a<m1){
+	  xu=x1;
+	  wu[m1]=x1;
+	}else {
+	  xu=x1;
+	  wu[a+1]=x1;
+	  if(x[a]>x1) x[a]=x1;
+	}
+      }else x0=x1;
+    }
+    x[k]=(x0+xu)/2;
+  }
+}
+}

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);
-    
-#ifdef HAVE_MM_MALLOC_H
-    if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) _mm_malloc(bytes,128);
-#else
-    if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) memalign(128,bytes);
-#endif
+    //    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,GRID_ALLOC_ALIGN);
+#else
+    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

@@ -6,8 +6,9 @@
 
     Copyright (C) 2015
 
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: paboyle <paboyle@ph.ed.ac.uk>
+    Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+    Author: paboyle <paboyle@ph.ed.ac.uk>
+    Author: Guido Cossu <guido.cossu@ed.ac.uk>
 
     This program is free software; you can redistribute it and/or modify
     it under the terms of the GNU General Public License as published by
@@ -48,7 +49,10 @@ public:
     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) : CartesianCommunicator(processor_grid,parent) {};
 
+    virtual ~GridBase() = default;
 
     // Physics Grid information.
     std::vector<int> _simd_layout;// Which dimensions get relayed out over simd lanes.
@@ -62,13 +66,12 @@ public:
     int _isites;
     int _fsites;                  // _isites*_osites = product(dimensions).
     int _gsites;
-    std::vector<int> _slice_block;   // subslice information
+    std::vector<int> _slice_block;// subslice information
     std::vector<int> _slice_stride;
     std::vector<int> _slice_nblock;
 
-    // Might need these at some point
-    //    std::vector<int> _lstart;     // local start of array in gcoors. _processor_coor[d]*_ldimensions[d]
-    //    std::vector<int> _lend;       // local end of array in gcoors    _processor_coor[d]*_ldimensions[d]+_ldimensions_[d]-1
+    std::vector<int> _lstart;     // local start of array in gcoors _processor_coor[d]*_ldimensions[d]
+    std::vector<int> _lend  ;     // local end of array in gcoors   _processor_coor[d]*_ldimensions[d]+_ldimensions_[d]-1
 
 public:
 
@@ -99,7 +102,7 @@ public:
     virtual int oIndex(std::vector<int> &coor)
     {
         int idx=0;
-	// Works with either global or local coordinates
+        // Works with either global or local coordinates
         for(int d=0;d<_ndimension;d++) idx+=_ostride[d]*(coor[d]%_rdimensions[d]);
         return idx;
     }
@@ -121,6 +124,12 @@ public:
       Lexicographic::CoorFromIndex(coor,Oindex,_rdimensions);
     }
 
+    inline void InOutCoorToLocalCoor (std::vector<int> &ocoor, std::vector<int> &icoor, std::vector<int> &lcoor) {
+      lcoor.resize(_ndimension);
+      for (int d = 0; d < _ndimension; d++)
+        lcoor[d] = ocoor[d] + _rdimensions[d] * icoor[d];
+    }
+
     //////////////////////////////////////////////////////////
     // SIMD lane addressing
     //////////////////////////////////////////////////////////
@@ -128,6 +137,7 @@ public:
     {
       Lexicographic::CoorFromIndex(coor,lane,_simd_layout);
     }
+
     inline int PermuteDim(int dimension){
       return _simd_layout[dimension]>1;
     }
@@ -145,15 +155,15 @@ public:
       // Distance should be either 0,1,2..
       //
       if ( _simd_layout[dimension] > 2 ) { 
-	for(int d=0;d<_ndimension;d++){
-	  if ( d != dimension ) assert ( (_simd_layout[d]==1)  );
-	}
-	permute_type = RotateBit; // How to specify distance; this is not just direction.
-	return permute_type;
+        for(int d=0;d<_ndimension;d++){
+          if ( d != dimension ) assert ( (_simd_layout[d]==1)  );
+        }
+        permute_type = RotateBit; // How to specify distance; this is not just direction.
+        return permute_type;
       }
 
       for(int d=_ndimension-1;d>dimension;d--){
-	if (_simd_layout[d]>1 ) permute_type++;
+        if (_simd_layout[d]>1 ) permute_type++;
       }
       return permute_type;
     }
@@ -168,11 +178,31 @@ public:
     inline int gSites(void) const { return _isites*_osites*_Nprocessors; }; 
     inline int Nd    (void) const { return _ndimension;};
 
+    inline const std::vector<int> LocalStarts(void)             { return _lstart;    };
     inline const std::vector<int> &FullDimensions(void)         { return _fdimensions;};
     inline const std::vector<int> &GlobalDimensions(void)       { return _gdimensions;};
     inline const std::vector<int> &LocalDimensions(void)        { return _ldimensions;};
     inline const std::vector<int> &VirtualLocalDimensions(void) { return _ldimensions;};
 
+    ////////////////////////////////////////////////////////////////
+    // Utility to print the full decomposition details 
+    ////////////////////////////////////////////////////////////////
+
+    void show_decomposition(){
+      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;             
+    } 
+
     ////////////////////////////////////////////////////////////////
     // Global addressing
     ////////////////////////////////////////////////////////////////
@@ -188,8 +218,8 @@ public:
       gidx=0;
       int mult=1;
       for(int mu=0;mu<_ndimension;mu++) {
-	gidx+=mult*gcoor[mu];
-	mult*=_gdimensions[mu];
+        gidx+=mult*gcoor[mu];
+        mult*=_gdimensions[mu];
       }
     }
     void GlobalCoorToProcessorCoorLocalCoor(std::vector<int> &pcoor,std::vector<int> &lcoor,const std::vector<int> &gcoor)
@@ -197,9 +227,9 @@ public:
       pcoor.resize(_ndimension);
       lcoor.resize(_ndimension);
       for(int mu=0;mu<_ndimension;mu++){
-	int _fld  = _fdimensions[mu]/_processors[mu];
-	pcoor[mu] = gcoor[mu]/_fld;
-	lcoor[mu] = gcoor[mu]%_fld;
+        int _fld  = _fdimensions[mu]/_processors[mu];
+        pcoor[mu] = gcoor[mu]/_fld;
+        lcoor[mu] = gcoor[mu]%_fld;
       }
     }
     void GlobalCoorToRankIndex(int &rank, int &o_idx, int &i_idx ,const std::vector<int> &gcoor)
@@ -211,9 +241,9 @@ public:
       /*
       std::vector<int> cblcoor(lcoor);
       for(int d=0;d<cblcoor.size();d++){
-	if( this->CheckerBoarded(d) ) {
-	  cblcoor[d] = lcoor[d]/2;
-	}
+        if( this->CheckerBoarded(d) ) {
+          cblcoor[d] = lcoor[d]/2;
+        }
       }
       */
       i_idx= iIndex(lcoor);
@@ -239,7 +269,7 @@ public:
     {
       RankIndexToGlobalCoor(rank,o_idx,i_idx ,fcoor);
       if(CheckerBoarded(0)){
-	fcoor[0] = fcoor[0]*2+cb;
+        fcoor[0] = fcoor[0]*2+cb;
       }
     }
     void ProcessorCoorLocalCoorToGlobalCoor(std::vector<int> &Pcoor,std::vector<int> &Lcoor,std::vector<int> &gcoor)

lib/cartesian/Cartesian_full.h

@@ -61,74 +61,104 @@ 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)
     {
-        ///////////////////////
-        // Grid information
-        ///////////////////////
-        _ndimension = dimensions.size();
-            
-        _fdimensions.resize(_ndimension);
-        _gdimensions.resize(_ndimension);
-        _ldimensions.resize(_ndimension);
-        _rdimensions.resize(_ndimension);
-        _simd_layout.resize(_ndimension);
-            
-        _ostride.resize(_ndimension);
-        _istride.resize(_ndimension);
-            
-        _fsites = _gsites = _osites = _isites = 1;
+      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);
+    }
 
-        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];
+    virtual ~GridCartesian() = default;
 
-	  //FIXME check for exact division
+    void Init(const std::vector<int> &dimensions,
+	      const std::vector<int> &simd_layout,
+	      const std::vector<int> &processor_grid)
+    {
+      ///////////////////////
+      // Grid information
+      ///////////////////////
+      _ndimension = dimensions.size();
 
-	  // Use a reduced simd grid
-	  _ldimensions[d]= _gdimensions[d]/_processors[d];  //local dimensions
-	  _rdimensions[d]= _ldimensions[d]/_simd_layout[d]; //overdecomposition
-	  _osites *= _rdimensions[d];
-	  _isites *= _simd_layout[d];
-                
-	  // Addressing support
-	  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];
-	  }
+      _fdimensions.resize(_ndimension);
+      _gdimensions.resize(_ndimension);
+      _ldimensions.resize(_ndimension);
+      _rdimensions.resize(_ndimension);
+      _simd_layout.resize(_ndimension);
+      _lstart.resize(_ndimension);
+      _lend.resize(_ndimension);
+
+      _ostride.resize(_ndimension);
+      _istride.resize(_ndimension);
+
+      _fsites = _gsites = _osites = _isites = 1;
+
+      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];
+
+        // Use a reduced simd grid
+        _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)
+        {
+          _ostride[d] = 1;
+          _istride[d] = 1;
         }
-        
-        ///////////////////////
-        // subplane information
-        ///////////////////////
-        _slice_block.resize(_ndimension);
-        _slice_stride.resize(_ndimension);
-        _slice_nblock.resize(_ndimension);
-            
-        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];
+        else
+        {
+          _ostride[d] = _ostride[d - 1] * _rdimensions[d - 1];
+          _istride[d] = _istride[d - 1] * _simd_layout[d - 1];
         }
+      }
 
+      ///////////////////////
+      // subplane information
+      ///////////////////////
+      _slice_block.resize(_ndimension);
+      _slice_stride.resize(_ndimension);
+      _slice_nblock.resize(_ndimension);
+
+      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];
+      }
     };
 };
-
-
 }
 #endif

lib/cartesian/Cartesian_red_black.h

@@ -112,151 +112,209 @@ 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,
-	      const std::vector<int> &checker_dim_mask,
-	      int checker_dim)
+              const std::vector<int> &simd_layout,
+              const std::vector<int> &processor_grid,
+              const std::vector<int> &checker_dim_mask,
+              int checker_dim)
     {
-    ///////////////////////
-    // Grid information
-    ///////////////////////
+      ///////////////////////
+      // 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);
       _ldimensions.resize(_ndimension);
       _rdimensions.resize(_ndimension);
       _simd_layout.resize(_ndimension);
-      
+      _lstart.resize(_ndimension);
+      _lend.resize(_ndimension);
+
       _ostride.resize(_ndimension);
       _istride.resize(_ndimension);
-      
+
       _fsites = _gsites = _osites = _isites = 1;
-	
-      _checker_dim_mask=checker_dim_mask;
 
-      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
-	}
-	_ldimensions[d] = _gdimensions[d]/_processors[d];
+      _checker_dim_mask = checker_dim_mask;
 
-	// Use a reduced simd grid
-	_simd_layout[d] = simd_layout[d];
-	_rdimensions[d]= _ldimensions[d]/_simd_layout[d];
-	assert(_rdimensions[d]>0);
+      for (int d = 0; d < _ndimension; d++)
+      {
+        _fdimensions[d] = dimensions[d];
+        _gdimensions[d] = _fdimensions[d];
+        _fsites = _fsites * _fdimensions[d];
+        _gsites = _gsites * _gdimensions[d];
 
-	// 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 (d == _checker_dim)
+        {
+          assert((_gdimensions[d] & 0x1) == 0);
+          _gdimensions[d] = _gdimensions[d] / 2; // Remove a checkerboard
+        }
+        _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;
 
-	_osites *= _rdimensions[d];
-	_isites *= _simd_layout[d];
-        
-	// Addressing support
-	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];
-	}
+        // Use a reduced simd grid
+        _simd_layout[d] = simd_layout[d];
+        _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);
+          }
+        }
 
+        _osites *= _rdimensions[d];
+        _isites *= _simd_layout[d];
+
+        // Addressing support
+        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];
+        }
       }
-            
+
       ////////////////////////////////////////////////////////////////////////////////////////////
       // subplane information
       ////////////////////////////////////////////////////////////////////////////////////////////
       _slice_block.resize(_ndimension);
       _slice_stride.resize(_ndimension);
       _slice_nblock.resize(_ndimension);
-        
-      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];
+
+      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];
       }
 
       ////////////////////////////////////////////////
       // 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;
     };
-        
+
     virtual int iIndex(std::vector<int> &lcoor)
     {
-        int idx=0;
-        for(int d=0;d<_ndimension;d++) {
-	  if( d==_checker_dim ) {
-	    idx+=_istride[d]*(lcoor[d]/(2*_rdimensions[d]));
-	  } else { 
-	    idx+=_istride[d]*(lcoor[d]/_rdimensions[d]);
-	  }
-	}
-        return idx;
+      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
@@ -60,6 +67,7 @@ void CartesianCommunicator::ShmBufferFreeAll(void) {
 /////////////////////////////////
 // Grid information queries
 /////////////////////////////////
+int                      CartesianCommunicator::Dimensions(void)        { return _ndimension; };
 int                      CartesianCommunicator::IsBoss(void)            { return _processor==0; };
 int                      CartesianCommunicator::BossRank(void)          { return 0; };
 int                      CartesianCommunicator::ThisRank(void)          { return _processor; };
@@ -88,24 +96,161 @@ 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)
+
+CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent) 
+{
+  _ndimension = processors.size();
+  
+  //////////////////////////////////////////////////////////////////////////////////////////////////////
+  // 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
+
+  std::vector<int> pcoor(_ndimension,0); 
+  std::vector<int> pdims(_ndimension,1); 
+
+  if(parent._processors.size()==4 && _ndimension==5){
+      for(int i=0;i<4;i++) pcoor[i+1]=parent._processor_coor[i];
+      for(int i=0;i<4;i++) pdims[i+1]=parent._processors[i];
+  } else {
+      assert(_ndimension == parent._ndimension);
+      for(int i=0;i<_ndimension;i++) pcoor[i]=parent._processor_coor[i];
+      for(int i=0;i<_ndimension;i++) pdims[i]=parent._processors[i];
+  }
+
+  for(int d=0;d<_ndimension;d++){
+    ccoor[d] = pcoor[d] % processors[d];
+    scoor[d] = pcoor[d] / processors[d];
+    ssize[d] = pdims[d] / processors[d];
+  }
+  int crank,srank;  // rank within subcomm ; rank of subcomm within blocks of subcomms
+  Lexicographic::IndexFromCoor(ccoor,crank,processors);
+  Lexicographic::IndexFromCoor(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;
+
+    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;
+  }
+
+  //////////////////////////////////////////////////////////////////////////////////////////////////////
+  // 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)
+{
+  //  if ( communicator_base != communicator_world ) {
+  //    std::cout << "Cartesian communicator created with a non-world communicator"<<std::endl;
+  //  }
+  _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],1,&communicator);
+  MPI_Comm_rank(communicator,&_processor);
+  MPI_Cart_coords(communicator,_processor,_ndimension,&_processor_coor[0]);
+
+  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();};
-
-double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
-						       void *xmit,
-						       int xmit_to_rank,
-						       void *recv,
-						       int recv_from_rank,
-						       int bytes)
+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 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;
@@ -119,8 +264,30 @@ 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 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;
+
+  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
   ////////////////////////////////////////////////////////////////////
@@ -117,11 +134,7 @@ 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;
 
@@ -135,11 +148,24 @@ class CartesianCommunicator {
   // Must call in Grid startup
   ////////////////////////////////////////////////
   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);
   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
@@ -148,6 +174,7 @@ class CartesianCommunicator {
   int  RankFromProcessorCoor(std::vector<int> &coor);
   void ProcessorCoorFromRank(int rank,std::vector<int> &coor);
   
+  int                      Dimensions(void)        ;
   int                      IsBoss(void)            ;
   int                      BossRank(void)          ;
   int                      ThisRank(void)          ;
@@ -155,6 +182,7 @@ class CartesianCommunicator {
   const std::vector<int> & ProcessorGrid(void)     ;
   int                      ProcessorCount(void)    ;
   int                      NodeCount(void)    ;
+  int                      RankCount(void)    ;
 
   ////////////////////////////////////////////////////////////////////////////////
   // very VERY rarely (Log, serial RNG) we need world without a grid
@@ -175,6 +203,8 @@ class CartesianCommunicator {
   void GlobalSumVector(ComplexF *c,int N);
   void GlobalSum(ComplexD &c);
   void GlobalSumVector(ComplexD *c,int N);
+  void GlobalXOR(uint32_t &);
+  void GlobalXOR(uint64_t &);
   
   template<class obj> void GlobalSum(obj &o){
     typedef typename obj::scalar_type scalar_type;
@@ -207,14 +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);
+				    void *xmit,
+				    int xmit_to_rank,
+				    void *recv,
+				    int recv_from_rank,
+				    int bytes,int dir);
   
-  void StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall);
+  
+  void StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int i);
   void StencilBarrier(void);
 
   ////////////////////////////////////////////////////////////
@@ -226,6 +263,27 @@ class CartesianCommunicator {
   // Broadcast a buffer and composite larger
   ////////////////////////////////////////////////////////////
   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)
     {

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);
@@ -83,6 +69,14 @@ 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);
@@ -202,6 +196,35 @@ 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];
+
+  CartesianCommunicator Comm(row,*this);
+  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.
@@ -222,5 +245,7 @@ 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 {
 
 ///////////////////////////////////////////////////////////////////////////////////////////////////
@@ -65,6 +66,7 @@ std::vector<int> CartesianCommunicator::MyGroup;
 std::vector<void *> CartesianCommunicator::ShmCommBufs;
 
 int CartesianCommunicator::NodeCount(void)    { return GroupSize;};
+int CartesianCommunicator::RankCount(void)    { return WorldSize;};
 
 
 #undef FORCE_COMMS
@@ -196,7 +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++){
@@ -209,11 +250,39 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
       int fd=shm_open(shm_name,O_RDWR|O_CREAT,0666);
       if ( fd < 0 ) {	perror("failed shm_open");	assert(0);      }
       ftruncate(fd, size);
+      
+      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);
 
-      void * ptr =  mmap(NULL,size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
-      if ( ptr == MAP_FAILED ) {       perror("failed mmap");      assert(0);    }
+      if ( ptr == (void * )MAP_FAILED ) {       perror("failed mmap");      assert(0);    }
       assert(((uint64_t)ptr&0x3F)==0);
-      ShmCommBufs[r] =ptr;
+
+// 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;
       
     }
   }
@@ -235,21 +304,32 @@ 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]);
       }
-      printf("shmid: 0x%x\n", shmids[r]);
     }
   }
   MPI_Barrier(ShmComm);
@@ -370,12 +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.
   ////////////////////////////////////////////////////////////////
@@ -509,6 +604,14 @@ 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);
@@ -590,13 +693,28 @@ void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &lis
   }
 }
 
-double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
-						       void *xmit,
-						       int dest,
-						       void *recv,
-						       int from,
-						       int bytes)
+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 dir)
+{
+  int ncomm  =communicator_halo.size(); 
+  int commdir=dir%ncomm;
+
   MPI_Request xrq;
   MPI_Request rrq;
 
@@ -615,26 +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);
 }