Release candidate

Conflicts:
	lib/qcd/action/fermion/Fermion.h
	lib/qcd/action/fermion/WilsonFermion.cc
	lib/util/Init.cc
	tests/Test_cayley_even_odd_vec.cc

.gitignore

@@ -9,6 +9,7 @@
 ################
 *~
 *#
+*.sublime-*
 
 # Precompiled Headers #
 #######################
@@ -47,7 +48,9 @@ Config.h.in
 config.log
 config.status
 .deps
-*.inc
+Make.inc
+eigen.inc
+Eigen.inc
 
 # http://www.gnu.org/software/autoconf #
 ########################################
@@ -101,4 +104,16 @@ lib/fftw/*
 # libtool macros #
 ##################
 m4/lt*
-m4/libtool.m4
+m4/libtool.m4
+
+# Buck files #
+##############
+.buck*
+buck-out
+BUCK
+make-bin-BUCK.sh
+
+# generated sources #
+#####################
+lib/qcd/spin/gamma-gen/*.h
+lib/qcd/spin/gamma-gen/*.cc

.travis.yml

@@ -102,5 +102,5 @@ script:
     - ../configure --enable-precision=single --enable-simd=SSE4 --enable-comms=mpi-auto
     - make -j4
     - if [[ "$TRAVIS_OS_NAME" == "linux" ]]; then mpirun.openmpi -n 2 ./benchmarks/Benchmark_dwf --threads 1 --mpi 2.1.1.1; fi
-    - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then mpirun -n 2 ./benchmarks/Benchmark_dwf --threads 1 --mpi 2.1.1.1; fi
+
 

Makefile.am

@@ -1,5 +1,12 @@
 # additional include paths necessary to compile the C++ library
-SUBDIRS = lib benchmarks tests
+SUBDIRS = lib benchmarks tests extras
+
+include $(top_srcdir)/doxygen.inc
+
+tests: all
+	$(MAKE) -C tests tests
+
+.PHONY: tests doxygen-run doxygen-doc $(DX_PS_GOAL) $(DX_PDF_GOAL)
 
 AM_CXXFLAGS += -I$(top_builddir)/include
 ACLOCAL_AMFLAGS = -I m4

README

@@ -1,44 +0,0 @@
-This library provides data parallel C++ container classes with internal memory layout
-that is transformed to map efficiently to SIMD architectures. CSHIFT facilities
-are provided, similar to HPF and cmfortran, and user control is given over the mapping of
-array indices to both MPI tasks and SIMD processing elements.
-
-* Identically shaped arrays then be processed with perfect data parallelisation.
-* Such identically shapped arrays are called conformable arrays.
-
-The transformation is based on the observation that Cartesian array processing involves
-identical processing to be performed on different regions of the Cartesian array.
-
-The library will (eventually) both geometrically decompose into MPI tasks and across SIMD lanes.
-
-Data parallel array operations can then be specified with a SINGLE data parallel paradigm, but
-optimally use MPI, OpenMP and SIMD parallelism under the hood. This is a significant simplification
-for most programmers.
-
-The layout transformations are parametrised by the SIMD vector length. This adapts according to the architecture.
-Presently SSE2 (128 bit) AVX, AVX2 (256 bit) and IMCI and AVX512 (512 bit) targets are supported.
-
-These are presented as 
-
-  vRealF, vRealD, vComplexF, vComplexD 
-
-internal vector data types. These may be useful in themselves for other programmers.
-The corresponding scalar types are named
-
-  RealF, RealD, ComplexF, ComplexD
-
-MPI parallelism is UNIMPLEMENTED and for now only OpenMP and SIMD parallelism is present in the library.
-
-   You can give `configure' initial values for configuration parameters
-by setting variables in the command line or in the environment.  Here
-is are examples:
-
-     ./configure CXX=clang++ CXXFLAGS="-std=c++11 -O3 -msse4" --enable-simd=SSE4
-
-     ./configure CXX=clang++ CXXFLAGS="-std=c++11 -O3 -mavx" --enable-simd=AVX1
-
-     ./configure CXX=clang++ CXXFLAGS="-std=c++11 -O3 -mavx2" --enable-simd=AVX2
-
-     ./configure CXX=icpc CXXFLAGS="-std=c++11 -O3 -mmic" --enable-simd=AVX512 --host=none
-     
-     

README

@@ -0,0 +1 @@
+README.md

README.md

@@ -16,11 +16,27 @@
 
 **Data parallel C++ mathematical object library.**
 
-Please send all pull requests to the `develop` branch.
-
 License: GPL v2.
 
-Last update 2016/08/03.
+Last update Nov 2016.
+
+_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
 This library provides data parallel C++ container classes with internal memory layout
@@ -29,7 +45,7 @@ are provided, similar to HPF and cmfortran, and user control is given over the m
 array indices to both MPI tasks and SIMD processing elements.
 
 * Identically shaped arrays then be processed with perfect data parallelisation.
-* Such identically shapped arrays are called conformable arrays.
+* Such identically shaped arrays are called conformable arrays.
 
 The transformation is based on the observation that Cartesian array processing involves
 identical processing to be performed on different regions of the Cartesian array.
@@ -42,7 +58,7 @@ optimally use MPI, OpenMP and SIMD parallelism under the hood. This is a signifi
 for most programmers.
 
 The layout transformations are parametrised by the SIMD vector length. This adapts according to the architecture.
-Presently SSE4 (128 bit) AVX, AVX2 (256 bit) and IMCI and AVX512 (512 bit) targets are supported (ARM NEON and BG/Q QPX on the way).
+Presently SSE4 (128 bit) AVX, AVX2, QPX (256 bit), IMCI, and AVX512 (512 bit) targets are supported (ARM NEON on the way).
 
 These are presented as `vRealF`, `vRealD`, `vComplexF`, and `vComplexD` internal vector data types. These may be useful in themselves for other programmers.
 The corresponding scalar types are named `RealF`, `RealD`, `ComplexF` and `ComplexD`.
@@ -50,7 +66,7 @@ The corresponding scalar types are named `RealF`, `RealD`, `ComplexF` and `Compl
 MPI, OpenMP, and SIMD parallelism are present in the library.
 Please see https://arxiv.org/abs/1512.03487 for more detail.
 
-### Installation
+### Quick start
 First, start by cloning the repository:
 
 ``` bash
@@ -71,12 +87,10 @@ mkdir build; cd build
 ../configure --enable-precision=double --enable-simd=AVX --enable-comms=mpi-auto --prefix=<path>
 ```
 
-where `--enable-precision=` set the default precision (`single` or `double`),
-`--enable-simd=` set the SIMD type (see possible values below), `--enable-
-comms=` set the protocol used for communications (`none`, `mpi`, `mpi-auto` or
-`shmem`), and `<path>` should be replaced by the prefix path where you want to
-install Grid. The `mpi-auto` communication option set `configure` to determine
-automatically how to link to MPI. Other options are available, use `configure
+where `--enable-precision=` set the default precision,
+`--enable-simd=` set the SIMD type, `--enable-
+comms=`, and `<path>` should be replaced by the prefix path where you want to
+install Grid. Other options are detailed in the next section, you can also use `configure
 --help` to display them. Like with any other program using GNU autotool, the
 `CXX`, `CXXFLAGS`, `LDFLAGS`, ... environment variables can be modified to
 customise the build.
@@ -92,25 +106,91 @@ To minimise the build time, only the tests at the root of the `tests` directory
 ``` bash
 make -C tests/<subdir> tests
 ```
+If you want to build all the tests at once just use `make tests`.
+
+### Build configuration options
+
+- `--prefix=<path>`: installation prefix for Grid.
+- `--with-gmp=<path>`: look for GMP in the UNIX prefix `<path>`
+- `--with-mpfr=<path>`: look for MPFR in the UNIX prefix `<path>`
+- `--with-fftw=<path>`: look for FFTW in the UNIX prefix `<path>`
+- `--enable-lapack[=<path>]`: enable LAPACK support in Lanczos eigensolver. A UNIX prefix containing the library can be specified (optional).
+- `--enable-mkl[=<path>]`: use Intel MKL for FFT (and LAPACK if enabled) routines. A UNIX prefix containing the library can be specified (optional).
+- `--enable-numa`: enable NUMA first touch optimisation
+- `--enable-simd=<code>`: setup Grid for the SIMD target `<code>` (default: `GEN`). A list of possible SIMD targets is detailed in a section below.
+- `--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 `).
+- `--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`)
+
+### Possible communication interfaces
+
+The following options can be use with the `--enable-comms=` option to target different communication interfaces:
+
+| `<comm>`       | Description                                                   |
+| -------------- | ------------------------------------------------------------- |
+| `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.  
 
 ### Possible SIMD types
 
 The following options can be use with the `--enable-simd=` option to target different SIMD instruction sets:
 
-| String      | Description                            |
+| `<code>`    | Description                            |
 | ----------- | -------------------------------------- |
 | `GEN`       | generic portable vector code           |
 | `SSE4`      | SSE 4.2 (128 bit)                      |
 | `AVX`       | AVX (256 bit)                          |
-| `AVXFMA4`   | AVX (256 bit) + FMA                    |
+| `AVXFMA`    | AVX (256 bit) + FMA                    |
+| `AVXFMA4`   | AVX (256 bit) + FMA4                   |
 | `AVX2`      | AVX 2 (256 bit)                        |
 | `AVX512`    | AVX 512 bit                            |
-| `AVX512MIC` | AVX 512 bit for Intel MIC architecture |
-| `ICMI`      | Intel ICMI instructions (512 bit)      |
+| `QPX`       | QPX (256 bit)                          |
 
 Alternatively, some CPU codenames can be directly used:
 
-| String      | Description                            |
+| `<code>`    | Description                            |
 | ----------- | -------------------------------------- |
-| `KNC`       | [Intel Knights Corner](http://ark.intel.com/products/codename/57721/Knights-Corner) |
-| `KNL`       | [Intel Knights Landing](http://ark.intel.com/products/codename/48999/Knights-Landing) |
+| `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                            |
+
+#### Notes:
+- We currently support AVX512 only for the Intel compiler. Support for GCC and clang will appear in future versions of Grid when the AVX512 support within GCC and clang will be more advanced.
+- For BG/Q only [bgclang](http://trac.alcf.anl.gov/projects/llvm-bgq) is supported. We do not presently plan to support more compilers for this platform.
+- BG/Q performances are currently rather poor. This is being investigated for future versions.
+- The vector size for the `GEN` target can be specified with the `configure` script option `--enable-gen-simd-width`.
+
+### Build setup for Intel Knights Landing platform
+
+The following configuration is recommended for the Intel Knights Landing platform:
+
+``` bash
+../configure --enable-precision=double\
+             --enable-simd=KNL        \
+             --enable-comms=mpi-auto \
+             --with-gmp=<path>        \
+             --with-mpfr=<path>       \
+             --enable-mkl             \
+             CXX=icpc MPICXX=mpiicpc
+```
+
+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=KNL        \
+             --enable-comms=mpi       \
+             --with-gmp=<path>        \
+             --with-mpfr=<path>       \
+             --enable-mkl             \
+             CXX=CC CC=cc
+```

VERSION

@@ -1,4 +1,6 @@
-Version : 0.5.0
+Version : 0.6.0
 
 - 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

benchmarks/Benchmark_comms.cc

@@ -42,16 +42,15 @@ int main (int argc, char ** argv)
 
   int Nloop=10;
   int nmu=0;
-  for(int mu=0;mu<4;mu++) if (mpi_layout[mu]>1) nmu++;
+  for(int mu=0;mu<Nd;mu++) if (mpi_layout[mu]>1) nmu++;
+
   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;
-
-
-
-  for(int lat=4;lat<=32;lat+=2){
-    for(int Ls=1;Ls<=16;Ls*=2){
+  int maxlat=24;
+  for(int lat=4;lat<=maxlat;lat+=4){
+    for(int Ls=8;Ls<=32;Ls*=2){
 
       std::vector<int> latt_size  ({lat*mpi_layout[0],
       				    lat*mpi_layout[1],
@@ -125,8 +124,8 @@ int main (int argc, char ** argv)
   std::cout<<GridLogMessage << "  L  "<<"\t\t"<<" Ls  "<<"\t\t"<<"bytes"<<"\t\t"<<"MB/s uni"<<"\t\t"<<"MB/s bidi"<<std::endl;
 
 
-  for(int lat=4;lat<=32;lat+=2){
-    for(int Ls=1;Ls<=16;Ls*=2){
+  for(int lat=4;lat<=maxlat;lat+=4){
+    for(int Ls=8;Ls<=32;Ls*=2){
 
       std::vector<int> latt_size  ({lat,lat,lat,lat});
 
@@ -194,128 +193,169 @@ int main (int argc, char ** argv)
     }
   }  
 
-#if 0
 
+  Nloop=10;
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
-  std::cout<<GridLogMessage << "= Benchmarking sequential persistent halo exchange in "<<nmu<<" dimensions"<<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;
 
+  for(int lat=4;lat<=maxlat;lat+=4){
+    for(int Ls=8;Ls<=32;Ls*=2){
 
-  for(int lat=4;lat<=32;lat+=2){
-    for(int Ls=1;Ls<=16;Ls*=2){
-
-      std::vector<int> latt_size  ({lat,lat,lat,lat});
+      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);
 
-      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<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));
+      }
 
       int ncomm;
       int bytes=lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD);
 
+      double start=usecond();
+      for(int i=0;i<Nloop;i++){
 
-      std::vector<CartesianCommunicator::CommsRequest_t> empty;
-      std::vector<std::vector<CartesianCommunicator::CommsRequest_t> > requests_fwd(Nd,empty);
-      std::vector<std::vector<CartesianCommunicator::CommsRequest_t> > requests_bwd(Nd,empty);
+	std::vector<CartesianCommunicator::CommsRequest_t> requests;
 
-      for(int mu=0;mu<4;mu++){
 	ncomm=0;
-	if (mpi_layout[mu]>1 ) {
-	  ncomm++;
-
-	  int comm_proc;
-	  int xmit_to_rank;
-	  int recv_from_rank;
-
-	  comm_proc=1;
-	  Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
-	  Grid.SendToRecvFromInit(requests_fwd[mu],
-				  (void *)&xbuf[mu][0],
-				  xmit_to_rank,
-				  (void *)&rbuf[mu][0],
-				  recv_from_rank,
-				  bytes);
-
-	  comm_proc = mpi_layout[mu]-1;
-	  Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
-	  Grid.SendToRecvFromInit(requests_bwd[mu],
-				  (void *)&xbuf[mu+4][0],
-				  xmit_to_rank,
-				  (void *)&rbuf[mu+4][0],
-				  recv_from_rank,
-				  bytes);
-
-	}
-      }
-
-      {
-	double start=usecond();
-	for(int i=0;i<Nloop;i++){
+	for(int mu=0;mu<4;mu++){
+	
+	  if (mpi_layout[mu]>1 ) {
 	  
-	  for(int mu=0;mu<4;mu++){
+	    ncomm++;
+	    int comm_proc=1;
+	    int xmit_to_rank;
+	    int recv_from_rank;
 	    
-	    if (mpi_layout[mu]>1 ) {
-	      
-	      Grid.SendToRecvFromBegin(requests_fwd[mu]);
-	      Grid.SendToRecvFromComplete(requests_fwd[mu]);
-	      Grid.SendToRecvFromBegin(requests_bwd[mu]);
-	      Grid.SendToRecvFromComplete(requests_bwd[mu]);
-	    }
-	  }
-	  Grid.Barrier();
-	}
+	    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);
 	
-	double stop=usecond();
-	
-	double dbytes    = bytes;
-	double xbytes    = Nloop*dbytes*2.0*ncomm;
-	double rbytes    = xbytes;
-	double bidibytes = xbytes+rbytes;
-	
-	double time = stop-start;
-	
-	std::cout<<GridLogMessage << lat<<"\t\t"<<Ls<<"\t\t"<<bytes<<"\t\t"<<xbytes/time<<"\t\t"<<bidibytes/time<<std::endl;
-
-      }
-
-
-      {
-	double start=usecond();
-	for(int i=0;i<Nloop;i++){
+	    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);
 	  
-	  for(int mu=0;mu<4;mu++){
-	    
-	    if (mpi_layout[mu]>1 ) {
-	      
-	      Grid.SendToRecvFromBegin(requests_fwd[mu]);
-	      Grid.SendToRecvFromBegin(requests_bwd[mu]);
-	      Grid.SendToRecvFromComplete(requests_fwd[mu]);
-	      Grid.SendToRecvFromComplete(requests_bwd[mu]);
-	    }
 	  }
-	  Grid.Barrier();
 	}
-	
-	double stop=usecond();
-	
-	double dbytes    = bytes;
-	double xbytes    = Nloop*dbytes*2.0*ncomm;
-	double rbytes    = xbytes;
-	double bidibytes = xbytes+rbytes;
-	
-	double time = stop-start;
-	
-	std::cout<<GridLogMessage << lat<<"\t\t"<<Ls<<"\t\t"<<bytes<<"\t\t"<<xbytes/time<<"\t\t"<<bidibytes/time<<std::endl;
+	Grid.StencilSendToRecvFromComplete(requests);
+	Grid.Barrier();
 
       }
+      double stop=usecond();
 
+      double dbytes    = bytes;
+      double xbytes    = Nloop*dbytes*2.0*ncomm;
+      double rbytes    = xbytes;
+      double bidibytes = xbytes+rbytes;
+
+      double time = stop-start; // microseconds
+
+      std::cout<<GridLogMessage << lat<<"\t\t"<<Ls<<"\t\t"<<bytes<<"\t\t"<<xbytes/time<<"\t\t"<<bidibytes/time<<std::endl;
     }
-  }
+  }    
 
-#endif
+
+
+  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;
+
+  for(int lat=4;lat<=maxlat;lat+=4){
+    for(int Ls=8;Ls<=32;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);
+
+      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));
+      }
+
+      int ncomm;
+      int bytes=lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD);
+
+      double start=usecond();
+      for(int i=0;i<Nloop;i++){
+
+	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);
+	    Grid.StencilSendToRecvFromComplete(requests);
+	    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);
+	    requests.resize(0);
+	  
+	  }
+	}
+	Grid.Barrier();
+
+      }
+      double stop=usecond();
+
+      double dbytes    = bytes;
+      double xbytes    = Nloop*dbytes*2.0*ncomm;
+      double rbytes    = xbytes;
+      double bidibytes = xbytes+rbytes;
+
+      double time = stop-start; // microseconds
+
+      std::cout<<GridLogMessage << lat<<"\t\t"<<Ls<<"\t\t"<<bytes<<"\t\t"<<xbytes/time<<"\t\t"<<bidibytes/time<<std::endl;
+    }
+  }    
 
   Grid_finalize();
 }

benchmarks/Benchmark_dwf.cc

@@ -37,26 +37,21 @@ struct scal {
   d internal;
 };
 
-  Gamma::GammaMatrix Gmu [] = {
-    Gamma::GammaX,
-    Gamma::GammaY,
-    Gamma::GammaZ,
-    Gamma::GammaT
+  Gamma::Algebra Gmu [] = {
+    Gamma::Algebra::GammaX,
+    Gamma::Algebra::GammaY,
+    Gamma::Algebra::GammaZ,
+    Gamma::Algebra::GammaT
   };
 
-bool overlapComms = false;
 typedef WilsonFermion5D<DomainWallVec5dImplR> WilsonFermion5DR;
 typedef WilsonFermion5D<DomainWallVec5dImplF> WilsonFermion5DF;
 typedef WilsonFermion5D<DomainWallVec5dImplD> WilsonFermion5DD;
 
-
 int main (int argc, char ** argv)
 {
   Grid_init(&argc,&argv);
 
-  if( GridCmdOptionExists(argv,argv+argc,"--asynch") ){
-    overlapComms = true;
-  }
 
   int threads = GridThread::GetThreads();
   std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;
@@ -76,35 +71,66 @@ int main (int argc, char ** argv)
 
   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;
 
   LatticeFermion src   (FGrid); random(RNG5,src);
+#if 0
+  src = zero;
+  {
+    std::vector<int> origin({0,0,0,latt4[2]-1,0});
+    SpinColourVectorF tmp;
+    tmp=zero;
+    tmp()(0)(0)=Complex(-2.0,0.0);
+    std::cout << " source site 0 " << tmp<<std::endl;
+    pokeSite(tmp,src,origin);
+  }
+#else
+  RealD N2 = 1.0/::sqrt(norm2(src));
+  src = src*N2;
+#endif
+
+
   LatticeFermion result(FGrid); result=zero;
   LatticeFermion    ref(FGrid);    ref=zero;
   LatticeFermion    tmp(FGrid);
   LatticeFermion    err(FGrid);
 
+  std::cout << GridLogMessage << "Drawing gauge field" << std::endl;
   LatticeGaugeField Umu(UGrid); 
-  random(RNG4,Umu);
-
-  LatticeGaugeField Umu5d(FGrid); 
+  SU3::HotConfiguration(RNG4,Umu); 
+  std::cout << GridLogMessage << "Random gauge initialised " << std::endl;
+#if 0
+  Umu=1.0;
+  for(int mu=0;mu<Nd;mu++){
+    LatticeColourMatrix ttmp(UGrid);
+    ttmp = PeekIndex<LorentzIndex>(Umu,mu);
+    //    if (mu !=2 ) ttmp = 0;
+    //    ttmp = ttmp* pow(10.0,mu);
+    PokeIndex<LorentzIndex>(Umu,ttmp,mu);
+  }
+  std::cout << GridLogMessage << "Forced to diagonal " << std::endl;
+#endif
 
+  ////////////////////////////////////
+  // Naive wilson implementation
+  ////////////////////////////////////
   // replicate across fifth dimension
+  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];
     }
   }
-
-  ////////////////////////////////////
-  // Naive wilson implementation
-  ////////////////////////////////////
-  std::vector<LatticeColourMatrix> U(4,FGrid);
   for(int mu=0;mu<Nd;mu++){
     U[mu] = PeekIndex<LorentzIndex>(Umu5d,mu);
   }
+  std::cout << GridLogMessage << "Setting up Cshift based reference " << std::endl;
 
   if (1)
   {
@@ -126,18 +152,27 @@ int main (int argc, char ** argv)
 
   RealD NP = UGrid->_Nprocessors;
 
-  for(int doasm=1;doasm<2;doasm++){
-
-    QCD::WilsonKernelsStatic::AsmOpt=doasm;
-
+  std::cout << GridLogMessage << "Creating action operator " << std::endl;
   DomainWallFermionR Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
-  
-  std::cout<<GridLogMessage << "Naive wilson implementation "<<std::endl;
-  std::cout << GridLogMessage<< "Calling Dw"<<std::endl;
-  int ncall =100;
-  if (1) {
 
+  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 DomainWallFermionR::Dhop                  "<<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;
+  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;
+
+  int ncall =1000;
+  if (1) {
+    FGrid->Barrier();
     Dw.ZeroCounters();
+    Dw.Dhop(src,result,0);
     double t0=usecond();
     for(int i=0;i<ncall;i++){
       __SSC_START;
@@ -145,95 +180,112 @@ int main (int argc, char ** argv)
       __SSC_STOP;
     }
     double t1=usecond();
+    FGrid->Barrier();
     
     double volume=Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
     double flops=1344*volume*ncall;
 
     std::cout<<GridLogMessage << "Called Dw "<<ncall<<" times in "<<t1-t0<<" us"<<std::endl;
-    std::cout<<GridLogMessage << "norm result "<< norm2(result)<<std::endl;
-    std::cout<<GridLogMessage << "norm ref    "<< norm2(ref)<<std::endl;
+    //    std::cout<<GridLogMessage << "norm result "<< norm2(result)<<std::endl;
+    //    std::cout<<GridLogMessage << "norm ref    "<< norm2(ref)<<std::endl;
     std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t1-t0)<<std::endl;
-    std::cout<<GridLogMessage << "mflop/s per node =  "<< flops/(t1-t0)/NP<<std::endl;
+    std::cout<<GridLogMessage << "mflop/s per rank =  "<< flops/(t1-t0)/NP<<std::endl;
     err = ref-result; 
     std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
+
+    /*
+    if(( norm2(err)>1.0e-4) ) { 
+      std::cout << "RESULT\n " << result<<std::endl;
+      std::cout << "REF   \n " << ref   <<std::endl;
+      std::cout << "ERR   \n " << err   <<std::endl;
+      FGrid->Barrier();
+      exit(-1);
+    }
+    */
+    assert (norm2(err)< 1.0e-4 );
     Dw.Report();
   }
 
   if (1)
   {
+
+    std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
+    std::cout << GridLogMessage<< "* Benchmarking WilsonFermion5D<DomainWallVec5dImplR>::Dhop "<<std::endl;
+    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;
+    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;
+
     typedef WilsonFermion5D<DomainWallVec5dImplR> WilsonFermion5DR;
     LatticeFermion ssrc(sFGrid);
     LatticeFermion sref(sFGrid);
     LatticeFermion sresult(sFGrid);
 
     WilsonFermion5DR sDw(Umu,*sFGrid,*sFrbGrid,*sUGrid,*sUrbGrid,M5);
-  
-    for(int x=0;x<latt4[0];x++){
-    for(int y=0;y<latt4[1];y++){
-    for(int z=0;z<latt4[2];z++){
-    for(int t=0;t<latt4[3];t++){
-    for(int s=0;s<Ls;s++){
-      std::vector<int> site({s,x,y,z,t});
-      SpinColourVector tmp;
-      peekSite(tmp,src,site);
-      pokeSite(tmp,ssrc,site);
-    }}}}}
+
+    localConvert(src,ssrc);
     std::cout<<GridLogMessage<< "src norms "<< norm2(src)<<" " <<norm2(ssrc)<<std::endl;
-    double t0=usecond();
+    FGrid->Barrier();
+    sDw.Dhop(ssrc,sresult,0);
     sDw.ZeroCounters();
+    double t0=usecond();
     for(int i=0;i<ncall;i++){
       __SSC_START;
       sDw.Dhop(ssrc,sresult,0);
       __SSC_STOP;
     }
     double t1=usecond();
+    FGrid->Barrier();
     double volume=Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
     double flops=1344*volume*ncall;
 
     std::cout<<GridLogMessage << "Called Dw s_inner "<<ncall<<" times in "<<t1-t0<<" us"<<std::endl;
     std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t1-t0)<<std::endl;
-    std::cout<<GridLogMessage << "mflop/s per node =  "<< flops/(t1-t0)/NP<<std::endl;
+    std::cout<<GridLogMessage << "mflop/s per rank =  "<< flops/(t1-t0)/NP<<std::endl;
+    //    std::cout<<GridLogMessage<< "res norms "<< norm2(result)<<" " <<norm2(sresult)<<std::endl;
     sDw.Report();
-  
-    if(0){
-      for(int i=0;i< PerformanceCounter::NumTypes(); i++ ){
-  sDw.Dhop(ssrc,sresult,0);
-  PerformanceCounter Counter(i);
-  Counter.Start();
-  sDw.Dhop(ssrc,sresult,0);
-  Counter.Stop();
-  Counter.Report();
-      }
+    RealD sum=0;
+
+    err=zero;
+    localConvert(sresult,err);
+    err = err - ref;
+    sum = norm2(err);
+    std::cout<<GridLogMessage<<" difference between normal ref and simd is "<<sum<<std::endl;
+    if(sum > 1.0e-4 ){
+      std::cout<< "sD REF\n " <<ref << std::endl;
+      std::cout<< "sD ERR   \n " <<err  <<std::endl;
     }
+    //    assert(sum < 1.0e-4);
 
-    std::cout<<GridLogMessage<< "res norms "<< norm2(result)<<" " <<norm2(sresult)<<std::endl;
-
-
-    RealF sum=0;
-    for(int x=0;x<latt4[0];x++){
-    for(int y=0;y<latt4[1];y++){
-    for(int z=0;z<latt4[2];z++){
-    for(int t=0;t<latt4[3];t++){
-    for(int s=0;s<Ls;s++){
-      std::vector<int> site({s,x,y,z,t});
-      SpinColourVector normal, simd;
-      peekSite(normal,result,site);
-      peekSite(simd,sresult,site);
-      sum=sum+norm2(normal-simd);
-      if (norm2(normal-simd) > 1.0e-6 ) {
-  std::cout << "site "<<x<<","<<y<<","<<z<<","<<t<<","<<s<<" "<<norm2(normal-simd)<<std::endl;
-  std::cout << "site "<<x<<","<<y<<","<<z<<","<<t<<","<<s<<" normal "<<normal<<std::endl;
-  std::cout << "site "<<x<<","<<y<<","<<z<<","<<t<<","<<s<<" simd   "<<simd<<std::endl;
-      }
-    }}}}}
-    std::cout<<GridLogMessage<<" difference between normal and simd is "<<sum<<std::endl;
-
-
-    if (1) {
+    err=zero;
+    localConvert(sresult,err);
+    err = err - result;
+    sum = norm2(err);
+    std::cout<<GridLogMessage<<" difference between normal result and simd is "<<sum<<std::endl;
+    if(sum > 1.0e-4 ){
+      std::cout<< "sD REF\n " <<result << std::endl;
+      std::cout<< "sD ERR   \n " << err  <<std::endl;
+    }
+    assert(sum < 1.0e-4);
+    
+    if(1){
+      std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
+      std::cout << GridLogMessage<< "* Benchmarking WilsonFermion5D<DomainWallVec5dImplR>::DhopEO "<<std::endl;
+      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;
+      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;
 
       LatticeFermion sr_eo(sFGrid);
-      LatticeFermion serr(sFGrid);
-
       LatticeFermion ssrc_e (sFrbGrid);
       LatticeFermion ssrc_o (sFrbGrid);
       LatticeFermion sr_e   (sFrbGrid);
@@ -241,29 +293,29 @@ int main (int argc, char ** argv)
 
       pickCheckerboard(Even,ssrc_e,ssrc);
       pickCheckerboard(Odd,ssrc_o,ssrc);
-
-      setCheckerboard(sr_eo,ssrc_o);
-      setCheckerboard(sr_eo,ssrc_e);
-      serr = sr_eo-ssrc; 
-      std::cout<<GridLogMessage << "EO src norm diff   "<< norm2(serr)<<std::endl;
+      //      setCheckerboard(sr_eo,ssrc_o);
+      //      setCheckerboard(sr_eo,ssrc_e);
 
       sr_e = zero;
       sr_o = zero;
 
+      FGrid->Barrier();
+      sDw.DhopEO(ssrc_o, sr_e, DaggerNo);
       sDw.ZeroCounters();
-      sDw.stat.init("DhopEO");
+      //      sDw.stat.init("DhopEO");
       double t0=usecond();
       for (int i = 0; i < ncall; i++) {
         sDw.DhopEO(ssrc_o, sr_e, DaggerNo);
       }
       double t1=usecond();
-      sDw.stat.print();
+      FGrid->Barrier();
+      //      sDw.stat.print();
 
       double volume=Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
       double flops=(1344.0*volume*ncall)/2;
 
       std::cout<<GridLogMessage << "sDeo mflop/s =   "<< flops/(t1-t0)<<std::endl;
-      std::cout<<GridLogMessage << "sDeo mflop/s per node   "<< flops/(t1-t0)/NP<<std::endl;
+      std::cout<<GridLogMessage << "sDeo mflop/s per rank   "<< flops/(t1-t0)/NP<<std::endl;
       sDw.Report();
 
       sDw.DhopEO(ssrc_o,sr_e,DaggerNo);
@@ -272,13 +324,26 @@ int main (int argc, char ** argv)
 
       pickCheckerboard(Even,ssrc_e,sresult);
       pickCheckerboard(Odd ,ssrc_o,sresult);
+
       ssrc_e = ssrc_e - sr_e;
+      RealD error = norm2(ssrc_e);
       std::cout<<GridLogMessage << "sE norm diff   "<< norm2(ssrc_e)<< "  vec nrm"<<norm2(sr_e) <<std::endl;
+
       ssrc_o = ssrc_o - sr_o;
+      error+= norm2(ssrc_o);
       std::cout<<GridLogMessage << "sO norm diff   "<< norm2(ssrc_o)<< "  vec nrm"<<norm2(sr_o) <<std::endl;
+
+      if(( error>1.0e-4) ) { 
+	setCheckerboard(ssrc,ssrc_o);
+	setCheckerboard(ssrc,ssrc_e);
+	std::cout<< "DIFF\n " <<ssrc << std::endl;
+	setCheckerboard(ssrc,sr_o);
+	setCheckerboard(ssrc,sr_e);
+	std::cout<< "CBRESULT\n " <<ssrc << std::endl;
+	std::cout<< "RESULT\n " <<sresult<< std::endl;
+      }
+      assert(error<1.0e-4);
     }
-
-
   }
 
   if (1)
@@ -286,28 +351,33 @@ int main (int argc, char ** argv)
     ref = zero;
     for(int mu=0;mu<Nd;mu++){
 
-      //    ref =  src - Gamma(Gamma::GammaX)* src ; // 1+gamma_x
+      //    ref =  src - Gamma(Gamma::Algebra::GammaX)* src ; // 1+gamma_x
       tmp = U[mu]*Cshift(src,mu+1,1);
       for(int i=0;i<ref._odata.size();i++){
-  ref._odata[i]+= tmp._odata[i] + Gamma(Gmu[mu])*tmp._odata[i]; ;
+	ref._odata[i]+= tmp._odata[i] + Gamma(Gmu[mu])*tmp._odata[i]; ;
       }
 
       tmp =adj(U[mu])*src;
       tmp =Cshift(tmp,mu+1,-1);
       for(int i=0;i<ref._odata.size();i++){
-  ref._odata[i]+= tmp._odata[i] - Gamma(Gmu[mu])*tmp._odata[i]; ;
+	ref._odata[i]+= tmp._odata[i] - Gamma(Gmu[mu])*tmp._odata[i]; ;
       }
     }
     ref = -0.5*ref;
   }
+  //  dump=1;
   Dw.Dhop(src,result,1);
-  std::cout << GridLogMessage << "Naive wilson implementation Dag" << std::endl;
+  std::cout << GridLogMessage << "Compare to naive wilson implementation Dag to verify correctness" << std::endl;
   std::cout<<GridLogMessage << "Called DwDag"<<std::endl;
-  std::cout<<GridLogMessage << "norm result "<< norm2(result)<<std::endl;
-  std::cout<<GridLogMessage << "norm ref    "<< norm2(ref)<<std::endl;
+  std::cout<<GridLogMessage << "norm dag result "<< norm2(result)<<std::endl;
+  std::cout<<GridLogMessage << "norm dag ref    "<< norm2(ref)<<std::endl;
   err = ref-result; 
-  std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
-
+  std::cout<<GridLogMessage << "norm dag diff   "<< norm2(err)<<std::endl;
+  if((norm2(err)>1.0e-4)){
+	std::cout<< "DAG RESULT\n "  <<ref     << std::endl;
+	std::cout<< "DAG sRESULT\n " <<result  << std::endl;
+	std::cout<< "DAG ERR   \n "  << err    <<std::endl;
+  }
   LatticeFermion src_e (FrbGrid);
   LatticeFermion src_o (FrbGrid);
   LatticeFermion r_e   (FrbGrid);
@@ -315,26 +385,43 @@ int main (int argc, char ** argv)
   LatticeFermion r_eo  (FGrid);
 
 
-  std::cout<<GridLogMessage << "Calling Deo and Doe"<<std::endl;
+  std::cout<<GridLogMessage << "Calling Deo and Doe and //assert Deo+Doe == Dunprec"<<std::endl;
   pickCheckerboard(Even,src_e,src);
   pickCheckerboard(Odd,src_o,src);
 
   std::cout<<GridLogMessage << "src_e"<<norm2(src_e)<<std::endl;
   std::cout<<GridLogMessage << "src_o"<<norm2(src_o)<<std::endl;
 
+
+  // 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;
+  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;
   {
     Dw.ZeroCounters();
+    FGrid->Barrier();
+    Dw.DhopEO(src_o,r_e,DaggerNo);
     double t0=usecond();
     for(int i=0;i<ncall;i++){
       Dw.DhopEO(src_o,r_e,DaggerNo);
     }
     double t1=usecond();
+    FGrid->Barrier();
     
     double volume=Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
     double flops=(1344.0*volume*ncall)/2;
 
     std::cout<<GridLogMessage << "Deo mflop/s =   "<< flops/(t1-t0)<<std::endl;
-    std::cout<<GridLogMessage << "Deo mflop/s per node   "<< flops/(t1-t0)/NP<<std::endl;
+    std::cout<<GridLogMessage << "Deo mflop/s per rank   "<< flops/(t1-t0)/NP<<std::endl;
     Dw.Report();
   }
   Dw.DhopEO(src_o,r_e,DaggerNo);
@@ -350,14 +437,19 @@ int main (int argc, char ** argv)
 
   err = r_eo-result; 
   std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
+  if((norm2(err)>1.0e-4)){
+	std::cout<< "Deo RESULT\n " <<r_eo << std::endl;
+	std::cout<< "Deo REF\n " <<result  << std::endl;
+	std::cout<< "Deo ERR   \n " << err <<std::endl;
+  }
 
   pickCheckerboard(Even,src_e,err);
   pickCheckerboard(Odd,src_o,err);
   std::cout<<GridLogMessage << "norm diff even  "<< norm2(src_e)<<std::endl;
   std::cout<<GridLogMessage << "norm diff odd   "<< norm2(src_o)<<std::endl;
 
-
-  }
+  //assert(norm2(src_e)<1.0e-4);
+  //assert(norm2(src_o)<1.0e-4);
 
   Grid_finalize();
 }

benchmarks/Benchmark_dwf_ntpf.cc

@@ -1,153 +0,0 @@
-    /*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./benchmarks/Benchmark_dwf.cc
-
-    Copyright (C) 2015
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-    *************************************************************************************/
-    /*  END LEGAL */
-#include <Grid/Grid.h>
-
-using namespace std;
-using namespace Grid;
-using namespace Grid::QCD;
-
-template<class d>
-struct scal {
-  d internal;
-};
-
-  Gamma::GammaMatrix Gmu [] = {
-    Gamma::GammaX,
-    Gamma::GammaY,
-    Gamma::GammaZ,
-    Gamma::GammaT
-  };
-
-bool overlapComms = false;
-
-
-int main (int argc, char ** argv)
-{
-  Grid_init(&argc,&argv);
-
-  if( GridCmdOptionExists(argv,argv+argc,"--asynch") ){
-    overlapComms = true;
-  }
-
-  int threads = GridThread::GetThreads();
-  std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;
-
-  std::vector<int> latt4 = GridDefaultLatt();
-  const int Ls=16;
-  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
-  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
-  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
-  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
-
-  std::vector<int> seeds4({1,2,3,4});
-  std::vector<int> seeds5({5,6,7,8});
-
-  GridParallelRNG          RNG4(UGrid);  RNG4.SeedFixedIntegers(seeds4);
-  GridParallelRNG          RNG5(FGrid);  RNG5.SeedFixedIntegers(seeds5);
-
-  LatticeFermion src   (FGrid); random(RNG5,src);
-  LatticeFermion result(FGrid); result=zero;
-  LatticeFermion    ref(FGrid);    ref=zero;
-  LatticeFermion    tmp(FGrid);
-  LatticeFermion    err(FGrid);
-
-  ColourMatrix cm = Complex(1.0,0.0);
-
-  LatticeGaugeField Umu(UGrid); 
-  random(RNG4,Umu);
-
-  LatticeGaugeField Umu5d(FGrid); 
-
-  // replicate across fifth dimension
-  for(int ss=0;ss<Umu._grid->oSites();ss++){
-    for(int s=0;s<Ls;s++){
-      Umu5d._odata[Ls*ss+s] = Umu._odata[ss];
-    }
-  }
-
-  ////////////////////////////////////
-  // Naive wilson implementation
-  ////////////////////////////////////
-  std::vector<LatticeColourMatrix> U(4,FGrid);
-  for(int mu=0;mu<Nd;mu++){
-    U[mu] = PeekIndex<LorentzIndex>(Umu5d,mu);
-  }
-
-  if (1)
-  {
-    ref = zero;
-    for(int mu=0;mu<Nd;mu++){
-
-      tmp = U[mu]*Cshift(src,mu+1,1);
-      ref=ref + tmp - Gamma(Gmu[mu])*tmp;
-
-      tmp =adj(U[mu])*src;
-      tmp =Cshift(tmp,mu+1,-1);
-      ref=ref + tmp + Gamma(Gmu[mu])*tmp;
-    }
-    ref = -0.5*ref;
-  }
-
-  RealD mass=0.1;
-  RealD M5  =1.8;
-
-  typename DomainWallFermionR::ImplParams params; 
-  params.overlapCommsCompute = overlapComms;
-  
-  RealD NP = UGrid->_Nprocessors;
-
-
-  QCD::WilsonKernelsStatic::AsmOpt=1;
-
-  DomainWallFermionR Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5,params);
-  
-  std::cout<<GridLogMessage << "Calling Dw"<<std::endl;
-  int ncall =50;
-  if (1) {
-
-    double t0=usecond();
-    for(int i=0;i<ncall;i++){
-      Dw.Dhop(src,result,0);
-    }
-    double t1=usecond();
-    
-    double volume=Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
-    double flops=1344*volume*ncall;
-
-    std::cout<<GridLogMessage << "Called Dw "<<ncall<<" times in "<<t1-t0<<" us"<<std::endl;
-    std::cout<<GridLogMessage << "norm result "<< norm2(result)<<std::endl;
-    std::cout<<GridLogMessage << "norm ref    "<< norm2(ref)<<std::endl;
-    std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t1-t0)<<std::endl;
-    std::cout<<GridLogMessage << "mflop/s per node =  "<< flops/(t1-t0)/NP<<std::endl;
-    err = ref-result; 
-    std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
-    //    Dw.Report();
-  }
-  Grid_finalize();
-}

benchmarks/Benchmark_dwf_sweep.cc

@@ -37,11 +37,11 @@ struct scal {
   d internal;
 };
 
-  Gamma::GammaMatrix Gmu [] = {
-    Gamma::GammaX,
-    Gamma::GammaY,
-    Gamma::GammaZ,
-    Gamma::GammaT
+  Gamma::Algebra Gmu [] = {
+    Gamma::Algebra::GammaX,
+    Gamma::Algebra::GammaY,
+    Gamma::Algebra::GammaZ,
+    Gamma::Algebra::GammaT
   };
 
 void benchDw(std::vector<int> & L, int Ls, int threads, int report =0 );
@@ -51,24 +51,26 @@ int main (int argc, char ** argv)
 {
   Grid_init(&argc,&argv);
 
+  std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
+  std::cout << GridLogMessage<< "* Kernel options --dslash-generic, --dslash-unroll, --dslash-asm" <<std::endl;
+  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;
+  std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
+
   const int Ls=8;
   int threads = GridThread::GetThreads();
   std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;
-
-  if ( getenv("ASMOPT") )  {
-    QCD::WilsonKernelsStatic::AsmOpt=1;
-  } else { 
-    QCD::WilsonKernelsStatic::AsmOpt=0;
-  }
-
   std::cout<<GridLogMessage << "=========================================================================="<<std::endl;
   std::cout<<GridLogMessage << "= Benchmarking DWF"<<std::endl;
   std::cout<<GridLogMessage << "=========================================================================="<<std::endl;
   std::cout<<GridLogMessage << "Volume \t\t\tProcs \t Dw \t eoDw \t sDw \t eosDw (Mflop/s)  "<<std::endl;
   std::cout<<GridLogMessage << "=========================================================================="<<std::endl;
 
-  int Lmax=32;
-  int dmin=0;
+  int Lmax=16;
+  int dmin=2;
   if ( getenv("LMAX") ) Lmax=atoi(getenv("LMAX"));
   if ( getenv("DMIN") ) dmin=atoi(getenv("DMIN"));
   for (int L=8;L<=Lmax;L*=2){

benchmarks/Benchmark_memory_asynch.cc

@@ -77,8 +77,7 @@ int main (int argc, char ** argv)
     }
 
     double start=usecond();
-PARALLEL_FOR_LOOP
-    for(int t=0;t<threads;t++){
+    parallel_for(int t=0;t<threads;t++){
 
       sum[t] = x[t]._odata[0];
       for(int i=0;i<Nloop;i++){

benchmarks/Benchmark_mooee.cc

@@ -0,0 +1,222 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./benchmarks/Benchmark_dwf.cc
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#include <Grid/Grid.h>
+
+using namespace std;
+using namespace Grid;
+using namespace Grid::QCD;
+
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  int threads = GridThread::GetThreads();
+  std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;
+
+  std::vector<int> latt4 = GridDefaultLatt();
+  const int Ls=16;
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
+  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
+
+  std::cout << GridLogMessage << "Making Vec5d innermost grids"<<std::endl;
+  GridCartesian         * sUGrid   = SpaceTimeGrid::makeFourDimDWFGrid(GridDefaultLatt(),GridDefaultMpi());
+  GridRedBlackCartesian * sUrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(sUGrid);
+  GridCartesian         * sFGrid   = SpaceTimeGrid::makeFiveDimDWFGrid(Ls,UGrid);
+  GridRedBlackCartesian * sFrbGrid = SpaceTimeGrid::makeFiveDimDWFRedBlackGrid(Ls,UGrid);
+
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+
+  GridParallelRNG          RNG4(UGrid);  RNG4.SeedFixedIntegers(seeds4);
+  std::cout << GridLogMessage << "Seeded"<<std::endl;
+
+  LatticeGaugeField Umu(UGrid); SU3::HotConfiguration(RNG4,Umu);
+
+  std::cout << GridLogMessage << "made random gauge fields"<<std::endl;
+
+  RealD mass=0.1;
+  RealD M5  =1.8;
+  RealD NP = UGrid->_Nprocessors;
+
+
+  if (1)
+  {
+    const int ncall=1000;
+
+    std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
+    std::cout << GridLogMessage<< "* Benchmarking DomainWallFermionR::Dhop "<<std::endl;
+    std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
+
+    GridParallelRNG RNG5(FGrid);
+    LatticeFermion src(FGrid); random(RNG5,src);
+    LatticeFermion result(FGrid);
+
+    DomainWallFermionR Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+    double t0,t1;
+
+    LatticeFermion r_eo(FGrid);
+    LatticeFermion src_e (FrbGrid);
+    LatticeFermion src_o (FrbGrid);
+    LatticeFermion r_e   (FrbGrid);
+    LatticeFermion r_o   (FrbGrid);
+    
+    pickCheckerboard(Even,src_e,src);
+    pickCheckerboard(Odd,src_o,src);
+    
+    setCheckerboard(r_eo,src_o);
+    setCheckerboard(r_eo,src_e);
+    
+    r_e = zero;
+    r_o = zero;
+
+
+#define BENCH_DW(A,in,out)			\
+    Dw.CayleyZeroCounters();			\
+    Dw. A (in,out);				\
+    FGrid->Barrier();				\
+    t0=usecond();				\
+    for(int i=0;i<ncall;i++){			\
+      Dw. A (in,out);				\
+    }						\
+    t1=usecond();				\
+    FGrid->Barrier();				\
+    Dw.CayleyReport();					\
+    std::cout<<GridLogMessage << "Called " #A " "<< (t1-t0)/ncall<<" us"<<std::endl;\
+    std::cout<<GridLogMessage << "******************"<<std::endl;
+
+#define BENCH_ZDW(A,in,out)			\
+    zDw.CayleyZeroCounters();			\
+    zDw. A (in,out);				\
+    FGrid->Barrier();				\
+    t0=usecond();				\
+    for(int i=0;i<ncall;i++){			\
+      zDw. A (in,out);				\
+    }						\
+    t1=usecond();				\
+    FGrid->Barrier();				\
+    zDw.CayleyReport();							\
+    std::cout<<GridLogMessage << "Called ZDw " #A " "<< (t1-t0)/ncall<<" us"<<std::endl;\
+    std::cout<<GridLogMessage << "******************"<<std::endl;
+
+#define BENCH_DW_SSC(A,in,out)			\
+    Dw.CayleyZeroCounters();			\
+    Dw. A (in,out);				\
+    FGrid->Barrier();				\
+    t0=usecond();				\
+    for(int i=0;i<ncall;i++){			\
+      __SSC_START ;				\
+      Dw. A (in,out);				\
+      __SSC_STOP ;				\
+    }						\
+    t1=usecond();				\
+    FGrid->Barrier();				\
+    Dw.CayleyReport();					\
+    std::cout<<GridLogMessage << "Called " #A " "<< (t1-t0)/ncall<<" us"<<std::endl;\
+    std::cout<<GridLogMessage << "******************"<<std::endl;
+
+#define BENCH_DW_MEO(A,in,out)			\
+    Dw.CayleyZeroCounters();			\
+    Dw. A (in,out,0);				\
+    FGrid->Barrier();				\
+    t0=usecond();				\
+    for(int i=0;i<ncall;i++){			\
+      Dw. A (in,out,0);				\
+    }						\
+    t1=usecond();				\
+    FGrid->Barrier();				\
+    Dw.CayleyReport();					\
+    std::cout<<GridLogMessage << "Called " #A " "<< (t1-t0)/ncall<<" us"<<std::endl;\
+    std::cout<<GridLogMessage << "******************"<<std::endl;
+
+    BENCH_DW_MEO(Dhop    ,src,result);
+    BENCH_DW_MEO(DhopEO  ,src_o,r_e);
+    BENCH_DW(Meooe   ,src_o,r_e);
+    BENCH_DW(Mooee   ,src_o,r_o);
+    BENCH_DW(MooeeInv,src_o,r_o);
+
+  }
+
+  if (1)
+  {
+    const int ncall=1000;
+
+    std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
+    std::cout << GridLogMessage<< "* Benchmarking DomainWallFermionVec5dR::Dhop "<<std::endl;
+    std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
+
+    GridParallelRNG RNG5(sFGrid);
+    LatticeFermion src(sFGrid); random(RNG5,src);
+    LatticeFermion sref(sFGrid);
+    LatticeFermion result(sFGrid);
+
+
+    std::cout<<GridLogMessage << "Constructing Vec5D Dw "<<std::endl;
+    DomainWallFermionVec5dR Dw(Umu,*sFGrid,*sFrbGrid,*sUGrid,*sUrbGrid,mass,M5);
+
+    RealD b=1.5;// Scale factor b+c=2, b-c=1
+    RealD c=0.5;
+    std::vector<ComplexD> gamma(Ls,std::complex<double>(1.0,0.0));
+    ZMobiusFermionVec5dR zDw(Umu,*sFGrid,*sFrbGrid,*sUGrid,*sUrbGrid,mass,M5,gamma,b,c);
+
+    std::cout<<GridLogMessage << "Calling Dhop "<<std::endl;
+    FGrid->Barrier();
+
+    double t0,t1;
+
+    LatticeFermion r_eo(sFGrid);
+    LatticeFermion src_e (sFrbGrid);
+    LatticeFermion src_o (sFrbGrid);
+    LatticeFermion r_e   (sFrbGrid);
+    LatticeFermion r_o   (sFrbGrid);
+    
+    pickCheckerboard(Even,src_e,src);
+    pickCheckerboard(Odd,src_o,src);
+    
+    setCheckerboard(r_eo,src_o);
+    setCheckerboard(r_eo,src_e);
+    
+    r_e = zero;
+    r_o = zero;
+
+    BENCH_DW_MEO(Dhop    ,src,result);
+    BENCH_DW_MEO(DhopEO  ,src_o,r_e);
+    BENCH_DW_SSC(Meooe   ,src_o,r_e);
+    BENCH_DW(Mooee   ,src_o,r_o);
+    BENCH_DW(MooeeInv,src_o,r_o);
+
+    BENCH_ZDW(Mooee   ,src_o,r_o);
+    BENCH_ZDW(MooeeInv,src_o,r_o);
+
+  }
+
+  Grid_finalize();
+}

benchmarks/Benchmark_staggered.cc

@@ -0,0 +1,134 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./benchmarks/Benchmark_staggered.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;
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  std::vector<int> latt_size   = GridDefaultLatt();
+  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);
+
+  int threads = GridThread::GetThreads();
+  std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;
+  std::cout<<GridLogMessage << "Grid floating point word size is REALF"<< sizeof(RealF)<<std::endl;
+  std::cout<<GridLogMessage << "Grid floating point word size is REALD"<< sizeof(RealD)<<std::endl;
+  std::cout<<GridLogMessage << "Grid floating point word size is REAL"<< sizeof(Real)<<std::endl;
+
+  std::vector<int> seeds({1,2,3,4});
+  GridParallelRNG          pRNG(&Grid);
+  pRNG.SeedFixedIntegers(seeds);
+  //  pRNG.SeedRandomDevice();
+
+  typedef typename ImprovedStaggeredFermionR::FermionField FermionField; 
+  typename ImprovedStaggeredFermionR::ImplParams params; 
+
+  FermionField src   (&Grid); random(pRNG,src);
+  FermionField result(&Grid); result=zero;
+  FermionField    ref(&Grid);    ref=zero;
+  FermionField    tmp(&Grid);    tmp=zero;
+  FermionField    err(&Grid);    tmp=zero;
+  LatticeGaugeField Umu(&Grid); random(pRNG,Umu);
+  std::vector<LatticeColourMatrix> U(4,&Grid);
+
+  double volume=1;
+  for(int mu=0;mu<Nd;mu++){
+    volume=volume*latt_size[mu];
+  }  
+
+  // Only one non-zero (y)
+#if 0
+  Umu=zero;
+  Complex cone(1.0,0.0);
+  for(int nn=0;nn<Nd;nn++){
+    random(pRNG,U[nn]);
+    if(1) {
+      if (nn!=2) { U[nn]=zero; std::cout<<GridLogMessage << "zeroing gauge field in dir "<<nn<<std::endl; }
+      //      else       { U[nn]= cone;std::cout<<GridLogMessage << "unit gauge field in dir "<<nn<<std::endl; }
+      else       { std::cout<<GridLogMessage << "random gauge field in dir "<<nn<<std::endl; }
+    }
+    PokeIndex<LorentzIndex>(Umu,U[nn],nn);
+  }
+#endif
+
+  for(int mu=0;mu<Nd;mu++){
+    U[mu] = PeekIndex<LorentzIndex>(Umu,mu);
+  }
+  ref = zero;
+  /*  
+  { // Naive wilson implementation
+    ref = zero;
+    for(int mu=0;mu<Nd;mu++){
+      //    ref =  src + Gamma(Gamma::GammaX)* src ; // 1-gamma_x
+      tmp = U[mu]*Cshift(src,mu,1);
+      for(int i=0;i<ref._odata.size();i++){
+	ref._odata[i]+= tmp._odata[i] - Gamma(Gmu[mu])*tmp._odata[i]; ;
+      }
+
+      tmp =adj(U[mu])*src;
+      tmp =Cshift(tmp,mu,-1);
+      for(int i=0;i<ref._odata.size();i++){
+	ref._odata[i]+= tmp._odata[i] + Gamma(Gmu[mu])*tmp._odata[i]; ;
+      }
+    }
+  }
+  ref = -0.5*ref;
+  */
+
+  RealD mass=0.1;
+  RealD c1=9.0/8.0;
+  RealD c2=-1.0/24.0;
+  RealD u0=1.0;
+  ImprovedStaggeredFermionR Ds(Umu,Umu,Grid,RBGrid,mass,c1,c2,u0,params);
+  
+  std::cout<<GridLogMessage << "Calling Ds"<<std::endl;
+  int ncall=1000;
+  double t0=usecond();
+  for(int i=0;i<ncall;i++){
+    Ds.Dhop(src,result,0);
+  }
+  double t1=usecond();
+  double flops=(16*(3*(6+8+8)) + 15*3*2)*volume*ncall; // == 66*16 +  == 1146
+  
+  std::cout<<GridLogMessage << "Called Ds"<<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;
+  err = ref-result; 
+  std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
+
+  Grid_finalize();
+}

benchmarks/Benchmark_wilson.cc

@@ -37,11 +37,11 @@ struct scal {
   d internal;
 };
 
-  Gamma::GammaMatrix Gmu [] = {
-    Gamma::GammaX,
-    Gamma::GammaY,
-    Gamma::GammaZ,
-    Gamma::GammaT
+  Gamma::Algebra Gmu [] = {
+    Gamma::Algebra::GammaX,
+    Gamma::Algebra::GammaY,
+    Gamma::Algebra::GammaZ,
+    Gamma::Algebra::GammaT
   };
 
 bool overlapComms = false;
@@ -106,7 +106,7 @@ int main (int argc, char ** argv)
   { // Naive wilson implementation
     ref = zero;
     for(int mu=0;mu<Nd;mu++){
-      //    ref =  src + Gamma(Gamma::GammaX)* src ; // 1-gamma_x
+      //    ref =  src + Gamma(Gamma::Algebra::GammaX)* src ; // 1-gamma_x
       tmp = U[mu]*Cshift(src,mu,1);
       for(int i=0;i<ref._odata.size();i++){
 	ref._odata[i]+= tmp._odata[i] - Gamma(Gmu[mu])*tmp._odata[i]; ;
@@ -159,7 +159,7 @@ int main (int argc, char ** argv)
     ref = zero;
     for(int mu=0;mu<Nd;mu++){
 
-      //    ref =  src - Gamma(Gamma::GammaX)* src ; // 1+gamma_x
+      //    ref =  src - Gamma(Gamma::Algebra::GammaX)* src ; // 1+gamma_x
       tmp = U[mu]*Cshift(src,mu,1);
       for(int i=0;i<ref._odata.size();i++){
 	ref._odata[i]+= tmp._odata[i] + Gamma(Gmu[mu])*tmp._odata[i]; ;

benchmarks/Benchmark_wilson_sweep.cc

@@ -30,11 +30,11 @@ struct scal {
   d internal;
 };
 
-Gamma::GammaMatrix Gmu [] = {
-  Gamma::GammaX,
-  Gamma::GammaY,
-  Gamma::GammaZ,
-  Gamma::GammaT
+Gamma::Algebra Gmu [] = {
+    Gamma::Algebra::GammaX,
+    Gamma::Algebra::GammaY,
+    Gamma::Algebra::GammaZ,
+    Gamma::Algebra::GammaT
 };
 
 bool overlapComms = false;
@@ -58,6 +58,19 @@ int main (int argc, char ** argv)
   std::vector<int> seeds({1,2,3,4});
   RealD mass = 0.1;
 
+  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 WilsonFermionR::Dhop                  "<<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;
+  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 << "============================================================================="<< std::endl;
   std::cout<<GridLogMessage << "= Benchmarking Wilson" << std::endl;
   std::cout<<GridLogMessage << "============================================================================="<< std::endl;

benchmarks/Benchmark_zmm.cc

@@ -1,175 +0,0 @@
-    /*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./tests/Test_zmm.cc
-
-    Copyright (C) 2015
-
-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 Grid;
-using namespace Grid::QCD;
-
-
-int bench(std::ofstream &os, std::vector<int> &latt4,int Ls);
-
-int main(int argc,char **argv)
-{
-  Grid_init(&argc,&argv);
-  std::ofstream os("zmm.dat");
-
-  os << "#V Ls Lxy Lzt C++ Asm OMP L1 " <<std::endl;
-  std::cout<<GridLogMessage << "====================================================================="<<std::endl;
-  std::cout<<GridLogMessage << "= Benchmarking ZMM"<<std::endl;
-  std::cout<<GridLogMessage << "====================================================================="<<std::endl;
-  std::cout<<GridLogMessage << "Volume \t\t\t\tC++DW/MFLOPs\tASM-DW/MFLOPs\tdiff"<<std::endl;
-  std::cout<<GridLogMessage << "====================================================================="<<std::endl;
-  for(int L=4;L<=32;L+=4){
-    for(int m=1;m<=2;m++){
-      for(int Ls=8;Ls<=16;Ls+=8){
-	std::vector<int> grid({L,L,m*L,m*L});
-  std::cout << GridLogMessage <<"\t";
-	for(int i=0;i<4;i++) { 
-	  std::cout << grid[i]<<"x";
-	}
-	std::cout << Ls<<"\t\t";
-	bench(os,grid,Ls);
-      }
-    }
-  }
-}
-
-int bench(std::ofstream &os, std::vector<int> &latt4,int Ls)
-{
-
-  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(latt4, GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
-  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
-  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
-  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
-
-  std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
-  std::vector<int> mpi_layout  = GridDefaultMpi();
-  int threads = GridThread::GetThreads();
-
-  std::vector<int> seeds4({1,2,3,4});
-  std::vector<int> seeds5({5,6,7,8});
-
-  GridSerialRNG sRNG; sRNG.SeedFixedIntegers(seeds4);
-
-  LatticeFermion src (FGrid);
-  LatticeFermion tmp (FGrid);
-  LatticeFermion srce(FrbGrid);
-
-  LatticeFermion resulto(FrbGrid); resulto=zero;
-  LatticeFermion resulta(FrbGrid); resulta=zero;
-  LatticeFermion junk(FrbGrid); junk=zero;
-  LatticeFermion diff(FrbGrid); 
-  LatticeGaugeField Umu(UGrid);
-
-  double mfc, mfa, mfo, mfl1;
-
-  GridParallelRNG          RNG4(UGrid);  RNG4.SeedFixedIntegers(seeds4);
-  GridParallelRNG          RNG5(FGrid);  RNG5.SeedFixedIntegers(seeds5);
-  random(RNG5,src);
-#if 1
-  random(RNG4,Umu);
-#else
-  int mmu=2;
-  std::vector<LatticeColourMatrix> U(4,UGrid);
-  for(int mu=0;mu<Nd;mu++){
-    U[mu] = PeekIndex<LorentzIndex>(Umu,mu);
-    if ( mu!=mmu ) U[mu] = zero;
-    if ( mu==mmu ) U[mu] = 1.0;
-    PokeIndex<LorentzIndex>(Umu,U[mu],mu);
-  }
-#endif
- pickCheckerboard(Even,srce,src);
-
-  RealD mass=0.1;
-  RealD M5  =1.8;
-  DomainWallFermionR Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
-
-  int ncall=50;
-  double t0=usecond();
-  for(int i=0;i<ncall;i++){
-    Dw.DhopOE(srce,resulto,0);
-  }
-  double t1=usecond();
-
-  double volume=Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
-  double flops=1344*volume/2;
-
-  mfc = flops*ncall/(t1-t0);
-  std::cout<<mfc<<"\t\t";
-
-  QCD::WilsonKernelsStatic::AsmOpt=1;
-  t0=usecond();
-  for(int i=0;i<ncall;i++){
-    Dw.DhopOE(srce,resulta,0);
-  }
-  t1=usecond();
-  mfa = flops*ncall/(t1-t0);
-  std::cout<<mfa<<"\t\t";
-  /*
-  int dag=DaggerNo;
-  t0=usecond();
-  for(int i=0;i<1;i++){
-    Dw.DhopInternalOMPbench(Dw.StencilEven,Dw.LebesgueEvenOdd,Dw.UmuOdd,srce,resulta,dag);
-  }
-  t1=usecond();
-  mfo = flops*100/(t1-t0);
-  std::cout<<GridLogMessage << "Called ASM-OMP Dw"<< " mflop/s =   "<< mfo<<std::endl;
-
-  t0=usecond();
-  for(int i=0;i<1;i++){
-    Dw.DhopInternalL1bench(Dw.StencilEven,Dw.LebesgueEvenOdd,Dw.UmuOdd,srce,resulta,dag);
-  }
-  t1=usecond();
-  mfl1= flops*100/(t1-t0);
-  std::cout<<GridLogMessage << "Called ASM-L1 Dw"<< " mflop/s =   "<< mfl1<<std::endl;
-  os << latt4[0]*latt4[1]*latt4[2]*latt4[3]<< " "<<Ls<<" "<< latt4[0] <<" " <<latt4[2]<< " "
-     << mfc<<" "
-     << mfa<<" "
-     << mfo<<" "
-     << mfl1<<std::endl;
-  */
-
-#if 0
-  for(int i=0;i< PerformanceCounter::NumTypes(); i++ ){
-    Dw.DhopOE(srce,resulta,0);
-    PerformanceCounter Counter(i);
-    Counter.Start();
-    Dw.DhopOE(srce,resulta,0);
-    Counter.Stop();
-    Counter.Report();
-  }
-#endif
-  //resulta = (-0.5) * resulta;
-
-  diff = resulto-resulta;
-  std::cout<<norm2(diff)<<std::endl;
-  return 0;
-}
-
-

benchmarks/Makefile.am

@@ -1 +1,11 @@
 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

benchmarks/simple_simd_test.cc

@@ -0,0 +1,11 @@
+#include <Grid/Grid.h>
+
+Grid::vRealD add(const Grid::vRealD &x, const Grid::vRealD &y)
+{
+	return x+y;
+}
+
+Grid::vRealD sub(const Grid::vRealD &x, const Grid::vRealD &y)
+{
+	return x-y;
+}

benchmarks/simple_su3_expr.cc

@@ -25,7 +25,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#include <Grid.h>
+#include <Grid/Grid.h>
 
 using namespace std;
 using namespace Grid;

benchmarks/simple_su3_test.cc

@@ -25,7 +25,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#include <Grid.h>
+#include <Grid/Grid.h>
 
 using namespace std;
 using namespace Grid;

bootstrap.sh

@@ -1,18 +1,12 @@
 #!/usr/bin/env bash
 
 EIGEN_URL='http://bitbucket.org/eigen/eigen/get/3.2.9.tar.bz2'
-FFTW_URL=http://www.fftw.org/fftw-3.3.4.tar.gz
 
 echo "-- deploying Eigen source..."
 wget ${EIGEN_URL} --no-check-certificate
 ./scripts/update_eigen.sh `basename ${EIGEN_URL}`
 rm `basename ${EIGEN_URL}`
 
-echo "-- copying fftw prototypes..."
-wget ${FFTW_URL}
-./scripts/update_fftw.sh `basename ${FFTW_URL}`
-rm `basename ${FFTW_URL}`
-
 echo '-- generating Make.inc files...'
 ./scripts/filelist
 echo '-- generating configure script...'

configure.ac

@@ -1,30 +1,41 @@
 AC_PREREQ([2.63])
-AC_INIT([Grid], [0.5.1-dev], [https://github.com/paboyle/Grid], [Grid])
+AC_INIT([Grid], [0.6.0], [https://github.com/paboyle/Grid], [Grid])
 AC_CANONICAL_BUILD
 AC_CANONICAL_HOST
 AC_CANONICAL_TARGET
 AM_INIT_AUTOMAKE(subdir-objects)
 AC_CONFIG_MACRO_DIR([m4])
 AC_CONFIG_SRCDIR([lib/Grid.h])
-AC_CONFIG_HEADERS([lib/Config.h])
+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
-AC_LANG(C++)
 CXXFLAGS="-O3 $CXXFLAGS"
 AC_PROG_CXX
 AC_PROG_RANLIB
 
-############ openmp  ###############
+############### Get compiler informations
+AC_LANG([C++])
+AX_CXX_COMPILE_STDCXX_11([noext],[mandatory])
+AX_COMPILER_VENDOR
+AC_DEFINE_UNQUOTED([CXX_COMP_VENDOR],["$ax_cv_cxx_compiler_vendor"],
+      [vendor of C++ compiler that will compile the code])
+AX_GXX_VERSION
+AC_DEFINE_UNQUOTED([GXX_VERSION],["$GXX_VERSION"],
+      [version of g++ that will compile the code])
+
+############### Checks for typedefs, structures, and compiler characteristics
+AC_TYPE_SIZE_T
+AC_TYPE_UINT32_T
+AC_TYPE_UINT64_T
+
+############### OpenMP 
 AC_OPENMP
-
 ac_openmp=no
-
 if test "${OPENMP_CXXFLAGS}X" != "X"; then
-ac_openmp=yes
-AM_CXXFLAGS="$OPENMP_CXXFLAGS $AM_CXXFLAGS"
-AM_LDFLAGS="$OPENMP_CXXFLAGS $AM_LDFLAGS"
+  ac_openmp=yes
+  AM_CXXFLAGS="$OPENMP_CXXFLAGS $AM_CXXFLAGS"
+  AM_LDFLAGS="$OPENMP_CXXFLAGS $AM_LDFLAGS"
 fi
 
 ############### Checks for header files
@@ -37,12 +48,7 @@ AC_CHECK_HEADERS(execinfo.h)
 AC_CHECK_DECLS([ntohll],[], [], [[#include <arpa/inet.h>]])
 AC_CHECK_DECLS([be64toh],[], [], [[#include <arpa/inet.h>]])
 
-############### Checks for typedefs, structures, and compiler characteristics
-AC_TYPE_SIZE_T
-AC_TYPE_UINT32_T
-AC_TYPE_UINT64_T
-
-############### GMP and MPFR #################
+############### GMP and MPFR
 AC_ARG_WITH([gmp],
     [AS_HELP_STRING([--with-gmp=prefix],
     [try this for a non-standard install prefix of the GMP library])],
@@ -54,10 +60,17 @@ AC_ARG_WITH([mpfr],
     [AM_CXXFLAGS="-I$with_mpfr/include $AM_CXXFLAGS"]
     [AM_LDFLAGS="-L$with_mpfr/lib $AM_LDFLAGS"])
 
-################## lapack ####################
+############### 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 $AM_LDFLAGS"])
+
+############### lapack 
 AC_ARG_ENABLE([lapack],
     [AC_HELP_STRING([--enable-lapack=yes|no|prefix], [enable LAPACK])], 
-    [ac_LAPACK=${enable_lapack}],[ac_LAPACK=no])
+    [ac_LAPACK=${enable_lapack}], [ac_LAPACK=no])
 
 case ${ac_LAPACK} in
     no)
@@ -67,10 +80,33 @@ case ${ac_LAPACK} in
     *)
         AM_CXXFLAGS="-I$ac_LAPACK/include $AM_CXXFLAGS"
         AM_LDFLAGS="-L$ac_LAPACK/lib $AM_LDFLAGS"
-        AC_DEFINE([USE_LAPACK],[1],[use LAPACK])
+        AC_DEFINE([USE_LAPACK],[1],[use LAPACK]);;
 esac
 
-################## first-touch ####################
+############### MKL
+AC_ARG_ENABLE([mkl],
+    [AC_HELP_STRING([--enable-mkl=yes|no|prefix], [enable Intel MKL for LAPACK & FFTW])],
+    [ac_MKL=${enable_mkl}], [ac_MKL=no])
+
+case ${ac_MKL} in
+    no)
+        ;;
+    yes)
+        AC_DEFINE([USE_MKL], [1], [Define to 1 if you use the Intel MKL]);;
+    *)
+        AM_CXXFLAGS="-I$ac_MKL/include $AM_CXXFLAGS"
+        AM_LDFLAGS="-L$ac_MKL/lib $AM_LDFLAGS"
+        AC_DEFINE([USE_MKL], [1], [Define to 1 if you use the Intel MKL]);;
+esac
+
+############### HDF5
+AC_ARG_WITH([hdf5],
+    [AS_HELP_STRING([--with-hdf5=prefix],
+    [try this for a non-standard install prefix of the HDF5 library])],
+    [AM_CXXFLAGS="-I$with_hdf5/include $AM_CXXFLAGS"]
+    [AM_LDFLAGS="-L$with_hdf5/lib $AM_LDFLAGS"])
+
+############### first-touch
 AC_ARG_ENABLE([numa],
     [AC_HELP_STRING([--enable-numa=yes|no|prefix], [enable first touch numa opt])], 
     [ac_NUMA=${enable_NUMA}],[ac_NUMA=no])
@@ -84,56 +120,56 @@ case ${ac_NUMA} in
         AC_DEFINE([GRID_NUMA],[1],[First touch numa locality]);;
 esac
 
-################## FFTW3 ####################
-AC_ARG_WITH([fftw],    
-            [AS_HELP_STRING([--with-fftw=prefix],
-            [try this for a non-standard install prefix of the FFTW3 library])],
-            [AM_CXXFLAGS="-I$with_fftw/include $AM_CXXFLAGS"]
-            [AM_LDFLAGS="-L$with_fftw/lib $AM_LDFLAGS"])
-
-################ Get compiler informations
-AC_LANG([C++])
-AX_CXX_COMPILE_STDCXX_11([noext],[mandatory])
-AX_COMPILER_VENDOR
-AC_DEFINE_UNQUOTED([CXX_COMP_VENDOR],["$ax_cv_cxx_compiler_vendor"],
-      [vendor of C++ compiler that will compile the code])
-AX_GXX_VERSION
-AC_DEFINE_UNQUOTED([GXX_VERSION],["$GXX_VERSION"],
-      [version of g++ that will compile the code])
-
 ############### Checks for library functions
 CXXFLAGS_CPY=$CXXFLAGS
 LDFLAGS_CPY=$LDFLAGS
 CXXFLAGS="$AM_CXXFLAGS $CXXFLAGS"
 LDFLAGS="$AM_LDFLAGS $LDFLAGS"
+
 AC_CHECK_FUNCS([gettimeofday])
-AC_CHECK_LIB([gmp],[__gmpf_init],
-             [AC_CHECK_LIB([mpfr],[mpfr_init],
-                 [AC_DEFINE([HAVE_LIBMPFR], [1], [Define to 1 if you have the `MPFR' library (-lmpfr).])]
-                 [have_mpfr=true]
-                 [LIBS="$LIBS -lmpfr"],
-                 [AC_MSG_ERROR([MPFR library not found])])]
-   	     [AC_DEFINE([HAVE_LIBGMP], [1], [Define to 1 if you have the `GMP' library (-lgmp).])]
-             [have_gmp=true]
-             [LIBS="$LIBS -lgmp"],
-             [AC_MSG_WARN([**** GMP library not found, Grid can still compile but RHMC will not work ****])])
+
+if test "${ac_MKL}x" != "nox"; then
+    AC_SEARCH_LIBS([mkl_set_interface_layer], [mkl_rt], [],
+                   [AC_MSG_ERROR("MKL enabled but library not found")])
+fi
+
+AC_SEARCH_LIBS([__gmpf_init], [gmp],
+               [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])]
+               [have_gmp=true])
 
 if test "${ac_LAPACK}x" != "nox"; then
-    AC_CHECK_LIB([lapack],[LAPACKE_sbdsdc],[],
-                 [AC_MSG_ERROR("LAPACK enabled but library not found")])
-fi
-AC_CHECK_LIB([fftw3],[fftw_execute],
-  [AC_DEFINE([HAVE_FFTW],[1],[Define to 1 if you have the `FFTW' library (-lfftw3).])]
-  [have_fftw=true]
-  [LIBS="$LIBS -lfftw3 -lfftw3f"],
-  [AC_MSG_WARN([**** FFTW library not found, Grid can still compile but FFT-based routines will not work ****])])
+    AC_SEARCH_LIBS([LAPACKE_sbdsdc], [lapack], [],
+                   [AC_MSG_ERROR("LAPACK enabled but library not found")])
+fi   
+
+AC_SEARCH_LIBS([fftw_execute], [fftw3],
+               [AC_SEARCH_LIBS([fftwf_execute], [fftw3f], [],
+                               [AC_MSG_ERROR("single precision FFTW library not found")])]
+               [AC_DEFINE([HAVE_FFTW], [1], [Define to 1 if you have the `FFTW' library])]
+               [have_fftw=true])
+
+AC_SEARCH_LIBS([H5Fopen], [hdf5_cpp],
+               [AC_DEFINE([HAVE_HDF5], [1], [Define to 1 if you have the `HDF5' library])]
+               [have_hdf5=true]
+               [LIBS="${LIBS} -lhdf5"], [], [-lhdf5])
+AM_CONDITIONAL(BUILD_HDF5, [ test "${have_hdf5}X" == "trueX" ])
+
 CXXFLAGS=$CXXFLAGS_CPY
 LDFLAGS=$LDFLAGS_CPY
 
 ############### SIMD instruction selection
-AC_ARG_ENABLE([simd],[AC_HELP_STRING([--enable-simd=SSE4|AVX|AVXFMA4|AVXFMA|AVX2|AVX512|AVX512MIC|IMCI|KNL|KNC],\
-	[Select instructions to be SSE4.0, AVX 1.0, AVX 2.0+FMA, AVX 512, IMCI])],\
-	[ac_SIMD=${enable_simd}],[ac_SIMD=GEN])
+AC_ARG_ENABLE([simd],[AC_HELP_STRING([--enable-simd=code],
+              [select SIMD target (cf. README.md)])], [ac_SIMD=${enable_simd}], [ac_SIMD=GEN])
+
+AC_ARG_ENABLE([gen-simd-width],
+            [AS_HELP_STRING([--enable-gen-simd-width=size],
+            [size (in bytes) of the generic SIMD vectors (default: 32)])],
+            [ac_gen_simd_width=$enable_gen_simd_width],
+            [ac_gen_simd_width=32])
 
 case ${ax_cv_cxx_compiler_vendor} in
   clang|gnu)
@@ -153,14 +189,20 @@ case ${ax_cv_cxx_compiler_vendor} in
       AVX2)
         AC_DEFINE([AVX2],[1],[AVX2 intrinsics])
         SIMD_FLAGS='-mavx2 -mfma';;
-      AVX512|AVX512MIC|KNL)
+      AVX512)
         AC_DEFINE([AVX512],[1],[AVX512 intrinsics])
         SIMD_FLAGS='-mavx512f -mavx512pf -mavx512er -mavx512cd';;
-      IMCI|KNC)
+      KNC)
         AC_DEFINE([IMCI],[1],[IMCI intrinsics for Knights Corner])
         SIMD_FLAGS='';;
+      KNL)
+        AC_DEFINE([AVX512],[1],[AVX512 intrinsics])
+        SIMD_FLAGS='-march=knl';;
       GEN)
-        AC_DEFINE([GENERIC_VEC],[1],[generic vector code])
+        AC_DEFINE([GEN],[1],[generic vector code])
+        AC_DEFINE_UNQUOTED([GEN_SIMD_WIDTH],[$ac_gen_simd_width],
+                           [generic SIMD vector width (in bytes)])
+        SIMD_GEN_WIDTH_MSG=" (width= $ac_gen_simd_width)"
         SIMD_FLAGS='';;
       QPX|BGQ)
         AC_DEFINE([QPX],[1],[QPX intrinsics for BG/Q])
@@ -176,26 +218,26 @@ case ${ax_cv_cxx_compiler_vendor} in
       AVX)
         AC_DEFINE([AVX1],[1],[AVX intrinsics])
         SIMD_FLAGS='-mavx -xavx';;
-      AVXFMA4)
-        AC_DEFINE([AVXFMA4],[1],[AVX intrinsics with FMA4])
-        SIMD_FLAGS='-mavx -mfma';;
       AVXFMA)
-        AC_DEFINE([AVXFMA],[1],[AVX intrinsics with FMA4])
-        SIMD_FLAGS='-mavx -mfma';;
+        AC_DEFINE([AVXFMA],[1],[AVX intrinsics with FMA3])
+        SIMD_FLAGS='-mavx -fma';;
       AVX2)
         AC_DEFINE([AVX2],[1],[AVX2 intrinsics])
         SIMD_FLAGS='-march=core-avx2 -xcore-avx2';;
       AVX512)
         AC_DEFINE([AVX512],[1],[AVX512 intrinsics])
         SIMD_FLAGS='-xcore-avx512';;
-      AVX512MIC|KNL)
-        AC_DEFINE([AVX512],[1],[AVX512 intrinsics for Knights Landing])
-        SIMD_FLAGS='-xmic-avx512';;
-      IMCI|KNC)
+      KNC)
         AC_DEFINE([IMCI],[1],[IMCI Intrinsics for Knights Corner])
         SIMD_FLAGS='';;
+      KNL)
+        AC_DEFINE([AVX512],[1],[AVX512 intrinsics for Knights Landing])
+        SIMD_FLAGS='-xmic-avx512';;
       GEN)
-        AC_DEFINE([GENERIC_VEC],[1],[generic vector code])
+        AC_DEFINE([GEN],[1],[generic vector code])
+        AC_DEFINE_UNQUOTED([GEN_SIMD_WIDTH],[$ac_gen_simd_width],
+                           [generic SIMD vector width (in bytes)])
+        SIMD_GEN_WIDTH_MSG=" (width= $ac_gen_simd_width)"
         SIMD_FLAGS='';;
       *)
         AC_MSG_ERROR(["SIMD option ${ac_SIMD} not supported by the Intel compiler"]);;
@@ -208,14 +250,18 @@ AM_CXXFLAGS="$SIMD_FLAGS $AM_CXXFLAGS"
 AM_CFLAGS="$SIMD_FLAGS $AM_CFLAGS"
 
 case ${ac_SIMD} in
-  AVX512|AVX512MIC|KNL)
+  AVX512|KNL)
     AC_DEFINE([TEST_ZMM],[1],[compile ZMM test]);;
   *)
 	;;
 esac
 
-############### precision selection
-AC_ARG_ENABLE([precision],[AC_HELP_STRING([--enable-precision=single|double],[Select default word size of Real])],[ac_PRECISION=${enable_precision}],[ac_PRECISION=double])
+############### Precision selection
+AC_ARG_ENABLE([precision],
+              [AC_HELP_STRING([--enable-precision=single|double],
+                              [Select default word size of Real])],
+              [ac_PRECISION=${enable_precision}],[ac_PRECISION=double])
+
 case ${ac_PRECISION} in
      single)
        AC_DEFINE([GRID_DEFAULT_PRECISION_SINGLE],[1],[GRID_DEFAULT_PRECISION is SINGLE] )
@@ -226,39 +272,56 @@ case ${ac_PRECISION} in
 esac
 
 ############### communication type selection
-AC_ARG_ENABLE([comms],[AC_HELP_STRING([--enable-comms=none|mpi|mpi-auto|shmem],[Select communications])],[ac_COMMS=${enable_comms}],[ac_COMMS=none])
+AC_ARG_ENABLE([comms],[AC_HELP_STRING([--enable-comms=none|mpi|mpi-auto|mpi3|mpi3-auto|shmem],
+              [Select communications])],[ac_COMMS=${enable_comms}],[ac_COMMS=none])
 
 case ${ac_COMMS} in
      none)
-       AC_DEFINE([GRID_COMMS_NONE],[1],[GRID_COMMS_NONE] )
+        AC_DEFINE([GRID_COMMS_NONE],[1],[GRID_COMMS_NONE] )
+        comms_type='none'
      ;;
-     mpi-auto)
-       AC_DEFINE([GRID_COMMS_MPI],[1],[GRID_COMMS_MPI] )
-       LX_FIND_MPI
-       if test "x$have_CXX_mpi" = 'xno'; then AC_MSG_ERROR(["MPI not found"]); fi
-       AM_CXXFLAGS="$MPI_CXXFLAGS $AM_CXXFLAGS"
-       AM_CFLAGS="$MPI_CFLAGS $AM_CFLAGS"
-       AM_LDFLAGS="`echo $MPI_CXXLDFLAGS | sed -E 's/-l@<:@^ @:>@+//g'` $AM_LDFLAGS"
-       LIBS="`echo $MPI_CXXLDFLAGS | sed -E 's/-L@<:@^ @:>@+//g'` $LIBS"
+     mpi3l*)
+       AC_DEFINE([GRID_COMMS_MPI3L],[1],[GRID_COMMS_MPI3L] )
+       comms_type='mpi3l'
      ;;
-     mpi)
-       AC_DEFINE([GRID_COMMS_MPI],[1],[GRID_COMMS_MPI] )
+     mpi3*)
+        AC_DEFINE([GRID_COMMS_MPI3],[1],[GRID_COMMS_MPI3] )
+        comms_type='mpi3'
+     ;;
+     mpi*)
+        AC_DEFINE([GRID_COMMS_MPI],[1],[GRID_COMMS_MPI] )
+        comms_type='mpi'
      ;;
      shmem)
-       AC_DEFINE([GRID_COMMS_SHMEM],[1],[GRID_COMMS_SHMEM] )
+        AC_DEFINE([GRID_COMMS_SHMEM],[1],[GRID_COMMS_SHMEM] )
+        comms_type='shmem'
      ;;
      *)
-     AC_MSG_ERROR([${ac_COMMS} unsupported --enable-comms option]); 
+        AC_MSG_ERROR([${ac_COMMS} unsupported --enable-comms option]); 
      ;;
 esac
-AM_CONDITIONAL(BUILD_COMMS_SHMEM,[ test "X${ac_COMMS}X" == "XshmemX" ])
-AM_CONDITIONAL(BUILD_COMMS_MPI,[ test "X${ac_COMMS}X" == "XmpiX" || test "X${ac_COMMS}X" == "Xmpi-autoX" ])
-AM_CONDITIONAL(BUILD_COMMS_NONE,[ test "X${ac_COMMS}X" == "XnoneX" ])
+case ${ac_COMMS} in
+    *-auto)
+        LX_FIND_MPI
+        if test "x$have_CXX_mpi" = 'xno'; then AC_MSG_ERROR(["The configure could not find the MPI compilation flags. N.B. The -auto mode is not supported by Cray wrappers. Use the non -auto version in this case."]); fi
+        AM_CXXFLAGS="$MPI_CXXFLAGS $AM_CXXFLAGS"
+        AM_CFLAGS="$MPI_CFLAGS $AM_CFLAGS"
+        AM_LDFLAGS="`echo $MPI_CXXLDFLAGS | sed -E 's/-l@<:@^ @:>@+//g'` $AM_LDFLAGS"
+        LIBS="`echo $MPI_CXXLDFLAGS | sed -E 's/-L@<:@^ @:>@+//g'` $LIBS";;
+    *)
+        ;;
+esac
+
+AM_CONDITIONAL(BUILD_COMMS_SHMEM, [ test "${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 "${comms_type}X" == "mpi3lX" ] )
+AM_CONDITIONAL(BUILD_COMMS_NONE,  [ test "${comms_type}X" == "noneX" ])
 
 ############### RNG selection
-AC_ARG_ENABLE([rng],[AC_HELP_STRING([--enable-rng=ranlux48|mt19937],\
-	[Select Random Number Generator to be used])],\
-	[ac_RNG=${enable_rng}],[ac_RNG=ranlux48])
+AC_ARG_ENABLE([rng],[AC_HELP_STRING([--enable-rng=ranlux48|mt19937|sitmo],\
+	            [Select Random Number Generator to be used])],\
+	            [ac_RNG=${enable_rng}],[ac_RNG=ranlux48])
 
 case ${ac_RNG} in
      ranlux48)
@@ -267,15 +330,19 @@ case ${ac_RNG} in
      mt19937)
       AC_DEFINE([RNG_MT19937],[1],[RNG_MT19937] )
      ;;
+     sitmo)
+      AC_DEFINE([RNG_SITMO],[1],[RNG_SITMO] )
+     ;;
      *)
       AC_MSG_ERROR([${ac_RNG} unsupported --enable-rng option]); 
      ;;
 esac
 
-############### timer option
+############### Timer option
 AC_ARG_ENABLE([timers],[AC_HELP_STRING([--enable-timers],\
-	[Enable system dependent high res timers])],\
-	[ac_TIMERS=${enable_timers}],[ac_TIMERS=yes])
+	            [Enable system dependent high res timers])],\
+	            [ac_TIMERS=${enable_timers}],[ac_TIMERS=yes])
+
 case ${ac_TIMERS} in
      yes)
       AC_DEFINE([TIMERS_ON],[1],[TIMERS_ON] )
@@ -289,7 +356,9 @@ case ${ac_TIMERS} in
 esac
 
 ############### Chroma regression test
-AC_ARG_ENABLE([chroma],[AC_HELP_STRING([--enable-chroma],[Expect chroma compiled under c++11 ])],ac_CHROMA=yes,ac_CHROMA=no)
+AC_ARG_ENABLE([chroma],[AC_HELP_STRING([--enable-chroma],
+              [Expect chroma compiled under c++11 ])],ac_CHROMA=yes,ac_CHROMA=no)
+
 case ${ac_CHROMA} in
      yes|no)
      ;;
@@ -297,15 +366,21 @@ case ${ac_CHROMA} in
        AC_MSG_ERROR([${ac_CHROMA} unsupported --enable-chroma option]); 
      ;;
 esac
+
 AM_CONDITIONAL(BUILD_CHROMA_REGRESSION,[ test "X${ac_CHROMA}X" == "XyesX" ])
 
 ############### Doxygen
-AC_PROG_DOXYGEN
-
-if test -n "$DOXYGEN"
-then
-AC_CONFIG_FILES([docs/doxy.cfg])
-fi
+DX_DOXYGEN_FEATURE([OFF])
+DX_DOT_FEATURE([OFF])
+DX_HTML_FEATURE([ON])
+DX_CHM_FEATURE([OFF])
+DX_CHI_FEATURE([OFF])
+DX_MAN_FEATURE([OFF])
+DX_RTF_FEATURE([OFF])
+DX_XML_FEATURE([OFF])
+DX_PDF_FEATURE([OFF])
+DX_PS_FEATURE([OFF])
+DX_INIT_DOXYGEN([$PACKAGE_NAME], [doxygen.cfg])
 
 ############### Ouput
 cwd=`pwd -P`; cd ${srcdir}; abs_srcdir=`pwd -P`; cd ${cwd}
@@ -322,41 +397,45 @@ AC_CONFIG_FILES(tests/IO/Makefile)
 AC_CONFIG_FILES(tests/core/Makefile)
 AC_CONFIG_FILES(tests/debug/Makefile)
 AC_CONFIG_FILES(tests/forces/Makefile)
+AC_CONFIG_FILES(tests/hadrons/Makefile)
 AC_CONFIG_FILES(tests/hmc/Makefile)
 AC_CONFIG_FILES(tests/solver/Makefile)
 AC_CONFIG_FILES(tests/qdpxx/Makefile)
 AC_CONFIG_FILES(benchmarks/Makefile)
+AC_CONFIG_FILES(extras/Makefile)
+AC_CONFIG_FILES(extras/Hadrons/Makefile)
 AC_OUTPUT
 
-echo "
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+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}
+architecture (build)        : $build_cpu
+os (build)                  : $build_os
+architecture (target)       : $target_cpu
+os (target)                 : $target_os
+compiler vendor             : ${ax_cv_cxx_compiler_vendor}
+compiler version            : ${ax_cv_gxx_version}
 ----- BUILD OPTIONS -----------------------------------
-- SIMD                          : ${ac_SIMD}
-- Threading                     : ${ac_openmp} 
-- Communications type           : ${ac_COMMS}
-- Default precision             : ${ac_PRECISION}
-- RNG choice                    : ${ac_RNG} 
-- GMP                           : `if test "x$have_gmp" = xtrue; then echo yes; else echo no; fi`
-- LAPACK                        : ${ac_LAPACK}
-- FFTW                          : `if test "x$have_fftw" = xtrue; then echo yes; else echo no; fi`
-- build DOXYGEN documentation   : `if test "x$enable_doc" = xyes; then echo yes; else echo no; fi`
-- graphs and diagrams           : `if test "x$enable_dot" = xyes; then echo yes; else echo no; fi`
+SIMD                        : ${ac_SIMD}${SIMD_GEN_WIDTH_MSG}
+Threading                   : ${ac_openmp} 
+Communications type         : ${comms_type}
+Default precision           : ${ac_PRECISION}
+RNG choice                  : ${ac_RNG} 
+GMP                         : `if test "x$have_gmp" = xtrue; then echo yes; else echo no; fi`
+LAPACK                      : ${ac_LAPACK}
+FFTW                        : `if test "x$have_fftw" = xtrue; then echo yes; else echo no; fi`
+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:
+CXXFLAGS:
 `echo ${AM_CXXFLAGS} ${CXXFLAGS} | tr ' ' '\n' | sed 's/^-/    -/g'`
-- LDFLAGS:
+LDFLAGS:
 `echo ${AM_LDFLAGS} ${LDFLAGS} | tr ' ' '\n' | sed 's/^-/    -/g'`
-- LIBS:
+LIBS:
 `echo ${LIBS} | tr ' ' '\n' | sed 's/^-/    -/g'`
--------------------------------------------------------
-"
+-------------------------------------------------------" > config.summary
+echo ""
+cat config.summary
+echo ""

doxygen.inc

@@ -0,0 +1,184 @@
+# Copyright (C) 2004 Oren Ben-Kiki
+# This file is distributed under the same terms as the Automake macro files.
+
+# Generate automatic documentation using Doxygen. Goals and variables values
+# are controlled by the various DX_COND_??? conditionals set by autoconf.
+#
+# The provided goals are:
+# doxygen-doc: Generate all doxygen documentation.
+# doxygen-run: Run doxygen, which will generate some of the documentation
+#              (HTML, CHM, CHI, MAN, RTF, XML) but will not do the post
+#              processing required for the rest of it (PS, PDF, and some MAN).
+# doxygen-man: Rename some doxygen generated man pages.
+# doxygen-ps: Generate doxygen PostScript documentation.
+# doxygen-pdf: Generate doxygen PDF documentation.
+#
+# Note that by default these are not integrated into the automake goals. If
+# doxygen is used to generate man pages, you can achieve this integration by
+# setting man3_MANS to the list of man pages generated and then adding the
+# dependency:
+#
+#   $(man3_MANS): doxygen-doc
+#
+# This will cause make to run doxygen and generate all the documentation.
+#
+# The following variable is intended for use in Makefile.am:
+#
+# DX_CLEANFILES = everything to clean.
+#
+# This is usually added to MOSTLYCLEANFILES.
+
+## --------------------------------- ##
+## Format-independent Doxygen rules. ##
+## --------------------------------- ##
+
+if DX_COND_doc
+
+## ------------------------------- ##
+## Rules specific for HTML output. ##
+## ------------------------------- ##
+
+if DX_COND_html
+
+DX_CLEAN_HTML = @DX_DOCDIR@/html
+
+endif DX_COND_html
+
+## ------------------------------ ##
+## Rules specific for CHM output. ##
+## ------------------------------ ##
+
+if DX_COND_chm
+
+DX_CLEAN_CHM = @DX_DOCDIR@/chm
+
+if DX_COND_chi
+
+DX_CLEAN_CHI = @DX_DOCDIR@/@PACKAGE@.chi
+
+endif DX_COND_chi
+
+endif DX_COND_chm
+
+## ------------------------------ ##
+## Rules specific for MAN output. ##
+## ------------------------------ ##
+
+if DX_COND_man
+
+DX_CLEAN_MAN = @DX_DOCDIR@/man
+
+endif DX_COND_man
+
+## ------------------------------ ##
+## Rules specific for RTF output. ##
+## ------------------------------ ##
+
+if DX_COND_rtf
+
+DX_CLEAN_RTF = @DX_DOCDIR@/rtf
+
+endif DX_COND_rtf
+
+## ------------------------------ ##
+## Rules specific for XML output. ##
+## ------------------------------ ##
+
+if DX_COND_xml
+
+DX_CLEAN_XML = @DX_DOCDIR@/xml
+
+endif DX_COND_xml
+
+## ----------------------------- ##
+## Rules specific for PS output. ##
+## ----------------------------- ##
+
+if DX_COND_ps
+
+DX_CLEAN_PS = @DX_DOCDIR@/@PACKAGE@.ps
+
+DX_PS_GOAL = doxygen-ps
+
+doxygen-ps: @DX_DOCDIR@/@PACKAGE@.ps
+
+@DX_DOCDIR@/@PACKAGE@.ps: @DX_DOCDIR@/@PACKAGE@.tag
+	cd @DX_DOCDIR@/latex; \
+	rm -f *.aux *.toc *.idx *.ind *.ilg *.log *.out; \
+	$(DX_LATEX) refman.tex; \
+	$(MAKEINDEX_PATH) refman.idx; \
+	$(DX_LATEX) refman.tex; \
+	countdown=5; \
+	while $(DX_EGREP) 'Rerun (LaTeX|to get cross-references right)' \
+	                  refman.log > /dev/null 2>&1 \
+	   && test $$countdown -gt 0; do \
+	    $(DX_LATEX) refman.tex; \
+	    countdown=`expr $$countdown - 1`; \
+	done; \
+	$(DX_DVIPS) -o ../@PACKAGE@.ps refman.dvi
+
+endif DX_COND_ps
+
+## ------------------------------ ##
+## Rules specific for PDF output. ##
+## ------------------------------ ##
+
+if DX_COND_pdf
+
+DX_CLEAN_PDF = @DX_DOCDIR@/@PACKAGE@.pdf
+
+DX_PDF_GOAL = doxygen-pdf
+
+doxygen-pdf: @DX_DOCDIR@/@PACKAGE@.pdf
+
+@DX_DOCDIR@/@PACKAGE@.pdf: @DX_DOCDIR@/@PACKAGE@.tag
+	cd @DX_DOCDIR@/latex; \
+	rm -f *.aux *.toc *.idx *.ind *.ilg *.log *.out; \
+	$(DX_PDFLATEX) refman.tex; \
+	$(DX_MAKEINDEX) refman.idx; \
+	$(DX_PDFLATEX) refman.tex; \
+	countdown=5; \
+	while $(DX_EGREP) 'Rerun (LaTeX|to get cross-references right)' \
+	                  refman.log > /dev/null 2>&1 \
+	   && test $$countdown -gt 0; do \
+	    $(DX_PDFLATEX) refman.tex; \
+	    countdown=`expr $$countdown - 1`; \
+	done; \
+	mv refman.pdf ../@PACKAGE@.pdf
+
+endif DX_COND_pdf
+
+## ------------------------------------------------- ##
+## Rules specific for LaTeX (shared for PS and PDF). ##
+## ------------------------------------------------- ##
+
+if DX_COND_latex
+
+DX_CLEAN_LATEX = @DX_DOCDIR@/latex
+
+endif DX_COND_latex
+
+.INTERMEDIATE: doxygen-run $(DX_PS_GOAL) $(DX_PDF_GOAL)
+
+doxygen-run: @DX_DOCDIR@/@PACKAGE@.tag
+
+doxygen-doc: doxygen-run $(DX_PS_GOAL) $(DX_PDF_GOAL)
+
+@DX_DOCDIR@/@PACKAGE@.tag: $(DX_CONFIG) $(pkginclude_HEADERS)
+	rm -rf @DX_DOCDIR@
+	$(DX_ENV) $(DX_DOXYGEN) $(srcdir)/$(DX_CONFIG)
+
+DX_CLEANFILES = \
+    @DX_DOCDIR@/@PACKAGE@.tag \
+    -r \
+    $(DX_CLEAN_HTML) \
+    $(DX_CLEAN_CHM) \
+    $(DX_CLEAN_CHI) \
+    $(DX_CLEAN_MAN) \
+    $(DX_CLEAN_RTF) \
+    $(DX_CLEAN_XML) \
+    $(DX_CLEAN_PS) \
+    $(DX_CLEAN_PDF) \
+    $(DX_CLEAN_LATEX)
+
+endif DX_COND_doc

extras/Hadrons/Application.cc

@@ -0,0 +1,317 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Application.cc
+
+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/Application.hpp>
+#include <Grid/Hadrons/GeneticScheduler.hpp>
+
+using namespace Grid;
+using namespace QCD;
+using namespace Hadrons;
+
+#define BIG_SEP "==============="
+#define SEP     "---------------"
+
+/******************************************************************************
+ *                       Application implementation                           *
+ ******************************************************************************/
+// constructors ////////////////////////////////////////////////////////////////
+Application::Application(void)
+{
+    LOG(Message) << "Modules available:" << std::endl;
+    auto list = ModuleFactory::getInstance().getBuilderList();
+    for (auto &m: list)
+    {
+        LOG(Message) << "  " << m << std::endl;
+    }
+    auto dim = GridDefaultLatt(), mpi = GridDefaultMpi(), loc(dim);
+    locVol_ = 1;
+    for (unsigned int d = 0; d < dim.size(); ++d)
+    {
+        loc[d]  /= mpi[d];
+        locVol_ *= loc[d];
+    }
+    LOG(Message) << "Global lattice: " << dim << std::endl;
+    LOG(Message) << "MPI partition : " << mpi << std::endl;
+    LOG(Message) << "Local lattice : " << loc << std::endl;
+}
+
+Application::Application(const Application::GlobalPar &par)
+: Application()
+{
+    setPar(par);
+}
+
+Application::Application(const std::string parameterFileName)
+: Application()
+{
+    parameterFileName_ = parameterFileName;
+}
+
+// environment shortcut ////////////////////////////////////////////////////////
+Environment & Application::env(void) const
+{
+    return Environment::getInstance();
+}
+
+// access //////////////////////////////////////////////////////////////////////
+void Application::setPar(const Application::GlobalPar &par)
+{
+    par_ = par;
+    env().setSeed(strToVec<int>(par_.seed));
+}
+
+const Application::GlobalPar & Application::getPar(void)
+{
+    return par_;
+}
+
+// execute /////////////////////////////////////////////////////////////////////
+void Application::run(void)
+{
+    if (!parameterFileName_.empty() and (env().getNModule() == 0))
+    {
+        parseParameterFile(parameterFileName_);
+    }
+    if (!scheduled_)
+    {
+        schedule();
+    }
+    printSchedule();
+    configLoop();
+}
+
+// parse parameter file ////////////////////////////////////////////////////////
+class ObjectId: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(ObjectId,
+                                    std::string, name,
+                                    std::string, type);
+};
+
+void Application::parseParameterFile(const std::string parameterFileName)
+{
+    XmlReader reader(parameterFileName);
+    GlobalPar par;
+    ObjectId  id;
+    
+    LOG(Message) << "Building application from '" << parameterFileName << "'..." << std::endl;
+    read(reader, "parameters", par);
+    setPar(par);
+    push(reader, "modules");
+    push(reader, "module");
+    do
+    {
+        read(reader, "id", id);
+        env().createModule(id.name, id.type, reader);
+    } while (reader.nextElement("module"));
+    pop(reader);
+    pop(reader);
+}
+
+void Application::saveParameterFile(const std::string parameterFileName)
+{
+    XmlWriter          writer(parameterFileName);
+    ObjectId           id;
+    const unsigned int nMod = env().getNModule();
+    
+    LOG(Message) << "Saving application to '" << parameterFileName << "'..." << std::endl;
+    write(writer, "parameters", getPar());
+    push(writer, "modules");
+    for (unsigned int i = 0; i < nMod; ++i)
+    {
+        push(writer, "module");
+        id.name = env().getModuleName(i);
+        id.type = env().getModule(i)->getRegisteredName();
+        write(writer, "id", id);
+        env().getModule(i)->saveParameters(writer, "options");
+        pop(writer);
+    }
+    pop(writer);
+    pop(writer);
+}
+
+// schedule computation ////////////////////////////////////////////////////////
+#define MEM_MSG(size)\
+sizeString((size)*locVol_) << " (" << sizeString(size)  << "/site)"
+
+#define DEFINE_MEMPEAK \
+auto memPeak = [this](const std::vector<unsigned int> &program)\
+{\
+    unsigned int memPeak;\
+    bool         msg;\
+    \
+    msg = HadronsLogMessage.isActive();\
+    HadronsLogMessage.Active(false);\
+    env().dryRun(true);\
+    memPeak = env().executeProgram(program);\
+    env().dryRun(false);\
+    env().freeAll();\
+    HadronsLogMessage.Active(true);\
+    \
+    return memPeak;\
+}
+
+void Application::schedule(void)
+{
+    DEFINE_MEMPEAK;
+    
+    // build module dependency graph
+    LOG(Message) << "Building module graph..." << std::endl;
+    auto graph = env().makeModuleGraph();
+    auto con = graph.getConnectedComponents();
+    
+    // constrained topological sort using a genetic algorithm
+    LOG(Message) << "Scheduling computation..." << std::endl;
+    LOG(Message) << "               #module= " << graph.size() << std::endl;
+    LOG(Message) << "       population size= " << par_.genetic.popSize << std::endl;
+    LOG(Message) << "       max. generation= " << par_.genetic.maxGen << std::endl;
+    LOG(Message) << "  max. cst. generation= " << par_.genetic.maxCstGen << std::endl;
+    LOG(Message) << "         mutation rate= " << par_.genetic.mutationRate << std::endl;
+    
+    unsigned int                               k = 0, gen, prevPeak, nCstPeak = 0;
+    std::random_device                         rd;
+    GeneticScheduler<unsigned int>::Parameters par;
+    
+    par.popSize      = par_.genetic.popSize;
+    par.mutationRate = par_.genetic.mutationRate;
+    par.seed         = rd();
+    memPeak_         = 0;
+    CartesianCommunicator::BroadcastWorld(0, &(par.seed), sizeof(par.seed));
+    for (unsigned int i = 0; i < con.size(); ++i)
+    {
+        GeneticScheduler<unsigned int> scheduler(con[i], memPeak, par);
+        
+        gen = 0;
+        do
+        {
+            LOG(Debug) << "Generation " << gen << ":" << std::endl;
+            scheduler.nextGeneration();
+            if (gen != 0)
+            {
+                if (prevPeak == scheduler.getMinValue())
+                {
+                    nCstPeak++;
+                }
+                else
+                {
+                    nCstPeak = 0;
+                }
+            }
+            
+            prevPeak = scheduler.getMinValue();
+            if (gen % 10 == 0)
+            {
+                LOG(Iterative) << "Generation " << gen << ": "
+                               << MEM_MSG(scheduler.getMinValue()) << std::endl;
+            }
+            
+            gen++;
+        } while ((gen < par_.genetic.maxGen)
+                 and (nCstPeak < par_.genetic.maxCstGen));
+        auto &t = scheduler.getMinSchedule();
+        if (scheduler.getMinValue() > memPeak_)
+        {
+            memPeak_ = scheduler.getMinValue();
+        }
+        for (unsigned int j = 0; j < t.size(); ++j)
+        {
+            program_.push_back(t[j]);
+        }
+    }
+    scheduled_ = true;
+}
+
+void Application::saveSchedule(const std::string filename)
+{
+    TextWriter               writer(filename);
+    std::vector<std::string> program;
+    
+    if (!scheduled_)
+    {
+        HADRON_ERROR("Computation not scheduled");
+    }
+    LOG(Message) << "Saving current schedule to '" << filename << "'..."
+                 << std::endl;
+    for (auto address: program_)
+    {
+        program.push_back(env().getModuleName(address));
+    }
+    write(writer, "schedule", program);
+}
+
+void Application::loadSchedule(const std::string filename)
+{
+    DEFINE_MEMPEAK;
+    
+    TextReader               reader(filename);
+    std::vector<std::string> program;
+    
+    LOG(Message) << "Loading schedule from '" << filename << "'..."
+                 << std::endl;
+    read(reader, "schedule", program);
+    program_.clear();
+    for (auto &name: program)
+    {
+        program_.push_back(env().getModuleAddress(name));
+    }
+    scheduled_ = true;
+    memPeak_   = memPeak(program_);
+}
+
+void Application::printSchedule(void)
+{
+    if (!scheduled_)
+    {
+        HADRON_ERROR("Computation not scheduled");
+    }
+    LOG(Message) << "Schedule (memory peak: " << MEM_MSG(memPeak_) << "):"
+                 << std::endl;
+    for (unsigned int i = 0; i < program_.size(); ++i)
+    {
+        LOG(Message) << std::setw(4) << i + 1 << ": "
+                     << env().getModuleName(program_[i]) << std::endl;
+    }
+}
+
+// loop on configurations //////////////////////////////////////////////////////
+void Application::configLoop(void)
+{
+    auto range = par_.trajCounter;
+    
+    for (unsigned int t = range.start; t < range.end; t += range.step)
+    {
+        LOG(Message) << BIG_SEP << " Starting measurement for trajectory " << t
+                     << " " << BIG_SEP << std::endl;
+        env().setTrajectory(t);
+        env().executeProgram(program_);
+    }
+    LOG(Message) << BIG_SEP << " End of measurement " << BIG_SEP << std::endl;
+    env().freeAll();
+}

extras/Hadrons/Application.hpp

@@ -0,0 +1,132 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Application.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_Application_hpp_
+#define Hadrons_Application_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Environment.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+#include <Grid/Hadrons/Modules.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                         Main program manager                               *
+ ******************************************************************************/
+class Application
+{
+public:
+    class TrajRange: Serializable
+    {
+    public:
+        GRID_SERIALIZABLE_CLASS_MEMBERS(TrajRange,
+                                        unsigned int, start,
+                                        unsigned int, end,
+                                        unsigned int, step);
+    };
+    class GeneticPar: Serializable
+    {
+    public:
+        GeneticPar(void):
+            popSize{20}, maxGen{1000}, maxCstGen{100}, mutationRate{.1} {};
+    public:
+        GRID_SERIALIZABLE_CLASS_MEMBERS(GeneticPar,
+                                        unsigned int, popSize,
+                                        unsigned int, maxGen,
+                                        unsigned int, maxCstGen,
+                                        double      , mutationRate);
+    };
+    class GlobalPar: Serializable
+    {
+    public:
+        GRID_SERIALIZABLE_CLASS_MEMBERS(GlobalPar,
+                                        TrajRange,   trajCounter,
+                                        GeneticPar,  genetic,
+                                        std::string, seed);
+    };
+public:
+    // constructors
+    Application(void);
+    Application(const GlobalPar &par);
+    Application(const std::string parameterFileName);
+    // destructor
+    virtual ~Application(void) = default;
+    // access
+    void              setPar(const GlobalPar &par);
+    const GlobalPar & getPar(void);
+    // module creation
+    template <typename M>
+    void createModule(const std::string name);
+    template <typename M>
+    void createModule(const std::string name, const typename M::Par &par);
+    // execute
+    void run(void);
+    // XML parameter file I/O
+    void parseParameterFile(const std::string parameterFileName);
+    void saveParameterFile(const std::string parameterFileName);
+    // schedule computation
+    void schedule(void);
+    void saveSchedule(const std::string filename);
+    void loadSchedule(const std::string filename);
+    void printSchedule(void);
+    // loop on configurations
+    void configLoop(void);
+private:
+    // environment shortcut
+    Environment & env(void) const;
+private:
+    long unsigned int         locVol_;
+    std::string               parameterFileName_{""};
+    GlobalPar                 par_;
+    std::vector<unsigned int> program_;
+    Environment::Size         memPeak_;
+    bool                      scheduled_{false};
+};
+
+/******************************************************************************
+ *                     Application template implementation                    *
+ ******************************************************************************/
+// module creation /////////////////////////////////////////////////////////////
+template <typename M>
+void Application::createModule(const std::string name)
+{
+    env().createModule<M>(name);
+}
+
+template <typename M>
+void Application::createModule(const std::string name,
+                               const typename M::Par &par)
+{
+    env().createModule<M>(name, par);
+}
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_Application_hpp_

extras/Hadrons/Environment.cc

@@ -0,0 +1,743 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Environment.cc
+
+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/Environment.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+using namespace Grid;
+using namespace QCD;
+using namespace Hadrons;
+
+/******************************************************************************
+ *                       Environment implementation                           *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+Environment::Environment(void)
+{
+    nd_ = GridDefaultLatt().size();
+    grid4d_.reset(SpaceTimeGrid::makeFourDimGrid(
+        GridDefaultLatt(), GridDefaultSimd(nd_, vComplex::Nsimd()),
+        GridDefaultMpi()));
+    gridRb4d_.reset(SpaceTimeGrid::makeFourDimRedBlackGrid(grid4d_.get()));
+    auto loc = getGrid()->LocalDimensions();
+    locVol_ = 1;
+    for (unsigned int d = 0; d < loc.size(); ++d)
+    {
+        locVol_ *= loc[d];
+    }
+    rng4d_.reset(new GridParallelRNG(grid4d_.get()));
+}
+
+// dry run /////////////////////////////////////////////////////////////////////
+void Environment::dryRun(const bool isDry)
+{
+    dryRun_ = isDry;
+}
+
+bool Environment::isDryRun(void) const
+{
+    return dryRun_;
+}
+
+// trajectory number ///////////////////////////////////////////////////////////
+void Environment::setTrajectory(const unsigned int traj)
+{
+    traj_ = traj;
+}
+
+unsigned int Environment::getTrajectory(void) const
+{
+    return traj_;
+}
+
+// grids ///////////////////////////////////////////////////////////////////////
+void Environment::createGrid(const unsigned int Ls)
+{
+    if (grid5d_.find(Ls) == grid5d_.end())
+    {
+        auto g = getGrid();
+        
+        grid5d_[Ls].reset(SpaceTimeGrid::makeFiveDimGrid(Ls, g));
+        gridRb5d_[Ls].reset(SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls, g));
+    }
+}
+
+GridCartesian * Environment::getGrid(const unsigned int Ls) const
+{
+    try
+    {
+        if (Ls == 1)
+        {
+            return grid4d_.get();
+        }
+        else
+        {
+            return grid5d_.at(Ls).get();
+        }
+    }
+    catch(std::out_of_range &)
+    {
+        HADRON_ERROR("no grid with Ls= " << Ls);
+    }
+}
+
+GridRedBlackCartesian * Environment::getRbGrid(const unsigned int Ls) const
+{
+    try
+    {
+        if (Ls == 1)
+        {
+            return gridRb4d_.get();
+        }
+        else
+        {
+            return gridRb5d_.at(Ls).get();
+        }
+    }
+    catch(std::out_of_range &)
+    {
+        HADRON_ERROR("no red-black 5D grid with Ls= " << Ls);
+    }
+}
+
+unsigned int Environment::getNd(void) const
+{
+    return nd_;
+}
+
+// random number generator /////////////////////////////////////////////////////
+void Environment::setSeed(const std::vector<int> &seed)
+{
+    rng4d_->SeedFixedIntegers(seed);
+}
+
+GridParallelRNG * Environment::get4dRng(void) const
+{
+    return rng4d_.get();
+}
+
+// module management ///////////////////////////////////////////////////////////
+void Environment::pushModule(Environment::ModPt &pt)
+{
+    std::string name = pt->getName();
+    
+    if (!hasModule(name))
+    {
+        std::vector<unsigned int> inputAddress;
+        unsigned int              address;
+        ModuleInfo                m;
+        
+        m.data = std::move(pt);
+        m.type = typeIdPt(*m.data.get());
+        m.name = name;
+        auto input  = m.data->getInput();
+        for (auto &in: input)
+        {
+            if (!hasObject(in))
+            {
+                addObject(in , -1);
+            }
+            m.input.push_back(objectAddress_[in]);
+        }
+        auto output = m.data->getOutput();
+        module_.push_back(std::move(m));
+        address              = static_cast<unsigned int>(module_.size() - 1);
+        moduleAddress_[name] = address;
+        for (auto &out: output)
+        {
+            if (!hasObject(out))
+            {
+                addObject(out, address);
+            }
+            else
+            {
+                if (object_[objectAddress_[out]].module < 0)
+                {
+                    object_[objectAddress_[out]].module = address;
+                }
+                else
+                {
+                    HADRON_ERROR("object '" + out
+                                 + "' is already produced by module '"
+                                 + module_[object_[getObjectAddress(out)].module].name
+                                 + "' (while pushing module '" + name + "')");
+                }
+            }
+        }
+    }
+    else
+    {
+        HADRON_ERROR("module '" + name + "' already exists");
+    }
+}
+
+unsigned int Environment::getNModule(void) const
+{
+    return module_.size();
+}
+
+void Environment::createModule(const std::string name, const std::string type,
+                               XmlReader &reader)
+{
+    auto &factory = ModuleFactory::getInstance();
+    auto pt       = factory.create(type, name);
+    
+    pt->parseParameters(reader, "options");
+    pushModule(pt);
+}
+
+ModuleBase * Environment::getModule(const unsigned int address) const
+{
+    if (hasModule(address))
+    {
+        return module_[address].data.get();
+    }
+    else
+    {
+        HADRON_ERROR("no module with address " + std::to_string(address));
+    }
+}
+
+ModuleBase * Environment::getModule(const std::string name) const
+{
+    return getModule(getModuleAddress(name));
+}
+
+unsigned int Environment::getModuleAddress(const std::string name) const
+{
+    if (hasModule(name))
+    {
+        return moduleAddress_.at(name);
+    }
+    else
+    {
+        HADRON_ERROR("no module with name '" + name + "'");
+    }
+}
+
+std::string Environment::getModuleName(const unsigned int address) const
+{
+    if (hasModule(address))
+    {
+        return module_[address].name;
+    }
+    else
+    {
+        HADRON_ERROR("no module with address " + std::to_string(address));
+    }
+}
+
+std::string Environment::getModuleType(const unsigned int address) const
+{
+    if (hasModule(address))
+    {
+        return typeName(module_[address].type);
+    }
+    else
+    {
+        HADRON_ERROR("no module with address " + std::to_string(address));
+    }
+}
+
+std::string Environment::getModuleType(const std::string name) const
+{
+    return getModuleType(getModuleAddress(name));
+}
+
+bool Environment::hasModule(const unsigned int address) const
+{
+    return (address < module_.size());
+}
+
+bool Environment::hasModule(const std::string name) const
+{
+    return (moduleAddress_.find(name) != moduleAddress_.end());
+}
+
+Graph<unsigned int> Environment::makeModuleGraph(void) const
+{
+    Graph<unsigned int> moduleGraph;
+    
+    for (unsigned int i = 0; i < module_.size(); ++i)
+    {
+        moduleGraph.addVertex(i);
+        for (auto &j: module_[i].input)
+        {
+            moduleGraph.addEdge(object_[j].module, i);
+        }
+    }
+    
+    return moduleGraph;
+}
+
+#define BIG_SEP "==============="
+#define SEP     "---------------"
+#define MEM_MSG(size)\
+sizeString((size)*locVol_) << " (" << sizeString(size)  << "/site)"
+
+Environment::Size
+Environment::executeProgram(const std::vector<unsigned int> &p)
+{
+    Size                                memPeak = 0, sizeBefore, sizeAfter;
+    std::vector<std::set<unsigned int>> freeProg;
+    bool                                continueCollect, nothingFreed;
+    
+    // build garbage collection schedule
+    freeProg.resize(p.size());
+    for (unsigned int i = 0; i < object_.size(); ++i)
+    {
+        auto pred = [i, this](const unsigned int j)
+        {
+            auto &in = module_[j].input;
+            auto it  = std::find(in.begin(), in.end(), i);
+            
+            return (it != in.end()) or (j == object_[i].module);
+        };
+        auto it = std::find_if(p.rbegin(), p.rend(), pred);
+        if (it != p.rend())
+        {
+            freeProg[p.rend() - it - 1].insert(i);
+        }
+    }
+    
+    // program execution
+    for (unsigned int i = 0; i < p.size(); ++i)
+    {
+        // execute module
+        if (!isDryRun())
+        {
+            LOG(Message) << SEP << " Measurement step " << i+1 << "/"
+                         << p.size() << " (module '" << module_[p[i]].name
+                         << "') " << SEP << std::endl;
+        }
+        (*module_[p[i]].data)();
+        sizeBefore = getTotalSize();
+        // print used memory after execution
+        if (!isDryRun())
+        {
+            LOG(Message) << "Allocated objects: " << MEM_MSG(sizeBefore)
+                         << std::endl;
+        }
+        if (sizeBefore > memPeak)
+        {
+            memPeak = sizeBefore;
+        }
+        // garbage collection for step i
+        if (!isDryRun())
+        {
+            LOG(Message) << "Garbage collection..." << std::endl;
+        }
+        nothingFreed = true;
+        do
+        {
+            continueCollect = false;
+            auto toFree = freeProg[i];
+            for (auto &j: toFree)
+            {
+                // continue garbage collection while there are still
+                // objects without owners
+                continueCollect = continueCollect or !hasOwners(j);
+                if(freeObject(j))
+                {
+                    // if an object has been freed, remove it from
+                    // the garbage collection schedule
+                    freeProg[i].erase(j);
+                    nothingFreed = false;
+                }
+            }
+        } while (continueCollect);
+        // any remaining objects in step i garbage collection schedule
+        // is scheduled for step i + 1
+        if (i + 1 < p.size())
+        {
+            for (auto &j: freeProg[i])
+            {
+                freeProg[i + 1].insert(j);
+            }
+        }
+        // print used memory after garbage collection if necessary
+        if (!isDryRun())
+        {
+            sizeAfter = getTotalSize();
+            if (sizeBefore != sizeAfter)
+            {
+                LOG(Message) << "Allocated objects: " << MEM_MSG(sizeAfter)
+                             << std::endl;
+            }
+            else
+            {
+                LOG(Message) << "Nothing to free" << std::endl;
+            }
+        }
+    }
+    
+    return memPeak;
+}
+
+Environment::Size Environment::executeProgram(const std::vector<std::string> &p)
+{
+    std::vector<unsigned int> pAddress;
+    
+    for (auto &n: p)
+    {
+        pAddress.push_back(getModuleAddress(n));
+    }
+    
+    return executeProgram(pAddress);
+}
+
+// general memory management ///////////////////////////////////////////////////
+void Environment::addObject(const std::string name, const int moduleAddress)
+{
+    if (!hasObject(name))
+    {
+        ObjInfo info;
+        
+        info.name   = name;
+        info.module = moduleAddress;
+        object_.push_back(std::move(info));
+        objectAddress_[name] = static_cast<unsigned int>(object_.size() - 1);
+    }
+    else
+    {
+        HADRON_ERROR("object '" + name + "' already exists");
+    }
+}
+
+void Environment::registerObject(const unsigned int address,
+                                 const unsigned int size, const unsigned int Ls)
+{
+    if (!hasRegisteredObject(address))
+    {
+        if (hasObject(address))
+        {
+            object_[address].size         = size;
+            object_[address].Ls           = Ls;
+            object_[address].isRegistered = true;
+        }
+        else
+        {
+            HADRON_ERROR("no object with address " + std::to_string(address));
+        }
+    }
+    else
+    {
+        HADRON_ERROR("object with address " + std::to_string(address)
+                     + " already registered");
+    }
+}
+
+void Environment::registerObject(const std::string name,
+                                 const unsigned int size, const unsigned int Ls)
+{
+    if (!hasObject(name))
+    {
+        addObject(name);
+    }
+    registerObject(getObjectAddress(name), size, Ls);
+}
+
+unsigned int Environment::getObjectAddress(const std::string name) const
+{
+    if (hasObject(name))
+    {
+        return objectAddress_.at(name);
+    }
+    else
+    {
+        HADRON_ERROR("no object with name '" + name + "'");
+    }
+}
+
+std::string Environment::getObjectName(const unsigned int address) const
+{
+    if (hasObject(address))
+    {
+        return object_[address].name;
+    }
+    else
+    {
+        HADRON_ERROR("no object with address " + std::to_string(address));
+    }
+}
+
+std::string Environment::getObjectType(const unsigned int address) const
+{
+    if (hasRegisteredObject(address))
+    {
+        return typeName(object_[address].type);
+    }
+    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));
+    }
+}
+
+std::string Environment::getObjectType(const std::string name) const
+{
+    return getObjectType(getObjectAddress(name));
+}
+
+Environment::Size Environment::getObjectSize(const unsigned int address) const
+{
+    if (hasRegisteredObject(address))
+    {
+        return object_[address].size;
+    }
+    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));
+    }
+}
+
+Environment::Size Environment::getObjectSize(const std::string name) const
+{
+    return getObjectSize(getObjectAddress(name));
+}
+
+unsigned int Environment::getObjectLs(const unsigned int address) const
+{
+    if (hasRegisteredObject(address))
+    {
+        return object_[address].Ls;
+    }
+    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));
+    }
+}
+
+unsigned int Environment::getObjectLs(const std::string name) const
+{
+    return getObjectLs(getObjectAddress(name));
+}
+
+bool Environment::hasObject(const unsigned int address) const
+{
+    return (address < object_.size());
+}
+
+bool Environment::hasObject(const std::string name) const
+{
+    auto it = objectAddress_.find(name);
+    
+    return ((it != objectAddress_.end()) and hasObject(it->second));
+}
+
+bool Environment::hasRegisteredObject(const unsigned int address) const
+{
+    if (hasObject(address))
+    {
+        return object_[address].isRegistered;
+    }
+    else
+    {
+        return false;
+    }
+}
+
+bool Environment::hasRegisteredObject(const std::string name) const
+{
+    if (hasObject(name))
+    {
+        return hasRegisteredObject(getObjectAddress(name));
+    }
+    else
+    {
+        return false;
+    }
+}
+
+bool Environment::hasCreatedObject(const unsigned int address) const
+{
+    if (hasObject(address))
+    {
+        return (object_[address].data != nullptr);
+    }
+    else
+    {
+        return false;
+    }
+}
+
+bool Environment::hasCreatedObject(const std::string name) const
+{
+    if (hasObject(name))
+    {
+        return hasCreatedObject(getObjectAddress(name));
+    }
+    else
+    {
+        return false;
+    }
+}
+
+bool Environment::isObject5d(const unsigned int address) const
+{
+    return (getObjectLs(address) > 1);
+}
+
+bool Environment::isObject5d(const std::string name) const
+{
+    return (getObjectLs(name) > 1);
+}
+
+Environment::Size Environment::getTotalSize(void) const
+{
+    Environment::Size size = 0;
+    
+    for (auto &o: object_)
+    {
+        if (o.isRegistered)
+        {
+            size += o.size;
+        }
+    }
+    
+    return size;
+}
+
+void Environment::addOwnership(const unsigned int owner,
+                               const unsigned int property)
+{
+    if (hasObject(property))
+    {
+        object_[property].owners.insert(owner);
+    }
+    else
+    {
+        HADRON_ERROR("no object with address " + std::to_string(property));
+    }
+    if (hasObject(owner))
+    {
+        object_[owner].properties.insert(property);
+    }
+    else
+    {
+        HADRON_ERROR("no object with address " + std::to_string(owner));
+    }
+}
+
+void Environment::addOwnership(const std::string owner,
+                               const std::string property)
+{
+    addOwnership(getObjectAddress(owner), getObjectAddress(property));
+}
+
+bool Environment::hasOwners(const unsigned int address) const
+{
+    
+    if (hasObject(address))
+    {
+        return (!object_[address].owners.empty());
+    }
+    else
+    {
+        HADRON_ERROR("no object with address " + std::to_string(address));
+    }
+}
+
+bool Environment::hasOwners(const std::string name) const
+{
+    return hasOwners(getObjectAddress(name));
+}
+
+bool Environment::freeObject(const unsigned int address)
+{
+    if (!hasOwners(address))
+    {
+        if (!isDryRun() and object_[address].isRegistered)
+        {
+            LOG(Message) << "Destroying object '" << object_[address].name
+                         << "'" << std::endl;
+        }
+        for (auto &p: object_[address].properties)
+        {
+            object_[p].owners.erase(address);
+        }
+        object_[address].size         = 0;
+        object_[address].Ls           = 0;
+        object_[address].isRegistered = false;
+        object_[address].type         = nullptr;
+        object_[address].owners.clear();
+        object_[address].properties.clear();
+        object_[address].data.reset(nullptr);
+        
+        return true;
+    }
+    else
+    {
+        return false;
+    }
+}
+
+bool Environment::freeObject(const std::string name)
+{
+    return freeObject(getObjectAddress(name));
+}
+
+void Environment::freeAll(void)
+{
+    for (unsigned int i = 0; i < object_.size(); ++i)
+    {
+        freeObject(i);
+    }
+}
+
+void Environment::printContent(void)
+{
+    LOG(Message) << "Modules: " << std::endl;
+    for (unsigned int i = 0; i < module_.size(); ++i)
+    {
+        LOG(Message) << std::setw(4) << i << ": "
+                     << getModuleName(i) << std::endl;
+    }
+    LOG(Message) << "Objects: " << std::endl;
+    for (unsigned int i = 0; i < object_.size(); ++i)
+    {
+        LOG(Message) << std::setw(4) << i << ": "
+                     << getObjectName(i) << std::endl;
+    }
+}

extras/Hadrons/Environment.hpp

@@ -0,0 +1,385 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Environment.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_Environment_hpp_
+#define Hadrons_Environment_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Graph.hpp>
+
+#ifndef SITE_SIZE_TYPE
+#define SITE_SIZE_TYPE unsigned int
+#endif
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                         Global environment                                 *
+ ******************************************************************************/
+// forward declaration of Module
+class ModuleBase;
+
+class Object
+{
+public:
+    Object(void) = default;
+    virtual ~Object(void) = default;
+};
+
+template <typename T>
+class Holder: public Object
+{
+public:
+    Holder(void) = default;
+    Holder(T *pt);
+    virtual ~Holder(void) = default;
+    T &       get(void) const;
+    T *       getPt(void) const;
+    void      reset(T *pt);
+private:
+    std::unique_ptr<T> objPt_{nullptr};
+};
+
+class Environment
+{
+    SINGLETON(Environment);
+public:
+    typedef SITE_SIZE_TYPE                         Size;
+    typedef std::unique_ptr<ModuleBase>            ModPt;
+    typedef std::unique_ptr<GridCartesian>         GridPt;
+    typedef std::unique_ptr<GridRedBlackCartesian> GridRbPt;
+    typedef std::unique_ptr<GridParallelRNG>       RngPt;
+    typedef std::unique_ptr<LatticeBase>           LatticePt;
+private:
+    struct ModuleInfo
+    {
+        const std::type_info      *type{nullptr};
+        std::string               name;
+        ModPt                     data{nullptr};
+        std::vector<unsigned int> input;
+    };
+    struct ObjInfo
+    {
+        Size                    size{0};
+        unsigned int            Ls{0};
+        bool                    isRegistered{false};
+        const std::type_info    *type{nullptr};
+        std::string             name;
+        int                     module{-1};
+        std::set<unsigned int>  owners, properties;
+        std::unique_ptr<Object> data{nullptr};
+    };
+public:
+    // dry run
+    void                    dryRun(const bool isDry);
+    bool                    isDryRun(void) const;
+    // trajectory number
+    void                    setTrajectory(const unsigned int traj);
+    unsigned int            getTrajectory(void) const;
+    // grids
+    void                    createGrid(const unsigned int Ls);
+    GridCartesian *         getGrid(const unsigned int Ls = 1) const;
+    GridRedBlackCartesian * getRbGrid(const unsigned int Ls = 1) const;
+    unsigned int            getNd(void) const;
+    // random number generator
+    void                    setSeed(const std::vector<int> &seed);
+    GridParallelRNG *       get4dRng(void) const;
+    // module management
+    void                    pushModule(ModPt &pt);
+    template <typename M>
+    void                    createModule(const std::string name);
+    template <typename M>
+    void                    createModule(const std::string name,
+                                         const typename M::Par &par);
+    void                    createModule(const std::string name,
+                                         const std::string type,
+                                         XmlReader &reader);
+    unsigned int            getNModule(void) const;
+    ModuleBase *            getModule(const unsigned int address) const;
+    ModuleBase *            getModule(const std::string name) const;
+    template <typename M>
+    M *                     getModule(const unsigned int address) const;
+    template <typename M>
+    M *                     getModule(const std::string name) const;
+    unsigned int            getModuleAddress(const std::string name) const;
+    std::string             getModuleName(const unsigned int address) const;
+    std::string             getModuleType(const unsigned int address) const;
+    std::string             getModuleType(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;
+    Size                    executeProgram(const std::vector<unsigned int> &p);
+    Size                    executeProgram(const std::vector<std::string> &p);
+    // general memory management
+    void                    addObject(const std::string name,
+                                      const int moduleAddress = -1);
+    void                    registerObject(const unsigned int address,
+                                           const unsigned int size,
+                                           const unsigned int Ls = 1);
+    void                    registerObject(const std::string name,
+                                           const unsigned int size,
+                                           const unsigned int Ls = 1);
+    template <typename T>
+    unsigned int            lattice4dSize(void) const;
+    template <typename T>
+    void                    registerLattice(const unsigned int address,
+                                            const unsigned int Ls = 1);
+    template <typename T>
+    void                    registerLattice(const std::string name,
+                                            const unsigned int Ls = 1);
+    template <typename T>
+    void                    setObject(const unsigned int address, T *object);
+    template <typename T>
+    void                    setObject(const std::string name, T *object);
+    template <typename T>
+    T *                     getObject(const unsigned int address) const;
+    template <typename T>
+    T *                     getObject(const std::string name) const;
+    template <typename T>
+    T *                     createLattice(const unsigned int address);
+    template <typename T>
+    T *                     createLattice(const std::string name);
+    unsigned int            getObjectAddress(const std::string name) const;
+    std::string             getObjectName(const unsigned int address) const;
+    std::string             getObjectType(const unsigned int address) const;
+    std::string             getObjectType(const std::string name) const;
+    Size                    getObjectSize(const unsigned int address) const;
+    Size                    getObjectSize(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;
+    bool                    hasObject(const std::string name) const;
+    bool                    hasRegisteredObject(const unsigned int address) const;
+    bool                    hasRegisteredObject(const std::string name) const;
+    bool                    hasCreatedObject(const unsigned int address) const;
+    bool                    hasCreatedObject(const std::string name) const;
+    bool                    isObject5d(const unsigned int address) const;
+    bool                    isObject5d(const std::string name) const;
+    Environment::Size       getTotalSize(void) const;
+    void                    addOwnership(const unsigned int owner,
+                                         const unsigned int property);
+    void                    addOwnership(const std::string owner,
+                                         const std::string property);
+    bool                    hasOwners(const unsigned int address) const;
+    bool                    hasOwners(const std::string name) const;
+    bool                    freeObject(const unsigned int address);
+    bool                    freeObject(const std::string name);
+    void                    freeAll(void);
+    void                    printContent(void);
+private:
+    // general
+    bool                                   dryRun_{false};
+    unsigned int                           traj_, locVol_;
+    // grids
+    GridPt                                 grid4d_;
+    std::map<unsigned int, GridPt>         grid5d_;
+    GridRbPt                               gridRb4d_;
+    std::map<unsigned int, GridRbPt>       gridRb5d_;
+    unsigned int                           nd_;
+    // random number generator
+    RngPt                                  rng4d_;
+    // module and related maps
+    std::vector<ModuleInfo>                module_;
+    std::map<std::string, unsigned int>    moduleAddress_;
+    // lattice store
+    std::map<unsigned int, LatticePt>      lattice_;
+    // object store
+    std::vector<ObjInfo>                   object_;
+    std::map<std::string, unsigned int>    objectAddress_;
+};
+
+/******************************************************************************
+ *                       Holder template implementation                       *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+template <typename T>
+Holder<T>::Holder(T *pt)
+: objPt_(pt)
+{}
+
+// access //////////////////////////////////////////////////////////////////////
+template <typename T>
+T & Holder<T>::get(void) const
+{
+    return &objPt_.get();
+}
+
+template <typename T>
+T * Holder<T>::getPt(void) const
+{
+    return objPt_.get();
+}
+
+template <typename T>
+void Holder<T>::reset(T *pt)
+{
+    objPt_.reset(pt);
+}
+
+/******************************************************************************
+ *                     Environment template implementation                    *
+ ******************************************************************************/
+// module management ///////////////////////////////////////////////////////////
+template <typename M>
+void Environment::createModule(const std::string name)
+{
+    ModPt pt(new M(name));
+    
+    pushModule(pt);
+}
+
+template <typename M>
+void Environment::createModule(const std::string name,
+                               const typename M::Par &par)
+{
+    ModPt pt(new M(name));
+    
+    static_cast<M *>(pt.get())->setPar(par);
+    pushModule(pt);
+}
+
+template <typename M>
+M * Environment::getModule(const unsigned int address) const
+{
+    if (auto *pt = dynamic_cast<M *>(getModule(address)))
+    {
+        return pt;
+    }
+    else
+    {
+        HADRON_ERROR("module '" + module_[address].name
+                     + "' does not have type " + typeid(M).name()
+                     + "(object type: " + getModuleType(address) + ")");
+    }
+}
+
+template <typename M>
+M * Environment::getModule(const std::string name) const
+{
+    return getModule<M>(getModuleAddress(name));
+}
+
+template <typename T>
+unsigned int Environment::lattice4dSize(void) const
+{
+    return sizeof(typename T::vector_object)/getGrid()->Nsimd();
+}
+
+template <typename T>
+void Environment::registerLattice(const unsigned int address,
+                                  const unsigned int Ls)
+{
+    createGrid(Ls);
+    registerObject(address, Ls*lattice4dSize<T>(), Ls);
+}
+
+template <typename T>
+void Environment::registerLattice(const std::string name, const unsigned int Ls)
+{
+    createGrid(Ls);
+    registerObject(name, Ls*lattice4dSize<T>(), Ls);
+}
+
+template <typename T>
+void Environment::setObject(const unsigned int address, T *object)
+{
+    if (hasRegisteredObject(address))
+    {
+        object_[address].data.reset(new Holder<T>(object));
+        object_[address].type = &typeid(T);
+    }
+    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>
+void Environment::setObject(const std::string name, T *object)
+{
+    setObject(getObjectAddress(name), object);
+}
+
+template <typename T>
+T * Environment::getObject(const unsigned int address) const
+{
+    if (hasRegisteredObject(address))
+    {
+        if (auto h = dynamic_cast<Holder<T> *>(object_[address].data.get()))
+        {
+            return h->getPt();
+        }
+        else
+        {
+            HADRON_ERROR("object with address " + std::to_string(address) +
+                         " does not have type '" + typeid(T).name() +
+                         "' (has type '" + getObjectType(address) + "')");
+        }
+    }
+    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>
+T * Environment::getObject(const std::string name) const
+{
+    return getObject<T>(getObjectAddress(name));
+}
+
+template <typename T>
+T * Environment::createLattice(const unsigned int address)
+{
+    GridCartesian *g = getGrid(getObjectLs(address));
+    
+    setObject(address, new T(g));
+    
+    return getObject<T>(address);
+}
+
+template <typename T>
+T * Environment::createLattice(const std::string name)
+{
+    return createLattice<T>(getObjectAddress(name));
+}
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_Environment_hpp_

extras/Hadrons/Factory.hpp

@@ -0,0 +1,106 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Factory.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_Factory_hpp_
+#define Hadrons_Factory_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                        abstract factory class                              *
+ ******************************************************************************/
+template <typename T>
+class Factory
+{
+public:
+    typedef std::function<std::unique_ptr<T>(const std::string)> Func;
+public:
+    // constructor
+    Factory(void) = default;
+    // destructor
+    virtual ~Factory(void) = default;
+    // registration
+    void registerBuilder(const std::string type, const Func &f);
+    // get builder list
+    std::vector<std::string> getBuilderList(void) const;
+    // factory
+    std::unique_ptr<T> create(const std::string type,
+                              const std::string name) const;
+private:
+    std::map<std::string, Func> builder_;
+};
+
+/******************************************************************************
+ *                         template implementation                            *
+ ******************************************************************************/
+// registration ////////////////////////////////////////////////////////////////
+template <typename T>
+void Factory<T>::registerBuilder(const std::string type, const Func &f)
+{
+    builder_[type] = f;
+}
+
+// get module list /////////////////////////////////////////////////////////////
+template <typename T>
+std::vector<std::string> Factory<T>::getBuilderList(void) const
+{
+    std::vector<std::string> list;
+    
+    for (auto &b: builder_)
+    {
+        list.push_back(b.first);
+    }
+    
+    return list;
+}
+
+// factory /////////////////////////////////////////////////////////////////////
+template <typename T>
+std::unique_ptr<T> Factory<T>::create(const std::string type,
+                                      const std::string name) const
+{
+    Func func;
+    
+    try
+    {
+        func = builder_.at(type);
+    }
+    catch (std::out_of_range &)
+    {
+        HADRON_ERROR("object of type '" + type + "' unknown");
+    }
+    
+    return func(name);
+}
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_Factory_hpp_

extras/Hadrons/GeneticScheduler.hpp

@@ -0,0 +1,329 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/GeneticScheduler.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_GeneticScheduler_hpp_
+#define Hadrons_GeneticScheduler_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Graph.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                   Scheduler based on a genetic algorithm                   *
+ ******************************************************************************/
+template <typename T>
+class GeneticScheduler
+{
+public:
+    typedef std::vector<T>                   Gene;
+    typedef std::pair<Gene *, Gene *>        GenePair;
+    typedef std::function<int(const Gene &)> ObjFunc;
+    struct Parameters
+    {
+        double       mutationRate;
+        unsigned int popSize, seed;
+    };
+public:
+    // constructor
+    GeneticScheduler(Graph<T> &graph, const ObjFunc &func,
+                     const Parameters &par);
+    // destructor
+    virtual ~GeneticScheduler(void) = default;
+    // access
+    const Gene & getMinSchedule(void);
+    int          getMinValue(void);
+    // breed a new generation
+    void nextGeneration(void);
+    // heuristic benchmarks
+    void benchmarkCrossover(const unsigned int nIt);
+    // print population
+    friend std::ostream & operator<<(std::ostream &out,
+                                     const GeneticScheduler<T> &s)
+    {
+        out << "[";
+        for (auto &p: s.population_)
+        {
+            out << p.first << ", ";
+        }
+        out << "\b\b]";
+        
+        return out;
+    }
+private:
+    // evolution steps
+    void initPopulation(void);
+    void doCrossover(void);
+    void doMutation(void);
+    // genetic operators
+    GenePair selectPair(void);
+    void     crossover(Gene &c1, Gene &c2, const Gene &p1, const Gene &p2);
+    void     mutation(Gene &m, const Gene &c);
+    
+private:
+    Graph<T>                 &graph_;
+    const ObjFunc            &func_;
+    const Parameters         par_;
+    std::multimap<int, Gene> population_;
+    std::mt19937             gen_;
+};
+
+/******************************************************************************
+ *                       template implementation                              *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+template <typename T>
+GeneticScheduler<T>::GeneticScheduler(Graph<T> &graph, const ObjFunc &func,
+                                      const Parameters &par)
+: graph_(graph)
+, func_(func)
+, par_(par)
+{
+    gen_.seed(par_.seed);
+}
+
+// access //////////////////////////////////////////////////////////////////////
+template <typename T>
+const typename GeneticScheduler<T>::Gene &
+GeneticScheduler<T>::getMinSchedule(void)
+{
+    return population_.begin()->second;
+}
+
+template <typename T>
+int GeneticScheduler<T>::getMinValue(void)
+{
+    return population_.begin()->first;
+}
+
+// breed a new generation //////////////////////////////////////////////////////
+template <typename T>
+void GeneticScheduler<T>::nextGeneration(void)
+{
+    // random initialization of the population if necessary
+    if (population_.size() != par_.popSize)
+    {
+        initPopulation();
+    }
+    LOG(Debug) << "Starting population:\n" << *this << std::endl;
+    
+    // random mutations
+    //PARALLEL_FOR_LOOP
+    for (unsigned int i = 0; i < par_.popSize; ++i)
+    {
+        doMutation();
+    }
+    LOG(Debug) << "After mutations:\n" << *this << std::endl;
+    
+    // mating
+    //PARALLEL_FOR_LOOP
+    for (unsigned int i = 0; i < par_.popSize/2; ++i)
+    {
+        doCrossover();
+    }
+    LOG(Debug) << "After mating:\n" << *this << std::endl;
+    
+    // grim reaper
+    auto it = population_.begin();
+    
+    std::advance(it, par_.popSize);
+    population_.erase(it, population_.end());
+    LOG(Debug) << "After grim reaper:\n" << *this << std::endl;
+}
+
+// evolution steps /////////////////////////////////////////////////////////////
+template <typename T>
+void GeneticScheduler<T>::initPopulation(void)
+{
+    population_.clear();
+    for (unsigned int i = 0; i < par_.popSize; ++i)
+    {
+        auto p = graph_.topoSort(gen_);
+        
+        population_.insert(std::make_pair(func_(p), p));
+    }
+}
+
+template <typename T>
+void GeneticScheduler<T>::doCrossover(void)
+{
+    auto p = selectPair();
+    Gene &p1 = *(p.first), &p2 = *(p.second);
+    Gene c1, c2;
+    
+    crossover(c1, c2, p1, p2);
+    PARALLEL_CRITICAL
+    {
+        population_.insert(std::make_pair(func_(c1), c1));
+        population_.insert(std::make_pair(func_(c2), c2));
+    }
+}
+
+template <typename T>
+void GeneticScheduler<T>::doMutation(void)
+{
+    std::uniform_real_distribution<double>      mdis(0., 1.);
+    std::uniform_int_distribution<unsigned int> pdis(0, population_.size() - 1);
+    
+    if (mdis(gen_) < par_.mutationRate)
+    {
+        Gene m;
+        auto it = population_.begin();
+        
+        std::advance(it, pdis(gen_));
+        mutation(m, it->second);
+        PARALLEL_CRITICAL
+        {
+            population_.insert(std::make_pair(func_(m), m));
+        }
+    }
+}
+
+// genetic operators ///////////////////////////////////////////////////////////
+template <typename T>
+typename GeneticScheduler<T>::GenePair GeneticScheduler<T>::selectPair(void)
+{
+    std::vector<double> prob;
+    unsigned int        ind;
+    Gene                *p1, *p2;
+    
+    for (auto &c: population_)
+    {
+        prob.push_back(1./c.first);
+    }
+    do
+    {
+        double probCpy;
+        
+        std::discrete_distribution<unsigned int> dis1(prob.begin(), prob.end());
+        auto rIt = population_.begin();
+        ind = dis1(gen_);
+        std::advance(rIt, ind);
+        p1 = &(rIt->second);
+        probCpy   = prob[ind];
+        prob[ind] = 0.;
+        std::discrete_distribution<unsigned int> dis2(prob.begin(), prob.end());
+        rIt = population_.begin();
+        std::advance(rIt, dis2(gen_));
+        p2 = &(rIt->second);
+        prob[ind] = probCpy;
+    } while (p1 == p2);
+    
+    return std::make_pair(p1, p2);
+}
+
+template <typename T>
+void GeneticScheduler<T>::crossover(Gene &c1, Gene &c2, const Gene &p1,
+                                    const Gene &p2)
+{
+    Gene                                        buf;
+    std::uniform_int_distribution<unsigned int> dis(0, p1.size() - 1);
+    unsigned int                                cut = dis(gen_);
+    
+    c1.clear();
+    buf = p2;
+    for (unsigned int i = 0; i < cut; ++i)
+    {
+        c1.push_back(p1[i]);
+        buf.erase(std::find(buf.begin(), buf.end(), p1[i]));
+    }
+    for (unsigned int i = 0; i < buf.size(); ++i)
+    {
+        c1.push_back(buf[i]);
+    }
+    c2.clear();
+    buf = p2;
+    for (unsigned int i = cut; i < p1.size(); ++i)
+    {
+        buf.erase(std::find(buf.begin(), buf.end(), p1[i]));
+    }
+    for (unsigned int i = 0; i < buf.size(); ++i)
+    {
+        c2.push_back(buf[i]);
+    }
+    for (unsigned int i = cut; i < p1.size(); ++i)
+    {
+        c2.push_back(p1[i]);
+    }
+}
+
+template <typename T>
+void GeneticScheduler<T>::mutation(Gene &m, const Gene &c)
+{
+    Gene                                        buf;
+    std::uniform_int_distribution<unsigned int> dis(0, c.size() - 1);
+    unsigned int                                cut = dis(gen_);
+    Graph<T>                                    g1 = graph_, g2 = graph_;
+    
+    for (unsigned int i = 0; i < cut; ++i)
+    {
+        g1.removeVertex(c[i]);
+    }
+    for (unsigned int i = cut; i < c.size(); ++i)
+    {
+        g2.removeVertex(c[i]);
+    }
+    if (g1.size() > 0)
+    {
+        buf = g1.topoSort(gen_);
+    }
+    if (g2.size() > 0)
+    {
+        m = g2.topoSort(gen_);
+    }
+    for (unsigned int i = cut; i < c.size(); ++i)
+    {
+        m.push_back(buf[i - cut]);
+    }
+}
+
+template <typename T>
+void GeneticScheduler<T>::benchmarkCrossover(const unsigned int nIt)
+{
+    Gene   p1, p2, c1, c2;
+    double neg = 0., eq = 0., pos = 0., total;
+    int    improvement;
+    
+    LOG(Message) << "Benchmarking crossover..." << std::endl;
+    for (unsigned int i = 0; i < nIt; ++i)
+    {
+        p1 = graph_.topoSort(gen_);
+        p2 = graph_.topoSort(gen_);
+        crossover(c1, c2, p1, p2);
+        improvement = (func_(c1) + func_(c2) - func_(p1) - func_(p2))/2;
+        if (improvement < 0) neg++; else if (improvement == 0) eq++; else pos++;
+    }
+    total = neg + eq + pos;
+    LOG(Message) << "  -: " << neg/total << "  =: " << eq/total
+                 << "  +: " << pos/total << std::endl;
+}
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_GeneticScheduler_hpp_

extras/Hadrons/Global.cc

@@ -0,0 +1,82 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Global.cc
+
+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/Global.hpp>
+
+using namespace Grid;
+using namespace QCD;
+using namespace Hadrons;
+
+HadronsLogger Hadrons::HadronsLogError(1,"Error");
+HadronsLogger Hadrons::HadronsLogWarning(1,"Warning");
+HadronsLogger Hadrons::HadronsLogMessage(1,"Message");
+HadronsLogger Hadrons::HadronsLogIterative(1,"Iterative");
+HadronsLogger Hadrons::HadronsLogDebug(1,"Debug");
+
+// pretty size formatting //////////////////////////////////////////////////////
+std::string Hadrons::sizeString(long unsigned int bytes)
+
+{
+    constexpr unsigned int bufSize = 256;
+    const char             *suffixes[7] = {"", "K", "M", "G", "T", "P", "E"};
+    char                   buf[256];
+    long unsigned int      s     = 0;
+    double                 count = bytes;
+    
+    while (count >= 1024 && s < 7)
+    {
+        s++;
+        count /= 1024;
+    }
+    if (count - floor(count) == 0.0)
+    {
+        snprintf(buf, bufSize, "%d %sB", (int)count, suffixes[s]);
+    }
+    else
+    {
+        snprintf(buf, bufSize, "%.1f %sB", count, suffixes[s]);
+    }
+    
+    return std::string(buf);
+}
+
+// type utilities //////////////////////////////////////////////////////////////
+constexpr unsigned int maxNameSize = 1024u;
+
+std::string Hadrons::typeName(const std::type_info *info)
+{
+    char        *buf;
+    std::string name;
+    
+    buf  = abi::__cxa_demangle(info->name(), nullptr, nullptr, nullptr);
+    name = buf;
+    free(buf);
+    
+    return name;
+}

extras/Hadrons/Global.hpp

@@ -0,0 +1,150 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Global.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_Global_hpp_
+#define Hadrons_Global_hpp_
+
+#include <set>
+#include <stack>
+#include <Grid/Grid.h>
+#include <cxxabi.h>
+
+#define BEGIN_HADRONS_NAMESPACE \
+namespace Grid {\
+using namespace QCD;\
+namespace Hadrons {\
+using Grid::operator<<;
+#define END_HADRONS_NAMESPACE }}
+
+#define BEGIN_MODULE_NAMESPACE(name)\
+namespace name {\
+using Grid::operator<<;
+#define END_MODULE_NAMESPACE }
+
+/* the 'using Grid::operator<<;' statement prevents a very nasty compilation
+ * 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
+
+BEGIN_HADRONS_NAMESPACE
+
+// type aliases
+#define 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 &,                             \
+                      const FermionField##suffix &)> SolverFn##suffix;
+
+// logger
+class HadronsLogger: public Logger
+{
+public:
+    HadronsLogger(int on, std::string nm): Logger("Hadrons", on, nm,
+                                                  GridLogColours, "BLACK"){};
+};
+
+#define LOG(channel) std::cout << HadronsLog##channel
+#define HADRON_ERROR(msg)\
+LOG(Error) << msg << " (" << __FUNCTION__ << " at " << __FILE__ << ":"\
+           << __LINE__ << ")" << std::endl;\
+abort();
+
+#define DEBUG_VAR(var) LOG(Debug) << #var << "= " << (var) << std::endl;
+
+extern HadronsLogger HadronsLogError;
+extern HadronsLogger HadronsLogWarning;
+extern HadronsLogger HadronsLogMessage;
+extern HadronsLogger HadronsLogIterative;
+extern HadronsLogger HadronsLogDebug;
+
+// singleton pattern
+#define SINGLETON(name)\
+public:\
+    name(const name &e) = delete;\
+    void operator=(const name &e) = delete;\
+    static name & getInstance(void)\
+    {\
+        static name e;\
+        return e;\
+    }\
+private:\
+    name(void);
+
+#define SINGLETON_DEFCTOR(name)\
+public:\
+    name(const name &e) = delete;\
+    void operator=(const name &e) = delete;\
+    static name & getInstance(void)\
+    {\
+        static name e;\
+        return e;\
+    }\
+private:\
+    name(void) = default;
+
+// pretty size formating
+std::string sizeString(long unsigned int bytes);
+
+// type utilities
+template <typename T>
+const std::type_info * typeIdPt(const T &x)
+{
+    return &typeid(x);
+}
+
+std::string typeName(const std::type_info *info);
+
+template <typename T>
+const std::type_info * typeIdPt(void)
+{
+    return &typeid(T);
+}
+
+template <typename T>
+std::string typeName(const T &x)
+{
+    return typeName(typeIdPt(x));
+}
+
+template <typename T>
+std::string typeName(void)
+{
+    return typeName(typeIdPt<T>());
+}
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_Global_hpp_

extras/Hadrons/Graph.hpp

@@ -0,0 +1,760 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Graph.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_Graph_hpp_
+#define Hadrons_Graph_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                          Oriented graph class                              *
+ ******************************************************************************/
+// I/O for edges
+template <typename T>
+std::ostream & operator<<(std::ostream &out, const std::pair<T, T> &e)
+{
+    out << "\""  << e.first << "\" -> \"" << e.second << "\"";
+    
+    return out;
+}
+
+// main class
+template <typename T>
+class Graph
+{
+public:
+    typedef std::pair<T, T> Edge;
+public:
+    // constructor
+    Graph(void);
+    // destructor
+    virtual ~Graph(void) = default;
+    // access
+    void           addVertex(const T &value);
+    void           addEdge(const Edge &e);
+    void           addEdge(const T &start, const T &end);
+    std::vector<T> getVertices(void) const;
+    void           removeVertex(const T &value);
+    void           removeEdge(const Edge &e);
+    void           removeEdge(const T &start, const T &end);
+    unsigned int   size(void) const;
+    // tests
+    bool gotValue(const T &value) const;
+    // graph topological manipulations
+    std::vector<T>              getAdjacentVertices(const T &value) const;
+    std::vector<T>              getChildren(const T &value) const;
+    std::vector<T>              getParents(const T &value) const;
+    std::vector<T>              getRoots(void) const;
+    std::vector<Graph<T>>       getConnectedComponents(void) const;
+    std::vector<T>              topoSort(void);
+    template <typename Gen>
+    std::vector<T>              topoSort(Gen &gen);
+    std::vector<std::vector<T>> allTopoSort(void);
+    // I/O
+    friend std::ostream & operator<<(std::ostream &out, const Graph<T> &g)
+    {
+        out << "{";
+        for (auto &e: g.edgeSet_)
+        {
+            out << e << ", ";
+        }
+        if (g.edgeSet_.size() != 0)
+        {
+            out << "\b\b";
+        }
+        out << "}";
+        
+        return out;
+    }
+private:
+    // vertex marking
+    void      mark(const T &value, const bool doMark = true);
+    void      markAll(const bool doMark = true);
+    void      unmark(const T &value);
+    void      unmarkAll(void);
+    bool      isMarked(const T &value) const;
+    const T * getFirstMarked(const bool isMarked = true) const;
+    template <typename Gen>
+    const T * getRandomMarked(const bool isMarked, Gen &gen);
+    const T * getFirstUnmarked(void) const;
+    template <typename Gen>
+    const T * getRandomUnmarked(Gen &gen);
+    // prune marked/unmarked vertices
+    void removeMarked(const bool isMarked = true);
+    void removeUnmarked(void);
+    // depth-first search marking
+    void depthFirstSearch(void);
+    void depthFirstSearch(const T &root);
+private:
+    std::map<T, bool>  isMarked_;
+    std::set<Edge>     edgeSet_;
+};
+
+// build depedency matrix from topological sorts
+template <typename T>
+std::map<T, std::map<T, bool>>
+makeDependencyMatrix(const std::vector<std::vector<T>> &topSort);
+
+/******************************************************************************
+ *                       template implementation                              *
+ ******************************************************************************
+ * in all the following V is the number of vertex and E is the number of edge
+ * in the worst case E = V^2
+ */
+
+// constructor /////////////////////////////////////////////////////////////////
+template <typename T>
+Graph<T>::Graph(void)
+{}
+
+// access //////////////////////////////////////////////////////////////////////
+// complexity: log(V)
+template <typename T>
+void Graph<T>::addVertex(const T &value)
+{
+    isMarked_[value] = false;
+}
+
+// complexity: O(log(V))
+template <typename T>
+void Graph<T>::addEdge(const Edge &e)
+{
+    addVertex(e.first);
+    addVertex(e.second);
+    edgeSet_.insert(e);
+}
+
+// complexity: O(log(V))
+template <typename T>
+void Graph<T>::addEdge(const T &start, const T &end)
+{
+    addEdge(Edge(start, end));
+}
+
+template <typename T>
+std::vector<T> Graph<T>::getVertices(void) const
+{
+    std::vector<T> vertex;
+    
+    for (auto &v: isMarked_)
+    {
+        vertex.push_back(v.first);
+    }
+    
+    return vertex;
+}
+
+// complexity: O(V*log(V))
+template <typename T>
+void Graph<T>::removeVertex(const T &value)
+{
+    // remove vertex from the mark table
+    auto vIt = isMarked_.find(value);
+    
+    if (vIt != isMarked_.end())
+    {
+        isMarked_.erase(vIt);
+    }
+    else
+    {
+        HADRON_ERROR("vertex " << value << " does not exists");
+    }
+
+    // remove all edges containing the vertex
+    auto pred = [&value](const Edge &e)
+    {
+        return ((e.first == value) or (e.second == value));
+    };
+    auto eIt = find_if(edgeSet_.begin(), edgeSet_.end(), pred);
+    
+    while (eIt != edgeSet_.end())
+    {
+        edgeSet_.erase(eIt);
+        eIt = find_if(edgeSet_.begin(), edgeSet_.end(), pred);
+    }
+}
+
+// complexity: O(log(V))
+template <typename T>
+void Graph<T>::removeEdge(const Edge &e)
+{
+    auto eIt = edgeSet_.find(e);
+    
+    if (eIt != edgeSet_.end())
+    {
+        edgeSet_.erase(eIt);
+    }
+    else
+    {
+        HADRON_ERROR("edge "  << e << " does not exists");
+    }
+}
+
+// complexity: O(log(V))
+template <typename T>
+void Graph<T>::removeEdge(const T &start, const T &end)
+{
+    removeEdge(Edge(start, end));
+}
+
+// complexity: O(1)
+template <typename T>
+unsigned int Graph<T>::size(void) const
+{
+    return isMarked_.size();
+}
+
+// tests ///////////////////////////////////////////////////////////////////////
+// complexity: O(log(V))
+template <typename T>
+bool Graph<T>::gotValue(const T &value) const
+{
+    auto it = isMarked_.find(value);
+    
+    if (it == isMarked_.end())
+    {
+        return false;
+    }
+    else
+    {
+        return true;
+    }
+}
+
+// vertex marking //////////////////////////////////////////////////////////////
+// complexity: O(log(V))
+template <typename T>
+void Graph<T>::mark(const T &value, const bool doMark)
+{
+    if (gotValue(value))
+    {
+        isMarked_[value] = doMark;
+    }
+    else
+    {
+        HADRON_ERROR("vertex " << value << " does not exists");
+    }
+}
+
+// complexity: O(V*log(V))
+template <typename T>
+void Graph<T>::markAll(const bool doMark)
+{
+    for (auto &v: isMarked_)
+    {
+        mark(v.first, doMark);
+    }
+}
+
+// complexity: O(log(V))
+template <typename T>
+void Graph<T>::unmark(const T &value)
+{
+    mark(value, false);
+}
+
+// complexity: O(V*log(V))
+template <typename T>
+void Graph<T>::unmarkAll(void)
+{
+    markAll(false);
+}
+
+// complexity: O(log(V))
+template <typename T>
+bool Graph<T>::isMarked(const T &value) const
+{
+    if (gotValue(value))
+    {
+        return isMarked_.at(value);
+    }
+    else
+    {
+        HADRON_ERROR("vertex " << value << " does not exists");
+        
+        return false;
+    }
+}
+
+// complexity: O(log(V))
+template <typename T>
+const T * Graph<T>::getFirstMarked(const bool isMarked) const
+{
+    auto pred = [&isMarked](const std::pair<T, bool> &v)
+    {
+        return (v.second == isMarked);
+    };
+    auto vIt = std::find_if(isMarked_.begin(), isMarked_.end(), pred);
+    
+    if (vIt != isMarked_.end())
+    {
+        return &(vIt->first);
+    }
+    else
+    {
+        return nullptr;
+    }
+}
+
+// complexity: O(log(V))
+template <typename T>
+template <typename Gen>
+const T * Graph<T>::getRandomMarked(const bool isMarked, Gen &gen)
+{
+    auto pred = [&isMarked](const std::pair<T, bool> &v)
+    {
+        return (v.second == isMarked);
+    };
+    std::uniform_int_distribution<unsigned int> dis(0, size() - 1);
+    auto                                        rIt = isMarked_.begin();
+    
+    std::advance(rIt, dis(gen));
+    auto vIt = std::find_if(rIt, isMarked_.end(), pred);
+    if (vIt != isMarked_.end())
+    {
+        return &(vIt->first);
+    }
+    else
+    {
+        vIt = std::find_if(isMarked_.begin(), rIt, pred);
+        if (vIt != rIt)
+        {
+            return &(vIt->first);
+        }
+        else
+        {
+            return nullptr;
+        }
+    }
+}
+
+// complexity: O(log(V))
+template <typename T>
+const T * Graph<T>::getFirstUnmarked(void) const
+{
+    return getFirstMarked(false);
+}
+
+// complexity: O(log(V))
+template <typename T>
+template <typename Gen>
+const T * Graph<T>::getRandomUnmarked(Gen &gen)
+{
+    return getRandomMarked(false, gen);
+}
+
+// prune marked/unmarked vertices //////////////////////////////////////////////
+// complexity: O(V^2*log(V))
+template <typename T>
+void Graph<T>::removeMarked(const bool isMarked)
+{
+    auto isMarkedCopy = isMarked_;
+    
+    for (auto &v: isMarkedCopy)
+    {
+        if (v.second == isMarked)
+        {
+            removeVertex(v.first);
+        }
+    }
+}
+
+// complexity: O(V^2*log(V))
+template <typename T>
+void Graph<T>::removeUnmarked(void)
+{
+    removeMarked(false);
+}
+
+// depth-first search marking //////////////////////////////////////////////////
+// complexity: O(V*log(V))
+template <typename T>
+void Graph<T>::depthFirstSearch(void)
+{
+    depthFirstSearch(isMarked_.begin()->first);
+}
+
+// complexity: O(V*log(V))
+template <typename T>
+void Graph<T>::depthFirstSearch(const T &root)
+{
+    std::vector<T> adjacentVertex;
+    
+    mark(root);
+    adjacentVertex = getAdjacentVertices(root);
+    for (auto &v: adjacentVertex)
+    {
+        if (!isMarked(v))
+        {
+            depthFirstSearch(v);
+        }
+    }
+}
+
+// graph topological manipulations /////////////////////////////////////////////
+// complexity: O(V*log(V))
+template <typename T>
+std::vector<T> Graph<T>::getAdjacentVertices(const T &value) const
+{
+    std::vector<T> adjacentVertex;
+    
+    auto pred = [&value](const Edge &e)
+    {
+        return ((e.first == value) or (e.second == value));
+    };
+    auto eIt = find_if(edgeSet_.begin(), edgeSet_.end(), pred);
+    
+    while (eIt != edgeSet_.end())
+    {
+        if (eIt->first == value)
+        {
+            adjacentVertex.push_back((*eIt).second);
+        }
+        else if (eIt->second == value)
+        {
+            adjacentVertex.push_back((*eIt).first);
+        }
+        eIt = find_if(++eIt, edgeSet_.end(), pred);
+    }
+    
+    return adjacentVertex;
+}
+
+// complexity: O(V*log(V))
+template <typename T>
+std::vector<T> Graph<T>::getChildren(const T &value) const
+{
+    std::vector<T> child;
+    
+    auto pred = [&value](const Edge &e)
+    {
+        return (e.first == value);
+    };
+    auto eIt = find_if(edgeSet_.begin(), edgeSet_.end(), pred);
+    
+    while (eIt != edgeSet_.end())
+    {
+        child.push_back((*eIt).second);
+        eIt = find_if(++eIt, edgeSet_.end(), pred);
+    }
+    
+    return child;
+}
+
+// complexity: O(V*log(V))
+template <typename T>
+std::vector<T> Graph<T>::getParents(const T &value) const
+{
+    std::vector<T> parent;
+    
+    auto pred = [&value](const Edge &e)
+    {
+        return (e.second == value);
+    };
+    auto eIt = find_if(edgeSet_.begin(), edgeSet_.end(), pred);
+    
+    while (eIt != edgeSet_.end())
+    {
+        parent.push_back((*eIt).first);
+        eIt = find_if(++eIt, edgeSet_.end(), pred);
+    }
+    
+    return parent;
+}
+
+// complexity: O(V^2*log(V))
+template <typename T>
+std::vector<T> Graph<T>::getRoots(void) const
+{
+    std::vector<T> root;
+    
+    for (auto &v: isMarked_)
+    {
+        auto parent = getParents(v.first);
+        
+        if (parent.size() == 0)
+        {
+            root.push_back(v.first);
+        }
+    }
+    
+    return root;
+}
+
+// complexity: O(V^2*log(V))
+template <typename T>
+std::vector<Graph<T>> Graph<T>::getConnectedComponents(void) const
+{
+    std::vector<Graph<T>> res;
+    Graph<T>              copy(*this);
+    
+    while (copy.size() > 0)
+    {
+        copy.depthFirstSearch();
+        res.push_back(copy);
+        res.back().removeUnmarked();
+        res.back().unmarkAll();
+        copy.removeMarked();
+        copy.unmarkAll();
+    }
+    
+    return res;
+}
+
+// topological sort using a directed DFS algorithm
+// complexity: O(V*log(V))
+template <typename T>
+std::vector<T> Graph<T>::topoSort(void)
+{
+    std::stack<T>     buf;
+    std::vector<T>    res;
+    const T           *vPt;
+    std::map<T, bool> tmpMarked(isMarked_);
+
+    // visit function
+    std::function<void(const T &)> visit = [&](const T &v)
+    {
+        if (tmpMarked.at(v))
+        {
+            HADRON_ERROR("cannot topologically sort a cyclic graph");
+        }
+        if (!isMarked(v))
+        {
+            std::vector<T> child = getChildren(v);
+
+            tmpMarked[v] = true;
+            for (auto &c: child)
+            {
+                visit(c);
+            }
+            mark(v);
+            tmpMarked[v] = false;
+            buf.push(v);
+        }
+    };
+    
+    // reset temporary marks
+    for (auto &v: tmpMarked)
+    {
+        tmpMarked.at(v.first) = false;
+    }
+    
+    // loop on unmarked vertices
+    unmarkAll();
+    vPt = getFirstUnmarked();
+    while (vPt)
+    {
+        visit(*vPt);
+        vPt = getFirstUnmarked();
+    }
+    unmarkAll();
+    
+    // create result vector
+    while (!buf.empty())
+    {
+        res.push_back(buf.top());
+        buf.pop();
+    }
+    
+    return res;
+}
+
+// random version of the topological sort
+// complexity: O(V*log(V))
+template <typename T>
+template <typename Gen>
+std::vector<T> Graph<T>::topoSort(Gen &gen)
+{
+    std::stack<T>     buf;
+    std::vector<T>    res;
+    const T           *vPt;
+    std::map<T, bool> tmpMarked(isMarked_);
+    
+    // visit function
+    std::function<void(const T &)> visit = [&](const T &v)
+    {
+        if (tmpMarked.at(v))
+        {
+            HADRON_ERROR("cannot topologically sort a cyclic graph");
+        }
+        if (!isMarked(v))
+        {
+            std::vector<T> child = getChildren(v);
+            
+            tmpMarked[v] = true;
+            std::shuffle(child.begin(), child.end(), gen);
+            for (auto &c: child)
+            {
+                visit(c);
+            }
+            mark(v);
+            tmpMarked[v] = false;
+            buf.push(v);
+        }
+    };
+    
+    // reset temporary marks
+    for (auto &v: tmpMarked)
+    {
+        tmpMarked.at(v.first) = false;
+    }
+    
+    // loop on unmarked vertices
+    unmarkAll();
+    vPt = getRandomUnmarked(gen);
+    while (vPt)
+    {
+        visit(*vPt);
+        vPt = getRandomUnmarked(gen);
+    }
+    unmarkAll();
+    
+    // create result vector
+    while (!buf.empty())
+    {
+        res.push_back(buf.top());
+        buf.pop();
+    }
+    
+    return res;
+}
+
+// generate all possible topological sorts
+// Y. L. Varol & D. Rotem, Comput. J. 24(1), pp. 83–84, 1981
+// http://comjnl.oupjournals.org/cgi/doi/10.1093/comjnl/24.1.83
+// complexity: O(V*log(V)) (from the paper, but really ?)
+template <typename T>
+std::vector<std::vector<T>> Graph<T>::allTopoSort(void)
+{
+    std::vector<std::vector<T>>    res;
+    std::map<T, std::map<T, bool>> iMat;
+    
+    // create incidence matrix
+    for (auto &v1: isMarked_)
+    for (auto &v2: isMarked_)
+    {
+        iMat[v1.first][v2.first] = false;
+    }
+    for (auto &v: isMarked_)
+    {
+        auto cVec = getChildren(v.first);
+        
+        for (auto &c: cVec)
+        {
+            iMat[v.first][c] = true;
+        }
+    }
+    
+    // generate initial topological sort
+    res.push_back(topoSort());
+    
+    // generate all other topological sorts by permutation
+    std::vector<T>            p = res[0];
+    const unsigned int        n = size();
+    std::vector<unsigned int> loc(n);
+    unsigned int              i, k, k1;
+    T                         obj_k, obj_k1;
+    bool                      isFinal;
+    
+    for (unsigned int j = 0; j < n; ++j)
+    {
+        loc[j] = j;
+    }
+    i = 0;
+    while (i < n-1)
+    {
+        k      = loc[i];
+        k1     = k + 1;
+        obj_k  = p[k];
+        if (k1 >= n)
+        {
+            isFinal = true;
+            obj_k1  = obj_k;
+        }
+        else
+        {
+            isFinal = false;
+            obj_k1  = p[k1];
+        }
+        if (iMat[res[0][i]][obj_k1] or isFinal)
+        {
+            for (unsigned int l = k; l >= i + 1; --l)
+            {
+                p[l]   = p[l-1];
+            }
+            p[i]   = obj_k;
+            loc[i] = i;
+            i++;
+        }
+        else
+        {
+            p[k]   = obj_k1;
+            p[k1]  = obj_k;
+            loc[i] = k1;
+            i      = 0;
+            res.push_back(p);
+        }
+    }
+    
+    return res;
+}
+
+// build depedency matrix from topological sorts ///////////////////////////////
+// complexity: something like O(V^2*log(V!))
+template <typename T>
+std::map<T, std::map<T, bool>>
+makeDependencyMatrix(const std::vector<std::vector<T>> &topSort)
+{
+    std::map<T, std::map<T, bool>> m;
+    const std::vector<T>           &vList = topSort[0];
+    
+    for (auto &v1: vList)
+    for (auto &v2: vList)
+    {
+        bool dep = true;
+        
+        for (auto &t: topSort)
+        {
+            auto i1 = std::find(t.begin(), t.end(), v1);
+            auto i2 = std::find(t.begin(), t.end(), v2);
+            
+            dep = dep and (i1 - i2 > 0);
+            if (!dep) break;
+        }
+        m[v1][v2] = dep;
+    }
+    
+    return m;
+}
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_Graph_hpp_

extras/Hadrons/HadronsXmlRun.cc

@@ -0,0 +1,80 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/HadronsXmlRun.cc
+
+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/Application.hpp>
+
+using namespace Grid;
+using namespace QCD;
+using namespace Hadrons;
+
+int main(int argc, char *argv[])
+{
+    // parse command line
+    std::string parameterFileName, scheduleFileName = "";
+    
+    if (argc < 2)
+    {
+        std::cerr << "usage: " << argv[0] << " <parameter file> [<precomputed schedule>] [Grid options]";
+        std::cerr << std::endl;
+        std::exit(EXIT_FAILURE);
+    }
+    parameterFileName = argv[1];
+    if (argc > 2)
+    {
+        if (argv[2][0] != '-')
+        {
+            scheduleFileName = argv[2];
+        }
+    }
+    
+    // initialization
+    Grid_init(&argc, &argv);
+    HadronsLogError.Active(GridLogError.isActive());
+    HadronsLogWarning.Active(GridLogWarning.isActive());
+    HadronsLogMessage.Active(GridLogMessage.isActive());
+    HadronsLogIterative.Active(GridLogIterative.isActive());
+    HadronsLogDebug.Active(GridLogDebug.isActive());
+    LOG(Message) << "Grid initialized" << std::endl;
+    
+    // execution
+    Application application(parameterFileName);
+    
+    application.parseParameterFile(parameterFileName);
+    if (!scheduleFileName.empty())
+    {
+        application.loadSchedule(scheduleFileName);
+    }
+    application.run();
+    
+    // epilogue
+    LOG(Message) << "Grid is finalizing now" << std::endl;
+    Grid_finalize();
+    
+    return EXIT_SUCCESS;
+}

extras/Hadrons/HadronsXmlSchedule.cc

@@ -0,0 +1,72 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/HadronsXmlSchedule.cc
+
+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/Application.hpp>
+
+using namespace Grid;
+using namespace QCD;
+using namespace Hadrons;
+
+int main(int argc, char *argv[])
+{
+    // parse command line
+    std::string parameterFileName, scheduleFileName;
+    
+    if (argc < 3)
+    {
+        std::cerr << "usage: " << argv[0] << " <parameter file> <schedule output> [Grid options]";
+        std::cerr << std::endl;
+        std::exit(EXIT_FAILURE);
+    }
+    parameterFileName = argv[1];
+    scheduleFileName  = argv[2];
+    
+    // initialization
+    Grid_init(&argc, &argv);
+    HadronsLogError.Active(GridLogError.isActive());
+    HadronsLogWarning.Active(GridLogWarning.isActive());
+    HadronsLogMessage.Active(GridLogMessage.isActive());
+    HadronsLogIterative.Active(GridLogIterative.isActive());
+    HadronsLogDebug.Active(GridLogDebug.isActive());
+    LOG(Message) << "Grid initialized" << std::endl;
+    
+    // execution
+    Application application;
+    
+    application.parseParameterFile(parameterFileName);
+    application.schedule();
+    application.printSchedule();
+    application.saveSchedule(scheduleFileName);
+    
+    // epilogue
+    LOG(Message) << "Grid is finalizing now" << std::endl;
+    Grid_finalize();
+    
+    return EXIT_SUCCESS;
+}

extras/Hadrons/Makefile.am

@@ -0,0 +1,29 @@
+lib_LIBRARIES = libHadrons.a
+bin_PROGRAMS  = HadronsXmlRun HadronsXmlSchedule
+
+include modules.inc
+
+libHadrons_a_SOURCES = \
+    $(modules_cc)      \
+    Application.cc     \
+    Environment.cc     \
+    Global.cc          \
+    Module.cc
+libHadrons_adir = $(pkgincludedir)/Hadrons
+nobase_libHadrons_a_HEADERS = \
+	$(modules_hpp)            \
+	Application.hpp           \
+	Environment.hpp           \
+	Factory.hpp               \
+	GeneticScheduler.hpp      \
+	Global.hpp                \
+	Graph.hpp                 \
+	Module.hpp                \
+	Modules.hpp               \
+	ModuleFactory.hpp
+
+HadronsXmlRun_SOURCES = HadronsXmlRun.cc
+HadronsXmlRun_LDADD   = libHadrons.a -lGrid
+
+HadronsXmlSchedule_SOURCES = HadronsXmlSchedule.cc
+HadronsXmlSchedule_LDADD   = libHadrons.a -lGrid

extras/Hadrons/Module.cc

@@ -0,0 +1,71 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Module.cc
+
+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/Module.hpp>
+
+using namespace Grid;
+using namespace QCD;
+using namespace Hadrons;
+
+/******************************************************************************
+ *                       ModuleBase implementation                            *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+ModuleBase::ModuleBase(const std::string name)
+: name_(name)
+, env_(Environment::getInstance())
+{}
+
+// access //////////////////////////////////////////////////////////////////////
+std::string ModuleBase::getName(void) const
+{
+    return name_;
+}
+
+Environment & ModuleBase::env(void) const
+{
+    return env_;
+}
+
+// get factory registration name if available
+std::string ModuleBase::getRegisteredName(void)
+{
+    HADRON_ERROR("module '" + getName() + "' has a type not registered"
+                 + " in the factory");
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+void ModuleBase::operator()(void)
+{
+    setup();
+    if (!env().isDryRun())
+    {
+        execute();
+    }
+}

extras/Hadrons/Module.hpp

@@ -0,0 +1,198 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Module.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_Module_hpp_
+#define Hadrons_Module_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Environment.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+// module registration macros
+#define MODULE_REGISTER(mod, base)\
+class mod: public base\
+{\
+public:\
+    typedef base Base;\
+    using Base::Base;\
+    virtual std::string getRegisteredName(void)\
+    {\
+        return std::string(#mod);\
+    }\
+};\
+class mod##ModuleRegistrar\
+{\
+public:\
+    mod##ModuleRegistrar(void)\
+    {\
+        ModuleFactory &modFac = ModuleFactory::getInstance();\
+        modFac.registerBuilder(#mod, [&](const std::string name)\
+                              {\
+                                  return std::unique_ptr<mod>(new mod(name));\
+                              });\
+    }\
+};\
+static mod##ModuleRegistrar mod##ModuleRegistrarInstance;
+
+#define MODULE_REGISTER_NS(mod, base, ns)\
+class mod: public base\
+{\
+public:\
+    typedef base Base;\
+    using Base::Base;\
+    virtual std::string getRegisteredName(void)\
+    {\
+        return std::string(#ns "::" #mod);\
+    }\
+};\
+class ns##mod##ModuleRegistrar\
+{\
+public:\
+    ns##mod##ModuleRegistrar(void)\
+    {\
+        ModuleFactory &modFac = ModuleFactory::getInstance();\
+        modFac.registerBuilder(#ns "::" #mod, [&](const std::string name)\
+                              {\
+                                  return std::unique_ptr<ns::mod>(new ns::mod(name));\
+                              });\
+    }\
+};\
+static ns##mod##ModuleRegistrar ns##mod##ModuleRegistrarInstance;
+
+#define ARG(...) __VA_ARGS__
+
+/******************************************************************************
+ *                            Module class                                    *
+ ******************************************************************************/
+// base class
+class ModuleBase
+{
+public:
+    // constructor
+    ModuleBase(const std::string name);
+    // destructor
+    virtual ~ModuleBase(void) = default;
+    // access
+    std::string getName(void) const;
+    Environment &env(void) const;
+    // get factory registration name if available
+    virtual std::string getRegisteredName(void);
+    // dependencies/products
+    virtual std::vector<std::string> getInput(void) = 0;
+    virtual std::vector<std::string> getOutput(void) = 0;
+    // parse parameters
+    virtual void parseParameters(XmlReader &reader, const std::string name) = 0;
+    virtual void saveParameters(XmlWriter &writer, const std::string name) = 0;
+    // setup
+    virtual void setup(void) {};
+    // execution
+    void operator()(void);
+    virtual void execute(void) = 0;
+private:
+    std::string name_;
+    Environment &env_;
+};
+
+// derived class, templating the parameter class
+template <typename P>
+class Module: public ModuleBase
+{
+public:
+    typedef P Par;
+public:
+    // constructor
+    Module(const std::string name);
+    // destructor
+    virtual ~Module(void) = default;
+    // parse parameters
+    virtual void parseParameters(XmlReader &reader, const std::string name);
+    virtual void saveParameters(XmlWriter &writer, const std::string name);
+    // parameter access
+    const P & par(void) const;
+    void      setPar(const P &par);
+private:
+    P par_;
+};
+
+// no parameter type
+class NoPar {};
+
+template <>
+class Module<NoPar>: public ModuleBase
+{
+public:
+    // constructor
+    Module(const std::string name): ModuleBase(name) {};
+    // destructor
+    virtual ~Module(void) = default;
+    // parse parameters (do nothing)
+    virtual void parseParameters(XmlReader &reader, const std::string name) {};
+    virtual void saveParameters(XmlWriter &writer, const std::string name)
+    {
+        push(writer, "options");
+        pop(writer);
+    };
+};
+
+/******************************************************************************
+ *                           Template implementation                          *
+ ******************************************************************************/
+template <typename P>
+Module<P>::Module(const std::string name)
+: ModuleBase(name)
+{}
+
+template <typename P>
+void Module<P>::parseParameters(XmlReader &reader, const std::string name)
+{
+    read(reader, name, par_);
+}
+
+template <typename P>
+void Module<P>::saveParameters(XmlWriter &writer, const std::string name)
+{
+    write(writer, name, par_);
+}
+
+template <typename P>
+const P & Module<P>::par(void) const
+{
+    return par_;
+}
+
+template <typename P>
+void Module<P>::setPar(const P &par)
+{
+    par_ = par;
+}
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_Module_hpp_

extras/Hadrons/ModuleFactory.hpp

@@ -0,0 +1,49 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/ModuleFactory.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_ModuleFactory_hpp_
+#define Hadrons_ModuleFactory_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Factory.hpp>
+#include <Grid/Hadrons/Module.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                            ModuleFactory                                   *
+ ******************************************************************************/
+class ModuleFactory: public Factory<ModuleBase>
+{
+    SINGLETON_DEFCTOR(ModuleFactory)
+};
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_ModuleFactory_hpp_

extras/Hadrons/Modules.hpp

@@ -0,0 +1,40 @@
+/*************************************************************************************
+
+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/Meson.hpp>
+#include <Grid/Hadrons/Modules/MGauge/Load.hpp>
+#include <Grid/Hadrons/Modules/MGauge/Random.hpp>
+#include <Grid/Hadrons/Modules/MGauge/Unit.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/Z2.hpp>
+#include <Grid/Hadrons/Modules/Quark.hpp>

extras/Hadrons/Modules/MAction/DWF.hpp

@@ -0,0 +1,134 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MAction/DWF.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_DWF_hpp_
+#define Hadrons_DWF_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                     Domain wall quark action                               *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MAction)
+
+class DWFPar: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(DWFPar,
+                                    std::string, gauge,
+                                    unsigned int, Ls,
+                                    double      , mass,
+                                    double      , M5);
+};
+
+template <typename FImpl>
+class TDWF: public Module<DWFPar>
+{
+public:
+    TYPE_ALIASES(FImpl,);
+public:
+    // constructor
+    TDWF(const std::string name);
+    // destructor
+    virtual ~TDWF(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(DWF, TDWF<FIMPL>, MAction);
+
+/******************************************************************************
+ *                        DWF template implementation                         *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+template <typename FImpl>
+TDWF<FImpl>::TDWF(const std::string name)
+: Module<DWFPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+template <typename FImpl>
+std::vector<std::string> TDWF<FImpl>::getInput(void)
+{
+    std::vector<std::string> in = {par().gauge};
+    
+    return in;
+}
+
+template <typename FImpl>
+std::vector<std::string> TDWF<FImpl>::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TDWF<FImpl>::setup(void)
+{
+    unsigned int size;
+    
+    size = 2*env().template lattice4dSize<typename FImpl::DoubledGaugeField>();
+    env().registerObject(getName(), size, par().Ls);
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TDWF<FImpl>::execute(void)
+{
+    LOG(Message) << "Setting up domain wall fermion matrix with m= "
+                 << par().mass << ", M5= " << par().M5 << " and Ls= "
+                 << par().Ls << " using gauge field '" << par().gauge << "'"
+                 << 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);
+    FMat *fMatPt = new DomainWallFermion<FImpl>(U, g5, grb5, g4, grb4,
+                                                par().mass, par().M5);
+    env().setObject(getName(), fMatPt);
+}
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_DWF_hpp_

extras/Hadrons/Modules/MAction/Wilson.hpp

@@ -0,0 +1,126 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MAction/Wilson.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_Wilson_hpp_
+#define Hadrons_Wilson_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                            TWilson quark action                            *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MAction)
+
+class WilsonPar: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(WilsonPar,
+                                    std::string, gauge,
+                                    double     , mass);
+};
+
+template <typename FImpl>
+class TWilson: public Module<WilsonPar>
+{
+public:
+    TYPE_ALIASES(FImpl,);
+public:
+    // constructor
+    TWilson(const std::string name);
+    // destructor
+    virtual ~TWilson(void) = default;
+    // dependencies/products
+    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(Wilson, TWilson<FIMPL>, MAction);
+
+/******************************************************************************
+ *                     TWilson template implementation                        *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+template <typename FImpl>
+TWilson<FImpl>::TWilson(const std::string name)
+: Module<WilsonPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+template <typename FImpl>
+std::vector<std::string> TWilson<FImpl>::getInput(void)
+{
+    std::vector<std::string> in = {par().gauge};
+    
+    return in;
+}
+
+template <typename FImpl>
+std::vector<std::string> TWilson<FImpl>::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TWilson<FImpl>::setup(void)
+{
+    unsigned int size;
+    
+    size = 2*env().template lattice4dSize<typename FImpl::DoubledGaugeField>();
+    env().registerObject(getName(), size);
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TWilson<FImpl>::execute()
+{
+    LOG(Message) << "Setting up TWilson fermion matrix with m= " << par().mass
+                 << " using gauge field '" << par().gauge << "'" << 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);
+    env().setObject(getName(), fMatPt);
+}
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_Wilson_hpp_

extras/Hadrons/Modules/MContraction/Baryon.hpp

@@ -0,0 +1,131 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MContraction/Baryon.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_Baryon_hpp_
+#define Hadrons_Baryon_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                               Baryon                                       *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MContraction)
+
+class BaryonPar: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(BaryonPar,
+                                    std::string, q1,
+                                    std::string, q2,
+                                    std::string, q3,
+                                    std::string, output);
+};
+
+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);
+    class Result: Serializable
+    {
+    public:
+        GRID_SERIALIZABLE_CLASS_MEMBERS(Result,
+                                        std::vector<std::vector<std::vector<Complex>>>, corr);
+    };
+public:
+    // constructor
+    TBaryon(const std::string name);
+    // destructor
+    virtual ~TBaryon(void) = default;
+    // dependency relation
+    virtual std::vector<std::string> getInput(void);
+    virtual std::vector<std::string> getOutput(void);
+    // execution
+    virtual void execute(void);
+};
+
+MODULE_REGISTER_NS(Baryon, ARG(TBaryon<FIMPL, FIMPL, FIMPL>), MContraction);
+
+/******************************************************************************
+ *                         TBaryon implementation                             *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+template <typename FImpl1, typename FImpl2, typename FImpl3>
+TBaryon<FImpl1, FImpl2, FImpl3>::TBaryon(const std::string name)
+: Module<BaryonPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+template <typename FImpl1, typename FImpl2, typename FImpl3>
+std::vector<std::string> TBaryon<FImpl1, FImpl2, FImpl3>::getInput(void)
+{
+    std::vector<std::string> input = {par().q1, par().q2, par().q3};
+    
+    return input;
+}
+
+template <typename FImpl1, typename FImpl2, typename FImpl3>
+std::vector<std::string> TBaryon<FImpl1, FImpl2, FImpl3>::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+template <typename FImpl1, typename FImpl2, typename FImpl3>
+void TBaryon<FImpl1, FImpl2, FImpl3>::execute(void)
+{
+    LOG(Message) << "Computing baryon contractions '" << getName() << "' using"
+                 << " quarks '" << par().q1 << "', '" << par().q2 << "', and '"
+                 << par().q3 << "'" << std::endl;
+    
+    XmlWriter             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);
+    LatticeComplex        c(env().getGrid());
+    Result                result;
+    
+    // FIXME: do contractions
+    
+    write(writer, "meson", result);
+}
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_Baryon_hpp_

extras/Hadrons/Modules/MContraction/Meson.hpp

@@ -0,0 +1,137 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MContraction/Meson.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_Meson_hpp_
+#define Hadrons_Meson_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                                TMeson                                       *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MContraction)
+
+class MesonPar: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(MesonPar,
+                                    std::string,    q1,
+                                    std::string,    q2,
+                                    std::string,    output,
+                                    Gamma::Algebra, gammaSource,
+                                    Gamma::Algebra, gammaSink);
+};
+
+template <typename FImpl1, typename FImpl2>
+class TMeson: public Module<MesonPar>
+{
+public:
+    TYPE_ALIASES(FImpl1, 1);
+    TYPE_ALIASES(FImpl2, 2);
+    class Result: Serializable
+    {
+    public:
+        GRID_SERIALIZABLE_CLASS_MEMBERS(Result, std::vector<Complex>, corr);
+    };
+public:
+    // constructor
+    TMeson(const std::string name);
+    // destructor
+    virtual ~TMeson(void) = default;
+    // dependencies/products
+    virtual std::vector<std::string> getInput(void);
+    virtual std::vector<std::string> getOutput(void);
+    // execution
+    virtual void execute(void);
+};
+
+MODULE_REGISTER_NS(Meson, ARG(TMeson<FIMPL, FIMPL>), MContraction);
+
+/******************************************************************************
+ *                           TMeson implementation                            *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+template <typename FImpl1, typename FImpl2>
+TMeson<FImpl1, FImpl2>::TMeson(const std::string name)
+: Module<MesonPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+template <typename FImpl1, typename FImpl2>
+std::vector<std::string> TMeson<FImpl1, FImpl2>::getInput(void)
+{
+    std::vector<std::string> input = {par().q1, par().q2};
+    
+    return input;
+}
+
+template <typename FImpl1, typename FImpl2>
+std::vector<std::string> TMeson<FImpl1, FImpl2>::getOutput(void)
+{
+    std::vector<std::string> output = {getName()};
+    
+    return output;
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+template <typename FImpl1, typename FImpl2>
+void TMeson<FImpl1, FImpl2>::execute(void)
+{
+    LOG(Message) << "Computing meson contractions '" << getName() << "' using"
+                 << " 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;
+    
+    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)
+    {
+        result.corr[t] = TensorRemove(buf[t]);
+    }
+    write(writer, "meson", result);
+}
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_Meson_hpp_

extras/Hadrons/Modules/MGauge/Load.cc

@@ -0,0 +1,78 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MGauge/Load.cc
+
+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/MGauge/Load.hpp>
+
+using namespace Grid;
+using namespace Hadrons;
+using namespace MGauge;
+
+/******************************************************************************
+*                           TLoad implementation                               *
+******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+TLoad::TLoad(const std::string name)
+: Module<LoadPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+std::vector<std::string> TLoad::getInput(void)
+{
+    std::vector<std::string> in;
+    
+    return in;
+}
+
+std::vector<std::string> TLoad::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+void TLoad::setup(void)
+{
+    env().registerLattice<LatticeGaugeField>(getName());
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+void TLoad::execute(void)
+{
+    NerscField  header;
+    std::string fileName = par().file + "."
+                           + std::to_string(env().getTrajectory());
+    
+    LOG(Message) << "Loading NERSC configuration from file '" << fileName
+                 << "'" << std::endl;
+    LatticeGaugeField &U = *env().createLattice<LatticeGaugeField>(getName());
+    NerscIO::readConfiguration(U, header, fileName);
+    LOG(Message) << "NERSC header:" << std::endl;
+    dump_nersc_header(header, LOG(Message));
+}

extras/Hadrons/Modules/MGauge/Load.hpp

@@ -0,0 +1,73 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MGauge/Load.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_Load_hpp_
+#define Hadrons_Load_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                         Load a NERSC configuration                         *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MGauge)
+
+class LoadPar: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(LoadPar,
+                                    std::string, file);
+};
+
+class TLoad: public Module<LoadPar>
+{
+public:
+    // constructor
+    TLoad(const std::string name);
+    // destructor
+    virtual ~TLoad(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(Load, TLoad, MGauge);
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_Load_hpp_

extras/Hadrons/Modules/MGauge/Random.cc

@@ -0,0 +1,69 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MGauge/Random.cc
+
+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/MGauge/Random.hpp>
+
+using namespace Grid;
+using namespace Hadrons;
+using namespace MGauge;
+
+/******************************************************************************
+*                           TRandom implementation                            *
+******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+TRandom::TRandom(const std::string name)
+: Module<NoPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+std::vector<std::string> TRandom::getInput(void)
+{
+    return std::vector<std::string>();
+}
+
+std::vector<std::string> TRandom::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+void TRandom::setup(void)
+{
+    env().registerLattice<LatticeGaugeField>(getName());
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+void TRandom::execute(void)
+{
+    LOG(Message) << "Generating random gauge configuration" << std::endl;
+    LatticeGaugeField &U = *env().createLattice<LatticeGaugeField>(getName());
+    SU3::HotConfiguration(*env().get4dRng(), U);
+}

extras/Hadrons/Modules/MGauge/Random.hpp

@@ -0,0 +1,66 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MGauge/Random.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_Random_hpp_
+#define Hadrons_Random_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                             Random gauge                                   *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MGauge)
+
+class TRandom: public Module<NoPar>
+{
+public:
+    // constructor
+    TRandom(const std::string name);
+    // destructor
+    virtual ~TRandom(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(Random, TRandom, MGauge);
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_Random_hpp_

extras/Hadrons/Modules/MGauge/Unit.cc

@@ -0,0 +1,69 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MGauge/Unit.cc
+
+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/MGauge/Unit.hpp>
+
+using namespace Grid;
+using namespace Hadrons;
+using namespace MGauge;
+
+/******************************************************************************
+*                            TUnit implementation                             *
+******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+TUnit::TUnit(const std::string name)
+: Module<NoPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+std::vector<std::string> TUnit::getInput(void)
+{
+    return std::vector<std::string>();
+}
+
+std::vector<std::string> TUnit::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+void TUnit::setup(void)
+{
+    env().registerLattice<LatticeGaugeField>(getName());
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+void TUnit::execute(void)
+{
+    LOG(Message) << "Creating unit gauge configuration" << std::endl;
+    LatticeGaugeField &U = *env().createLattice<LatticeGaugeField>(getName());
+    SU3::ColdConfiguration(*env().get4dRng(), U);
+}

extras/Hadrons/Modules/MGauge/Unit.hpp

@@ -0,0 +1,66 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MGauge/Unit.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_Unit_hpp_
+#define Hadrons_Unit_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                              Unit gauge                                    *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MGauge)
+
+class TUnit: public Module<NoPar>
+{
+public:
+    // constructor
+    TUnit(const std::string name);
+    // destructor
+    virtual ~TUnit(void) = default;
+    // dependencies/products
+    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(Unit, TUnit, MGauge);
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_Unit_hpp_

extras/Hadrons/Modules/MSolver/RBPrecCG.hpp

@@ -0,0 +1,132 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MSolver/RBPrecCG.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_RBPrecCG_hpp_
+#define Hadrons_RBPrecCG_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                     Schur red-black preconditioned CG                      *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MSolver)
+
+class RBPrecCGPar: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(RBPrecCGPar,
+                                    std::string, action,
+                                    double     , residual);
+};
+
+template <typename FImpl>
+class TRBPrecCG: public Module<RBPrecCGPar>
+{
+public:
+    TYPE_ALIASES(FImpl,);
+public:
+    // constructor
+    TRBPrecCG(const std::string name);
+    // destructor
+    virtual ~TRBPrecCG(void) = default;
+    // dependencies/products
+    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(RBPrecCG, TRBPrecCG<FIMPL>, MSolver);
+
+/******************************************************************************
+ *                      TRBPrecCG template implementation                     *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+template <typename FImpl>
+TRBPrecCG<FImpl>::TRBPrecCG(const std::string name)
+: Module(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+template <typename FImpl>
+std::vector<std::string> TRBPrecCG<FImpl>::getInput(void)
+{
+    std::vector<std::string> in = {par().action};
+    
+    return in;
+}
+
+template <typename FImpl>
+std::vector<std::string> TRBPrecCG<FImpl>::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TRBPrecCG<FImpl>::setup(void)
+{
+    auto Ls = env().getObjectLs(par().action);
+    
+    env().registerObject(getName(), 0, Ls);
+    env().addOwnership(getName(), par().action);
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TRBPrecCG<FImpl>::execute(void)
+{
+    auto &mat   = *(env().template getObject<FMat>(par().action));
+    auto solver = [&mat, this](FermionField &sol, const FermionField &source)
+    {
+        ConjugateGradient<FermionField>           cg(par().residual, 10000);
+        SchurRedBlackDiagMooeeSolve<FermionField> schurSolver(cg);
+        
+        schurSolver(mat, source, sol);
+    };
+    
+    LOG(Message) << "setting up Schur red-black preconditioned CG for"
+                 << " action '" << par().action << "' with residual "
+                 << par().residual << std::endl;
+    env().setObject(getName(), new SolverFn(solver));
+}
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_RBPrecCG_hpp_

extras/Hadrons/Modules/MSource/Point.hpp

@@ -0,0 +1,135 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MSource/Point.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_Point_hpp_
+#define Hadrons_Point_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/*
+ 
+ Point source
+ ------------
+ * src_x = delta_x,position
+ 
+ * options:
+ - position: space-separated integer sequence (e.g. "0 1 1 0")
+ 
+ */
+
+/******************************************************************************
+ *                                  TPoint                                     *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MSource)
+
+class PointPar: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(PointPar,
+                                    std::string, position);
+};
+
+template <typename FImpl>
+class TPoint: public Module<PointPar>
+{
+public:
+    TYPE_ALIASES(FImpl,);
+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>, MSource);
+
+/******************************************************************************
+ *                       TPoint template 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)
+{
+    env().template registerLattice<PropagatorField>(getName());
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TPoint<FImpl>::execute(void)
+{
+    std::vector<int> position = strToVec<int>(par().position);
+    typename SitePropagator::scalar_object id;
+    
+    LOG(Message) << "Creating point source at position [" << par().position
+                 << "]" << std::endl;
+    PropagatorField &src = *env().template createLattice<PropagatorField>(getName());
+    id  = 1.;
+    src = zero;
+    pokeSite(id, src, position);
+}
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_Point_hpp_

extras/Hadrons/Modules/MSource/SeqGamma.hpp

@@ -0,0 +1,163 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MSource/SeqGamma.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_SeqGamma_hpp_
+#define Hadrons_SeqGamma_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/*
+ 
+ Sequential source
+ -----------------------------
+ * src_x = q_x * theta(x_3 - tA) * theta(tB - x_3) * gamma * exp(i x.mom)
+ 
+ * options:
+ - q: input propagator (string)
+ - tA: begin timeslice (integer)
+ - tB: end timesilce (integer)
+ - gamma: gamma product to insert (integer)
+ - mom: momentum insertion, space-separated float sequence (e.g ".1 .2 1. 0.")
+ 
+ */
+
+/******************************************************************************
+ *                         SeqGamma                                 *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MSource)
+
+class SeqGammaPar: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(SeqGammaPar,
+                                    std::string,    q,
+                                    unsigned int,   tA,
+                                    unsigned int,   tB,
+                                    Gamma::Algebra, gamma,
+                                    std::string,    mom);
+};
+
+template <typename FImpl>
+class TSeqGamma: public Module<SeqGammaPar>
+{
+public:
+    TYPE_ALIASES(FImpl,);
+public:
+    // constructor
+    TSeqGamma(const std::string name);
+    // destructor
+    virtual ~TSeqGamma(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(SeqGamma, TSeqGamma<FIMPL>, MSource);
+
+/******************************************************************************
+ *                         TSeqGamma implementation                           *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+template <typename FImpl>
+TSeqGamma<FImpl>::TSeqGamma(const std::string name)
+: Module<SeqGammaPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+template <typename FImpl>
+std::vector<std::string> TSeqGamma<FImpl>::getInput(void)
+{
+    std::vector<std::string> in = {par().q};
+    
+    return in;
+}
+
+template <typename FImpl>
+std::vector<std::string> TSeqGamma<FImpl>::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TSeqGamma<FImpl>::setup(void)
+{
+    env().template registerLattice<PropagatorField>(getName());
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TSeqGamma<FImpl>::execute(void)
+{
+    if (par().tA == par().tB)
+    {
+        LOG(Message) << "Generating gamma_" << par().gamma
+                     << " sequential source at t= " << par().tA << std::endl;
+    }
+    else
+    {
+        LOG(Message) << "Generating gamma_" << par().gamma
+                     << " sequential source for "
+                     << par().tA << " <= t <= " << par().tB << std::endl;
+    }
+    PropagatorField &src = *env().template createLattice<PropagatorField>(getName());
+    PropagatorField &q   = *env().template getObject<PropagatorField>(par().q);
+    Lattice<iScalar<vInteger>> t(env().getGrid());
+    LatticeComplex             ph(env().getGrid()), coor(env().getGrid());
+    Gamma                      g(par().gamma);
+    std::vector<Real>          p;
+    Complex                    i(0.0,1.0);
+    
+    p  = strToVec<Real>(par().mom);
+    ph = zero;
+    for(unsigned int mu = 0; mu < env().getNd(); mu++)
+    {
+        LatticeCoordinate(coor, mu);
+        ph = ph + p[mu]*coor;
+    }
+    ph = exp(i*ph);
+    LatticeCoordinate(t, Tp);
+    src = where((t >= par().tA) and (t <= par().tB), ph*(g*q), 0.*q);
+}
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_SeqGamma_hpp_

extras/Hadrons/Modules/MSource/Z2.hpp

@@ -0,0 +1,151 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MSource/Z2.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_Z2_hpp_
+#define Hadrons_Z2_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/*
+ 
+ Z_2 stochastic source
+ -----------------------------
+ * src_x = eta_x * theta(x_3 - tA) * theta(tB - x_3)
+ 
+ the eta_x are independent uniform random numbers in {+/- 1 +/- i}
+ 
+ * options:
+ - tA: begin timeslice (integer)
+ - tB: end timesilce (integer)
+ 
+ */
+ 
+/******************************************************************************
+ *                          Z2 stochastic source                              *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MSource)
+
+class Z2Par: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(Z2Par,
+                                    unsigned int, tA,
+                                    unsigned int, tB);
+};
+
+template <typename FImpl>
+class TZ2: public Module<Z2Par>
+{
+public:
+    TYPE_ALIASES(FImpl,);
+public:
+    // constructor
+    TZ2(const std::string name);
+    // destructor
+    virtual ~TZ2(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(Z2, TZ2<FIMPL>, MSource);
+
+/******************************************************************************
+ *                       TZ2 template implementation                          *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+template <typename FImpl>
+TZ2<FImpl>::TZ2(const std::string name)
+: Module<Z2Par>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+template <typename FImpl>
+std::vector<std::string> TZ2<FImpl>::getInput(void)
+{
+    std::vector<std::string> in;
+    
+    return in;
+}
+
+template <typename FImpl>
+std::vector<std::string> TZ2<FImpl>::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TZ2<FImpl>::setup(void)
+{
+    env().template registerLattice<PropagatorField>(getName());
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TZ2<FImpl>::execute(void)
+{
+    Lattice<iScalar<vInteger>> t(env().getGrid());
+    LatticeComplex             eta(env().getGrid());
+    Complex                    shift(1., 1.);
+    
+    if (par().tA == par().tB)
+    {
+        LOG(Message) << "Generating Z_2 wall source at t= " << par().tA
+        << std::endl;
+    }
+    else
+    {
+        LOG(Message) << "Generating Z_2 band for " << par().tA << " <= t <= "
+        << par().tB << std::endl;
+    }
+    PropagatorField &src = *env().template createLattice<PropagatorField>(getName());
+    LatticeCoordinate(t, Tp);
+    bernoulli(*env().get4dRng(), eta);
+    eta = (2.*eta - shift)*(1./::sqrt(2.));
+    eta = where((t >= par().tA) and (t <= par().tB), eta, 0.*eta);
+    src = 1.;
+    src = src*eta;
+}
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_Z2_hpp_

extras/Hadrons/Modules/Quark.hpp

@@ -0,0 +1,185 @@
+/*************************************************************************************
+
+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_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                               TQuark                                       *
+ ******************************************************************************/
+class QuarkPar: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(QuarkPar,
+                                    std::string, source,
+                                    std::string, solver);
+};
+
+template <typename FImpl>
+class TQuark: public Module<QuarkPar>
+{
+public:
+    TYPE_ALIASES(FImpl,);
+public:
+    // constructor
+    TQuark(const std::string name);
+    // destructor
+    virtual ~TQuark(void) = default;
+    // dependencies/products
+    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:
+    unsigned int Ls_;
+    SolverFn     *solver_{nullptr};
+};
+
+MODULE_REGISTER(Quark, TQuark<FIMPL>);
+
+/******************************************************************************
+ *                          TQuark implementation                             *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+template <typename FImpl>
+TQuark<FImpl>::TQuark(const std::string name)
+: Module(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+template <typename FImpl>
+std::vector<std::string> TQuark<FImpl>::getInput(void)
+{
+    std::vector<std::string> in = {par().source, par().solver};
+    
+    return in;
+}
+
+template <typename FImpl>
+std::vector<std::string> TQuark<FImpl>::getOutput(void)
+{
+    std::vector<std::string> out = {getName(), getName() + "_5d"};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TQuark<FImpl>::setup(void)
+{
+    Ls_ = env().getObjectLs(par().solver);
+    env().template registerLattice<PropagatorField>(getName());
+    if (Ls_ > 1)
+    {
+        env().template registerLattice<PropagatorField>(getName() + "_5d", Ls_);
+    }
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TQuark<FImpl>::execute(void)
+{
+    LOG(Message) << "Computing quark propagator '" << getName() << "'"
+                 << std::endl;
+    
+    FermionField    source(env().getGrid(Ls_)), sol(env().getGrid(Ls_)),
+                    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);
+    SolverFn        &solver  = *env().template getObject<SolverFn>(par().solver);
+    if (Ls_ > 1)
+    {
+        env().template createLattice<PropagatorField>(getName());
+    }
+    
+    LOG(Message) << "Inverting using solver '" << par().solver
+                 << "' on source '" << par().source << "'" << std::endl;
+    for (unsigned int s = 0; s < Ns; ++s)
+    for (unsigned int c = 0; c < Nc; ++c)
+    {
+        LOG(Message) << "Inversion for spin= " << s << ", color= " << c
+        << std::endl;
+        // source conversion for 4D sources
+        if (!env().isObject5d(par().source))
+        {
+            if (Ls_ == 1)
+            {
+                PropToFerm(source, fullSrc, s, c);
+            }
+            else
+            {
+                source = zero;
+                PropToFerm(tmp, fullSrc, s, c);
+                InsertSlice(tmp, source, 0, 0);
+                InsertSlice(tmp, source, Ls_-1, 0);
+                axpby_ssp_pplus(source, 0., source, 1., source, 0, 0);
+                axpby_ssp_pminus(source, 0., source, 1., source, Ls_-1, Ls_-1);
+            }
+        }
+        // source conversion for 5D sources
+        else
+        {
+            if (Ls_ != env().getObjectLs(par().source))
+            {
+                HADRON_ERROR("Ls mismatch between quark action and source");
+            }
+            else
+            {
+                PropToFerm(source, fullSrc, s, c);
+            }
+        }
+        sol = zero;
+        solver(sol, source);
+        FermToProp(prop, sol, s, c);
+        // create 4D propagators from 5D one if necessary
+        if (Ls_ > 1)
+        {
+            PropagatorField &p4d =
+                *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);
+            ExtractSlice(tmp, sol, 0, 0);
+            FermToProp(p4d, tmp, s, c);
+        }
+    }
+}
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_Quark_hpp_

extras/Hadrons/Modules/templates/Module.cc.template

@@ -0,0 +1,39 @@
+#include <Grid/Hadrons/Modules/___FILEBASENAME___.hpp>
+
+using namespace Grid;
+using namespace Hadrons;
+
+/******************************************************************************
+*                  T___FILEBASENAME___ implementation                             *
+******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+T___FILEBASENAME___::T___FILEBASENAME___(const std::string name)
+: Module<___FILEBASENAME___Par>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+std::vector<std::string> T___FILEBASENAME___::getInput(void)
+{
+    std::vector<std::string> in;
+    
+    return in;
+}
+
+std::vector<std::string> T___FILEBASENAME___::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+void T___FILEBASENAME___::setup(void)
+{
+
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+void T___FILEBASENAME___::execute(void)
+{
+
+}

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

@@ -0,0 +1,40 @@
+#ifndef Hadrons____FILEBASENAME____hpp_
+#define Hadrons____FILEBASENAME____hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                         ___FILEBASENAME___                                 *
+ ******************************************************************************/
+class ___FILEBASENAME___Par: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(___FILEBASENAME___Par, 
+                                    unsigned int, i);
+};
+
+class T___FILEBASENAME___: public Module<___FILEBASENAME___Par>
+{
+public:
+    // constructor
+    T___FILEBASENAME___(const std::string name);
+    // destructor
+    virtual ~T___FILEBASENAME___(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(___FILEBASENAME___, T___FILEBASENAME___);
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons____FILEBASENAME____hpp_

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

@@ -0,0 +1,40 @@
+#include <Grid/Hadrons/Modules/___NAMESPACE___/___FILEBASENAME___.hpp>
+
+using namespace Grid;
+using namespace Hadrons;
+using namespace ___NAMESPACE___;
+
+/******************************************************************************
+*                  T___FILEBASENAME___ implementation                             *
+******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+T___FILEBASENAME___::T___FILEBASENAME___(const std::string name)
+: Module<___FILEBASENAME___Par>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+std::vector<std::string> T___FILEBASENAME___::getInput(void)
+{
+    std::vector<std::string> in;
+    
+    return in;
+}
+
+std::vector<std::string> T___FILEBASENAME___::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+void T___FILEBASENAME___::setup(void)
+{
+
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+void T___FILEBASENAME___::execute(void)
+{
+
+}

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

@@ -0,0 +1,44 @@
+#ifndef Hadrons____FILEBASENAME____hpp_
+#define Hadrons____FILEBASENAME____hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                         ___FILEBASENAME___                                 *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(___NAMESPACE___)
+
+class ___FILEBASENAME___Par: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(___FILEBASENAME___Par,
+                                    unsigned int, i);
+};
+
+class T___FILEBASENAME___: public Module<___FILEBASENAME___Par>
+{
+public:
+    // constructor
+    T___FILEBASENAME___(const std::string name);
+    // destructor
+    virtual ~T___FILEBASENAME___(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(___FILEBASENAME___, T___FILEBASENAME___, ___NAMESPACE___);
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons____FILEBASENAME____hpp_

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

@@ -0,0 +1,81 @@
+#ifndef Hadrons____FILEBASENAME____hpp_
+#define Hadrons____FILEBASENAME____hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                         ___FILEBASENAME___                                 *
+ ******************************************************************************/
+class ___FILEBASENAME___Par: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(___FILEBASENAME___Par,
+                                    unsigned int, i);
+};
+
+template <typename FImpl>
+class T___FILEBASENAME___: public Module<___FILEBASENAME___Par>
+{
+public:
+    // constructor
+    T___FILEBASENAME___(const std::string name);
+    // destructor
+    virtual ~T___FILEBASENAME___(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(___FILEBASENAME___, T___FILEBASENAME___<FIMPL>);
+
+/******************************************************************************
+ *                 T___FILEBASENAME___ implementation                             *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+template <typename FImpl>
+T___FILEBASENAME___<FImpl>::T___FILEBASENAME___(const std::string name)
+: Module<___FILEBASENAME___Par>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+template <typename FImpl>
+std::vector<std::string> T___FILEBASENAME___<FImpl>::getInput(void)
+{
+    std::vector<std::string> in;
+    
+    return in;
+}
+
+template <typename FImpl>
+std::vector<std::string> T___FILEBASENAME___<FImpl>::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void T___FILEBASENAME___<FImpl>::setup(void)
+{
+    
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void T___FILEBASENAME___<FImpl>::execute(void)
+{
+    
+}
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons____FILEBASENAME____hpp_

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

@@ -0,0 +1,85 @@
+#ifndef Hadrons____FILEBASENAME____hpp_
+#define Hadrons____FILEBASENAME____hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                         ___FILEBASENAME___                                 *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(___NAMESPACE___)
+
+class ___FILEBASENAME___Par: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(___FILEBASENAME___Par,
+                                    unsigned int, i);
+};
+
+template <typename FImpl>
+class T___FILEBASENAME___: public Module<___FILEBASENAME___Par>
+{
+public:
+    // constructor
+    T___FILEBASENAME___(const std::string name);
+    // destructor
+    virtual ~T___FILEBASENAME___(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(___FILEBASENAME___, T___FILEBASENAME___<FIMPL>, ___NAMESPACE___);
+
+/******************************************************************************
+ *                 T___FILEBASENAME___ implementation                             *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+template <typename FImpl>
+T___FILEBASENAME___<FImpl>::T___FILEBASENAME___(const std::string name)
+: Module<___FILEBASENAME___Par>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+template <typename FImpl>
+std::vector<std::string> T___FILEBASENAME___<FImpl>::getInput(void)
+{
+    std::vector<std::string> in;
+    
+    return in;
+}
+
+template <typename FImpl>
+std::vector<std::string> T___FILEBASENAME___<FImpl>::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void T___FILEBASENAME___<FImpl>::setup(void)
+{
+    
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void T___FILEBASENAME___<FImpl>::execute(void)
+{
+    
+}
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons____FILEBASENAME____hpp_

extras/Hadrons/add_module.sh

@@ -0,0 +1,31 @@
+#!/usr/bin/env bash
+
+if (( $# != 1 && $# != 2)); then
+    echo "usage: `basename $0` <module name> [<namespace>]" 1>&2
+    exit 1
+fi
+NAME=$1
+NS=$2
+
+if (( $# == 1 )); then
+	if [ -e "Modules/${NAME}.cc" ] || [ -e "Modules/${NAME}.hpp" ]; then
+	    echo "error: files Modules/${NAME}.* already exists" 1>&2
+	    exit 1
+	fi
+	sed "s/___FILEBASENAME___/${NAME}/g" Modules/templates/Module.cc.template > Modules/${NAME}.cc
+	sed "s/___FILEBASENAME___/${NAME}/g" Modules/templates/Module.hpp.template > Modules/${NAME}.hpp
+elif (( $# == 2 )); then
+	mkdir -p Modules/${NS}
+	if [ -e "Modules/${NS}/${NAME}.cc" ] || [ -e "Modules/${NS}/${NAME}.hpp" ]; then
+	    echo "error: files Modules/${NS}/${NAME}.* already exists" 1>&2
+	    exit 1
+	fi
+	TMPCC=".${NS}.${NAME}.tmp.cc"
+	TMPHPP=".${NS}.${NAME}.tmp.hpp"
+	sed "s/___FILEBASENAME___/${NAME}/g" Modules/templates/Module_in_NS.cc.template  > ${TMPCC}
+	sed "s/___FILEBASENAME___/${NAME}/g" Modules/templates/Module_in_NS.hpp.template > ${TMPHPP}
+	sed "s/___NAMESPACE___/${NS}/g" ${TMPCC}  > Modules/${NS}/${NAME}.cc
+	sed "s/___NAMESPACE___/${NS}/g" ${TMPHPP} > Modules/${NS}/${NAME}.hpp
+	rm -f ${TMPCC} ${TMPHPP}
+fi
+./make_module_list.sh

extras/Hadrons/add_module_template.sh

@@ -0,0 +1,28 @@
+#!/usr/bin/env bash
+
+if (( $# != 1 && $# != 2)); then
+    echo "usage: `basename $0` <module name> [<namespace>]" 1>&2
+    exit 1
+fi
+NAME=$1
+NS=$2
+
+if (( $# == 1 )); then
+	if [ -e "Modules/${NAME}.cc" ] || [ -e "Modules/${NAME}.hpp" ]; then
+	    echo "error: files Modules/${NAME}.* already exists" 1>&2
+	    exit 1
+	fi
+	sed "s/___FILEBASENAME___/${NAME}/g" Modules/templates/Module_tmp.hpp.template > Modules/${NAME}.hpp
+elif (( $# == 2 )); then
+	mkdir -p Modules/${NS}
+	if [ -e "Modules/${NS}/${NAME}.cc" ] || [ -e "Modules/${NS}/${NAME}.hpp" ]; then
+	    echo "error: files Modules/${NS}/${NAME}.* already exists" 1>&2
+	    exit 1
+	fi
+	TMPCC=".${NS}.${NAME}.tmp.cc"
+	TMPHPP=".${NS}.${NAME}.tmp.hpp"
+	sed "s/___FILEBASENAME___/${NAME}/g" Modules/templates/Module_tmp_in_NS.hpp.template > ${TMPHPP}
+	sed "s/___NAMESPACE___/${NS}/g" ${TMPHPP} > Modules/${NS}/${NAME}.hpp
+	rm -f ${TMPCC} ${TMPHPP}
+fi
+./make_module_list.sh

extras/Hadrons/make_module_list.sh

@@ -0,0 +1,12 @@
+#!/usr/bin/env bash
+
+echo 'modules_cc =\' > modules.inc
+find Modules -name '*.cc' -type f -print | sed 's/^/  /;$q;s/$/ \\/' >> modules.inc
+echo '' >> modules.inc
+echo 'modules_hpp =\' >> modules.inc
+find Modules -name '*.hpp' -type f -print | sed 's/^/  /;$q;s/$/ \\/' >> modules.inc
+echo '' >> modules.inc
+rm -f Modules.hpp
+for f in `find Modules -name '*.hpp'`; do
+	echo "#include <Grid/Hadrons/${f}>" >> Modules.hpp
+done

extras/Hadrons/modules.inc

@@ -0,0 +1,19 @@
+modules_cc =\
+  Modules/MGauge/Load.cc \
+  Modules/MGauge/Random.cc \
+  Modules/MGauge/Unit.cc
+
+modules_hpp =\
+  Modules/MAction/DWF.hpp \
+  Modules/MAction/Wilson.hpp \
+  Modules/MContraction/Baryon.hpp \
+  Modules/MContraction/Meson.hpp \
+  Modules/MGauge/Load.hpp \
+  Modules/MGauge/Random.hpp \
+  Modules/MGauge/Unit.hpp \
+  Modules/MSolver/RBPrecCG.hpp \
+  Modules/MSource/Point.hpp \
+  Modules/MSource/SeqGamma.hpp \
+  Modules/MSource/Z2.hpp \
+  Modules/Quark.hpp
+

extras/Makefile.am

@@ -0,0 +1 @@
+SUBDIRS = Hadrons

lib/Grid.h

@@ -38,53 +38,10 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_H
 #define GRID_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 <Grid/serialisation/Serialisation.h>
-#include "Config.h"
-#include <Grid/Timer.h>
-#include <Grid/PerfCount.h>
-#include <Grid/Log.h>
-#include <Grid/AlignedAllocator.h>
-#include <Grid/Simd.h>
-#include <Grid/Threads.h>
-#include <Grid/Lexicographic.h>
-#include <Grid/Init.h>
-#include <Grid/Communicator.h> 
-#include <Grid/Cartesian.h>    
-#include <Grid/Tensors.h>      
-#include <Grid/Lattice.h>      
-#include <Grid/Cshift.h>       
-#include <Grid/Stencil.h>      
-#include <Grid/Algorithms.h>   
-#include <Grid/parallelIO/BinaryIO.h>
-#include <Grid/qcd/QCD.h>
-#include <Grid/parallelIO/NerscIO.h>
-
-#include <Grid/FFT.h>
-
-#include <Grid/qcd/hmc/NerscCheckpointer.h>
-#include <Grid/qcd/hmc/HmcRunner.h>
-
-
+#include <Grid/GridCore.h>
+#include <Grid/GridQCDcore.h>
+#include <Grid/qcd/action/Action.h>
+#include <Grid/qcd/smearing/Smearing.h>
+#include <Grid/qcd/hmc/HMC_aggregate.h>
 
 #endif

lib/GridCore.h

@@ -0,0 +1,81 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/Grid.h
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: azusayamaguchi <ayamaguc@YAMAKAZE.local>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+//
+//  Grid.h
+//  simd
+//
+//  Created by Peter Boyle on 09/05/2014.
+//  Copyright (c) 2014 University of Edinburgh. All rights reserved.
+//
+
+#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/perfmon/Timer.h>
+#include <Grid/perfmon/PerfCount.h>
+#include <Grid/log/Log.h>
+#include <Grid/allocator/AlignedAllocator.h>
+#include <Grid/simd/Simd.h>
+#include <Grid/serialisation/Serialisation.h>
+#include <Grid/threads/Threads.h>
+#include <Grid/util/Util.h>
+#include <Grid/communicator/Communicator.h> 
+#include <Grid/cartesian/Cartesian.h>    
+#include <Grid/tensors/Tensors.h>      
+#include <Grid/lattice/Lattice.h>      
+#include <Grid/cshift/Cshift.h>       
+#include <Grid/stencil/Stencil.h>      
+#include <Grid/parallelIO/BinaryIO.h>
+#include <Grid/algorithms/Algorithms.h>   
+
+#endif

lib/GridQCDcore.h

@@ -2,12 +2,12 @@
 
     Grid physics library, www.github.com/paboyle/Grid 
 
-    Source file: ./lib/qcd/hmc/HMC.cc
+    Source file: ./lib/Grid.h
 
     Copyright (C) 2015
 
 Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: neo <cossu@post.kek.jp>
+Author: azusayamaguchi <ayamaguc@YAMAKAZE.local>
 Author: paboyle <paboyle@ph.ed.ac.uk>
 
     This program is free software; you can redistribute it and/or modify
@@ -27,10 +27,16 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#include <Grid.h>
+#ifndef GRID_QCD_CORE_H
+#define GRID_QCD_CORE_H
 
-namespace Grid{
-  namespace QCD{
+/////////////////////////
+// Core Grid QCD headers
+/////////////////////////
+#include <Grid/GridCore.h>
+#include <Grid/qcd/QCD.h>
+#include <Grid/qcd/spin/Spin.h>
+#include <Grid/qcd/utils/Utils.h>
+#include <Grid/qcd/representations/Representations.h>
 
-  }
-}
+#endif

lib/Hadrons

@@ -0,0 +1 @@
+../extras/Hadrons

lib/Init.cc

@@ -1,346 +0,0 @@
-    /*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/Init.cc
-
-    Copyright (C) 2015
-
-Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Peter Boyle <peterboyle@MacBook-Pro.local>
-Author: paboyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-    *************************************************************************************/
-    /*  END LEGAL */
-/****************************************************************************/
-/* pab: Signal magic. Processor state dump is x86-64 specific               */
-/****************************************************************************/
-#include <stdlib.h>
-#include <stdio.h>
-#include <stdint.h>
-#include <unistd.h>
-#include <sys/mman.h>
-#include <sys/stat.h> 
-#include <sys/time.h>
-#include <signal.h>
-#include <iostream>
-#include <iterator>
-#include <Grid.h>
-#include <algorithm>
-#include <iterator>
-
-namespace Grid {
-
-//////////////////////////////////////////////////////
-// Convenience functions to access stadard command line arg
-// driven parallelism controls
-//////////////////////////////////////////////////////
-static std::vector<int> Grid_default_latt;
-static std::vector<int> Grid_default_mpi;
-
-int GridThread::_threads =1;
-int GridThread::_hyperthreads=1;
-int GridThread::_cores=1;
-
-const std::vector<int> &GridDefaultLatt(void)     {return Grid_default_latt;};
-const std::vector<int> &GridDefaultMpi(void)      {return Grid_default_mpi;};
-const std::vector<int> GridDefaultSimd(int dims,int nsimd)
-{
-    std::vector<int> layout(dims);
-    int nn=nsimd;
-    for(int d=dims-1;d>=0;d--){
-      if ( nn>=2) {
-	layout[d]=2;
-	nn/=2;
-      } else { 
-	layout[d]=1;
-      }
-    }
-    assert(nn==1);
-    return layout;
-}
-  
-////////////////////////////////////////////////////////////
-// Command line parsing assist for stock controls
-////////////////////////////////////////////////////////////
-std::string GridCmdOptionPayload(char ** begin, char ** end, const std::string & option)
-{
-  char ** itr = std::find(begin, end, option);
-  if (itr != end && ++itr != end) {
-    std::string payload(*itr);
-    return payload;
-  }
-  return std::string("");
-}
-bool GridCmdOptionExists(char** begin, char** end, const std::string& option)
-{
-  return std::find(begin, end, option) != end;
-}
-  // Comma separated list
-void GridCmdOptionCSL(std::string str,std::vector<std::string> & vec)
-{
-  size_t pos = 0;
-  std::string token;
-  std::string delimiter(",");
-
-  vec.resize(0);
-  while ((pos = str.find(delimiter)) != std::string::npos) {
-    token = str.substr(0, pos);
-    vec.push_back(token);
-    str.erase(0, pos + delimiter.length());
-  }
-  token = str;
-  vec.push_back(token);
-  return;
-}
-
-void GridCmdOptionIntVector(std::string &str,std::vector<int> & vec)
-{
-  vec.resize(0);
-  std::stringstream ss(str);
-  int i;
-  while (ss >> i){
-    vec.push_back(i);
-    if(std::ispunct(ss.peek()))
-      ss.ignore();
-  }    
-  return;
-}
-
-
-void GridParseLayout(char **argv,int argc,
-		     std::vector<int> &latt,
-		     std::vector<int> &mpi)
-{
-  mpi =std::vector<int>({1,1,1,1});
-  latt=std::vector<int>({8,8,8,8});
-
-  GridThread::SetMaxThreads();
-
-  std::string arg;
-  if( GridCmdOptionExists(argv,argv+argc,"--mpi") ){
-    arg = GridCmdOptionPayload(argv,argv+argc,"--mpi");
-    GridCmdOptionIntVector(arg,mpi);
-  }
-  if( GridCmdOptionExists(argv,argv+argc,"--grid") ){
-    arg= GridCmdOptionPayload(argv,argv+argc,"--grid");
-    GridCmdOptionIntVector(arg,latt);
-  }
-  if( GridCmdOptionExists(argv,argv+argc,"--threads") ){
-    std::vector<int> ompthreads(0);
-#ifndef GRID_OMP
-    std::cout << GridLogWarning << "'--threads' option used but Grid was"
-              << " not compiled with thread support" << std::endl;
-#endif
-    arg= GridCmdOptionPayload(argv,argv+argc,"--threads");
-    GridCmdOptionIntVector(arg,ompthreads);
-    assert(ompthreads.size()==1);
-    GridThread::SetThreads(ompthreads[0]);
-  }
-
-  if( GridCmdOptionExists(argv,argv+argc,"--cores") ){
-    std::vector<int> cores(0);
-    arg= GridCmdOptionPayload(argv,argv+argc,"--cores");
-    GridCmdOptionIntVector(arg,cores);
-    GridThread::SetCores(cores[0]);
-  }
-
-}
-
-std::string GridCmdVectorIntToString(const std::vector<int> & vec){
-  std::ostringstream oss;
-  std::copy(vec.begin(), vec.end(),std::ostream_iterator<int>(oss, " "));
-  return oss.str();
-}
-/////////////////////////////////////////////////////////
-//
-/////////////////////////////////////////////////////////
-void Grid_init(int *argc,char ***argv)
-{
-  CartesianCommunicator::Init(argc,argv);
-
-  // Parse command line args.
-
-  GridLogger::StopWatch.Start();
-
-  std::string arg;
-  std::vector<std::string> logstreams;
-  std::string defaultLog("Error,Warning,Message,Performance");
-
-  GridCmdOptionCSL(defaultLog,logstreams);
-  GridLogConfigure(logstreams);
-
-  if( GridCmdOptionExists(*argv,*argv+*argc,"--help") ){
-    std::cout<<GridLogMessage<<"--help : this message"<<std::endl;
-    std::cout<<GridLogMessage<<"--debug-signals : catch sigsegv and print a blame report"<<std::endl;
-    std::cout<<GridLogMessage<<"--debug-stdout  : print stdout from EVERY node"<<std::endl;    
-    std::cout<<GridLogMessage<<"--decomposition : report on default omp,mpi and simd decomposition"<<std::endl;    
-    std::cout<<GridLogMessage<<"--mpi n.n.n.n   : default MPI decomposition"<<std::endl;    
-    std::cout<<GridLogMessage<<"--threads n     : default number of OMP threads"<<std::endl;
-    std::cout<<GridLogMessage<<"--grid n.n.n.n  : default Grid size"<<std::endl;    
-    std::cout<<GridLogMessage<<"--log list      : comma separted list of streams from Error,Warning,Message,Performance,Iterative,Integrator,Debug,Colours"<<std::endl;
-    exit(EXIT_SUCCESS);
-  }
-
-  if( GridCmdOptionExists(*argv,*argv+*argc,"--log") ){
-    arg = GridCmdOptionPayload(*argv,*argv+*argc,"--log");
-    GridCmdOptionCSL(arg,logstreams);
-    GridLogConfigure(logstreams);
-  }
-
-  if( GridCmdOptionExists(*argv,*argv+*argc,"--debug-signals") ){
-    Grid_debug_handler_init();
-  }
-  if( !GridCmdOptionExists(*argv,*argv+*argc,"--debug-stdout") ){
-    Grid_quiesce_nodes();
-  }
-  if( GridCmdOptionExists(*argv,*argv+*argc,"--dslash-opt") ){
-    QCD::WilsonKernelsStatic::HandOpt=1;
-  }
-  if( GridCmdOptionExists(*argv,*argv+*argc,"--lebesgue") ){
-    LebesgueOrder::UseLebesgueOrder=1;
-  }
-
-  if( GridCmdOptionExists(*argv,*argv+*argc,"--cacheblocking") ){
-    arg= GridCmdOptionPayload(*argv,*argv+*argc,"--cacheblocking");
-    GridCmdOptionIntVector(arg,LebesgueOrder::Block);
-  }
-  GridParseLayout(*argv,*argc,
-		  Grid_default_latt,
-		  Grid_default_mpi);
-  if( GridCmdOptionExists(*argv,*argv+*argc,"--decomposition") ){
-    std::cout<<GridLogMessage<<"Grid Decomposition\n";
-    std::cout<<GridLogMessage<<"\tOpenMP threads : "<<GridThread::GetThreads()<<std::endl;
-    std::cout<<GridLogMessage<<"\tMPI tasks      : "<<GridCmdVectorIntToString(GridDefaultMpi())<<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<<"\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::string COL_RED    = GridLogColours.colour["RED"];
-  std::string COL_PURPLE = GridLogColours.colour["PURPLE"];
-  std::string COL_BLACK  = GridLogColours.colour["BLACK"];
-  std::string COL_GREEN  = GridLogColours.colour["GREEN"];
-  std::string COL_BLUE   = GridLogColours.colour["BLUE"];
-  std::string COL_YELLOW = GridLogColours.colour["YELLOW"];
-  std::string COL_BACKGROUND = GridLogColours.colour["NORMAL"];
-
-  
-  std::cout <<std::endl;
-  std::cout <<COL_RED  << "__|__|__|__|__"<<             "|__|__|_"<<COL_PURPLE<<"_|__|__|"<<                "__|__|__|__|__"<<std::endl; 
-  std::cout <<COL_RED  << "__|__|__|__|__"<<             "|__|__|_"<<COL_PURPLE<<"_|__|__|"<<                "__|__|__|__|__"<<std::endl; 
-  std::cout <<COL_RED  << "__|__|  |  |  "<<             "|  |  | "<<COL_PURPLE<<" |  |  |"<<                "  |  |  | _|__"<<std::endl; 
-  std::cout <<COL_RED  << "__|__         "<<             "        "<<COL_PURPLE<<"        "<<                "          _|__"<<std::endl; 
-  std::cout <<COL_RED  << "__|_  "<<COL_GREEN<<" GGGG   "<<COL_RED<<" RRRR   "<<COL_BLUE  <<" III    "<<COL_PURPLE<<"DDDD  "<<COL_PURPLE<<"    _|__"<<std::endl;
-  std::cout <<COL_RED  << "__|_  "<<COL_GREEN<<"G       "<<COL_RED<<" R   R  "<<COL_BLUE  <<"  I     "<<COL_PURPLE<<"D   D "<<COL_PURPLE<<"    _|__"<<std::endl;
-  std::cout <<COL_RED  << "__|_  "<<COL_GREEN<<"G       "<<COL_RED<<" R   R  "<<COL_BLUE  <<"  I     "<<COL_PURPLE<<"D    D"<<COL_PURPLE<<"    _|__"<<std::endl;
-  std::cout <<COL_BLUE << "__|_  "<<COL_GREEN<<"G  GG   "<<COL_RED<<" RRRR   "<<COL_BLUE  <<"  I     "<<COL_PURPLE<<"D    D"<<COL_GREEN <<"    _|__"<<std::endl;
-  std::cout <<COL_BLUE << "__|_  "<<COL_GREEN<<"G   G   "<<COL_RED<<" R  R   "<<COL_BLUE  <<"  I     "<<COL_PURPLE<<"D   D "<<COL_GREEN <<"    _|__"<<std::endl;
-  std::cout <<COL_BLUE << "__|_  "<<COL_GREEN<<" GGGG   "<<COL_RED<<" R   R  "<<COL_BLUE  <<" III    "<<COL_PURPLE<<"DDDD  "<<COL_GREEN <<"    _|__"<<std::endl;
-  std::cout <<COL_BLUE << "__|__         "<<             "        "<<COL_GREEN <<"        "<<                "          _|__"<<std::endl; 
-  std::cout <<COL_BLUE << "__|__|__|__|__"<<             "|__|__|_"<<COL_GREEN <<"_|__|__|"<<                "__|__|__|__|__"<<std::endl; 
-  std::cout <<COL_BLUE << "__|__|__|__|__"<<             "|__|__|_"<<COL_GREEN <<"_|__|__|"<<                "__|__|__|__|__"<<std::endl; 
-  std::cout <<COL_BLUE << "  |  |  |  |  "<<             "|  |  | "<<COL_GREEN <<" |  |  |"<<                "  |  |  |  |  "<<std::endl; 
-  std::cout << std::endl;
-  std::cout << std::endl;
-  std::cout <<COL_YELLOW<< std::endl;
-  std::cout << "Copyright (C) 2015 Peter Boyle, Azusa Yamaguchi, Guido Cossu, Antonin Portelli and other authors"<<std::endl;
-  std::cout << std::endl;
-  std::cout << "This program is free software; you can redistribute it and/or modify"<<std::endl;
-  std::cout << "it under the terms of the GNU General Public License as published by"<<std::endl;
-  std::cout << "the Free Software Foundation; either version 2 of the License, or"<<std::endl;
-  std::cout << "(at your option) any later version."<<std::endl;
-  std::cout << std::endl;
-  std::cout << "This program is distributed in the hope that it will be useful,"<<std::endl;
-  std::cout << "but WITHOUT ANY WARRANTY; without even the implied warranty of"<<std::endl;
-  std::cout << "MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the"<<std::endl;
-  std::cout << "GNU General Public License for more details."<<std::endl;
-  std::cout << COL_BACKGROUND <<std::endl;
-  std::cout << std::endl;
-}
-
-  
-void Grid_finalize(void)
-{
-#ifdef GRID_COMMS_MPI
-  MPI_Finalize();
-  Grid_unquiesce_nodes();
-#endif
-}
-
-void * Grid_backtrace_buffer[_NBACKTRACE];
-
-void Grid_sa_signal_handler(int sig,siginfo_t *si,void * ptr)
-{
-  printf("Caught signal %d\n",si->si_signo);
-  printf("  mem address %llx\n",(unsigned long long)si->si_addr);
-  printf("         code %d\n",si->si_code);
-
-  // Linux/Posix
-#ifdef __linux__
-  // And x86 64bit
-#ifdef __x86_64__
-  ucontext_t * uc= (ucontext_t *)ptr;
-  struct sigcontext *sc = (struct sigcontext *)&uc->uc_mcontext;
-  printf("  instruction %llx\n",(unsigned long long)sc->rip);
-#define REG(A)  printf("  %s %lx\n",#A,sc-> A);
-  REG(rdi);
-  REG(rsi);
-  REG(rbp);
-  REG(rbx);
-  REG(rdx);
-  REG(rax);
-  REG(rcx);
-  REG(rsp);
-  REG(rip);
-
-
-  REG(r8);
-  REG(r9);
-  REG(r10);
-  REG(r11);
-  REG(r12);
-  REG(r13);
-  REG(r14);
-  REG(r15);
-#endif
-#endif
-  BACKTRACE();
-  exit(0);
-  return;
-};
-#ifdef GRID_FPE
-#define _GNU_SOURCE
-#include <fenv.h>
-#endif
-void Grid_debug_handler_init(void)
-{
-  struct sigaction sa,osa;
-  sigemptyset (&sa.sa_mask);
-  sa.sa_sigaction= Grid_sa_signal_handler;
-  sa.sa_flags    = SA_SIGINFO;
-  sigaction(SIGSEGV,&sa,NULL);
-  sigaction(SIGTRAP,&sa,NULL);
-#ifdef GRID_FPE
-  feenableexcept( FE_INVALID|FE_OVERFLOW|FE_DIVBYZERO);
-  sigaction(SIGFPE,&sa,NULL);
-#endif
-}
-}

lib/Makefile.am

@@ -1,14 +1,34 @@
 extra_sources=
+extra_headers=
 if BUILD_COMMS_MPI
   extra_sources+=communicator/Communicator_mpi.cc
+  extra_sources+=communicator/Communicator_base.cc
+endif
+
+if BUILD_COMMS_MPI3
+  extra_sources+=communicator/Communicator_mpi3.cc
+  extra_sources+=communicator/Communicator_base.cc
+endif
+
+if BUILD_COMMS_MPI3L
+  extra_sources+=communicator/Communicator_mpi3_leader.cc
+  extra_sources+=communicator/Communicator_base.cc
 endif
 
 if BUILD_COMMS_SHMEM
   extra_sources+=communicator/Communicator_shmem.cc
+  extra_sources+=communicator/Communicator_base.cc
 endif
 
 if BUILD_COMMS_NONE
   extra_sources+=communicator/Communicator_none.cc
+  extra_sources+=communicator/Communicator_base.cc
+endif
+
+if BUILD_HDF5
+  extra_sources+=serialisation/Hdf5IO.cc 
+  extra_headers+=serialisation/Hdf5IO.h
+  extra_headers+=serialisation/Hdf5Type.h
 endif
 
 #
@@ -19,6 +39,9 @@ include Eigen.inc
 
 lib_LIBRARIES = libGrid.a
 
-libGrid_a_SOURCES              = $(CCFILES) $(extra_sources)
+CCFILES += $(extra_sources)
+HFILES  += $(extra_headers)
+
+libGrid_a_SOURCES              = $(CCFILES)
 libGrid_adir                   = $(pkgincludedir)
 nobase_dist_pkginclude_HEADERS = $(HFILES) $(eigen_files) Config.h

lib/Old/Tensor_peek.h

@@ -1,154 +0,0 @@
-    /*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/Old/Tensor_peek.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_MATH_PEEK_H
-#define GRID_MATH_PEEK_H
-namespace Grid {
-
-//////////////////////////////////////////////////////////////////////////////
-// Peek on a specific index; returns a scalar in that index, tensor inherits rest
-//////////////////////////////////////////////////////////////////////////////
-// If we hit the right index, return scalar with no further recursion
-
-//template<int Level> inline ComplexF peekIndex(const ComplexF arg) { return arg;}
-//template<int Level> inline ComplexD peekIndex(const ComplexD arg) { return arg;}
-//template<int Level> inline RealF peekIndex(const RealF arg) { return arg;}
-//template<int Level> inline RealD peekIndex(const RealD arg) { return arg;}
-#if 0
-// Scalar peek, no indices
-template<int Level,class vtype,typename std::enable_if< iScalar<vtype>::TensorLevel == Level >::type * =nullptr> inline 
-  auto peekIndex(const iScalar<vtype> &arg) ->  iScalar<vtype> 
-{
-  return arg;
-}
-// Vector peek, one index
-template<int Level,class vtype,int N,typename std::enable_if< iScalar<vtype>::TensorLevel == Level >::type * =nullptr> inline 
-  auto peekIndex(const iVector<vtype,N> &arg,int i) -> iScalar<vtype> // Index matches
-{
-  iScalar<vtype> ret;                              // return scalar
-  ret._internal = arg._internal[i];
-  return ret;
-}
-// Matrix peek, two indices
-template<int Level,class vtype,int N,typename std::enable_if< iScalar<vtype>::TensorLevel == Level >::type * =nullptr> inline 
-  auto peekIndex(const iMatrix<vtype,N> &arg,int i,int j) ->  iScalar<vtype>
-{
-  iScalar<vtype> ret;                              // return scalar
-  ret._internal = arg._internal[i][j];
-  return ret;
-}
-
-/////////////
-// No match peek for scalar,vector,matrix must forward on either 0,1,2 args. Must have 9 routines with notvalue
-/////////////
-// scalar
-template<int Level,class vtype,typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  auto peekIndex(const iScalar<vtype> &arg) -> iScalar<decltype(peekIndex<Level>(arg._internal))>
-{
-  iScalar<decltype(peekIndex<Level>(arg._internal))> ret;
-  ret._internal= peekIndex<Level>(arg._internal);
-  return ret;
-}
-template<int Level,class vtype, typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  auto peekIndex(const iScalar<vtype> &arg,int i) ->  iScalar<decltype(peekIndex<Level>(arg._internal,i))> 
-{
-  iScalar<decltype(peekIndex<Level>(arg._internal,i))> ret;
-  ret._internal=peekIndex<Level>(arg._internal,i);
-  return ret;
-}
-template<int Level,class vtype, typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  auto peekIndex(const iScalar<vtype> &arg,int i,int j) ->  iScalar<decltype(peekIndex<Level>(arg._internal,i,j))>
-{
-  iScalar<decltype(peekIndex<Level>(arg._internal,i,j))> ret;
-  ret._internal=peekIndex<Level>(arg._internal,i,j);
-  return ret;
-}
-// vector
-template<int Level,class vtype,int N, typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-auto peekIndex(const iVector<vtype,N> &arg) ->   iVector<decltype(peekIndex<Level>(arg._internal[0])),N>
-{
-  iVector<decltype(peekIndex<Level>(arg._internal[0])),N> ret;
-  for(int ii=0;ii<N;ii++){
-    ret._internal[ii]=peekIndex<Level>(arg._internal[ii]);
-  }
-  return ret;
-}
-template<int Level,class vtype,int N, typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  auto peekIndex(const iVector<vtype,N> &arg,int i) ->  iVector<decltype(peekIndex<Level>(arg._internal[0],i)),N>
-{
-  iVector<decltype(peekIndex<Level>(arg._internal[0],i)),N> ret;
-  for(int ii=0;ii<N;ii++){
-    ret._internal[ii]=peekIndex<Level>(arg._internal[ii],i);
-  }
-  return ret;
-}
-template<int Level,class vtype,int N, typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  auto peekIndex(const iVector<vtype,N> &arg,int i,int j) ->  iVector<decltype(peekIndex<Level>(arg._internal[0],i,j)),N> 
-{
-  iVector<decltype(peekIndex<Level>(arg._internal[0],i,j)),N> ret;
-  for(int ii=0;ii<N;ii++){
-    ret._internal[ii]=peekIndex<Level>(arg._internal[ii],i,j);
-  }
-  return ret;
-}
-
-// matrix
-template<int Level,class vtype,int N, typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-auto peekIndex(const iMatrix<vtype,N> &arg) ->   iMatrix<decltype(peekIndex<Level>(arg._internal[0][0])),N> 
-{
-  iMatrix<decltype(peekIndex<Level>(arg._internal[0][0])),N> ret;
-  for(int ii=0;ii<N;ii++){
-  for(int jj=0;jj<N;jj++){
-    ret._internal[ii][jj]=peekIndex<Level>(arg._internal[ii][jj]);// Could avoid this because peeking a scalar is dumb
-  }}
-  return ret;
-}
-template<int Level,class vtype,int N, typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  auto peekIndex(const iMatrix<vtype,N> &arg,int i) ->   iMatrix<decltype(peekIndex<Level>(arg._internal[0][0],i)),N>
-{
-  iMatrix<decltype(peekIndex<Level>(arg._internal[0][0],i)),N> ret;
-  for(int ii=0;ii<N;ii++){
-  for(int jj=0;jj<N;jj++){
-    ret._internal[ii][jj]=peekIndex<Level>(arg._internal[ii][jj],i);
-  }}
-  return ret;
-}
-template<int Level,class vtype,int N, typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  auto peekIndex(const iMatrix<vtype,N> &arg,int i,int j) ->   iMatrix<decltype(peekIndex<Level>(arg._internal[0][0],i,j)),N>
-{
-  iMatrix<decltype(peekIndex<Level>(arg._internal[0][0],i,j)),N> ret;
-  for(int ii=0;ii<N;ii++){
-  for(int jj=0;jj<N;jj++){
-    ret._internal[ii][jj]=peekIndex<Level>(arg._internal[ii][jj],i,j);
-  }}
-  return ret;
-}
-#endif
-
-
-}
-#endif

lib/Old/Tensor_poke.h

@@ -1,127 +0,0 @@
-    /*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/Old/Tensor_poke.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_MATH_POKE_H
-#define GRID_MATH_POKE_H
-namespace Grid {
-
-//////////////////////////////////////////////////////////////////////////////
-// Poke a specific index; 
-//////////////////////////////////////////////////////////////////////////////
-#if 0
-// Scalar poke
-template<int Level,class vtype,typename std::enable_if< iScalar<vtype>::TensorLevel == Level >::type * =nullptr> inline 
-  void pokeIndex(iScalar<vtype> &ret, const iScalar<vtype> &arg)
-{
-  ret._internal = arg._internal;
-}
-// Vector poke, one index
-template<int Level,class vtype,int N,typename std::enable_if< iScalar<vtype>::TensorLevel == Level >::type * =nullptr> inline 
-  void pokeIndex(iVector<vtype,N> &ret, const iScalar<vtype> &arg,int i)
-{
-  ret._internal[i] = arg._internal;
-}
-//Matrix poke, two indices
-template<int Level,class vtype,int N,typename std::enable_if< iScalar<vtype>::TensorLevel == Level >::type * =nullptr> inline 
-  void pokeIndex(iMatrix<vtype,N> &ret, const iScalar<vtype> &arg,int i,int j)
-{
-  ret._internal[i][j] = arg._internal;
-}
-
-/////////////
-// No match poke for scalar,vector,matrix must forward on either 0,1,2 args. Must have 9 routines with notvalue
-/////////////
-// scalar
-template<int Level,class vtype,typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-void pokeIndex(iScalar<vtype> &ret, const iScalar<decltype(peekIndex<Level>(ret._internal))>  &arg)
-{
-  pokeIndex<Level>(ret._internal,arg._internal);
-}
-template<int Level,class vtype,typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  void pokeIndex(iScalar<vtype> &ret, const iScalar<decltype(peekIndex<Level>(ret._internal,0))> &arg, int i)
-		 
-{
-  pokeIndex<Level>(ret._internal,arg._internal,i);
-}
-template<int Level,class vtype,typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  void pokeIndex(iScalar<vtype> &ret, const iScalar<decltype(peekIndex<Level>(ret._internal,0,0))> &arg,int i,int j)
-{
-  pokeIndex<Level>(ret._internal,arg._internal,i,j);
-}
-
-// Vector
-template<int Level,class vtype,int N,typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  void pokeIndex(iVector<vtype,N> &ret, iVector<decltype(peekIndex<Level>(ret._internal)),N>  &arg)
-{
-  for(int ii=0;ii<N;ii++){
-    pokeIndex<Level>(ret._internal[ii],arg._internal[ii]);
-  }
-}
-template<int Level,class vtype,int N,typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  void pokeIndex(iVector<vtype,N> &ret, const iVector<decltype(peekIndex<Level>(ret._internal,0)),N> &arg,int i)
-{
-  for(int ii=0;ii<N;ii++){
-    pokeIndex<Level>(ret._internal[ii],arg._internal[ii],i);
-  }
-}
-template<int Level,class vtype,int N,typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  void pokeIndex(iVector<vtype,N> &ret, const iVector<decltype(peekIndex<Level>(ret._internal,0,0)),N> &arg,int i,int j)
-{
-  for(int ii=0;ii<N;ii++){
-    pokeIndex<Level>(ret._internal[ii],arg._internal[ii],i,j);
-  }
-}
-
-// Matrix
-template<int Level,class vtype,int N,typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  void pokeIndex(iMatrix<vtype,N> &ret, const iMatrix<decltype(peekIndex<Level>(ret._internal)),N> &arg)		 
-{
-  for(int ii=0;ii<N;ii++){
-  for(int jj=0;jj<N;jj++){
-    pokeIndex<Level>(ret._internal[ii][jj],arg._internal[ii][jj]);
-  }}
-}
-template<int Level,class vtype,int N,typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  void pokeIndex(iMatrix<vtype,N> &ret, const iMatrix<decltype(peekIndex<Level>(ret._internal,0)),N> &arg,int i)
-{
-  for(int ii=0;ii<N;ii++){
-  for(int jj=0;jj<N;jj++){
-    pokeIndex<Level>(ret._internal[ii][jj],arg._internal[ii][jj],i);
-  }}
-}
-template<int Level,class vtype,int N,typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  void pokeIndex(iMatrix<vtype,N> &ret, const iMatrix<decltype(peekIndex<Level>(ret._internal,0,0)),N> &arg, int i,int j)
-{
-  for(int ii=0;ii<N;ii++){
-  for(int jj=0;jj<N;jj++){
-    pokeIndex<Level>(ret._internal[ii][jj],arg._internal[ii][jj],i,j);
-  }}
-}
-#endif
-
-}
-#endif

lib/Stencil.h

@@ -1,917 +0,0 @@
-   /*************************************************************************************
-
-     Grid physics library, www.github.com/paboyle/Grid 
-
-     Source file: ./lib/Stencil.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_STENCIL_H
- #define GRID_STENCIL_H
-
- #include <thread>
-
- #include <Grid/stencil/Lebesgue.h>   // subdir aggregate
-
- //////////////////////////////////////////////////////////////////////////////////////////
- // Must not lose sight that goal is to be able to construct really efficient
- // gather to a point stencil code. CSHIFT is not the best way, so need
- // additional stencil support.
- //
- // Stencil based code will pre-exchange haloes and use a table lookup for neighbours.
- // This will be done with generality to allow easier efficient implementations.
- // Overlap of comms and compute could be semi-automated by tabulating off-node connected,
- // and 
- //
- // Lattice <foo> could also allocate haloes which get used for stencil code.
- //
- // Grid could create a neighbour index table for a given stencil.
- //
- // Could also implement CovariantCshift, to fuse the loops and enhance performance.
- //
- //
- // General stencil computation:
- //
- // Generic services
- // 0) Prebuild neighbour tables
- // 1) Compute sizes of all haloes/comms buffers; allocate them.
- //
- // 2) Gather all faces, and communicate.
- // 3) Loop over result sites, giving nbr index/offnode info for each
- // 
- // Could take a 
- // SpinProjectFaces 
- // start comms
- // complete comms 
- // Reconstruct Umu
- //
- // Approach.
- //
- //////////////////////////////////////////////////////////////////////////////////////////
-
- namespace Grid {
-
-   struct StencilEntry { 
-     uint32_t _offset;
-     uint32_t _byte_offset;
-     uint16_t _is_local;
-     uint16_t _permute;
-     uint32_t _around_the_world; //256 bits, 32 bytes, 1/2 cacheline
-   };
-
-   template<class vobj,class cobj>
-   class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal fill in.
-   public:
-
-       typedef uint32_t StencilInteger;
-       typedef typename cobj::vector_type vector_type;
-       typedef typename cobj::scalar_type scalar_type;
-       typedef typename cobj::scalar_object scalar_object;
-
-       //////////////////////////////////////////
-       // Comms packet queue for asynch thread
-       //////////////////////////////////////////
-
-       struct Packet {
-	 void * send_buf;
-	 void * recv_buf;
-	 Integer to_rank;
-	 Integer from_rank;
-	 Integer bytes;
-	 volatile Integer done;
-       };
-
-       std::vector<Packet> Packets;
-       
- #define SEND_IMMEDIATE
- #define SERIAL_SENDS
-
-       void AddPacket(void *xmit,void * rcv, Integer to,Integer from,Integer bytes){
- #ifdef SEND_IMMEDIATE
-	 commtime-=usecond();
-	 _grid->SendToRecvFrom(xmit,to,rcv,from,bytes);
-	 commtime+=usecond();
- #endif
-	 Packet p;
-	 p.send_buf = xmit;
-	 p.recv_buf = rcv;
-	 p.to_rank  = to;
-	 p.from_rank= from;
-	 p.bytes    = bytes;
-	 p.done     = 0;
-	 comms_bytes+=2.0*bytes;
-	 Packets.push_back(p);
-
-       }
-
- #ifdef SERIAL_SENDS
-       void Communicate(void ) { 
-	 commtime-=usecond();
-	 for(int i=0;i<Packets.size();i++){
- #ifndef SEND_IMMEDIATE
-	   _grid->SendToRecvFrom(
-				 Packets[i].send_buf,
-				 Packets[i].to_rank,
-				 Packets[i].recv_buf,
-				 Packets[i].from_rank,
-				 Packets[i].bytes);
- #endif
-	   Packets[i].done = 1;
-	 }
-	 commtime+=usecond();
-       }
- #else
-       void Communicate(void ) { 
-	 typedef CartesianCommunicator::CommsRequest_t CommsRequest_t;
-	 std::vector<std::vector<CommsRequest_t> > reqs(Packets.size());
-	 commtime-=usecond();
-	 const int concurrency=2;
-	 for(int i=0;i<Packets.size();i+=concurrency){
-	   for(int ii=0;ii<concurrency;ii++){
-	     int j = i+ii;
-	     if ( j<Packets.size() ) {
- #ifndef SEND_IMMEDIATE
-	       _grid->SendToRecvFromBegin(reqs[j],
-					  Packets[j].send_buf,
-					  Packets[j].to_rank,
-					  Packets[j].recv_buf,
-					  Packets[j].from_rank,
-					  Packets[j].bytes);
- #endif
-	     }
-	   }
-	   for(int ii=0;ii<concurrency;ii++){
-	     int j = i+ii;
-	     if ( j<Packets.size() ) {
- #ifndef SEND_IMMEDIATE
-	       _grid->SendToRecvFromComplete(reqs[i]);
- #endif
-	     }
-	   }
-	   for(int ii=0;ii<concurrency;ii++){
-	     int j = i+ii;
-	     if ( j<Packets.size() ) {
-	       Packets[j].done = 1;
-	     }
-	   }
-	 }
-	 commtime+=usecond();
-       }
- #endif
-
-       ///////////////////////////////////////////
-       // Simd merge queue for asynch comms
-       ///////////////////////////////////////////
-       struct Merge {
-	 cobj * mpointer;
-	 std::vector<scalar_object *> rpointers;
-	 Integer buffer_size;
-	 Integer packet_id;
-       };
-
-       std::vector<Merge> Mergers;
-
-       void AddMerge(cobj *merge_p,std::vector<scalar_object *> &rpointers,Integer buffer_size,Integer packet_id) {
-	 Merge m;
-	 m.mpointer = merge_p;
-	 m.rpointers= rpointers;
-	 m.buffer_size = buffer_size;
-	 m.packet_id   = packet_id;
- #ifdef SEND_IMMEDIATE
-	 mergetime-=usecond();
- PARALLEL_FOR_LOOP
-	 for(int o=0;o<m.buffer_size;o++){
-	   merge1(m.mpointer[o],m.rpointers,o);
-	 }
-	 mergetime+=usecond();
- #else
-	 Mergers.push_back(m);
- #endif
-
-       }
-
-       void CommsMerge(void ) { 
-	 //PARALLEL_NESTED_LOOP2 
-	 for(int i=0;i<Mergers.size();i++){	
-
-	 spintime-=usecond();
-	 int packet_id = Mergers[i].packet_id;
-	 while(! Packets[packet_id].done ); // spin for completion
-	 spintime+=usecond();
-
- #ifndef SEND_IMMEDIATE
-	 mergetime-=usecond();
- PARALLEL_FOR_LOOP
-	   for(int o=0;o<Mergers[i].buffer_size;o++){
-	     merge1(Mergers[i].mpointer[o],Mergers[i].rpointers,o);
-	   }
-	 mergetime+=usecond();
- #endif
-
-	 }
-       }
-
-       ////////////////////////////////////////
-       // Basic Grid and stencil info
-       ////////////////////////////////////////
-
-       int                               _checkerboard;
-       int                               _npoints; // Move to template param?
-       GridBase *                        _grid;
-
-       // npoints of these
-       std::vector<int>                  _directions;
-       std::vector<int>                  _distances;
-       std::vector<int>                  _comm_buf_size;
-       std::vector<int>                  _permute_type;
-
-       // npoints x Osites() of these
-       // Flat vector, change layout for cache friendly.
-       Vector<StencilEntry>  _entries;
-
-       inline StencilEntry * GetEntry(int &ptype,int point,int osite) { ptype = _permute_type[point]; return & _entries[point+_npoints*osite]; }
-
-       void PrecomputeByteOffsets(void){
-	 for(int i=0;i<_entries.size();i++){
-	   if( _entries[i]._is_local ) {
-	     _entries[i]._byte_offset = _entries[i]._offset*sizeof(vobj);
-	   } else { 
-	     _entries[i]._byte_offset =(uint64_t)&comm_buf[0]+ _entries[i]._offset*sizeof(cobj);
-	   }
-	 }
-       };
-
-       inline uint64_t Touch(int ent) {
-	 //	 _mm_prefetch((char *)&_entries[ent],_MM_HINT_T0);
-       }
-       inline uint64_t GetInfo(int &ptype,int &local,int &perm,int point,int ent,uint64_t base) {
-	 //_mm_prefetch((char *)&_entries[ent+1],_MM_HINT_T0);
-	 uint64_t cbase = (uint64_t) 0;
-	 local = _entries[ent]._is_local;
-	 perm  = _entries[ent]._permute;
-	 if (perm)  ptype = _permute_type[point]; 
-	 if (local) return  base + _entries[ent]._byte_offset;
-	 else       return cbase + _entries[ent]._byte_offset;
-       }
-       inline uint64_t GetPFInfo(int ent,uint64_t base) {
-	 uint64_t cbase =  (uint64_t) 0;
-	 int local = _entries[ent]._is_local;
-	 if (local) return  base + _entries[ent]._byte_offset;
-	 else       return cbase + _entries[ent]._byte_offset;
-       }
-
-       // Comms buffers
-       std::vector<Vector<scalar_object> > u_simd_send_buf;
-       std::vector<Vector<scalar_object> > u_simd_recv_buf;
-       Vector<cobj>          u_send_buf;
-       Vector<cobj>          comm_buf;
-       int u_comm_offset;
-       int _unified_buffer_size;
-
-       /////////////////////////////////////////
-       // Timing info; ugly; possibly temporary
-       /////////////////////////////////////////
- #define TIMING_HACK
- #ifdef TIMING_HACK
-       double jointime;
-       double gathertime;
-       double commtime;
-       double halogtime;
-       double mergetime;
-       double spintime;
-       double comms_bytes;
-       double gathermtime;
-       double splicetime;
-       double nosplicetime;
-       double calls;
-
-       void ZeroCounters(void) {
-         gathertime = 0.;
-         jointime = 0.;
-         commtime = 0.;
-         halogtime = 0.;
-         mergetime = 0.;
-         spintime = 0.;
-         gathermtime = 0.;
-         splicetime = 0.;
-         nosplicetime = 0.;
-         comms_bytes = 0.;
-         calls = 0.;
-       };
-
-       void Report(void) {
-#define PRINTIT(A)	\
- std::cout << GridLogMessage << " Stencil " << #A << " "<< A/calls<<std::endl;
-	 if ( calls > 0. ) {
- std::cout << GridLogMessage << " Stencil calls "<<calls<<std::endl;
-       PRINTIT(jointime);
-       PRINTIT(gathertime);
-       PRINTIT(commtime);
-       PRINTIT(halogtime);
-       PRINTIT(mergetime);
-       PRINTIT(spintime);
-       PRINTIT(comms_bytes);
-       PRINTIT(gathermtime);
-       PRINTIT(splicetime);
-       PRINTIT(nosplicetime);
-	 }
-       };
- #endif
-
-   CartesianStencil(GridBase *grid,
-				      int npoints,
-				      int checkerboard,
-				      const std::vector<int> &directions,
-				      const std::vector<int> &distances) 
-     :   _permute_type(npoints), _comm_buf_size(npoints)
-     {
-       _npoints = npoints;
-       _grid    = grid;
-       _directions = directions;
-       _distances  = distances;
-       _unified_buffer_size=0;
-
-       int osites  = _grid->oSites();
-
-       _entries.resize(_npoints* osites);
-       for(int ii=0;ii<npoints;ii++){
-
-	 int i = ii; // reverse direction to get SIMD comms done first
-	 int point = i;
-
-
-	 int dimension    = directions[i];
-	 int displacement = distances[i];
-	 int shift = displacement;
-
-	 int fd = _grid->_fdimensions[dimension];
-	 int rd = _grid->_rdimensions[dimension];
-	 _permute_type[point]=_grid->PermuteType(dimension);
-
-	 _checkerboard = checkerboard;
-
-	 // the permute type
-	 int simd_layout     = _grid->_simd_layout[dimension];
-	 int comm_dim        = _grid->_processors[dimension] >1 ;
-	 int splice_dim      = _grid->_simd_layout[dimension]>1 && (comm_dim);
-	 int rotate_dim      = _grid->_simd_layout[dimension]>2;
-
-	 assert ( (rotate_dim && comm_dim) == false) ; // Do not think spread out is supported
-
-	 int sshift[2];
-
-	 // Underlying approach. For each local site build
-	 // up a table containing the npoint "neighbours" and whether they 
-	 // live in lattice or a comms buffer.
-	 if ( !comm_dim ) {
-	   sshift[0] = _grid->CheckerBoardShiftForCB(_checkerboard,dimension,shift,Even);
-	   sshift[1] = _grid->CheckerBoardShiftForCB(_checkerboard,dimension,shift,Odd);
-
-	   if ( sshift[0] == sshift[1] ) {
-	     Local(point,dimension,shift,0x3);
-	   } else {
-	     Local(point,dimension,shift,0x1);// if checkerboard is unfavourable take two passes
-	     Local(point,dimension,shift,0x2);// both with block stride loop iteration
-	   }
-	 } else { // All permute extract done in comms phase prior to Stencil application
-	   //        So tables are the same whether comm_dim or splice_dim
-	   sshift[0] = _grid->CheckerBoardShiftForCB(_checkerboard,dimension,shift,Even);
-	   sshift[1] = _grid->CheckerBoardShiftForCB(_checkerboard,dimension,shift,Odd);
-
-	   if ( sshift[0] == sshift[1] ) {
-	     Comms(point,dimension,shift,0x3);
-	   } else {
-	     Comms(point,dimension,shift,0x1);// if checkerboard is unfavourable take two passes
-	     Comms(point,dimension,shift,0x2);// both with block stride loop iteration
-	   }
-	 }
-       }
-       u_send_buf.resize(_unified_buffer_size);
-       comm_buf.resize(_unified_buffer_size);
-
-       PrecomputeByteOffsets();
-
-       const int Nsimd = grid->Nsimd();
-       u_simd_send_buf.resize(Nsimd);
-       u_simd_recv_buf.resize(Nsimd);
-       for(int l=0;l<Nsimd;l++){
-	 u_simd_send_buf[l].resize(_unified_buffer_size);
-	 u_simd_recv_buf[l].resize(_unified_buffer_size);
-       }
-     }
-
-     void Local     (int point, int dimension,int shiftpm,int cbmask)
-     {
-       int fd = _grid->_fdimensions[dimension];
-       int rd = _grid->_rdimensions[dimension];
-       int ld = _grid->_ldimensions[dimension];
-       int gd = _grid->_gdimensions[dimension];
-       int ly = _grid->_simd_layout[dimension];
-
-       // Map to always positive shift modulo global full dimension.
-       int shift = (shiftpm+fd)%fd;
-
-       // the permute type
-       int permute_dim =_grid->PermuteDim(dimension);
-
-       for(int x=0;x<rd;x++){       
-
-	 int o   = 0;
-	 int bo  = x * _grid->_ostride[dimension];
-
-	 int cb= (cbmask==0x2)? Odd : Even;
-
-	 int sshift = _grid->CheckerBoardShiftForCB(_checkerboard,dimension,shift,cb);
-	 int sx     = (x+sshift)%rd;
-
-	 int wraparound=0;
-	 if ( (shiftpm==-1) && (sx>x)  ) {
-	   wraparound = 1;
-	 }
-	 if ( (shiftpm== 1) && (sx<x)  ) {
-	   wraparound = 1;
-	 }
-
-	 int permute_slice=0;
-	 if(permute_dim){
-	   int wrap = sshift/rd;
-	   int  num = sshift%rd;
-	   if ( x< rd-num ) permute_slice=wrap;
-	   else permute_slice = (wrap+1)%ly;
-	 }
-
-	 CopyPlane(point,dimension,x,sx,cbmask,permute_slice,wraparound);
-
-       }
-     }
-
-     void Comms     (int point,int dimension,int shiftpm,int cbmask)
-     {
-       GridBase *grid=_grid;
-       const int Nsimd = grid->Nsimd();
-
-       int fd              = _grid->_fdimensions[dimension];
-       int ld              = _grid->_ldimensions[dimension];
-       int rd              = _grid->_rdimensions[dimension];
-       int pd              = _grid->_processors[dimension];
-       int simd_layout     = _grid->_simd_layout[dimension];
-       int comm_dim        = _grid->_processors[dimension] >1 ;
-
-       assert(comm_dim==1);
-       int shift = (shiftpm + fd) %fd;
-       assert(shift>=0);
-       assert(shift<fd);
-
-       int buffer_size = _grid->_slice_nblock[dimension]*_grid->_slice_block[dimension]; // done in reduced dims, so SIMD factored
-
-       _comm_buf_size[point] = buffer_size; // Size of _one_ plane. Multiple planes may be gathered and
-					    // send to one or more remote nodes.
-
-       int cb= (cbmask==0x2)? Odd : Even;
-       int sshift= _grid->CheckerBoardShiftForCB(_checkerboard,dimension,shift,cb);
-
-       for(int x=0;x<rd;x++){       
-
-	 int permute_type=grid->PermuteType(dimension);
-
-	 int sx        =  (x+sshift)%rd;
-
-	 int offnode = 0;
-	 if ( simd_layout > 1 ) {
-
-	   for(int i=0;i<Nsimd;i++){
-
-	     int inner_bit = (Nsimd>>(permute_type+1));
-	     int ic= (i&inner_bit)? 1:0;
-	     int my_coor          = rd*ic + x;
-	     int nbr_coor         = my_coor+sshift;
-	     int nbr_proc = ((nbr_coor)/ld) % pd;// relative shift in processors
-
-	     if ( nbr_proc ) { 
-	       offnode =1;
-	     }
-	   }
-
-	 } else { 
-	   int comm_proc = ((x+sshift)/rd)%pd;
-	   offnode = (comm_proc!= 0);
-	 }
-
-
-	 int wraparound=0;
-	 if ( (shiftpm==-1) && (sx>x) && (grid->_processor_coor[dimension]==0) ) {
-	   wraparound = 1;
-	 }
-	 if ( (shiftpm== 1) && (sx<x) && (grid->_processor_coor[dimension]==grid->_processors[dimension]-1) ) {
-	   wraparound = 1;
-	 }
-	 if (!offnode) {
-
-	   int permute_slice=0;
-	   CopyPlane(point,dimension,x,sx,cbmask,permute_slice,wraparound); 
-
-	 } else {
-
-	   int words = buffer_size;
-	   if (cbmask != 0x3) words=words>>1;
-
-	   int rank           = grid->_processor;
-	   int recv_from_rank;
-	   int xmit_to_rank;
-
-	   int unified_buffer_offset = _unified_buffer_size;
-	   _unified_buffer_size    += words;
-
-	   ScatterPlane(point,dimension,x,cbmask,unified_buffer_offset,wraparound); // permute/extract/merge is done in comms phase
-
-	 }
-       }
-     }
-   // Routine builds up integer table for each site in _offsets, _is_local, _permute
-   void CopyPlane(int point, int dimension,int lplane,int rplane,int cbmask,int permute,int wrap)
-     {
-       int rd = _grid->_rdimensions[dimension];
-
-       if ( !_grid->CheckerBoarded(dimension) ) {
-
-	 int o   = 0;                                     // relative offset to base within plane
-	 int ro  = rplane*_grid->_ostride[dimension]; // base offset for start of plane 
-	 int lo  = lplane*_grid->_ostride[dimension]; // offset in buffer
-
-	 // Simple block stride gather of SIMD objects
-	 for(int n=0;n<_grid->_slice_nblock[dimension];n++){
-	   for(int b=0;b<_grid->_slice_block[dimension];b++){
-	     int idx=point+(lo+o+b)*_npoints;
-	     _entries[idx]._offset  =ro+o+b;
-	     _entries[idx]._permute=permute;
-	     _entries[idx]._is_local=1;
-	     _entries[idx]._around_the_world=wrap;
-	   }
-	   o +=_grid->_slice_stride[dimension];
-	 }
-
-       } else {
-
-	 int ro  = rplane*_grid->_ostride[dimension]; // base offset for start of plane 
-	 int lo  = lplane*_grid->_ostride[dimension]; // base offset for start of plane 
-	 int o   = 0;                                     // relative offset to base within plane
-
-	 for(int n=0;n<_grid->_slice_nblock[dimension];n++){
-	   for(int b=0;b<_grid->_slice_block[dimension];b++){
-
-	     int ocb=1<<_grid->CheckerBoardFromOindex(o+b);
-
-	     if ( ocb&cbmask ) {
-	       int idx = point+(lo+o+b)*_npoints;
-	       _entries[idx]._offset =ro+o+b;
-	       _entries[idx]._is_local=1;
-	       _entries[idx]._permute=permute;
-	       _entries[idx]._around_the_world=wrap;
-	     }
-
-	   }
-	   o +=_grid->_slice_stride[dimension];
-	 }
-
-       }
-     }
-   // Routine builds up integer table for each site in _offsets, _is_local, _permute
-    void ScatterPlane (int point,int dimension,int plane,int cbmask,int offset, int wrap)
-     {
-       int rd = _grid->_rdimensions[dimension];
-
-       if ( !_grid->CheckerBoarded(dimension) ) {
-
-	 int so  = plane*_grid->_ostride[dimension]; // base offset for start of plane 
-	 int o   = 0;                                    // relative offset to base within plane
-	 int bo  = 0;                                    // offset in buffer
-
-	 // Simple block stride gather of SIMD objects
-	 for(int n=0;n<_grid->_slice_nblock[dimension];n++){
-	   for(int b=0;b<_grid->_slice_block[dimension];b++){
-	     int idx=point+(so+o+b)*_npoints;
-	     _entries[idx]._offset  =offset+(bo++);
-	     _entries[idx]._is_local=0;
-	     _entries[idx]._permute=0;
-	     _entries[idx]._around_the_world=wrap;
-	   }
-	   o +=_grid->_slice_stride[dimension];
-	 }
-
-       } else { 
-
-	 int so  = plane*_grid->_ostride[dimension]; // base offset for start of plane 
-	 int o   = 0;                                      // relative offset to base within plane
-	 int bo  = 0;                                      // offset in buffer
-
-	 for(int n=0;n<_grid->_slice_nblock[dimension];n++){
-	   for(int b=0;b<_grid->_slice_block[dimension];b++){
-
-	     int ocb=1<<_grid->CheckerBoardFromOindex(o+b);// Could easily be a table lookup
-	     if ( ocb & cbmask ) {
-	       int idx = point+(so+o+b)*_npoints;
-	       _entries[idx]._offset  =offset+(bo++);
-	       _entries[idx]._is_local=0;
-	       _entries[idx]._permute =0;
-	       _entries[idx]._around_the_world=wrap;
-	     }
-	   }
-	   o +=_grid->_slice_stride[dimension];
-	 }
-       }
-     }
-
-
-
-       template<class compressor>
-       void HaloExchange(const Lattice<vobj> &source,compressor &compress) 
-       {
-	 calls++;
-	 Mergers.resize(0);
-         Packets.resize(0);
-         HaloGather(source,compress);
-	 this->Communicate();
-	 CommsMerge(); // spins
-       }
-#if 0
-       // Overlapping comms and compute typically slows down compute and  is useless
-       // unless memory bandwidth greatly exceeds network
-       template<class compressor>
-       std::thread HaloExchangeBegin(const Lattice<vobj> &source,compressor &compress) {
-	 Mergers.resize(0); 
-	 Packets.resize(0);
-	 HaloGather(source,compress);
-	 return std::thread([&] { this->Communicate(); });
-       }
-       void HaloExchangeComplete(std::thread &thr) 
-       {
-	 CommsMerge(); // spins
-	 jointime-=usecond();
-	 thr.join();
-	 jointime+=usecond();
-       }
-#endif
-       template<class compressor>
-       void HaloGatherDir(const Lattice<vobj> &source,compressor &compress,int point)
-       {
-	   int dimension    = _directions[point];
-	   int displacement = _distances[point];
-
-	   int fd = _grid->_fdimensions[dimension];
-	   int rd = _grid->_rdimensions[dimension];
-
-
-	   // Map to always positive shift modulo global full dimension.
-	   int shift = (displacement+fd)%fd;
-
-	   //     	  int checkerboard = _grid->CheckerBoardDestination(source.checkerboard,shift);
-	   assert (source.checkerboard== _checkerboard);
-
-	   // the permute type
-	   int simd_layout     = _grid->_simd_layout[dimension];
-	   int comm_dim        = _grid->_processors[dimension] >1 ;
-	   int splice_dim      = _grid->_simd_layout[dimension]>1 && (comm_dim);
-
-	   // Gather phase
-	   int sshift [2];
-	   if ( comm_dim ) {
-	     sshift[0] = _grid->CheckerBoardShiftForCB(_checkerboard,dimension,shift,Even);
-	     sshift[1] = _grid->CheckerBoardShiftForCB(_checkerboard,dimension,shift,Odd);
-	     if ( sshift[0] == sshift[1] ) {
-	       if (splice_dim) {
-		 splicetime-=usecond();
-		 GatherSimd(source,dimension,shift,0x3,compress);
-		 splicetime+=usecond();
-	       } else { 
-		 nosplicetime-=usecond();
-		 Gather(source,dimension,shift,0x3,compress);
-		 nosplicetime+=usecond();
-	       }
-	     } else {
-	       if(splice_dim){
-		 splicetime-=usecond();
-		 GatherSimd(source,dimension,shift,0x1,compress);// if checkerboard is unfavourable take two passes
-		 GatherSimd(source,dimension,shift,0x2,compress);// both with block stride loop iteration
-		 splicetime+=usecond();
-	       } else {
-		 nosplicetime-=usecond();
-		 Gather(source,dimension,shift,0x1,compress);
-		 Gather(source,dimension,shift,0x2,compress);
-		 nosplicetime+=usecond();
-	       }
-	     }
-	   }
-       }
-
-       template<class compressor>
-       void HaloGather(const Lattice<vobj> &source,compressor &compress)
-       {
-	 // conformable(source._grid,_grid);
-	 assert(source._grid==_grid);
-	 halogtime-=usecond();
-
-	 assert (comm_buf.size() == _unified_buffer_size );
-	 u_comm_offset=0;
-
-	 // Gather all comms buffers
-	 for(int point = 0 ; point < _npoints; point++) {
-	   compress.Point(point);
-	   HaloGatherDir(source,compress,point);
-	 }
-
-	 assert(u_comm_offset==_unified_buffer_size);
-	 halogtime+=usecond();
-       }
-
-       template<class compressor>
-	 void Gather(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor & compress)
-	 {
-	   typedef typename cobj::vector_type vector_type;
-	   typedef typename cobj::scalar_type scalar_type;
-
-	   GridBase *grid=_grid;
-	   assert(rhs._grid==_grid);
-	   //	  conformable(_grid,rhs._grid);
-
-	   int fd              = _grid->_fdimensions[dimension];
-	   int rd              = _grid->_rdimensions[dimension];
-	   int pd              = _grid->_processors[dimension];
-	   int simd_layout     = _grid->_simd_layout[dimension];
-	   int comm_dim        = _grid->_processors[dimension] >1 ;
-	   assert(simd_layout==1);
-	   assert(comm_dim==1);
-	   assert(shift>=0);
-	   assert(shift<fd);
-
-	   int buffer_size = _grid->_slice_nblock[dimension]*_grid->_slice_block[dimension];
-
-	   int cb= (cbmask==0x2)? Odd : Even;
-	   int sshift= _grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,cb);
-
-	   for(int x=0;x<rd;x++){       
-
-	     int sx        = (x+sshift)%rd;
-	     int comm_proc = ((x+sshift)/rd)%pd;
-
-	     if (comm_proc) {
-
-	       int words = buffer_size;
-	       if (cbmask != 0x3) words=words>>1;
-
-	       int bytes = words * sizeof(cobj);
-
-	       gathertime-=usecond();
-	       Gather_plane_simple (rhs,u_send_buf,dimension,sx,cbmask,compress,u_comm_offset);
-	       gathertime+=usecond();
-
-	       int rank           = _grid->_processor;
-	       int recv_from_rank;
-	       int xmit_to_rank;
-	       _grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
-	       assert (xmit_to_rank   != _grid->ThisRank());
-	       assert (recv_from_rank != _grid->ThisRank());
-
-	       //      FIXME Implement asynchronous send & also avoid buffer copy
-	       AddPacket((void *)&u_send_buf[u_comm_offset],
-			 (void *)  &comm_buf[u_comm_offset],
-			 xmit_to_rank,
-			 recv_from_rank,
-			 bytes);
-
-	       u_comm_offset+=words;
-	     }
-	   }
-	 }
-
-
-       template<class compressor>
-	 void  GatherSimd(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor &compress)
-	 {
-	   const int Nsimd = _grid->Nsimd();
-
-	   int fd = _grid->_fdimensions[dimension];
-	   int rd = _grid->_rdimensions[dimension];
-	   int ld = _grid->_ldimensions[dimension];
-	   int pd              = _grid->_processors[dimension];
-	   int simd_layout     = _grid->_simd_layout[dimension];
-	   int comm_dim        = _grid->_processors[dimension] >1 ;
-
-	   assert(comm_dim==1);
-	   // This will not work with a rotate dim
-	   assert(simd_layout==2);
-	   assert(shift>=0);
-	   assert(shift<fd);
-
-	   int permute_type=_grid->PermuteType(dimension);
-
-	   ///////////////////////////////////////////////
-	   // Simd direction uses an extract/merge pair
-	   ///////////////////////////////////////////////
-	   int buffer_size = _grid->_slice_nblock[dimension]*_grid->_slice_block[dimension];
-	   int words = sizeof(cobj)/sizeof(vector_type);
-
-	   assert(cbmask==0x3); // Fixme think there is a latent bug if not true
-
-	   int bytes = buffer_size*sizeof(scalar_object);
-
-	   std::vector<scalar_object *> rpointers(Nsimd);
-	   std::vector<scalar_object *> spointers(Nsimd);
-
-	   ///////////////////////////////////////////
-	   // Work out what to send where
-	   ///////////////////////////////////////////
-
-	   int cb    = (cbmask==0x2)? Odd : Even;
-	   int sshift= _grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,cb);
-
-	   // loop over outer coord planes orthog to dim
-	   for(int x=0;x<rd;x++){       
-
-	     int any_offnode = ( ((x+sshift)%fd) >= rd );
-
-	     if ( any_offnode ) {
-
-	       for(int i=0;i<Nsimd;i++){       
-		 spointers[i] = &u_simd_send_buf[i][u_comm_offset];
-	       }
-
-	       int sx   = (x+sshift)%rd;
-
-	       gathermtime-=usecond();
-	       Gather_plane_extract<cobj>(rhs,spointers,dimension,sx,cbmask,compress);
-	       gathermtime+=usecond();
-
-	       for(int i=0;i<Nsimd;i++){
-
-		 // FIXME 
-		 // This logic is hard coded to simd_layout ==2 and not allowing >2
-		 //		std::cout << "GatherSimd : lane 1st elem " << i << u_simd_send_buf[i ][u_comm_offset]<<std::endl;
-
-		 int inner_bit = (Nsimd>>(permute_type+1));
-		 int ic= (i&inner_bit)? 1:0;
-
-		 int my_coor          = rd*ic + x;
-		 int nbr_coor         = my_coor+sshift;
-		 int nbr_proc = ((nbr_coor)/ld) % pd;// relative shift in processors
-		 int nbr_lcoor= (nbr_coor%ld);
-		 int nbr_ic   = (nbr_lcoor)/rd;    // inner coord of peer
-		 int nbr_ox   = (nbr_lcoor%rd);    // outer coord of peer
-		 int nbr_lane = (i&(~inner_bit));
-
-		 if (nbr_ic) nbr_lane|=inner_bit;
-		 assert (sx == nbr_ox);
-
-		 auto rp = &u_simd_recv_buf[i       ][u_comm_offset];
-		 auto sp = &u_simd_send_buf[nbr_lane][u_comm_offset];
-
-		 void *vrp = (void *)rp;
-		 void *vsp = (void *)sp;
-
-
-		 if(nbr_proc){
-
-		   int recv_from_rank;
-		   int xmit_to_rank;
-
-		   _grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank); 
-
-		   AddPacket( vsp,vrp,xmit_to_rank,recv_from_rank,bytes);
-
-		   rpointers[i] = rp;
-
-		 } else { 
-
-		   rpointers[i] = sp;
-
-		 }
-	       }
-
-	       AddMerge(&comm_buf[u_comm_offset],rpointers,buffer_size,Packets.size()-1);
-
-	       u_comm_offset     +=buffer_size;
-	     }
-	   }
-	 }
-
-   };
- }
-#endif

lib/algorithms/Algorithms.h

@@ -42,15 +42,14 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/algorithms/iterative/ConjugateResidual.h>
 #include <Grid/algorithms/iterative/NormalEquations.h>
 #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/ImplicitlyRestartedLanczos.h>
-
 #include <Grid/algorithms/CoarsenedMatrix.h>
+#include <Grid/algorithms/FFT.h>
 
 // Eigen/lanczos
 // EigCg

lib/algorithms/CoarsenedMatrix.h

@@ -267,8 +267,7 @@ namespace Grid {
       SimpleCompressor<siteVector> compressor;
       Stencil.HaloExchange(in,compressor);
 
-PARALLEL_FOR_LOOP
-      for(int ss=0;ss<Grid()->oSites();ss++){
+      parallel_for(int ss=0;ss<Grid()->oSites();ss++){
         siteVector res = zero;
 	siteVector nbr;
 	int ptype;
@@ -282,7 +281,7 @@ PARALLEL_FOR_LOOP
 	  } else if(SE->_is_local) { 
 	    nbr = in._odata[SE->_offset];
 	  } else {
-	    nbr = Stencil.comm_buf[SE->_offset];
+	    nbr = Stencil.CommBuf()[SE->_offset];
 	  }
 	  res = res + A[point]._odata[ss]*nbr;
 	}
@@ -380,8 +379,7 @@ PARALLEL_FOR_LOOP
 	  Subspace.ProjectToSubspace(oProj,oblock);
 	  //	  blockProject(iProj,iblock,Subspace.subspace);
 	  //	  blockProject(oProj,oblock,Subspace.subspace);
-PARALLEL_FOR_LOOP
-	  for(int ss=0;ss<Grid()->oSites();ss++){
+	  parallel_for(int ss=0;ss<Grid()->oSites();ss++){
 	    for(int j=0;j<nbasis;j++){
 	      if( disp!= 0 ) {
 		A[p]._odata[ss](j,i) = oProj._odata[ss](j);

lib/algorithms/FFT.h

@@ -29,9 +29,15 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef _GRID_FFT_H_
 #define _GRID_FFT_H_
 
-#ifdef HAVE_FFTW	
+#ifdef HAVE_FFTW
+#ifdef USE_MKL
+#include <fftw/fftw3.h>
+#else
 #include <fftw3.h>
 #endif
+#endif
+
+
 namespace Grid {
 
   template<class scalar> struct FFTW { };
@@ -98,179 +104,202 @@ namespace Grid {
 #define FFTW_BACKWARD (+1)
 #endif
 
-  class FFT { 
+  class FFT {
   private:
-
+    
     GridCartesian *vgrid;
     GridCartesian *sgrid;
-
+    
     int Nd;
     double flops;
     double flops_call;
     uint64_t usec;
-
+    
     std::vector<int> dimensions;
     std::vector<int> processors;
     std::vector<int> processor_coor;
-
+    
   public:
-
+    
     static const int forward=FFTW_FORWARD;
     static const int backward=FFTW_BACKWARD;
-
+    
     double Flops(void) {return flops;}
     double MFlops(void) {return flops/usec;}
+    double USec(void)   {return (double)usec;}    
 
-    FFT ( GridCartesian * grid ) : 
-      vgrid(grid),
-      Nd(grid->_ndimension),
-      dimensions(grid->_fdimensions),
-      processors(grid->_processors),
-      processor_coor(grid->_processor_coor)
+    FFT ( GridCartesian * grid ) :
+    vgrid(grid),
+    Nd(grid->_ndimension),
+    dimensions(grid->_fdimensions),
+    processors(grid->_processors),
+    processor_coor(grid->_processor_coor)
     {
       flops=0;
       usec =0;
       std::vector<int> layout(Nd,1);
       sgrid = new GridCartesian(dimensions,layout,processors);
     };
-
-    ~FFT ( void)  { 
-      delete sgrid; 
+    
+    ~FFT ( void)  {
+      delete sgrid;
     }
     
     template<class vobj>
-    void FFT_dim(Lattice<vobj> &result,const Lattice<vobj> &source,int dim, int inverse){
+    void FFT_dim_mask(Lattice<vobj> &result,const Lattice<vobj> &source,std::vector<int> mask,int sign){
 
       conformable(result._grid,vgrid);
       conformable(source._grid,vgrid);
+      Lattice<vobj> tmp(vgrid);
+      tmp = source;
+      for(int d=0;d<Nd;d++){
+	if( mask[d] ) {
+	  FFT_dim(result,tmp,d,sign);
+	  tmp=result;
+	}
+      }
+    }
+
+    template<class vobj>
+    void FFT_all_dim(Lattice<vobj> &result,const Lattice<vobj> &source,int sign){
+      std::vector<int> mask(Nd,1);
+      FFT_dim_mask(result,source,mask,sign);
+    }
+
+
+    template<class vobj>
+    void FFT_dim(Lattice<vobj> &result,const Lattice<vobj> &source,int dim, int sign){
+#ifndef HAVE_FFTW
+      assert(0);
+#else
+      conformable(result._grid,vgrid);
+      conformable(source._grid,vgrid);
 
       int L = vgrid->_ldimensions[dim];
       int G = vgrid->_fdimensions[dim];
-
+      
       std::vector<int> layout(Nd,1);
       std::vector<int> pencil_gd(vgrid->_fdimensions);
-
-      pencil_gd[dim] = G*processors[dim];    
-
+      
+      pencil_gd[dim] = G*processors[dim];
+      
       // Pencil global vol LxLxGxLxL per node
       GridCartesian pencil_g(pencil_gd,layout,processors);
-
+      
       // Construct pencils
       typedef typename vobj::scalar_object sobj;
       typedef typename sobj::scalar_type   scalar;
+      
+      Lattice<sobj> pgbuf(&pencil_g);
+      
 
-      Lattice<vobj> ssource(vgrid); ssource =source;
-      Lattice<sobj> pgsource(&pencil_g);
-      Lattice<sobj> pgresult(&pencil_g); pgresult=zero;
-
-#ifndef HAVE_FFTW	
-      assert(0);
-#else 
       typedef typename FFTW<scalar>::FFTW_scalar FFTW_scalar;
       typedef typename FFTW<scalar>::FFTW_plan   FFTW_plan;
-
-      {
-	int Ncomp = sizeof(sobj)/sizeof(scalar);
-	int Nlow  = 1;
-	for(int d=0;d<dim;d++){
-	  Nlow*=vgrid->_ldimensions[d];
-	}
-
-	int rank = 1;  /* 1d transforms */
-	int n[] = {G}; /* 1d transforms of length G */
-	int howmany = Ncomp;
-	int odist,idist,istride,ostride;
-	idist   = odist   = 1;          /* Distance between consecutive FT's */
-	istride = ostride = Ncomp*Nlow; /* distance between two elements in the same FT */
-	int *inembed = n, *onembed = n;
-
-	
-	int sign = FFTW_FORWARD;
-	if (inverse) sign = FFTW_BACKWARD;
-
-	FFTW_plan p;
-	{
-	  FFTW_scalar *in = (FFTW_scalar *)&pgsource._odata[0];
-	  FFTW_scalar *out= (FFTW_scalar *)&pgresult._odata[0];
-	  p = FFTW<scalar>::fftw_plan_many_dft(rank,n,howmany,
-					       in,inembed,
-					       istride,idist,
-					       out,onembed,
-					       ostride, odist,
-					       sign,FFTW_ESTIMATE);
-	}
-
-	double add,mul,fma;
-	FFTW<scalar>::fftw_flops(p,&add,&mul,&fma);
-	flops_call = add+mul+2.0*fma;
-
-	GridStopWatch timer;
-
-	// Barrel shift and collect global pencil
-	for(int p=0;p<processors[dim];p++) { 
-
-	  for(int idx=0;idx<sgrid->lSites();idx++) { 
-
-	    std::vector<int> lcoor(Nd);
-    	    sgrid->LocalIndexToLocalCoor(idx,lcoor);
-
-	    sobj s;
-
-	    peekLocalSite(s,ssource,lcoor);
-
-	    lcoor[dim]+=p*L;
-	   
-	    pokeLocalSite(s,pgsource,lcoor);
-	  }
-
-	  ssource = Cshift(ssource,dim,L);
-	}
-	
-	// Loop over orthog coords
-	int NN=pencil_g.lSites();
-
-	GridStopWatch Timer;
-	Timer.Start();
-
-PARALLEL_FOR_LOOP
-	for(int idx=0;idx<NN;idx++) { 
-
-	  std::vector<int> lcoor(Nd);
-	  pencil_g.LocalIndexToLocalCoor(idx,lcoor);
-
-	  if ( lcoor[dim] == 0 ) {  // restricts loop to plane at lcoor[dim]==0
-	    FFTW_scalar *in = (FFTW_scalar *)&pgsource._odata[idx];
-	    FFTW_scalar *out= (FFTW_scalar *)&pgresult._odata[idx];
-	    FFTW<scalar>::fftw_execute_dft(p,in,out);
-	  }
-	}
-
-        Timer.Stop();
-	usec += Timer.useconds();
-	flops+= flops_call*NN;
-
-        int pc = processor_coor[dim];
-        for(int idx=0;idx<sgrid->lSites();idx++) { 
-	  std::vector<int> lcoor(Nd);
-	  sgrid->LocalIndexToLocalCoor(idx,lcoor);
-	  std::vector<int> gcoor = lcoor;
-	  // extract the result
-	  sobj s;
-	  gcoor[dim] = lcoor[dim]+L*pc;
-	  peekLocalSite(s,pgresult,gcoor);
-	  pokeLocalSite(s,result,lcoor);
-	}
-      	  
-	FFTW<scalar>::fftw_destroy_plan(p);
+      
+      int Ncomp = sizeof(sobj)/sizeof(scalar);
+      int Nlow  = 1;
+      for(int d=0;d<dim;d++){
+        Nlow*=vgrid->_ldimensions[d];
       }
+      
+      int rank = 1;  /* 1d transforms */
+      int n[] = {G}; /* 1d transforms of length G */
+      int howmany = Ncomp;
+      int odist,idist,istride,ostride;
+      idist   = odist   = 1;          /* Distance between consecutive FT's */
+      istride = ostride = Ncomp*Nlow; /* distance between two elements in the same FT */
+      int *inembed = n, *onembed = n;
+      
+      scalar div;
+	  if ( sign == backward ) div = 1.0/G;
+	  else if ( sign == forward ) div = 1.0;
+	  else assert(0);
+      
+      FFTW_plan p;
+      {
+        FFTW_scalar *in = (FFTW_scalar *)&pgbuf._odata[0];
+        FFTW_scalar *out= (FFTW_scalar *)&pgbuf._odata[0];
+        p = FFTW<scalar>::fftw_plan_many_dft(rank,n,howmany,
+                                             in,inembed,
+                                             istride,idist,
+                                             out,onembed,
+                                             ostride, odist,
+                                             sign,FFTW_ESTIMATE);
+      }
+      
+      // Barrel shift and collect global pencil
+      std::vector<int> lcoor(Nd), gcoor(Nd);
+      result = source;
+      for(int p=0;p<processors[dim];p++) {
+        PARALLEL_REGION
+        {
+          std::vector<int> cbuf(Nd);
+          sobj s;
+          
+          PARALLEL_FOR_LOOP_INTERN
+          for(int idx=0;idx<sgrid->lSites();idx++) {
+            sgrid->LocalIndexToLocalCoor(idx,cbuf);
+            peekLocalSite(s,result,cbuf);
+            cbuf[dim]+=p*L;
+            pokeLocalSite(s,pgbuf,cbuf);
+          }
+        }
+        if (p != processors[dim] - 1)
+        {
+          result = Cshift(result,dim,L);
+        }
+      }
+      
+      // Loop over orthog coords
+      int NN=pencil_g.lSites();
+      GridStopWatch timer;
+      timer.Start();
+      PARALLEL_REGION
+      {
+        std::vector<int> cbuf(Nd);
+        
+        PARALLEL_FOR_LOOP_INTERN
+        for(int idx=0;idx<NN;idx++) {
+          pencil_g.LocalIndexToLocalCoor(idx, cbuf);
+          if ( cbuf[dim] == 0 ) {  // restricts loop to plane at lcoor[dim]==0
+            FFTW_scalar *in = (FFTW_scalar *)&pgbuf._odata[idx];
+            FFTW_scalar *out= (FFTW_scalar *)&pgbuf._odata[idx];
+            FFTW<scalar>::fftw_execute_dft(p,in,out);
+          }
+        }
+      }
+      timer.Stop();
+      
+      // performance counting
+      double add,mul,fma;
+      FFTW<scalar>::fftw_flops(p,&add,&mul,&fma);
+      flops_call = add+mul+2.0*fma;
+      usec += timer.useconds();
+      flops+= flops_call*NN;
+      
+      // writing out result
+      int pc = processor_coor[dim];
+      PARALLEL_REGION
+      {
+        std::vector<int> clbuf(Nd), cgbuf(Nd);
+        sobj s;
+        
+        PARALLEL_FOR_LOOP_INTERN
+        for(int idx=0;idx<sgrid->lSites();idx++) {
+          sgrid->LocalIndexToLocalCoor(idx,clbuf);
+          cgbuf = clbuf;
+          cgbuf[dim] = clbuf[dim]+L*pc;
+          peekLocalSite(s,pgbuf,cgbuf);
+          pokeLocalSite(s,result,clbuf);
+        }
+      }
+      result = result*div;
+      
+      // destroying plan
+      FFTW<scalar>::fftw_destroy_plan(p);
 #endif
-
-
     }
-
   };
-
-
 }
 
 #endif

lib/algorithms/approx/MultiShiftFunction.cc

@@ -25,7 +25,7 @@ Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#include <Grid.h>
+#include <Grid/GridCore.h>
 
 namespace Grid {
 double MultiShiftFunction::approx(double x)

lib/algorithms/approx/Remez.cc

@@ -20,7 +20,7 @@
 #include<iomanip>
 #include<cassert>
 
-#include<algorithms/approx/Remez.h>
+#include<Grid/algorithms/approx/Remez.h>
 
 // Constructor
 AlgRemez::AlgRemez(double lower, double upper, long precision) 

lib/algorithms/iterative/ConjugateGradient.h

@@ -45,6 +45,8 @@ class ConjugateGradient : public OperatorFunction<Field> {
                            // Defaults true.
   RealD Tolerance;
   Integer MaxIterations;
+  Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
+  
   ConjugateGradient(RealD tol, Integer maxit, bool err_on_no_conv = true)
       : Tolerance(tol),
         MaxIterations(maxit),
@@ -154,14 +156,15 @@ class ConjugateGradient : public OperatorFunction<Field> {
                   << LinalgTimer.Elapsed();
         std::cout << std::endl;
 
-        if (ErrorOnNoConverge) assert(true_residual / Tolerance < 1000.0);
-
+        if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
+	IterationsToComplete = k;	
         return;
       }
     }
     std::cout << GridLogMessage << "ConjugateGradient did NOT converge"
               << std::endl;
     if (ErrorOnNoConverge) assert(0);
+    IterationsToComplete = k;
   }
 };
 }

lib/algorithms/iterative/ConjugateGradientMixedPrec.h

@@ -35,6 +35,7 @@ namespace Grid {
   class MixedPrecisionConjugateGradient : public LinearFunction<FieldD> {
   public:                                                
     RealD   Tolerance;
+    RealD   InnerTolerance; //Initial tolerance for inner CG. Defaults to Tolerance but can be changed
     Integer MaxInnerIterations;
     Integer MaxOuterIterations;
     GridBase* SinglePrecGrid; //Grid for single-precision fields
@@ -42,12 +43,16 @@ namespace Grid {
     LinearOperatorBase<FieldF> &Linop_f;
     LinearOperatorBase<FieldD> &Linop_d;
 
+    Integer TotalInnerIterations; //Number of inner CG iterations
+    Integer TotalOuterIterations; //Number of restarts
+    Integer TotalFinalStepIterations; //Number of CG iterations in final patch-up step
+
     //Option to speed up *inner single precision* solves using a LinearFunction that produces a guess
     LinearFunction<FieldF> *guesser;
     
     MixedPrecisionConjugateGradient(RealD tol, Integer maxinnerit, Integer maxouterit, GridBase* _sp_grid, LinearOperatorBase<FieldF> &_Linop_f, LinearOperatorBase<FieldD> &_Linop_d) :
       Linop_f(_Linop_f), Linop_d(_Linop_d),
-      Tolerance(tol), MaxInnerIterations(maxinnerit), MaxOuterIterations(maxouterit), SinglePrecGrid(_sp_grid),
+      Tolerance(tol), InnerTolerance(tol), MaxInnerIterations(maxinnerit), MaxOuterIterations(maxouterit), SinglePrecGrid(_sp_grid),
       OuterLoopNormMult(100.), guesser(NULL){ };
 
     void useGuesser(LinearFunction<FieldF> &g){
@@ -55,6 +60,8 @@ namespace Grid {
     }
   
     void operator() (const FieldD &src_d_in, FieldD &sol_d){
+      TotalInnerIterations = 0;
+	
       GridStopWatch TotalTimer;
       TotalTimer.Start();
     
@@ -74,7 +81,7 @@ namespace Grid {
       FieldD src_d(DoublePrecGrid);
       src_d = src_d_in; //source for next inner iteration, computed from residual during operation
     
-      RealD inner_tol = Tolerance;
+      RealD inner_tol = InnerTolerance;
     
       FieldF src_f(SinglePrecGrid);
       src_f.checkerboard = cb;
@@ -89,7 +96,9 @@ namespace Grid {
 
       GridStopWatch PrecChangeTimer;
     
-      for(Integer outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++){
+      Integer &outer_iter = TotalOuterIterations; //so it will be equal to the final iteration count
+      
+      for(outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++){
 	//Compute double precision rsd and also new RHS vector.
 	Linop_d.HermOp(sol_d, tmp_d);
 	RealD norm = axpy_norm(src_d, -1., tmp_d, src_d_in); //src_d is residual vector
@@ -117,6 +126,7 @@ namespace Grid {
 	InnerCGtimer.Start();
 	CG_f(Linop_f, src_f, sol_f);
 	InnerCGtimer.Stop();
+	TotalInnerIterations += CG_f.IterationsToComplete;
       
 	//Convert sol back to double and add to double prec solution
 	PrecChangeTimer.Start();
@@ -131,9 +141,11 @@ namespace Grid {
     
       ConjugateGradient<FieldD> CG_d(Tolerance, MaxInnerIterations);
       CG_d(Linop_d, src_d_in, sol_d);
+      TotalFinalStepIterations = CG_d.IterationsToComplete;
 
       TotalTimer.Stop();
-      std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Total " << TotalTimer.Elapsed() << " Precision change " << PrecChangeTimer.Elapsed() << " Inner CG total " << InnerCGtimer.Elapsed() << std::endl;
+      std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Inner CG iterations " << TotalInnerIterations << " Restarts " << TotalOuterIterations << " Final CG iterations " << TotalFinalStepIterations << std::endl;
+      std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Total time " << TotalTimer.Elapsed() << " Precision change " << PrecChangeTimer.Elapsed() << " Inner CG total " << InnerCGtimer.Elapsed() << std::endl;
     }
   };
 

lib/algorithms/iterative/ImplicitlyRestartedLanczos.h

@@ -31,7 +31,11 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 
 #include <string.h> //memset
 #ifdef USE_LAPACK
-#include <lapacke.h>
+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);
 #endif
 #include "DenseMatrix.h"
 #include "EigenSort.h"
@@ -1076,10 +1080,10 @@ say con = 2
 **/
 
 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
+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)

lib/algorithms/iterative/Matrix.h

@@ -36,7 +36,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <iomanip>
 #include <complex>
 #include <typeinfo>
-#include <Grid.h>
+#include <Grid/Grid.h>
 
 
 /** Sign function **/

lib/allocator/AlignedAllocator.cc

@@ -0,0 +1,66 @@
+
+
+
+#include <Grid/GridCore.h>
+
+namespace Grid {
+
+int PointerCache::victim;
+
+  PointerCache::PointerCacheEntry PointerCache::Entries[PointerCache::Ncache];
+
+void *PointerCache::Insert(void *ptr,size_t bytes) {
+
+  if (bytes < 4096 ) return NULL;
+
+#ifdef GRID_OMP
+  assert(omp_in_parallel()==0);
+#endif 
+
+  void * ret = NULL;
+  int v = -1;
+
+  for(int e=0;e<Ncache;e++) {
+    if ( Entries[e].valid==0 ) {
+      v=e; 
+      break;
+    }
+  }
+
+  if ( v==-1 ) {
+    v=victim;
+    victim = (victim+1)%Ncache;
+  }
+
+  if ( Entries[v].valid ) {
+    ret = Entries[v].address;
+    Entries[v].valid = 0;
+    Entries[v].address = NULL;
+    Entries[v].bytes = 0;
+  }
+
+  Entries[v].address=ptr;
+  Entries[v].bytes  =bytes;
+  Entries[v].valid  =1;
+
+  return ret;
+}
+
+void *PointerCache::Lookup(size_t bytes) {
+
+ if (bytes < 4096 ) return NULL;
+
+#ifdef _OPENMP
+  assert(omp_in_parallel()==0);
+#endif 
+
+  for(int e=0;e<Ncache;e++){
+    if ( Entries[e].valid && ( Entries[e].bytes == bytes ) ) {
+      Entries[e].valid = 0;
+      return Entries[e].address;
+    }
+  }
+  return NULL;
+}
+
+}

lib/allocator/AlignedAllocator.h

@@ -1,4 +1,4 @@
-    /*************************************************************************************
+/*************************************************************************************
 
     Grid physics library, www.github.com/paboyle/Grid 
 
@@ -40,19 +40,34 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <mm_malloc.h>
 #endif
 
-#ifdef GRID_COMMS_SHMEM
-extern "C" { 
-#include <mpp/shmem.h>
-extern void * shmem_align(size_t, size_t);
-extern void  shmem_free(void *);
-}
-#endif
-
 namespace Grid {
 
+  class PointerCache {
+  private:
+
+    static const int Ncache=8;
+    static int victim;
+
+    typedef struct { 
+      void *address;
+      size_t bytes;
+      int valid;
+    } PointerCacheEntry;
+    
+    static PointerCacheEntry Entries[Ncache];
+
+  public:
+
+
+    static void *Insert(void *ptr,size_t bytes) ;
+    static void *Lookup(size_t bytes) ;
+
+  };
+
 ////////////////////////////////////////////////////////////////////
 // A lattice of something, but assume the something is SIMDized.
 ////////////////////////////////////////////////////////////////////
+
 template<typename _Tp>
 class alignedAllocator {
 public: 
@@ -65,28 +80,85 @@ public:
   typedef _Tp        value_type;
 
   template<typename _Tp1>  struct rebind { typedef alignedAllocator<_Tp1> other; };
-
   alignedAllocator() throw() { }
-
   alignedAllocator(const alignedAllocator&) throw() { }
-
   template<typename _Tp1> alignedAllocator(const alignedAllocator<_Tp1>&) throw() { }
-
   ~alignedAllocator() throw() { }
-
   pointer       address(reference __x)       const { return &__x; }
-  //  const_pointer address(const_reference __x) const { return &__x; }
-
   size_type  max_size() const throw() { return size_t(-1) / sizeof(_Tp); }
 
   pointer allocate(size_type __n, const void* _p= 0)
   { 
+    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
+
+    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
+    if ( __freeme ) _mm_free((void *)__freeme); 
+#else
+    if ( __freeme ) free((void *)__freeme);
+#endif
+  }
+  void construct(pointer __p, const _Tp& __val) { };
+  void construct(pointer __p) { };
+  void destroy(pointer __p) { };
+};
+template<typename _Tp>  inline bool operator==(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return true; }
+template<typename _Tp>  inline bool operator!=(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return false; }
+
+//////////////////////////////////////////////////////////////////////////////////////////
+// MPI3 : comms must use shm region
+// SHMEM: comms must use symmetric heap
+//////////////////////////////////////////////////////////////////////////////////////////
 #ifdef GRID_COMMS_SHMEM
-
-    _Tp *ptr = (_Tp *) shmem_align(__n*sizeof(_Tp),64);
-
-
+extern "C" { 
+#include <mpp/shmem.h>
+extern void * shmem_align(size_t, size_t);
+extern void  shmem_free(void *);
+}
 #define PARANOID_SYMMETRIC_HEAP
+#endif
+
+template<typename _Tp>
+class commAllocator {
+public: 
+  typedef std::size_t     size_type;
+  typedef std::ptrdiff_t  difference_type;
+  typedef _Tp*       pointer;
+  typedef const _Tp* const_pointer;
+  typedef _Tp&       reference;
+  typedef const _Tp& const_reference;
+  typedef _Tp        value_type;
+
+  template<typename _Tp1>  struct rebind { typedef commAllocator<_Tp1> other; };
+  commAllocator() throw() { }
+  commAllocator(const commAllocator&) throw() { }
+  template<typename _Tp1> commAllocator(const commAllocator<_Tp1>&) throw() { }
+  ~commAllocator() throw() { }
+  pointer       address(reference __x)       const { return &__x; }
+  size_type  max_size() const throw() { return size_t(-1) / sizeof(_Tp); }
+
+#ifdef GRID_COMMS_SHMEM
+  pointer allocate(size_type __n, const void* _p= 0)
+  {
+#ifdef CRAY
+    _Tp *ptr = (_Tp *) shmem_align(__n*sizeof(_Tp),64);
+#else
+    _Tp *ptr = (_Tp *) shmem_align(64,__n*sizeof(_Tp));
+#endif
 #ifdef PARANOID_SYMMETRIC_HEAP
     static void * bcast;
     static long  psync[_SHMEM_REDUCE_SYNC_SIZE];
@@ -96,54 +168,47 @@ public:
 
     if ( bcast != ptr ) {
       std::printf("inconsistent alloc pe %d %lx %lx \n",shmem_my_pe(),bcast,ptr);std::fflush(stdout);
-      BACKTRACEFILE();
+      //      BACKTRACEFILE();
       exit(0);
     }
-
     assert( bcast == (void *) ptr);
-
 #endif 
+    return ptr;
+  }
+  void deallocate(pointer __p, size_type) { 
+    shmem_free((void *)__p);
+  }
 #else
-
+  pointer allocate(size_type __n, const void* _p= 0) 
+  {
 #ifdef HAVE_MM_MALLOC_H
     _Tp * ptr = (_Tp *) _mm_malloc(__n*sizeof(_Tp),128);
 #else
     _Tp * ptr = (_Tp *) memalign(128,__n*sizeof(_Tp));
 #endif
-
-#endif
-    _Tp tmp;
-#ifdef GRID_NUMA
-#pragma omp parallel for schedule(static)
-  for(int i=0;i<__n;i++){
-    ptr[i]=tmp;
-  }
-#endif 
     return ptr;
   }
-
   void deallocate(pointer __p, size_type) { 
-#ifdef GRID_COMMS_SHMEM
-    shmem_free((void *)__p);
-#else
 #ifdef HAVE_MM_MALLOC_H
     _mm_free((void *)__p); 
 #else
     free((void *)__p);
-#endif
 #endif
   }
+#endif
   void construct(pointer __p, const _Tp& __val) { };
   void construct(pointer __p) { };
-
   void destroy(pointer __p) { };
 };
+template<typename _Tp>  inline bool operator==(const commAllocator<_Tp>&, const commAllocator<_Tp>&){ return true; }
+template<typename _Tp>  inline bool operator!=(const commAllocator<_Tp>&, const commAllocator<_Tp>&){ return false; }
 
-template<typename _Tp>  inline bool
-operator==(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return true; }
-
-template<typename _Tp>  inline bool
-operator!=(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return false; }
+////////////////////////////////////////////////////////////////////////////////
+// Template typedefs
+////////////////////////////////////////////////////////////////////////////////
+template<class T> using Vector     = std::vector<T,alignedAllocator<T> >;           
+template<class T> using commVector = std::vector<T,commAllocator<T> >;              
+template<class T> using Matrix     = std::vector<std::vector<T,alignedAllocator<T> > >;
     
 }; // namespace Grid
 #endif

lib/cartesian/Cartesian_base.h

@@ -52,7 +52,7 @@ public:
 
     // Physics Grid information.
     std::vector<int> _simd_layout;// Which dimensions get relayed out over simd lanes.
-    std::vector<int> _fdimensions;// Global dimensions of array prior to cb removal
+    std::vector<int> _fdimensions;// (full) Global dimensions of array prior to cb removal
     std::vector<int> _gdimensions;// Global dimensions of array after cb removal
     std::vector<int> _ldimensions;// local dimensions of array with processor images removed
     std::vector<int> _rdimensions;// Reduced local dimensions with simd lane images and processor images removed 
@@ -77,7 +77,7 @@ public:
     // GridCartesian / GridRedBlackCartesian
     ////////////////////////////////////////////////////////////////
     virtual int CheckerBoarded(int dim)=0;
-    virtual int CheckerBoard(std::vector<int> site)=0;
+    virtual int CheckerBoard(const std::vector<int> &site)=0;
     virtual int CheckerBoardDestination(int source_cb,int shift,int dim)=0;
     virtual int CheckerBoardShift(int source_cb,int dim,int shift,int osite)=0;
     virtual int CheckerBoardShiftForCB(int source_cb,int dim,int shift,int cb)=0;
@@ -121,7 +121,6 @@ public:
       Lexicographic::CoorFromIndex(coor,Oindex,_rdimensions);
     }
 
-
     //////////////////////////////////////////////////////////
     // SIMD lane addressing
     //////////////////////////////////////////////////////////
@@ -207,16 +206,16 @@ public:
       std::vector<int> lcoor;
       GlobalCoorToProcessorCoorLocalCoor(pcoor,lcoor,gcoor);
       rank = RankFromProcessorCoor(pcoor);
-
+      /*
       std::vector<int> cblcoor(lcoor);
       for(int d=0;d<cblcoor.size();d++){
 	if( this->CheckerBoarded(d) ) {
 	  cblcoor[d] = lcoor[d]/2;
 	}
       }
-
-      i_idx= iIndex(cblcoor);// this does not imply divide by 2 on checker dim
-      o_idx= oIndex(lcoor);  // this implies divide by 2 on checkerdim
+      */
+      i_idx= iIndex(lcoor);
+      o_idx= oIndex(lcoor);
     }
 
     void RankIndexToGlobalCoor(int rank, int o_idx, int i_idx , std::vector<int> &gcoor)

lib/cartesian/Cartesian_full.h

@@ -49,7 +49,7 @@ public:
     virtual int CheckerBoarded(int dim){
       return 0;
     }
-    virtual int CheckerBoard(std::vector<int> site){
+    virtual int CheckerBoard(const std::vector<int> &site){
         return 0;
     }
     virtual int CheckerBoardDestination(int cb,int shift,int dim){

lib/cartesian/Cartesian_red_black.h

@@ -49,7 +49,7 @@ public:
       if( dim==_checker_dim) return 1;
       else return 0;
     }
-    virtual int CheckerBoard(std::vector<int> site){
+    virtual int CheckerBoard(const std::vector<int> &site){
       int linear=0;
       assert(site.size()==_ndimension);
       for(int d=0;d<_ndimension;d++){ 

lib/communicator/Communicator_base.cc

@@ -0,0 +1,129 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/communicator/Communicator_none.cc
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#include <Grid/GridCore.h>
+
+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;
+
+/////////////////////////////////
+// Alloc, free shmem region
+/////////////////////////////////
+void *CartesianCommunicator::ShmBufferMalloc(size_t bytes){
+  //  bytes = (bytes+sizeof(vRealD))&(~(sizeof(vRealD)-1));// align up bytes
+  void *ptr = (void *)heap_top;
+  heap_top  += bytes;
+  heap_bytes+= bytes;
+  if (heap_bytes >= MAX_MPI_SHM_BYTES) {
+    std::cout<< " ShmBufferMalloc exceeded shared heap size -- try increasing with --shm <MB> flag" <<std::endl;
+    std::cout<< " Parameter specified in units of MB (megabytes) " <<std::endl;
+    std::cout<< " Current value is " << (MAX_MPI_SHM_BYTES/(1024*1024)) <<std::endl;
+    assert(heap_bytes<MAX_MPI_SHM_BYTES);
+  }
+  return ptr;
+}
+void CartesianCommunicator::ShmBufferFreeAll(void) { 
+  heap_top  =(size_t)ShmBufferSelf();
+  heap_bytes=0;
+}
+
+/////////////////////////////////
+// Grid information queries
+/////////////////////////////////
+int                      CartesianCommunicator::IsBoss(void)            { return _processor==0; };
+int                      CartesianCommunicator::BossRank(void)          { return 0; };
+int                      CartesianCommunicator::ThisRank(void)          { return _processor; };
+const std::vector<int> & CartesianCommunicator::ThisProcessorCoor(void) { return _processor_coor; };
+const std::vector<int> & CartesianCommunicator::ProcessorGrid(void)     { return _processors; };
+int                      CartesianCommunicator::ProcessorCount(void)    { return _Nprocessors; };
+
+////////////////////////////////////////////////////////////////////////////////
+// very VERY rarely (Log, serial RNG) we need world without a grid
+////////////////////////////////////////////////////////////////////////////////
+
+void CartesianCommunicator::GlobalSum(ComplexF &c)
+{
+  GlobalSumVector((float *)&c,2);
+}
+void CartesianCommunicator::GlobalSumVector(ComplexF *c,int N)
+{
+  GlobalSumVector((float *)c,2*N);
+}
+void CartesianCommunicator::GlobalSum(ComplexD &c)
+{
+  GlobalSumVector((double *)&c,2);
+}
+void CartesianCommunicator::GlobalSumVector(ComplexD *c,int N)
+{
+  GlobalSumVector((double *)c,2*N);
+}
+
+#if !defined( GRID_COMMS_MPI3) && !defined (GRID_COMMS_MPI3L)
+
+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)
+{
+  SendToRecvFromBegin(list,xmit,xmit_to_rank,recv,recv_from_rank,bytes);
+  return 2.0*bytes;
+}
+void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall)
+{
+  SendToRecvFromComplete(waitall);
+}
+void CartesianCommunicator::StencilBarrier(void){};
+
+commVector<uint8_t> CartesianCommunicator::ShmBufStorageVector;
+
+void *CartesianCommunicator::ShmBufferSelf(void) { return ShmCommBuf; }
+
+void *CartesianCommunicator::ShmBuffer(int rank) {
+  return NULL;
+}
+void *CartesianCommunicator::ShmBufferTranslate(int rank,void * local_p) { 
+  return NULL;
+}
+void CartesianCommunicator::ShmInitGeneric(void){
+  ShmBufStorageVector.resize(MAX_MPI_SHM_BYTES);
+  ShmCommBuf=(void *)&ShmBufStorageVector[0];
+}
+
+#endif
+  
+}
+

lib/communicator/Communicator_base.h

@@ -1,3 +1,4 @@
+
     /*************************************************************************************
 
     Grid physics library, www.github.com/paboyle/Grid 
@@ -34,123 +35,203 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifdef GRID_COMMS_MPI
 #include <mpi.h>
 #endif
+#ifdef GRID_COMMS_MPI3
+#include <mpi.h>
+#endif
+#ifdef GRID_COMMS_MPI3L
+#include <mpi.h>
+#endif
 #ifdef GRID_COMMS_SHMEM
 #include <mpp/shmem.h>
 #endif
+
 namespace Grid {
+
 class CartesianCommunicator {
   public:    
 
+  // 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;
+
   // Communicator should know nothing of the physics grid, only processor grid.
+  int              _Nprocessors;     // How many in all
+  std::vector<int> _processors;      // Which dimensions get relayed out over processors lanes.
+  int              _processor;       // linear processor rank
+  std::vector<int> _processor_coor;  // linear processor coordinate
+  unsigned long _ndimension;
 
-    int              _Nprocessors;     // How many in all
-    std::vector<int> _processors;      // Which dimensions get relayed out over processors lanes.
-    int              _processor;       // linear processor rank
-    std::vector<int> _processor_coor;  // linear processor coordinate
-    unsigned long _ndimension;
-
-#ifdef GRID_COMMS_MPI
-    MPI_Comm communicator;
-    typedef MPI_Request CommsRequest_t;
+#if defined (GRID_COMMS_MPI) || defined (GRID_COMMS_MPI3) || defined (GRID_COMMS_MPI3L)
+  static MPI_Comm communicator_world;
+         MPI_Comm communicator;
+  typedef MPI_Request CommsRequest_t;
 #else 
-    typedef int CommsRequest_t;
+  typedef int CommsRequest_t;
 #endif
 
-    static void Init(int *argc, char ***argv);
+  ////////////////////////////////////////////////////////////////////
+  // Helper functionality for SHM Windows common to all other impls
+  ////////////////////////////////////////////////////////////////////
+  // Longer term; drop this in favour of a master / slave model with 
+  // cartesian communicator on a subset of ranks, slave ranks controlled
+  // by group leader with data xfer via shared memory
+  ////////////////////////////////////////////////////////////////////
+#ifdef GRID_COMMS_MPI3
 
-    // Constructor
-    CartesianCommunicator(const std::vector<int> &pdimensions_in);
+  static int ShmRank;
+  static int ShmSize;
+  static int GroupRank;
+  static int GroupSize;
+  static int WorldRank;
+  static int WorldSize;
 
-    // Wraps MPI_Cart routines
-    void ShiftedRanks(int dim,int shift,int & source, int & dest);
-    int  RankFromProcessorCoor(std::vector<int> &coor);
-    void ProcessorCoorFromRank(int rank,std::vector<int> &coor);
+  std::vector<int>  WorldDims;
+  std::vector<int>  GroupDims;
+  std::vector<int>  ShmDims;
+  
+  std::vector<int> GroupCoor;
+  std::vector<int> ShmCoor;
+  std::vector<int> WorldCoor;
 
-    /////////////////////////////////
-    // Grid information queries
-    /////////////////////////////////
-    int                      IsBoss(void)            { return _processor==0; };
-    int                      BossRank(void)          { return 0; };
-    int                      ThisRank(void)          { return _processor; };
-    const std::vector<int> & ThisProcessorCoor(void) { return _processor_coor; };
-    const std::vector<int> & ProcessorGrid(void)     { return _processors; };
-    int                      ProcessorCount(void)    { return _Nprocessors; };
+  static std::vector<int> GroupRanks; 
+  static std::vector<int> MyGroup;
+  static int ShmSetup;
+  static MPI_Win ShmWindow; 
+  static MPI_Comm ShmComm;
+  
+  std::vector<int>  LexicographicToWorldRank;
+  
+  static std::vector<void *> ShmCommBufs;
 
-    ////////////////////////////////////////////////////////////
-    // Reduction
-    ////////////////////////////////////////////////////////////
-    void GlobalSum(RealF &);
-    void GlobalSumVector(RealF *,int N);
+#else 
+  static void ShmInitGeneric(void);
+  static commVector<uint8_t> ShmBufStorageVector;
+#endif 
 
-    void GlobalSum(RealD &);
-    void GlobalSumVector(RealD *,int N);
+  /////////////////////////////////
+  // Grid information and queries
+  // Implemented in Communicator_base.C
+  /////////////////////////////////
+  static void * ShmCommBuf;
 
-    void GlobalSum(uint32_t &);
-    void GlobalSum(uint64_t &);
+  // Isend/Irecv/Wait, or Sendrecv blocking
+  enum CommunicatorPolicy_t { CommunicatorPolicyConcurrent, CommunicatorPolicySequential };
+  static CommunicatorPolicy_t CommunicatorPolicy;
+  static void SetCommunicatorPolicy(CommunicatorPolicy_t policy ) { CommunicatorPolicy = policy; }
 
-    void GlobalSum(ComplexF &c)
-    {
-      GlobalSumVector((float *)&c,2);
-    }
-    void GlobalSumVector(ComplexF *c,int N)
-    {
-      GlobalSumVector((float *)c,2*N);
-    }
+  size_t heap_top;
+  size_t heap_bytes;
 
-    void GlobalSum(ComplexD &c)
-    {
-      GlobalSumVector((double *)&c,2);
-    }
-    void GlobalSumVector(ComplexD *c,int N)
-    {
-      GlobalSumVector((double *)c,2*N);
-    }
-    
-    template<class obj> void GlobalSum(obj &o){
-      typedef typename obj::scalar_type scalar_type;
-      int words = sizeof(obj)/sizeof(scalar_type);
-      scalar_type * ptr = (scalar_type *)& o;
-      GlobalSumVector(ptr,words);
-    }
-    ////////////////////////////////////////////////////////////
-    // Face exchange, buffer swap in translational invariant way
-    ////////////////////////////////////////////////////////////
-    void SendToRecvFrom(void *xmit,
-			int xmit_to_rank,
-			void *recv,
-			int recv_from_rank,
-			int bytes);
+  void *ShmBufferSelf(void);
+  void *ShmBuffer(int rank);
+  void *ShmBufferTranslate(int rank,void * local_p);
+  void *ShmBufferMalloc(size_t bytes);
+  void ShmBufferFreeAll(void) ;
+  
+  ////////////////////////////////////////////////
+  // Must call in Grid startup
+  ////////////////////////////////////////////////
+  static void Init(int *argc, char ***argv);
+  
+  ////////////////////////////////////////////////
+  // Constructor of any given grid
+  ////////////////////////////////////////////////
+  CartesianCommunicator(const std::vector<int> &pdimensions_in);
+  
+  ////////////////////////////////////////////////////////////////////////////////////////
+  // Wraps MPI_Cart routines, or implements equivalent on other impls
+  ////////////////////////////////////////////////////////////////////////////////////////
+  void ShiftedRanks(int dim,int shift,int & source, int & dest);
+  int  RankFromProcessorCoor(std::vector<int> &coor);
+  void ProcessorCoorFromRank(int rank,std::vector<int> &coor);
+  
+  int                      IsBoss(void)            ;
+  int                      BossRank(void)          ;
+  int                      ThisRank(void)          ;
+  const std::vector<int> & ThisProcessorCoor(void) ;
+  const std::vector<int> & ProcessorGrid(void)     ;
+  int                      ProcessorCount(void)    ;
+  int                      NodeCount(void)    ;
 
-    void SendRecvPacket(void *xmit,
-			void *recv,
-			int xmit_to_rank,
-			int recv_from_rank,
-			int bytes);
+  ////////////////////////////////////////////////////////////////////////////////
+  // very VERY rarely (Log, serial RNG) we need world without a grid
+  ////////////////////////////////////////////////////////////////////////////////
+  static int  RankWorld(void) ;
+  static void BroadcastWorld(int root,void* data, int bytes);
+  
+  ////////////////////////////////////////////////////////////
+  // Reduction
+  ////////////////////////////////////////////////////////////
+  void GlobalSum(RealF &);
+  void GlobalSumVector(RealF *,int N);
+  void GlobalSum(RealD &);
+  void GlobalSumVector(RealD *,int N);
+  void GlobalSum(uint32_t &);
+  void GlobalSum(uint64_t &);
+  void GlobalSum(ComplexF &c);
+  void GlobalSumVector(ComplexF *c,int N);
+  void GlobalSum(ComplexD &c);
+  void GlobalSumVector(ComplexD *c,int N);
+  
+  template<class obj> void GlobalSum(obj &o){
+    typedef typename obj::scalar_type scalar_type;
+    int words = sizeof(obj)/sizeof(scalar_type);
+    scalar_type * ptr = (scalar_type *)& o;
+    GlobalSumVector(ptr,words);
+  }
+  
+  ////////////////////////////////////////////////////////////
+  // Face exchange, buffer swap in translational invariant way
+  ////////////////////////////////////////////////////////////
+  void SendToRecvFrom(void *xmit,
+		      int xmit_to_rank,
+		      void *recv,
+		      int recv_from_rank,
+		      int bytes);
+  
+  void SendRecvPacket(void *xmit,
+		      void *recv,
+		      int xmit_to_rank,
+		      int recv_from_rank,
+		      int bytes);
+  
+  void SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+			   void *xmit,
+			   int xmit_to_rank,
+			   void *recv,
+			   int recv_from_rank,
+			   int bytes);
+  
+  void SendToRecvFromComplete(std::vector<CommsRequest_t> &waitall);
 
-    void SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
-			 void *xmit,
-			 int xmit_to_rank,
-			 void *recv,
-			 int recv_from_rank,
-			 int bytes);
-    void SendToRecvFromComplete(std::vector<CommsRequest_t> &waitall);
+  double StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+				  void *xmit,
+				  int xmit_to_rank,
+				  void *recv,
+				  int recv_from_rank,
+				  int bytes);
+  
+  void StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall);
+  void StencilBarrier(void);
 
-    ////////////////////////////////////////////////////////////
-    // Barrier
-    ////////////////////////////////////////////////////////////
-    void Barrier(void);
-
-    ////////////////////////////////////////////////////////////
-    // Broadcast a buffer and composite larger
-    ////////////////////////////////////////////////////////////
-    void Broadcast(int root,void* data, int bytes);
-    template<class obj> void Broadcast(int root,obj &data)
+  ////////////////////////////////////////////////////////////
+  // Barrier
+  ////////////////////////////////////////////////////////////
+  void Barrier(void);
+  
+  ////////////////////////////////////////////////////////////
+  // Broadcast a buffer and composite larger
+  ////////////////////////////////////////////////////////////
+  void Broadcast(int root,void* data, int bytes);
+  
+  template<class obj> void Broadcast(int root,obj &data)
     {
       Broadcast(root,(void *)&data,sizeof(data));
     };
 
-    static void BroadcastWorld(int root,void* data, int bytes);
-
 }; 
 }
 

lib/communicator/Communicator_mpi.cc

@@ -25,26 +25,34 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#include "Grid.h"
+#include <Grid/GridCore.h>
+#include <Grid/GridQCDcore.h>
+#include <Grid/qcd/action/ActionCore.h>
 #include <mpi.h>
 
 namespace Grid {
 
-  // Should error check all MPI calls.
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+// Info that is setup once and indept of cartesian layout
+///////////////////////////////////////////////////////////////////////////////////////////////////
+MPI_Comm CartesianCommunicator::communicator_world;
+
+// Should error check all MPI calls.
 void CartesianCommunicator::Init(int *argc, char ***argv) {
   int flag;
+  int provided;
   MPI_Initialized(&flag); // needed to coexist with other libs apparently
   if ( !flag ) {
-    MPI_Init(argc,argv);
+    MPI_Init_thread(argc,argv,MPI_THREAD_MULTIPLE,&provided);
+    if ( provided != MPI_THREAD_MULTIPLE ) {
+      QCD::WilsonKernelsStatic::Comms = QCD::WilsonKernelsStatic::CommsThenCompute;
+    }
   }
+  MPI_Comm_dup (MPI_COMM_WORLD,&communicator_world);
+  ShmInitGeneric();
 }
 
-  int Rank(void) {
-    int pe;
-    MPI_Comm_rank(MPI_COMM_WORLD,&pe);
-    return pe;
-  }
-
 CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
 {
   _ndimension = processors.size();
@@ -54,7 +62,7 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
   _processors = processors;
   _processor_coor.resize(_ndimension);
   
-  MPI_Cart_create(MPI_COMM_WORLD, _ndimension,&_processors[0],&periodic[0],1,&communicator);
+  MPI_Cart_create(communicator_world, _ndimension,&_processors[0],&periodic[0],1,&communicator);
   MPI_Comm_rank(communicator,&_processor);
   MPI_Cart_coords(communicator,_processor,_ndimension,&_processor_coor[0]);
 
@@ -67,7 +75,6 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
   
   assert(Size==_Nprocessors);
 }
-
 void CartesianCommunicator::GlobalSum(uint32_t &u){
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_SUM,communicator);
   assert(ierr==0);
@@ -151,25 +158,34 @@ void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &lis
 						int from,
 						int bytes)
 {
-  MPI_Request xrq;
-  MPI_Request rrq;
-  int rank = _processor;
+  int myrank = _processor;
   int ierr;
-  ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
-  ierr|=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
-  
-  assert(ierr==0);
+  if ( CommunicatorPolicy == CommunicatorPolicyConcurrent ) { 
+    MPI_Request xrq;
+    MPI_Request rrq;
 
-  list.push_back(xrq);
-  list.push_back(rrq);
+    ierr =MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
+    ierr|=MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
+    
+    assert(ierr==0);
+    list.push_back(xrq);
+    list.push_back(rrq);
+  } else { 
+    // Give the CPU to MPI immediately; can use threads to overlap optionally
+    ierr=MPI_Sendrecv(xmit,bytes,MPI_CHAR,dest,myrank,
+		      recv,bytes,MPI_CHAR,from, from,
+		      communicator,MPI_STATUS_IGNORE);
+    assert(ierr==0);
+  }
 }
 void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
 {
-  int nreq=list.size();
-  std::vector<MPI_Status> status(nreq);
-  int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
-
-  assert(ierr==0);
+  if ( CommunicatorPolicy == CommunicatorPolicyConcurrent ) { 
+    int nreq=list.size();
+    std::vector<MPI_Status> status(nreq);
+    int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
+    assert(ierr==0);
+  }
 }
 
 void CartesianCommunicator::Barrier(void)
@@ -187,14 +203,22 @@ void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
 		     communicator);
   assert(ierr==0);
 }
-
+  ///////////////////////////////////////////////////////
+  // Should only be used prior to Grid Init finished.
+  // Check for this?
+  ///////////////////////////////////////////////////////
+int CartesianCommunicator::RankWorld(void){ 
+  int r; 
+  MPI_Comm_rank(communicator_world,&r);
+  return r;
+}
 void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
 {
   int ierr= MPI_Bcast(data,
 		      bytes,
 		      MPI_BYTE,
 		      root,
-		      MPI_COMM_WORLD);
+		      communicator_world);
   assert(ierr==0);
 }
 

lib/communicator/Communicator_mpi3.cc

@@ -0,0 +1,686 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/communicator/Communicator_mpi.cc
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#include <Grid/GridCore.h>
+
+#include <mpi.h>
+
+#include <semaphore.h>
+#include <fcntl.h>
+#include <unistd.h>
+#include <limits.h>
+#include <sys/types.h>
+#include <sys/ipc.h>
+#include <sys/shm.h>
+#include <sys/mman.h>
+//#include <zlib.h>
+#ifndef SHM_HUGETLB
+#define SHM_HUGETLB 04000
+#endif
+
+namespace Grid {
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+// Info that is setup once and indept of cartesian layout
+///////////////////////////////////////////////////////////////////////////////////////////////////
+int CartesianCommunicator::ShmSetup = 0;
+
+int CartesianCommunicator::ShmRank;
+int CartesianCommunicator::ShmSize;
+int CartesianCommunicator::GroupRank;
+int CartesianCommunicator::GroupSize;
+int CartesianCommunicator::WorldRank;
+int CartesianCommunicator::WorldSize;
+
+MPI_Comm CartesianCommunicator::communicator_world;
+MPI_Comm CartesianCommunicator::ShmComm;
+MPI_Win  CartesianCommunicator::ShmWindow;
+
+std::vector<int> CartesianCommunicator::GroupRanks;  
+std::vector<int> CartesianCommunicator::MyGroup;
+std::vector<void *> CartesianCommunicator::ShmCommBufs;
+
+int CartesianCommunicator::NodeCount(void)    { return GroupSize;};
+
+
+#undef FORCE_COMMS
+void *CartesianCommunicator::ShmBufferSelf(void)
+{
+  return ShmCommBufs[ShmRank];
+}
+void *CartesianCommunicator::ShmBuffer(int rank)
+{
+  int gpeer = GroupRanks[rank];
+#ifdef FORCE_COMMS
+  return NULL;
+#endif
+  if (gpeer == MPI_UNDEFINED){
+    return NULL;
+  } else { 
+    return ShmCommBufs[gpeer];
+  }
+}
+void *CartesianCommunicator::ShmBufferTranslate(int rank,void * local_p)
+{
+  static int count =0;
+  int gpeer = GroupRanks[rank];
+  assert(gpeer!=ShmRank); // never send to self
+  assert(rank!=WorldRank);// never send to self
+#ifdef FORCE_COMMS
+  return NULL;
+#endif
+  if (gpeer == MPI_UNDEFINED){
+    return NULL;
+  } else { 
+    uint64_t offset = (uint64_t)local_p - (uint64_t)ShmCommBufs[ShmRank];
+    uint64_t remote = (uint64_t)ShmCommBufs[gpeer]+offset;
+    return (void *) remote;
+  }
+}
+
+void CartesianCommunicator::Init(int *argc, char ***argv) {
+
+  int flag;
+  int provided;
+  //  mtrace();
+
+  MPI_Initialized(&flag); // needed to coexist with other libs apparently
+  if ( !flag ) {
+    MPI_Init_thread(argc,argv,MPI_THREAD_MULTIPLE,&provided);
+    assert (provided == MPI_THREAD_MULTIPLE);
+  }
+
+  Grid_quiesce_nodes();
+
+  MPI_Comm_dup (MPI_COMM_WORLD,&communicator_world);
+  MPI_Comm_rank(communicator_world,&WorldRank);
+  MPI_Comm_size(communicator_world,&WorldSize);
+
+  if ( WorldRank == 0 ) {
+    std::cout << GridLogMessage<< "Initialising MPI "<< WorldRank <<"/"<<WorldSize <<std::endl;
+  }
+
+  /////////////////////////////////////////////////////////////////////
+  // Split into groups that can share memory
+  /////////////////////////////////////////////////////////////////////
+  MPI_Comm_split_type(communicator_world, MPI_COMM_TYPE_SHARED, 0, MPI_INFO_NULL,&ShmComm);
+  MPI_Comm_rank(ShmComm     ,&ShmRank);
+  MPI_Comm_size(ShmComm     ,&ShmSize);
+  GroupSize = WorldSize/ShmSize;
+
+  /////////////////////////////////////////////////////////////////////
+  // find world ranks in our SHM group (i.e. which ranks are on our node)
+  /////////////////////////////////////////////////////////////////////
+  MPI_Group WorldGroup, ShmGroup;
+  MPI_Comm_group (communicator_world, &WorldGroup); 
+  MPI_Comm_group (ShmComm, &ShmGroup);
+  
+  std::vector<int> world_ranks(WorldSize); 
+  GroupRanks.resize(WorldSize); 
+  for(int r=0;r<WorldSize;r++) world_ranks[r]=r;
+  
+  MPI_Group_translate_ranks (WorldGroup,WorldSize,&world_ranks[0],ShmGroup, &GroupRanks[0]); 
+
+  ///////////////////////////////////////////////////////////////////
+  // Identify who is in my group and noninate the leader
+  ///////////////////////////////////////////////////////////////////
+  int g=0;
+  MyGroup.resize(ShmSize);
+  for(int rank=0;rank<WorldSize;rank++){
+    if(GroupRanks[rank]!=MPI_UNDEFINED){
+      assert(g<ShmSize);
+      MyGroup[g++] = rank;
+    }
+  }
+  
+  std::sort(MyGroup.begin(),MyGroup.end(),std::less<int>());
+  int myleader = MyGroup[0];
+  
+  std::vector<int> leaders_1hot(WorldSize,0);
+  std::vector<int> leaders_group(GroupSize,0);
+  leaders_1hot [ myleader ] = 1;
+    
+  ///////////////////////////////////////////////////////////////////
+  // global sum leaders over comm world
+  ///////////////////////////////////////////////////////////////////
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&leaders_1hot[0],WorldSize,MPI_INT,MPI_SUM,communicator_world);
+  assert(ierr==0);
+  ///////////////////////////////////////////////////////////////////
+  // find the group leaders world rank
+  ///////////////////////////////////////////////////////////////////
+  int group=0;
+  for(int l=0;l<WorldSize;l++){
+    if(leaders_1hot[l]){
+      leaders_group[group++] = l;
+    }
+  }
+  ///////////////////////////////////////////////////////////////////
+  // Identify the rank of the group in which I (and my leader) live
+  ///////////////////////////////////////////////////////////////////
+  GroupRank=-1;
+  for(int g=0;g<GroupSize;g++){
+    if (myleader == leaders_group[g]){
+      GroupRank=g;
+    }
+  }
+  assert(GroupRank!=-1);
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  // allocate the shared window for our group
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  MPI_Barrier(ShmComm);
+
+  ShmCommBuf = 0;
+  ShmCommBufs.resize(ShmSize);
+
+#if 1
+  char shm_name [NAME_MAX];
+  if ( ShmRank == 0 ) {
+    for(int r=0;r<ShmSize;r++){
+
+      size_t size = CartesianCommunicator::MAX_MPI_SHM_BYTES;
+
+      sprintf(shm_name,"/Grid_mpi3_shm_%d_%d",GroupRank,r);
+
+      shm_unlink(shm_name);
+      int fd=shm_open(shm_name,O_RDWR|O_CREAT,0660);
+      if ( fd < 0 ) {	perror("failed shm_open");	assert(0);      }
+      ftruncate(fd, size);
+
+      void * ptr =  mmap(NULL,size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
+      if ( ptr == MAP_FAILED ) {       perror("failed mmap");      assert(0);    }
+      assert(((uint64_t)ptr&0x3F)==0);
+      ShmCommBufs[r] =ptr;
+      
+    }
+  }
+
+  MPI_Barrier(ShmComm);
+
+  if ( ShmRank != 0 ) { 
+    for(int r=0;r<ShmSize;r++){
+      size_t size = CartesianCommunicator::MAX_MPI_SHM_BYTES ;
+    
+      sprintf(shm_name,"/Grid_mpi3_shm_%d_%d",GroupRank,r);
+
+      int fd=shm_open(shm_name,O_RDWR,0660);
+      if ( fd<0 ) {	perror("failed shm_open");	assert(0);      }
+
+      void * ptr =  mmap(NULL,size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
+      if ( ptr == MAP_FAILED ) {       perror("failed mmap");      assert(0);    }
+      assert(((uint64_t)ptr&0x3F)==0);
+      ShmCommBufs[r] =ptr;
+    }
+  }
+
+#else
+  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) {
+	int errsv = errno;
+	printf("Errno %d\n",errsv);
+	perror("shmget");
+	exit(1);
+      }
+      printf("shmid: 0x%x\n", shmids[r]);
+    }
+  }
+  MPI_Barrier(ShmComm);
+  MPI_Bcast(&shmids[0],ShmSize*sizeof(int),MPI_BYTE,0,ShmComm);
+  MPI_Barrier(ShmComm);
+
+  for(int r=0;r<ShmSize;r++){
+    ShmCommBufs[r] = (uint64_t *)shmat(shmids[r], NULL,0);
+    if (ShmCommBufs[r] == (uint64_t *)-1) {
+      perror("Shared memory attach failure");
+      shmctl(shmids[r], IPC_RMID, NULL);
+      exit(2);
+    }
+    printf("shmaddr: %p\n", ShmCommBufs[r]);
+  }
+  MPI_Barrier(ShmComm);
+  // Mark for clean up
+  for(int r=0;r<ShmSize;r++){
+    shmctl(shmids[r], IPC_RMID,(struct shmid_ds *)NULL);
+  }
+  MPI_Barrier(ShmComm);
+
+#endif
+  ShmCommBuf         = ShmCommBufs[ShmRank];
+
+  MPI_Barrier(ShmComm);
+  if ( ShmRank == 0 ) {
+    for(int r=0;r<ShmSize;r++){
+      uint64_t * check = (uint64_t *) ShmCommBufs[r];
+      check[0] = GroupRank;
+      check[1] = r;
+      check[2] = 0x5A5A5A;
+    }
+  }
+
+  MPI_Barrier(ShmComm);
+  for(int r=0;r<ShmSize;r++){
+    uint64_t * check = (uint64_t *) ShmCommBufs[r];
+    
+    assert(check[0]==GroupRank);
+    assert(check[1]==r);
+    assert(check[2]==0x5A5A5A);
+
+  }
+  MPI_Barrier(ShmComm);
+
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Verbose for now
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  if (WorldRank == 0){
+    std::cout<<GridLogMessage<< "Grid MPI-3 configuration: detected ";
+    std::cout<< WorldSize << " Ranks " ;
+    std::cout<< GroupSize << " Nodes " ;
+    std::cout<< " with "<< ShmSize  << " ranks-per-node "<<std::endl;
+    
+    std::cout<<GridLogMessage     <<"Grid MPI-3 configuration: allocated shared memory region of size ";
+    std::cout<<std::hex << MAX_MPI_SHM_BYTES <<" ShmCommBuf address = "<<ShmCommBuf << std::dec<<std::endl;
+
+    for(int g=0;g<GroupSize;g++){
+      std::cout<<GridLogMessage<<" Node "<<g<<" led by MPI rank "<<leaders_group[g]<<std::endl;
+    }
+
+    std::cout<<GridLogMessage<<" Boss Node Shm Pointers are {";
+    for(int g=0;g<ShmSize;g++){
+      std::cout<<std::hex<<ShmCommBufs[g]<<std::dec;
+      if(g!=ShmSize-1) std::cout<<",";
+      else std::cout<<"}"<<std::endl;
+    }
+  }
+  
+  for(int g=0;g<GroupSize;g++){
+    if ( (ShmRank == 0) && (GroupRank==g) )  std::cout<<GridLogMessage<<"["<<g<<"] Node Group "<<g<<" is ranks {";
+    for(int r=0;r<ShmSize;r++){
+      if ( (ShmRank == 0) && (GroupRank==g) ) {
+	std::cout<<MyGroup[r];
+	if(r<ShmSize-1) std::cout<<",";
+	else std::cout<<"}"<<std::endl<<std::flush;
+      }
+      MPI_Barrier(communicator_world);
+    }
+  }
+
+  assert(ShmSetup==0);  ShmSetup=1;
+}
+
+////////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Want to implement some magic ... Group sub-cubes into those on same node
+////////////////////////////////////////////////////////////////////////////////////////////////////////////
+void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &dest,int &source)
+{
+  std::vector<int> coor = _processor_coor; // my coord
+  assert(std::abs(shift) <_processors[dim]);
+
+  coor[dim] = (_processor_coor[dim] + shift + _processors[dim])%_processors[dim];
+  Lexicographic::IndexFromCoor(coor,source,_processors);
+  source = LexicographicToWorldRank[source];
+
+  coor[dim] = (_processor_coor[dim] - shift + _processors[dim])%_processors[dim];
+  Lexicographic::IndexFromCoor(coor,dest,_processors);
+  dest = LexicographicToWorldRank[dest];
+
+}// rank is world rank.
+
+int CartesianCommunicator::RankFromProcessorCoor(std::vector<int> &coor)
+{
+  int rank;
+  Lexicographic::IndexFromCoor(coor,rank,_processors);
+  rank = LexicographicToWorldRank[rank];
+  return rank;
+}// rank is world rank
+
+void  CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor)
+{
+  int lr=-1;
+  for(int r=0;r<WorldSize;r++){// map world Rank to lexico and then to coor
+    if( LexicographicToWorldRank[r]==rank) lr = r;
+  }
+  assert(lr!=-1);
+  Lexicographic::CoorFromIndex(coor,lr,_processors);
+}
+CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
+{ 
+  int ierr;
+  communicator=communicator_world;
+  _ndimension = processors.size();
+
+  ////////////////////////////////////////////////////////////////
+  // Assert power of two shm_size.
+  ////////////////////////////////////////////////////////////////
+  int log2size = -1;
+  for(int i=0;i<=MAXLOG2RANKSPERNODE;i++){  
+    if ( (0x1<<i) == ShmSize ) {
+      log2size = i;
+      break;
+    }
+  }
+  assert(log2size != -1);
+
+  ////////////////////////////////////////////////////////////////
+  // Identify subblock of ranks on node spreading across dims
+  // in a maximally symmetrical way
+  ////////////////////////////////////////////////////////////////
+  std::vector<int> WorldDims = processors;
+
+  ShmDims.resize  (_ndimension,1);
+  GroupDims.resize(_ndimension);
+  ShmCoor.resize  (_ndimension);
+  GroupCoor.resize(_ndimension);
+  WorldCoor.resize(_ndimension);
+
+  int dim = 0;
+  for(int l2=0;l2<log2size;l2++){
+    while ( (WorldDims[dim] / ShmDims[dim]) <= 1 ) dim=(dim+1)%_ndimension;
+    ShmDims[dim]*=2;
+    dim=(dim+1)%_ndimension;
+  }
+
+  ////////////////////////////////////////////////////////////////
+  // Establish torus of processes and nodes with sub-blockings
+  ////////////////////////////////////////////////////////////////
+  for(int d=0;d<_ndimension;d++){
+    GroupDims[d] = WorldDims[d]/ShmDims[d];
+  }
+
+  ////////////////////////////////////////////////////////////////
+  // Verbose
+  ////////////////////////////////////////////////////////////////
+#if 0
+  std::cout<< GridLogMessage << "MPI-3 usage "<<std::endl;
+  std::cout<< GridLogMessage << "SHM   ";
+  for(int d=0;d<_ndimension;d++){
+    std::cout<< ShmDims[d] <<" ";
+  }
+  std::cout<< std::endl;
+
+  std::cout<< GridLogMessage << "Group ";
+  for(int d=0;d<_ndimension;d++){
+    std::cout<< GroupDims[d] <<" ";
+  }
+  std::cout<< std::endl;
+
+  std::cout<< GridLogMessage<<"World ";
+  for(int d=0;d<_ndimension;d++){
+    std::cout<< WorldDims[d] <<" ";
+  }
+  std::cout<< std::endl;
+#endif
+  ////////////////////////////////////////////////////////////////
+  // Check processor counts match
+  ////////////////////////////////////////////////////////////////
+  _Nprocessors=1;
+  _processors = processors;
+  _processor_coor.resize(_ndimension);
+  for(int i=0;i<_ndimension;i++){
+    _Nprocessors*=_processors[i];
+  }
+  assert(WorldSize==_Nprocessors);
+      
+  ////////////////////////////////////////////////////////////////
+  // Establish mapping between lexico physics coord and WorldRank
+  ////////////////////////////////////////////////////////////////
+  Lexicographic::CoorFromIndex(GroupCoor,GroupRank,GroupDims);
+  Lexicographic::CoorFromIndex(ShmCoor,ShmRank,ShmDims);
+  for(int d=0;d<_ndimension;d++){
+    WorldCoor[d] = GroupCoor[d]*ShmDims[d]+ShmCoor[d];
+  }
+  _processor_coor = WorldCoor;
+  _processor      = WorldRank;
+
+  ///////////////////////////////////////////////////////////////////
+  // global sum Lexico to World mapping
+  ///////////////////////////////////////////////////////////////////
+  int lexico;
+  LexicographicToWorldRank.resize(WorldSize,0);
+  Lexicographic::IndexFromCoor(WorldCoor,lexico,WorldDims);
+  LexicographicToWorldRank[lexico] = WorldRank;
+  ierr=MPI_Allreduce(MPI_IN_PLACE,&LexicographicToWorldRank[0],WorldSize,MPI_INT,MPI_SUM,communicator);
+  assert(ierr==0);
+
+  for(int i=0;i<WorldSize;i++){
+
+    int wr = LexicographicToWorldRank[i];
+    //    int wr = i;
+
+    std::vector<int> coor(_ndimension);
+    ProcessorCoorFromRank(wr,coor); // from world rank
+    int ck = RankFromProcessorCoor(coor);
+    assert(ck==wr);
+
+    if ( wr == WorldRank ) { 
+      for(int j=0;j<coor.size();j++) {
+	assert(coor[j] == _processor_coor[j]);
+      }
+    }
+    /*
+    std::cout << GridLogMessage<< " Lexicographic "<<i;
+    std::cout << " MPI rank      "<<wr;
+    std::cout << " Coor          ";
+    for(int j=0;j<coor.size();j++) std::cout << coor[j];
+    std::cout<< std::endl;
+    */
+    /////////////////////////////////////////////////////
+    // Check everyone agrees on everyone elses coords
+    /////////////////////////////////////////////////////
+    std::vector<int> mcoor = coor;
+    this->Broadcast(0,(void *)&mcoor[0],mcoor.size()*sizeof(int));
+    for(int d = 0 ; d< _ndimension; d++) {
+      assert(coor[d] == mcoor[d]);
+    }
+  }
+};
+void CartesianCommunicator::GlobalSum(uint32_t &u){
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_SUM,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalSum(uint64_t &u){
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_SUM,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalSum(float &f){
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalSumVector(float *f,int N)
+{
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,f,N,MPI_FLOAT,MPI_SUM,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalSum(double &d)
+{
+  int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_SUM,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalSumVector(double *d,int N)
+{
+  int ierr = MPI_Allreduce(MPI_IN_PLACE,d,N,MPI_DOUBLE,MPI_SUM,communicator);
+  assert(ierr==0);
+}
+// Basic Halo comms primitive
+void CartesianCommunicator::SendToRecvFrom(void *xmit,
+					   int dest,
+					   void *recv,
+					   int from,
+					   int bytes)
+{
+  std::vector<CommsRequest_t> reqs(0);
+  //    unsigned long  xcrc = crc32(0L, Z_NULL, 0);
+  //    unsigned long  rcrc = crc32(0L, Z_NULL, 0);
+  //    xcrc = crc32(xcrc,(unsigned char *)xmit,bytes);
+  SendToRecvFromBegin(reqs,xmit,dest,recv,from,bytes);
+  SendToRecvFromComplete(reqs);
+  //    rcrc = crc32(rcrc,(unsigned char *)recv,bytes);
+  //    printf("proc %d SendToRecvFrom %d bytes %lx %lx\n",_processor,bytes,xcrc,rcrc);
+}
+void CartesianCommunicator::SendRecvPacket(void *xmit,
+					   void *recv,
+					   int sender,
+					   int receiver,
+					   int bytes)
+{
+  MPI_Status stat;
+  assert(sender != receiver);
+  int tag = sender;
+  if ( _processor == sender ) {
+    MPI_Send(xmit, bytes, MPI_CHAR,receiver,tag,communicator);
+  }
+  if ( _processor == receiver ) { 
+    MPI_Recv(recv, bytes, MPI_CHAR,sender,tag,communicator,&stat);
+  }
+}
+// Basic Halo comms primitive
+void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+						void *xmit,
+						int dest,
+						void *recv,
+						int from,
+						int bytes)
+{
+  int myrank = _processor;
+  int ierr;
+
+  if ( CommunicatorPolicy == CommunicatorPolicyConcurrent ) { 
+    MPI_Request xrq;
+    MPI_Request rrq;
+
+    ierr =MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
+    ierr|=MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
+    
+    assert(ierr==0);
+    list.push_back(xrq);
+    list.push_back(rrq);
+  } else { 
+    // Give the CPU to MPI immediately; can use threads to overlap optionally
+    ierr=MPI_Sendrecv(xmit,bytes,MPI_CHAR,dest,myrank,
+		      recv,bytes,MPI_CHAR,from, from,
+		      communicator,MPI_STATUS_IGNORE);
+    assert(ierr==0);
+  }
+}
+
+double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+						       void *xmit,
+						       int dest,
+						       void *recv,
+						       int from,
+						       int bytes)
+{
+  MPI_Request xrq;
+  MPI_Request rrq;
+
+  int ierr;
+  int gdest = GroupRanks[dest];
+  int gfrom = GroupRanks[from];
+  int gme   = GroupRanks[_processor];
+
+  assert(dest != _processor);
+  assert(from != _processor);
+  assert(gme  == ShmRank);
+  double off_node_bytes=0.0;
+
+#ifdef FORCE_COMMS
+  gdest = MPI_UNDEFINED;
+  gfrom = MPI_UNDEFINED;
+#endif
+  if ( gfrom ==MPI_UNDEFINED) {
+    ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&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);
+    assert(ierr==0);
+    list.push_back(xrq);
+    off_node_bytes+=bytes;
+  }
+
+  if ( CommunicatorPolicy == CommunicatorPolicySequential ) { 
+    this->StencilSendToRecvFromComplete(list);
+  }
+
+  return off_node_bytes;
+}
+void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall)
+{
+  SendToRecvFromComplete(waitall);
+}
+void CartesianCommunicator::StencilBarrier(void)
+{
+  MPI_Barrier  (ShmComm);
+}
+void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
+{
+  int nreq=list.size();
+
+  if (nreq==0) return;
+
+  std::vector<MPI_Status> status(nreq);
+  int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
+  assert(ierr==0);
+  list.resize(0);
+}
+void CartesianCommunicator::Barrier(void)
+{
+  int ierr = MPI_Barrier(communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
+{
+  int ierr=MPI_Bcast(data,
+		     bytes,
+		     MPI_BYTE,
+		     root,
+		     communicator);
+  assert(ierr==0);
+}
+int CartesianCommunicator::RankWorld(void){ 
+  int r; 
+  MPI_Comm_rank(communicator_world,&r);
+  return r;
+}
+void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
+{
+  int ierr= MPI_Bcast(data,
+		      bytes,
+		      MPI_BYTE,
+		      root,
+		      communicator_world);
+  assert(ierr==0);
+}
+
+}
+

lib/communicator/Communicator_mpi3_leader.cc

@@ -0,0 +1,988 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/communicator/Communicator_mpi.cc
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#include "Grid.h"
+#include <mpi.h>
+//#include <numaif.h>
+
+////////////////////////////////////////////////////////////////////////////////////////////////////////////////
+/// Workarounds:
+/// i) bloody mac os doesn't implement unnamed semaphores since it is "optional" posix.
+///    darwin dispatch semaphores don't seem to be multiprocess.
+///
+/// ii) openmpi under --mca shmem posix works with two squadrons per node; 
+///     openmpi under default mca settings (I think --mca shmem mmap) on MacOS makes two squadrons map the SAME
+///     memory as each other, despite their living on different communicators. This appears to be a bug in OpenMPI.
+///
+////////////////////////////////////////////////////////////////////////////////////////////////////////////////
+#include <semaphore.h>
+#include <fcntl.h>
+#include <unistd.h>
+#include <limits.h>
+typedef sem_t *Grid_semaphore;
+
+
+#error  /*THis is deprecated*/
+
+#if 0 
+#define SEM_INIT(S)      S = sem_open(sem_name,0,0600,0); assert ( S != SEM_FAILED );
+#define SEM_INIT_EXCL(S) sem_unlink(sem_name); S = sem_open(sem_name,O_CREAT|O_EXCL,0600,0); assert ( S != SEM_FAILED );
+#define SEM_POST(S) assert ( sem_post(S) == 0 ); 
+#define SEM_WAIT(S) assert ( sem_wait(S) == 0 );
+#else
+#define SEM_INIT(S)      ;
+#define SEM_INIT_EXCL(S) ;
+#define SEM_POST(S) ;
+#define SEM_WAIT(S) ;
+#endif
+#include <sys/mman.h>
+
+namespace Grid {
+
+enum { COMMAND_ISEND, COMMAND_IRECV, COMMAND_WAITALL, COMMAND_SENDRECV };
+
+struct Descriptor {
+  uint64_t buf;
+  size_t bytes;
+  int rank;
+  int tag;
+  int command;
+  uint64_t xbuf;
+  uint64_t rbuf;
+  int xtag;
+  int rtag;
+  int src;
+  int dest;
+  MPI_Request request;
+};
+
+const int pool = 48;
+
+class SlaveState {
+public:
+  volatile int head;
+  volatile int start;
+  volatile int tail;
+  volatile Descriptor Descrs[pool];
+};
+
+class Slave {
+public:
+  Grid_semaphore  sem_head;
+  Grid_semaphore  sem_tail;
+  SlaveState *state;
+  MPI_Comm squadron;
+  uint64_t     base;
+  int universe_rank;
+  int vertical_rank;
+  char sem_name [NAME_MAX];
+  ////////////////////////////////////////////////////////////
+  // Descriptor circular pointers
+  ////////////////////////////////////////////////////////////
+  Slave() {};
+
+  void Init(SlaveState * _state,MPI_Comm _squadron,int _universe_rank,int _vertical_rank);
+
+  void SemInit(void) {
+    sprintf(sem_name,"/Grid_mpi3_sem_head_%d",universe_rank);
+    SEM_INIT(sem_head);
+    sprintf(sem_name,"/Grid_mpi3_sem_tail_%d",universe_rank);
+    SEM_INIT(sem_tail);
+  }  
+  void SemInitExcl(void) {
+    sprintf(sem_name,"/Grid_mpi3_sem_head_%d",universe_rank);
+    SEM_INIT_EXCL(sem_head);
+    sprintf(sem_name,"/Grid_mpi3_sem_tail_%d",universe_rank);
+    SEM_INIT_EXCL(sem_tail);
+  }  
+  void WakeUpDMA(void) { 
+    SEM_POST(sem_head);
+  };
+  void WakeUpCompute(void) { 
+    SEM_POST(sem_tail);
+  };
+  void WaitForCommand(void) { 
+    SEM_WAIT(sem_head);
+  };
+  void WaitForComplete(void) { 
+    SEM_WAIT(sem_tail);
+  };
+  void EventLoop (void) {
+    //    std::cout<< " Entering event loop "<<std::endl;
+    while(1){
+      WaitForCommand();
+      //      std::cout << "Getting command "<<std::endl;
+#if 0
+      _mm_monitor((void *)&state->head,0,0);
+      int s=state->start;
+      if ( s != state->head ) {
+	_mm_mwait(0,0);
+      }
+#endif
+      Event();
+    }
+  }
+
+  int Event (void) ;
+
+  uint64_t QueueCommand(int command,void *buf, int bytes, int hashtag, MPI_Comm comm,int u_rank) ;
+  void QueueSendRecv(void *xbuf, void *rbuf, int bytes, int xtag, int rtag, MPI_Comm comm,int dest,int src) ;
+
+  void WaitAll() {
+    //    std::cout << "Queueing WAIT command  "<<std::endl;
+    QueueCommand(COMMAND_WAITALL,0,0,0,squadron,0);
+    //    std::cout << "Waking up DMA "<<std::endl;
+    WakeUpDMA();
+    //    std::cout << "Waiting from semaphore "<<std::endl;
+    WaitForComplete();
+    //    std::cout << "Checking FIFO is empty "<<std::endl;
+    while ( state->tail != state->head );
+  }
+};
+
+////////////////////////////////////////////////////////////////////////
+// One instance of a data mover.
+// Master and Slave must agree on location in shared memory
+////////////////////////////////////////////////////////////////////////
+
+class MPIoffloadEngine { 
+public:
+
+  static std::vector<Slave> Slaves;
+
+  static int ShmSetup;
+  
+  static int UniverseRank;
+  static int UniverseSize;
+  
+  static MPI_Comm communicator_universe;
+  static MPI_Comm communicator_cached;
+
+  static MPI_Comm HorizontalComm;
+  static int HorizontalRank;
+  static int HorizontalSize;
+  
+  static MPI_Comm VerticalComm;
+  static MPI_Win  VerticalWindow; 
+  static int VerticalSize;
+  static int VerticalRank;
+  
+  static std::vector<void *> VerticalShmBufs;
+  static std::vector<std::vector<int> > UniverseRanks;
+  static std::vector<int> UserCommunicatorToWorldRanks; 
+  
+  static MPI_Group WorldGroup, CachedGroup;
+  
+  static void CommunicatorInit (MPI_Comm &communicator_world,
+				MPI_Comm &ShmComm,
+				void * &ShmCommBuf);
+
+  static void MapCommRankToWorldRank(int &hashtag, int & comm_world_peer,int tag, MPI_Comm comm,int commrank);
+
+  /////////////////////////////////////////////////////////
+  // routines for master proc must handle any communicator
+  /////////////////////////////////////////////////////////
+
+  static void QueueSend(int slave,void *buf, int bytes, int tag, MPI_Comm comm,int rank) {
+     //    std::cout<< " Queueing send  "<< bytes<< " slave "<< slave << " to comm "<<rank  <<std::endl;
+    Slaves[slave].QueueCommand(COMMAND_ISEND,buf,bytes,tag,comm,rank);
+    //    std::cout << "Queued send command to rank "<< rank<< " via "<<slave <<std::endl;
+    Slaves[slave].WakeUpDMA();
+    //    std::cout << "Waking up DMA "<< slave<<std::endl;
+  };
+
+  static void QueueSendRecv(int slave,void *xbuf, void *rbuf, int bytes, int xtag, int rtag, MPI_Comm comm,int dest,int src) 
+  {
+    Slaves[slave].QueueSendRecv(xbuf,rbuf,bytes,xtag,rtag,comm,dest,src);
+    Slaves[slave].WakeUpDMA();
+  }
+
+  static void QueueRecv(int slave, void *buf, int bytes, int tag, MPI_Comm comm,int rank) {
+    //    std::cout<< " Queueing recv "<< bytes<< " slave "<< slave << " from comm "<<rank  <<std::endl;
+    Slaves[slave].QueueCommand(COMMAND_IRECV,buf,bytes,tag,comm,rank);
+    //    std::cout << "Queued recv command from rank "<< rank<< " via "<<slave <<std::endl;
+    Slaves[slave].WakeUpDMA();
+    //    std::cout << "Waking up DMA "<< slave<<std::endl;
+  };
+
+  static void WaitAll() {
+    for(int s=1;s<VerticalSize;s++) {
+      //      std::cout << "Waiting for slave "<< s<<std::endl;
+      Slaves[s].WaitAll();
+    }
+    //    std::cout << " Wait all Complete "<<std::endl;
+  };
+
+  static void GetWork(int nwork, int me, int & mywork, int & myoff,int units){
+    int basework = nwork/units;
+    int backfill = units-(nwork%units);
+    if ( me >= units ) { 
+      mywork = myoff = 0;
+    } else { 
+      mywork = (nwork+me)/units;
+      myoff  = basework * me;
+      if ( me > backfill ) 
+	myoff+= (me-backfill);
+    }
+    return;
+  };
+
+  static void QueueRoundRobinSendRecv(void *xbuf, void *rbuf, int bytes, int xtag, int rtag, MPI_Comm comm,int dest,int src) {
+    uint8_t * cxbuf = (uint8_t *) xbuf;
+    uint8_t * crbuf = (uint8_t *) rbuf;
+    static int rrp=0;
+    int procs = VerticalSize-1;
+    int myoff=0;
+    int mywork=bytes;
+    QueueSendRecv(rrp+1,&cxbuf[myoff],&crbuf[myoff],mywork,xtag,rtag,comm,dest,src);
+    rrp = rrp+1;
+    if ( rrp == (VerticalSize-1) ) rrp = 0;
+  }
+
+  static void QueueMultiplexedSendRecv(void *xbuf, void *rbuf, int bytes, int xtag, int rtag, MPI_Comm comm,int dest,int src) {
+    uint8_t * cxbuf = (uint8_t *) xbuf;
+    uint8_t * crbuf = (uint8_t *) rbuf;
+    int mywork, myoff, procs;
+    procs = VerticalSize-1;
+    for(int s=0;s<procs;s++) {
+      GetWork(bytes,s,mywork,myoff,procs);
+      QueueSendRecv(s+1,&cxbuf[myoff],&crbuf[myoff],mywork,xtag,rtag,comm,dest,src);
+    }
+  };
+  static void QueueMultiplexedSend(void *buf, int bytes, int tag, MPI_Comm comm,int rank) {
+    uint8_t * cbuf = (uint8_t *) buf;
+    int mywork, myoff, procs;
+    procs = VerticalSize-1;
+    for(int s=0;s<procs;s++) {
+      GetWork(bytes,s,mywork,myoff,procs);
+      QueueSend(s+1,&cbuf[myoff],mywork,tag,comm,rank);
+    }
+  };
+
+  static void QueueMultiplexedRecv(void *buf, int bytes, int tag, MPI_Comm comm,int rank) {
+    uint8_t * cbuf = (uint8_t *) buf;
+    int mywork, myoff, procs;
+    procs = VerticalSize-1;
+    for(int s=0;s<procs;s++) {
+      GetWork(bytes,s,mywork,myoff,procs);
+      QueueRecv(s+1,&cbuf[myoff],mywork,tag,comm,rank);
+    }
+  };
+
+};
+
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+// Info that is setup once and indept of cartesian layout
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+std::vector<Slave> MPIoffloadEngine::Slaves;
+    
+int MPIoffloadEngine::UniverseRank;
+int MPIoffloadEngine::UniverseSize;
+
+MPI_Comm  MPIoffloadEngine::communicator_universe;
+MPI_Comm  MPIoffloadEngine::communicator_cached;
+MPI_Group MPIoffloadEngine::WorldGroup;
+MPI_Group MPIoffloadEngine::CachedGroup;
+
+MPI_Comm MPIoffloadEngine::HorizontalComm;
+int      MPIoffloadEngine::HorizontalRank;
+int      MPIoffloadEngine::HorizontalSize;
+
+MPI_Comm MPIoffloadEngine::VerticalComm;
+int      MPIoffloadEngine::VerticalSize;
+int      MPIoffloadEngine::VerticalRank;
+MPI_Win  MPIoffloadEngine::VerticalWindow; 
+std::vector<void *>            MPIoffloadEngine::VerticalShmBufs;
+std::vector<std::vector<int> > MPIoffloadEngine::UniverseRanks;
+std::vector<int>               MPIoffloadEngine::UserCommunicatorToWorldRanks; 
+
+int CartesianCommunicator::NodeCount(void)    { return HorizontalSize;};
+int MPIoffloadEngine::ShmSetup = 0;
+
+void MPIoffloadEngine::CommunicatorInit (MPI_Comm &communicator_world,
+					 MPI_Comm &ShmComm,
+					 void * &ShmCommBuf)
+{      
+  int flag;
+  assert(ShmSetup==0);  
+  
+  //////////////////////////////////////////////////////////////////////
+  // Universe is all nodes prior to squadron grouping
+  //////////////////////////////////////////////////////////////////////
+  MPI_Comm_dup (MPI_COMM_WORLD,&communicator_universe);
+  MPI_Comm_rank(communicator_universe,&UniverseRank);
+  MPI_Comm_size(communicator_universe,&UniverseSize);
+  
+  /////////////////////////////////////////////////////////////////////
+  // Split into groups that can share memory (Verticals)
+  /////////////////////////////////////////////////////////////////////
+#undef MPI_SHARED_MEM_DEBUG
+#ifdef  MPI_SHARED_MEM_DEBUG
+  MPI_Comm_split(communicator_universe,(UniverseRank/4),UniverseRank,&VerticalComm);
+#else 
+  MPI_Comm_split_type(communicator_universe, MPI_COMM_TYPE_SHARED, 0, MPI_INFO_NULL,&VerticalComm);
+#endif
+  MPI_Comm_rank(VerticalComm     ,&VerticalRank);
+  MPI_Comm_size(VerticalComm     ,&VerticalSize);
+  
+  //////////////////////////////////////////////////////////////////////
+  // Split into horizontal groups by rank in squadron
+  //////////////////////////////////////////////////////////////////////
+  MPI_Comm_split(communicator_universe,VerticalRank,UniverseRank,&HorizontalComm);
+  MPI_Comm_rank(HorizontalComm,&HorizontalRank);
+  MPI_Comm_size(HorizontalComm,&HorizontalSize);
+  assert(HorizontalSize*VerticalSize==UniverseSize);
+  
+  ////////////////////////////////////////////////////////////////////////////////
+  // What is my place in the world
+  ////////////////////////////////////////////////////////////////////////////////
+  int WorldRank=0;
+  if(VerticalRank==0) WorldRank = HorizontalRank;
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&WorldRank,1,MPI_INT,MPI_SUM,VerticalComm);
+  assert(ierr==0);
+  
+  ////////////////////////////////////////////////////////////////////////////////
+  // Where is the world in the universe?
+  ////////////////////////////////////////////////////////////////////////////////
+  UniverseRanks = std::vector<std::vector<int> >(HorizontalSize,std::vector<int>(VerticalSize,0));
+  UniverseRanks[WorldRank][VerticalRank] = UniverseRank;
+  for(int w=0;w<HorizontalSize;w++){
+    ierr=MPI_Allreduce(MPI_IN_PLACE,&UniverseRanks[w][0],VerticalSize,MPI_INT,MPI_SUM,communicator_universe);
+    assert(ierr==0);
+  }
+  
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  // allocate the shared window for our group, pass back Shm info to CartesianCommunicator
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  VerticalShmBufs.resize(VerticalSize);
+
+#undef MPI_SHARED_MEM
+#ifdef MPI_SHARED_MEM
+  ierr = MPI_Win_allocate_shared(CartesianCommunicator::MAX_MPI_SHM_BYTES,1,MPI_INFO_NULL,VerticalComm,&ShmCommBuf,&VerticalWindow);
+  ierr|= MPI_Win_lock_all (MPI_MODE_NOCHECK, VerticalWindow);
+  assert(ierr==0);
+  //  std::cout<<"SHM "<<ShmCommBuf<<std::endl;
+
+  for(int r=0;r<VerticalSize;r++){
+    MPI_Aint sz;
+    int dsp_unit;
+    MPI_Win_shared_query (VerticalWindow, r, &sz, &dsp_unit, &VerticalShmBufs[r]);
+    //    std::cout<<"SHM "<<r<<" " <<VerticalShmBufs[r]<<std::endl;
+  }
+#else 
+  char shm_name [NAME_MAX];
+  MPI_Barrier(VerticalComm);
+
+  if ( VerticalRank == 0 ) {
+    for(int r=0;r<VerticalSize;r++){
+
+      size_t size = CartesianCommunicator::MAX_MPI_SHM_BYTES;
+      if ( r>0 ) size = sizeof(SlaveState);
+
+      sprintf(shm_name,"/Grid_mpi3_shm_%d_%d",WorldRank,r);
+      
+      shm_unlink(shm_name);
+
+      int fd=shm_open(shm_name,O_RDWR|O_CREAT,0600);
+      if ( fd < 0 ) {
+	perror("failed shm_open");
+	assert(0);
+      }
+
+      ftruncate(fd, size);
+
+      VerticalShmBufs[r] = mmap(NULL,size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
+      if ( VerticalShmBufs[r] == MAP_FAILED ) { 
+	perror("failed mmap");
+	assert(0);
+      }
+
+      /*
+      for(uint64_t page=0;page<size;page+=4096){
+	void *pages = (void *) ( page + (uint64_t)VerticalShmBufs[r] );
+	int status;
+	int flags=MPOL_MF_MOVE_ALL;
+	int nodes=1; // numa domain == MCDRAM
+	unsigned long count=1;
+	ierr= move_pages(0,count, &pages,&nodes,&status,flags);
+	if (ierr && (page==0)) perror("numa relocate command failed");
+      }
+      */
+      uint64_t * check = (uint64_t *) VerticalShmBufs[r];
+      check[0] = WorldRank;
+      check[1] = r;
+
+      //      std::cout<<"SHM "<<r<<" " <<VerticalShmBufs[r]<<std::endl;
+    }
+  }
+
+  MPI_Barrier(VerticalComm);
+
+  if ( VerticalRank != 0 ) { 
+  for(int r=0;r<VerticalSize;r++){
+
+    size_t size = CartesianCommunicator::MAX_MPI_SHM_BYTES ;
+    if ( r>0 ) size = sizeof(SlaveState);
+    
+    sprintf(shm_name,"/Grid_mpi3_shm_%d_%d",WorldRank,r);
+    
+    int fd=shm_open(shm_name,O_RDWR|O_CREAT,0600);
+    if ( fd<0 ) {
+      perror("failed shm_open");
+      assert(0);
+    }
+    VerticalShmBufs[r] = mmap(NULL,size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
+
+    uint64_t * check = (uint64_t *) VerticalShmBufs[r];
+    assert(check[0]== WorldRank);
+    assert(check[1]== r);
+    //    std::cerr<<"SHM "<<r<<" " <<VerticalShmBufs[r]<<std::endl;
+  }
+  }
+#endif
+  MPI_Barrier(VerticalComm);
+
+  //////////////////////////////////////////////////////////////////////
+  // Map rank of leader on node in their in new world, to the
+  // rank in this vertical plane's horizontal communicator
+  //////////////////////////////////////////////////////////////////////
+  communicator_world = HorizontalComm;
+  ShmComm            = VerticalComm;
+  ShmCommBuf         = VerticalShmBufs[0];
+  MPI_Comm_group (communicator_world, &WorldGroup); 
+  
+  ///////////////////////////////////////////////////////////
+  // Start the slave data movers
+  ///////////////////////////////////////////////////////////
+  if ( VerticalRank != 0 ) {
+    Slave indentured;
+    indentured.Init( (SlaveState *) VerticalShmBufs[VerticalRank], VerticalComm, UniverseRank,VerticalRank);
+    indentured.SemInitExcl();// init semaphore in shared memory
+    MPI_Barrier(VerticalComm);
+    MPI_Barrier(VerticalComm);
+    indentured.EventLoop();
+    assert(0);
+  } else {
+    Slaves.resize(VerticalSize);
+    for(int i=1;i<VerticalSize;i++){
+      Slaves[i].Init((SlaveState *)VerticalShmBufs[i],VerticalComm, UniverseRanks[HorizontalRank][i],i);
+    }
+    MPI_Barrier(VerticalComm);
+    for(int i=1;i<VerticalSize;i++){
+      Slaves[i].SemInit();// init semaphore in shared memory
+    }
+    MPI_Barrier(VerticalComm);
+  }
+  
+  ///////////////////////////////////////////////////////////
+  // Verbose for now
+  ///////////////////////////////////////////////////////////
+  
+  ShmSetup=1;
+  
+  if (UniverseRank == 0){
+      
+    std::cout<<GridLogMessage << "Grid MPI-3 configuration: detected ";
+    std::cout<<UniverseSize   << " Ranks " ;
+    std::cout<<HorizontalSize << " Nodes " ;
+    std::cout<<VerticalSize   << " with ranks-per-node "<<std::endl;
+    
+    std::cout<<GridLogMessage << "Grid MPI-3 configuration: using one lead process per node " << std::endl;
+    std::cout<<GridLogMessage << "Grid MPI-3 configuration: reduced communicator has size " << HorizontalSize << std::endl;
+    
+    for(int g=0;g<HorizontalSize;g++){
+      std::cout<<GridLogMessage<<" Node "<<g<<" led by MPI rank "<< UniverseRanks[g][0]<<std::endl;
+    }
+    
+    for(int g=0;g<HorizontalSize;g++){
+      std::cout<<GridLogMessage<<" { ";
+      for(int s=0;s<VerticalSize;s++){
+	std::cout<< UniverseRanks[g][s];
+	if ( s<VerticalSize-1 ) {
+	  std::cout<<",";
+	}
+      }
+      std::cout<<" } "<<std::endl;
+    }
+  }
+};
+
+  ///////////////////////////////////////////////////////////////////////////////////////////////
+  // Map the communicator into communicator_world, and find the neighbour.
+  // Cache the mappings; cache size is 1.
+  ///////////////////////////////////////////////////////////////////////////////////////////////
+void MPIoffloadEngine::MapCommRankToWorldRank(int &hashtag, int & comm_world_peer,int tag, MPI_Comm comm,int rank) {
+
+  if ( comm == HorizontalComm ) {
+    comm_world_peer = rank;
+    //    std::cout << " MapCommRankToWorldRank  horiz " <<rank<<"->"<<comm_world_peer<<std::endl;
+  } else if ( comm == communicator_cached ) {
+    comm_world_peer = UserCommunicatorToWorldRanks[rank];
+    //    std::cout << " MapCommRankToWorldRank  cached " <<rank<<"->"<<comm_world_peer<<std::endl;
+  } else { 
+    
+    int size;
+
+    MPI_Comm_size(comm,&size);
+
+    UserCommunicatorToWorldRanks.resize(size);
+
+    std::vector<int> cached_ranks(size); 
+
+    for(int r=0;r<size;r++) {
+      cached_ranks[r]=r;
+    }
+
+    communicator_cached=comm;
+    
+    MPI_Comm_group(communicator_cached, &CachedGroup);
+    
+    MPI_Group_translate_ranks(CachedGroup,size,&cached_ranks[0],WorldGroup, &UserCommunicatorToWorldRanks[0]); 
+    
+    comm_world_peer = UserCommunicatorToWorldRanks[rank];
+    //    std::cout << " MapCommRankToWorldRank  cache miss " <<rank<<"->"<<comm_world_peer<<std::endl;
+    
+    assert(comm_world_peer != MPI_UNDEFINED);
+  }
+
+  assert( (tag & (~0xFFFFL)) ==0); 
+  
+  uint64_t icomm = (uint64_t)comm;
+  int comm_hash = ((icomm>>0 )&0xFFFF)^((icomm>>16)&0xFFFF)
+                ^ ((icomm>>32)&0xFFFF)^((icomm>>48)&0xFFFF);
+  
+  //  hashtag = (comm_hash<<15) | tag;      
+  hashtag = tag;      
+
+};
+
+void Slave::Init(SlaveState * _state,MPI_Comm _squadron,int _universe_rank,int _vertical_rank)
+{
+  squadron=_squadron;
+  universe_rank=_universe_rank;
+  vertical_rank=_vertical_rank;
+  state   =_state;
+  //  std::cout << "state "<<_state<<" comm "<<_squadron<<" universe_rank"<<universe_rank <<std::endl;
+  state->head = state->tail = state->start = 0;
+  base = (uint64_t)MPIoffloadEngine::VerticalShmBufs[0];
+  int rank; MPI_Comm_rank(_squadron,&rank);
+}
+#define PERI_PLUS(A) ( (A+1)%pool )
+int Slave::Event (void) {
+
+  static int tail_last;
+  static int head_last;
+  static int start_last;
+  int ierr;
+  MPI_Status stat;
+  static int i=0;
+
+  ////////////////////////////////////////////////////
+  // Try to advance the start pointers
+  ////////////////////////////////////////////////////
+  int s=state->start;
+  if ( s != state->head ) {
+    switch ( state->Descrs[s].command ) {
+    case COMMAND_ISEND:
+      ierr = MPI_Isend((void *)(state->Descrs[s].buf+base), 
+		       state->Descrs[s].bytes, 
+		       MPI_CHAR,
+		       state->Descrs[s].rank,
+		       state->Descrs[s].tag,
+		       MPIoffloadEngine::communicator_universe,
+		       (MPI_Request *)&state->Descrs[s].request);
+      assert(ierr==0);
+      state->start = PERI_PLUS(s);
+      return 1;
+      break;
+
+    case COMMAND_IRECV:
+      ierr=MPI_Irecv((void *)(state->Descrs[s].buf+base), 
+		     state->Descrs[s].bytes, 
+		     MPI_CHAR,
+		     state->Descrs[s].rank,
+		     state->Descrs[s].tag,
+		     MPIoffloadEngine::communicator_universe,
+		     (MPI_Request *)&state->Descrs[s].request);
+
+      //      std::cout<< " Request is "<<state->Descrs[s].request<<std::endl;
+      //      std::cout<< " Request0 is "<<state->Descrs[0].request<<std::endl;
+      assert(ierr==0);
+      state->start = PERI_PLUS(s);
+      return 1;
+      break;
+
+    case COMMAND_SENDRECV:
+
+      //      fprintf(stderr,"Sendrecv ->%d %d : <-%d %d \n",state->Descrs[s].dest, state->Descrs[s].xtag+i*10,state->Descrs[s].src, state->Descrs[s].rtag+i*10);
+
+      ierr=MPI_Sendrecv((void *)(state->Descrs[s].xbuf+base), state->Descrs[s].bytes, MPI_CHAR, state->Descrs[s].dest, state->Descrs[s].xtag+i*10,
+			(void *)(state->Descrs[s].rbuf+base), state->Descrs[s].bytes, MPI_CHAR, state->Descrs[s].src , state->Descrs[s].rtag+i*10,
+			MPIoffloadEngine::communicator_universe,MPI_STATUS_IGNORE);
+
+      assert(ierr==0);
+
+      //      fprintf(stderr,"Sendrecv done %d %d\n",ierr,i);
+      //      MPI_Barrier(MPIoffloadEngine::HorizontalComm);
+      //      fprintf(stderr,"Barrier\n");
+      i++;
+
+      state->start = PERI_PLUS(s);
+
+      return 1;
+      break;
+
+    case COMMAND_WAITALL:
+
+      for(int t=state->tail;t!=s; t=PERI_PLUS(t) ){
+	if ( state->Descrs[t].command != COMMAND_SENDRECV ) {
+	  MPI_Wait((MPI_Request *)&state->Descrs[t].request,MPI_STATUS_IGNORE);
+	}
+      };
+      s=PERI_PLUS(s);
+      state->start = s;
+      state->tail  = s;
+
+      WakeUpCompute();
+
+      return 1;
+      break;
+
+    default:
+      assert(0);
+      break;
+    }
+  }
+  return 0;
+}
+  //////////////////////////////////////////////////////////////////////////////
+  // External interaction with the queue
+  //////////////////////////////////////////////////////////////////////////////
+  
+void Slave::QueueSendRecv(void *xbuf, void *rbuf, int bytes, int xtag, int rtag, MPI_Comm comm,int dest,int src) 
+{
+  int head =state->head;
+  int next = PERI_PLUS(head);
+  
+  // Set up descriptor
+  int worldrank;
+  int hashtag;
+  MPI_Comm    communicator;
+  MPI_Request request;
+  uint64_t relative;
+  
+  relative = (uint64_t)xbuf - base;
+  state->Descrs[head].xbuf    = relative;
+  
+  relative= (uint64_t)rbuf - base;
+  state->Descrs[head].rbuf    = relative;
+  
+  state->Descrs[head].bytes  = bytes;
+  
+  MPIoffloadEngine::MapCommRankToWorldRank(hashtag,worldrank,xtag,comm,dest);
+  state->Descrs[head].dest   = MPIoffloadEngine::UniverseRanks[worldrank][vertical_rank];
+  state->Descrs[head].xtag    = hashtag;
+  
+  MPIoffloadEngine::MapCommRankToWorldRank(hashtag,worldrank,rtag,comm,src);
+  state->Descrs[head].src    = MPIoffloadEngine::UniverseRanks[worldrank][vertical_rank];
+  state->Descrs[head].rtag    = hashtag;
+  
+  state->Descrs[head].command= COMMAND_SENDRECV;
+  
+  // Block until FIFO has space
+  while( state->tail==next );
+  
+  // Msync on weak order architectures
+  
+  // Advance pointer
+  state->head = next;
+  
+};
+uint64_t Slave::QueueCommand(int command,void *buf, int bytes, int tag, MPI_Comm comm,int commrank) 
+{
+  /////////////////////////////////////////
+  // Spin; if FIFO is full until not full
+  /////////////////////////////////////////
+  int head =state->head;
+  int next = PERI_PLUS(head);
+    
+  // Set up descriptor
+  int worldrank;
+  int hashtag;
+  MPI_Comm    communicator;
+  MPI_Request request;
+  
+  MPIoffloadEngine::MapCommRankToWorldRank(hashtag,worldrank,tag,comm,commrank);
+
+  uint64_t relative= (uint64_t)buf - base;
+  state->Descrs[head].buf    = relative;
+  state->Descrs[head].bytes  = bytes;
+  state->Descrs[head].rank   = MPIoffloadEngine::UniverseRanks[worldrank][vertical_rank];
+  state->Descrs[head].tag    = hashtag;
+  state->Descrs[head].command= command;
+
+  /*  
+  if ( command == COMMAND_ISEND ) { 
+  std::cout << "QueueSend from "<< universe_rank <<" to commrank " << commrank 
+            << " to worldrank " << worldrank <<std::endl;
+  std::cout << " via VerticalRank "<< vertical_rank <<" to universerank " << MPIoffloadEngine::UniverseRanks[worldrank][vertical_rank]<<std::endl;
+  std::cout << " QueueCommand "<<buf<<"["<<bytes<<"]" << std::endl;
+  } 
+  if ( command == COMMAND_IRECV ) { 
+  std::cout << "QueueRecv on "<< universe_rank <<" from commrank " << commrank 
+            << " from worldrank " << worldrank <<std::endl;
+  std::cout << " via VerticalRank "<< vertical_rank <<" from universerank " << MPIoffloadEngine::UniverseRanks[worldrank][vertical_rank]<<std::endl;
+  std::cout << " QueueSend "<<buf<<"["<<bytes<<"]" << std::endl;
+  } 
+  */
+  // Block until FIFO has space
+  while( state->tail==next );
+
+  // Msync on weak order architectures
+  // Advance pointer
+  state->head = next;
+
+  return 0;
+}
+  
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+// Info that is setup once and indept of cartesian layout
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+MPI_Comm CartesianCommunicator::communicator_world;
+
+void CartesianCommunicator::Init(int *argc, char ***argv) 
+{
+  int flag;
+  MPI_Initialized(&flag); // needed to coexist with other libs apparently
+  if ( !flag ) {
+    MPI_Init(argc,argv);
+  }
+  communicator_world = MPI_COMM_WORLD;
+  MPI_Comm ShmComm;
+  MPIoffloadEngine::CommunicatorInit (communicator_world,ShmComm,ShmCommBuf);
+}
+void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
+{
+  int ierr=MPI_Cart_shift(communicator,dim,shift,&source,&dest);
+  assert(ierr==0);
+}
+int CartesianCommunicator::RankFromProcessorCoor(std::vector<int> &coor)
+{
+  int rank;
+  int ierr=MPI_Cart_rank  (communicator, &coor[0], &rank);
+  assert(ierr==0);
+  return rank;
+}
+void  CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor)
+{
+  coor.resize(_ndimension);
+  int ierr=MPI_Cart_coords  (communicator, rank, _ndimension,&coor[0]);
+  assert(ierr==0);
+}
+
+CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
+{ 
+  _ndimension = processors.size();
+  std::vector<int> periodic(_ndimension,1);
+
+  _Nprocessors=1;
+  _processors = processors;
+
+  for(int i=0;i<_ndimension;i++){
+    _Nprocessors*=_processors[i];
+  }
+
+  int Size; 
+  MPI_Comm_size(communicator_world,&Size);
+  assert(Size==_Nprocessors);
+
+  _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]);
+};
+
+void CartesianCommunicator::GlobalSum(uint32_t &u){
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_SUM,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalSum(uint64_t &u){
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_SUM,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalSum(float &f){
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalSumVector(float *f,int N)
+{
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,f,N,MPI_FLOAT,MPI_SUM,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalSum(double &d)
+{
+  int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_SUM,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalSumVector(double *d,int N)
+{
+  int ierr = MPI_Allreduce(MPI_IN_PLACE,d,N,MPI_DOUBLE,MPI_SUM,communicator);
+  assert(ierr==0);
+}
+
+// Basic Halo comms primitive
+void CartesianCommunicator::SendToRecvFrom(void *xmit,
+					   int dest,
+					   void *recv,
+					   int from,
+					   int bytes)
+{
+  std::vector<CommsRequest_t> reqs(0);
+  SendToRecvFromBegin(reqs,xmit,dest,recv,from,bytes);
+  SendToRecvFromComplete(reqs);
+}
+
+void CartesianCommunicator::SendRecvPacket(void *xmit,
+					   void *recv,
+					   int sender,
+					   int receiver,
+					   int bytes)
+{
+  MPI_Status stat;
+  assert(sender != receiver);
+  int tag = sender;
+  if ( _processor == sender ) {
+    MPI_Send(xmit, bytes, MPI_CHAR,receiver,tag,communicator);
+  }
+  if ( _processor == receiver ) { 
+    MPI_Recv(recv, bytes, MPI_CHAR,sender,tag,communicator,&stat);
+  }
+}
+
+// Basic Halo comms primitive
+void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+						void *xmit,
+						int dest,
+						void *recv,
+						int from,
+						int bytes)
+{
+  MPI_Request xrq;
+  MPI_Request rrq;
+  int rank = _processor;
+  int ierr;
+  ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
+  ierr|=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
+  
+  assert(ierr==0);
+
+  list.push_back(xrq);
+  list.push_back(rrq);
+}
+
+void CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+						       void *xmit,
+						       int dest,
+						       void *recv,
+						       int from,
+						       int bytes)
+{
+  uint64_t xmit_i = (uint64_t) xmit;
+  uint64_t recv_i = (uint64_t) recv;
+  uint64_t shm    = (uint64_t) ShmCommBuf;
+  // assert xmit and recv lie in shared memory region
+  assert( (xmit_i >= shm) && (xmit_i+bytes <= shm+MAX_MPI_SHM_BYTES) );
+  assert( (recv_i >= shm) && (recv_i+bytes <= shm+MAX_MPI_SHM_BYTES) );
+  assert(from!=_processor);
+  assert(dest!=_processor);
+
+  MPIoffloadEngine::QueueMultiplexedSendRecv(xmit,recv,bytes,_processor,from,communicator,dest,from);
+
+  //MPIoffloadEngine::QueueRoundRobinSendRecv(xmit,recv,bytes,_processor,from,communicator,dest,from);
+
+  //MPIoffloadEngine::QueueMultiplexedSend(xmit,bytes,_processor,communicator,dest);
+  //MPIoffloadEngine::QueueMultiplexedRecv(recv,bytes,from,communicator,from);
+}
+
+void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list)
+{
+  MPIoffloadEngine::WaitAll();
+  //this->Barrier();
+}
+
+void CartesianCommunicator::StencilBarrier(void) { }
+
+void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
+{
+  int nreq=list.size();
+  std::vector<MPI_Status> status(nreq);
+  int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
+  assert(ierr==0);
+}
+
+void CartesianCommunicator::Barrier(void)
+{
+  int ierr = MPI_Barrier(communicator);
+  assert(ierr==0);
+}
+
+void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
+{
+  int ierr=MPI_Bcast(data,
+		     bytes,
+		     MPI_BYTE,
+		     root,
+		     communicator);
+  assert(ierr==0);
+}
+
+void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
+{
+  int ierr= MPI_Bcast(data,
+		      bytes,
+		      MPI_BYTE,
+		      root,
+		      communicator_world);
+  assert(ierr==0);
+}
+
+void *CartesianCommunicator::ShmBufferSelf(void) { return ShmCommBuf; }
+
+void *CartesianCommunicator::ShmBuffer(int rank) {
+  return NULL;
+}
+void *CartesianCommunicator::ShmBufferTranslate(int rank,void * local_p) { 
+  return NULL;
+}
+
+
+};
+

lib/communicator/Communicator_none.cc

@@ -25,15 +25,19 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#include "Grid.h"
+#include <Grid/GridCore.h>
+
 namespace Grid {
 
+///////////////////////////////////////////////////////////////////////////////////////////////////
+// Info that is setup once and indept of cartesian layout
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
 void CartesianCommunicator::Init(int *argc, char *** arv)
 {
+  ShmInitGeneric();
 }
 
-int Rank(void ){ return 0; };
-
 CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
 {
   _processors = processors;
@@ -84,35 +88,23 @@ void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &lis
 {
   assert(0);
 }
+
 void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
 {
   assert(0);
 }
 
-void CartesianCommunicator::Barrier(void)
-{
-}
-
-void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
-{
-}
-void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
-{
-}
-
-
+int  CartesianCommunicator::RankWorld(void){return 0;}
+void CartesianCommunicator::Barrier(void){}
+void CartesianCommunicator::Broadcast(int root,void* data, int bytes) {}
+void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes) { }
+int  CartesianCommunicator::RankFromProcessorCoor(std::vector<int> &coor) {  return 0;}
+void CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor){  coor = _processor_coor; }
 void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
 {
   source =0;
   dest=0;
 }
-int CartesianCommunicator::RankFromProcessorCoor(std::vector<int> &coor)
-{
-  return 0;
-}
-void  CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor)
-{
-}
 
 
 }