Test script for Frontier on I/O

May finally clean up the poor MPI2 IO performance issue that has been persistent.

CLAUDE.md

@@ -0,0 +1,156 @@
+# CLAUDE.md
+
+This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
+
+## What This Is
+
+Grid is a data-parallel C++ library for lattice QCD. It provides SIMD-vectorised lattice containers, MPI-based domain decomposition, GPU acceleration (CUDA/HIP/SYCL), and a full suite of QCD algorithms including HMC.
+
+## Build
+
+Uses GNU Autotools. The bootstrap step only needs to run once (or after `configure.ac` changes).
+
+```bash
+./bootstrap.sh                    # downloads Eigen 3.4.0, generates configure
+mkdir build && cd build
+../configure [options]
+make -j$(nproc)
+make check                        # run root-level tests
+make install
+```
+
+Key configure options:
+
+| Option | Common values |
+|--------|---------------|
+| `--enable-simd=` | `AVX2`, `AVX512`, `KNL`, `A64FX`, `NEONv8`, `GPU` |
+| `--enable-comms=` | `mpi-auto`, `mpi3-auto`, `none` |
+| `--enable-accelerator=` | `cuda`, `hip`, `sycl` |
+| `--enable-shm=` | `shmopen`, `hugetlbfs`, `nvlink` |
+| `--enable-Nc=` | `3` (default), `2`, `4`, `5` |
+| `--with-gmp=`, `--with-mpfr=`, `--with-fftw=`, `--with-lime=` | paths to libs |
+| `--enable-hdf5`, `--enable-mkl`, `--enable-lapack` | optional features |
+
+Platform recipes from `README.md`:
+- **KNL**: `--enable-simd=KNL --enable-comms=mpi3-auto --enable-mkl`
+- **Skylake/Haswell**: `--enable-simd=AVX512` or `AVX2` + `--enable-comms=mpi3-auto`
+- **AMD EPYC**: `--enable-simd=AVX2 --enable-comms=mpi3`
+- **A64FX (Fugaku)**: `--enable-simd=A64FX --enable-comms=mpi3 --enable-shm=shmget` (see `SVE_README.txt`)
+
+Required external libs: GMP, MPFR, OpenSSL, zlib.
+
+### Use `systems/` for real machines
+
+`systems/<machine>/` holds the known-good build for each production platform (`Frontier`, `Aurora`, `Perlmutter`, `Summit`, `Tursa`, `Lumi`, `Booster`, `Crusher`, `SDCC-*`, `mac-arm`, …). Each contains a `config-command` (the exact `../../configure` invocation) and a `sourceme.sh` (module loads and env). **Prefer copying/adapting these over hand-rolling configure flags** — they encode compiler workarounds, `LDFLAGS`, and shared-memory settings that are easy to get wrong. `systems/WorkArounds.txt` records known vendor bugs.
+
+Note the GPU builds use `--enable-simd=GPU --enable-gen-simd-width=64`, so `Nsimd` is *not* 1 on device (it is `64/sizeof(scalar)`).
+
+### Regenerating `Make.inc` — required after adding or deleting source files
+
+`Make.inc` files are generated, not tracked in git (`.gitignore`d). `scripts/filelist` walks `Grid/`, `tests/*`, `benchmarks/`, `examples/`, and `HMC/` and writes the file lists and per-test `bin_PROGRAMS` rules. Every new `.cc`/`.h` in `Grid/`, and every new `Test_*.cc` / `Benchmark_*.cc` / `Example_*.cc`, is invisible to the build until you run:
+
+```bash
+./scripts/filelist    # from the source root, then re-run configure/make
+```
+
+`bootstrap.sh` runs it for you on the first setup.
+
+## Running Tests and Benchmarks
+
+```bash
+# From build directory
+make check                          # runs only Test_simd, Test_cshift, Test_stencil, Test_dwf_mixedcg_prec
+make -C tests/<subdir> tests        # build (not run) the tests in a subdirectory
+make bench                          # build benchmarks
+./tests/core/Test_simd              # run a single test binary directly
+mpirun -n 4 ./tests/core/Test_cshift --grid 16.16.16.16 --mpi 1.1.1.4
+```
+
+`make check` is a thin smoke test — building a subdirectory with `make -C tests/<subdir> tests` and running the relevant binaries directly is the normal development loop. Test binaries take Grid's standard command-line arguments (`--grid`, `--mpi`, `--accelerator-threads`, `--threads`, `--debug-signals`, `--log`); see `Grid/util/Init.cc`.
+
+Test subdirectories and their focus: `core` (SIMD, stencil, comms), `solver` (CG, GMRES, eigensolvers), `hmc` (MD integrators), `forces` (fermion forces), `lanczos`, `IO`, `smearing`, `sp2n`, `debug`.
+
+Tests and benchmarks that need optional fermion representations are guarded by `disable_tests_without_instantiations.h` / `disable_benchmarks_without_instantiations.h`, so a `--disable-fermion-reps --disable-gparity` build silently compiles them to no-ops.
+
+## Architecture
+
+### Layer stack (bottom to top)
+
+1. **SIMD layer** (`Grid/simd/`) — platform-specific intrinsics wrapped into `vRealF`, `vComplexD`, etc. The SIMD width and layout are compile-time constants controlled by `--enable-simd`.
+
+2. **Tensor layer** (`Grid/tensors/`) — Lorentz/colour/spin tensor algebra built on top of SIMD types. `iMatrix`, `iVector`, `iScalar` templates compose into QCD types like `ColourMatrix`, `SpinColourVector`.
+
+3. **Lattice layer** (`Grid/lattice/`) — `Lattice<T>` container: a site-local tensor replicated across a distributed Cartesian grid. All arithmetic is site-parallel and expression-template-fused.
+
+4. **Cartesian/comms layer** (`Grid/cartesian/`, `Grid/communicator/`) — `GridCartesian` holds the MPI topology and local/global geometry. `Grid/cshift/` implements nearest-neighbour halo exchange; `Grid/stencil/` is the optimised multi-hop stencil used by Dirac operators.
+
+5. **Algorithm layer** (`Grid/algorithms/`) — iterative solvers (CG, GMRES, BiCGSTAB, mixed-precision), eigensolvers (Lanczos, LAPACK), FFT, smearing.
+
+6. **QCD layer** (`Grid/qcd/`) — gauge and fermion actions, HMC integrators, observables.
+
+### QCD subsystem (`Grid/qcd/`)
+
+- `action/fermion/` — Wilson, Clover, DWF (Mobius), Staggered, twisted-mass, G-parity variants
+- `action/gauge/` — Wilson gauge, Symanzik, Iwasaki, DBW2, plaquette+rect
+- `representations/` — Fundamental, Adjoint, Two-index, Sp(2n)
+- `hmc/` — Leapfrog, OMF2/OMF4 integrators; pseudofermion refreshment; Metropolis accept/reject
+- `smearing/` — APE, Stout, HEX, gradient flow
+- `observables/` — Polyakov loop, plaquette, topological charge
+
+### GPU acceleration and the view/memory-manager discipline
+
+GPU support is injected via macros in `Grid/threads/Accelerator.h` — `accelerator_for(i, n, nsimd, {...})`, `accelerator_forNB` (non-blocking, must be followed by `accelerator_barrier()`), `accelerator_for2dNB`, and `accelerator_inline`. On a CPU build these degrade to `thread_for` (OpenMP). Unified virtual memory is on by default (`--enable-unified=yes`); device-aware MPI (`--enable-accelerator-aware-mpi`) avoids device→host copies on transfers.
+
+Lattice data is **not** directly addressable inside a kernel. You must open a view with the correct access mode so `Grid/allocator/MemoryManager.h` can move/mark the data:
+
+```cpp
+autoView(out_v, out, AcceleratorWriteDiscard);   // RAII; closes at end of scope
+autoView(in_v,  in,  AcceleratorRead);
+accelerator_for(ss, grid->oSites(), Nsimd, {
+    coalescedWrite(out_v[ss], coalescedRead(in_v[ss]));
+});
+```
+
+Modes are `AcceleratorRead/Write/WriteDiscard` and `CpuRead/Write/WriteDiscard`. Getting the mode wrong (e.g. `AcceleratorRead` on a field you write) produces stale-data bugs that only appear on GPU builds. Inside kernels use `coalescedRead`/`coalescedWrite` rather than raw `operator[]` — they map the SIMD lane onto `threadIdx.x` so accesses stay coalesced.
+
+### Repo-local debugging skills (`skills/`)
+
+`skills/` contains hard-won, Grid-specific playbooks written as invocable skill files. Consult them before debugging in these areas rather than reasoning from first principles:
+
+| File | Covers |
+|---|---|
+| `gpu-memory-performance.md` | `acceleratorThreads()`, LambdaApply thread mapping, `coalescedRead` idiom, fused vs staged HBM access |
+| `gpu-runtime-correctness.md` | GPU runtime returning early from sync, silent wrong answers |
+| `communication-overlap.md` | 7-phase halo pipeline, per-packet events, host-staging vs GPU-direct RDMA |
+| `mpi-heterogeneous.md` | `MPI_Sendrecv` device-buffer aliasing, deterministic reductions |
+| `compiler-validation.md` | Isolating GPU compiler codegen bugs, minimal reproducers |
+| `correctness-verification.md` | Double-run fingerprinting, per-packet checksums, flight recorder |
+| `hang-diagnosis.md` | Diagnosing MPI/accelerator hangs |
+
+### Memory and I/O
+
+- `Grid/allocator/` — aligned/NUMA-aware allocators; caching allocator via `--enable-alloc-cache`
+- `Grid/parallelIO/` — distributed parallel reader/writer for ILDG (via LIME), SciDAC, and native binary formats
+- `Grid/serialisation/` — text, binary, HDF5, XML/JSON serialisation of arbitrary Grid objects
+
+### Executables
+
+- `HMC/` — production HMC driver programmes (e.g. `Mobius2p1f.cc`, `DWF_plus_DSDR_nf2plus1_Shamir_Gparity.cc`)
+- `benchmarks/` — `Benchmark_dwf`, `Benchmark_ITT`, `Benchmark_comms`, `Benchmark_memory_bandwidth`, … used to qualify a new machine
+- `examples/` — small, readable programmes (`Example_plaquette.cc`, `Example_Mobius_spectrum.cc`) that are the best starting point for learning the API
+
+Each of these directories auto-builds every top-level `.cc` as its own binary via `scripts/filelist`.
+
+Every programme is wrapped in `Grid_init(&argc, &argv)` / `Grid_finalize()` (`Grid/util/Init.h`).
+
+## Key Conventions
+
+- **C++17** is required throughout.
+- Template structure: most classes are templated on `<_FImpl>` (fermion impl) or `<Gimpl>` (gauge impl), which encode the representation and precision. Instantiation is controlled by `--enable-fermion-instantiations`; the explicit instantiation `.cc` files live under `Grid/qcd/action/fermion/instantiation/<Impl>/` and are selected by `scripts/filelist`.
+- The `RealD`/`RealF`/`ComplexD`/`ComplexF` typedefs are used everywhere; avoid raw `double`/`float`.
+- Use `GRID_ASSERT(cond)` (defined in `Grid/GridStd.h`), not bare `assert` — it prints a Grid-formatted message and aborts cleanly under MPI.
+- Logging is stream-based, not macro-based: `std::cout << GridLogMessage << ... << std::endl;`. Channels declared in `Grid/log/Log.h` include `GridLogError`, `GridLogWarning`, `GridLogDebug`, `GridLogPerformance`, `GridLogIterative`, `GridLogSolver`, `GridLogHMC`, `GridLogComms`, `GridLogMemory`, `GridLogDslash`, `GridLogIRL`, `GridLogMG`. A subset is switched on at runtime with e.g. `--log Error,Warning,Message,Performance,Iterative,Integrator,Debug,Colours` (names given without the `GridLog` prefix).
+- Performance-critical paths use `GRID_TRACE(name)` from `Grid/perfmon/Tracing.h` (compiled out unless `--enable-tracing` selects a backend) and the `GridStopWatch` timers in `Grid/perfmon/Timer.h`.
+- Reductions across MPI ranks go through `GridBase::GlobalSum` / `GlobalMax`; never reduce with bare MPI calls inside library code.
+- Everything lives in `NAMESPACE_BEGIN(Grid)` / `NAMESPACE_END(Grid)` macros; follow the surrounding file rather than writing `namespace Grid { }`.
+- British spelling is used in identifiers and comments (`colour`, `neighbour`, `serialisation`).

Grid/Makefile.am

@@ -54,22 +54,24 @@ Version.h: version-cache
 include Make.inc
 include Eigen.inc
 
-extra_sources+=$(WILS_FERMION_FILES)
-extra_sources+=$(STAG_FERMION_FILES)
+if BUILD_FERMION_INSTANTIATIONS
+  extra_sources+=$(WILS_FERMION_FILES)
+  extra_sources+=$(STAG_FERMION_FILES)
 if BUILD_ZMOBIUS
-  extra_sources+=$(ZWILS_FERMION_FILES)
+    extra_sources+=$(ZWILS_FERMION_FILES)
 endif
 if BUILD_GPARITY
-  extra_sources+=$(GP_FERMION_FILES)
+    extra_sources+=$(GP_FERMION_FILES)
 endif
 if BUILD_FERMION_REPS
-  extra_sources+=$(ADJ_FERMION_FILES)
-  extra_sources+=$(TWOIND_FERMION_FILES)
+    extra_sources+=$(ADJ_FERMION_FILES)
+    extra_sources+=$(TWOIND_FERMION_FILES)
 endif
 if BUILD_SP
     extra_sources+=$(SP_FERMION_FILES)
 if BUILD_FERMION_REPS
-    extra_sources+=$(SP_TWOIND_FERMION_FILES)
+      extra_sources+=$(SP_TWOIND_FERMION_FILES)
+endif
 endif
 endif
 

Grid/algorithms/FFT.h

@@ -1,6 +1,6 @@
 /*************************************************************************************
 
-    Grid physics library, www.github.com/paboyle/Grid 
+    Grid physics library, www.github.com/paboyle/Grid
 
     Source file: ./lib/Cshift.h
 
@@ -65,17 +65,16 @@ public:
   typedef hipfftDoubleComplex FFTW_scalar;
   typedef hipfftHandle        FFTW_plan;
   static FFTW_plan fftw_plan_many_dft(int rank, int *n,int howmany,
-				      FFTW_scalar *in, int *inembed,		
-				      int istride, int idist,		
-				      FFTW_scalar *out, int *onembed,		
-				      int ostride, int odist,		
+				      FFTW_scalar *in, int *inembed,
+				      int istride, int idist,
+				      FFTW_scalar *out, int *onembed,
+				      int ostride, int odist,
 				      int sign, unsigned flags) {
     FFTW_plan p;
     auto rv = hipfftPlanMany(&p,rank,n,n,istride,idist,n,ostride,odist,HIPFFT_Z2Z,howmany);
     GRID_ASSERT(rv==HIPFFT_SUCCESS);
     return p;
-  }	  
-    
+  }
   inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out, int sign) {
     hipfftResult rv;
     if ( sign == forward ) rv =hipfftExecZ2Z(p,in,out,HIPFFT_FORWARD);
@@ -83,29 +82,25 @@ public:
     accelerator_barrier();
     GRID_ASSERT(rv==HIPFFT_SUCCESS);
   }
-  inline static void fftw_destroy_plan(const FFTW_plan p) {
-    hipfftDestroy(p);
-  }
+  inline static void fftw_destroy_plan(const FFTW_plan p) { hipfftDestroy(p); }
 };
 template<> struct FFTW<ComplexF> {
 public:
   static const int forward=FFTW_FORWARD;
   static const int backward=FFTW_BACKWARD;
-  typedef hipfftComplex      FFTW_scalar;
-  typedef hipfftHandle        FFTW_plan;
-
+  typedef hipfftComplex FFTW_scalar;
+  typedef hipfftHandle  FFTW_plan;
   static FFTW_plan fftw_plan_many_dft(int rank, int *n,int howmany,
-				      FFTW_scalar *in, int *inembed,		
-				      int istride, int idist,		
-				      FFTW_scalar *out, int *onembed,		
-				      int ostride, int odist,		
+				      FFTW_scalar *in, int *inembed,
+				      int istride, int idist,
+				      FFTW_scalar *out, int *onembed,
+				      int ostride, int odist,
 				      int sign, unsigned flags) {
     FFTW_plan p;
     auto rv = hipfftPlanMany(&p,rank,n,n,istride,idist,n,ostride,odist,HIPFFT_C2C,howmany);
     GRID_ASSERT(rv==HIPFFT_SUCCESS);
     return p;
-  }	  
-    
+  }
   inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out, int sign) {
     hipfftResult rv;
     if ( sign == forward ) rv =hipfftExecC2C(p,in,out,HIPFFT_FORWARD);
@@ -113,9 +108,7 @@ public:
     accelerator_barrier();
     GRID_ASSERT(rv==HIPFFT_SUCCESS);
   }
-  inline static void fftw_destroy_plan(const FFTW_plan p) {
-    hipfftDestroy(p);
-  }
+  inline static void fftw_destroy_plan(const FFTW_plan p) { hipfftDestroy(p); }
 };
 #endif
 
@@ -126,53 +119,45 @@ public:
   static const int backward=FFTW_BACKWARD;
   typedef cufftDoubleComplex FFTW_scalar;
   typedef cufftHandle        FFTW_plan;
-
   static FFTW_plan fftw_plan_many_dft(int rank, int *n,int howmany,
-				      FFTW_scalar *in, int *inembed,		
-				      int istride, int idist,		
-				      FFTW_scalar *out, int *onembed,		
-				      int ostride, int odist,		
+				      FFTW_scalar *in, int *inembed,
+				      int istride, int idist,
+				      FFTW_scalar *out, int *onembed,
+				      int ostride, int odist,
 				      int sign, unsigned flags) {
     FFTW_plan p;
     cufftPlanMany(&p,rank,n,n,istride,idist,n,ostride,odist,CUFFT_Z2Z,howmany);
     return p;
-  }	  
-    
+  }
   inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out, int sign) {
     if ( sign == forward ) cufftExecZ2Z(p,in,out,CUFFT_FORWARD);
     else                   cufftExecZ2Z(p,in,out,CUFFT_INVERSE);
     accelerator_barrier();
   }
-  inline static void fftw_destroy_plan(const FFTW_plan p) {
-    cufftDestroy(p);
-  }
+  inline static void fftw_destroy_plan(const FFTW_plan p) { cufftDestroy(p); }
 };
 template<> struct FFTW<ComplexF> {
 public:
   static const int forward=FFTW_FORWARD;
   static const int backward=FFTW_BACKWARD;
   typedef cufftComplex FFTW_scalar;
-  typedef cufftHandle        FFTW_plan;
-
+  typedef cufftHandle  FFTW_plan;
   static FFTW_plan fftw_plan_many_dft(int rank, int *n,int howmany,
-				      FFTW_scalar *in, int *inembed,		
-				      int istride, int idist,		
-				      FFTW_scalar *out, int *onembed,		
-				      int ostride, int odist,		
+				      FFTW_scalar *in, int *inembed,
+				      int istride, int idist,
+				      FFTW_scalar *out, int *onembed,
+				      int ostride, int odist,
 				      int sign, unsigned flags) {
     FFTW_plan p;
     cufftPlanMany(&p,rank,n,n,istride,idist,n,ostride,odist,CUFFT_C2C,howmany);
     return p;
-  }	  
-    
+  }
   inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out, int sign) {
     if ( sign == forward ) cufftExecC2C(p,in,out,CUFFT_FORWARD);
     else                   cufftExecC2C(p,in,out,CUFFT_INVERSE);
     accelerator_barrier();
   }
-  inline static void fftw_destroy_plan(const FFTW_plan p) {
-    cufftDestroy(p);
-  }
+  inline static void fftw_destroy_plan(const FFTW_plan p) { cufftDestroy(p); }
 };
 #endif
 
@@ -183,313 +168,325 @@ public:
   typedef fftw_complex FFTW_scalar;
   typedef fftw_plan    FFTW_plan;
   static FFTW_plan fftw_plan_many_dft(int rank, int *n,int howmany,
-				      FFTW_scalar *in, int *inembed,		
-				      int istride, int idist,		
-				      FFTW_scalar *out, int *onembed,		
-				      int ostride, int odist,		
+				      FFTW_scalar *in, int *inembed,
+				      int istride, int idist,
+				      FFTW_scalar *out, int *onembed,
+				      int ostride, int odist,
 				      int sign, unsigned flags) {
     return ::fftw_plan_many_dft(rank,n,howmany,in,inembed,istride,idist,out,onembed,ostride,odist,sign,flags);
-  }	  
-    
+  }
   inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out, int sign) {
     ::fftw_execute_dft(p,in,out);
   }
-  inline static void fftw_destroy_plan(const FFTW_plan p) {
-    ::fftw_destroy_plan(p);
-  }
+  inline static void fftw_destroy_plan(const FFTW_plan p) { ::fftw_destroy_plan(p); }
 };
 template<> struct FFTW<ComplexF> {
 public:
   typedef fftwf_complex FFTW_scalar;
   typedef fftwf_plan    FFTW_plan;
   static FFTW_plan fftw_plan_many_dft(int rank, int *n,int howmany,
-				      FFTW_scalar *in, int *inembed,		
-				      int istride, int idist,		
-				      FFTW_scalar *out, int *onembed,		
-				      int ostride, int odist,		
+				      FFTW_scalar *in, int *inembed,
+				      int istride, int idist,
+				      FFTW_scalar *out, int *onembed,
+				      int ostride, int odist,
 				      int sign, unsigned flags) {
     return ::fftwf_plan_many_dft(rank,n,howmany,in,inembed,istride,idist,out,onembed,ostride,odist,sign,flags);
-  }	  
-    
+  }
   inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out, int sign) {
     ::fftwf_execute_dft(p,in,out);
   }
-  inline static void fftw_destroy_plan(const FFTW_plan p) {
-    ::fftwf_destroy_plan(p);
-  }
+  inline static void fftw_destroy_plan(const FFTW_plan p) { ::fftwf_destroy_plan(p); }
 };
 #endif
 #endif
 
-class FFT {
-private:
-    
-  double flops;
-  double flops_call;
-  uint64_t usec;
-    
-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;}    
+struct FFTbase {
+  double        flops;
+  double        flops_call;
+  uint64_t      usec;
+  GridCartesian *_grid;
 
-  FFT ( GridCartesian * grid ) 
-  {
-    flops=0;
-    usec =0;
-  };
-    
-  ~FFT ( void)  {
-    //    delete sgrid;
+  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; }
+
+  FFTbase(GridCartesian *grid) : _grid(grid), flops(0), flops_call(0), usec(0) {}
+};
+
+// Barrel-shift gather, FFT execute, and insert. Called by both FFT and PlannedFFT.
+// The caller is responsible for plan acquisition and destruction.
+template<class vobj>
+static void FFT_dim_execute(
+    Lattice<vobj>       &result,
+    const Lattice<vobj> &source,
+    int dim, int sign,
+    typename FFTW<typename vobj::scalar_type>::FFTW_plan p,
+    GridCartesian *grid,
+    double &flops, double &flops_call, uint64_t &usec)
+{
+  typedef typename vobj::scalar_type   scalar;
+  typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_type   scalar_type;
+  typedef typename vobj::vector_type   vector_type;
+  typedef typename FFTW<scalar>::FFTW_scalar FFTW_scalar;
+
+  const int Ndim = grid->Nd();
+  int L = grid->_ldimensions[dim];
+  int G = grid->_fdimensions[dim];
+  int Ncomp = sizeof(sobj) / sizeof(scalar);
+  int64_t Nlow = 1, Nhigh = 1;
+  for (int d = 0; d < dim;      d++) Nlow  *= grid->_ldimensions[d];
+  for (int d = dim+1; d < Ndim; d++) Nhigh *= grid->_ldimensions[d];
+  int64_t Nperp = Nlow * Nhigh;
+
+  deviceVector<scalar> pgbuf(Nperp * Ncomp * G);
+  scalar *pgbuf_v = &pgbuf[0];
+  int howmany = Ncomp * Nperp;
+
+  scalar div;
+  if      (sign == FFTW_BACKWARD) div = 1.0 / G;
+  else if (sign == FFTW_FORWARD)  div = 1.0;
+  else GRID_ASSERT(0);
+
+  double t_pencil = 0, t_fft = 0, t_copy = 0, t_shift = 0;
+  double t_total  = -usecond();
+
+  result = source;
+  int pc = grid->_processor_coor[dim];
+
+  const Coordinate ldims      = grid->_ldimensions;
+  const Coordinate rdims      = grid->_rdimensions;
+  const Coordinate sdims      = grid->_simd_layout;
+  const Coordinate processors = grid->_processors;
+
+  Coordinate pgdims(Ndim);
+  pgdims[0] = G;
+  for (int d = 0, dd = 1; d < Ndim; d++)
+    if (d != dim) pgdims[dd++] = ldims[d];
+  int64_t pgvol = 1;
+  for (int d = 0; d < Ndim; d++) pgvol *= pgdims[d];
+
+  const int Nsimd = vobj::Nsimd();
+  t_pencil = -usecond();
+  for (int p_idx = 0; p_idx < processors[dim]; p_idx++) {
+    t_copy -= usecond();
+    autoView(r_v, result, AcceleratorRead);
+    accelerator_for(idx, grid->oSites(), vobj::Nsimd(), {
+#ifdef GRID_SIMT
+    {
+      int lane = acceleratorSIMTlane(Nsimd);
+#else
+    for (int lane = 0; lane < Nsimd; lane++) {
+#endif
+      Coordinate icoor, ocoor, pgcoor;
+      Lexicographic::CoorFromIndex(icoor, lane, sdims);
+      Lexicographic::CoorFromIndex(ocoor, idx,  rdims);
+      pgcoor[0] = ocoor[dim] + icoor[dim]*rdims[dim] + ((pc+p_idx)%processors[dim])*L;
+      for (int d = 0, dd = 1; d < Ndim; d++)
+        if (d != dim) { pgcoor[dd] = ocoor[d] + icoor[d]*rdims[d]; dd++; }
+      int64_t pgidx;
+      Lexicographic::IndexFromCoor(pgcoor, pgidx, pgdims);
+      vector_type *from = (vector_type *)&r_v[idx];
+      scalar_type  stmp;
+      for (int w = 0; w < Ncomp; w++) {
+        stmp = getlane(from[w], lane);
+        pgbuf_v[pgidx + w*pgvol] = stmp;
+      }
+#ifdef GRID_SIMT
+    }
+#else
+    }
+#endif
+    });
+    t_copy += usecond();
+    if (p_idx != processors[dim] - 1) {
+      Lattice<vobj> temp(grid);
+      t_shift -= usecond();
+      temp = Cshift(result, dim, L); result = temp;
+      t_shift += usecond();
+    }
   }
-    
-  template<class vobj>
-  void FFT_dim_mask(Lattice<vobj> &result,const Lattice<vobj> &source,Coordinate mask,int sign){
+  t_pencil += usecond();
 
-    //    vgrid=result.Grid();
-    //    conformable(result.Grid(),vgrid);
-    //    conformable(source.Grid(),vgrid);
-    const int Ndim = source.Grid()->Nd();
+  FFTW_scalar *in  = (FFTW_scalar *)pgbuf_v;
+  FFTW_scalar *out = (FFTW_scalar *)pgbuf_v;
+  t_fft = -usecond();
+  FFTW<scalar>::fftw_execute_dft(p, in, out, sign);
+  t_fft += usecond();
+
+  flops_call = 5.0 * howmany * G * log2(G);
+  usec  = t_fft;
+  flops = flops_call;
+
+  result = Zero();
+  double t_insert = -usecond();
+  {
+    autoView(r_v, result, AcceleratorWrite);
+    accelerator_for(idx, grid->oSites(), Nsimd, {
+#ifdef GRID_SIMT
+    {
+      int lane = acceleratorSIMTlane(Nsimd);
+#else
+    for (int lane = 0; lane < Nsimd; lane++) {
+#endif
+      Coordinate icoor(Ndim), ocoor(Ndim), pgcoor(Ndim);
+      Lexicographic::CoorFromIndex(icoor, lane, sdims);
+      Lexicographic::CoorFromIndex(ocoor, idx,  rdims);
+      pgcoor[0] = ocoor[dim] + icoor[dim]*rdims[dim] + pc*L;
+      for (int d = 0, dd = 1; d < Ndim; d++)
+        if (d != dim) { pgcoor[dd] = ocoor[d] + icoor[d]*rdims[d]; dd++; }
+      int64_t pgidx;
+      Lexicographic::IndexFromCoor(pgcoor, pgidx, pgdims);
+      vector_type *to  = (vector_type *)&r_v[idx];
+      scalar_type  stmp;
+      for (int w = 0; w < Ncomp; w++) {
+        stmp = pgbuf_v[pgidx + w*pgvol];
+        putlane(to[w], stmp, lane);
+      }
+#ifdef GRID_SIMT
+    }
+#else
+    }
+#endif
+    });
+  }
+  result = result * div;
+  t_insert += usecond();
+  t_total  += usecond();
+
+  std::cout << GridLogPerformance << " FFT took   " << t_total/1.0e6  << " s" << std::endl;
+  std::cout << GridLogPerformance << " FFT pencil " << t_pencil/1.0e6 << " s" << std::endl;
+  std::cout << GridLogPerformance << "  of which copy " << t_copy/1.0e6  << " s" << std::endl;
+  std::cout << GridLogPerformance << "  of which shift" << t_shift/1.0e6 << " s" << std::endl;
+  std::cout << GridLogPerformance << " FFT kernels " << t_fft/1.0e6    << " s" << std::endl;
+  std::cout << GridLogPerformance << " FFT insert  " << t_insert/1.0e6 << " s" << std::endl;
+}
+
+class FFT : public FFTbase {
+public:
+  FFT(GridCartesian *grid) : FFTbase(grid) {}
+  ~FFT() {}
+
+  template<class vobj>
+  void FFT_dim_mask(Lattice<vobj> &result, const Lattice<vobj> &source, Coordinate mask, int sign) {
+    const int Ndim = _grid->Nd();
     Lattice<vobj> tmp = source;
-    for(int d=0;d<Ndim;d++){
-      if( mask[d] ) {
-	FFT_dim(result,tmp,d,sign);
-	tmp=result;
+    for (int d = 0; d < Ndim; 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){
-    const int Ndim = source.Grid()->Nd();
-    Coordinate mask(Ndim,1);
-    FFT_dim_mask(result,source,mask,sign);
+  void FFT_all_dim(Lattice<vobj> &result, const Lattice<vobj> &source, int sign) {
+    Coordinate mask(_grid->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){
-    const int Ndim = source.Grid()->Nd();
-    GridBase *grid = source.Grid();
-    conformable(result.Grid(),source.Grid());
+  void FFT_dim(Lattice<vobj> &result, const Lattice<vobj> &source, int dim, int sign) {
+    GRID_ASSERT(source.Grid() == _grid);
+    GRID_ASSERT(result.Grid() == _grid);
+    conformable(result.Grid(), source.Grid());
 
-    int L = grid->_ldimensions[dim];
-    int G = grid->_fdimensions[dim];
-      
-    Coordinate layout(Ndim,1);
-    
-    // Construct pencils
-    typedef typename vobj::scalar_object sobj;
     typedef typename vobj::scalar_type   scalar;
-    typedef typename vobj::scalar_type   scalar_type;
-    typedef typename vobj::vector_type   vector_type;
-      
-    //std::cout << "CPU view" << std::endl;
-    
+    typedef typename vobj::scalar_object sobj;
     typedef typename FFTW<scalar>::FFTW_scalar FFTW_scalar;
     typedef typename FFTW<scalar>::FFTW_plan   FFTW_plan;
-      
-    int Ncomp = sizeof(sobj)/sizeof(scalar);
-    int64_t Nlow  = 1;
-    int64_t Nhigh = 1;
 
-    for(int d=0;d<dim;d++){
-      Nlow*=grid->_ldimensions[d];
-    }
-    for(int d=dim+1;d<Ndim;d++){
-      Nhigh*=grid->_ldimensions[d];
-    }
-    int64_t Nperp=Nlow*Nhigh;
-    
-    deviceVector<scalar> pgbuf; // Layout is [perp][component][dim]
-    pgbuf.resize(Nperp*Ncomp*G);
-    scalar *pgbuf_v = &pgbuf[0];
-      
-    int rank = 1;  /* 1d transforms */
-    int n[] = {G}; /* 1d transforms of length G */
+    const int Ndim = _grid->Nd();
+    int G     = _grid->_fdimensions[dim];
+    int Ncomp = sizeof(sobj) / sizeof(scalar);
+    int64_t Nperp = 1;
+    for (int d = 0; d < Ndim; d++)
+      if (d != dim) Nperp *= _grid->_ldimensions[d];
+    int n[]     = {G};
     int howmany = Ncomp * Nperp;
-    int odist,idist,istride,ostride;
-    idist   = odist   = G;            /* Distance between consecutive FT's */
-    istride = ostride = 1;            /* 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 GRID_ASSERT(0);
 
-    double t_pencil=0;
-    double t_fft   =0;
-    double t_total =-usecond();
-    //    std::cout << GridLogPerformance<<"Making FFTW plan" << std::endl;
-    /*
-     *
-     */
-    FFTW_plan p;
-    {
-      FFTW_scalar *in = (FFTW_scalar *)&pgbuf_v[0];
-      FFTW_scalar *out= (FFTW_scalar *)&pgbuf_v[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::cout << GridLogPerformance<<"Making pencil" << std::endl;
-    Coordinate lcoor(Ndim), gcoor(Ndim);
-    double t_copy=0;
-    double t_shift=0;
-    t_pencil = -usecond();
-    result = source;
-    int pc = grid->_processor_coor[dim];
-
-    const Coordinate ldims = grid->_ldimensions;
-    const Coordinate rdims = grid->_rdimensions;
-    const Coordinate sdims = grid->_simd_layout;
-
-    Coordinate processors = grid->_processors;
-    Coordinate pgdims(Ndim);
-    pgdims[0] = G;
-    for(int d=0, dd=1;d<Ndim;d++){
-      if ( d!=dim ) pgdims[dd++] = ldims[d];
-    }
-    int64_t pgvol=1;
-    for(int d=0;d<Ndim;d++) pgvol*=pgdims[d];
-    
-    const int Nsimd = vobj::Nsimd();
-    for(int p=0;p<processors[dim];p++) {
-      t_copy-=usecond();
-      autoView(r_v,result,AcceleratorRead);
-      accelerator_for(idx, grid->oSites(), vobj::Nsimd(), {
-#ifdef GRID_SIMT
-      {
-	int lane=acceleratorSIMTlane(Nsimd); // buffer lane
-#else
-      for(int lane=0;lane<Nsimd;lane++) {
-#endif
-	Coordinate icoor;
-	Coordinate ocoor;
-	Coordinate pgcoor;
-
-	Lexicographic::CoorFromIndex(icoor,lane,sdims);
-	Lexicographic::CoorFromIndex(ocoor,idx,rdims);
-
-	pgcoor[0] = ocoor[dim] + icoor[dim]*rdims[dim] + ((pc+p)%processors[dim])*L;
-	for(int d=0,dd=1;d<Ndim;d++){
-	  if ( d!=dim ) {
-	    pgcoor[dd] = ocoor[d] + icoor[d]*rdims[d];
-	    dd++;
-	  }
-	}
-
-	// Map coordinates in lattice layout to FFTW index
-	int64_t pgidx;
-	Lexicographic::IndexFromCoor(pgcoor,pgidx,pgdims);
-
-	vector_type *from = (vector_type *)&r_v[idx];
-	scalar_type stmp;
-	for(int w=0;w<Ncomp;w++){
-	  int64_t pg_idx = pgidx + w*pgvol;
-	  stmp = getlane(from[w], lane);
-	  pgbuf_v[pg_idx] = stmp;
-	}
-#ifdef GRID_SIMT
-      }
-#else
-      }
-#endif
-      });
-
-      t_copy+=usecond();
-      if (p != processors[dim] - 1) {
-	Lattice<vobj> temp(grid);
-	t_shift-=usecond();
-	temp = Cshift(result,dim,L); result = temp;
-	t_shift+=usecond();
-      }
-    }
-    t_pencil += usecond();
-      
-    FFTW_scalar *in = (FFTW_scalar *)pgbuf_v;
-    FFTW_scalar *out= (FFTW_scalar *)pgbuf_v;
-    t_fft = -usecond();
-    FFTW<scalar>::fftw_execute_dft(p,in,out,sign);
-    t_fft += usecond();
-    
-    // performance counting
-    flops_call = 5.0*howmany*G*log2(G);
-    usec = t_fft;
-    flops= flops_call;
-
-    result = Zero();
-    
-    double t_insert = -usecond();
-    {
-      autoView(r_v,result,AcceleratorWrite);
-      accelerator_for(idx,grid->oSites(),Nsimd,{
-#ifdef GRID_SIMT
-      {
-	int lane=acceleratorSIMTlane(Nsimd); // buffer lane
-#else
-      for(int lane=0;lane<Nsimd;lane++) {
-#endif
-	Coordinate icoor(Ndim);
-	Coordinate ocoor(Ndim);
-	Coordinate pgcoor(Ndim);
-
-	Lexicographic::CoorFromIndex(icoor,lane,sdims);
-	Lexicographic::CoorFromIndex(ocoor,idx,rdims);
-
-	pgcoor[0] = ocoor[dim] + icoor[dim]*rdims[dim] + pc*L;
-	for(int d=0,dd=1;d<Ndim;d++){
-	  if ( d!=dim ) {
-	    pgcoor[dd] = ocoor[d] + icoor[d]*rdims[d];
-	    dd++;
-	  }
-	}
-	// Map coordinates in lattice layout to FFTW index
-	int64_t pgidx;
-	Lexicographic::IndexFromCoor(pgcoor,pgidx,pgdims);
-
-	vector_type *to = (vector_type *)&r_v[idx];
-	scalar_type stmp;
-	for(int w=0;w<Ncomp;w++){
-	  int64_t pg_idx = pgidx + w*pgvol;
-	  stmp = pgbuf_v[pg_idx];
-	  putlane(to[w], stmp, lane);
-	}
-	
-#ifdef GRID_SIMT
-      }
-#else
-      }
-#endif
-      });
-    }
-
-    result = result*div;
-
-    t_insert +=usecond();
-    
-    // destroying plan
+    deviceVector<scalar> dummy(2);
+    FFTW_scalar *buf = (FFTW_scalar *)&dummy[0];
+    FFTW_plan p = FFTW<scalar>::fftw_plan_many_dft(1, n, howmany,
+                                                    buf, n, 1, G,
+                                                    buf, n, 1, G,
+                                                    sign, FFTW_ESTIMATE);
+    FFT_dim_execute(result, source, dim, sign, p, _grid, flops, flops_call, usec);
     FFTW<scalar>::fftw_destroy_plan(p);
+  }
+};
 
-    t_total +=usecond();
+template<class vobj>
+class PlannedFFT : public FFTbase {
+private:
+  typedef typename vobj::scalar_type   scalar;
+  typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::vector_type   vector_type;
+  typedef typename FFTW<scalar>::FFTW_scalar FFTW_scalar;
+  typedef typename FFTW<scalar>::FFTW_plan   FFTW_plan;
 
-    std::cout <<GridLogPerformance<< " FFT took   "<<t_total/1.0e6 <<" s" << std::endl;
-    std::cout <<GridLogPerformance<< " FFT pencil "<<t_pencil/1.0e6 <<" s" << std::endl;
-    std::cout <<GridLogPerformance<< "  of which copy "<<t_copy/1.0e6 <<" s" << std::endl;
-    std::cout <<GridLogPerformance<< "  of which shift"<<t_shift/1.0e6 <<" s" << std::endl;
-    std::cout <<GridLogPerformance<< " FFT kernels "<<t_fft/1.0e6 <<" s" << std::endl;
-    std::cout <<GridLogPerformance<< " FFT insert  "<<t_insert/1.0e6 <<" s" << std::endl;
-    
+  std::vector<FFTW_plan> forward_plans;
+  std::vector<FFTW_plan> backward_plans;
+
+  void PlanCreate() {
+    const int Ndim = _grid->Nd();
+    forward_plans.resize(Ndim);
+    backward_plans.resize(Ndim);
+
+    for (int d = 0; d < Ndim; d++) {
+      int G       = _grid->_fdimensions[d];
+      int Ncomp   = sizeof(sobj) / sizeof(scalar);
+      int64_t Nperp = 1;
+      for (int dd = 0; dd < Ndim; dd++)
+        if (dd != d) Nperp *= _grid->_ldimensions[dd];
+      int howmany = Ncomp * (int)Nperp;
+      int n[]     = {G};
+
+      deviceVector<scalar> dummy(2);
+      FFTW_scalar *buf = (FFTW_scalar *)&dummy[0];
+
+      forward_plans[d]  = FFTW<scalar>::fftw_plan_many_dft(1, n, howmany, buf, n, 1, G, buf, n, 1, G, FFTW_FORWARD,  FFTW_ESTIMATE);
+      backward_plans[d] = FFTW<scalar>::fftw_plan_many_dft(1, n, howmany, buf, n, 1, G, buf, n, 1, G, FFTW_BACKWARD, FFTW_ESTIMATE);
+    }
+  }
+
+  void PlanDestroy() {
+    for (auto p : forward_plans)  FFTW<scalar>::fftw_destroy_plan(p);
+    for (auto p : backward_plans) FFTW<scalar>::fftw_destroy_plan(p);
+    forward_plans.clear();
+    backward_plans.clear();
+  }
+
+public:
+  PlannedFFT(GridCartesian *grid) : FFTbase(grid) { PlanCreate(); }
+  ~PlannedFFT() { PlanDestroy(); }
+
+  void FFT_dim_mask(Lattice<vobj> &result, const Lattice<vobj> &source, Coordinate mask, int sign) {
+    const int Ndim = _grid->Nd();
+    Lattice<vobj> tmp = source;
+    for (int d = 0; d < Ndim; d++) {
+      if (mask[d]) {
+        FFT_dim(result, tmp, d, sign);
+        tmp = result;
+      }
+    }
+  }
+
+  void FFT_all_dim(Lattice<vobj> &result, const Lattice<vobj> &source, int sign) {
+    Coordinate mask(_grid->Nd(), 1);
+    FFT_dim_mask(result, source, mask, sign);
+  }
+
+  void FFT_dim(Lattice<vobj> &result, const Lattice<vobj> &source, int dim, int sign) {
+    GRID_ASSERT(source.Grid() == _grid);
+    GRID_ASSERT(result.Grid() == _grid);
+    GRID_ASSERT((int)forward_plans.size() == _grid->Nd());
+    conformable(result.Grid(), source.Grid());
+    FFTW_plan p = (sign == forward ? forward_plans : backward_plans)[dim];
+    FFT_dim_execute(result, source, dim, sign, p, _grid, flops, flops_call, usec);
   }
 };
 

Grid/algorithms/blas/BatchedBlas.h

@@ -28,6 +28,7 @@ Author: Peter Boyle <pboyle@bnl.gov>
 #pragma once
 
 #ifdef GRID_HIP
+#include <hip/hip_version.h>
 #include <hipblas/hipblas.h>
 #endif
 #ifdef GRID_CUDA
@@ -255,17 +256,30 @@ public:
     if ( OpB == GridBLAS_OP_N ) hOpB = HIPBLAS_OP_N;
     if ( OpB == GridBLAS_OP_T ) hOpB = HIPBLAS_OP_T;
     if ( OpB == GridBLAS_OP_C ) hOpB = HIPBLAS_OP_C;
+#if defined(HIP_VERSION_MAJOR) && (HIP_VERSION_MAJOR >=7)
     auto err = hipblasZgemmBatched(gridblasHandle,
 				   hOpA,
 				   hOpB,
 				   m,n,k,
-				   (hipblasDoubleComplex *) &alpha_p[0],
-				   (hipblasDoubleComplex **)&Amk[0], lda,
-				   (hipblasDoubleComplex **)&Bkn[0], ldb,
-				   (hipblasDoubleComplex *) &beta_p[0],
-				   (hipblasDoubleComplex **)&Cmn[0], ldc,
+				   (hipDoubleComplex *) &alpha_p[0],
+				   (hipDoubleComplex **)&Amk[0], lda,
+				   (hipDoubleComplex **)&Bkn[0], ldb,
+				   (hipDoubleComplex *) &beta_p[0],
+				   (hipDoubleComplex **)&Cmn[0], ldc,
 				   batchCount);
-    //	 std::cout << " hipblas return code " <<(int)err<<std::endl;
+#else
+    auto err = hipblasZgemmBatched(gridblasHandle,
+                                   hOpA,
+                                   hOpB,
+                                   m,n,k,
+                                   (hipblasDoubleComplex *) &alpha_p[0],
+                                   (hipblasDoubleComplex **)&Amk[0], lda,
+                                   (hipblasDoubleComplex **)&Bkn[0], ldb,
+                                   (hipblasDoubleComplex *) &beta_p[0],
+                                   (hipblasDoubleComplex **)&Cmn[0], ldc,
+                                   batchCount);
+#endif
+    //    std::cout << " hipblas return code " <<(int)err<<" "<<__LINE__<<std::endl;
     GRID_ASSERT(err==HIPBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_CUDA
@@ -503,17 +517,31 @@ public:
     if ( OpB == GridBLAS_OP_N ) hOpB = HIPBLAS_OP_N;
     if ( OpB == GridBLAS_OP_T ) hOpB = HIPBLAS_OP_T;
     if ( OpB == GridBLAS_OP_C ) hOpB = HIPBLAS_OP_C;
+#if defined(HIP_VERSION_MAJOR) && (HIP_VERSION_MAJOR >=7)
     auto err = hipblasCgemmBatched(gridblasHandle,
 				   hOpA,
 				   hOpB,
 				   m,n,k,
-				   (hipblasComplex *) &alpha_p[0],
-				   (hipblasComplex **)&Amk[0], lda,
-				   (hipblasComplex **)&Bkn[0], ldb,
-				   (hipblasComplex *) &beta_p[0],
-				   (hipblasComplex **)&Cmn[0], ldc,
+				   (hipComplex *) &alpha_p[0],
+				   (hipComplex **)&Amk[0], lda,
+				   (hipComplex **)&Bkn[0], ldb,
+				   (hipComplex *) &beta_p[0],
+				   (hipComplex **)&Cmn[0], ldc,
 				   batchCount);
+#else
+    auto err = hipblasCgemmBatched(gridblasHandle,
+                                   hOpA,
+                                   hOpB,
+                                   m,n,k,
+                                   (hipblasComplex *) &alpha_p[0],
+                                   (hipblasComplex **)&Amk[0], lda,
+                                   (hipblasComplex **)&Bkn[0], ldb,
+                                   (hipblasComplex *) &beta_p[0],
+                                   (hipblasComplex **)&Cmn[0], ldc,
+                                   batchCount);
 
+#endif
+    //    std::cout << " hipblas return code " <<(int)err<<" "<<__LINE__<<std::endl;
     GRID_ASSERT(err==HIPBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_CUDA
@@ -550,6 +578,7 @@ public:
 				(void **)&Cmn[0], CUDA_C_32F, ldc,
 				batchCount, compute_precision, CUBLAS_GEMM_DEFAULT);
     }
+    //    std::cout << " hipblas return code " <<(int)err<<" "<<__LINE__<<std::endl;
     GRID_ASSERT(err==CUBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_SYCL
@@ -720,6 +749,7 @@ public:
 				   (float *) &beta_p[0],
 				   (float **)&Cmn[0], ldc,
 				   batchCount);
+    //    std::cout << " hipblas return code " <<(int)err<<" "<<__LINE__<<std::endl;
     GRID_ASSERT(err==HIPBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_CUDA
@@ -880,6 +910,7 @@ public:
 				   (double *) &beta_p[0],
 				   (double **)&Cmn[0], ldc,
 				   batchCount);
+    //    std::cout << " hipblas return code " <<(int)err<<" "<<__LINE__<<std::endl;
     GRID_ASSERT(err==HIPBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_CUDA
@@ -1094,11 +1125,20 @@ public:
     GRID_ASSERT(info.size()==batchCount);
 
 #ifdef GRID_HIP
+#if defined(HIP_VERSION_MAJOR) && (HIP_VERSION_MAJOR >=7)
     auto err = hipblasZgetrfBatched(gridblasHandle,(int)n,
-				    (hipblasDoubleComplex **)&Ann[0], (int)n,
+				    (hipDoubleComplex **)&Ann[0], (int)n,
 				    (int*) &ipiv[0],
 				    (int*) &info[0],
 				    (int)batchCount);
+#else
+    auto err = hipblasZgetrfBatched(gridblasHandle,(int)n,
+                                    (hipblasDoubleComplex **)&Ann[0], (int)n,
+                                    (int*) &ipiv[0],
+                                    (int*) &info[0],
+                                    (int)batchCount);
+#endif
+    //    std::cout << " hipblas return code " <<(int)err<<" "<<__LINE__<<std::endl;
     GRID_ASSERT(err==HIPBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_CUDA
@@ -1124,11 +1164,21 @@ public:
     GRID_ASSERT(info.size()==batchCount);
 
 #ifdef GRID_HIP
+#if defined(HIP_VERSION_MAJOR) && (HIP_VERSION_MAJOR >=7)
     auto err = hipblasCgetrfBatched(gridblasHandle,(int)n,
-				    (hipblasComplex **)&Ann[0], (int)n,
+				    (hipComplex **)&Ann[0], (int)n,
 				    (int*) &ipiv[0],
 				    (int*) &info[0],
 				    (int)batchCount);
+#else
+    auto err = hipblasCgetrfBatched(gridblasHandle,(int)n,
+                                    (hipblasComplex **)&Ann[0], (int)n,
+                                    (int*) &ipiv[0],
+                                    (int*) &info[0],
+                                    (int)batchCount);
+#endif
+
+    //    std::cout << " hipblas return code " <<(int)err<<" "<<__LINE__<<std::endl;
     GRID_ASSERT(err==HIPBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_CUDA
@@ -1201,12 +1251,23 @@ public:
     GRID_ASSERT(Cnn.size()==batchCount);
 
 #ifdef GRID_HIP
+#if defined(HIP_VERSION_MAJOR) && (HIP_VERSION_MAJOR >=7)
     auto err = hipblasZgetriBatched(gridblasHandle,(int)n,
-				    (hipblasDoubleComplex **)&Ann[0], (int)n,
+				    (hipDoubleComplex **)&Ann[0], (int)n,
 				    (int*) &ipiv[0],
-				    (hipblasDoubleComplex **)&Cnn[0], (int)n,
+				    (hipDoubleComplex **)&Cnn[0], (int)n,
 				    (int*) &info[0],
 				    (int)batchCount);
+#else
+    auto err = hipblasZgetriBatched(gridblasHandle,(int)n,
+                                    (hipblasDoubleComplex **)&Ann[0], (int)n,
+                                    (int*) &ipiv[0],
+                                    (hipblasDoubleComplex **)&Cnn[0], (int)n,
+                                    (int*) &info[0],
+                                    (int)batchCount);
+
+#endif
+    //    std::cout << " hipblas return code " <<(int)err<<" "<<__LINE__<<std::endl;
     GRID_ASSERT(err==HIPBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_CUDA
@@ -1235,12 +1296,22 @@ public:
     GRID_ASSERT(Cnn.size()==batchCount);
 
 #ifdef GRID_HIP
+#if defined(HIP_VERSION_MAJOR) && (HIP_VERSION_MAJOR >=7)
     auto err = hipblasCgetriBatched(gridblasHandle,(int)n,
-				    (hipblasComplex **)&Ann[0], (int)n,
+				    (hipComplex **)&Ann[0], (int)n,
 				    (int*) &ipiv[0],
-				    (hipblasComplex **)&Cnn[0], (int)n,
+				    (hipComplex **)&Cnn[0], (int)n,
 				    (int*) &info[0],
 				    (int)batchCount);
+#else
+    auto err = hipblasCgetriBatched(gridblasHandle,(int)n,
+                                    (hipblasComplex **)&Ann[0], (int)n,
+                                    (int*) &ipiv[0],
+                                    (hipblasComplex **)&Cnn[0], (int)n,
+                                    (int*) &info[0],
+                                    (int)batchCount);
+#endif
+    //    std::cout << " hipblas return code " <<(int)err<<" "<<__LINE__<<std::endl;
     GRID_ASSERT(err==HIPBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_CUDA

Grid/algorithms/iterative/AdefMrhs.h

@@ -92,8 +92,8 @@ class TwoLevelCGmrhs
   // Vector case
   virtual void operator() (std::vector<Field> &src, std::vector<Field> &x)
   {
-    //    SolveSingleSystem(src,x);
-    SolvePrecBlockCG(src,x);
+    SolveSingleSystem(src,x);
+	// SolvePrecBlockCG(src,x);
   }
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////

Grid/algorithms/multigrid/Aggregates.h

@@ -97,7 +97,7 @@ public:
 
     RealD scale;
 
-    ConjugateGradient<FineField> CG(1.0e-3,400,false);
+    ConjugateGradient<FineField> CG(1.0e-4,2000,false);
     FineField noise(FineGrid);
     FineField Mn(FineGrid);
 
@@ -131,7 +131,10 @@ public:
     RealD scale;
 
     TrivialPrecon<FineField> simple_fine;
-    PrecGeneralisedConjugateResidualNonHermitian<FineField> GCR(0.001,30,DiracOp,simple_fine,12,12);
+    //    PrecGeneralisedConjugateResidualNonHermitian<FineField> GCR(0.001,10,DiracOp,simple_fine,30,30);
+    //    PrecGeneralisedConjugateResidualNonHermitian<FineField> GCR(0.001,10,DiracOp,simple_fine,12,12);
+    //    PrecGeneralisedConjugateResidualNonHermitian<FineField> GCR(0.001,30,DiracOp,simple_fine,12,12);
+    PrecGeneralisedConjugateResidualNonHermitian<FineField> GCR(0.001,30,DiracOp,simple_fine,10,10);
     FineField noise(FineGrid);
     FineField src(FineGrid);
     FineField guess(FineGrid);
@@ -146,16 +149,16 @@ public:
       
       DiracOp.Op(noise,Mn); std::cout<<GridLogMessage << "noise   ["<<b<<"] <n|Op|n> "<<innerProduct(noise,Mn)<<std::endl;
 
-      for(int i=0;i<2;i++){
+      for(int i=0;i<3;i++){
 	//  void operator() (const Field &src, Field &psi){
 #if 1
-	std::cout << GridLogMessage << " inverting on noise "<<std::endl;
+	if (i==0)std::cout << GridLogMessage << " inverting on noise "<<std::endl;
 	src = noise;
 	guess=Zero();
 	GCR(src,guess);
 	subspace[b] = guess;
 #else
-	std::cout << GridLogMessage << " inverting on zero "<<std::endl;
+	if (i==0)std::cout << GridLogMessage << " inverting on zero "<<std::endl;
 	src=Zero();
 	guess = noise;
 	GCR(src,guess);
@@ -167,7 +170,7 @@ public:
 
       }
 
-      DiracOp.Op(noise,Mn); std::cout<<GridLogMessage << "filtered["<<b<<"] <f|Op|f> "<<innerProduct(noise,Mn)<<std::endl;
+      DiracOp.Op(noise,Mn); std::cout<<GridLogMessage << "filtered["<<b<<"] <f|Op|f> "<<innerProduct(noise,Mn)<<" <f|OpDagOp|f>"<<norm2(Mn)<<std::endl;
       subspace[b]   = noise;
 
     }

Grid/allocator/MemoryManager.cc

@@ -63,12 +63,10 @@ void MemoryManager::PrintBytes(void)
   std::cout << " MemoryManager : "<<(total_device>>20)<<" accelerator Mbytes "<<std::endl;
   std::cout << " MemoryManager : "<<(total_host>>20)  <<" cpu         Mbytes "<<std::endl;
   uint64_t cacheBytes;
-  cacheBytes = CacheBytes[Cpu];
-  std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" cpu cache Mbytes "<<std::endl;
-  cacheBytes = CacheBytes[Acc];
-  std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" acc cache Mbytes "<<std::endl;
-  cacheBytes = CacheBytes[Shared];
-  std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" shared cache Mbytes "<<std::endl;
+  cacheBytes = HostCacheBytes();
+  std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" cpu alloc cache Mbytes "<<std::endl;
+  cacheBytes = DeviceCacheBytes();
+  std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" acc alloc cache Mbytes "<<std::endl;
   
 #ifdef GRID_CUDA
   cuda_mem();

Grid/communicator/Communicator_base.h

@@ -238,6 +238,16 @@ public:
   }
   void AllToAll(int dim  ,void *in,void *out,uint64_t words,uint64_t bytes);
   void AllToAll(void  *in,void *out,uint64_t words         ,uint64_t bytes);
+
+  ////////////////////////////////////////////////////////////////////////////
+  // Variable count all to all. Counts and displacements are in units of
+  // "bytes" sized words and are indexed by rank within this communicator.
+  // For exchanges that are a permutation but do not divide evenly between
+  // ranks; AllToAll above is the uniform count special case.
+  ////////////////////////////////////////////////////////////////////////////
+  void AllToAllV(void *in ,const std::vector<int> &sendcounts,const std::vector<int> &senddispls,
+                 void *out,const std::vector<int> &recvcounts,const std::vector<int> &recvdispls,
+                 uint64_t bytes);
   
   template<class obj> void Broadcast(int root,obj &data)
   {

Grid/communicator/Communicator_mpi3.cc

@@ -945,5 +945,25 @@ void CartesianCommunicator::AllToAll(void  *in,void *out,uint64_t words,uint64_t
   MPI_Alltoall(in,iwords,object,out,iwords,object,communicator);
   MPI_Type_free(&object);
 }
+void CartesianCommunicator::AllToAllV(void *in ,const std::vector<int> &sendcounts,const std::vector<int> &senddispls,
+                                      void *out,const std::vector<int> &recvcounts,const std::vector<int> &recvdispls,
+                                      uint64_t bytes)
+{
+  FlightRecorder::StepLog("AllToAllV");
+  GRID_ASSERT(sendcounts.size()==(size_t)_Nprocessors);
+  GRID_ASSERT(senddispls.size()==(size_t)_Nprocessors);
+  GRID_ASSERT(recvcounts.size()==(size_t)_Nprocessors);
+  GRID_ASSERT(recvdispls.size()==(size_t)_Nprocessors);
+  // MPI counts are "int"; the caller sizes the word to keep them in range
+  int ibytes = bytes;
+  GRID_ASSERT(bytes == (uint64_t)ibytes);
+  MPI_Datatype object;
+  MPI_Type_contiguous(ibytes,MPI_BYTE,&object);
+  MPI_Type_commit(&object);
+  int ierr = MPI_Alltoallv(in ,(int *)&sendcounts[0],(int *)&senddispls[0],object,
+                           out,(int *)&recvcounts[0],(int *)&recvdispls[0],object,communicator);
+  GRID_ASSERT(ierr==0);
+  MPI_Type_free(&object);
+}
 
 NAMESPACE_END(Grid);

Grid/communicator/Communicator_none.cc

@@ -113,6 +113,17 @@ void CartesianCommunicator::AllToAll(void  *in,void *out,uint64_t words,uint64_t
 {
   bcopy(in,out,bytes*words);
 }
+void CartesianCommunicator::AllToAllV(void *in ,const std::vector<int> &sendcounts,const std::vector<int> &senddispls,
+                                      void *out,const std::vector<int> &recvcounts,const std::vector<int> &recvdispls,
+                                      uint64_t bytes)
+{
+  // Single rank: the exchange degenerates to a copy of our own segment
+  GRID_ASSERT(sendcounts.size()==1);
+  GRID_ASSERT(recvcounts.size()==1);
+  GRID_ASSERT(sendcounts[0]==recvcounts[0]);
+  bcopy((char *)in +(uint64_t)senddispls[0]*bytes,
+        (char *)out+(uint64_t)recvdispls[0]*bytes,bytes*(uint64_t)sendcounts[0]);
+}
 
 int  CartesianCommunicator::RankWorld(void){return 0;}
 void CartesianCommunicator::Barrier(void){}

Grid/lattice/Lattice_base.h

@@ -289,7 +289,7 @@ public:
   ///////////////////////////////////////////
   // move constructor
   ///////////////////////////////////////////
-  Lattice(Lattice && r){ 
+  Lattice(Lattice && r) noexcept { 
     this->_grid = r.Grid();
     this->_odata      = r._odata;
     this->_odata_size = r._odata_size;
@@ -330,7 +330,7 @@ public:
   ///////////////////////////////////////////
   // Move assignment possible if same type
   ///////////////////////////////////////////
-  inline Lattice<vobj> & operator = (Lattice<vobj> && r){
+  inline Lattice<vobj> & operator = (Lattice<vobj> && r) noexcept {
 
     resize(0); // deletes if appropriate
     this->_grid       = r.Grid();

Grid/lattice/Lattice_reduction_gpu.h

@@ -197,12 +197,15 @@ __global__ void reduceKernel(const vobj *lat, sobj *buffer, Iterator n) {
 /////////////////////////////////////////////////////////////////////////////////////////////////////////
 // Possibly promote to double and sum
 /////////////////////////////////////////////////////////////////////////////////////////////////////////
+
+#undef GRID_REDUCTION_TIMING
+
 template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osites) 
+inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osites)
 {
   typedef typename vobj::scalar_objectD sobj;
   typedef decltype(lat) Iterator;
-  
+
   Integer nsimd= vobj::Nsimd();
   Integer size = osites*nsimd;
 
@@ -211,41 +214,188 @@ inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osi
   GRID_ASSERT(ok);
 
   Integer smemSize = numThreads * sizeof(sobj);
-  // Move out of UVM
-  // Turns out I had messed up the synchronise after move to compute stream
-  // as running this on the default stream fools the synchronise
   deviceVector<sobj> buffer(numBlocks);
   sobj *buffer_v = &buffer[0];
   sobj result;
+
+#ifdef GRID_REDUCTION_TIMING
+  RealD t_kernel = -usecond();
+#endif
   reduceKernel<<< numBlocks, numThreads, smemSize, computeStream >>>(lat, buffer_v, size);
   accelerator_barrier();
+#ifdef GRID_REDUCTION_TIMING
+  t_kernel += usecond();
+  RealD t_d2h = -usecond();
+#endif
   acceleratorCopyFromDevice(buffer_v,&result,sizeof(result));
+#ifdef GRID_REDUCTION_TIMING
+  t_d2h += usecond();
+  std::cout << GridLogDebug << "  sumD_gpu_small"
+            << " sizeof(sobj)=" << sizeof(sobj)
+            << " blocks=" << numBlocks << " threads=" << numThreads
+            << " kernel+barrier=" << t_kernel << " us"
+            << " D2H=" << t_d2h << " us" << std::endl;
+#endif
   return result;
 }
 
+// Fused pack+reduce: reads R words of each vobj at word offset 'base',
+// accumulates directly into iVector<iScalar<scalarD>,R> without staging
+// through an intermediate bundle buffer.  One HBM pass instead of three.
+template <int R, class vobj, class sobj, class Iterator>
+__device__ void packReduceBlocks(
+    const iScalar<typename vobj::vector_type> *idat,
+    sobj *g_odata, Iterator osites, int base, int words)
+{
+  constexpr Iterator nsimd = vobj::Nsimd();
+  Iterator blockSize = blockDim.x;
+
+  extern __shared__ __align__(COALESCE_GRANULARITY) unsigned char shmem_pointer[];
+  sobj *sdata = (sobj *)shmem_pointer;
+
+  Iterator tid      = threadIdx.x;
+  Iterator i        = blockIdx.x * (blockSize * 2) + threadIdx.x;
+  Iterator gridSize = blockSize * 2 * gridDim.x;
+  sobj mySum = Zero();
+
+  while (i < osites * nsimd) {
+    Iterator lane = i % nsimd;
+    Iterator ss   = i / nsimd;
+    sobj tmpD; zeroit(tmpD);
+    for (int k = 0; k < R; k++) {
+      auto w = extractLane(lane, idat[ss * words + base + k]);
+      iScalar<typename vobj::scalar_typeD> wd; wd = w;
+      tmpD._internal[k] = wd;
+    }
+    mySum += tmpD;
+
+    if (i + blockSize < osites * nsimd) {
+      lane = (i + blockSize) % nsimd;
+      ss   = (i + blockSize) / nsimd;
+      sobj tmpD2; zeroit(tmpD2);
+      for (int k = 0; k < R; k++) {
+        auto w = extractLane(lane, idat[ss * words + base + k]);
+        iScalar<typename vobj::scalar_typeD> wd; wd = w;
+        tmpD2._internal[k] = wd;
+      }
+      mySum += tmpD2;
+    }
+    i += gridSize;
+  }
+
+  reduceBlock(sdata, mySum, tid);
+  if (tid == 0) g_odata[blockIdx.x] = sdata[0];
+}
+
+template <int R, class vobj, class sobj, class Iterator>
+__global__ void packReduceKernel(
+    const iScalar<typename vobj::vector_type> *idat,
+    sobj *buffer, Iterator osites, int base, int words)
+{
+  Iterator blockSize = blockDim.x;
+
+  packReduceBlocks<R, vobj, sobj>(idat, buffer, osites, base, words);
+
+  if (gridDim.x > 1) {
+    const Iterator tid = threadIdx.x;
+    __shared__ bool amLast;
+    extern __shared__ __align__(COALESCE_GRANULARITY) unsigned char shmem_pointer[];
+    sobj *smem = (sobj *)shmem_pointer;
+
+    acceleratorFence();
+
+    if (tid == 0) {
+      unsigned int ticket = atomicInc(&retirementCount, gridDim.x);
+      amLast = (ticket == gridDim.x - 1);
+    }
+    acceleratorSynchroniseAll();
+
+    if (amLast) {
+      Iterator i = tid;
+      sobj mySum = Zero();
+      while (i < (Iterator)gridDim.x) {
+        mySum += buffer[i];
+        i += blockSize;
+      }
+      reduceBlock(smem, mySum, tid);
+      if (tid == 0) {
+        buffer[0] = smem[0];
+        retirementCount = 0;
+      }
+    }
+  }
+}
+
+template<int R, class vobj>
+inline void sumD_gpu_reduce_words(const vobj *lat, Integer osites,
+                                   typename vobj::scalar_typeD *ret_p, int base)
+{
+  typedef typename vobj::vector_type  vector;
+  typedef typename vobj::scalar_typeD scalarD;
+  using BundleScalarD = iVector<iScalar<scalarD>, R>;
+
+  constexpr int Nsimd = vobj::Nsimd();
+  const int words = sizeof(vobj) / sizeof(vector);
+  const iScalar<vector> *idat = (const iScalar<vector> *)lat;
+
+  Integer size = (Integer)osites * Nsimd;
+  Integer numThreads, numBlocks;
+  int ok = getNumBlocksAndThreads(size, sizeof(BundleScalarD), numThreads, numBlocks);
+  GRID_ASSERT(ok);
+
+  Integer smemSize = numThreads * sizeof(BundleScalarD);
+  deviceVector<BundleScalarD> buffer(numBlocks);
+  BundleScalarD *buffer_v = &buffer[0];
+  BundleScalarD result;
+
+#ifdef GRID_REDUCTION_TIMING
+  RealD t_kernel = -usecond();
+#endif
+  packReduceKernel<R, vobj, BundleScalarD, Integer>
+    <<<numBlocks, numThreads, smemSize, computeStream>>>
+    (idat, buffer_v, osites, base, words);
+  accelerator_barrier();
+#ifdef GRID_REDUCTION_TIMING
+  t_kernel += usecond();
+  RealD t_d2h = -usecond();
+#endif
+  acceleratorCopyFromDevice(buffer_v, &result, sizeof(result));
+#ifdef GRID_REDUCTION_TIMING
+  t_d2h += usecond();
+  std::cout << GridLogDebug << " sumD_gpu_reduce_words R=" << R
+            << " base=" << base
+            << " kernel=" << t_kernel << " D2H=" << t_d2h << " us" << std::endl;
+#endif
+
+  for (int k = 0; k < R; k++)
+    ret_p[base + k] = TensorRemove(result._internal[k]);
+}
+
 template <class vobj>
 inline typename vobj::scalar_objectD sumD_gpu_large(const vobj *lat, Integer osites)
 {
-  typedef typename vobj::vector_type  vector;
-  typedef typename vobj::scalar_typeD scalarD;
-  typedef typename vobj::scalar_objectD sobj;
-  sobj ret;
+  typedef typename vobj::vector_type    vector;
+  typedef typename vobj::scalar_typeD   scalarD;
+  typedef typename vobj::scalar_objectD sobjD;
+
+  const int words = sizeof(vobj) / sizeof(vector);
+  sobjD ret; zeroit(ret);
   scalarD *ret_p = (scalarD *)&ret;
-  
-  const int words = sizeof(vobj)/sizeof(vector);
 
-  deviceVector<vector> buffer(osites);
-  vector *dat = (vector *)lat;
-  vector *buf = &buffer[0];
-  iScalar<vector> *tbuf =(iScalar<vector> *)  &buffer[0];
-  for(int w=0;w<words;w++) {
+#ifdef GRID_REDUCTION_TIMING
+  RealD t_large = -usecond();
+#endif
+  int w = 0;
+  while (w + 12 <= words) { sumD_gpu_reduce_words<12>(lat, osites, ret_p, w); w += 12; }
+  while (w +  4 <= words) { sumD_gpu_reduce_words< 4>(lat, osites, ret_p, w); w +=  4; }
+  while (w      <  words) { sumD_gpu_reduce_words< 1>(lat, osites, ret_p, w); w +=  1; }
+#ifdef GRID_REDUCTION_TIMING
+  t_large += usecond();
+  std::cout << GridLogDebug << "sumD_gpu_large"
+            << " sizeof(sobjD)=" << sizeof(sobjD)
+            << " words=" << words << " total=" << t_large << " us" << std::endl;
+#endif
 
-    accelerator_for(ss,osites,1,{
-	buf[ss] = dat[ss*words+w];
-      });
-      
-    ret_p[w] = sumD_gpu_small(tbuf,osites);
-  }
   return ret;
 }
 
@@ -288,5 +438,11 @@ inline typename vobj::scalar_object sum_gpu_large(const vobj *lat, Integer osite
   result = sumD_gpu_large(lat,osites);
   return result;
 }
+template<class Word> Word checksum_gpu(Word *vec,uint64_t L)
+{
+  Word w;
+  bzero(&w,sizeof(w));
+  return w;
+}
 
 NAMESPACE_END(Grid);

Grid/lattice/Lattice_reduction_sycl.h

@@ -6,28 +6,27 @@ NAMESPACE_BEGIN(Grid);
 
 
 template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu_tensor(const vobj *lat, Integer osites) 
+inline typename vobj::scalar_objectD sumD_gpu_tensor(const vobj *lat, Integer osites)
 {
-  typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_object  sobj;
   typedef typename vobj::scalar_objectD sobjD;
 
-  sobj identity; zeroit(identity);
-  sobj ret; zeroit(ret);
-  Integer nsimd= vobj::Nsimd();
-  { 
-    sycl::buffer<sobj, 1> abuff(&ret, {1});
+  sobjD identity; zeroit(identity);
+  sobjD ret;      zeroit(ret);
+  {
+    sycl::buffer<sobjD, 1> abuff(&ret, {1});
     theGridAccelerator->submit([&](sycl::handler &cgh) {
-      auto Reduction = sycl::reduction(abuff,cgh,identity,std::plus<>());
-      cgh.parallel_for(sycl::range<1>{osites},
-                      Reduction,
-                      [=] (sycl::id<1> item, auto &sum) {
-                        auto osite   = item[0];
-                        sum +=Reduce(lat[osite]);
-                      });
+      auto Reduction = sycl::reduction(abuff, cgh, identity, std::plus<>());
+      cgh.parallel_for(sycl::range<1>{(size_t)osites},
+                       Reduction,
+                       [=](sycl::id<1> item, auto &sum) {
+                         sobj s = Reduce(lat[item[0]]);
+                         sobjD sd; sd = s;
+                         sum += sd;
+                       });
     });
   }
-  sobjD dret; convertType(dret,ret);
-  return dret;
+  return ret;
 }
 
 template <class vobj>

Grid/lattice/Lattice_slicesum_core.h

@@ -1,7 +1,6 @@
 #pragma once
 
 #if defined(GRID_CUDA)
-
 #include <cub/cub.cuh>
 #define gpucub cub
 #define gpuError_t cudaError_t
@@ -57,8 +56,13 @@ inline void sliceSumReduction_cub_small(const vobj *Data,
   //copy offsets to device
   acceleratorCopyToDeviceAsynch(&offsets[0],d_offsets,sizeof(int)*(rd+1),computeStream);
   
+#if defined(__CUDACC__) && (__CUDACC_VER_MAJOR__ >= 13)
+  #define GRID_CUB_SUM_OP ::cuda::std::plus<>{}
+#else
+  #define GRID_CUB_SUM_OP ::gpucub::Sum()
+#endif
   
-  gpuError_t gpuErr = gpucub::DeviceSegmentedReduce::Reduce(temp_storage_array, temp_storage_bytes, rb_p,d_out, rd, d_offsets, d_offsets+1, ::gpucub::Sum(), zero_init, computeStream);
+  gpuError_t gpuErr = gpucub::DeviceSegmentedReduce::Reduce(temp_storage_array, temp_storage_bytes, rb_p,d_out, rd, d_offsets, d_offsets+1, GRID_CUB_SUM_OP, zero_init, computeStream);
   if (gpuErr!=gpuSuccess) {
     std::cout << GridLogError << "Lattice_slicesum_gpu.h: Encountered error during gpucub::DeviceSegmentedReduce::Reduce (setup)! Error: " << gpuErr <<std::endl;
     exit(EXIT_FAILURE);
@@ -82,11 +86,13 @@ inline void sliceSumReduction_cub_small(const vobj *Data,
   });
   
   //issue segmented reductions in computeStream
-  gpuErr = gpucub::DeviceSegmentedReduce::Reduce(temp_storage_array, temp_storage_bytes, rb_p, d_out, rd, d_offsets, d_offsets+1,::gpucub::Sum(), zero_init, computeStream);
+  gpuErr = gpucub::DeviceSegmentedReduce::Reduce(temp_storage_array, temp_storage_bytes, rb_p, d_out, rd, d_offsets, d_offsets+1, GRID_CUB_SUM_OP, zero_init, computeStream);
   if (gpuErr!=gpuSuccess) {
     std::cout << GridLogError << "Lattice_slicesum_gpu.h: Encountered error during gpucub::DeviceSegmentedReduce::Reduce! Error: " << gpuErr <<std::endl;
     exit(EXIT_FAILURE);
   }
+
+#undef GRID_CUB_SUM_OP
   
   acceleratorCopyFromDeviceAsynch(d_out,&lvSum[0],rd*sizeof(vobj),computeStream);
   

Grid/parallelIO/BinaryIO.cc

@@ -2,3 +2,7 @@
 
 int                    Grid::BinaryIO::latticeWriteMaxRetry = -1;
 Grid::BinaryIO::IoPerf Grid::BinaryIO::lastPerf;
+
+// Target size of a single contiguous file extent under BINARYIO_AGGREGATE.
+// 4MB is around the knee for Lustre; exposed so it can be swept at runtime.
+uint64_t               Grid::BinaryIO::aggregateTargetBytes = 4*1024*1024;

Grid/parallelIO/BinaryIO.h

@@ -39,6 +39,7 @@
 #endif
 
 #include <arpa/inet.h>
+#include <sys/stat.h>
 #include <algorithm>
 
 NAMESPACE_BEGIN(Grid);
@@ -87,6 +88,7 @@ class BinaryIO {
 
   static IoPerf lastPerf;
   static int latticeWriteMaxRetry;
+  static uint64_t aggregateTargetBytes;
 
   /////////////////////////////////////////////////////////////////////////////
   // more byte manipulation helpers
@@ -253,12 +255,392 @@ class BinaryIO {
   // Read or Write distributed lexico array of ANY object to a specific location in file 
   //////////////////////////////////////////////////////////////////////////////////////
 
+  static const int BINARYIO_AGGREGATE     = 0x20;
   static const int BINARYIO_MASTER_APPEND = 0x10;
   static const int BINARYIO_UNORDERED     = 0x08;
   static const int BINARYIO_LEXICOGRAPHIC = 0x04;
   static const int BINARYIO_READ          = 0x02;
   static const int BINARYIO_WRITE         = 0x01;
 
+#ifdef USE_MPI_IO
+  /////////////////////////////////////////////////////////////////////////////
+  // Aggregation: self controlled transposition onto an I/O friendly layout.
+  //
+  // Under BINARYIO_LEXICOGRAPHIC the subarray file view handed to MPI-IO has
+  // contiguous runs of only lLattice[0]*sizeof(fobj) bytes -- a few KB for
+  // typical local volumes. Rather than rely on collective buffering to repair
+  // that, redistribute the payload ourselves so every rank owns a contiguous
+  // range of the global lexicographic site ordering, then issue large plain
+  // contiguous writes.
+  //
+  // "Un-splitting" the nunsplit fastest dimensions means the row of ranks
+  // sharing the remaining process coordinates collectively owns whole global
+  // hyperplanes. All data movement is then confined to that row communicator.
+  // Every rank still owns exactly lSites() sites afterwards, so the exchange is
+  // a pure permutation and needs no divisibility condition on the process grid.
+  /////////////////////////////////////////////////////////////////////////////
+  struct AggregationPlan {
+    int      nunsplit{0};  // number of fastest dimensions un-split
+    int      rowsize{0};   // ranks in the aggregation (row) communicator
+    int      rowrank{0};   // our logical (lexicographic) index within the row
+    uint64_t lsites{0};    // sites per rank -- invariant under the permutation
+    uint64_t chunk{0};     // sites in one globally contiguous run owned by the row
+    std::unique_ptr<CartesianCommunicator> rowcomm;
+    // counts and displacements are indexed by rank within rowcomm
+    std::vector<int>      sendcounts, senddispls, recvcounts, recvdispls;
+    std::vector<uint64_t> scatter;     // recv slot -> slot in the aggregated buffer
+    std::vector<uint64_t> extentGsite; // global lex site index of extent start
+    std::vector<uint64_t> extentLocal; // offset of extent within aggregated buffer
+    std::vector<uint64_t> extentSites; // sites in this extent
+  };
+
+  static inline void BuildAggregationPlan(GridBase *grid,uint64_t fobjSize,AggregationPlan &p)
+  {
+    int ndim           = grid->Dimensions();
+    Coordinate psizes  = grid->ProcessorGrid();
+    Coordinate pcoor   = grid->ThisProcessorCoor();
+    Coordinate gLattice= grid->GlobalDimensions();
+    Coordinate lLattice= grid->LocalDimensions();
+    Coordinate lstart  = grid->LocalStarts();
+
+    uint64_t lsites = grid->lSites();
+    p.lsites = lsites;
+
+    //////////////////////////////////////////////////////////////////////////
+    // Un-splitting dims 0..k-1 gives the row a contiguous run of
+    //     chunk(k) = prod_{d<k} gLattice[d] * lLattice[k]
+    // sites, and each rank writes extents of min(chunk,lsites). Take the
+    // smallest k that reaches the target so we disturb as few dimensions --
+    // and move as little data -- as possible.
+    //////////////////////////////////////////////////////////////////////////
+    int k = ndim-1;
+    for(int trial=1; trial<ndim; trial++){
+      uint64_t chunk = lLattice[trial];
+      for(int d=0; d<trial; d++) chunk *= gLattice[d];
+      if ( std::min(chunk,lsites)*fobjSize >= aggregateTargetBytes ) { k = trial; break; }
+    }
+    p.nunsplit = k;
+
+    //////////////////////////////////////////////////////////////////////////
+    // The box the row collectively owns, expressed in global coordinates.
+    // Restricting the global lexicographic order to this box preserves the
+    // ordering, so the row index below is monotone in the global index.
+    //////////////////////////////////////////////////////////////////////////
+    Coordinate B(ndim), S(ndim);
+    for(int d=0; d<ndim; d++){
+      if ( d<k ) { B[d] = gLattice[d]; S[d] = 0;         }
+      else       { B[d] = lLattice[d]; S[d] = lstart[d]; }
+    }
+
+    uint64_t chunk = lLattice[k];
+    for(int d=0; d<k; d++) chunk *= gLattice[d];
+    p.chunk = chunk;
+
+    //////////////////////////////////////////////////////////////////////////
+    // Row communicator: the ranks sharing the process coordinates of the slow
+    // (still split) dimensions. This is the sub-division the Cartesian
+    // communicator already performs for AllToAll(dim,...), widened from one
+    // dimension to the k fastest.
+    //////////////////////////////////////////////////////////////////////////
+    Coordinate row(ndim,1);
+    for(int d=0; d<k; d++) row[d] = psizes[d];
+    int srank;
+    p.rowcomm.reset(new CartesianCommunicator(row,*grid,srank));
+    p.rowsize = p.rowcomm->ProcessorCount();
+
+    //////////////////////////////////////////////////////////////////////////
+    // Our logical index in the row is the forward lexicographic index of the
+    // un-split process coordinates, so that increasing logical index means
+    // increasing global lexicographic position in the file. The communicator
+    // numbers its own ranks by the reversed (MPI) convention, so build the map
+    // between the two rather than assuming either.
+    //////////////////////////////////////////////////////////////////////////
+    int64_t logical=0, lstride=1;
+    for(int d=0; d<k; d++){ logical += pcoor[d]*lstride; lstride *= psizes[d]; }
+    GRID_ASSERT(lstride == (int64_t)p.rowsize);
+    p.rowrank = (int)logical;
+
+    std::vector<uint64_t> commOf(p.rowsize,0);
+    commOf[p.rowrank] = (uint64_t)p.rowcomm->ThisRank();
+    p.rowcomm->GlobalSumVector(&commOf[0],p.rowsize);
+
+    uint64_t mystart = (uint64_t)p.rowrank * lsites;
+    uint64_t myend   = mystart + lsites;
+
+    Coordinate lcoor(ndim), bcoor(ndim), gcoor(ndim);
+
+    //////////////////////////////////////////////////////////////////////////
+    // Send side. Walking our local sites in local lexicographic order walks
+    // the row index monotonically, so the send buffer is iodata untouched and
+    // we need only the per destination counts.
+    //////////////////////////////////////////////////////////////////////////
+    std::vector<int> sendLogical(p.rowsize,0);
+    for(uint64_t L=0; L<lsites; L++){
+      Lexicographic::CoorFromIndex(lcoor,L,lLattice);
+      for(int d=0; d<ndim; d++) bcoor[d] = (d<k) ? (lstart[d]+lcoor[d]) : lcoor[d];
+      int64_t ri; Lexicographic::IndexFromCoor(bcoor,ri,B);
+      sendLogical[ ri/(int64_t)lsites ]++;
+    }
+    p.sendcounts.assign(p.rowsize,0);
+    p.senddispls.assign(p.rowsize,0);
+    { int64_t disp=0;
+      for(int d=0; d<p.rowsize; d++){          // send buffer is in logical order
+        int c = (int)commOf[d];
+        p.sendcounts[c] = sendLogical[d];
+        p.senddispls[c] = (int)disp;
+        disp += sendLogical[d];
+      }
+    }
+
+    //////////////////////////////////////////////////////////////////////////
+    // Receive side. For each slot of our aggregated range work out which rank
+    // of the row owns it. Within one source the slots arrive in increasing row
+    // index order, which is the order the source sends them in.
+    //////////////////////////////////////////////////////////////////////////
+    std::vector<int> recvLogical(p.rowsize,0), recvDisplLogical(p.rowsize,0);
+    std::vector<int> source(lsites);
+    for(uint64_t pos=0; pos<lsites; pos++){
+      Lexicographic::CoorFromIndex(bcoor,(int64_t)(mystart+pos),B);
+      int64_t j=0, jstride=1;
+      for(int d=0; d<k; d++){ j += (bcoor[d]/lLattice[d])*jstride; jstride *= psizes[d]; }
+      source[pos] = (int)j;
+      recvLogical[j]++;
+    }
+    p.recvcounts.assign(p.rowsize,0);
+    p.recvdispls.assign(p.rowsize,0);
+    { int64_t disp=0;
+      for(int s=0; s<p.rowsize; s++){         // recv buffer is in logical order
+        int c = (int)commOf[s];
+        recvDisplLogical[s] = (int)disp;
+        p.recvcounts[c] = recvLogical[s];
+        p.recvdispls[c] = (int)disp;
+        disp += recvLogical[s];
+      }
+    }
+    p.scatter.resize(lsites);
+    {
+      std::vector<int> fill(p.rowsize,0);
+      for(uint64_t pos=0; pos<lsites; pos++){
+        int j = source[pos];
+        p.scatter[ recvDisplLogical[j] + fill[j]++ ] = pos;
+      }
+    }
+
+    //////////////////////////////////////////////////////////////////////////
+    // The two sides are derived independently; make them check each other.
+    //////////////////////////////////////////////////////////////////////////
+    {
+      std::vector<uint64_t> sendc(p.rowsize),recvc(p.rowsize);
+      for(int c=0;c<p.rowsize;c++) sendc[c]=(uint64_t)p.sendcounts[c];
+      p.rowcomm->AllToAll(&sendc[0],&recvc[0],1,sizeof(uint64_t));
+      for(int c=0;c<p.rowsize;c++) GRID_ASSERT((int)recvc[c]==p.recvcounts[c]);
+    }
+
+    //////////////////////////////////////////////////////////////////////////
+    // Decompose our range into globally contiguous file extents.
+    //////////////////////////////////////////////////////////////////////////
+    for(uint64_t c = mystart/chunk; c <= (myend-1)/chunk; c++){
+      uint64_t lo = std::max(mystart, c*chunk);
+      uint64_t hi = std::min(myend, (c+1)*chunk);
+      Lexicographic::CoorFromIndex(bcoor,(int64_t)(c*chunk),B);
+      for(int d=0;d<ndim;d++) gcoor[d] = (d<k) ? bcoor[d] : bcoor[d]+S[d];
+      int64_t gbase; Lexicographic::IndexFromCoor(gcoor,gbase,gLattice);
+      p.extentGsite.push_back( (uint64_t)gbase + (lo - c*chunk) );
+      p.extentLocal.push_back( lo - mystart );
+      p.extentSites.push_back( hi - lo );
+    }
+  }
+
+  static inline void ReportAggregationPlan(GridBase *grid,const AggregationPlan &p,uint64_t fobjSize,const char *what)
+  {
+    if ( !grid->IsBoss() ) return;
+    std::cout << GridLogMessage << "IOobject: aggregate " << what
+              << " un-splitting " << p.nunsplit << " fastest dimensions, row of "
+              << p.rowsize << " ranks" << std::endl;
+    std::cout << GridLogMessage << "IOobject: aggregate " << p.extentSites.size()
+              << " extent(s)/rank, first " << p.extentSites[0]*fobjSize/1024./1024. << " MB"
+              << " (target " << aggregateTargetBytes/1024./1024. << " MB)" << std::endl;
+    std::cout << GridLogMessage << "IOobject: aggregate buffer overhead "
+              << p.lsites*fobjSize/1024./1024. << " MB/rank" << std::endl;
+  }
+
+  ////////////////////////////////////////////////////////////////////////////
+  // Stage timings.  The interesting quantity is the slowest rank, since every
+  // stage is followed sooner or later by a synchronisation, so reduce with
+  // GlobalMax rather than reporting whatever the boss happened to see.
+  ////////////////////////////////////////////////////////////////////////////
+  static inline void ReportStages(GridBase *grid,const char *what,
+                                  const std::vector<const char *> &names,
+                                  std::vector<RealD> &useconds)
+  {
+    GRID_ASSERT(names.size()==useconds.size());
+    for(uint64_t i=0;i<useconds.size();i++) grid->GlobalMax(useconds[i]);
+    if ( grid->IsBoss() ) {
+      std::cout << GridLogMessage << "IOobject: aggregate " << what << " stages (max over ranks, s):";
+      for(uint64_t i=0;i<names.size();i++)
+        std::cout << " " << names[i] << " " << useconds[i]/1.0e6;
+      std::cout << std::endl;
+    }
+  }
+
+  template<class fobj>
+  static inline void AggregateExchange(GridBase *grid,AggregationPlan &p,std::vector<fobj> &iodata,
+                                       std::vector<fobj> &aggregated,int forward)
+  {
+    uint64_t lsites = p.lsites;
+    GridStopWatch talloc,tperm,tcomm;
+
+    talloc.Start();
+    std::vector<fobj> tmp(lsites);
+    talloc.Stop();
+
+    if ( forward ) { // iodata (local order) -> aggregated (lexicographic order)
+      tcomm.Start();
+      p.rowcomm->AllToAllV(&iodata[0],p.sendcounts,p.senddispls,
+                           &tmp[0],   p.recvcounts,p.recvdispls,sizeof(fobj));
+      tcomm.Stop();
+      tperm.Start();
+      thread_for(s,lsites,{ aggregated[p.scatter[s]] = tmp[s]; });
+      tperm.Stop();
+    } else {         // aggregated -> iodata, the exact mirror
+      tperm.Start();
+      thread_for(s,lsites,{ tmp[s] = aggregated[p.scatter[s]]; });
+      tperm.Stop();
+      tcomm.Start();
+      p.rowcomm->AllToAllV(&tmp[0],   p.recvcounts,p.recvdispls,
+                           &iodata[0],p.sendcounts,p.senddispls,sizeof(fobj));
+      tcomm.Stop();
+    }
+
+    std::vector<RealD> us = { (RealD)talloc.useconds(), (RealD)tperm.useconds(), (RealD)tcomm.useconds() };
+    ReportStages(grid,forward?"exchange (write)":"exchange (read)",
+                 {"alloc","permute","alltoallv"},us);
+  }
+
+  template<class fobj>
+  static inline void AggregateWrite(GridBase *grid,AggregationPlan &p,std::vector<fobj> &aggregated,
+                                    std::string file,uint64_t offset)
+  {
+    //////////////////////////////////////////////////////////////////////////
+    // All ranks write concurrently into a shared file, so the file must exist
+    // before any of them open it for update, but it does NOT have to be the
+    // right length first: the extents tile the record exactly, so writing them
+    // extends a short file to precisely offset+payload.
+    //
+    // Records are created in sequence, so this payload ends the file: the
+    // length must end up precisely offset+payload. Anything beyond is left
+    // over from whatever the file previously held and must not survive -- a
+    // shorter new record written over a longer old one would otherwise leave
+    // a trailing fragment of the previous contents masquerading as data.
+    // That is the only case needing a truncate, so stat first and truncate
+    // afterwards only when the size actually came out wrong. Measured on
+    // Frontier, an unconditional truncate up front cost 0.22 to 5.4 s per
+    // record -- 15 to 25% of a 19 GB write and 100% of a small one -- while
+    // create, open and close together cost a few milliseconds. It is per
+    // record, so multi record files do not amortise it away.
+    //
+    // ::truncate is used because the C++ standard library cannot express this.
+    // std::filebuf has no length operation at all; ios::trunc only truncates to
+    // zero; seeking past the end and writing a byte can grow a file but never
+    // shrink one; and there is no portable way to recover a descriptor from a
+    // stream in order to call ftruncate. C++17 does finally offer
+    // std::filesystem::resize_file, but that would be Grid's first <filesystem>
+    // dependency and needs -lstdc++fs on the older toolchains still in use.
+    //////////////////////////////////////////////////////////////////////////
+    GridStopWatch tcreate,ttrunc,tbar,topen,twrite,tclose,tskew;
+    uint64_t need = offset + (uint64_t)grid->_gsites*sizeof(fobj);
+
+    tcreate.Start();
+    if ( grid->IsBoss() ) {
+      // opening for update needs the file to exist; create one only if not
+      std::fstream probe(file,std::ios::binary|std::ios::out|std::ios::in);
+      if ( !probe.is_open() ) {
+        std::ofstream create(file,std::ios::binary|std::ios::out);
+        create.close();
+      }
+    }
+    tcreate.Stop();
+
+    tbar.Start();
+    grid->Barrier();
+    tbar.Stop();
+
+    std::ofstream fout;
+    fout.exceptions( std::fstream::failbit | std::fstream::badbit );
+    try {
+      topen.Start();
+      fout.open(file,std::ios::binary|std::ios::out|std::ios::in);
+      topen.Stop();
+      twrite.Start();
+      for(uint64_t e=0;e<p.extentSites.size();e++){
+        fout.seekp(offset + p.extentGsite[e]*sizeof(fobj));
+        fout.write((char *)&aggregated[p.extentLocal[e]],p.extentSites[e]*sizeof(fobj));
+      }
+      twrite.Stop();
+      tclose.Start();
+      fout.close();   // flushes the stream buffer; does not force writeback
+      tclose.Stop();
+    } catch (const std::fstream::failure& exc) {
+      std::cout << GridLogError << "Error in aggregate write to " << file << std::endl;
+      std::cout << GridLogError << "Exception description: " << exc.what() << std::endl;
+      GridAbort();
+    }
+
+    ////////////////////////////////////////////////////////////////////////
+    // Timed apart from the truncate that follows it.  seek+write above is the
+    // slowest rank; this barrier is what the fastest rank then waits, so the
+    // pair separates the write cost from the spread across ranks. Folding it
+    // into the truncate makes a millisecond stat look like a second.
+    ////////////////////////////////////////////////////////////////////////
+    tskew.Start();
+    grid->Barrier();                 // every extent must be on its way first
+    tskew.Stop();
+
+    ttrunc.Start();
+    if ( grid->IsBoss() ) {
+      struct stat sb;
+      int ierr = ::stat(file.c_str(),&sb);
+      GRID_ASSERT(ierr==0);
+      if ( (uint64_t)sb.st_size != need ) {   // only when a longer record preceded us
+        ierr = ::truncate(file.c_str(),(off_t)need);
+        GRID_ASSERT(ierr==0);
+      }
+    }
+    grid->Barrier();
+    ttrunc.Stop();
+
+    std::vector<RealD> us = { (RealD)tcreate.useconds(), (RealD)tbar.useconds(),
+                              (RealD)topen.useconds(),   (RealD)twrite.useconds(),
+                              (RealD)tclose.useconds(),  (RealD)tskew.useconds(),
+                              (RealD)ttrunc.useconds() };
+    ReportStages(grid,"write",{"create","barrier","open","seek+write","close","skew","stat+truncate"},us);
+  }
+
+  template<class fobj>
+  static inline void AggregateRead(GridBase *grid,AggregationPlan &p,std::vector<fobj> &aggregated,
+                                   std::string file,uint64_t offset)
+  {
+    GridStopWatch topen,tread,tclose;
+    std::ifstream fin;
+    topen.Start();
+    fin.open(file,std::ios::binary|std::ios::in);
+    topen.Stop();
+    tread.Start();
+    for(uint64_t e=0;e<p.extentSites.size();e++){
+      fin.seekg(offset + p.extentGsite[e]*sizeof(fobj));
+      fin.read((char *)&aggregated[p.extentLocal[e]],p.extentSites[e]*sizeof(fobj));
+      GRID_ASSERT(fin.fail()==0);
+    }
+    tread.Stop();
+    tclose.Start();
+    fin.close();
+    tclose.Stop();
+
+    std::vector<RealD> us = { (RealD)topen.useconds(), (RealD)tread.useconds(), (RealD)tclose.useconds() };
+    ReportStages(grid,"read",{"open","seek+read","close"},us);
+  }
+#endif
+
   template<class word,class fobj>
   static inline void IOobject(word w,
 			      GridBase *grid,
@@ -302,6 +684,18 @@ class BinaryIO {
       lStart[d] = 0;
     }
 
+    //////////////////////////////////////////////////////////////////////////////
+    // Aggregate the lexicographic layout onto contiguous per rank extents
+    // ourselves rather than leaving it to MPI-IO collective buffering
+    //////////////////////////////////////////////////////////////////////////////
+    int aggregate = (control & BINARYIO_AGGREGATE)
+                 && (control & BINARYIO_LEXICOGRAPHIC)
+                 && !(control & BINARYIO_MASTER_APPEND)
+                 && (nrank > 1);
+#ifndef USE_MPI_IO
+    GRID_ASSERT(aggregate==0); // BINARYIO_AGGREGATE requires MPI
+#endif
+
 #ifdef USE_MPI_IO
     std::vector<int> distribs(ndim,MPI_DISTRIBUTE_BLOCK);
     std::vector<int> dargs   (ndim,MPI_DISTRIBUTE_DFLT_DARG);
@@ -329,6 +723,8 @@ class BinaryIO {
     ierr = MPI_Type_contiguous(numword,mpiword,&mpiObject);    GRID_ASSERT(ierr==0);
     ierr = MPI_Type_commit(&mpiObject);
 
+    // The subarray view is what aggregation exists to avoid; do not build it
+    if ( !aggregate ) {
     //////////////////////////////////////////////////////////////////////////////
     // File global array data type
     //////////////////////////////////////////////////////////////////////////////
@@ -340,6 +736,7 @@ class BinaryIO {
     //////////////////////////////////////////////////////////////////////////////
     ierr=MPI_Type_create_subarray(ndim,&lLattice[0],&lLattice[0],&lStart[0],MPI_ORDER_FORTRAN, mpiObject,&localArray);    GRID_ASSERT(ierr==0);
     ierr=MPI_Type_commit(&localArray);    GRID_ASSERT(ierr==0);
+    }
 #endif
 
     //////////////////////////////////////////////////////////////////////////////
@@ -358,7 +755,19 @@ class BinaryIO {
 
       timer.Start();
 
-      if ( (control & BINARYIO_LEXICOGRAPHIC) && (nrank > 1) ) {
+      if ( aggregate ) {
+#ifdef USE_MPI_IO
+        std::cout<< GridLogMessage<<"IOobject: aggregate read I/O "<< file<< std::endl;
+        AggregationPlan plan;
+        BuildAggregationPlan(grid,sizeof(fobj),plan);
+        ReportAggregationPlan(grid,plan,sizeof(fobj),"read");
+        std::vector<fobj> aggregated(lsites);
+        AggregateRead(grid,plan,aggregated,file,offset);
+        AggregateExchange(grid,plan,iodata,aggregated,0);
+#else
+        GRID_ASSERT(0);
+#endif
+      } else if ( (control & BINARYIO_LEXICOGRAPHIC) && (nrank > 1) ) {
 #ifdef USE_MPI_IO
 	std::cout<< GridLogMessage<<"IOobject: MPI read I/O "<< file<< std::endl;
 	ierr=MPI_File_open(grid->communicator,(char *) file.c_str(), MPI_MODE_RDONLY, MPI_INFO_NULL, &fh);    GRID_ASSERT(ierr==0);
@@ -416,7 +825,25 @@ class BinaryIO {
       grid->Barrier();
 
       timer.Start();
-      if ( (control & BINARYIO_LEXICOGRAPHIC) && (nrank > 1) ) {
+      if ( aggregate ) {
+#ifdef USE_MPI_IO
+        std::cout << GridLogMessage <<"IOobject: aggregate write I/O " << file << std::endl;
+        AggregationPlan plan;
+        BuildAggregationPlan(grid,sizeof(fobj),plan);
+        ReportAggregationPlan(grid,plan,sizeof(fobj),"write");
+        std::vector<fobj> aggregated(lsites);
+        AggregateExchange(grid,plan,iodata,aggregated,1);
+        AggregateWrite(grid,plan,aggregated,file,offset);
+        ////////////////////////////////////////////////////////////////////////
+        // Not every rank ends at the end of the payload, so the position can
+        // not be recovered from a file handle. Callers (Lime record chaining)
+        // rely on this being the first byte past the record.
+        ////////////////////////////////////////////////////////////////////////
+        offset = offset + (uint64_t)grid->_gsites*sizeof(fobj);
+#else
+        GRID_ASSERT(0);
+#endif
+      } else if ( (control & BINARYIO_LEXICOGRAPHIC) && (nrank > 1) ) {
 #ifdef USE_MPI_IO
         std::cout << GridLogMessage <<"IOobject: MPI write I/O " << file << std::endl;
         ierr = MPI_File_open(grid->communicator, (char *)file.c_str(), MPI_MODE_RDWR | MPI_MODE_CREATE, MPI_INFO_NULL, &fh);
@@ -461,12 +888,26 @@ class BinaryIO {
         
 	std::ofstream fout; 
 	fout.exceptions ( std::fstream::failbit | std::fstream::badbit );
+
+        ////////////////////////////////////////////////////////////////////
+        // Grid's model is that the boss rank performs the metadata
+        // operations and every other rank only seeks and writes into a file
+        // that already exists. Opening with ios::out on all ranks broke that:
+        // it is O_TRUNC, so a rank opening late truncated the file back to
+        // zero after an earlier rank had written its segment, leaving a hole
+        // in its place. The barriers around this block are outside it and do
+        // not order the opens against the writes. Let the boss create and
+        // empty the file, then everyone opens for update only. Same resulting
+        // length, one metadata operation instead of one per rank, no race.
+        ////////////////////////////////////////////////////////////////////
+        if ( !offset && grid->IsBoss() ) { // offset zero: this record starts the file
+          std::ofstream create(file,std::ios::binary|std::ios::out);
+          create.close();
+        }
+        grid->Barrier();
+
 	try {
-	  if (offset) { // Must already exist and contain data
-	    fout.open(file,std::ios::binary|std::ios::out|std::ios::in);
-	  } else {     // Allow create
-	    fout.open(file,std::ios::binary|std::ios::out);
-	  }
+	  fout.open(file,std::ios::binary|std::ios::out|std::ios::in);
 	} catch (const std::fstream::failure& exc) {
 	  std::cout << GridLogError << "Error in opening the file " << file << " for output" <<std::endl;
 	  std::cout << GridLogError << "Exception description: " << exc.what() << std::endl;
@@ -477,7 +918,7 @@ class BinaryIO {
 	  exit(1);
 #endif
 	}
-	
+
 	if ( control & BINARYIO_MASTER_APPEND )  {
 	  try {
 	    fout.seekp(0,fout.end);
@@ -548,7 +989,7 @@ class BinaryIO {
 				       uint32_t &nersc_csum,
 				       uint32_t &scidac_csuma,
 				       uint32_t &scidac_csumb,
-				       int control=BINARYIO_LEXICOGRAPHIC
+				       int control=BINARYIO_LEXICOGRAPHIC|BINARYIO_AGGREGATE
 				       )
   {
     typedef typename vobj::scalar_object sobj;
@@ -587,7 +1028,7 @@ class BinaryIO {
 					  uint32_t &nersc_csum,
 					  uint32_t &scidac_csuma,
 					  uint32_t &scidac_csumb,
-					  int control=BINARYIO_LEXICOGRAPHIC)
+					  int control=BINARYIO_LEXICOGRAPHIC|BINARYIO_AGGREGATE)
   {
     typedef typename vobj::scalar_object sobj;
     typedef typename vobj::Realified::scalar_type word;    word w=0;
@@ -674,7 +1115,7 @@ class BinaryIO {
     std::cout << GridLogMessage << "RNG read I/O on file " << file << std::endl;
 
     std::vector<RNGstate> iodata(lsites);
-    IOobject(w,grid,iodata,file,offset,format,BINARYIO_READ|BINARYIO_LEXICOGRAPHIC,
+    IOobject(w,grid,iodata,file,offset,format,BINARYIO_READ|BINARYIO_LEXICOGRAPHIC|BINARYIO_AGGREGATE,
 	     nersc_csum,scidac_csuma,scidac_csumb);
 
     timer.Start();
@@ -753,7 +1194,7 @@ class BinaryIO {
     });
     timer.Stop();
 
-    IOobject(w,grid,iodata,file,offset,format,BINARYIO_WRITE|BINARYIO_LEXICOGRAPHIC,
+    IOobject(w,grid,iodata,file,offset,format,BINARYIO_WRITE|BINARYIO_LEXICOGRAPHIC|BINARYIO_AGGREGATE,
 	     nersc_csum,scidac_csuma,scidac_csumb);
     iodata.resize(1);
     {

Grid/perfmon/PerfCount.h

@@ -93,8 +93,7 @@ inline uint64_t cyclecount(void){
 }
 #elif defined __x86_64__
 inline uint64_t cyclecount(void){ 
-  uint64_t ret = __rdtsc();
-  return (uint64_t)ret;
+  return (uint64_t)0;
 }
 #else
 

Grid/qcd/QCD.h

@@ -596,16 +596,32 @@ template<int Index,class vobj> inline vobj transposeColour(const vobj &lhs){
 //////////////////////////////////////////
 // Trace lattice and non-lattice
 //////////////////////////////////////////
+#define GRID_UNOP(name)   name
+#define GRID_DEF_UNOP(op, name)						\
+  template <typename T1, typename std::enable_if<is_lattice<T1>::value||is_lattice_expr<T1>::value,T1>::type * = nullptr> \
+  inline auto op(const T1 &arg) ->decltype(LatticeUnaryExpression<GRID_UNOP(name),T1>(GRID_UNOP(name)(), arg)) \
+  {									\
+    return     LatticeUnaryExpression<GRID_UNOP(name),T1>(GRID_UNOP(name)(), arg); \
+  }
+
 template<int Index,class vobj>
 inline auto traceSpin(const Lattice<vobj> &lhs) -> Lattice<decltype(traceIndex<SpinIndex>(vobj()))>
 {
   return traceIndex<SpinIndex>(lhs);
 }
+
+GridUnopClass(UnaryTraceSpin, traceIndex<SpinIndex>(a));
+GRID_DEF_UNOP(traceSpin, UnaryTraceSpin);
+
 template<int Index,class vobj>
 inline auto traceColour(const Lattice<vobj> &lhs) -> Lattice<decltype(traceIndex<ColourIndex>(vobj()))>
 {
   return traceIndex<ColourIndex>(lhs);
 }
+
+GridUnopClass(UnaryTraceColour, traceIndex<ColourIndex>(a));
+GRID_DEF_UNOP(traceColour, UnaryTraceColour);
+
 template<int Index,class vobj>
 inline auto traceSpin(const vobj &lhs) -> Lattice<decltype(traceIndex<SpinIndex>(lhs))>
 {
@@ -617,6 +633,8 @@ inline auto traceColour(const vobj &lhs) -> Lattice<decltype(traceIndex<ColourIn
   return traceIndex<ColourIndex>(lhs);
 }
 
+#undef GRID_UNOP
+#undef GRID_DEF_UNOP
 //////////////////////////////////////////
 // Current types
 //////////////////////////////////////////

Grid/qcd/modules/ObservableModules.h

@@ -103,6 +103,18 @@ class PolyakovMod: public ObservableModule<PolyakovLogger<Impl>, NoParameters>{
   PolyakovMod(): ObsBase(NoParameters()){}
 };
 
+template < class Impl >
+class SpatialPolyakovMod: public ObservableModule<SpatialPolyakovLogger<Impl>, NoParameters>{
+  typedef ObservableModule<SpatialPolyakovLogger<Impl>, NoParameters> ObsBase;
+  using ObsBase::ObsBase; // for constructors
+
+  // acquire resource
+  virtual void initialize(){
+    this->ObservablePtr.reset(new SpatialPolyakovLogger<Impl>());
+  }
+  public:
+  SpatialPolyakovMod(): ObsBase(NoParameters()){}
+};
 
 template < class Impl >
 class TopologicalChargeMod: public ObservableModule<TopologicalCharge<Impl>, TopologyObsParameters>{

Grid/qcd/observables/polyakov_loop.h

@@ -2,11 +2,12 @@
 
 Grid physics library, www.github.com/paboyle/Grid
 
-Source file: ./lib/qcd/modules/polyakov_line.h
+Source file: ./Grid/qcd/observables/polyakov_loop.h
 
-Copyright (C) 2017
+Copyright (C) 2025
 
 Author: David Preti <david.preti@csic.es>
+Author: Alexis Verney-Provatas <2414441@swansea.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
@@ -60,4 +61,43 @@ class PolyakovLogger : public HmcObservable<typename Impl::Field> {
   }
 };
 
+template <class Impl>
+class SpatialPolyakovLogger : public HmcObservable<typename Impl::Field> {
+ public:
+    // here forces the Impl to be of gauge fields
+    // if not the compiler will complain
+    INHERIT_GIMPL_TYPES(Impl);
+
+     // necessary for HmcObservable compatibility
+    typedef typename Impl::Field Field;
+
+    void TrajectoryComplete(int traj,
+                            Field &U,
+                            GridSerialRNG &sRNG,
+                            GridParallelRNG &pRNG) {
+
+    // Save current numerical output precision
+    int def_prec = std::cout.precision();
+
+    // Assume that the dimensions are D=3+1
+    int Ndim = 3;
+    ComplexD polyakov;
+   
+    // Iterate over the spatial directions and print the average spatial polyakov loop
+    // over them 
+    for (int idx=0; idx<Ndim; idx++) {
+        polyakov = WilsonLoops<Impl>::avgPolyakovLoop(U, idx);
+    
+        std::cout << GridLogMessage
+            << std::setprecision(std::numeric_limits<Real>::digits10 + 1)
+            << "Polyakov Loop in the " << idx << " spatial direction : [ " << traj << " ] "<< polyakov << std::endl;
+
+    }
+
+    // Return to original output precision
+    std::cout.precision(def_prec);
+
+  }
+};
+
 NAMESPACE_END(Grid);

Grid/qcd/utils/Sp2n.impl.h

@@ -254,9 +254,9 @@ static void testGenerators(GroupName::Sp) {
   }
 }
 
-template <int N>
-static Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > >
-ProjectOnGeneralGroup(const Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > &Umu, GroupName::Sp) {
+template <class vtype, int N>
+static Lattice<iScalar<iScalar<iMatrix<vtype, N> > > >
+ProjectOnGeneralGroup(const Lattice<iScalar<iScalar<iMatrix<vtype, N> > > > &Umu, GroupName::Sp) {
   return ProjectOnSpGroup(Umu);
 }
 

Grid/qcd/utils/WilsonLoops.h

@@ -177,25 +177,43 @@ public:
   }
 
   //////////////////////////////////////////////////
-  // average over all x,y,z the temporal loop
+  // average Polyakov loop in mu direction over all directions != mu
   //////////////////////////////////////////////////
-  static ComplexD avgPolyakovLoop(const GaugeField &Umu) {  //assume Nd=4
-    GaugeMat Ut(Umu.Grid()), P(Umu.Grid());
+  static ComplexD avgPolyakovLoop(const GaugeField &Umu, const int mu) {  //assume Nd=4
+    
+    // Protect against bad value of mu [0, 3]
+    if ((mu < 0 ) || (mu > 3)) {
+      std::cout << GridLogError << "Index is not an integer inclusively between 0 and 3." << std::endl;
+      exit(1);
+    }
+
+    // U_loop is U_{mu}
+    GaugeMat U_loop(Umu.Grid()), P(Umu.Grid());
     ComplexD out;
     int T = Umu.Grid()->GlobalDimensions()[3];
     int X = Umu.Grid()->GlobalDimensions()[0];
     int Y = Umu.Grid()->GlobalDimensions()[1];
     int Z = Umu.Grid()->GlobalDimensions()[2];
 
-    Ut = peekLorentz(Umu,3); //Select temporal direction
-    P = Ut;
-    for (int t=1;t<T;t++){ 
-      P = Gimpl::CovShiftForward(Ut,3,P);
+    // Number of sites in mu direction
+    int N_mu = Umu.Grid()->GlobalDimensions()[mu];
+
+    U_loop = peekLorentz(Umu, mu); //Select direction
+    P = U_loop;
+    for (int t=1;t<N_mu;t++){ 
+      P = Gimpl::CovShiftForward(U_loop,mu,P);
     }
    RealD norm = 1.0/(Nc*X*Y*Z*T);
    out = sum(trace(P))*norm;
    return out;   
-}
+  }  
+
+  /////////////////////////////////////////////////
+  // overload for temporal Polyakov loop
+  /////////////////////////////////////////////////
+  static ComplexD avgPolyakovLoop(const GaugeField &Umu) { 
+    return avgPolyakovLoop(Umu, 3);
+  }
 
   //////////////////////////////////////////////////
   // average over traced single links

Grid/simd/Simd.h

@@ -113,6 +113,14 @@ accelerator_inline RealD    adj(const RealD  & r){ return r; }
 accelerator_inline ComplexD adj(const ComplexD& r){ return(conjugate(r)); }
 accelerator_inline ComplexF adj(const ComplexF& r ){ return(conjugate(r)); }
 
+#if defined(GRID_CUDA) || defined(GRID_HIP)
+//Provide for convenience
+inline std::complex<double> conjugate(const std::complex<double>& r){ return(conj(r)); }
+inline std::complex<float>  conjugate(const std::complex<float>& r) { return(conj(r)); }
+inline std::complex<double> adj(const std::complex<double>& r)      { return(conj(r)); }
+inline std::complex<float>  adj(const std::complex<float>& r)       { return(conj(r)); }
+#endif
+
 accelerator_inline RealF real(const RealF  & r){ return r; }
 accelerator_inline RealD real(const RealD  & r){ return r; }
 accelerator_inline RealF real(const ComplexF  & r){ return r.real(); }

Grid/threads/Accelerator.cc

@@ -3,7 +3,7 @@
 NAMESPACE_BEGIN(Grid);
 int      world_rank; // Use to control world rank for print guarding
 int      acceleratorAbortOnGpuError=1;
-uint32_t accelerator_threads=2;
+uint32_t accelerator_threads=8;
 uint32_t acceleratorThreads(void)       {return accelerator_threads;};
 void     acceleratorThreads(uint32_t t) {accelerator_threads = t;};
 

Grid/threads/Accelerator.h

@@ -96,7 +96,9 @@ void     acceleratorInit(void);
 
 #ifdef GRID_CUDA
 
+NAMESPACE_END(Grid);
 #include <cuda.h>
+NAMESPACE_BEGIN(Grid);
 
 #ifdef __CUDA_ARCH__
 #define GRID_SIMT
@@ -432,22 +434,20 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
 
 #define accelerator_for2dNB( iter1, num1, iter2, num2, nsimd, ... )	\
   {									\
-    typedef uint64_t Iterator;						\
-    auto lambda = [=] accelerator					\
-      (Iterator iter1,Iterator iter2,Iterator lane ) mutable {		\
-      { __VA_ARGS__;}							\
-    };									\
-    int nt=acceleratorThreads();					\
-    dim3 hip_threads(nsimd, nt, 1);					 \
-    dim3 hip_blocks ((num1+nt-1)/nt,num2,1); \
-    if(hip_threads.x * hip_threads.y * hip_threads.z <= 64){ \
-      hipLaunchKernelGGL(LambdaApply64,hip_blocks,hip_threads,		\
-   	                 0,computeStream,						\
-			 num1,num2,nsimd, lambda);			\
-    } else { \
-      hipLaunchKernelGGL(LambdaApply,hip_blocks,hip_threads,		\
-			 0,computeStream,				\
-			 num1,num2,nsimd, lambda);			\
+    if (num1*num2) { \
+      typedef uint64_t Iterator;						\
+      auto lambda = [=] accelerator					\
+        (Iterator iter1,Iterator iter2,Iterator lane ) mutable {		\
+        { __VA_ARGS__;}							\
+      };									\
+      int nt=acceleratorThreads();					\
+      dim3 hip_threads(nsimd, nt, 1);					 \
+      dim3 hip_blocks ((num1+nt-1)/nt,num2,1); \
+      if(hip_threads.x * hip_threads.y * hip_threads.z <= 64){ \
+        LambdaApply64<<<hip_blocks,hip_threads,0,computeStream>>>(num1,num2,nsimd,lambda);			\
+      } else { \
+        LambdaApply<<<hip_blocks,hip_threads,0,computeStream>>>(num1,num2,nsimd,lambda);			\
+      } \
     } \
   }
 

HMC/FTHMC2p1f.cc

@@ -24,7 +24,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 #if Nc == 3
 #include <Grid/qcd/smearing/GaugeConfigurationMasked.h>
@@ -230,3 +234,4 @@ int main(int argc, char **argv)
 #endif
 } // main
 
+#endif

HMC/FTHMC2p1f_3GeV.cc

@@ -25,7 +25,11 @@ directory
 *************************************************************************************/
 /*  END LEGAL */
 
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 #if Nc == 3
 #include <Grid/qcd/smearing/GaugeConfigurationMasked.h>
@@ -231,5 +235,4 @@ int main(int argc, char **argv)
 #endif
 } // main
 
-
-
+#endif

HMC/HMC2p1f_3GeV.cc

@@ -24,7 +24,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 #if Nc == 3
 #include <Grid/qcd/smearing/GaugeConfigurationMasked.h>
@@ -230,5 +234,4 @@ int main(int argc, char **argv)
 #endif
 } // main
 
-
-
+#endif

HMC/Mobius2p1f.cc

@@ -27,7 +27,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 int main(int argc, char **argv) {
   using namespace Grid;
@@ -195,5 +199,4 @@ int main(int argc, char **argv) {
   Grid_finalize();
 } // main
 
-
-
+#endif

HMC/Mobius2p1fEOFA.cc

@@ -28,7 +28,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 #ifdef GRID_DEFAULT_PRECISION_DOUBLE
 #define MIXED_PRECISION
@@ -449,5 +453,4 @@ int main(int argc, char **argv) {
   Grid_finalize();
 } // main
 
-
-
+#endif

HMC/Mobius2p1fEOFA_F1.cc

@@ -28,7 +28,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 #ifdef GRID_DEFAULT_PRECISION_DOUBLE
 #define MIXED_PRECISION
@@ -442,5 +446,4 @@ int main(int argc, char **argv) {
   Grid_finalize();
 } // main
 
-
-
+#endif

HMC/Mobius2p1fIDSDRGparityEOFA_40ID.cc

@@ -28,7 +28,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 using namespace Grid;
 
@@ -918,3 +922,5 @@ int main(int argc, char **argv) {
   return 0;
 #endif
 } // main
+
+#endif

HMC/Mobius2p1fIDSDRGparityEOFA_48ID.cc

@@ -28,7 +28,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 using namespace Grid;
 
@@ -873,3 +877,5 @@ int main(int argc, char **argv) {
   return 0;
 #endif
 } // main
+
+#endif

HMC/Mobius2p1fRHMC.cc

@@ -27,7 +27,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 int main(int argc, char **argv) {
   using namespace Grid;
@@ -193,5 +197,4 @@ int main(int argc, char **argv) {
   Grid_finalize();
 } // main
 
-
-
+#endif

HMC/Mobius2p1f_DD_EOFA_96I_3level.cc

@@ -27,7 +27,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -512,5 +516,4 @@ int main(int argc, char **argv) {
   Grid_finalize();
 } // main
 
-
-
+#endif

HMC/Mobius2p1f_DD_EOFA_96I_double.cc

@@ -27,7 +27,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 int main(int argc, char **argv) {
   using namespace Grid;
@@ -345,5 +349,4 @@ int main(int argc, char **argv) {
   Grid_finalize();
 } // main
 
-
-
+#endif

HMC/Mobius2p1f_DD_EOFA_96I_mixed.cc

@@ -27,7 +27,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -516,5 +520,4 @@ int main(int argc, char **argv) {
   Grid_finalize();
 } // main
 
-
-
+#endif

HMC/Mobius2p1f_DD_EOFA_96I_mshift.cc

@@ -27,7 +27,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -567,5 +571,4 @@ int main(int argc, char **argv) {
   Grid_finalize();
 } // main
 
-
-
+#endif

HMC/Mobius2p1f_DD_RHMC.cc

@@ -27,7 +27,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 int main(int argc, char **argv) {
   using namespace Grid;
@@ -263,5 +267,4 @@ int main(int argc, char **argv) {
   Grid_finalize();
 } // main
 
-
-
+#endif

HMC/Mobius2p1f_DD_RHMC_96I.cc

@@ -27,7 +27,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 int main(int argc, char **argv) {
   using namespace Grid;
@@ -417,5 +421,4 @@ int main(int argc, char **argv) {
   Grid_finalize();
 } // main
 
-
-
+#endif

HMC/Mobius2p1f_DD_RHMC_96I_mixed.cc

@@ -27,7 +27,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -452,5 +456,4 @@ int main(int argc, char **argv) {
   Grid_finalize();
 } // main
 
-
-
+#endif

HMC/Mobius2p1f_EOFA_96I_hmc.cc

@@ -27,7 +27,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -462,5 +466,4 @@ int main(int argc, char **argv) {
   Grid_finalize();
 } // main
 
-
-
+#endif

HMC/Mobius2p1f_EOFA_96I_hmc_double.cc

@@ -27,7 +27,11 @@ See the full license in the file "LICENSE" in the top level distribution
 directory
 *************************************************************************************/
 /*  END LEGAL */
-#include <Grid/Grid.h>
+
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+#include<Grid/Grid.h>
 
 
 
@@ -264,5 +268,4 @@ int main(int argc, char **argv) {
   Grid_finalize();
 } // main
 
-
-
+#endif

HMC/disable_examples_without_instantiations.h

@@ -0,0 +1,16 @@
+#include <Grid/Grid.h>
+#pragma once
+
+
+#ifndef ENABLE_FERMION_INSTANTIATIONS
+#include <iostream>
+
+int main(void) {
+  std::cout << "This build of Grid was configured to exclude fermion instantiations, "
+	    << "which this example relies on. "
+	    << "Please reconfigure and rebuild Grid with --enable-fermion-instantiations"
+	    << "to run this example."
+	    << std::endl;
+  return 1;
+}
+#endif

benchmarks/Benchmark_ITT.cc

@@ -26,6 +26,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_benchmarks_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace Grid;
@@ -731,3 +734,5 @@ int main (int argc, char ** argv)
 
   Grid_finalize();
 }
+
+#endif

benchmarks/Benchmark_dwf.cc

@@ -20,6 +20,9 @@
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_benchmarks_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 #ifdef GRID_CUDA
 #define CUDA_PROFILE
@@ -439,3 +442,4 @@ void Benchmark(int Ls, Coordinate Dirichlet,bool sloppy)
   GRID_ASSERT(norm2(src_e)<1.0e-4);
   GRID_ASSERT(norm2(src_o)<1.0e-4);
 }
+#endif

benchmarks/Benchmark_dwf_fp32.cc

@@ -20,6 +20,10 @@
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_benchmarks_without_instantiations.h"
+
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 #ifdef GRID_CUDA
 #define CUDA_PROFILE
@@ -439,3 +443,5 @@ void Benchmark(int Ls, Coordinate Dirichlet,bool sloppy)
   GRID_ASSERT(norm2(src_e)<1.0e-4);
   GRID_ASSERT(norm2(src_o)<1.0e-4);
 }
+
+#endif

benchmarks/Benchmark_dwf_fp32_paranoid.cc

@@ -20,6 +20,9 @@
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_benchmarks_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 #ifdef GRID_CUDA
 #define CUDA_PROFILE
@@ -385,3 +388,5 @@ int main (int argc, char ** argv)
   Grid_finalize();
   exit(0);
 }
+
+#endif

benchmarks/Benchmark_dwf_sweep.cc

@@ -26,6 +26,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_benchmarks_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -238,5 +241,4 @@ void benchDw(std::vector<int> & latt4, int Ls, int threads,int report )
   }
 }
 
-
-
+#endif

benchmarks/Benchmark_gparity.cc

@@ -1,3 +1,7 @@
+#include "disable_benchmarks_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+
 #include <Grid/Grid.h>
 #include <sstream>
 using namespace std;
@@ -155,3 +159,4 @@ int main (int argc, char ** argv)
   Grid_finalize();
 }
 
+#endif

benchmarks/Benchmark_halo.cc

@@ -20,6 +20,9 @@
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_benchmarks_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 #ifdef GRID_CUDA
 #define CUDA_PROFILE
@@ -129,3 +132,5 @@ int main (int argc, char ** argv)
   Grid_finalize();
   exit(0);
 }
+
+#endif

benchmarks/Benchmark_mooee.cc

@@ -26,6 +26,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_benchmarks_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -149,3 +152,5 @@ int main (int argc, char ** argv)
 
   Grid_finalize();
 }
+
+#endif

benchmarks/Benchmark_schur.cc

@@ -26,6 +26,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_benchmarks_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -172,5 +175,4 @@ void benchDw(std::vector<int> & latt4, int Ls)
   //  Dw.Report();
 }
 
-
-
+#endif

benchmarks/Benchmark_staggered.cc

@@ -26,6 +26,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_benchmarks_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -110,3 +113,5 @@ int main (int argc, char ** argv)
 
   Grid_finalize();
 }
+
+#endif

benchmarks/Benchmark_staggeredF.cc

@@ -26,6 +26,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_benchmarks_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -112,3 +115,5 @@ int main (int argc, char ** argv)
 
   Grid_finalize();
 }
+
+#endif

benchmarks/Benchmark_usqcd.cc

@@ -26,6 +26,10 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_benchmarks_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
+
 #include <Grid/Grid.h>
 #include <Grid/algorithms/blas/BatchedBlas.h>
 
@@ -978,3 +982,5 @@ int main (int argc, char ** argv)
   Grid_finalize();
   fclose(FP);
 }
+
+#endif

benchmarks/Benchmark_wilson.cc

@@ -26,6 +26,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_benchmarks_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -258,3 +261,5 @@ int main (int argc, char ** argv)
 
   Grid_finalize();
 }
+
+#endif

benchmarks/Benchmark_wilson_sweep.cc

@@ -19,6 +19,9 @@ Author: Richard Rollins <rprollins@users.noreply.github.com>
     See the full license in the file "LICENSE" in the top level distribution directory
 *************************************************************************************/
 /*  END LEGAL */
+#include "disable_benchmarks_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -161,3 +164,5 @@ void bench_wilson_eo (
   double flops = (single_site_flops * volume * ncall)/2.0;
   std::cout << flops/(t1-t0) << "\t\t";
 }
+
+#endif

benchmarks/disable_benchmarks_without_instantiations.h

@@ -0,0 +1,16 @@
+#include <Grid/Grid.h>
+
+#pragma once
+
+#ifndef ENABLE_FERMION_INSTANTIATIONS
+#include <iostream>
+
+int main(void) {
+  std::cout << "This build of Grid was configured to exclude fermion instantiations, "
+	    << "which this benchmark relies on. "
+	    << "Please reconfigure and rebuild Grid with --enable-fermion-instantiations"
+	    << "to run this benchmark."
+	    << std::endl;
+  return 1;
+}
+#endif

configure.ac

@@ -172,6 +172,12 @@ case ${ac_TRACING} in
 esac
 
 ############### fermions
+AC_ARG_ENABLE([fermion-instantiations],
+     [AS_HELP_STRING([--enable-fermion-instantiations=yes|no],[enable fermion instantiations])],
+     [ac_FERMION_REPS=${enable_fermion_instantiations}], [ac_FERMION_INSTANTIATIONS=yes])
+
+AM_CONDITIONAL(BUILD_FERMION_INSTANTIATIONS, [ test "${ac_FERMION_INSTANTIATIONS}X" == "yesX" ])
+
 AC_ARG_ENABLE([fermion-reps],
      [AS_HELP_STRING([--enable-fermion-reps=yes|no],[enable extra fermion representation support])],
      [ac_FERMION_REPS=${enable_fermion_reps}], [ac_FERMION_REPS=yes])
@@ -194,6 +200,9 @@ AM_CONDITIONAL(BUILD_ZMOBIUS, [ test "${ac_ZMOBIUS}X" == "yesX" ])
 case ${ac_FERMION_REPS} in
    yes) AC_DEFINE([ENABLE_FERMION_REPS],[1],[non QCD fermion reps]);;
 esac
+case ${ac_FERMION_INSTANTIATIONS} in
+   yes) AC_DEFINE([ENABLE_FERMION_INSTANTIATIONS],[1],[enable fermions]);;
+esac
 case ${ac_GPARITY} in
    yes) AC_DEFINE([ENABLE_GPARITY],[1],[fermion actions with GPARITY BCs]);;
 esac

examples/Example_Mobius_spectrum.cc

@@ -3,6 +3,9 @@
  * without regression / tests being applied
  */
 
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -310,5 +313,4 @@ int main (int argc, char ** argv)
   Grid_finalize();
 }
 
-
-
+#endif

examples/Example_christoph.cc

@@ -3,6 +3,9 @@
  * without regression / tests being applied
  */
 
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -432,5 +435,4 @@ int main (int argc, char ** argv)
   Grid_finalize();
 }
 
-
-
+#endif

examples/Example_wall_wall_3pt.cc

@@ -3,6 +3,9 @@
  * without regression / tests being applied
  */
 
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -535,5 +538,4 @@ int main (int argc, char ** argv)
   Grid_finalize();
 }
 
-
-
+#endif

examples/Example_wall_wall_spectrum.cc

@@ -3,6 +3,9 @@
  * without regression / tests being applied
  */
 
+#include "disable_examples_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -429,5 +432,4 @@ int main (int argc, char ** argv)
   Grid_finalize();
 }
 
-
-
+#endif

examples/disable_examples_without_instantiations.h

@@ -0,0 +1,15 @@
+#include <Grid/Grid.h>
+#pragma once
+
+#ifndef ENABLE_FERMION_INSTANTIATIONS
+#include <iostream>
+
+int main(void) {
+  std::cout << "This build of Grid was configured to exclude fermion instantiations, "
+	    << "which this example relies on. "
+	    << "Please reconfigure and rebuild Grid with --enable-fermion-instantiations"
+	    << "to run this example."
+	    << std::endl;
+  return 1;
+}
+#endif

scripts/filelist

@@ -11,7 +11,7 @@ CCFILES=`find . -name '*.cc' -not -path '*/instantiation/*/*' -not -path '*/gamm
 
 ZWILS_FERMION_FILES=` find . -name '*.cc' -path '*/instantiation/*' -path '*/instantiation/ZWilsonImpl*' `
 WILS_FERMION_FILES=`  find . -name '*.cc' -path '*/instantiation/*' -path '*/instantiation/WilsonImpl*' `
-STAG_FERMION_FILES=`  find . -name '*.cc' -path '*/instantiation/*' -path '*/instantiation/Staggered*' `
+STAG_FERMION_FILES=`  find . -name '*.cc' -path '*/instantiation/*' -path '*/instantiation/StaggeredImpl*' `
 GP_FERMION_FILES=`    find . -name '*.cc' -path '*/instantiation/*' -path '*/instantiation/Gparity*' `
 ADJ_FERMION_FILES=`   find . -name '*.cc' -path '*/instantiation/*' -path '*/instantiation/WilsonAdj*' `
 TWOIND_FERMION_FILES=`find . -name '*.cc' -path '*/instantiation/*' -path '*/instantiation/WilsonTwoIndex*'`

skills/communication-overlap.md

@@ -0,0 +1,196 @@
+---
+name: communication-overlap
+description: Design and implement communication/computation overlap pipelines for GPU+MPI codes — per-packet event tracking, host-staging through pinned memory, internode/intranode bandwidth separation, and the 7-phase pipeline pattern that replaces broken accelerator-aware MPI paths.
+user-invocable: true
+allowed-tools:
+  - Read
+  - Bash(grep -r)
+---
+
+# Communication/Computation Overlap Pipeline Design
+
+## Why GPU-Direct MPI Is Often Not the Right Default
+
+GPU-direct RDMA (passing GPU buffer pointers directly to MPI) is appealing because it eliminates explicit D2H/H2D copies. In practice on several leadership systems:
+
+- **Bandwidth**: RDMA at 30% of wirespeed has been observed on Pontevecchio/Aurora. The overhead of staging through pinned host memory can be *lower* total latency than slow RDMA.
+- **Correctness**: Device buffer aliasing in `MPI_Sendrecv` (see `mpi-heterogeneous.md`) makes direct GPU-to-GPU transfer unreliable.
+- **Overlap**: Host-staging enables fine-grained overlap — each packet's D2H can be issued as a separate asynchronous event, and the corresponding MPI send can fire as soon as *that packet* arrives in host memory, not after all packets are ready.
+
+The pipeline pattern below was developed to replace broken MPICH accelerator-aware paths. It achieves genuine computation/communication overlap by tracking per-packet GPU events.
+
+## The 7-Phase Pipeline
+
+Given a set of halo exchange operations (each identified by a `packet_index`):
+
+### Phase 0: Prepare data on device
+Pack halo data into contiguous GPU buffers. One buffer per direction/neighbour.
+
+### Phase 1: Post receives + start D2H
+Post all `MPI_Irecv` calls immediately (into pinned host buffers). Simultaneously, start asynchronous D2H copies for all send buffers:
+
+```cpp
+for (auto &pkt : send_packets) {
+    MPI_Irecv(pkt.host_recv_buf, pkt.bytes, MPI_BYTE,
+               pkt.src_rank, pkt.tag, comm, &pkt.recv_req);
+    
+    acceleratorCopyFromDeviceAsync(pkt.device_send_buf,
+                                   pkt.host_send_buf,
+                                   pkt.bytes, &pkt.d2h_event);
+}
+```
+
+The key: `pkt.d2h_event` is a per-packet GPU event (e.g. `cudaEvent_t`, `hipEvent_t`, or SYCL event). We can poll individual packet completion rather than waiting for all.
+
+### Phase 2: Fire sends as D2H completes (packet by packet)
+Poll packet D2H events. As each packet becomes ready in host memory, immediately fire the corresponding `MPI_Isend`. Also start intranode D2D copies at this point — these are deferred until now to avoid competing with the internode D2H on PCIe bandwidth:
+
+```cpp
+bool all_sent = false;
+while (!all_sent) {
+    all_sent = true;
+    for (auto &pkt : send_packets) {
+        if (!pkt.sent && acceleratorEventIsComplete(pkt.d2h_event)) {
+            MPI_Isend(pkt.host_send_buf, pkt.bytes, MPI_BYTE,
+                       pkt.dst_rank, pkt.tag, comm, &pkt.send_req);
+            pkt.sent = true;
+            start_intranode_copy(pkt);  // now safe, D2H is done
+        }
+        if (!pkt.sent) all_sent = false;
+    }
+}
+```
+
+### Phase 3: Poll receives + start H2D as each arrives
+`MPI_Test` individual receive requests. As each completes, immediately start the H2D copy into device-resident halo buffer:
+
+```cpp
+bool all_recvd = false;
+while (!all_recvd) {
+    all_recvd = true;
+    for (auto &pkt : recv_packets) {
+        if (!pkt.h2d_started) {
+            int flag = 0;
+            MPI_Test(&pkt.recv_req, &flag, MPI_STATUS_IGNORE);
+            if (flag) {
+                acceleratorCopyToDeviceAsync(pkt.host_recv_buf,
+                                              pkt.device_recv_buf,
+                                              pkt.bytes, &pkt.h2d_event);
+                pkt.h2d_started = true;
+            }
+        }
+        if (!pkt.h2d_started) all_recvd = false;
+    }
+}
+```
+
+### Phase 4: Wait for all sends
+```cpp
+std::vector<MPI_Request> send_reqs;
+for (auto &pkt : send_packets) send_reqs.push_back(pkt.send_req);
+MPI_Waitall(send_reqs.size(), send_reqs.data(), MPI_STATUSES_IGNORE);
+```
+
+### Phase 5: Wait for all H2D copies
+```cpp
+for (auto &pkt : recv_packets) acceleratorEventWait(pkt.h2d_event);
+```
+
+### Phase 6: Run interior computation
+The interior (non-halo) computation can run from Phase 1 onwards, overlapped with all of the above:
+
+```cpp
+// Launched in Phase 1, runs in parallel with the pipeline
+accelerator_for(ss, interior_sites, ...) { compute_interior(ss); }
+```
+
+Synchronise with interior before using the full field:
+```cpp
+accelerator_barrier();  // interior kernel done
+// Halo H2D is also complete (Phase 5 above)
+// Now safe to use full field
+```
+
+## Per-Packet Event Tracking Data Structure
+
+```cpp
+struct Packet {
+    // Buffers
+    void *device_send_buf;
+    void *host_send_buf;    // pinned
+    void *device_recv_buf;
+    void *host_recv_buf;    // pinned
+    size_t bytes;
+    
+    // MPI
+    int src_rank, dst_rank, tag;
+    MPI_Request send_req, recv_req;
+    
+    // GPU events (one per packet, not one global barrier)
+    AcceleratorEvent d2h_event;
+    AcceleratorEvent h2d_event;
+    
+    // State flags
+    bool sent = false;
+    bool h2d_started = false;
+};
+```
+
+The critical design point: `d2h_event` and `h2d_event` are **per-packet**, not global. This allows the MPI send for packet 0 to fire while packet 1's D2H is still in progress.
+
+## Internode vs Intranode Separation
+
+PCIe (GPU-to-CPU) and NVLink/xGMI (GPU-to-GPU within a node) are separate bandwidth resources. They do not compete with each other, but they *do* compete with each other for transactions if both are active simultaneously.
+
+Strategy: complete all internode D2H copies first (to maximise NIC injection bandwidth), then start intranode D2D copies (which use NVLink/xGMI and do not contend with PCIe for internode traffic):
+
+```cpp
+// In Phase 2: start intranode D2D only after D2H is confirmed complete
+if (pkt.is_intranode && pkt.d2h_done) {
+    // Use peer access (cudaMemcpyPeerAsync / hipMemcpyPeerAsync)
+    // rather than staging through host for intranode
+    cudaMemcpyPeerAsync(pkt.peer_recv_buf, pkt.dst_device,
+                         pkt.device_send_buf, pkt.src_device,
+                         pkt.bytes, computeStream);
+}
+```
+
+Grid reference: `Grid/communicator/Communicator_mpi3.cc` — search for `NVLINK_GET` and `ACCELERATOR_AWARE_MPI` conditional blocks.
+
+## Pinned Memory Allocation
+
+All host staging buffers must be pinned (page-locked) for async D2H/H2D:
+
+```cpp
+// CUDA
+cudaMallocHost(&host_buf, bytes);
+cudaFreeHost(host_buf);
+
+// HIP  
+hipHostMalloc(&host_buf, bytes, hipHostMallocDefault);
+hipHostFree(host_buf);
+
+// SYCL
+host_buf = sycl::malloc_host(bytes, *queue);
+sycl::free(host_buf, *queue);
+```
+
+Pre-allocate at startup. Repeated `cudaMallocHost` in the hot path adds latency from the OS memory manager.
+
+## Checksumming in the Pipeline
+
+Insert checksum computation before D2H (on the GPU-resident data) and verification after H2D (on the received GPU-resident data). See `correctness-verification.md` for the checksum pattern. The salting (`packet_index + 1000 * tag`) detects packet transposition — critical for diagnosing MPI buffer aliasing bugs where two packets' contents are swapped.
+
+## Smoke Test for a New System
+
+Before running physics, validate the pipeline on a synthetic benchmark:
+
+```cpp
+// Send a buffer of known values, receive and check
+// Run at multiple message sizes: 4KB, 64KB, 1MB, 16MB
+// Run at multiple process counts: 2, 8, 64, 512
+// Verify checksums on every packet
+// Measure bandwidth: should be ≥ 80% of FDR/HDR/NDR peak for host-staged
+```
+
+Any bandwidth below 50% of theoretical, or any checksum failure, indicates a problem in the communication stack that must be resolved before production runs.

skills/compiler-validation.md

@@ -0,0 +1,154 @@
+---
+name: compiler-validation
+description: Identify GPU compiler code generation bugs, distinguish them from hardware and runtime bugs, construct minimal reproducers, and validate correctness of generated assembly for performance-critical HPC kernels.
+user-invocable: true
+allowed-tools:
+  - Read
+  - Bash(grep -r)
+  - Bash(objdump)
+---
+
+# Compiler Validation for GPU HPC Codes
+
+## Why Compiler Bugs Are Distinct
+
+Compiler bugs have a unique diagnostic signature: they produce *deterministically wrong* results. The same input always produces the same wrong output. This distinguishes them from:
+
+- Hardware bugs: usually stochastic (wrong answer sometimes, correct answer other times)
+- Runtime bugs (premature barrier, buffer aliasing): often stochastic or history-dependent
+- Race conditions: non-deterministic
+
+**The determinism test**: run the same kernel 100 times with the same input. If the wrong answer is always the same wrong answer, suspect the compiler.
+
+## The Minimal Reproducer Protocol
+
+When a kernel produces wrong results, isolate the compiler as quickly as possible:
+
+**Step 1: Eliminate the physics**. Reduce the failing kernel to the smallest possible computation that still exhibits the bug. Replace QCD fields with `double` arrays. Replace lattice operations with scalar arithmetic. The goal is a 20-line CUDA/HIP/SYCL file that any compiler engineer can compile and run.
+
+**Step 2: Binary search over optimisation levels**. Compile at `-O0` (or equivalent). If the answer becomes correct, the bug is in an optimisation pass. Then test `-O1`, `-O2`, `-O3` individually to find which optimisation level introduces the bug.
+
+```bash
+# HIP example
+hipcc -O0 minimal_repro.cc -o test_O0 && ./test_O0   # should be correct
+hipcc -O1 minimal_repro.cc -o test_O1 && ./test_O1   # compare
+hipcc -O2 minimal_repro.cc -o test_O2 && ./test_O2   # compare
+```
+
+**Step 3: Identify the optimisation pass**. For LLVM-based compilers (clang, hipcc, dpcpp, nvcc via ptxas):
+```bash
+# Disable individual optimisation passes:
+hipcc -O2 -mllvm -disable-loop-unrolling minimal_repro.cc -o test
+hipcc -O2 -fno-vectorize minimal_repro.cc -o test
+hipcc -O2 -fno-slp-vectorize minimal_repro.cc -o test
+```
+
+**Step 4: Inspect the generated code**. For CUDA/HIP, use `--generate-line-info` and `cuobjdump` or `roc-obj-extract` to get annotated assembly:
+```bash
+# CUDA
+nvcc -O2 --generate-line-info --keep minimal_repro.cu
+cuobjdump --dump-ptx minimal_repro.o
+# HIP/ROCm
+hipcc -O2 --save-temps minimal_repro.cc
+llvm-objdump -d minimal_repro.o
+# SYCL/DPC++
+icpx -O2 -fsycl -Xclang -ast-dump minimal_repro.cc 2>&1 | grep -A5 "suspicious_expr"
+```
+
+Look for: incorrect register spill/fill sequences, loop trip count miscalculation, vectorisation across iteration boundaries, incorrect address arithmetic.
+
+## Known Compiler Bug Patterns in GPU Code
+
+### Register Pressure / Spill Bugs
+High register usage forces spills to local memory. Some compiler versions generate incorrect spill/fill code — the value is written to local memory but a stale register value is read back instead of the spilled value.
+
+**Signature**: Wrong answer with high-register-count kernels; becomes correct when `--maxrregcount=N` forces lower register count (more spilling) or higher (`--maxrregcount=256`, fewer spills).
+
+**Diagnostic**: Check register usage:
+```bash
+nvcc -O2 --ptxas-options=-v minimal_repro.cu 2>&1 | grep "registers"
+hipcc -O2 --offload-arch=gfx90a --save-temps minimal_repro.cc
+llvm-mc --arch=amdgcn minimal_repro.s 2>&1 | grep "VGPRs"
+```
+
+### Vectorisation Across Loop Boundaries
+The compiler vectorises two successive loop iterations as a SIMD unit when they have a data dependency that the compiler has incorrectly determined does not exist.
+
+**Signature**: Wrong answer that becomes correct when the loop body is extracted to a non-inlined function (disabling auto-vectorisation across iterations).
+
+### Incorrect Constant Propagation
+The compiler evaluates a compile-time expression incorrectly, substituting a wrong constant. Common in template-heavy code where `sizeof(T)` or `alignof(T)` is used in arithmetic that the compiler folds at compile time.
+
+**Signature**: Wrong array index or wrong stride. Inspecting the generated assembly shows a literal constant where you expect a computed value.
+
+## Stress Patterns for Compiler Validation
+
+These patterns exercise the compiler in ways that commonly expose bugs:
+
+```cpp
+// 1. Aliased pointer write followed by immediate read
+// (tests correct handling of write-after-write in register allocation)
+__global__ void alias_stress(double *a, double *b, int n) {
+    int i = blockIdx.x * blockDim.x + threadIdx.x;
+    if (i < n) {
+        a[i] = a[i] * 2.0;
+        b[i] = a[i] + 1.0;  // must read the updated value, not the original
+    }
+}
+
+// 2. Mixed-precision accumulation
+// (tests correct type promotion in FMA sequences)
+__global__ void precision_stress(float *in, double *out, int n) {
+    double acc = 0.0;
+    for (int i = 0; i < n; i++) acc += (double)in[i];
+    *out = acc;
+}
+
+// 3. Large struct in shared memory
+// (tests alignment and offset calculation for non-power-of-2-sized objects)
+struct S { double x[3]; };  // sizeof = 24 bytes, not a power of 2
+__global__ void struct_stress(S *in, S *out, int n) {
+    extern __shared__ S smem[];
+    int tid = threadIdx.x;
+    smem[tid] = in[tid];
+    __syncthreads();
+    out[tid] = smem[(tid + 1) % blockDim.x];
+}
+```
+
+## Separating Compiler from Runtime/Hardware
+
+When results are deterministically wrong:
+
+| Test | Compiler bug | Runtime/hardware bug |
+|---|---|---|
+| Recompile at -O0 | Fixes it | No effect |
+| Run on CPU (host code equivalent) | Fixes it | No effect |
+| Reorder loop iterations | Changes wrong answer | No effect or different pattern |
+| Different compiler version | Fixes or changes wrong answer | No effect |
+| Different GPU of same model | Same wrong answer | Different or no error |
+| Different GPU model | Fixes it (ISA-specific codegen bug) | May or may not fix |
+
+## Reporting to Compiler Teams
+
+A compiler bug report needs:
+1. Minimal reproducer (< 50 lines)
+2. Compiler version (`hipcc --version`, `nvcc --version`, `icpx --version`)
+3. GPU model and driver version
+4. Exact wrong and correct answers (hexfloat for reproducibility)
+5. Which compile flags change the behaviour
+6. Generated assembly for the correct and incorrect variants
+
+File with: LLVM Bugzilla (for hipcc/clang/dpcpp backends), NVIDIA bug portal (nvcc/ptxas), or vendor-specific developer forum. The minimal reproducer is the single most important element — without it, compiler teams cannot prioritise.
+
+## Pragmatic In-Production Workaround
+
+When a compiler bug is confirmed but the fix is not yet available, the lowest-risk workaround is to mark the affected function with reduced optimisation:
+
+```cpp
+#pragma clang optimize off  // clang/hipcc/dpcpp
+void __attribute__((optimize("O0"))) affected_kernel_host_wrapper() { ... }
+// For device code, use per-file compilation flags via CMake/Makefile
+```
+
+Document the workaround with a comment referencing the compiler bug report number so it can be removed when the compiler is updated.

skills/correctness-verification.md

@@ -0,0 +1,169 @@
+---
+name: correctness-verification
+description: Implement application-level correctness verification for HPC codes on unreliable hardware — double-run pattern, deterministic reductions, per-packet checksums, and flight recorder step logging.
+user-invocable: true
+allowed-tools:
+  - Read
+  - Bash(grep -r)
+---
+
+# Correctness Verification Infrastructure for HPC Codes
+
+## The Problem
+
+Leadership computing facilities sometimes have hardware or firmware bugs below the level visible to application code. The accelerator runtime can return from `q.wait()` or `cudaDeviceSynchronize()` before work is actually complete, or silently produce wrong answers in DMA transfers. Standard testing does not catch these because they are non-deterministic and often topology-dependent (fail only at specific process counts or on specific node configurations).
+
+The symptoms look like numerical instabilities, random MPI hangs, or wrong physics results — not like crashes. Without deliberate infrastructure, diagnosing root cause takes months.
+
+## The Double-Run Pattern
+
+The most reliable correctness check for non-deterministic hardware bugs is to run every computation twice and compare bit-identical fingerprints.
+
+**Key constraint**: the second run must use a *deterministic* code path. Non-deterministic floating-point ordering (e.g. from MPI_Allreduce with different reduction trees on retry) produces false mismatches. See `mpi-heterogeneous.md` for how to make reductions deterministic.
+
+```cpp
+// Pseudocode: double-run a step and compare CRC fingerprints
+void run_step_verified(State &state) {
+    state.save_checkpoint();
+    
+    uint64_t crc_a = run_step_and_fingerprint(state);
+    state.restore_checkpoint();
+    uint64_t crc_b = run_step_and_fingerprint(state);
+    
+    if (crc_a != crc_b) {
+        report_mismatch("step", crc_a, crc_b);
+        // Policy: abort, retry from checkpoint, or continue with alarm
+    }
+}
+```
+
+**Fingerprinting**: XOR-fold a CRC32 over all floating-point data after each step. XOR is order-independent, so it works across distributed nodes without communication. For field data:
+
+```cpp
+uint64_t fingerprint(const double *data, size_t n) {
+    uint64_t acc = 0;
+    for (size_t i = 0; i < n; i++) {
+        uint64_t bits;
+        memcpy(&bits, &data[i], sizeof(bits));
+        acc ^= crc32(bits);
+    }
+    return acc;
+}
+```
+
+On GPU, compute the XOR reduction on-device (avoids D2H transfer of the full field):
+
+```cpp
+// SYCL
+uint64_t svm_xor(uint64_t *vec, uint64_t L) {
+    uint64_t ret = 0;
+    { sycl::buffer<uint64_t,1> abuff(&ret, {1});
+      theGridAccelerator->submit([&](sycl::handler &cgh) {
+        auto R = sycl::reduction(abuff, cgh, uint64_t(0), std::bit_xor<>());
+        cgh.parallel_for(sycl::range<1>{L}, R,
+          [=](sycl::id<1> i, auto &sum) { sum ^= vec[i]; });
+      }); }
+    theGridAccelerator->wait();
+    return ret;
+}
+```
+
+## Per-Packet Communication Checksums
+
+Silent data corruption in MPI buffers (documented in MPICH with device-resident buffers; see `mpi-heterogeneous.md`) requires per-packet verification, not just end-to-end. The pattern:
+
+1. Before packing a send buffer, compute a GPU-side checksum of the payload.
+2. Append the checksum to the host staging buffer alongside the data.
+3. After receiving and copying to device, recompute the checksum on-device and compare.
+
+Salt each checksum with `packet_index + 1000 * mpi_tag` to detect transposition (packet A landing in packet B's slot):
+
+```cpp
+uint64_t salt = (uint64_t)packet_index + 1000ULL * mpi_tag;
+checksum_send = checksum_gpu(payload_gpu, payload_words) ^ salt;
+// ... transmit payload + checksum_send ...
+checksum_recv = checksum_gpu(payload_gpu_recv, payload_words) ^ salt;
+assert(checksum_recv == checksum_send);
+```
+
+Grid reference: `Grid/communicator/Communicator_mpi3.cc`, `#ifdef GRID_CHECKSUM_COMMS`.
+
+## Flight Recorder: Step-Level Logging
+
+Maintain a monotonic counter that names the current operation. On a hang, this is the only way to know *which* operation the process is stuck in without a debugger.
+
+```cpp
+struct FlightRecorder {
+    std::atomic<uint64_t> step_counter{0};
+    const char *step_name = "init";
+    
+    void step_log(const char *name) {
+        step_name = name;
+        step_counter.fetch_add(1, std::memory_order_relaxed);
+    }
+};
+extern FlightRecorder gRecorder;
+```
+
+In Record mode, also store floating-point norms and communication checksums to vectors. In Verify mode, compare against stored values:
+
+```cpp
+void norm_log(double val) {
+    if (mode == Record) norm_log_vec.push_back(val);
+    if (mode == Verify) {
+        double expected = norm_log_vec[norm_counter];
+        if (val != expected) {  // bit-exact for deterministic paths
+            std::cerr << "MISMATCH at step " << step_counter
+                      << " (" << step_name << "): "
+                      << std::hexfloat << val << " vs " << expected << "\n";
+            print_backtrace();
+        }
+        norm_counter++;
+    }
+}
+```
+
+Grid reference: `Grid/util/FlightRecorder.h`, `Grid/util/FlightRecorder.cc`.
+
+## Signal Handler for Hang Detection
+
+Install a SIGHUP handler that dumps the current flight recorder state. This is async-safe only if the handler writes to a pre-allocated buffer using `write()` (not `printf`):
+
+```cpp
+static char hang_buf[4096];
+
+static void sighup_handler(int) {
+    int n = snprintf(hang_buf, sizeof(hang_buf),
+        "rank=%d step=%llu name=%s\n",
+        mpi_rank,
+        (unsigned long long)gRecorder.step_counter.load(),
+        gRecorder.step_name);
+    write(STDERR_FILENO, hang_buf, n);
+    // Optional: call backtrace_symbols_fd (async-safe on Linux)
+    void *frames[64];
+    int depth = backtrace(frames, 64);
+    backtrace_symbols_fd(frames, depth, STDERR_FILENO);
+}
+
+// In main():
+signal(SIGHUP, sighup_handler);
+```
+
+To diagnose a hang across all ranks: `kill -HUP $(pgrep my_app)` or via job scheduler.
+
+## What to Verify at Each Step
+
+| Data type | Fingerprint method | Frequency |
+|---|---|---|
+| Lattice fields | XOR of CRC32 over float64 words | Every algorithmic step |
+| Communication buffers | GPU XOR reduction, salted | Every MPI operation |
+| Scalar reductions | Bit-exact match of double | Every GlobalSum |
+| Iteration counters | Exact integer match | Every solver iteration |
+
+## When to Abort vs Continue
+
+- **Abort immediately**: communication checksum mismatch (data is corrupt, continuing will silently propagate errors).
+- **Log and continue**: norm mismatch in Verify mode if you need to map out which operations are unreliable.
+- **Retry from checkpoint**: double-run mismatch when the underlying bug is non-deterministic (second retry will usually pass).
+
+Track the mismatch rate over a production run. A rate above ~1/1000 steps indicates a systemic hardware issue that should be escalated to the facility.

skills/gpu-memory-performance.md

@@ -0,0 +1,205 @@
+---
+name: gpu-memory-performance
+description: Diagnose and fix GPU memory bandwidth and occupancy problems in Grid HPC kernels — acceleratorThreads() pitfalls, LambdaApply thread mapping, coalescedRead/Write idiom, when to use accelerator_for vs a hand-rolled __global__ kernel, and fused vs staged HBM access patterns.
+user-invocable: true
+allowed-tools:
+  - Read
+  - Bash(grep -r)
+---
+
+# GPU Memory Performance in Grid
+
+## Nsimd on GPU builds
+
+With `GEN_SIMD_WIDTH=64B` (the typical production setting), `Nsimd` is **not 1**:
+
+| Scalar type | `sizeof` | `Nsimd = 64 / sizeof` |
+|---|---|---|
+| `ComplexD` | 16 B | **4** |
+| `ComplexF` | 8 B  | **8** |
+| `RealD`    | 8 B  | **8** |
+| `RealF`    | 4 B  | **16** |
+
+So for `LatticePropagatorD` (scalar type `ComplexD`), `Nsimd=4` and the SIMD lane runs `threadIdx.x ∈ {0,1,2,3}`.  True `Nsimd=1` scalar-GPU builds are the exception, not the rule.
+
+## The acceleratorThreads() Trap
+
+`acceleratorThreads()` is a runtime-settable global (default **8**) that controls the `blockDim.y` of every `accelerator_for` launch.  It is NOT the SIMD width — it is the number of sites processed per block in the y-dimension.
+
+```cpp
+// Grid/threads/Accelerator.cc
+uint32_t accelerator_threads = 8;
+```
+
+With `accelerator_for(ss, osites, nsimd, ...)`, the launch is:
+
+```
+dim3 threads(nsimd, acceleratorThreads(), 1)
+dim3 blocks ((osites + acceleratorThreads() - 1) / acceleratorThreads(), 1, 1)
+```
+
+Total threads per block = `Nsimd × acceleratorThreads()`.  With `GEN_SIMD_WIDTH=64B` and `Nsimd=8` (fp32 / ComplexF):
+
+| `acceleratorThreads()` | threads/block | AMD wavefront | note |
+|---|---|---|---|
+| 2 (original default) | 16 | 25% | sub-wavefront, poor |
+| 4 | 32 | 50% | half wavefront |
+| **8 (current default)** | **64** | **100%** | **one full wavefront — Dslash sweet spot** |
+| 16 | 128 | 200% | two wavefronts/block; register-pressure cliff for heavy kernels |
+| 32 | 256 | 400% | severe register spill for stencil kernels |
+
+**Why 8 and not higher?**  Compute-heavy kernels like the Domain Wall Dslash carry many live registers per thread (spinors + gauge links + projections).  Doubling `acceleratorThreads` from 8→16 doubles the register demand per block, which on AMD GFX90A triggers a hard occupancy cliff: `Benchmark_dwf_fp32` drops from 1.7 TF/s (nt=8) to ~300 GF/s (nt=16).  The sweet spot is one full wavefront per block, which with `Nsimd=8` (fp32) means `nt=8`.
+
+For fp64 work (`Nsimd=4`), `nt=8` gives 32 threads = half a wavefront (AMD pads to 64 with idle lanes).  Kernels that are not register-limited (e.g. simple lattice arithmetic) can benefit from `--accelerator-threads 16` at runtime.  Reduction kernels bypass `acceleratorThreads()` entirely via `getNumBlocksAndThreads`.
+
+**Why was the default ever 2?**  Before the `threadIdx.x`/`threadIdx.y` remap (see LambdaApply section below), the site index lived in `threadIdx.x` — the fast, coalescing dimension.  Increasing `acceleratorThreads` widened the block in the *site* direction, so adjacent threads in a warp hit adjacent sites, each stride `sizeof(vobj)` apart in AoS memory — breaking coalescing.  On early NVIDIA ports with `Nsimd≈8`, `nt=2` gave 16 threads = 50% of a 32-thread warp; NVIDIA recovers this via multiple concurrent blocks per SM, so occupancy was barely tolerable.  AMD has no such multiplier when blocks are already sub-wavefront.  After the remap put the site index in `threadIdx.y` and the SIMD lane in `threadIdx.x`, coalescing became independent of `acceleratorThreads`, removing the constraint.
+
+**Diagnostic**: observed bandwidth << peak, kernel time >> expected from data volume.  Check with `--accelerator-threads 32` at runtime.  A large speedup confirms occupancy starvation.
+
+**Fix options** (in order of preference):
+1. Kernel needs its own thread count — use `getNumBlocksAndThreads` and launch a `__global__` kernel directly (see below).
+2. Temporarily acceptable: set `--accelerator-threads 32` at the application level.  Note this affects every `accelerator_for` site in the binary.
+
+## LambdaApply Thread Mapping
+
+`accelerator_for` and `accelerator_for2d` go through `LambdaApply`:
+
+```cpp
+// HIP/CUDA LambdaApply kernel:
+uint64_t x = threadIdx.y + blockDim.y * blockIdx.x;  // iter1 (site index)
+uint64_t y = threadIdx.z + blockDim.z * blockIdx.y;  // iter2
+uint64_t z = threadIdx.x;                             // lane (SIMD lane)
+Lambda(x, y, z);
+```
+
+`threadIdx.x` is the **fast** (lane) dimension — consecutive thread IDs within a warp/wavefront correspond to consecutive lane values on the **same** site, not consecutive sites.
+
+With `GEN_SIMD_WIDTH=64B` and `Nsimd=4` (PropagatorD), a 64-thread AMD wavefront contains 64/4 = 16 sites, each processed by 4 lanes.  Adjacent threads within the wavefront read different lanes of the same site — this is a broadcast pattern (hardware handles this efficiently), not a stride.  The stride between consecutive *sites* (`sizeof(vobj)`) only appears between groups of `Nsimd` threads, spaced `Nsimd` apart in threadIdx.x — not between adjacent threads within a warp.
+
+## coalescedRead / coalescedWrite
+
+These are Grid's canonical way to read/write one SIMD lane from a vector type inside a `GRID_SIMT` kernel:
+
+```cpp
+// accelerator_for(ss, osites, Nsimd, {
+//   lane = acceleratorSIMTlane(Nsimd) = threadIdx.x  ∈ {0..Nsimd-1}
+auto scalar_val = coalescedRead(field[ss]);       // extractLane(lane, field[ss])
+coalescedWrite(field[ss], scalar_val);             // insertLane(lane, field[ss], scalar_val)
+```
+
+For `vobj` aggregate types, `coalescedRead` calls `extractLane(lane, vobj)` which recurses through the tensor hierarchy and returns `vobj::scalar_object`.
+
+For `vsimd` (raw SIMD vector) types, it casts to `scalar_type*` and indexes with `lane`.
+
+## Coalescing the Iteration Structure
+
+For an AoS input array where each site is `words` 16-byte elements, adjacent threads reading the same site's consecutive words achieve coalesced access:
+
+```cpp
+// Good: k varies across threads in a block → consecutive 16-byte reads
+accelerator_for2d(k, R, ss, osites, Nsimd, {
+    coalescedWrite(out[ss]._internal[k],
+                   coalescedRead(idat[ss * words + base + k]));
+});
+// dim3(Nsimd, nt, 1): threadIdx.y = k (consecutive words, coalesced)
+//                      threadIdx.x = lane (SIMD sub-lane, coalesced for Nsimd>1)
+```
+
+```cpp
+// Bad: each thread reads all R words of its site serially
+accelerator_for(ss, osites, 1, {
+    Bundle b;
+    for (int k = 0; k < R; k++)
+        b._internal[k] = idat[ss * words + base + k];  // serial, not coalesced across threads
+    out[ss] = b;                                         // bulk struct write
+});
+```
+
+The bad pattern also accumulates a large struct in registers (192 bytes for R=12), increasing register pressure and reducing occupancy further.
+
+## When to Use a __global__ Kernel Instead of accelerator_for
+
+`accelerator_for` is correct for site-parallel work where `acceleratorThreads()` is tuned appropriately.  Use a direct `__global__` kernel when:
+
+- The kernel requires a **specific thread count** for correctness or performance (reductions, shared-memory algorithms).
+- The optimal thread count depends on `sizeof(sobj)` and `sharedMemPerBlock`, not on a runtime global.
+- You need the retirement-count pattern for cross-block final reduction.
+
+Pattern: use `getNumBlocksAndThreads` to pick `numThreads` and `numBlocks`:
+
+```cpp
+Integer numThreads, numBlocks;
+int ok = getNumBlocksAndThreads(n, sizeof(sobj), numThreads, numBlocks);
+// starts at warpSize (32/64), doubles while 2*threads*sizeof(sobj) < sharedMemPerBlock
+// gives 64–256 threads/block → correct occupancy independent of acceleratorThreads()
+Integer smemSize = numThreads * sizeof(sobj);
+myKernel<<<numBlocks, numThreads, smemSize, computeStream>>>(args...);
+```
+
+Grid's `reduceKernel` uses this pattern and achieves ~400 GB/s on MI250X.
+
+## Fused vs Staged HBM Access
+
+A staged pack+reduce reads the data **three times**:
+
+```
+pack kernel:   reads vobj array (N bytes), writes bundle buffer (N bytes)
+reduce kernel: reads bundle buffer (N bytes), writes tiny result buffer
+```
+
+Total HBM: 3N bytes for N bytes of useful input.
+
+A fused kernel reads the data **once**:
+
+```
+packReduceKernel: reads R words of vobj array (N bytes), reduces in-place
+```
+
+Total HBM: N bytes.  Register pressure increases (R words held per thread) but the 3× HBM saving dominates for large objects.
+
+The fused pattern in Grid's `sumD_gpu_reduce_words<R>`:
+
+```cpp
+template <int R, class vobj, class sobj, class Iterator>
+__device__ void packReduceBlocks(
+    const iScalar<typename vobj::vector_type> *idat,
+    sobj *g_odata, Iterator osites, int base, int words)
+{
+  // sobj = iVector<iScalar<scalarD>, R>  (R double-precision scalars per site)
+  constexpr Iterator nsimd = vobj::Nsimd();
+  ...
+  while (i < osites * nsimd) {
+    Iterator lane = i % nsimd;
+    Iterator ss   = i / nsimd;
+    sobj tmpD; zeroit(tmpD);
+    for (int k = 0; k < R; k++) {
+      auto w = extractLane(lane, idat[ss * words + base + k]);
+      iScalar<typename vobj::scalar_typeD> wd; wd = w;   // float→double promotion
+      tmpD._internal[k] = wd;
+    }
+    mySum += tmpD;
+    ...
+  }
+  reduceBlock(sdata, mySum, tid);
+}
+```
+
+Launched with `getNumBlocksAndThreads` → 128 threads/block for R=12 (`BundleScalarD`=192 B, sharedMem=64 KB) → correct occupancy without depending on `acceleratorThreads()`.
+
+## Observed Numbers on MI250X (32^4 LatticePropagatorD, Nsimd=4, GEN_SIMD_WIDTH=64B)
+
+| Configuration | pack µs/group | reduce µs/group | total µs | GB/s |
+|---|---|---|---|---|
+| acceleratorThreads=2 (8 threads/block), staged | 10,080 | 470 | 126,909 | 50 |
+| acceleratorThreads=16 (64 threads/block), staged | 342 | 310 | 8,251 | 297 |
+| acceleratorThreads=16, fused (128 threads/block via getNumBlocksAndThreads) | — | 349 | 4,584 | 546 |
+
+The fused kernel at 349 µs/group reads 201 MB at 576 GB/s — 36% of MI250X HBM peak.  The remaining gap from peak is the in-kernel serial loop over R=12 words and the 12 serial kernel launches.
+
+## Quick Checklist When a Kernel Is Slow
+
+1. **Check Nsimd**: `GEN_SIMD_WIDTH=64B` → Nsimd=4 (ComplexD), 8 (ComplexF).  Total threads/block = `Nsimd × acceleratorThreads()`.  With old default nt=2 and Nsimd=4: 8 threads = 12.5% of AMD wavefront.
+2. Check threads per block: for `accelerator_for` kernels use `--accelerator-threads 32` and measure; a large speedup confirms occupancy starvation.
+3. Check for bulk struct accumulation in registers (`Bundle b; for(...) b._internal[k] = ...;`).  Replace with per-element writes via `coalescedWrite`.
+4. Check for staged HBM access (pack → buffer → reduce).  Count the passes; fuse if ≥ 2 passes over the same data.
+5. For reduction kernels, always use `getNumBlocksAndThreads` rather than `accelerator_for` so thread count is independent of `acceleratorThreads()`.

skills/gpu-runtime-correctness.md

@@ -0,0 +1,101 @@
+---
+name: gpu-runtime-correctness
+description: Detect and work around GPU runtime correctness failures — premature completion signalling, infinite poll hangs, stale completion flags, and the double-wait diagnostic pattern. Covers CUDA, HIP/ROCm, and SYCL/Level Zero runtimes.
+user-invocable: true
+allowed-tools:
+  - Read
+  - Bash(grep -r)
+---
+
+# GPU Runtime Correctness
+
+## The Completion Signalling Problem
+
+GPU runtimes expose a synchronisation primitive — `cudaDeviceSynchronize()`, `hipDeviceSynchronize()`, `q.wait()` — that is supposed to block until all previously submitted GPU work is complete. On several production systems, this guarantee has been violated in two distinct ways:
+
+### Failure Mode A: Premature Return
+The wait returns before the GPU work is done. The subsequent CPU code reads stale data from the output buffer. This is the most dangerous failure because it looks like a numerical instability, not a crash. Results are wrong but the program exits normally.
+
+**Identifying Premature Return**: Insert a second, independent wait immediately after the first. If a second `q.wait()` "fixes" incorrect results that appeared with a single `q.wait()`, the first wait was returning prematurely.
+
+```cpp
+// Diagnostic version — if this stabilises results, you have premature return
+accelerator_barrier();  // first wait
+accelerator_barrier();  // second wait (diagnostic)
+```
+
+Production fix: submit a trivially cheap no-op kernel after the real work and wait for it. The no-op kernel cannot complete until all previous commands in the queue are done (command queue ordering guarantee), so waiting for the no-op is a stronger barrier than waiting for the queue itself:
+
+```cpp
+// Lightweight fence kernel
+template<class T>
+__global__ void noop_kernel(T *p) { if (threadIdx.x == 0) (void)(*p); }
+
+void strong_barrier(T *device_ptr) {
+    noop_kernel<<<1, 1, 0, computeStream>>>(device_ptr);
+    cudaStreamSynchronize(computeStream);  // wait for the no-op
+}
+```
+
+### Failure Mode B: Infinite Poll
+The wait enters a polling loop that never terminates. The process consumes 100% CPU in a runtime library. The GPU has either stopped signalling progress entirely, or the completion flag is in a memory region that has become incoherent.
+
+This is distinct from Failure Mode A: with premature return the CPU proceeds; with infinite poll the CPU is stuck.
+
+**Identifying Infinite Poll**: `top` shows the MPI rank at 100% CPU. `perf top -p PID` or `strace -p PID` shows the process burning cycles inside the GPU runtime library (e.g. `libze_intel_gpu.so`, `libamdhip64.so`).
+
+**Documented instances**: 
+- Intel Level Zero on Pontevecchio (Aurora): both premature return *and* infinite poll have been observed as independent bugs on the same system.
+- The two failure modes can co-exist and have overlapping symptoms at the application level.
+
+## Completion Signalling Architecture
+
+Understanding why these bugs happen requires knowing how completion signalling works:
+
+```
+GPU command processor
+  → signals completion by writing to a host-visible memory address
+  → CPU runtime polls that address (or uses OS event notification via ioctl)
+```
+
+A premature return means the memory write happened before the actual work completed (e.g. the signal is on a different command stream that has not been serialised with the work stream). An infinite poll means the memory write never happens (hardware or driver bug preventing the signal from being written).
+
+**Implication**: `accelerator_barrier()` is not an unconditional correctness guarantee on all production systems. Application-level verification (double-run, checksums) is necessary as a second line of defence.
+
+## The Double-Wait Pattern in Practice
+
+The double-wait is a pragmatic workaround when premature return is suspected but not yet confirmed. It adds latency but does not change correctness if the barrier is working properly, so it is safe to enable in production:
+
+```cpp
+#ifdef WORKAROUND_PREMATURE_BARRIER
+#define accelerator_barrier() do { \
+    real_accelerator_barrier();    \
+    real_accelerator_barrier();    \
+} while(0)
+#endif
+```
+
+Monitor whether this changes observed behaviour. If double-wait eliminates wrong answers, you have confirmed premature return. If it does not help but inserting a no-op kernel does, the issue is with the wait primitive specifically, not with the underlying completion signal.
+
+## SYCL/Level Zero Specifics
+
+Level Zero (the backend for Intel GPU runtimes) separates command submission from synchronisation. A `q.wait()` should wait for all previously submitted command lists to retire. Documented bugs include:
+
+- `q.wait()` returning before the associated fence in Level Zero has been signalled.
+- `q.wait()` entering an `ioctl(i915, I915_GEM_WAIT)` call that never returns (kernel driver bug, not runtime bug).
+
+The latter requires a node reboot and cannot be worked around in application code. Detect it by checking process state (`D` in `ps aux`) and the kernel function via `/proc/PID/wchan`.
+
+## Stream Ordering and Compute Streams
+
+All GPU work must be submitted to the *same* stream/queue if you rely on in-order execution guarantees. Mixing default stream and non-default streams invalidates ordering assumptions on some backends.
+
+Grid uses `computeStream` (CUDA/HIP) or `theGridAccelerator` (SYCL) consistently throughout. If mixing Grid with third-party GPU code, ensure the third-party code is directed to the same stream, or insert explicit inter-stream barriers.
+
+## Checklist for New GPU Code
+
+1. Every kernel launch is followed by an `accelerator_barrier()` before reading device-side output on the host.
+2. All device-to-host copies use an explicit stream synchronisation after the copy, not before.
+3. If results are non-deterministic across runs, insert a second barrier and observe whether reproducibility improves.
+4. For correctness-critical operations (reductions that will be compared against reference values), add the double-run checksum test from `correctness-verification.md`.
+5. If the process hangs at 100% CPU in a runtime library function, this is a driver/runtime bug — there is no application-level fix beyond scheduling a node reboot.

skills/hang-diagnosis.md

@@ -0,0 +1,102 @@
+---
+name: hang-diagnosis
+description: Diagnose and isolate process hangs on HPC systems — distinguishing kernel-level ioctl hangs, infinite poll loops, collective deadlocks, and GPU completion signalling failures using async-safe signal handlers and flight recorder step counters.
+user-invocable: true
+allowed-tools:
+  - Read
+  - Bash(grep -r)
+  - Bash(strace)
+  - Bash(gdb)
+---
+
+# Hang Diagnosis on HPC Systems
+
+## Taxonomy of Hangs
+
+Not all hangs are the same. Misidentifying the type leads to wrong mitigation. The four distinct classes encountered on production leadership systems:
+
+### 1. Kernel-level ioctl hang (never returns)
+The process is in `D` (uninterruptible sleep) state. `strace` shows it blocked in an `ioctl` syscall. The GPU device driver has entered an unrecoverable state.
+
+**Diagnosis**: `ps aux | grep D` — the process shows `D` state. `cat /proc/PID/wchan` shows `i915_gem_wait_for_error` or similar.
+
+**Resolution**: Only a driver reload or node reboot recovers it. Log the node identifier and request replacement from the facility scheduler.
+
+### 2. Infinite poll loop (`q.wait()` or `cudaDeviceSynchronize()` never returns)
+The process is in `R` (running) state, consuming 100% CPU. A polling loop inside the runtime is checking a completion flag that never becomes true, either because the hardware never sets it or because the flag is in a memory region not visible to the polling thread.
+
+**Diagnosis**: `top` shows the rank at 100% CPU. `strace -p PID` shows repeated `futex` or `read` syscalls with zero-length results, or no syscalls at all (pure spinloop). `perf top -p PID` shows the process burning cycles in a single tight loop in a runtime library (e.g., `ze_intel_gpu.so`).
+
+**Resolution**: The double-wait workaround — submit a trivially cheap kernel after the operation under test to act as a fence, then wait for the trivial kernel. See `gpu-runtime-correctness.md`.
+
+### 3. Collective deadlock
+One or more ranks are blocked in an MPI call, usually `MPI_Allreduce` or `MPI_Barrier`, while others are not. Root cause: a topology-dependent bug in the MPI library's collective algorithm where some ranks' contributions never arrive.
+
+**Diagnosis**: Flight recorder step logs show some ranks at step N (inside the collective) while others are at step N+1 or stuck at step N with different `step_name` strings. The hung ranks will show `D` or `S` state in `ps`.
+
+**Resolution**: Replace `MPI_Allreduce` with a deterministic point-to-point tree reduction. See `mpi-heterogeneous.md`.
+
+### 4. Premature return from wait (silent wrong answer, not a hang)
+The runtime returns from `q.wait()` before the GPU work is complete. The next operation reads stale data. This is not a hang — it manifests as a wrong answer or non-deterministic floating-point results. It is listed here because it is the most confusing failure mode: the code appears to run correctly and completes normally.
+
+**Diagnosis**: Double-run with checksum (see `correctness-verification.md`). Insert a second `q.wait()` after the first and observe if results become reproducible. If inserting the second wait "fixes" wrong answers, the first wait was returning prematurely.
+
+## Flight Recorder for Hang Localization
+
+The most important diagnostic tool is knowing *which operation* a process is in when it hangs. Maintain a named step counter:
+
+```cpp
+// Call at the start of every major operation
+FlightRecorder::StepLog("MPI_Allreduce::norm");
+// ... do the operation ...
+FlightRecorder::StepLog("MPI_Allreduce::done");
+```
+
+On SIGHUP, dump rank, step counter value, and step name to stderr in an async-safe manner:
+
+```cpp
+static void sighup_handler(int) {
+    char buf[256];
+    int n = snprintf(buf, sizeof(buf), "rank %d: step %llu '%s'\n",
+        comm_rank,
+        (unsigned long long)step_counter,
+        step_name);
+    write(2, buf, n);
+    // backtrace_symbols_fd is async-safe on Linux glibc
+    void *frames[32];
+    backtrace_symbols_fd(frames, backtrace(frames, 32), 2);
+}
+signal(SIGHUP, sighup_handler);
+```
+
+Broadcast SIGHUP to all ranks from outside the job:
+```bash
+# In a separate shell while the job is hung
+squeue --job $JOBID -o "%i %N" | awk '{print $2}' | \
+  xargs -I{} ssh {} "pkill -SIGHUP -f my_application"
+```
+
+The step names from all ranks will reveal which collective operation has diverged.
+
+## Distinguishing Driver Hang from MPI Hang
+
+| Symptom | Driver hang | MPI hang |
+|---|---|---|
+| Process state | `D` (ioctl) or `R` (spinloop) | `S` (blocked in syscall) |
+| `strace` | blocked `ioctl` or tight loop | blocked `recvmsg` / `read` |
+| Scope | single rank / single node | subset of ranks, pattern-dependent |
+| Recovery | reboot node | cancel job |
+| Flight recorder | step name is a GPU operation | step name is a collective |
+
+## Reducing Diagnostic Time
+
+1. **Name every collective operation** in the flight recorder before calling it.
+2. **Separate GPU work from MPI work** in the code so the step name unambiguously identifies which subsystem is hung.
+3. **Log node identifiers** alongside step names so flaky nodes can be identified and blacklisted.
+4. **Request flight recorder dumps from all ranks simultaneously** (SIGHUP broadcast) rather than attaching a debugger — attaching `gdb` to one rank of a hung MPI job usually deadlocks the debugger too.
+
+## What Not to Do
+
+- Do not `kill -9` a hung rank immediately — get the flight recorder dump first, otherwise diagnostic information is lost.
+- Do not assume the first rank that prints an error is the faulty one — collective hangs are frequently caused by the *last* rank to arrive at the barrier.
+- Do not use `MPI_Abort` in the hang handler — it may itself hang on some implementations. Use `_exit(1)` to force termination.

skills/mpi-heterogeneous.md

@@ -0,0 +1,182 @@
+---
+name: mpi-heterogeneous
+description: Diagnose and work around MPI correctness bugs on heterogeneous (CPU+GPU) systems — device buffer aliasing in MPI_Sendrecv, AARCH64 PLT corruption from libfabric, topology-dependent allreduce hangs, mixed-ABI HIP runtime from wrong GTL library (Frontier/ROCm), and deterministic point-to-point reduction trees as a replacement for MPI_Allreduce.
+user-invocable: true
+allowed-tools:
+  - Read
+  - Bash(grep -r)
+---
+
+# MPI Correctness on Heterogeneous HPC Systems
+
+## The Core Problem
+
+MPI libraries were designed for CPU-resident buffers. When GPU-resident buffers are passed directly (GPU-aware MPI / GPU direct RDMA), several correctness assumptions break:
+
+- **Buffer aliasing**: The MPI library may internally alias send/receive buffer addresses for `MPI_Sendrecv` in ways that are safe for CPU memory but wrong for GPU memory with different cache coherency rules.
+- **RDMA bandwidth**: GPU direct RDMA on some fabrics operates at a fraction of peak wirespeed (documented at ~30% on Pontevecchio/Aurora), making host-staging mandatory for performance even when correctness is not an issue.
+- **Collective tree topology**: `MPI_Allreduce` implementations may select reduction trees based on process count or communicator topology that expose rank-ordering bugs, causing hangs on some configurations but not others.
+
+## Bug Class 1: Device Buffer Aliasing in MPI_Sendrecv
+
+**Symptom**: `MPI_Sendrecv` with GPU-resident send and receive buffers produces wrong results. The received data matches neither the expected payload nor a host-staged copy. The failure is *deterministic* for a given problem size and process count, but *history-dependent* — earlier sends affect which alias is selected.
+
+**Root cause**: The MPI library internally reuses GPU buffer addresses for temporary staging without proper device memory ordering. When the same physical GPU memory pages appear in both the send and receive paths, writes from one path corrupt the other.
+
+**Diagnosis**: 
+1. Enable per-packet checksumming (see `correctness-verification.md`). If the checksum on the received packet does not match the sent checksum, the data was corrupted in transit.
+2. Replace `MPI_Sendrecv` with separate `MPI_Isend` + `MPI_Irecv` + `MPI_Waitall`. If this fixes the problem, the bug is in the `MPI_Sendrecv` implementation's internal buffer handling.
+3. Stage through host memory (`cudaMemcpy`/`hipMemcpy` to a host buffer, then `MPI_Sendrecv` on host buffers, then copy back). If this fixes the problem, confirms GPU-specific aliasing.
+
+**Reported as**: MPICH issue #7302. Affects MPICH on Intel Pontevecchio (Aurora) with device-resident buffers.
+
+**Workaround**: Do not use `MPI_Sendrecv` with GPU buffers. Use asynchronous send/receive pairs or host-staging. See `communication-overlap.md` for the full pipeline pattern.
+
+## Bug Class 2: PLT Corruption on AARCH64 (libfabric)
+
+**Symptom**: Application crashes or hangs on first `MPI_Comm_dup` call on AARCH64 systems (e.g. NVIDIA Grace/H200). Backtrace shows a bad instruction in the PLT (Procedure Linkage Table) for `MPI_Comm_dup` — specifically a `br x15` instruction that should instead be a proper trampoline.
+
+**Root cause**: `libfabric`'s memory registration cache monitor patches PLT entries at runtime to intercept memory allocation calls. Its AARCH64 trampoline generation writes an incorrect instruction sequence, leaving `br x15` (branch to whatever happens to be in x15) in the PLT entry. The next call through that PLT entry executes garbage.
+
+**Diagnosis**:
+```bash
+# Check if the PLT entry is corrupted
+objdump -d /proc/PID/exe | grep -A5 "MPI_Comm_dup@plt"
+# Look for "br x15" — this should be a proper stub, not a register branch
+```
+
+Or check the disassembly of the live process:
+```bash
+gdb -p PID -batch -ex "disassemble 'MPI_Comm_dup@plt'"
+```
+
+**Workaround**:
+```bash
+export FI_MR_CACHE_MONITOR=disabled
+```
+This prevents libfabric from patching PLT entries. It may reduce MR cache performance but restores correctness.
+
+**Reported as**: libfabric issue #11451. Affects systems using AARCH64 + libfabric OFI provider (Cray Slingshot, AWS EFA) with memory registration cache enabled.
+
+## Bug Class 3: Topology-Dependent Allreduce Hangs
+
+**Symptom**: `MPI_Allreduce` hangs indefinitely on some node configurations but completes correctly on others. The failure correlates with process count (e.g. fails at 512 ranks, works at 256) or network topology (fails when crossing specific router boundaries).
+
+**Root cause**: The MPI library's collective selection algorithm picks a reduction tree implementation that assumes symmetric participation from all ranks. A bug in one rank's contribution path (e.g. a GPU-side buffer not yet flushed when MPI reads it, due to premature barrier — see `gpu-runtime-correctness.md`) causes that rank to send wrong or incomplete data, and the tree-reduction protocol deadlocks waiting for data that never arrives correctly.
+
+**Diagnosis**: Flight recorder step logging (see `hang-diagnosis.md`). SIGHUP broadcast to all ranks. Ranks that are hung will show step name `MPI_Allreduce::...`; ranks that completed will show the next step. The hung ranks are the *recipients* of the stale data, not necessarily the *cause*.
+
+**Workaround — deterministic P2P reduction tree**:
+
+Replace `MPI_Allreduce` with an explicit point-to-point binary tree reduction. This is slower for large communicators but:
+1. Is immune to topology-dependent collective bugs.
+2. Is deterministic in floating-point ordering (the tree is fixed, not chosen at runtime).
+3. Makes the hang location explicit — each P2P operation is a named step in the flight recorder.
+
+```cpp
+// Binary tree reduction: rank 0 collects, then broadcasts
+void GlobalSumP2P(double *data, int count, MPI_Comm comm) {
+    int rank, size;
+    MPI_Comm_rank(comm, &rank); MPI_Comm_size(comm, &size);
+    
+    // Reduce phase: even ranks receive from odd neighbours
+    for (int stride = 1; stride < size; stride *= 2) {
+        if (rank % (2*stride) == 0) {
+            int partner = rank + stride;
+            if (partner < size) {
+                std::vector<double> tmp(count);
+                MPI_Recv(tmp.data(), count, MPI_DOUBLE, partner, 0, comm, MPI_STATUS_IGNORE);
+                for (int i = 0; i < count; i++) data[i] += tmp[i];
+            }
+        } else if (rank % stride == 0) {
+            int partner = rank - stride;
+            MPI_Send(data, count, MPI_DOUBLE, partner, 0, comm);
+            break;
+        }
+    }
+    // Broadcast phase
+    for (int stride = /* highest power of 2 <= size */; stride >= 1; stride /= 2) {
+        if (rank % (2*stride) == 0) {
+            int partner = rank + stride;
+            if (partner < size)
+                MPI_Send(data, count, MPI_DOUBLE, partner, 0, comm);
+        } else if (rank % stride == 0) {
+            int partner = rank - stride;
+            MPI_Recv(data, count, MPI_DOUBLE, partner, 0, comm, MPI_STATUS_IGNORE);
+        }
+    }
+}
+```
+
+Grid reference: `USE_GRID_REDUCTION` macro in `Grid/communicator/Communicator_mpi3.cc`.
+
+## Bug Class 4: Mixed HIP ABI from Wrong GTL Library (Frontier / ROCm)
+
+**Symptom**: `HIPFFT_PARSE_ERROR` (error code 12) returned by `hipfftPlanMany` / `hipfftMakePlanMany` / `hipfftPlan1d` for FFT sizes G < 32, but G ≥ 32 succeeds. The failure only occurs with an empty rocFFT kernel cache (`~/.cache/rocfft`); a warm cache may mask it. Host-side operations and GPU kernels that do not invoke rocFFT JIT work correctly.
+
+**Root cause — mixed HIP ABI**: rocFFT uses JIT compilation (via `libamd_comgr`) for small transforms (G < 32); for G ≥ 32 it uses pre-compiled device code bundled in the library, so the JIT path is never exercised. When two HIP runtime versions are loaded in the same process — e.g. `libamdhip64.so.7` (ROCm 7) and `libamdhip64.so.6` (ROCm 6) — the rocFFT JIT cannot complete successfully.
+
+The hidden source of the old library is the Cray MPI GPU Transport Layer. On Frontier, `cray-mpich`'s `libmpi_gtl_hsa.so` may be compiled against `libamdhip64.so.6` (ROCm 6 ABI) even when the loaded ROCm module is 7.0.2. Because `LD_LIBRARY_PATH` picks up the GTL directory before the ROCm 7 library directory, `libamdhip64.so.6` is pulled in first, and both ABI versions end up resident in the process.
+
+**Diagnosis**:
+```bash
+# Check which libamdhip64 versions are actually linked into your binary at runtime
+ldd --verbose ./your_binary 2>&1 | grep amdhip
+# Bad output — two different .so versions:
+#   libamdhip64.so.6 => /opt/rocm-6.4.2/lib/libamdhip64.so.6
+#   libamdhip64.so.7 => /opt/rocm-7.0.2/lib/libamdhip64.so.7
+# Good output — only one:
+#   libamdhip64.so.7 => /opt/rocm-7.0.2/lib/libamdhip64.so.7
+```
+
+If two versions appear, the problem is the GTL/LD_LIBRARY_PATH ordering.
+
+**Fix — correct module stack and LD_LIBRARY_PATH ordering (Frontier)**:
+```bash
+module load cce/21.0.0
+module load cpe/26.03
+module load rocm/7.0.2
+# Prepend CRAY_LD_LIBRARY_PATH so the ROCm-7-aware GTL is found first
+export LD_LIBRARY_PATH=$CRAY_LD_LIBRARY_PATH:$LD_LIBRARY_PATH
+# Ensure ROCm 7 LLVM libs (needed by libamd_comgr JIT) are on the path
+export LD_LIBRARY_PATH=/opt/rocm-7.0.2/lib/llvm/lib/:$LD_LIBRARY_PATH
+```
+
+The critical step is prepending `CRAY_LD_LIBRARY_PATH`: this ensures the GTL library built against the ROCm 7 ABI is resolved before any older version that may appear further down `LD_LIBRARY_PATH`. Without this step, a stale symlink or directory ordering can silently load the wrong `libmpi_gtl_hsa.so`.
+
+**Reproducer**: `tests/debug/Test_hipfft_repro.cc` — standalone hipFFT test (no Grid headers) that sweeps G and howmany values matching realistic Grid lattice geometries. Compile with:
+```bash
+hipcc -o Test_hipfft_repro Test_hipfft_repro.cc -lhipfft
+rm -rf ~/.cache/rocfft   # empty cache required to trigger JIT path
+./Test_hipfft_repro
+```
+
+**Reference**: `systems/WorkArounds.txt`, Frontier section — GPU mapping, XPMEM, and `FI_MR_CACHE_MONITOR=disabled` settings for Frontier are documented there.
+
+**Systems affected**: Frontier (ORNL, MI250X). Likely applies to any Cray PE system where the loaded `cray-mpich` GTL was compiled against an older ROCm ABI than the runtime ROCm module. LumiG (CSC, MI250X) uses the same Cray PE and may exhibit the same issue.
+
+## Compile-Time Guard Structure
+
+Recommended macro structure to switch between the workaround paths:
+
+```cpp
+// In configure / CMake, expose as options:
+// ACCELERATOR_AWARE_MPI  — use GPU direct (fast, potentially broken)
+// GRID_CHECKSUM_COMMS    — per-packet checksums (overhead: ~5%)
+// USE_GRID_REDUCTION     — P2P tree instead of MPI_Allreduce (slower, deterministic)
+// FI_MR_CACHE_MONITOR    — libfabric PLT workaround (env var, not compile-time)
+```
+
+On a known-good system, enable `ACCELERATOR_AWARE_MPI` and disable the others. On a system with known bugs, disable `ACCELERATOR_AWARE_MPI` and enable `GRID_CHECKSUM_COMMS` + `USE_GRID_REDUCTION` as needed.
+
+## Escalation Checklist
+
+Before concluding a bug is in your code:
+
+1. [ ] Can you reproduce with a minimal reproducer (two MPI ranks, no physics code)?
+2. [ ] Does the failure rate correlate with buffer size, process count, or network route?
+3. [ ] Does staging through host memory eliminate the failure?
+4. [ ] Is the failure deterministic for a given input (same answer, always wrong) or stochastic?
+5. [ ] Does the failure appear on a different MPI implementation (e.g. OpenMPI vs MPICH)?
+
+Deterministic wrong answers that reproduce with minimal reproducers and disappear with host-staging are strong evidence of an MPI library bug. File with the MPI library issue tracker with the minimal reproducer.

systems/Booster/config-command

@@ -8,7 +8,6 @@ LIME=/p/home/jusers/boyle2/juwels/gm2dwf/boyle/
     --disable-gparity \
     --disable-fermion-reps \
     --with-lime=$LIME \
-    --enable-accelerator-cshift \
     --disable-unified \
     CXX=nvcc \
     LDFLAGS="-cudart shared " \

systems/Frontier/aggregate_io_scaling.slurm

@@ -0,0 +1,104 @@
+#!/bin/bash -l
+
+# Node count scan.  Everything that could confound the comparison is held
+# fixed and only the number of Lustre clients changes:
+#
+#   local volume   8.8.32.128        = 262144 sites x 576 B = 151 MB / rank
+#   aggregation    k=2, row of 16, 8 extents of 18 MB        (identical at
+#                                                             every node count)
+#   old path       contiguous run = lLattice[0]*576 = 4.6 KB (identical too)
+#
+# So the ratio of contiguous extent between the two layout preserving paths is
+# a constant 4000x across the whole scan, and any change in their relative
+# bandwidth is a property of the client count alone.  That is the variable the
+# 4 node runs could not move.
+#
+# Trim the tail of the run list if you want a smaller allocation; the
+# geometries are independent.
+
+#SBATCH --job-name=ioScale
+#SBATCH --nodes=128
+#SBATCH --ntasks-per-node=8
+#SBATCH --cpus-per-task=7
+#SBATCH --gpus-per-node=8
+#SBATCH --time=02:00:00
+#SBATCH --account=phy157_dwf
+#SBATCH --gpu-bind=none
+#SBATCH --exclusive
+#SBATCH --mem=0
+
+cat << EOF > select_gpu
+#!/bin/bash
+export GPU_MAP=(0 1 2 3 7 6 5 4)
+export NUMA_MAP=(3 3 1 1 2 2 0 0)
+export GPU=\${GPU_MAP[\$SLURM_LOCALID]}
+export NUMA=\${NUMA_MAP[\$SLURM_LOCALID]}
+unset ROCR_VISIBLE_DEVICES
+export HIP_VISIBLE_DEVICES=\$GPU
+
+rank=\$SLURM_PROCID
+CMD="numactl -m \$NUMA -N \$NUMA  \$*"
+exec \$CMD
+EOF
+
+chmod +x ./select_gpu
+SELECT=$SLURM_SUBMIT_DIR/select_gpu
+
+root=$HOME/ParallelIO/systems/Frontier/
+source ${root}/sourceme-rocm7.2.sh
+
+export OMP_NUM_THREADS=7
+export MPICH_GPU_SUPPORT_ENABLED=1
+
+WORK=/lustre/orion/phy157/proj-shared/phy157_dwf/paboyle/aggscale.$SLURM_JOB_ID
+mkdir -p $WORK
+cd       $WORK
+
+# The largest rung writes three files of 154 GB, so budget ~700 GB for the
+# whole scan and check the quota before submitting.
+#
+# Do NOT stripe by default.  Measured on Frontier at 128 nodes / 1024 ranks,
+# 154.6 GB record, best of 3 (MB/s):
+#
+#                     default layout    lfs setstripe -c -1 -S 8M
+#   raw                    97673                81728
+#   MPI-IO lexicographic   31675                77534
+#   aggregate             189535                86063
+#
+# The two layout preserving paths want opposite things.  Wide striping is what
+# rescues MPI-IO at scale, 32 -> 78 GB/s, which is presumably what careful
+# striping has been buying in production.  It costs the aggregate path a factor
+# of 2.2, and costs raw as well.  With the default layout MPI-IO plateaus at
+# ~30 GB/s from 64 nodes upward -- the 20-40 GB/s seen in production -- while
+# the aggregate path keeps scaling.  Uncomment only to reproduce that.
+#lfs setstripe -c -1 -S 8M $WORK
+lfs getstripe -d $WORK      # record the inherited layout; needed to read the numbers
+
+BIN=$SLURM_SUBMIT_DIR/tests/IO/Test_aggregate_io
+ARGS="--aggregate-target 4194304 --io-no-correctness --io-reps 3"
+
+run () {   # run <nodes> <grid> <mpi> <comment>
+  local nodes=$1 nranks=$(( $1 * 8 ))
+  echo
+  echo "==================================================================="
+  echo "=== SCAN nodes=$nodes ranks=$nranks  grid=$2  mpi=$3   $4"
+  echo "==================================================================="
+  srun -N$nodes -n$nranks --ntasks-per-node=8 $SELECT $BIN --grid $2 --mpi $3 \
+       --accelerator-threads 8 --shm 2048 --device-mem 32000 $ARGS
+  echo "=== exit $?"
+}
+
+#            nodes  global lattice      decomposition   record
+run    4     32.32.64.128    4.4.2.1   "4.8 GB"
+run    8     32.32.64.256    4.4.2.2   "9.7 GB"
+run   16     32.32.64.512    4.4.2.4   "19.3 GB"
+run   32     32.32.128.512   4.4.4.4   "38.6 GB"
+run   64     32.32.128.1024  4.4.4.8   "77.3 GB"
+run  128     32.32.256.1024  4.4.8.8   "154.6 GB"
+
+echo
+echo "=== done.  Files left in $WORK"
+echo "=== For cold read numbers, submit a SECOND job on a fresh allocation"
+echo "=== with WORK pointed at this directory and --io-read-only added to"
+echo "=== ARGS.  Reading back what this job wrote only measures page cache."
+ls -l $WORK

systems/Frontier/config-command

@@ -13,8 +13,8 @@ CLIME=`spack find --paths c-lime@2-3-9 | grep c-lime| cut -c 15-`
 --with-mpfr=/opt/cray/pe/gcc/mpfr/3.1.4/ \
 --disable-fermion-reps \
 CXX=hipcc MPICXX=mpicxx \
-CXXFLAGS="-fPIC -I${ROCM_PATH}/include/ -I${MPICH_DIR}/include -L/lib64 " \
- LDFLAGS="-L/lib64 -L${ROCM_PATH}/lib -L${MPICH_DIR}/lib -lmpi -L${CRAY_MPICH_ROOTDIR}/gtl/lib -lmpi_gtl_hsa -lhipblas -lrocblas -lhipfft"
+CXXFLAGS="-fPIC -I${ROCM_PATH}/include/ -I${MPICH_DIR}/include " \
+ LDFLAGS="-L${ROCM_PATH}/lib -L${MPICH_DIR}/lib -lmpi -lmpi_gtl_hsa -lhipblas -lrocblas -lhipfft -lamdhip64"
 
 
 

systems/Frontier/sourceme-rocm6.4.sh

@@ -0,0 +1,14 @@
+
+echo spack
+. /autofs/nccs-svm1_home1/paboyle/spack/share/spack/setup-env.sh
+
+export CLIME=`spack find --paths c-lime | grep ^c-lime | awk '{print $2}' `
+export MPFR=`spack find --paths mpfr    | grep ^mpfr  | awk '{print $2}' `
+export OPENSSL=`spack find --paths openssl | grep openssl | awk  '{print $2}' `
+export GMP=`spack find --paths gmp      | grep ^gmp | awk '{print $2}' `
+
+module load cce/20.0.0
+module load cpe/25.09
+module load rocm/6.4.0
+export LD_LIBRARY_PATH=$CRAY_LD_LIBRARY_PATH:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/opt/rocm-6.4.0/lib/llvm/lib/:$LD_LIBRARY_PATH

systems/Frontier/sourceme-rocm7.0.sh

@@ -0,0 +1,10 @@
+
+echo spack
+. /autofs/nccs-svm1_home1/paboyle/spack/share/spack/setup-env.sh
+
+
+module load cce/21.0.0
+module load cpe/26.03
+module load rocm/7.0.2
+export LD_LIBRARY_PATH=$CRAY_LD_LIBRARY_PATH:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/opt/rocm-7.0.2/lib/llvm/lib/:$LD_LIBRARY_PATH

systems/Frontier/sourceme-rocm7.2.sh

@@ -0,0 +1,14 @@
+
+echo spack
+. /autofs/nccs-svm1_home1/paboyle/spack/share/spack/setup-env.sh
+
+export CLIME=`spack find --paths c-lime | grep ^c-lime | awk '{print $2}' `
+export MPFR=`spack find --paths mpfr    | grep ^mpfr  | awk '{print $2}' `
+export OPENSSL=`spack find --paths openssl | grep openssl | awk  '{print $2}' `
+export GMP=`spack find --paths gmp      | grep ^gmp | awk '{print $2}' `
+
+module load cce/21.0.0
+module load cpe/26.03
+module load rocm/7.2.0
+export LD_LIBRARY_PATH=$CRAY_LD_LIBRARY_PATH:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/opt/rocm-7.2.0/lib/llvm/lib/:$LD_LIBRARY_PATH

systems/Frontier/sourceme.sh

@@ -1,25 +1,10 @@
 
 echo spack
-. /autofs/nccs-svm1_home1/paboyle/Crusher/Grid/spack/share/spack/setup-env.sh
+. /autofs/nccs-svm1_home1/paboyle/spack/share/spack/setup-env.sh
 
-module load cce/15.0.1
-module load rocm/5.3.0
-module load cray-fftw
-module load craype-accel-amd-gfx90a
 
-#Ugly hacks to get down level software working on current system
-export LD_LIBRARY_PATH=/opt/cray/libfabric/1.20.1/lib64/:$LD_LIBRARY_PATH
-export LD_LIBRARY_PATH=/opt/gcc/mpfr/3.1.4/lib:$LD_LIBRARY_PATH
-export LD_LIBRARY_PATH=`pwd`/:$LD_LIBRARY_PATH
-ln -s /opt/rocm-6.0.0/lib/libamdhip64.so.6 .
-
-#echo spack load c-lime
-#spack load c-lime
-#module load emacs 
-##module load PrgEnv-gnu
-##module load cray-mpich
-##module load cray-fftw
-##module load craype-accel-amd-gfx90a
-##export LD_LIBRARY_PATH=/opt/gcc/mpfr/3.1.4/lib:$LD_LIBRARY_PATH
-#Hack for lib
-##export LD_LIBRARY_PATH=`pwd`/:$LD_LIBRARY_PATH
+module load cce/21.0.0
+module load cpe/26.03
+module load rocm/7.0.2
+export LD_LIBRARY_PATH=$CRAY_LD_LIBRARY_PATH:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/opt/rocm-7.0.2/lib/llvm/lib/:$LD_LIBRARY_PATH

systems/Perlmutter/config-command

@@ -1,17 +1,16 @@
 DIR=`pwd`
-PREFIX=$DIR/../Prequisites/install/
+
 ../../configure \
     --enable-comms=mpi \
     --enable-simd=GPU \
     --enable-shm=nvlink \
     --enable-gen-simd-width=64 \
+    --with-gmp=$GMP \
+    --with-mpfr=$MPFR \
     --enable-accelerator=cuda \
-    --enable-setdevice \
-    --disable-accelerator-cshift \
-    --with-gmp=$PREFIX \
     --disable-fermion-reps \
     --disable-unified \
     --disable-gparity \
     CXX=nvcc \
     LDFLAGS="-cudart shared " \
-    CXXFLAGS="-ccbin CC -gencode arch=compute_80,code=sm_80 -std=c++14 -cudart shared"
+    CXXFLAGS="-ccbin CC -gencode arch=compute_80,code=sm_80 -std=c++17 -cudart shared"

systems/Perlmutter/dwf384.slurm

@@ -0,0 +1,39 @@
+#!/bin/bash
+##SBATCH -A m5294_g
+#SBATCH -A mp13_g
+#m3886_g
+#SBATCH -C gpu
+#SBATCH -q premium
+#SBATCH -t 00:20
+#SBATCH -c 32
+#SBATCH -N 384
+#SBATCH -n 1536
+#SBATCH --ntasks-per-node=4
+#SBATCH --gpus-per-task=1
+#SBATCH --exclusive
+#SBATCH --gpu-bind=none
+
+export SLURM_CPU_BIND="cores"
+export MPICH_GPU_SUPPORT_ENABLED=1
+export MPICH_RDMA_ENABLED_CUDA=1
+export MPICH_GPU_IPC_ENABLED=1
+export MPICH_GPU_EAGER_REGISTER_HOST_MEM=0
+export MPICH_GPU_NO_ASYNC_MEMCPY=0
+#export MPICH_SMP_SINGLE_COPY_MODE=CMA
+
+cat << EOF > select_gpu
+#!/bin/bash
+export GPU_MAP=(0 1 2 3)
+export NUMA_MAP=( 0 1 2 3 )
+export GPU=\$SLURM_LOCALID
+export NUMA=\$SLURM_LOCALID
+export CUDA_VISIBLE_DEVICES=\$GPU
+exec numactl -m \$NUMA -N \$NUMA \$*
+EOF
+
+chmod +x ./select_gpu
+
+export VOL=128.128.128.288
+OPT="--comms-overlap --shm-mpi 0"
+srun ./select_gpu ./benchmarks/Benchmark_dwf_fp32 --mpi 4.8.4.12 --grid $VOL --device-mem 16000 --accelerator-threads 8 --shm 2048 $OPT
+

systems/Perlmutter/dwf4.slurm

@@ -1,5 +1,6 @@
 #!/bin/bash
-#SBATCH -A m3886_g
+#SBATCH -A m5294_g
+#m3886_g
 #SBATCH -C gpu
 #SBATCH -q debug
 #SBATCH -t 0:20:00
@@ -19,9 +20,21 @@ export MPICH_GPU_EAGER_REGISTER_HOST_MEM=0
 export MPICH_GPU_NO_ASYNC_MEMCPY=0
 #export MPICH_SMP_SINGLE_COPY_MODE=CMA
 
-OPT="--comms-sequential --shm-mpi 1"
-VOL=64.64.64.64
-srun ./benchmarks/Benchmark_dwf_fp32 --mpi 2.2.1.1 --grid $VOL --accelerator-threads 8 --shm 2048 $OPT
-#srun ./benchmarks/Benchmark_dwf_fp32 --mpi 2.1.1.4 --grid $VOL --accelerator-threads 8 --shm 2048 $OPT
-#srun ./benchmarks/Benchmark_dwf_fp32 --mpi 1.1.1.8 --grid $VOL --accelerator-threads 8 --shm 2048 $OPT
+cat << EOF > select_gpu
+#!/bin/bash
+export GPU_MAP=(0 1 2 3)
+export NUMA_MAP=( 0 1 2 3 )
+export GPU=\$SLURM_LOCALID
+export NUMA=\$SLURM_LOCALID
+export CUDA_VISIBLE_DEVICES=\$GPU
+exec numactl -m \$NUMA -N \$NUMA \$*
+EOF
+
+chmod +x ./select_gpu
+
+OPT="--comms-sequential --shm-mpi 0"
+VOL=64.64.32.32
+srun ./select_gpu ./benchmarks/Benchmark_dwf_fp32 --mpi 2.2.1.1 --grid $VOL --device-mem 16000 --accelerator-threads 8 --shm 2048 $OPT
+OPT="--comms-overlap --shm-mpi 0"
+srun ./select_gpu ./benchmarks/Benchmark_dwf_fp32 --mpi 2.2.1.1 --grid $VOL --device-mem 16000 --accelerator-threads 8 --shm 2048 $OPT
 

systems/Perlmutter/sourceme.sh

@@ -1,4 +1,6 @@
-
 export CRAY_ACCEL_TARGET=nvidia80
+source /global/homes/p/pboyle/spack/share/spack/setup-env.sh
+export MPFR=`spack find --paths mpfr | grep mpfr | cut -c 13-`
+export GMP=`spack find --paths gmp   | grep gmp  | cut -c 12-`
 
-module load PrgEnv-gnu cpe-cuda cudatoolkit/11.4
+module load PrgEnv-gnu cpe-cuda cudatoolkit/12.0

systems/Perlmutter/tests/cg32.slurm

@@ -0,0 +1,44 @@
+#!/bin/bash
+##SBATCH -A m5294_g
+#SBATCH -A mp13_g
+#m3886_g
+#SBATCH -C gpu
+#SBATCH -q premium
+#SBATCH -t 00:10
+#SBATCH -c 32
+#SBATCH -N 32
+#SBATCH -n 128
+#SBATCH --ntasks-per-node=4
+#SBATCH --gpus-per-task=1
+#SBATCH --exclusive
+#SBATCH --gpu-bind=none
+
+export SLURM_CPU_BIND="cores"
+export MPICH_GPU_SUPPORT_ENABLED=1
+export MPICH_RDMA_ENABLED_CUDA=1
+export MPICH_GPU_IPC_ENABLED=1
+export MPICH_GPU_EAGER_REGISTER_HOST_MEM=0
+export MPICH_GPU_NO_ASYNC_MEMCPY=0
+#export MPICH_SMP_SINGLE_COPY_MODE=CMA
+
+cat << EOF > select_gpu
+#!/bin/bash
+export GPU_MAP=(0 1 2 3)
+export NUMA_MAP=( 0 1 2 3 )
+export GPU=\$SLURM_LOCALID
+export NUMA=\$SLURM_LOCALID
+export CUDA_VISIBLE_DEVICES=\$GPU
+exec numactl -m \$NUMA -N \$NUMA \$*
+EOF
+
+chmod +x ./select_gpu
+
+OPT="--comms-overlap --shm-mpi 0"
+#
+# Local volume WAS 32.16.32.24
+#
+# 384 nodes
+#srun ./select_gpu ./Test_dwf_mixedcg_prec --seconds 300 --grid 128.128.128.288 --mpi 4.8.4.12 --device-mem 16000 --accelerator-threads 8 --shm 2048 $OPT > job.log
+# 32 nodes, same volume per node
+srun ./select_gpu ./Test_dwf_mixedcg_prec --seconds 300 --grid 64.32.64.96 --mpi 2.2.2.4 --device-mem 16000 --accelerator-threads 8 --shm 2048 $OPT > job.log
+

systems/Perlmutter/tests/cg384.slurm

@@ -0,0 +1,38 @@
+#!/bin/bash
+##SBATCH -A m5294_g
+#SBATCH -A mp13_g
+#m3886_g
+#SBATCH -C gpu
+#SBATCH -q premium
+#SBATCH -t 00:10
+#SBATCH -c 32
+#SBATCH -N 384
+#SBATCH -n 1536
+#SBATCH --ntasks-per-node=4
+#SBATCH --gpus-per-task=1
+#SBATCH --exclusive
+#SBATCH --gpu-bind=none
+
+export SLURM_CPU_BIND="cores"
+export MPICH_GPU_SUPPORT_ENABLED=1
+export MPICH_RDMA_ENABLED_CUDA=1
+export MPICH_GPU_IPC_ENABLED=1
+export MPICH_GPU_EAGER_REGISTER_HOST_MEM=0
+export MPICH_GPU_NO_ASYNC_MEMCPY=0
+#export MPICH_SMP_SINGLE_COPY_MODE=CMA
+
+cat << EOF > select_gpu
+#!/bin/bash
+export GPU_MAP=(0 1 2 3)
+export NUMA_MAP=( 0 1 2 3 )
+export GPU=\$SLURM_LOCALID
+export NUMA=\$SLURM_LOCALID
+export CUDA_VISIBLE_DEVICES=\$GPU
+exec numactl -m \$NUMA -N \$NUMA \$*
+EOF
+
+chmod +x ./select_gpu
+
+OPT="--comms-overlap --shm-mpi 0"
+srun ./select_gpu ./Test_dwf_mixedcg_prec --seconds 300 --grid 128.128.128.288 --mpi 4.8.4.12 --device-mem 16000 --accelerator-threads 8 --shm 2048 $OPT > job.log
+

systems/Perlmutter/usqcd4.slurm

@@ -0,0 +1,39 @@
+#!/bin/bash
+#SBATCH -A m5294_g
+#m3886_g
+#SBATCH -C gpu
+#SBATCH -q debug
+#SBATCH -t 0:30:00
+#SBATCH -c 32
+#SBATCH -N 4
+#SBATCH -n 16
+#SBATCH --ntasks-per-node=4
+#SBATCH --gpus-per-task=1
+#SBATCH --exclusive
+#SBATCH --gpu-bind=none
+
+export SLURM_CPU_BIND="cores"
+export MPICH_GPU_SUPPORT_ENABLED=1
+export MPICH_RDMA_ENABLED_CUDA=1
+export MPICH_GPU_IPC_ENABLED=1
+export MPICH_GPU_EAGER_REGISTER_HOST_MEM=0
+export MPICH_GPU_NO_ASYNC_MEMCPY=0
+#export MPICH_SMP_SINGLE_COPY_MODE=CMA
+
+cat << EOF > select_gpu
+#!/bin/bash
+export GPU_MAP=(0 1 2 3)
+export NUMA_MAP=( 0 1 2 3 )
+export GPU=\$SLURM_LOCALID
+export NUMA=\$SLURM_LOCALID
+export CUDA_VISIBLE_DEVICES=\$GPU
+exec numactl -m \$NUMA -N \$NUMA \$*
+EOF
+
+chmod +x ./select_gpu
+
+OPT="--comms-sequential --shm-mpi 0"
+VOL=64.64.32.32
+srun ./select_gpu ./benchmarks/Benchmark_usqcd --mpi 2.2.2.2 --grid $VOL --device-mem 16000 --accelerator-threads 8 --shm 2048 $OPT > usqcd.log
+
+

systems/mac-arm/config-command-mpi

@@ -3,7 +3,10 @@
 CXX=mpicxx ../../configure \
 	   --enable-simd=GEN \
 	   --enable-comms=mpi-auto \
-	   --enable-Sp=yes \
+	   --enable-Sp=no \
+	   --disable-fermion-reps \
+	   --disable-gparity \
+	   --with-fftw=$FFTW \
 	   --enable-unified=yes \
 	   --prefix /Users/peterboyle/QCD/vtk/Grid/install \
 	   --with-lime=$CLIME \

systems/mac-arm/sourceme.sh

@@ -0,0 +1,11 @@
+source /Users/peterboyle/QCD//Spack/spack//share/spack/setup-env.sh
+
+export FFTW=`spack find --paths fftw    | grep ^fftw   | awk '{print $2}' `
+#export HDF5=`spack find --paths hdf5+cxx   | grep ^hdf5   | awk '{print $2}' `
+export CLIME=`spack find --paths c-lime | grep ^c-lime | awk '{print $2}' `
+export MPFR=`spack find --paths mpfr    | grep ^mpfr  | awk '{print $2}' `
+export OPENSSL=`spack find --paths openssl | grep openssl | awk  '{print $2}' `
+export GMP=`spack find --paths gmp      | grep ^gmp | awk '{print $2}' `
+
+export LD_LIBRARY_PATH=$MPFR/lib:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=$GMP/lib:$LD_LIBRARY_PATH

tests/IO/Test_aggregate_io.cc

@@ -0,0 +1,458 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: ./tests/IO/Test_aggregate_io.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 */
+
+// Correctness and performance test for BINARYIO_AGGREGATE.
+//
+// Correctness, per aggregateTargetBytes:
+//   1. write via the MPI-IO lexicographic path      -> ref.bin
+//   2. write via the aggregate path                 -> agg.bin
+//   3. the two files must be byte identical         <- proves the layout matches
+//   4. read agg.bin back through the aggregate path <- proves the mirror inverts
+//   5. write and read the non-lexicographic path    -> raw.bin.  Its layout is
+//      different by construction (each rank owns one contiguous segment in
+//      rank order) so it cannot be compared byte for byte, but the NERSC and
+//      SciDAC checksums are computed from the global site index and are
+//      therefore layout independent: they must match the other two paths.
+//   6. a record written over a longer pre-existing file must leave the file at
+//      exactly offset+payload, with no trailing fragment of the old contents
+//
+// Performance: three paths, both directions, timed with the client page cache
+// dropped before every read so that a read back reports filesystem bandwidth
+// rather than memory bandwidth.  The non-lexicographic path is the zero
+// overhead reference: no transposition, no layout independence, one disjoint
+// contiguous segment per rank, which is the arrangement that reaches full
+// filesystem bandwidth on a leadership machine.  It is the upper bound the
+// other two are trying to approach.
+//
+// Options:
+//   --aggregate-target <bytes>   sweep this one target only (default: sweep
+//                                1, 1024, 64K, 4M)
+//   --io-reps <n>                repetitions in the performance section
+//                                (default 3; 0 disables it)
+//   --io-no-correctness          skip the correctness section, which reads the
+//                                whole file on one rank and is not affordable
+//                                at very large volume
+//   --io-read-only               time reads only, of files left in place by an
+//                                earlier job.  Reading back what this job just
+//                                wrote measures the client page cache; a fresh
+//                                allocation pointed at the same directory is
+//                                the only way to get a cold read without root.
+//
+// The exchange is only meaningfully exercised when the fast dimensions are
+// split across ranks; --mpi 1.1.X.Y leaves the rows of size one and the test
+// then passes vacuously.  Non-uniform AllToAllV counts additionally need an
+// odd process factor in a fast dimension and a small local volume.
+
+#include <Grid/Grid.h>
+#include <fcntl.h>
+
+using namespace Grid;
+
+/////////////////////////////////////////////////////////////////////////////
+// Compare in chunks.  Slurping both files into memory is fine for a few MB
+// and fatal for the multi-GB records this test is meant to reach.
+/////////////////////////////////////////////////////////////////////////////
+static bool FilesIdentical(std::string a,std::string b)
+{
+  std::ifstream fa(a,std::ios::binary), fb(b,std::ios::binary);
+  if ( !fa.good() || !fb.good() ) {
+    std::cout<<GridLogMessage<<"  could not open "<<a<<" and/or "<<b<<std::endl;
+    return false;
+  }
+  fa.seekg(0,std::ios::end); fb.seekg(0,std::ios::end);
+  uint64_t sa = (uint64_t)fa.tellg(), sb = (uint64_t)fb.tellg();
+  if ( sa != sb ) {
+    std::cout<<GridLogMessage<<"  size mismatch "<<sa<<" vs "<<sb<<std::endl;
+    return false;
+  }
+  fa.seekg(0,std::ios::beg); fb.seekg(0,std::ios::beg);
+
+  const uint64_t chunk = 8*1024*1024;
+  std::vector<char> va(chunk), vb(chunk);
+  uint64_t done=0;
+  while ( done < sa ) {
+    uint64_t n = std::min(chunk,sa-done);
+    fa.read(&va[0],n);
+    fb.read(&vb[0],n);
+    for(uint64_t i=0;i<n;i++){
+      if ( va[i]!=vb[i] ) {
+        std::cout<<GridLogMessage<<"  first differing byte at "<<done+i<<" of "<<sa<<std::endl;
+        return false;
+      }
+    }
+    done += n;
+  }
+  return true;
+}
+
+/////////////////////////////////////////////////////////////////////////////
+// Reading back a file we have just written measures the client page cache,
+// not the filesystem: the earlier runs of this test reported 8 GB/s on reads
+// and ~1 GB/s on writes for the same data.  POSIX_FADV_DONTNEED asks the
+// kernel to drop the cached pages for the file.  It is advisory and every
+// rank must do it, since each client caches independently, so treat this as
+// best effort rather than a guarantee of a cold read.
+/////////////////////////////////////////////////////////////////////////////
+static void DropCache(GridBase *grid,std::string file)
+{
+  grid->Barrier();
+  int fd = ::open(file.c_str(),O_RDONLY);
+  if ( fd >= 0 ) {
+#ifdef POSIX_FADV_DONTNEED
+    ::posix_fadvise(fd,0,0,POSIX_FADV_DONTNEED);
+#endif
+    ::close(fd);
+  }
+  grid->Barrier();
+}
+
+static uint64_t OptionU64(int argc,char **argv,const char *opt,uint64_t def)
+{
+  if ( GridCmdOptionExists(argv,argv+argc,opt) ) {
+    std::string arg = GridCmdOptionPayload(argv,argv+argc,opt);
+    return (uint64_t)std::stoull(arg);
+  }
+  return def;
+}
+
+int main(int argc,char **argv)
+{
+  Grid_init(&argc,&argv);
+
+  Coordinate latt = GridDefaultLatt();
+  Coordinate simd = GridDefaultSimd(Nd,vComplexD::Nsimd());
+  Coordinate mpi  = GridDefaultMpi();
+  GridCartesian grid(latt,simd,mpi);
+
+  typedef vLorentzColourMatrixD vobj;
+  typedef LorentzColourMatrixD  sobj;
+
+  GridParallelRNG pRNG(&grid);
+  pRNG.SeedFixedIntegers(std::vector<int>({1,2,3,4}));
+  LatticeGaugeFieldD Umu(&grid);
+  random(pRNG,Umu);
+
+  BinarySimpleMunger<sobj,sobj> munge;
+  const std::string format("IEEE64BIG");
+
+  const int lex = BinaryIO::BINARYIO_LEXICOGRAPHIC;
+  const int agg = BinaryIO::BINARYIO_LEXICOGRAPHIC|BinaryIO::BINARYIO_AGGREGATE;
+  const int raw = 0;   // no BINARYIO_LEXICOGRAPHIC: contiguous segment per rank
+
+  uint64_t payload = (uint64_t)grid._gsites*sizeof(sobj);
+
+  std::vector<uint64_t> targets = {1, 1024, 64*1024, 4*1024*1024};
+  if ( GridCmdOptionExists(argv,argv+argc,"--aggregate-target") ) {
+    targets.clear();
+    targets.push_back(OptionU64(argc,argv,"--aggregate-target",4*1024*1024));
+  }
+  uint64_t reps        = OptionU64(argc,argv,"--io-reps",3);
+  bool     correctness = !GridCmdOptionExists(argv,argv+argc,"--io-no-correctness");
+  // Read only: time reads of files left by an earlier job.  The only way to
+  // get a cold client cache without root is to read on an allocation that did
+  // not write the data, so run one job to write and a second, pointed at the
+  // same directory, with this flag.
+  bool     readonly    = GridCmdOptionExists(argv,argv+argc,"--io-read-only");
+  if ( readonly ) correctness = false;
+
+  std::cout<<GridLogMessage<<"Record payload "<<payload<<" bytes = "
+           <<payload/1024./1024.<<" MB, "
+           <<payload/(RealD)grid.ProcessorCount()/1024./1024.<<" MB/rank"<<std::endl;
+
+  int failures=0;
+
+  //////////////////////////////////////////////////////////////////////////
+  // Correctness
+  //////////////////////////////////////////////////////////////////////////
+  if ( correctness ) for(auto target : targets){
+
+    std::cout<<GridLogMessage<<"=== correctness, aggregateTargetBytes = "<<target<<" ==="<<std::endl;
+
+    uint32_t n1,a1,b1, n2,a2,b2, n3,a3,b3;
+    uint64_t off;
+
+    // Start from a clean slate.  The aggregate path sets the file length to
+    // exactly offset+payload; the MPI-IO path (MPI_MODE_CREATE) leaves any
+    // pre-existing tail in place.  Comparing stale files would therefore
+    // report a size mismatch that says nothing about the payload.
+    if ( grid.IsBoss() ) { ::unlink("ref.bin"); ::unlink("agg.bin"); ::unlink("raw.bin"); }
+    grid.Barrier();
+
+    off=0;
+    BinaryIO::writeLatticeObject<vobj,sobj>(Umu,"ref.bin",munge,off,format,n1,a1,b1,lex);
+
+    BinaryIO::aggregateTargetBytes = target;
+    off=0;
+    BinaryIO::writeLatticeObject<vobj,sobj>(Umu,"agg.bin",munge,off,format,n2,a2,b2,agg);
+
+    grid.Barrier();
+
+    if ( grid.IsBoss() ) {
+      if ( !FilesIdentical("ref.bin","agg.bin") ) {
+        std::cout<<GridLogError<<"  FAIL: aggregate file differs from lexicographic file"<<std::endl;
+        failures++;
+      } else {
+        std::cout<<GridLogMessage<<"  files byte identical"<<std::endl;
+      }
+    }
+    if ( (n1!=n2)||(a1!=a2)||(b1!=b2) ) {
+      std::cout<<GridLogError<<"  FAIL: checksum mismatch between paths"<<std::endl;
+      failures++;
+    }
+    // writeLatticeObject takes offset by value, so the out-parameter that
+    // IOobject sets never reaches us here and cannot be checked directly.
+    // The observable equivalent is the file length: both paths must leave the
+    // record ending at exactly offset+payload.
+    if ( grid.IsBoss() ) {
+      for(auto f : {std::string("ref.bin"),std::string("agg.bin")}){
+        std::ifstream fs(f,std::ios::binary|std::ios::ate);
+        uint64_t sz = (uint64_t)fs.tellg();
+        if ( sz != payload ) {
+          std::cout<<GridLogError<<"  FAIL: "<<f<<" is "<<sz<<" bytes, expected "<<payload<<std::endl;
+          failures++;
+        }
+      }
+    }
+
+    LatticeGaugeFieldD Uchk(&grid);
+    DropCache(&grid,"agg.bin");
+    off=0;
+    BinaryIO::readLatticeObject<vobj,sobj>(Uchk,"agg.bin",munge,off,format,n3,a3,b3,agg);
+    if ( (n3!=n1)||(a3!=a1)||(b3!=b1) ) {
+      std::cout<<GridLogError<<"  FAIL: read back checksum mismatch"<<std::endl;
+      failures++;
+    }
+    Uchk = Uchk - Umu;
+    RealD residual = norm2(Uchk);
+    std::cout<<GridLogMessage<<"  read back residual "<<residual<<std::endl;
+    if ( residual != 0.0 ) {
+      std::cout<<GridLogError<<"  FAIL: read back does not reproduce the field"<<std::endl;
+      failures++;
+    }
+
+    ////////////////////////////////////////////////////////////////////////
+    // Non-lexicographic.  Different file layout by construction, so compare
+    // by checksum and by round trip rather than by bytes.
+    ////////////////////////////////////////////////////////////////////////
+    uint32_t n4,a4,b4, n5,a5,b5;
+    off=0;
+    BinaryIO::writeLatticeObject<vobj,sobj>(Umu,"raw.bin",munge,off,format,n4,a4,b4,raw);
+    grid.Barrier();
+    if ( (n4!=n1)||(a4!=a1)||(b4!=b1) ) {
+      std::cout<<GridLogError<<"  FAIL: raw path checksum differs; the NERSC and"
+               <<" SciDAC checksums are layout independent and must agree"<<std::endl;
+      failures++;
+    }
+    DropCache(&grid,"raw.bin");
+    off=0;
+    BinaryIO::readLatticeObject<vobj,sobj>(Uchk,"raw.bin",munge,off,format,n5,a5,b5,raw);
+    if ( (n5!=n1)||(a5!=a1)||(b5!=b1) ) {
+      std::cout<<GridLogError<<"  FAIL: raw read back checksum mismatch"<<std::endl;
+      failures++;
+    }
+    Uchk = Uchk - Umu;
+    RealD rawresidual = norm2(Uchk);
+    std::cout<<GridLogMessage<<"  raw read back residual "<<rawresidual<<std::endl;
+    if ( rawresidual != 0.0 ) {
+      std::cout<<GridLogError<<"  FAIL: raw read back does not reproduce the field"<<std::endl;
+      failures++;
+    }
+  }
+
+  //////////////////////////////////////////////////////////////////////////
+  // Truncation.  offset!=0 is the case that matters: it is what ILDG and
+  // NERSC use, and it is the branch that used to only ever grow the file.
+  // The oversized starting file is made by extending a sparse one rather
+  // than writing padding from a single rank, which does not scale.
+  //////////////////////////////////////////////////////////////////////////
+  if ( correctness ) {
+    BinaryIO::aggregateTargetBytes = 4*1024*1024;
+    for(uint64_t testOffset : {(uint64_t)0, (uint64_t)1024}){
+
+      uint64_t expect = testOffset + payload;
+
+      if ( grid.IsBoss() ) {
+        { std::ofstream create("trunc.bin",std::ios::binary|std::ios::out); create.close(); }
+        int ierr = ::truncate("trunc.bin",(off_t)(expect+65536));
+        GRID_ASSERT(ierr==0);
+      }
+      grid.Barrier();
+
+      uint32_t n,a,b;
+      uint64_t off = testOffset;
+      BinaryIO::writeLatticeObject<vobj,sobj>(Umu,"trunc.bin",munge,off,format,n,a,b,agg);
+      grid.Barrier();
+
+      if ( grid.IsBoss() ) {
+        std::ifstream f("trunc.bin",std::ios::binary|std::ios::ate);
+        uint64_t sz = (uint64_t)f.tellg();
+        f.close();
+        if ( sz != expect ) {
+          std::cout<<GridLogError<<"  FAIL: offset "<<testOffset<<" left file "<<sz
+                   <<" bytes, expected "<<expect<<std::endl;
+          failures++;
+        } else {
+          std::cout<<GridLogMessage<<"  truncation ok at offset "<<testOffset
+                   <<": file is exactly "<<sz<<" bytes"<<std::endl;
+        }
+      }
+    }
+  }
+
+  //////////////////////////////////////////////////////////////////////////
+  // Performance.  Four numbers per repetition: write and read, old path and
+  // new.  Reads are preceded by a cache drop; writes are not, so a write
+  // number is "time to hand the data to the client cache and close", the
+  // same convention for both paths.
+  //////////////////////////////////////////////////////////////////////////
+  if ( reps ) {
+
+    LatticeGaugeFieldD Uio(&grid);
+    uint32_t n,a,b;
+
+    // The NERSC and SciDAC checksums are computed from the global site index,
+    // so all three layouts must produce the same values.  This costs nothing
+    // and is the only correctness check available at a volume where the byte
+    // for byte comparison (single rank, whole file) is unaffordable.
+    uint32_t cn[6],ca[6],cb[6];
+    auto agreeing = [&](const char *what,int lo,int hi){
+      for(int i=lo+1;i<=hi;i++){
+        if ( (cn[i]!=cn[lo])||(ca[i]!=ca[lo])||(cb[i]!=cb[lo]) ) {
+          std::cout<<GridLogError<<"  FAIL: "<<what<<" checksums disagree between paths"<<std::endl;
+          return false;
+        }
+      }
+      return true;
+    };
+
+    for(auto target : targets){
+
+      BinaryIO::aggregateTargetBytes = target;
+      std::cout<<GridLogMessage<<"=== performance, aggregateTargetBytes = "<<target<<" ==="<<std::endl;
+
+      std::vector<RealD> wref,wagg,wraw,rref,ragg,rraw;
+
+      for(uint64_t rep=0;rep<reps;rep++){
+
+        uint64_t off;
+
+        // Unlink only before the first repetition.  Lustre metadata cost is
+        // per file, not per byte, so rep 0 reports "create the file and write
+        // it" and the later reps report the steady state of overwriting an
+        // existing file -- which is what a multi record file does for every
+        // record after the first, and what production actually looks like.
+        if ( (rep==0) && !readonly && grid.IsBoss() ) {
+          ::unlink("ref.bin"); ::unlink("agg.bin"); ::unlink("raw.bin");
+        }
+        grid.Barrier();
+
+        if ( !readonly ) {
+
+        off=0;
+        BinaryIO::writeLatticeObject<vobj,sobj>(Umu,"ref.bin",munge,off,format,n,a,b,lex);
+        wref.push_back(BinaryIO::lastPerf.mbytesPerSecond);
+        cn[0]=n; ca[0]=a; cb[0]=b;
+
+        off=0;
+        BinaryIO::writeLatticeObject<vobj,sobj>(Umu,"agg.bin",munge,off,format,n,a,b,agg);
+        wagg.push_back(BinaryIO::lastPerf.mbytesPerSecond);
+        cn[1]=n; ca[1]=a; cb[1]=b;
+
+        off=0;
+        BinaryIO::writeLatticeObject<vobj,sobj>(Umu,"raw.bin",munge,off,format,n,a,b,raw);
+        wraw.push_back(BinaryIO::lastPerf.mbytesPerSecond);
+        cn[2]=n; ca[2]=a; cb[2]=b;
+        if ( !agreeing("write",0,2) ) failures++;
+
+        }  // !readonly
+
+        DropCache(&grid,"ref.bin");
+        off=0;
+        BinaryIO::readLatticeObject<vobj,sobj>(Uio,"ref.bin",munge,off,format,n,a,b,lex);
+        rref.push_back(BinaryIO::lastPerf.mbytesPerSecond);
+        cn[3]=n; ca[3]=a; cb[3]=b;
+
+        DropCache(&grid,"agg.bin");
+        off=0;
+        BinaryIO::readLatticeObject<vobj,sobj>(Uio,"agg.bin",munge,off,format,n,a,b,agg);
+        ragg.push_back(BinaryIO::lastPerf.mbytesPerSecond);
+        cn[4]=n; ca[4]=a; cb[4]=b;
+
+        DropCache(&grid,"raw.bin");
+        off=0;
+        BinaryIO::readLatticeObject<vobj,sobj>(Uio,"raw.bin",munge,off,format,n,a,b,raw);
+        rraw.push_back(BinaryIO::lastPerf.mbytesPerSecond);
+        cn[5]=n; ca[5]=a; cb[5]=b;
+        if ( !agreeing("read back",readonly?3:0,5) ) failures++;
+      }
+
+      if ( grid.IsBoss() ) {
+        auto report = [&](const char *name,std::vector<RealD> &v){
+          if ( v.empty() ) return;
+          RealD best=0, sum=0;
+          for(auto x : v){ if(x>best) best=x; sum+=x; }
+          // First sample includes file creation, later ones do not; quote both
+          // rather than a mean that mixes the two.
+          std::cout<<GridLogMessage<<"  PERF target="<<target<<" "<<name
+                   <<" best "<<best<<" MB/s, mean "<<sum/v.size()
+                   <<" MB/s, first(cold create) "<<v[0]<<" MB/s, samples";
+          for(auto x : v) std::cout<<" "<<x;
+          std::cout<<std::endl;
+        };
+        report("write raw       ",wraw);   // zero overhead reference
+        report("write MPI-IO    ",wref);
+        report("write aggregate ",wagg);
+        report("read  raw       ",rraw);
+        report("read  MPI-IO    ",rref);
+        report("read  aggregate ",ragg);
+
+        // Fraction of the zero overhead reference that each layout preserving
+        // path achieves.  This is the number the whole exercise is about.
+        auto best = [](std::vector<RealD> &v){ RealD m=0; for(auto x:v) if(x>m) m=x; return m; };
+        if ( !wraw.empty() && best(wraw) > 0 ) {
+          std::cout<<GridLogMessage<<"  PERF target="<<target
+                   <<" write fraction of raw: MPI-IO "<<best(wref)/best(wraw)
+                   <<"  aggregate "<<best(wagg)/best(wraw)<<std::endl;
+        }
+        if ( !rraw.empty() && best(rraw) > 0 ) {
+          std::cout<<GridLogMessage<<"  PERF target="<<target
+                   <<" read  fraction of raw: MPI-IO "<<best(rref)/best(rraw)
+                   <<"  aggregate "<<best(ragg)/best(rraw)<<std::endl;
+        }
+      }
+    }
+  }
+
+  if ( grid.IsBoss() ) {
+    if ( failures ) std::cout<<GridLogError  <<failures<<" FAILURE(S)"<<std::endl;
+    else            std::cout<<GridLogMessage<<"ALL AGGREGATE IO TESTS PASSED"<<std::endl;
+  }
+
+  Grid_finalize();
+  return failures!=0;
+}

tests/Test_cayley_even_odd_vec.cc

@@ -25,6 +25,9 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_tests_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -273,8 +276,6 @@ void  TestWhat(What & Ddwf,
 
   err = phi-chi;
   std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<< std::endl;
-
-  
 }
 
-
+#endif

tests/Test_compressed_lanczos_hot_start.cc

@@ -30,6 +30,9 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
  *  Reimplement the badly named "multigrid" lanczos as compressed Lanczos using the features 
  *  in Grid that were intended to be used to support blocked Aggregates, from
  */
+#include "disable_tests_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 #include <Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h>
 #include <Grid/algorithms/iterative/LocalCoherenceLanczos.h>
@@ -256,3 +259,4 @@ int main (int argc, char ** argv) {
   Grid_finalize();
 }
 
+#endif

tests/Test_dwf_dslash_repro.cc

@@ -25,6 +25,9 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_tests_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -237,3 +240,5 @@ int main (int argc, char ** argv)
   
   Grid_finalize();
 }
+
+#endif

tests/Test_dwf_mixedcg_prec.cc

@@ -25,6 +25,9 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_tests_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -222,3 +225,5 @@ int main (int argc, char ** argv)
   
   Grid_finalize();
 }
+
+#endif

tests/Test_dwf_mixedcg_prec_halfcomms.cc

@@ -25,6 +25,9 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
+#include "disable_tests_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
+
 #include <Grid/Grid.h>
 
 using namespace std;
@@ -118,3 +121,4 @@ int main (int argc, char ** argv)
   Grid_finalize();
 }
 #endif
+#endif

tests/Test_fft_memory.cc

@@ -0,0 +1,261 @@
+/*************************************************************************************
+
+ Test_fft_memory.cc
+
+ Memory growth test for PlannedFFT on a spin-colour matrix (propagator) field.
+
+ The test creates a single PlannedFFT object (which allocates FFTW plans once),
+ then repeatedly applies FFT_all_dim to the same propagator 400 times.
+
+ If PlannedFFT is working correctly the RSS should remain flat after the first
+ iteration — no new plans, no new deviceVector allocations beyond the per-call
+ pencil buffer which is freed at the end of each FFT_dim_execute call.
+
+ Build exactly like any other Grid test, e.g.:
+   make Test_fft_memory
+ or compile manually:
+   $(CXX) $(CXXFLAGS) Test_fft_memory.cc -o Test_fft_memory $(LDFLAGS)
+
+*************************************************************************************/
+
+#include <Grid/Grid.h>
+using namespace Grid;
+
+// --------------------------------------------------------------------------
+// Helper: read RSS (resident set size) in kB from /proc/self/status.
+// Returns 0 on platforms where /proc is unavailable.
+// --------------------------------------------------------------------------
+static long getCPURSSKb()
+{
+  long rss = 0;
+  FILE *fp = fopen("/proc/self/status", "r");
+  if (!fp) return -1;
+  char line[256];
+  while (fgets(line, sizeof(line), fp)) {
+    if (strncmp(line, "VmRSS:", 6) == 0) {
+      sscanf(line + 6, "%ld", &rss);
+      break;
+    }
+  }
+  fclose(fp);
+  return rss;
+}
+
+static long getGPUUsedMb()
+{
+#if defined(GRID_CUDA)
+  size_t free_bytes  = 0;
+  size_t total_bytes = 0;
+  cudaError_t err = cudaMemGetInfo(&free_bytes, &total_bytes);
+  if (err != cudaSuccess) return -1;
+  return (long)((total_bytes - free_bytes) / (1024 * 1024));
+ 
+#elif defined(GRID_HIP)
+  size_t free_bytes  = 0;
+  size_t total_bytes = 0;
+  hipError_t err = hipMemGetInfo(&free_bytes, &total_bytes);
+  if (err != hipSuccess) return -1;
+  return (long)((total_bytes - free_bytes) / (1024 * 1024));
+ 
+#else
+  return -1;   // CPU-only build: no GPU to query
+#endif
+}
+
+// ============================================================
+//  Convenience struct — one snapshot of both sides
+// ============================================================
+struct MemSnapshot {
+  long cpu_rss_kb;   // host RSS in kB  (-1 if unavailable)
+  long gpu_used_mb;  // device used in MB (-1 if no GPU)
+};
+ 
+static MemSnapshot takeSnapshot()
+{
+  MemSnapshot s;
+  s.cpu_rss_kb  = getCPURSSKb();
+  s.gpu_used_mb = getGPUUsedMb();
+  return s;
+}
+ 
+// ============================================================
+//  Pretty-print one row of the monitoring table
+// ============================================================
+static void printRow(int iter,
+                     const MemSnapshot &now,
+                     const MemSnapshot &prev)
+{
+  long cpu_delta = (now.cpu_rss_kb  >= 0 && prev.cpu_rss_kb  >= 0)
+                 ? now.cpu_rss_kb  - prev.cpu_rss_kb  : 0;
+  long gpu_delta = (now.gpu_used_mb >= 0 && prev.gpu_used_mb >= 0)
+                 ? now.gpu_used_mb - prev.gpu_used_mb : 0;
+ 
+  // Sign prefix so deltas are unambiguous
+  auto sign = [](long v) -> const char* { return v >= 0 ? "+" : ""; };
+ 
+  std::cout << GridLogMessage
+            << std::setw(6)  << iter
+            << "  CPU: " << std::setw(10) << now.cpu_rss_kb  << " kB"
+            << "  (" << sign(cpu_delta) << std::setw(7) << cpu_delta << " kB)"
+            << "  GPU: " << std::setw(7)  << now.gpu_used_mb << " MB"
+            << "  (" << sign(gpu_delta) << std::setw(5) << gpu_delta << " MB)"
+            << "\n";
+}
+
+// ============================================================
+
+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;
+
+  // ------------------------------------------------------------------
+  // Grid setup — use whatever lattice/mpi/simd was passed on the CLI,
+  // e.g.  --grid 8.8.8.8 --mpi 1.1.1.1
+  // ------------------------------------------------------------------
+  Coordinate latt_size   = GridDefaultLatt();
+  Coordinate simd_layout = GridDefaultSimd(Nd, vComplexD::Nsimd());
+  Coordinate mpi_layout  = GridDefaultMpi();
+
+  GridCartesian GRID(latt_size, simd_layout, mpi_layout);
+
+  int vol = 1;
+  for (int d = 0; d < (int)latt_size.size(); d++) vol *= latt_size[d];
+
+  std::cout << GridLogMessage << "Lattice : ";
+  for (int d = 0; d < Nd; d++) std::cout << latt_size[d] << " ";
+  std::cout << std::endl;
+
+  // ------------------------------------------------------------------
+  // Propagator field: SpinColourMatrix = 12x12 complex, i.e.
+  // LatticePropagatorD (= Lattice<iSpinColourMatrix<vComplexD>>).
+  // This is the standard QCD quark propagator type.
+  // ------------------------------------------------------------------
+  LatticePropagatorD prop(&GRID);
+  
+  // ------------------------------------------------------------------
+  // Fill the propagator with a momentum-space plane wave,
+  // following the pattern from Test_fft.cc.
+  // We set each spin-colour component (a,b) to exp(i * sum_mu p_mu x_mu)
+  // with a fixed momentum p = (1,2,1,2).
+  // ------------------------------------------------------------------
+  Coordinate pvec({1, 2, 1, 2});
+
+  LatticeComplexD phase(&GRID);
+  LatticeComplexD coor(&GRID);
+  ComplexD ci(0.0, 1.0);
+
+  phase = Zero();
+  for (int mu = 0; mu < Nd; mu++) {
+    RealD TwoPiL = M_PI * 2.0 / latt_size[mu];
+    LatticeCoordinate(coor, mu);
+    phase = phase + (TwoPiL * pvec[mu]) * coor;
+  }
+  phase = exp(phase * ci);   // e^{i p.x}
+
+  // Broadcast the phase into every spin-colour matrix entry
+  prop = Zero();
+  prop = prop + phase;
+
+  std::cout << GridLogMessage
+            << "Propagator norm2 = " << norm2(prop) << std::endl;
+
+  // ------------------------------------------------------------------
+  // Baseline snapshot BEFORE PlannedFFT construction
+  // ------------------------------------------------------------------
+  MemSnapshot snap_before_plan = takeSnapshot();
+  std::cout << GridLogMessage
+            << "[mem] Before PlannedFFT construction"
+            << "  CPU: " << snap_before_plan.cpu_rss_kb  << " kB"
+            << "  GPU: " << snap_before_plan.gpu_used_mb << " MB"
+            << std::endl;
+ 
+  // ------------------------------------------------------------------
+  // Create the PlannedFFT — plans are allocated here ONCE for all
+  // dimensions and stored inside the object.
+  // ------------------------------------------------------------------
+  PlannedFFT<iSpinColourMatrix<vComplexD>> plannedFFT(&GRID);
+
+  // ------------------------------------------------------------------
+  // Snapshot AFTER plan construction — this is the true baseline
+  // for the loop, because cufftPlanMany itself grabs device memory.
+  // ------------------------------------------------------------------
+  MemSnapshot snap_after_plan = takeSnapshot();
+  std::cout << GridLogMessage
+            << "[mem] After  PlannedFFT construction"
+            << "  CPU: " << snap_after_plan.cpu_rss_kb  << " kB"
+            << "  GPU: " << snap_after_plan.gpu_used_mb << " MB"
+            << "  (plan overhead:"
+            << "  CPU +" << snap_after_plan.cpu_rss_kb - snap_before_plan.cpu_rss_kb << " kB"
+            << "  GPU +" << snap_after_plan.gpu_used_mb - snap_before_plan.gpu_used_mb << " MB)"
+            << std::endl;
+
+  MemoryManager::Print();
+  // ------------------------------------------------------------------
+  // 400-iteration loop.
+  // Each iteration computes the full 4d forward FFT of `prop`.
+  // We deliberately do NOT cache the result — we always start from
+  // the same `prop` so the FFT is recomputed identically each time.
+  // The point is to watch memory, not correctness.
+  // ------------------------------------------------------------------
+  const int Niter = 40;
+  const int Niter2 = 32;
+
+  // Print header for the memory table
+  std::cout << GridLogMessage
+              << "\n"
+              << std::setw(6)  << "iter"
+              << "  CPU: " << std::setw(10) << "RSS[kB]"
+              << "  (  delta  )"
+              << "  GPU: " << std::setw(7)  << "used[MB]"
+              << "  (delta)"
+              << "\n";
+ 
+  MemSnapshot snap_prev = snap_after_plan;
+
+  for (int i = 0; i < Niter; i++) {
+    std::vector<LatticePropagatorD> G;
+
+    for (int j = 0; j < Niter2; j++) {
+      LatticePropagatorD prop_fft(&GRID);
+
+      // Full 4d forward FFT using the pre-built plans
+      plannedFFT.FFT_all_dim(prop_fft, prop, FFT::forward);
+
+      G.push_back(prop_fft);
+    }
+ 
+    // cudaMemGetInfo reflects the state *after* any pooled frees have
+    // been committed, so this is accurate without an explicit sync —
+    // FFT_dim_execute already calls accelerator_barrier() internally.
+    MemSnapshot snap_now = takeSnapshot();
+    printRow(i, snap_now, snap_prev);
+    MemoryManager::Print();
+    snap_prev = snap_now;
+  }
+
+  // ------------------------------------------------------------------
+  // Summary
+  // ------------------------------------------------------------------
+  MemSnapshot snap_final = takeSnapshot();
+ 
+  long cpu_growth = snap_final.cpu_rss_kb  - snap_after_plan.cpu_rss_kb;
+  long gpu_growth = snap_final.gpu_used_mb - snap_after_plan.gpu_used_mb;
+ 
+  std::cout << GridLogMessage
+            << "\n==== Memory summary (baseline = after plan construction) ====\n"
+            << "  CPU RSS growth over " << Niter << " FFTs : "
+            << cpu_growth << " kB"
+            << (cpu_growth == 0 ? "  OK" : "  *** GROWING ***") << "\n"
+            << "  GPU used growth over " << Niter << " FFTs : "
+            << gpu_growth << " MB"
+            << (gpu_growth == 0 ? "  OK" : "  *** GROWING ***") << "\n"
+            << "  Note: first-call watermark from pool fill is expected and benign.\n"
+            << "        A leak shows as continuous growth beyond iter ~2-3.\n";
+ 
+  Grid_finalize();
+  return 0;
+}

tests/Test_meson_field.cc

@@ -24,6 +24,8 @@ with this program; if not, write to the Free Software Foundation, Inc.,
 
 See the full license in the file "LICENSE" in the top level distribution directory
 *************************************************************************************/
+#include "disable_tests_without_instantiations.h"
+#ifdef ENABLE_FERMION_INSTANTIATIONS
 
 #include <Grid/Grid.h>
 #include <Grid/qcd/utils/A2Autils.h>
@@ -157,3 +159,5 @@ int main(int argc, char *argv[])
   
   return EXIT_SUCCESS;
 }
+
+#endif

tests/core/Test_planned_fft.cc

@@ -0,0 +1,321 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: ./tests/core/Test_planned_fft.cc
+
+    Copyright (C) 2015
+
+Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+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/Grid.h>
+
+using namespace Grid;
+
+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;
+
+  Coordinate latt_size   = GridDefaultLatt();
+  Coordinate simd_layout = GridDefaultSimd(Nd,vComplexD::Nsimd());
+  Coordinate mpi_layout  = GridDefaultMpi();
+
+  int vol = 1;
+  for(int d=0;d<latt_size.size();d++) vol *= latt_size[d];
+
+  GridCartesian         GRID(latt_size,simd_layout,mpi_layout);
+  GridRedBlackCartesian RBGRID(&GRID);
+
+  LatticeComplexD     one(&GRID);
+  LatticeComplexD      zz(&GRID);
+  LatticeComplexD       C(&GRID);
+  LatticeComplexD  Ctilde(&GRID);
+  LatticeComplexD  Cref  (&GRID);
+  LatticeComplexD  Csav  (&GRID);
+  LatticeComplexD    coor(&GRID);
+
+  LatticeSpinMatrixD    S(&GRID);
+  LatticeSpinMatrixD    Stilde(&GRID);
+
+  Coordinate p({1,3,2,3});
+
+  one = ComplexD(1.0,0.0);
+  zz  = ComplexD(0.0,0.0);
+  ComplexD ci(0.0,1.0);
+
+  std::cout<<"*************************************************"<<std::endl;
+  std::cout<<"Testing Fourier form of known plane wave         "<<std::endl;
+  std::cout<<"*************************************************"<<std::endl;
+  C=Zero();
+  for(int mu=0;mu<4;mu++){
+    RealD TwoPiL =  M_PI * 2.0/ latt_size[mu];
+    LatticeCoordinate(coor,mu);
+    C = C + (TwoPiL * p[mu]) * coor;
+  }
+  C = exp(C*ci);
+  Csav = C;
+  S=Zero();
+  S = S+C;
+
+  // PlannedFFT is templated on the lattice element type (vector_object), not the Lattice<> itself.
+  PlannedFFT<LatticeComplexD::vector_object>    theFFT(&GRID);
+  PlannedFFT<LatticeSpinMatrixD::vector_object> theFFT_spin(&GRID);
+
+  Ctilde=C;
+  std::cout<<" Benchmarking PlannedFFT of LatticeComplex  "<<std::endl;
+  theFFT.FFT_dim(Ctilde,Ctilde,0,FFTbase::forward); std::cout << theFFT.MFlops()<<" Mflops "<<std::endl;
+  theFFT.FFT_dim(Ctilde,Ctilde,1,FFTbase::forward); std::cout << theFFT.MFlops()<<" Mflops "<<std::endl;
+  theFFT.FFT_dim(Ctilde,Ctilde,2,FFTbase::forward); std::cout << theFFT.MFlops()<<" Mflops "<<std::endl;
+  theFFT.FFT_dim(Ctilde,Ctilde,3,FFTbase::forward); std::cout << theFFT.MFlops()<<" Mflops "<<std::endl;
+
+  TComplexD cVol;
+  cVol()()() = vol;
+
+  Cref=Zero();
+  pokeSite(cVol,Cref,p);
+
+  Cref=Cref-Ctilde;
+  std::cout << "diff scalar "<<norm2(Cref) << std::endl;
+
+  C=Csav;
+  theFFT.FFT_all_dim(Ctilde,C,FFTbase::forward);
+  theFFT.FFT_all_dim(Cref,Ctilde,FFTbase::backward);
+
+  std::cout << norm2(C) << " " << norm2(Ctilde) << " " << norm2(Cref)<< " vol " << vol<< std::endl;
+
+  Cref= Cref - C;
+  std::cout << " invertible check " << norm2(Cref)<<std::endl;
+
+  Stilde=S;
+  std::cout<<" Benchmarking PlannedFFT of LatticeSpinMatrix  "<<std::endl;
+  theFFT_spin.FFT_dim(Stilde,Stilde,0,FFTbase::forward); std::cout << theFFT_spin.MFlops()<<" mflops "<<std::endl;
+  theFFT_spin.FFT_dim(Stilde,Stilde,1,FFTbase::forward); std::cout << theFFT_spin.MFlops()<<" mflops "<<std::endl;
+  theFFT_spin.FFT_dim(Stilde,Stilde,2,FFTbase::forward); std::cout << theFFT_spin.MFlops()<<" mflops "<<std::endl;
+  theFFT_spin.FFT_dim(Stilde,Stilde,3,FFTbase::forward); std::cout << theFFT_spin.MFlops()<<" mflops "<<std::endl;
+
+  SpinMatrixD Sp;
+  Sp = Zero(); Sp = Sp+cVol;
+
+  S=Zero();
+  pokeSite(Sp,S,p);
+
+  S= S-Stilde;
+  std::cout << "diff FT[SpinMat] "<<norm2(S) << std::endl;
+
+  std::vector<int> seeds({1,2,3,4});
+  GridSerialRNG          sRNG;  sRNG.SeedFixedIntegers(seeds);
+  GridParallelRNG          pRNG(&GRID);
+  pRNG.SeedFixedIntegers(seeds);
+
+  LatticeGaugeFieldD Umu(&GRID);
+  SU<Nc>::ColdConfiguration(pRNG,Umu);
+
+  ////////////////////////////////////////////////////
+  // Wilson test
+  ////////////////////////////////////////////////////
+  {
+    LatticeFermionD    src(&GRID); gaussian(pRNG,src);
+    LatticeFermionD    tmp(&GRID);
+    LatticeFermionD    ref(&GRID);
+
+    RealD mass=0.01;
+    WilsonFermionD Dw(Umu,GRID,RBGRID,mass);
+
+    Dw.M(src,tmp);
+
+    std::cout << "Dw src = " <<norm2(src)<<std::endl;
+    std::cout << "Dw tmp = " <<norm2(tmp)<<std::endl;
+
+    Dw.FreePropagator(tmp,ref,mass);
+
+    std::cout << "Dw ref = " <<norm2(ref)<<std::endl;
+
+    ref = ref - src;
+    std::cout << "Dw ref-src = " <<norm2(ref)<<std::endl;
+  }
+
+  ////////////////////////////////////////////////////
+  // Dwf matrix — verify Fourier representation using PlannedFFT<LatticeFermionD>
+  ////////////////////////////////////////////////////
+  {
+    std::cout<<"****************************************"<<std::endl;
+    std::cout<<"Testing Fourier representation of Ddwf"<<std::endl;
+    std::cout<<"****************************************"<<std::endl;
+
+    const int Ls=16;
+    const int sdir=0;
+    RealD mass=0.01;
+    RealD M5  =1.0;
+    Gamma G5(Gamma::Algebra::Gamma5);
+
+    GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,&GRID);
+    GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,&GRID);
+
+    DomainWallFermionD Ddwf(Umu,*FGrid,*FrbGrid,GRID,RBGRID,mass,M5);
+
+    GridParallelRNG          RNG5(FGrid);      RNG5.SeedFixedIntegers(seeds);
+    LatticeFermionD    src5(FGrid); gaussian(RNG5,src5);
+    LatticeFermionD    src5_p(FGrid);
+    LatticeFermionD    result5(FGrid);
+    LatticeFermionD    ref5(FGrid);
+    LatticeFermionD    tmp5(FGrid);
+
+    Ddwf.M(src5,tmp5);
+    ref5 = tmp5;
+
+    PlannedFFT<LatticeFermionD::vector_object> theFFT5(FGrid);
+
+    theFFT5.FFT_dim(result5,tmp5,1,FFTbase::forward); tmp5 = result5;
+    std::cout<<"Fourier xformed Ddwf 1 "<<norm2(result5)<<std::endl;
+    theFFT5.FFT_dim(result5,tmp5,2,FFTbase::forward); tmp5 = result5;
+    std::cout<<"Fourier xformed Ddwf 2 "<<norm2(result5)<<std::endl;
+    theFFT5.FFT_dim(result5,tmp5,3,FFTbase::forward); tmp5 = result5;
+    std::cout<<"Fourier xformed Ddwf 3 "<<norm2(result5)<<std::endl;
+    theFFT5.FFT_dim(result5,tmp5,4,FFTbase::forward);
+    std::cout<<"Fourier xformed Ddwf 4 "<<norm2(result5)<<std::endl;
+    result5 = result5*ComplexD(::sqrt(1.0/vol),0.0);
+
+    std::cout<<"Fourier xformed Ddwf "<<norm2(result5)<<std::endl;
+
+    tmp5 = src5;
+    theFFT5.FFT_dim(src5_p,tmp5,1,FFTbase::forward); tmp5 = src5_p;
+    theFFT5.FFT_dim(src5_p,tmp5,2,FFTbase::forward); tmp5 = src5_p;
+    theFFT5.FFT_dim(src5_p,tmp5,3,FFTbase::forward); tmp5 = src5_p;
+    theFFT5.FFT_dim(src5_p,tmp5,4,FFTbase::forward); src5_p = src5_p*ComplexD(::sqrt(1.0/vol),0.0);
+
+    std::cout<<"Fourier xformed src5"<< norm2(src5)<<" -> "<<norm2(src5_p)<<std::endl;
+
+    Gamma::Algebra Gmu [] = {
+      Gamma::Algebra::GammaX,
+      Gamma::Algebra::GammaY,
+      Gamma::Algebra::GammaZ,
+      Gamma::Algebra::GammaT,
+      Gamma::Algebra::Gamma5
+    };
+    LatticeFermionD    Kinetic(FGrid); Kinetic = Zero();
+    LatticeComplexD    kmu(FGrid);
+    LatticeInteger     scoor(FGrid);
+    LatticeComplexD    sk (FGrid); sk = Zero();
+    LatticeComplexD    sk2(FGrid); sk2= Zero();
+    LatticeComplexD    W(FGrid); W= Zero();
+    LatticeComplexD    one5(FGrid); one5 =ComplexD(1.0,0.0);
+
+    for(int mu=0;mu<Nd;mu++) {
+      LatticeCoordinate(kmu,mu+1);
+      RealD TwoPiL =  M_PI * 2.0/ latt_size[mu];
+      kmu = TwoPiL * kmu;
+      sk2 = sk2 + 2.0*sin(kmu*0.5)*sin(kmu*0.5);
+      sk  = sk  +     sin(kmu)    *sin(kmu);
+      Kinetic = Kinetic + sin(kmu)*ci*(Gamma(Gmu[mu])*src5_p);
+    }
+    std::cout << " src5    "<<norm2(src5_p)<<std::endl;
+    std::cout << " Kinetic "<<norm2(Kinetic)<<std::endl;
+
+    W = one5 - M5 + sk2;
+    std::cout << " W "<<norm2(W)<<std::endl;
+    Kinetic = Kinetic + W * src5_p;
+    std::cout << " Kinetic "<<norm2(Kinetic)<<std::endl;
+
+    LatticeCoordinate(scoor,sdir);
+
+    tmp5 = Cshift(src5_p,sdir,+1);
+    tmp5 = (tmp5 - G5*tmp5)*0.5;
+    tmp5 = where(scoor==Integer(Ls-1),mass*tmp5,-tmp5);
+    Kinetic = Kinetic + tmp5;
+
+    tmp5 = Cshift(src5_p,sdir,-1);
+    tmp5 = (tmp5 + G5*tmp5)*0.5;
+    tmp5 = where(scoor==Integer(0),mass*tmp5,-tmp5);
+    Kinetic = Kinetic + tmp5;
+
+    std::cout<<"Momentum space Ddwf  "<< norm2(Kinetic)<<std::endl;
+    std::cout<<"Stencil Ddwf         "<< norm2(result5)<<std::endl;
+
+    result5 = result5 - Kinetic;
+    std::cout<<"diff "<< norm2(result5)<<std::endl;
+    GRID_ASSERT(norm2(result5)<1.0e-4);
+  }
+
+  ////////////////////////////////////////////////////
+  // Dwf prop
+  ////////////////////////////////////////////////////
+  {
+    std::cout<<"****************************************"<<std::endl;
+    std::cout << "Testing Ddwf Ht Mom space 4d propagator \n";
+    std::cout<<"****************************************"<<std::endl;
+
+    LatticeFermionD    src(&GRID); gaussian(pRNG,src);
+    LatticeFermionD    tmp(&GRID);
+    LatticeFermionD    ref(&GRID);
+    LatticeFermionD    diff(&GRID);
+
+    Coordinate point(4,0);
+    src=Zero();
+    SpinColourVectorD ferm; gaussian(sRNG,ferm);
+    pokeSite(ferm,src,point);
+
+    const int Ls=32;
+    GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,&GRID);
+    GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,&GRID);
+
+    RealD mass=0.01;
+    RealD M5  =0.8;
+    DomainWallFermionD Ddwf(Umu,*FGrid,*FrbGrid,GRID,RBGRID,mass,M5);
+
+    std::cout << " Solving by FFT and Feynman rules" <<std::endl;
+    bool fiveD = false;
+    Ddwf.FreePropagator(src,ref,mass,fiveD);
+
+    Gamma G5(Gamma::Algebra::Gamma5);
+
+    LatticeFermionD    src5(FGrid); src5=Zero();
+    LatticeFermionD    tmp5(FGrid);
+    LatticeFermionD    result5(FGrid); result5=Zero();
+    LatticeFermionD    result4(&GRID);
+    const int sdir=0;
+
+    tmp =   (src + G5*src)*0.5;      InsertSlice(tmp,src5,   0,sdir);
+    tmp =   (src - G5*src)*0.5;      InsertSlice(tmp,src5,Ls-1,sdir);
+
+    std::cout << " Solving by Conjugate Gradient (CGNE)" <<std::endl;
+    Ddwf.Mdag(src5,tmp5);
+    src5=tmp5;
+    MdagMLinearOperator<DomainWallFermionD,LatticeFermionD> HermOp(Ddwf);
+    ConjugateGradient<LatticeFermionD> CG(1.0e-8,10000);
+    CG(HermOp,src5,result5);
+
+    ExtractSlice(tmp,result5,0   ,sdir);   result4 =         (tmp-G5*tmp)*0.5;
+    ExtractSlice(tmp,result5,Ls-1,sdir);   result4 = result4+(tmp+G5*tmp)*0.5;
+
+    std::cout << " Taking difference" <<std::endl;
+    std::cout << "Ddwf result4 "<<norm2(result4)<<std::endl;
+    std::cout << "Ddwf ref     "<<norm2(ref)<<std::endl;
+
+    diff = ref - result4;
+    std::cout << "result - ref     "<<norm2(diff)<<std::endl;
+    GRID_ASSERT(norm2(diff)<1.0e-4);
+  }
+
+  Grid_finalize();
+}

tests/debug/Test_general_coarse_hdcg.cc

@@ -128,6 +128,10 @@ int main (int argc, char ** argv)
   typedef HermOpAdaptor<LatticeFermionD> HermFineMatrix;
   HermFineMatrix FineHermOp(HermOpEO);
 
+  LatticeFermionD src(FrbGrid); 
+  src = ComplexD(1.0);
+  PowerMethod<LatticeFermionD>       PM;   PM(HermOpEO,src);
+
   ////////////////////////////////////////////////////////////
   ///////////// Coarse basis and Little Dirac Operator ///////
   ////////////////////////////////////////////////////////////
@@ -150,7 +154,7 @@ int main (int argc, char ** argv)
   std::cout << "**************************************"<<std::endl;
   std::cout << "Create Subspace"<<std::endl;
   std::cout << "**************************************"<<std::endl;
-  Aggregates.CreateSubspaceChebyshevNew(RNG5,HermOpEO,95.); 
+  Aggregates.CreateSubspaceChebyshev(RNG5,HermOpEO,nbasis,35.,0.01,500);// <== last run
 
   std::cout << "**************************************"<<std::endl;
   std::cout << "Refine Subspace"<<std::endl;
@@ -185,7 +189,7 @@ int main (int argc, char ** argv)
   std::cout << "**************************************"<<std::endl;
 
   typedef HermitianLinearOperator<MultiGeneralCoarsenedMatrix_t,CoarseVector> MrhsHermMatrix;
-  Chebyshev<CoarseVector>      IRLCheby(0.05,40.0,101);  // 1 iter
+  Chebyshev<CoarseVector>      IRLCheby(0.01,16.0,201);  // 1 iter
   MrhsHermMatrix MrhsCoarseOp     (mrhs);
 
   CoarseVector pm_src(CoarseMrhs);
@@ -193,10 +197,10 @@ int main (int argc, char ** argv)
   PowerMethod<CoarseVector>       cPM;
   cPM(MrhsCoarseOp,pm_src);
 
-  int Nk=nrhs;
-  int Nm=Nk*3;
-  //  int Nk=36;
-  //  int Nm=144;
+  //  int Nk=16;
+  //  int Nm=Nk*3;
+  int Nk=32;
+  int Nm=128;
   int Nstop=Nk;
   int Nconv_test_interval=1;
   
@@ -210,7 +214,7 @@ int main (int argc, char ** argv)
 							  nrhs,
 							  Nk,
 							  Nm,
-							  1e-4,10);
+							  1e-4,100);
 
   int Nconv;
   std::vector<RealD>            eval(Nm);
@@ -231,8 +235,6 @@ int main (int argc, char ** argv)
   std::cout << "**************************************"<<std::endl;
   std::cout << " Recompute coarse evecs  "<<std::endl;
   std::cout << "**************************************"<<std::endl;
-  evec.resize(Nm,Coarse5d);
-  eval.resize(Nm);
   for(int r=0;r<nrhs;r++){
     random(CRNG,c_src[r]);
   }
@@ -243,7 +245,7 @@ int main (int argc, char ** argv)
   // Deflation guesser object
   ///////////////////////
   std::cout << "**************************************"<<std::endl;
-  std::cout << " Reimport coarse evecs  "<<std::endl;
+  std::cout << " Reimport coarse evecs "<<evec.size()<<" "<<eval.size()<<std::endl;
   std::cout << "**************************************"<<std::endl;
   MultiRHSDeflation<CoarseVector> MrhsGuesser;
   MrhsGuesser.ImportEigenBasis(evec,eval);
@@ -252,9 +254,11 @@ int main (int argc, char ** argv)
   // Extra HDCG parameters
   //////////////////////////
   int maxit=3000;
-  ConjugateGradient<CoarseVector>  CG(2.0e-1,maxit,false);
-  RealD lo=2.0;
-  int ord = 9;
+  //  ConjugateGradient<CoarseVector>  CG(2.0e-1,maxit,false);
+  //  ConjugateGradient<CoarseVector>  CG(1.0e-2,maxit,false);
+  ConjugateGradient<CoarseVector>  CG(5.0e-2,maxit,false);
+  RealD lo=0.2;
+  int ord = 7;
 
   DoNothingGuesser<CoarseVector> DoNothing;
   HPDSolver<CoarseVector> HPDSolveMrhs(MrhsCoarseOp,CG,DoNothing);
@@ -300,6 +304,19 @@ int main (int argc, char ** argv)
     ConjugateGradient<LatticeFermionD>  CGfine(1.0e-8,30000,false);
     CGfine(HermOpEO, src, result);
   }
+  {
+    std::cout << "**************************************"<<std::endl;
+    std::cout << "Calling MdagM CG"<<std::endl;
+    std::cout << "**************************************"<<std::endl;
+      
+    LatticeFermion result(FGrid); result=Zero();
+    LatticeFermion    src(FGrid); random(RNG5,src);
+    result=Zero();
+
+    MdagMLinearOperator<MobiusFermionD, LatticeFermionD> HermOp(Ddwf);
+    ConjugateGradient<LatticeFermionD>  CGfine(1.0e-8,30000,false);
+    CGfine(HermOp, src, result);
+  }
 #endif  
   Grid_finalize();
   return 0;

tests/debug/Test_general_coarse_pvdagm.cc

@@ -368,7 +368,10 @@ int main (int argc, char ** argv)
   TrivialPrecon<CoarseVector> simple;
   NonHermitianLinearOperator<LittleDiracOperator,CoarseVector> LinOpCoarse(LittleDiracOpPV);
   //  PrecGeneralisedConjugateResidualNonHermitian<CoarseVector>  L2PGCR(1.0e-4, 100, LinOpCoarse,simple,10,10); 
-  PrecGeneralisedConjugateResidualNonHermitian<CoarseVector>  L2PGCR(3.0e-2, 100, LinOpCoarse,simple,10,10); 
+  //  PrecGeneralisedConjugateResidualNonHermitian<CoarseVector>  L2PGCR(3.0e-2, 100, LinOpCoarse,simple,12,12);  // 35 outer
+  //  PrecGeneralisedConjugateResidualNonHermitian<CoarseVector>  L2PGCR(5.0e-2, 100, LinOpCoarse,simple,12,12);  // 36 outer, 12s
+  //  PrecGeneralisedConjugateResidualNonHermitian<CoarseVector>  L2PGCR(1.0e-1, 100, LinOpCoarse,simple,12,12);  // 36 ; 11s   
+  PrecGeneralisedConjugateResidualNonHermitian<CoarseVector>  L2PGCR(3.0e-1, 100, LinOpCoarse,simple,12,12);     
   L2PGCR.Level(3);
   c_res=Zero();
   L2PGCR(c_src,c_res);
@@ -400,7 +403,7 @@ int main (int argc, char ** argv)
 			    LinOpCoarse,
 			    L2PGCR);
   
-  PrecGeneralisedConjugateResidualNonHermitian<LatticeFermion> L1PGCR(1.0e-8,1000,PVdagM,TwoLevelPrecon,16,16);
+  PrecGeneralisedConjugateResidualNonHermitian<LatticeFermion> L1PGCR(1.0e-8,100,PVdagM,TwoLevelPrecon,10,10);
   L1PGCR.Level(1);
 
   f_res=Zero();