FOr pvdagm preconditioners if they work

Arm complex instructions on M3/M4 NEON v8.3 and simplify A64FX code paths/broaden.
Grid_vector_types and Simd.h mainly reorg and prep for sComplexD and sComplexF alternat Nsimd=1 types

BLAS_benchmark/BatchBlasBench.cc

@@ -120,7 +120,7 @@ inline void acceleratorMemSet(void *base,int value,size_t bytes) { cudaMemset(ba
 	     cudaGetErrorString( err ));				\
       printf("File %s Line %d\n",__FILE__,__LINE__);			\
       fflush(stdout);							\
-      if (acceleratorAbortOnGpuError) assert(err==cudaSuccess);		\
+      if (acceleratorAbortOnGpuError) GRID_ASSERT(err==cudaSuccess);		\
     }									\
   }
 
@@ -168,7 +168,7 @@ public:
     if ( (_Tp*)ptr == (_Tp *) NULL ) {
       printf("Grid Device Allocator got NULL for %lu bytes\n",(unsigned long) bytes );
     }
-    assert( ( (_Tp*)ptr != (_Tp *)NULL ) );
+    GRID_ASSERT( ( (_Tp*)ptr != (_Tp *)NULL ) );
     return ptr;
   }
 
@@ -276,11 +276,11 @@ public:
   {
 #ifdef GRID_HIP
     auto err = hipDeviceSynchronize();
-    assert(err==hipSuccess);
+    GRID_ASSERT(err==hipSuccess);
 #endif
 #ifdef GRID_CUDA
     auto err = cudaDeviceSynchronize();
-    assert(err==cudaSuccess);
+    GRID_ASSERT(err==cudaSuccess);
 #endif
 #ifdef GRID_SYCL
     accelerator_barrier();
@@ -305,8 +305,8 @@ public:
   {
     RealD t2=usecond();
 
-    assert(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
-    assert(OpB!=GridBLAS_OP_T);
+    GRID_ASSERT(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
+    GRID_ASSERT(OpB!=GridBLAS_OP_T);
 
     int lda = m; // m x k column major
     int ldb = k; // k x n column major
@@ -341,7 +341,7 @@ public:
 			    (hipblasDoubleComplex *) Bkn, ldb,
 			    (hipblasDoubleComplex *) &beta_p[0],
 			    (hipblasDoubleComplex *) Cmn, ldc);
-    assert(err==HIPBLAS_STATUS_SUCCESS);
+    GRID_ASSERT(err==HIPBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_CUDA
     cublasOperation_t hOpA;
@@ -361,7 +361,7 @@ public:
 			   (cuDoubleComplex *) Bkn, ldb,
 			   (cuDoubleComplex *) &beta_p[0],
 			   (cuDoubleComplex *) Cmn, ldc);
-    assert(err==CUBLAS_STATUS_SUCCESS);
+    GRID_ASSERT(err==CUBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_SYCL
       int64_t m64=m;
@@ -433,8 +433,8 @@ public:
   {
     RealD t2=usecond();
 
-    assert(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
-    assert(OpB!=GridBLAS_OP_T);
+    GRID_ASSERT(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
+    GRID_ASSERT(OpB!=GridBLAS_OP_T);
 
     int lda = m; // m x k column major
     int ldb = k; // k x n column major
@@ -469,7 +469,7 @@ public:
 			    (hipblasComplex *) Bkn, ldb,
 			    (hipblasComplex *) &beta_p[0],
 			    (hipblasComplex *) Cmn, ldc);
-    assert(err==HIPBLAS_STATUS_SUCCESS);
+    GRID_ASSERT(err==HIPBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_CUDA
     cublasOperation_t hOpA;
@@ -489,7 +489,7 @@ public:
 			   (cuComplex *) Bkn, ldb,
 			   (cuComplex *) &beta_p[0],
 			   (cuComplex *) Cmn, ldc);
-    assert(err==CUBLAS_STATUS_SUCCESS);
+    GRID_ASSERT(err==CUBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_SYCL
       int64_t m64=m;
@@ -595,11 +595,11 @@ public:
   {
     RealD t2=usecond();
     int32_t batchCount = Amk.size();
-    assert(Bkn.size()==batchCount);
-    assert(Cmn.size()==batchCount);
+    GRID_ASSERT(Bkn.size()==batchCount);
+    GRID_ASSERT(Cmn.size()==batchCount);
 
-    assert(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
-    assert(OpB!=GridBLAS_OP_T);
+    GRID_ASSERT(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
+    GRID_ASSERT(OpB!=GridBLAS_OP_T);
 
     int lda = m; // m x k column major
     int ldb = k; // k x n column major
@@ -636,7 +636,7 @@ public:
 				   (hipblasDoubleComplex **)&Cmn[0], ldc,
 				   batchCount);
     //	 std::cout << " hipblas return code " <<(int)err<<std::endl;
-    assert(err==HIPBLAS_STATUS_SUCCESS);
+    GRID_ASSERT(err==HIPBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_CUDA
     cublasOperation_t hOpA;
@@ -657,7 +657,7 @@ public:
 				  (cuDoubleComplex *) &beta_p[0],
 				  (cuDoubleComplex **)&Cmn[0], ldc,
 				  batchCount);
-    assert(err==CUBLAS_STATUS_SUCCESS);
+    GRID_ASSERT(err==CUBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_SYCL
       int64_t m64=m;
@@ -804,8 +804,8 @@ public:
     RealD t2=usecond();
     int32_t batchCount = Amk.size();
 
-    assert(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
-    assert(OpB!=GridBLAS_OP_T);
+    GRID_ASSERT(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
+    GRID_ASSERT(OpB!=GridBLAS_OP_T);
 
     int lda = m; // m x k column major
     int ldb = k; // k x n column major
@@ -821,8 +821,8 @@ public:
     acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(ComplexF));
     RealD t0=usecond();
 
-    assert(Bkn.size()==batchCount);
-    assert(Cmn.size()==batchCount);
+    GRID_ASSERT(Bkn.size()==batchCount);
+    GRID_ASSERT(Cmn.size()==batchCount);
 #ifdef GRID_HIP
     hipblasOperation_t hOpA;
     hipblasOperation_t hOpB;
@@ -843,7 +843,7 @@ public:
 				   (hipblasComplex **)&Cmn[0], ldc,
 				   batchCount);
 
-    assert(err==HIPBLAS_STATUS_SUCCESS);
+    GRID_ASSERT(err==HIPBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_CUDA
     cublasOperation_t hOpA;
@@ -864,7 +864,7 @@ public:
 				  (cuComplex *) &beta_p[0],
 				  (cuComplex **)&Cmn[0], ldc,
 				  batchCount);
-    assert(err==CUBLAS_STATUS_SUCCESS);
+    GRID_ASSERT(err==CUBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_SYCL
       int64_t m64=m;

CLAUDE.md

@@ -0,0 +1,183 @@
+# 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 |
+
+GPU builds additionally need `--enable-gen-simd-width=64` (sets 512-bit SIMD width for GPU warp/wavefront sizing) and `--enable-unified=no --enable-shm=nvlink` for multi-GPU runs.
+
+To speed up compilation, `--disable-fermion-reps --disable-gparity` skips instantiating G-parity and higher-representation fermion operators.
+
+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`)
+
+Complete, working `configure` invocations for specific HPC systems (Frontier/ROCm, Perlmutter/CUDA, Summit, SDCC-A100, etc.) live in `systems/<platform>/config-command`. These are the canonical references for production builds.
+
+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                          # root-level tests (Test_simd, Test_cshift, etc.)
+make -C tests/<subdir> tests        # build tests in a subdirectory
+make tests                          # build all tests across all subdirectories
+./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, and multigrid.
+
+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
+### Multigrid (`Grid/algorithms/multigrid/`)
+
+Aggregation-based algebraic multigrid for Wilson-type fermions. Key files: `CoarsenedMatrix.h` (coarse operator), `GeneralCoarsenedMatrix.h` and `GeneralCoarsenedMatrixMultiRHS.h` (general coarsening supporting multi-RHS solves), `Aggregates.h` (near-null vector construction), `Geometry.h` (coarse-grid geometry). `MultiGrid.h` is the top-level include.
+
+### GPU acceleration
+
+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 |
+
+The key loop macros (defined in `Grid/threads/Accelerator.h`) are:
+- `accelerator_for(iter, num, nsimd, {...})` — maps to CUDA/HIP kernel or OpenMP loop; `nsimd` is the innermost SIMD lane count
+- `accelerator_forNB(...)` — non-blocking variant (no implicit barrier)
+- `accelerator_for2dNB(iter1, num1, iter2, num2, nsimd, {...})` — 2D kernel launch
+- `thread_for(iter, num, {...})` — CPU OpenMP loop (never dispatches to GPU)
+
+On CPU builds, `accelerator_for` aliases to `thread_for`.
+
+### Solver patterns
+
+`SchurRedBlack` (`Grid/algorithms/iterative/SchurRedBlack.h`) implements red-black (even/odd) preconditioning for fermion operators. Most production fermion solves use `SchurRedBlackDiagMooeeSolve` or similar wrappers that internally call a `ConjugateGradient` on the Schur complement.
+
+Mixed-precision solvers (`ConjugateGradientMixedPrec`, `BiCGSTABMixedPrec`) drive a double-precision outer loop with single-precision inner solves.
+
+### 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`.
+- **Tensor indices are positional, not labelled.** The `Grid/tensors/` arithmetic recurses structurally over the `iScalar`/`iVector`/`iMatrix` nest: each level defines only the {scalar,vector,matrix}² products at its own level, with element types resolved by automatic type deduction, so every colour/spin/lorentz combination composes from ~200 lines (versus the pre-C++11 QDP++/PETE approach of machine-generating every case). An index's meaning derives entirely from its nesting depth counted from the outside; `iScalar` is the identity/broadcast case at every level. Never insert or remove a nesting level casually — the multiplication tables contract by position.
+- **Multigrid coarsening deepens the tensor nest by one level.** A coarse site vector is `iVector<CComplex,nbasis>`, and `innerProduct` on it returns `iScalar<CComplex>` — one level deeper than the fine block scalar. So the block-inner-product scalar type gains one `iScalar` wrapper per MG level (fine: `vTComplex`; level 2: `iScalar<vTComplex>`; see `examples/Example_pvdagm_3level.cc`). When calling `blockInnerProduct`/`blockZAXPY`/`blockOrthogonalise` on coarse fields, the coarse scalar type must match `decltype(innerProduct(siteVector(),siteVector()))` exactly; a wrong depth fails to compile (no viable `operator=` deep in the instantiation chain) rather than mis-contracting.
+- 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/DisableWarnings.h

@@ -51,11 +51,13 @@ directory
 #pragma nv_diag_suppress cast_to_qualified_type
  //disables nvcc specific warning in many files
 #pragma nv_diag_suppress esa_on_defaulted_function_ignored
+#pragma nv_diag_suppress declared_but_not_referenced
 #pragma nv_diag_suppress extra_semicolon
 #else
  //disables nvcc specific warning in json.hpp
 #pragma diag_suppress unsigned_compare_with_zero
 #pragma diag_suppress cast_to_qualified_type
+#pragma diag_suppress declared_but_not_referenced
  //disables nvcc specific warning in many files
 #pragma diag_suppress esa_on_defaulted_function_ignored
 #pragma diag_suppress extra_semicolon

Grid/GridStd.h

@@ -1,9 +1,17 @@
 #ifndef GRID_STD_H
 #define GRID_STD_H
 
+///////////////////
+// Grid config
+///////////////////
+#include "Config.h"
+
 ///////////////////
 // Std C++ dependencies
 ///////////////////
+#define _NBACKTRACE (256)
+extern void * Grid_backtrace_buffer[_NBACKTRACE];
+
 #include <cassert>
 #include <complex>
 #include <memory>
@@ -15,7 +23,9 @@
 #include <random>
 #include <functional>
 #include <stdio.h>
+#include <string.h>
 #include <stdlib.h>
+#include <unistd.h>
 #include <strings.h>
 #include <stdio.h>
 #include <signal.h>
@@ -23,11 +33,36 @@
 #include <sys/time.h>
 #include <chrono>
 #include <zlib.h>
+#ifdef HAVE_EXECINFO_H
+#include <execinfo.h>
+#endif
+
+void GridAbort(void);
+
+#define ASSLOG(A) ::write(STDERR_FILENO,A,::strlen(A));
+#ifdef HAVE_EXECINFO_H
+#define GRID_ASSERT(b) if(!(b)) {					\
+    fflush(stdout); \
+    ASSLOG(" GRID_ASSERT failure: ");					\
+    ASSLOG(__FILE__);							\
+    ASSLOG(" : ");							\
+    ASSLOG(#b);								\
+    ASSLOG(" : ");							\
+    int symbols = backtrace(Grid_backtrace_buffer,_NBACKTRACE);		\
+    backtrace_symbols_fd(Grid_backtrace_buffer,symbols,STDERR_FILENO);	\
+    GridAbort();							\
+  };
+#else
+#define GRID_ASSERT(b) if(!(b)) {					\
+    ASSLOG(" GRID_ASSERT failure: ");					\
+    ASSLOG(__FILE__);							\
+    ASSLOG(" : ");							\
+    ASSLOG(#b);								\
+    ASSLOG(" : ");							\
+    GridAbort();							\
+  };
+#endif
 
-///////////////////
-// Grid config
-///////////////////
-#include "Config.h"
 
 #ifdef TOFU
 #undef GRID_COMMS_THREADS

Grid/Makefile.am

@@ -54,21 +54,25 @@ 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+=$(SP_FERMION_FILES)
-    extra_sources+=$(SP_TWOIND_FERMION_FILES)
+if BUILD_FERMION_REPS
+      extra_sources+=$(SP_TWOIND_FERMION_FILES)
+endif
+endif
 endif
 
 lib_LIBRARIES = libGrid.a

Grid/Namespace.h

@@ -29,8 +29,8 @@ directory
 #pragma once
 
 #include <type_traits>
-#include <cassert>
 #include <exception>
+#include <cassert>
 
 #define NAMESPACE_BEGIN(A) namespace A {
 #define NAMESPACE_END(A)   }

Grid/algorithms/Algorithms.h

@@ -51,6 +51,8 @@ NAMESPACE_CHECK(approx);
 #include <Grid/algorithms/deflation/MultiRHSBlockProject.h>
 #include <Grid/algorithms/deflation/MultiRHSDeflation.h>
 #include <Grid/algorithms/deflation/MultiRHSBlockCGLinalg.h>
+// Not really deflation, but useful
+#include <Grid/algorithms/blas/MomentumProject.h>
 NAMESPACE_CHECK(deflation);
 #include <Grid/algorithms/iterative/ConjugateGradient.h>
 NAMESPACE_CHECK(ConjGrad);

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
 
@@ -28,6 +28,15 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef _GRID_FFT_H_
 #define _GRID_FFT_H_
 
+#ifdef GRID_CUDA
+#include <cufft.h>
+#endif
+
+#ifdef GRID_HIP
+#include <hipfft/hipfft.h>
+#endif
+
+#if !defined(GRID_CUDA) && !defined(GRID_HIP)
 #ifdef HAVE_FFTW
 #if defined(USE_MKL) || defined(GRID_SYCL)
 #include <fftw/fftw3.h>
@@ -35,259 +44,449 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <fftw3.h>
 #endif
 #endif
+#endif
 
 NAMESPACE_BEGIN(Grid);
 
-template<class scalar> struct FFTW { };
-
-#ifdef HAVE_FFTW	
-template<> struct FFTW<ComplexD> {
-public:
-
-  typedef fftw_complex FFTW_scalar;
-  typedef fftw_plan    FFTW_plan;
-
-  static FFTW_plan fftw_plan_many_dft(int rank, const int *n,int howmany,
-				      FFTW_scalar *in, const int *inembed,		
-				      int istride, int idist,		
-				      FFTW_scalar *out, const int *onembed,		
-				      int ostride, int odist,		
-				      int sign, unsigned flags) {
-    return ::fftw_plan_many_dft(rank,n,howmany,in,inembed,istride,idist,out,onembed,ostride,odist,sign,flags);
-  }	  
-    
-  static void fftw_flops(const FFTW_plan p,double *add, double *mul, double *fmas){
-    ::fftw_flops(p,add,mul,fmas);
-  }
-
-  inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out) {
-    ::fftw_execute_dft(p,in,out);
-  }
-  inline static void fftw_destroy_plan(const FFTW_plan p) {
-    ::fftw_destroy_plan(p);
-  }
-};
-
-template<> struct FFTW<ComplexF> {
-public:
-
-  typedef fftwf_complex FFTW_scalar;
-  typedef fftwf_plan    FFTW_plan;
-
-  static FFTW_plan fftw_plan_many_dft(int rank, const int *n,int howmany,
-				      FFTW_scalar *in, const int *inembed,		
-				      int istride, int idist,		
-				      FFTW_scalar *out, const int *onembed,		
-				      int ostride, int odist,		
-				      int sign, unsigned flags) {
-    return ::fftwf_plan_many_dft(rank,n,howmany,in,inembed,istride,idist,out,onembed,ostride,odist,sign,flags);
-  }	  
-    
-  static void fftw_flops(const FFTW_plan p,double *add, double *mul, double *fmas){
-    ::fftwf_flops(p,add,mul,fmas);
-  }
-
-  inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out) {
-    ::fftwf_execute_dft(p,in,out);
-  }
-  inline static void fftw_destroy_plan(const FFTW_plan p) {
-    ::fftwf_destroy_plan(p);
-  }
-};
-
-#endif
-
 #ifndef FFTW_FORWARD
 #define FFTW_FORWARD (-1)
 #define FFTW_BACKWARD (+1)
+#define FFTW_ESTIMATE (0)
 #endif
 
-class FFT {
-private:
-    
-  GridCartesian *vgrid;
-  GridCartesian *sgrid;
-    
-  int Nd;
-  double flops;
-  double flops_call;
-  uint64_t usec;
-    
-  Coordinate dimensions;
-  Coordinate processors;
-  Coordinate processor_coor;
-    
+template<class scalar> struct FFTW {
+};
+
+#ifdef GRID_HIP
+template<> struct FFTW<ComplexD> {
 public:
-    
   static const int forward=FFTW_FORWARD;
   static const int backward=FFTW_BACKWARD;
-    
-  double Flops(void) {return flops;}
-  double MFlops(void) {return flops/usec;}
-  double USec(void)   {return (double)usec;}    
-
-  FFT ( GridCartesian * grid ) :
-    vgrid(grid),
-    Nd(grid->_ndimension),
-    dimensions(grid->_fdimensions),
-    processors(grid->_processors),
-    processor_coor(grid->_processor_coor)
-  {
-    flops=0;
-    usec =0;
-    Coordinate layout(Nd,1);
-    sgrid = new GridCartesian(dimensions,layout,processors,*grid);
-  };
-    
-  ~FFT ( void)  {
-    delete sgrid;
+  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,
+				      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;
   }
-    
-  template<class vobj>
-  void FFT_dim_mask(Lattice<vobj> &result,const Lattice<vobj> &source,Coordinate mask,int sign){
+  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);
+    else                   rv =hipfftExecZ2Z(p,in,out,HIPFFT_BACKWARD);
+    accelerator_barrier();
+    GRID_ASSERT(rv==HIPFFT_SUCCESS);
+  }
+  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;
+  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,
+				      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);
+    else                   rv =hipfftExecC2C(p,in,out,HIPFFT_BACKWARD);
+    accelerator_barrier();
+    GRID_ASSERT(rv==HIPFFT_SUCCESS);
+  }
+  inline static void fftw_destroy_plan(const FFTW_plan p) { hipfftDestroy(p); }
+};
+#endif
 
-    conformable(result.Grid(),vgrid);
-    conformable(source.Grid(),vgrid);
-    Lattice<vobj> tmp(vgrid);
-    tmp = source;
-    for(int d=0;d<Nd;d++){
-      if( mask[d] ) {
-	FFT_dim(result,tmp,d,sign);
-	tmp=result;
+#ifdef GRID_CUDA
+template<> struct FFTW<ComplexD> {
+public:
+  static const int forward=FFTW_FORWARD;
+  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,
+				      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); }
+};
+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;
+  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,
+				      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); }
+};
+#endif
+
+#if !defined(GRID_CUDA) && !defined(GRID_HIP)
+#ifdef HAVE_FFTW
+template<> struct FFTW<ComplexD> {
+public:
+  typedef fftw_complex FFTW_scalar;
+  typedef fftw_plan    FFTW_plan;
+  static FFTW_plan fftw_plan_many_dft(int rank, int *n,int howmany,
+				      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); }
+};
+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,
+				      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); }
+};
+#endif
+#endif
+
+struct FFTbase {
+  double        flops;
+  double        flops_call;
+  uint64_t      usec;
+  GridCartesian *_grid;
+
+  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();
+    }
+  }
+  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();
+
+  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;
       }
     }
   }
 
   template<class vobj>
-  void FFT_all_dim(Lattice<vobj> &result,const Lattice<vobj> &source,int sign){
-    Coordinate mask(Nd,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){
-#ifndef HAVE_FFTW
-    assert(0);
-#else
-    conformable(result.Grid(),vgrid);
-    conformable(source.Grid(),vgrid);
+  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 = vgrid->_ldimensions[dim];
-    int G = vgrid->_fdimensions[dim];
-      
-    Coordinate layout(Nd,1);
-    Coordinate pencil_gd(vgrid->_fdimensions);
-      
-    pencil_gd[dim] = G*processors[dim];
-      
-    // Pencil global vol LxLxGxLxL per node
-    GridCartesian pencil_g(pencil_gd,layout,processors,*vgrid);
-      
-    // Construct pencils
+    typedef typename vobj::scalar_type   scalar;
     typedef typename vobj::scalar_object sobj;
-    typedef typename sobj::scalar_type   scalar;
-      
-    Lattice<sobj> pgbuf(&pencil_g);
-    autoView(pgbuf_v , pgbuf, CpuWrite);
-
     typedef typename FFTW<scalar>::FFTW_scalar FFTW_scalar;
     typedef typename FFTW<scalar>::FFTW_plan   FFTW_plan;
-      
-    int Ncomp = sizeof(sobj)/sizeof(scalar);
-    int Nlow  = 1;
-    for(int d=0;d<dim;d++){
-      Nlow*=vgrid->_ldimensions[d];
-    }
-      
-    int rank = 1;  /* 1d transforms */
-    int n[] = {G}; /* 1d transforms of length G */
-    int howmany = Ncomp;
-    int odist,idist,istride,ostride;
-    idist   = odist   = 1;          /* Distance between consecutive FT's */
-    istride = ostride = Ncomp*Nlow; /* distance between two elements in the same FT */
-    int *inembed = n, *onembed = n;
-      
-    scalar div;
-    if ( sign == backward ) div = 1.0/G;
-    else if ( sign == forward ) div = 1.0;
-    else assert(0);
-      
-    FFTW_plan p;
-    {
-      FFTW_scalar *in = (FFTW_scalar *)&pgbuf_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
-    Coordinate lcoor(Nd), gcoor(Nd);
-    result = source;
-    int pc = processor_coor[dim];
-    for(int p=0;p<processors[dim];p++) {
-      {
-	autoView(r_v,result,CpuRead);
-	autoView(p_v,pgbuf,CpuWrite);
-	thread_for(idx, sgrid->lSites(),{
-          Coordinate cbuf(Nd);
-          sobj s;
-	  sgrid->LocalIndexToLocalCoor(idx,cbuf);
-	  peekLocalSite(s,r_v,cbuf);
-	  cbuf[dim]+=((pc+p) % processors[dim])*L;
-	  pokeLocalSite(s,p_v,cbuf);
-        });
-      }
-      if (p != processors[dim] - 1) {
-	result = Cshift(result,dim,L);
-      }
-    }
-      
-    // Loop over orthog coords
-    int NN=pencil_g.lSites();
-    GridStopWatch timer;
-    timer.Start();
-    thread_for( idx,NN,{
-        Coordinate cbuf(Nd);
-	pencil_g.LocalIndexToLocalCoor(idx, cbuf);
-	if ( cbuf[dim] == 0 ) {  // restricts loop to plane at lcoor[dim]==0
-	  FFTW_scalar *in = (FFTW_scalar *)&pgbuf_v[idx];
-	  FFTW_scalar *out= (FFTW_scalar *)&pgbuf_v[idx];
-	  FFTW<scalar>::fftw_execute_dft(p,in,out);
-	}
-    });
-    timer.Stop();
-      
-    // performance counting
-    double add,mul,fma;
-    FFTW<scalar>::fftw_flops(p,&add,&mul,&fma);
-    flops_call = add+mul+2.0*fma;
-    usec += timer.useconds();
-    flops+= flops_call*NN;
-      
-    // writing out result
-    {
-      autoView(pgbuf_v,pgbuf,CpuRead);
-      autoView(result_v,result,CpuWrite);
-      thread_for(idx,sgrid->lSites(),{
-	Coordinate clbuf(Nd), cgbuf(Nd);
-	sobj s;
-	sgrid->LocalIndexToLocalCoor(idx,clbuf);
-	cgbuf = clbuf;
-	cgbuf[dim] = clbuf[dim]+L*pc;
-	peekLocalSite(s,pgbuf_v,cgbuf);
-	pokeLocalSite(s,result_v,clbuf);
-      });
-    }
-    result = result*div;
-      
-    // destroying plan
+
+    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;
+
+    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);
-#endif
+  }
+};
+
+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::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/LinearOperator.h

@@ -64,7 +64,7 @@ public:
 //
 // I'm not entirely happy with implementation; to share the Schur code between herm and non-herm
 // while still having a "OpAndNorm" in the abstract base I had to implement it in both cases
-// with an assert trap in the non-herm. This isn't right; there must be a better C++ way to
+// with an GRID_ASSERT trap in the non-herm. This isn't right; there must be a better C++ way to
 // do it, but I fear it required multiple inheritance and mixed in abstract base classes
 /////////////////////////////////////////////////////////////////////////////////////////////
 
@@ -148,22 +148,22 @@ public:
   // Support for coarsening to a multigrid
   void OpDiag (const Field &in, Field &out) {
     _Mat.Mdiag(in,out);
-    assert(0);
+    GRID_ASSERT(0);
   }
   void OpDir  (const Field &in, Field &out,int dir,int disp) {
     _Mat.Mdir(in,out,dir,disp);
-    assert(0);
+    GRID_ASSERT(0);
   }
   void OpDirAll  (const Field &in, std::vector<Field> &out){
-    assert(0);
+    GRID_ASSERT(0);
   };
   void Op     (const Field &in, Field &out){
     _Mat.M(in,out);
-    assert(0);
+    GRID_ASSERT(0);
   }
   void AdjOp     (const Field &in, Field &out){
     _Mat.Mdag(in,out);
-    assert(0);
+    GRID_ASSERT(0);
   }
   void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
     HermOp(in,out);
@@ -188,13 +188,13 @@ public:
   ShiftedHermOpLinearOperator(LinearOperatorBase<Field> &Mat,RealD shift): _Mat(Mat), _shift(shift){};
   // Support for coarsening to a multigrid
   void OpDiag (const Field &in, Field &out) {
-    assert(0);
+    GRID_ASSERT(0);
   }
   void OpDir  (const Field &in, Field &out,int dir,int disp) {
-    assert(0);
+    GRID_ASSERT(0);
   }
   void OpDirAll  (const Field &in, std::vector<Field> &out){
-    assert(0);
+    GRID_ASSERT(0);
   };
   void Op     (const Field &in, Field &out){
     HermOp(in,out);
@@ -271,10 +271,42 @@ public:
     _Mat.Mdag(in,out);
   }
   void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
-    assert(0);
+    GRID_ASSERT(0);
   }
   void HermOp(const Field &in, Field &out){
-    assert(0);
+    GRID_ASSERT(0);
+  }
+};
+template<class Matrix,class Field>
+class ShiftedNonHermitianLinearOperator : public LinearOperatorBase<Field> {
+  Matrix &_Mat;
+  RealD shift;
+public:
+  ShiftedNonHermitianLinearOperator(Matrix &Mat,RealD shft): _Mat(Mat),shift(shft){};
+  // Support for coarsening to a multigrid
+  void OpDiag (const Field &in, Field &out) {
+    _Mat.Mdiag(in,out);
+    out = out + shift*in;
+  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {
+    _Mat.Mdir(in,out,dir,disp);
+  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){
+    _Mat.MdirAll(in,out);
+  };
+  void Op     (const Field &in, Field &out){
+    _Mat.M(in,out);
+    out = out + shift * in;
+  }
+  void AdjOp     (const Field &in, Field &out){
+    _Mat.Mdag(in,out);
+    out = out + shift * in;
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
+    GRID_ASSERT(0);
+  }
+  void HermOp(const Field &in, Field &out){
+    GRID_ASSERT(0);
   }
 };
 
@@ -313,13 +345,13 @@ class SchurOperatorBase :  public LinearOperatorBase<Field> {
   }
   // Support for coarsening to a multigrid
   void OpDiag (const Field &in, Field &out) {
-    assert(0); // must coarsen the unpreconditioned system
+    GRID_ASSERT(0); // must coarsen the unpreconditioned system
   }
   void OpDir  (const Field &in, Field &out,int dir,int disp) {
-    assert(0);
+    GRID_ASSERT(0);
   }
   void OpDirAll  (const Field &in, std::vector<Field> &out){
-    assert(0);
+    GRID_ASSERT(0);
   };
 };
 template<class Matrix,class Field>
@@ -415,10 +447,10 @@ class NonHermitianSchurOperatorBase :  public LinearOperatorBase<Field>
     MpcDag(tmp,out);
   }
   virtual void HermOpAndNorm(const Field& in, Field& out, RealD& n1, RealD& n2) {
-    assert(0);
+    GRID_ASSERT(0);
   }
   virtual void HermOp(const Field& in, Field& out) {
-    assert(0);
+    GRID_ASSERT(0);
   }
   void Op(const Field& in, Field& out) {
     Mpc(in, out);
@@ -428,13 +460,13 @@ class NonHermitianSchurOperatorBase :  public LinearOperatorBase<Field>
   }
   // Support for coarsening to a multigrid
   void OpDiag(const Field& in, Field& out) {
-    assert(0); // must coarsen the unpreconditioned system
+    GRID_ASSERT(0); // must coarsen the unpreconditioned system
   }
   void OpDir(const Field& in, Field& out, int dir, int disp) {
-    assert(0);
+    GRID_ASSERT(0);
   }
   void OpDirAll(const Field& in, std::vector<Field>& out){
-    assert(0);
+    GRID_ASSERT(0);
   };
 };
 
@@ -548,7 +580,7 @@ class SchurStaggeredOperator :  public SchurOperatorBase<Field> {
  public:
   SchurStaggeredOperator (Matrix &Mat): _Mat(Mat), tmp(_Mat.RedBlackGrid()) 
   { 
-    assert( _Mat.isTrivialEE() );
+    GRID_ASSERT( _Mat.isTrivialEE() );
     mass = _Mat.Mass();
   }
   virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
@@ -579,7 +611,7 @@ class SchurStaggeredOperator :  public SchurOperatorBase<Field> {
     Mpc(in,out);
   }
   virtual void MpcDagMpc(const Field &in, Field &out) {
-    assert(0);// Never need with staggered
+    GRID_ASSERT(0);// Never need with staggered
   }
 };
 template<class Matrix,class Field> using SchurStagOperator = SchurStaggeredOperator<Matrix,Field>;
@@ -591,7 +623,7 @@ template<class Field> class OperatorFunction {
 public:
   virtual void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) = 0;
   virtual void operator() (LinearOperatorBase<Field> &Linop, const std::vector<Field> &in,std::vector<Field> &out) {
-    assert(in.size()==out.size());
+    GRID_ASSERT(in.size()==out.size());
     for(int k=0;k<in.size();k++){
       (*this)(Linop,in[k],out[k]);
     }
@@ -605,7 +637,7 @@ public:
 
   virtual void operator() (const std::vector<Field> &in, std::vector<Field> &out)
   {
-    assert(in.size() == out.size());
+    GRID_ASSERT(in.size() == out.size());
 
     for (unsigned int i = 0; i < in.size(); ++i)
     {

Grid/algorithms/approx/Chebyshev.h

@@ -269,7 +269,9 @@ public:
     RealD xscale = 2.0/(hi-lo);
     RealD mscale = -(hi+lo)/(hi-lo);
     Linop.HermOp(T0,y);
+    grid->Barrier();
     axpby(T1,xscale,mscale,y,in);
+    grid->Barrier();
 
     // sum = .5 c[0] T0 + c[1] T1
     //    out = ()*T0 + Coeffs[1]*T1;

Grid/algorithms/approx/Remez.cc

@@ -121,7 +121,7 @@ double AlgRemez::generateApprox(int num_degree, int den_degree,
   // Reallocate arrays, since degree has changed
   if (num_degree != n || den_degree != d) allocate(num_degree,den_degree);
 
-  assert(a_len<=SUM_MAX);
+  GRID_ASSERT(a_len<=SUM_MAX);
 
   step = new bigfloat[num_degree+den_degree+2];
 
@@ -151,9 +151,9 @@ double AlgRemez::generateApprox(int num_degree, int den_degree,
     equations();
     if (delta < tolerance) {
       std::cout<<"Delta too small, try increasing precision\n";
-      assert(0);
+      GRID_ASSERT(0);
     };    
-    assert( delta>= tolerance);
+    GRID_ASSERT( delta>= tolerance);
 
     search(step);
   }

Grid/algorithms/approx/Remez.h

@@ -134,7 +134,7 @@ class AlgRemez
   virtual ~AlgRemez();
 
   int getDegree(void){ 
-    assert(n==d);
+    GRID_ASSERT(n==d);
     return n;
   }
   // Reset the bounds of the approximation

Grid/algorithms/approx/RemezGeneral.cc

@@ -28,11 +28,11 @@ void AlgRemezGeneral::setupPolyProperties(int num_degree, int den_degree, PolyTy
   pow_n = num_degree;
   pow_d = den_degree;
 
-  if(pow_n % 2 == 0 && num_type_in == PolyType::Odd) assert(0);
-  if(pow_n % 2 == 1 && num_type_in == PolyType::Even) assert(0);
+  if(pow_n % 2 == 0 && num_type_in == PolyType::Odd) GRID_ASSERT(0);
+  if(pow_n % 2 == 1 && num_type_in == PolyType::Even) GRID_ASSERT(0);
 
-  if(pow_d % 2 == 0 && den_type_in == PolyType::Odd) assert(0);
-  if(pow_d % 2 == 1 && den_type_in == PolyType::Even) assert(0);
+  if(pow_d % 2 == 0 && den_type_in == PolyType::Odd) GRID_ASSERT(0);
+  if(pow_d % 2 == 1 && den_type_in == PolyType::Even) GRID_ASSERT(0);
 
   num_type = num_type_in;
   den_type = den_type_in;
@@ -112,9 +112,9 @@ double AlgRemezGeneral::generateApprox(const int num_degree, const int den_degre
     equations();
     if (delta < tolerance) {
       std::cout<<"Iteration " << iter-1 << " delta too small (" << delta << "<" << tolerance << "), try increasing precision\n";
-      assert(0);
+      GRID_ASSERT(0);
     };    
-    assert( delta>= tolerance );
+    GRID_ASSERT( delta>= tolerance );
 
     search();
   }
@@ -278,7 +278,7 @@ void AlgRemezGeneral::equations(){
       if(num_pows[j] != -1){ *aa++ = z; t++; }
       z *= x;
     }
-    assert(t == n+1);
+    GRID_ASSERT(t == n+1);
 
     z = (bigfloat)1l;
     t = 0;
@@ -286,7 +286,7 @@ void AlgRemezGeneral::equations(){
       if(den_pows[j] != -1){ *aa++ = -y * z; t++; }
       z *= x;
     }
-    assert(t == d);
+    GRID_ASSERT(t == d);
 
     B[i] = y * z;		// Right hand side vector
   }

Grid/algorithms/approx/RemezGeneral.h

@@ -106,7 +106,7 @@ class AlgRemezGeneral{
 		  bigfloat (*f)(bigfloat x, void *data), void *data);
 
   inline int getDegree(void) const{ 
-    assert(n==d);
+    GRID_ASSERT(n==d);
     return n;
   }
   // Reset the bounds of the approximation

Grid/algorithms/approx/ZMobius.cc

@@ -74,7 +74,7 @@ bigfloat epsilonMobius(bigfloat x, void* data){
 void computeZmobiusOmega(std::vector<ComplexD> &omega_out, const int Ls_out,
 			 const std::vector<RealD> &omega_in, const int Ls_in,
 			 const RealD lambda_bound){
-  assert(omega_in.size() == Ls_in);
+  GRID_ASSERT(omega_in.size() == Ls_in);
   omega_out.resize(Ls_out);
 
   //Use the Remez algorithm to generate the appropriate rational polynomial

Grid/algorithms/blas/BatchedInverse.h

@@ -0,0 +1,282 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: BatchedInverse.h
+
+    Copyright (C) 2026
+
+Author: Peter Boyle <pboyle@bnl.gov>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#pragma once
+
+#include <Grid/algorithms/blas/BatchedBlas.h>
+
+#ifdef GRID_HIP
+#include <rocsolver/rocsolver.h>
+#endif
+// GRID_CUDA: batched LU inversion lives in cuBLAS (getrfBatched/getriBatched);
+//            cublas_v2.h already included via BatchedBlas.h.
+// GRID_SYCL: oneapi/mkl.hpp already included via BatchedBlas.h (lapack::getrf/getri).
+
+NAMESPACE_BEGIN(Grid);
+
+///////////////////////////////////////////////////////////////////////////////
+// GridBLASInverse: cross-platform batched dense matrix inversion.
+//
+// HIGH LEVEL contract (deliberately NOT a getrf/getrs interface): invert a
+// batch of dense N x N matrices IN PLACE,
+//
+//     A[i] <- A[i]^{-1}       i = 0 .. batchCount-1
+//
+// Layout: column major, lda = N, contiguous per batch element; pointer list
+// exactly as GridBLAS::gemmBatched (deviceVector<T*> of device pointers).
+// Each backend chooses HOW:
+//   HIP  : rocSOLVER getrf_batched + getri_batched
+//   CUDA : cuBLAS getrfBatched + getriBatched (out-of-place getri; workspace
+//          hidden here, result copied back so the surface stays in-place)
+//   SYCL : oneMKL LAPACK getrf + getri per batch element (USM, in-order queue)
+//   CPU  : Eigen PartialPivLU (the correctness oracle for all of the above)
+//
+// The int32 vendor-batched entry points bound N < 2^31 (asserted); the huge
+// single-matrix ILP64 path (getrf_64 + blocked identity-getrs harvest, proven
+// in the dense coarse-coarse setup at N=69120) migrates here as a batch==1
+// large-N dispatch in a follow-up -- the recursive Schur leaves are the
+// batched consumers this surface is shaped for.
+//
+// NB GPU-backend call signatures are written to vendor documentation but the
+// air-gapped development loop compiles only the CPU/Eigen path; verify the
+// rocSOLVER/cuBLAS/oneMKL calls against headers on first device compile.
+// Semantics are locked by the CPU unit test (Test_batched_blas).
+///////////////////////////////////////////////////////////////////////////////
+class GridBLASInverse {
+public:
+
+#ifdef GRID_HIP
+  // rocSOLVER runs on a rocblas_handle (distinct type from hipblasHandle_t)
+  static rocblas_handle & Handle(void) {
+    static rocblas_handle h;
+    static int init = 0;
+    if ( !init ) {
+      auto st = rocblas_create_handle(&h);
+      GRID_ASSERT(st == rocblas_status_success);
+      init = 1;
+    }
+    return h;
+  }
+#endif
+#ifdef GRID_CUDA
+  // cuBLAS batched LU shares the GridBLAS handle
+  static cublasHandle_t & Handle(void) {
+    GridBLAS::Init();
+    return GridBLAS::gridblasHandle;
+  }
+#endif
+#ifdef GRID_SYCL
+  static sycl::queue * & Handle(void) {
+    GridBLAS::Init();
+    return GridBLAS::gridblasHandle;
+  }
+#endif
+
+  GridBLASInverse() {};
+  ~GridBLASInverse() {};
+
+  void inverseBatched(int64_t N, deviceVector<ComplexF*> &Amat)
+  {
+    int32_t batchCount = Amat.size();
+    GRID_ASSERT(batchCount > 0);
+
+#ifdef GRID_HIP
+    GRID_ASSERT( N < 2147483647L );
+    rocblas_int n   = (rocblas_int)N;
+    rocblas_int lda = (rocblas_int)N;
+
+    deviceVector<rocblas_int> ipiv((uint64_t)batchCount*N);
+    deviceVector<rocblas_int> info(batchCount);
+
+    auto st1 = rocsolver_cgetrf_batched(Handle(), n, n,
+					(rocblas_float_complex *const *)&Amat[0], lda,
+					&ipiv[0], (rocblas_stride)N,
+					&info[0], batchCount);
+    GRID_ASSERT(st1 == rocblas_status_success);
+    auto st2 = rocsolver_cgetri_batched(Handle(), n,
+					(rocblas_float_complex *const *)&Amat[0], lda,
+					&ipiv[0], (rocblas_stride)N,
+					&info[0], batchCount);
+    GRID_ASSERT(st2 == rocblas_status_success);
+    accelerator_barrier();
+    std::vector<rocblas_int> info_h(batchCount);
+    acceleratorCopyFromDevice(&info[0],&info_h[0],batchCount*sizeof(rocblas_int));
+    for(int i=0;i<batchCount;i++) GRID_ASSERT(info_h[i]==0); // singular pivot => abort loudly
+#endif
+#ifdef GRID_CUDA
+    GRID_ASSERT( N < 2147483647L );
+    int n = (int)N;
+
+    deviceVector<int> ipiv((uint64_t)batchCount*N);
+    deviceVector<int> info(batchCount);
+
+    auto st1 = cublasCgetrfBatched(Handle(), n,
+				   (cuComplex **)&Amat[0], n,
+				   &ipiv[0], &info[0], batchCount);
+    GRID_ASSERT(st1 == CUBLAS_STATUS_SUCCESS);
+
+    // getri is OUT of place: hidden workspace keeps the surface in-place
+    deviceVector<ComplexF>  work((uint64_t)batchCount*N*N);
+    deviceVector<ComplexF*> Cptr(batchCount);
+    std::vector<ComplexF*>  Cptr_h(batchCount);
+    std::vector<ComplexF*>  Aptr_h(batchCount);
+    for(int i=0;i<batchCount;i++) Cptr_h[i] = &work[(uint64_t)i*N*N];
+    acceleratorCopyToDevice(&Cptr_h[0],&Cptr[0],batchCount*sizeof(ComplexF*));
+    acceleratorCopyFromDevice(&Amat[0],&Aptr_h[0],batchCount*sizeof(ComplexF*));
+
+    auto st2 = cublasCgetriBatched(Handle(), n,
+				   (const cuComplex *const *)&Amat[0], n,
+				   &ipiv[0],
+				   (cuComplex **)&Cptr[0], n,
+				   &info[0], batchCount);
+    GRID_ASSERT(st2 == CUBLAS_STATUS_SUCCESS);
+    accelerator_barrier();
+    std::vector<int> info_h(batchCount);
+    acceleratorCopyFromDevice(&info[0],&info_h[0],batchCount*sizeof(int));
+    for(int i=0;i<batchCount;i++) GRID_ASSERT(info_h[i]==0);
+    for(int i=0;i<batchCount;i++)
+      acceleratorCopyDeviceToDevice(Cptr_h[i],Aptr_h[i],(uint64_t)N*N*sizeof(ComplexF));
+#endif
+#ifdef GRID_SYCL
+    // Per-element oneMKL LAPACK on the in-order queue; group API optimisation later.
+    sycl::queue *q = Handle();
+    std::vector<ComplexF*> Aptr_h(batchCount);
+    acceleratorCopyFromDevice(&Amat[0],&Aptr_h[0],batchCount*sizeof(ComplexF*));
+
+    int64_t lwf = oneapi::mkl::lapack::getrf_scratchpad_size<std::complex<float> >(*q,N,N,N);
+    int64_t lwi = oneapi::mkl::lapack::getri_scratchpad_size<std::complex<float> >(*q,N,N);
+    deviceVector<ComplexF> scratchf(lwf);
+    deviceVector<ComplexF> scratchi(lwi);
+    deviceVector<int64_t>  ipiv(N);
+    for(int i=0;i<batchCount;i++){
+      oneapi::mkl::lapack::getrf(*q,N,N,(std::complex<float>*)Aptr_h[i],N,&ipiv[0],
+				 (std::complex<float>*)&scratchf[0],lwf);
+      oneapi::mkl::lapack::getri(*q,N,  (std::complex<float>*)Aptr_h[i],N,&ipiv[0],
+				 (std::complex<float>*)&scratchi[0],lwi);
+    }
+    q->wait();
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
+    // Reference implementation; the oracle the unit test locks semantics with.
+    thread_for (p, batchCount, {
+      Eigen::Map<Eigen::MatrixXcf> eA(Amat[p],N,N);
+      Eigen::PartialPivLU<Eigen::MatrixXcf> lu(eA);
+      eA = lu.inverse();
+      });
+#endif
+  }
+
+  void inverseBatched(int64_t N, deviceVector<ComplexD*> &Amat)
+  {
+    int32_t batchCount = Amat.size();
+    GRID_ASSERT(batchCount > 0);
+
+#ifdef GRID_HIP
+    GRID_ASSERT( N < 2147483647L );
+    rocblas_int n   = (rocblas_int)N;
+    rocblas_int lda = (rocblas_int)N;
+
+    deviceVector<rocblas_int> ipiv((uint64_t)batchCount*N);
+    deviceVector<rocblas_int> info(batchCount);
+
+    auto st1 = rocsolver_zgetrf_batched(Handle(), n, n,
+					(rocblas_double_complex *const *)&Amat[0], lda,
+					&ipiv[0], (rocblas_stride)N,
+					&info[0], batchCount);
+    GRID_ASSERT(st1 == rocblas_status_success);
+    auto st2 = rocsolver_zgetri_batched(Handle(), n,
+					(rocblas_double_complex *const *)&Amat[0], lda,
+					&ipiv[0], (rocblas_stride)N,
+					&info[0], batchCount);
+    GRID_ASSERT(st2 == rocblas_status_success);
+    accelerator_barrier();
+    std::vector<rocblas_int> info_h(batchCount);
+    acceleratorCopyFromDevice(&info[0],&info_h[0],batchCount*sizeof(rocblas_int));
+    for(int i=0;i<batchCount;i++) GRID_ASSERT(info_h[i]==0);
+#endif
+#ifdef GRID_CUDA
+    GRID_ASSERT( N < 2147483647L );
+    int n = (int)N;
+
+    deviceVector<int> ipiv((uint64_t)batchCount*N);
+    deviceVector<int> info(batchCount);
+
+    auto st1 = cublasZgetrfBatched(Handle(), n,
+				   (cuDoubleComplex **)&Amat[0], n,
+				   &ipiv[0], &info[0], batchCount);
+    GRID_ASSERT(st1 == CUBLAS_STATUS_SUCCESS);
+
+    deviceVector<ComplexD>  work((uint64_t)batchCount*N*N);
+    deviceVector<ComplexD*> Cptr(batchCount);
+    std::vector<ComplexD*>  Cptr_h(batchCount);
+    std::vector<ComplexD*>  Aptr_h(batchCount);
+    for(int i=0;i<batchCount;i++) Cptr_h[i] = &work[(uint64_t)i*N*N];
+    acceleratorCopyToDevice(&Cptr_h[0],&Cptr[0],batchCount*sizeof(ComplexD*));
+    acceleratorCopyFromDevice(&Amat[0],&Aptr_h[0],batchCount*sizeof(ComplexD*));
+
+    auto st2 = cublasZgetriBatched(Handle(), n,
+				   (const cuDoubleComplex *const *)&Amat[0], n,
+				   &ipiv[0],
+				   (cuDoubleComplex **)&Cptr[0], n,
+				   &info[0], batchCount);
+    GRID_ASSERT(st2 == CUBLAS_STATUS_SUCCESS);
+    accelerator_barrier();
+    std::vector<int> info_h(batchCount);
+    acceleratorCopyFromDevice(&info[0],&info_h[0],batchCount*sizeof(int));
+    for(int i=0;i<batchCount;i++) GRID_ASSERT(info_h[i]==0);
+    for(int i=0;i<batchCount;i++)
+      acceleratorCopyDeviceToDevice(Cptr_h[i],Aptr_h[i],(uint64_t)N*N*sizeof(ComplexD));
+#endif
+#ifdef GRID_SYCL
+    sycl::queue *q = Handle();
+    std::vector<ComplexD*> Aptr_h(batchCount);
+    acceleratorCopyFromDevice(&Amat[0],&Aptr_h[0],batchCount*sizeof(ComplexD*));
+
+    int64_t lwf = oneapi::mkl::lapack::getrf_scratchpad_size<std::complex<double> >(*q,N,N,N);
+    int64_t lwi = oneapi::mkl::lapack::getri_scratchpad_size<std::complex<double> >(*q,N,N);
+    deviceVector<ComplexD> scratchf(lwf);
+    deviceVector<ComplexD> scratchi(lwi);
+    deviceVector<int64_t>  ipiv(N);
+    for(int i=0;i<batchCount;i++){
+      oneapi::mkl::lapack::getrf(*q,N,N,(std::complex<double>*)Aptr_h[i],N,&ipiv[0],
+				 (std::complex<double>*)&scratchf[0],lwf);
+      oneapi::mkl::lapack::getri(*q,N,  (std::complex<double>*)Aptr_h[i],N,&ipiv[0],
+				 (std::complex<double>*)&scratchi[0],lwi);
+    }
+    q->wait();
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
+    thread_for (p, batchCount, {
+      Eigen::Map<Eigen::MatrixXcd> eA(Amat[p],N,N);
+      Eigen::PartialPivLU<Eigen::MatrixXcd> lu(eA);
+      eA = lu.inverse();
+      });
+#endif
+  }
+};
+
+NAMESPACE_END(Grid);

Grid/algorithms/blas/MomentumProject.h

@@ -0,0 +1,300 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: MomentumProject.h
+
+    Copyright (C) 2025
+
+Author: Peter Boyle <pboyle@bnl.gov>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+/* 
+   MultiMomProject
+
+   Import vectors -> nxyz x (ncomponent x nt)
+   Import complex phases -> nmom x nxy
+
+   apply = via (possibly batched) GEMM
+*/
+template<class Field, class ComplexField>
+class MomentumProject
+{
+public:
+
+  typedef typename Field::scalar_type   scalar;
+  typedef typename Field::scalar_object scalar_object;
+
+  GridBase *grid;
+  uint64_t nmom;
+  uint64_t nxyz;
+  uint64_t nt;
+  uint64_t nbtw;
+  uint64_t words;
+
+  deviceVector<scalar> BLAS_V;      // 
+  deviceVector<scalar> BLAS_M;      // 
+  deviceVector<scalar> BLAS_P;      // 
+  
+  MomentumProject(){};
+ ~MomentumProject(){ Deallocate(); };
+  
+  void Deallocate(void)
+  {
+    grid=nullptr;
+    nmom=0;
+    nxyz=0;
+    nt=0;
+    nbtw=0;
+    words=0;
+    BLAS_V.resize(0);
+    BLAS_M.resize(0);
+    BLAS_P.resize(0);
+  }
+  void Allocate(int _nmom,GridBase *_grid)
+  {
+    grid=_grid;
+    Coordinate ldims = grid->LocalDimensions();
+
+    nmom=_nmom;
+    nt   = ldims[grid->Nd()-1];
+    nxyz = grid->lSites()/nt;
+    words = sizeof(scalar_object)/sizeof(scalar);
+    nbtw = nt * words;
+
+    BLAS_V.resize (nxyz * nt * words );
+    BLAS_M.resize (nmom * nxyz       );
+    BLAS_P.resize (nmom * nt * words );
+  }
+  void ImportMomenta(const std::vector <ComplexField> &momenta)
+  {
+    GRID_ASSERT(momenta.size()==nmom);
+    //    might as well just make the momenta here
+    typedef typename Field::vector_object vobj;
+
+    int nd = grid->_ndimension;
+
+    uint64_t sz = BLAS_M.size();
+
+    GRID_ASSERT(momenta.size()==nmom)
+    GRID_ASSERT(momenta[0].Grid()==grid);
+    GRID_ASSERT(sz = nxyz * nmom);
+    
+    Coordinate rdimensions = grid->_rdimensions;
+    Coordinate ldims       = grid->LocalDimensions();
+    int64_t osites         = grid->oSites();
+    Coordinate simd        = grid->_simd_layout;
+    const int Nsimd        = vobj::Nsimd();
+    uint64_t lwords        = words; // local variable for copy in to GPU
+    int64_t Nxyz = nxyz;
+    auto blasData_p  = &BLAS_M[0];
+    for(int m=0;m<momenta.size();m++){
+
+      autoView( Data   , momenta[m], AcceleratorRead);
+      auto Data_p  = &Data[0];
+
+      accelerator_for(xyz,nxyz,1,{
+	  //////////////////////////////////////////
+	  // isite -- map lane within buffer to lane within lattice
+	  ////////////////////////////////////////////
+	    Coordinate lcoor(nd,0);
+	    Lexicographic::CoorFromIndex(lcoor,xyz,ldims);
+	    
+	    Coordinate icoor(nd);
+	    Coordinate ocoor(nd);
+	    for (int d = 0; d < nd; d++) {
+	      icoor[d] = lcoor[d]/rdimensions[d];
+	      ocoor[d] = lcoor[d]%rdimensions[d];
+	    }
+	    int64_t osite;
+	    int64_t isite;
+	    Lexicographic::IndexFromCoor(ocoor,osite,rdimensions);
+	    Lexicographic::IndexFromCoor(icoor,isite,simd);
+	    
+	    // BLAS_M[nmom][slice_vol]
+	    // Fortran Column major BLAS layout is M_xyz,mom
+	    scalar data = extractLane(isite,Data[osite]);
+	    uint64_t idx = xyz+m*Nxyz;
+	    blasData_p[idx] = data;
+	});
+    }
+  }
+  void ImportVector(Field &vec)
+  {
+    typedef typename Field::vector_object vobj;
+
+    int nd = grid->_ndimension;
+
+    uint64_t sz = BLAS_V.size();
+
+    GRID_ASSERT(sz = nxyz * words * nt);
+    
+    Coordinate rdimensions = grid->_rdimensions;
+    Coordinate ldims= grid->LocalDimensions();
+    int64_t osites = grid->oSites();
+    Coordinate simd = grid->_simd_layout;
+    const int Nsimd = vobj::Nsimd();
+    uint64_t lwords= words; // local variable for copy in to GPU
+
+    auto blasData_p  = &BLAS_V[0];
+    autoView( Data   , vec, AcceleratorRead);
+    auto Data_p  = &Data[0];
+
+    int64_t nwords = words;// for capture
+    int64_t Nt     = nt;// for capture
+    
+    accelerator_for(sf,osites,Nsimd,{
+#ifdef GRID_SIMT
+        {
+	  int lane=acceleratorSIMTlane(Nsimd); // buffer lane
+#else
+	  for(int lane=0;lane<Nsimd;lane++) {
+#endif
+	  //////////////////////////////////////////
+	  // isite -- map lane within buffer to lane within lattice
+	  ////////////////////////////////////////////
+	    Coordinate lcoor(nd,0);
+	    Coordinate icoor(nd);
+	    Coordinate ocoor(nd);
+	    
+	    Lexicographic::CoorFromIndex(icoor,lane,simd);
+	    Lexicographic::CoorFromIndex(ocoor,sf,rdimensions);
+
+	  
+	    int64_t l_xyz = 0;
+	    for (int d = 0; d < nd; d++) {
+	      lcoor[d] = rdimensions[d]*icoor[d] + ocoor[d];
+	    }
+	    uint64_t l_t   = lcoor[nd-1];
+
+	    Coordinate xyz_coor = lcoor;
+	    xyz_coor[nd-1] =0;
+	    Lexicographic::IndexFromCoor(xyz_coor,l_xyz,ldims);
+
+	    
+	    scalar_object data = extractLane(lane,Data[sf]);
+	    scalar *data_words = (scalar *) &data;
+	    for(int w = 0 ; w < nwords; w++) {
+	      // BLAS_V[slice_vol][nt][words]
+	      // Fortran Column major BLAS layout is V_(t,w)_xyz
+	      uint64_t idx = w+l_t*nwords + l_xyz * nwords * Nt;
+	      blasData_p[idx] = data_words[w];
+	    }
+#ifdef GRID_SIMT
+	}
+#else
+	}
+#endif
+	});
+  }
+  void ExportMomentumProjection(std::vector<typename Field::scalar_object> &projection)
+  {
+    projection.resize(nmom*nt);
+    acceleratorCopyFromDevice(&BLAS_P[0],(scalar *)&projection[0],BLAS_P.size()*sizeof(scalar));
+    // Could decide on a layout late?
+  }
+
+  // Row major layout "C" order:
+  // BLAS_V[slice_vol][nt][words]
+  // BLAS_M[nmom][slice_vol]
+  // BLAS_P[nmom][nt][words]
+  //
+  // Fortran Column major BLAS layout is V_(w,t)_xyz
+  // Fortran Column major BLAS layout is M_xyz,mom
+  // Fortran Column major BLAS layout is P_(w,t),mom
+  //
+  // Projected
+  //
+  // P = (V * M)_(w,t),mom
+  //
+  void Project(Field &data,std::vector< typename Field::scalar_object > & projected_gdata)
+  {
+    double t_import=0;
+    double t_export=0;
+    double t_gemm  =0;
+    double t_allreduce=0;
+    t_import-=usecond();
+    this->ImportVector(data);
+
+    std::vector< typename Field::scalar_object > projected_planes;
+
+    deviceVector<scalar *> Vd(1);
+    deviceVector<scalar *> Md(1);
+    deviceVector<scalar *> Pd(1);
+
+    scalar * Vh = & BLAS_V[0];
+    scalar * Mh = & BLAS_M[0];
+    scalar * Ph = & BLAS_P[0];
+
+    acceleratorPut(Vd[0],Vh);
+    acceleratorPut(Md[0],Mh);
+    acceleratorPut(Pd[0],Ph);
+    t_import+=usecond();
+
+    GridBLAS BLAS;
+
+    /////////////////////////////////////////
+    // P_im = VMmx . Vxi
+    /////////////////////////////////////////
+    t_gemm-=usecond();
+    BLAS.gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N, 
+    		     words*nt,nmom,nxyz,
+		     scalar(1.0),
+		     Vd,
+		     Md,
+		     scalar(0.0),  // wipe out result
+		     Pd);
+    BLAS.synchronise();
+    t_gemm+=usecond();
+
+    t_export-=usecond();
+    ExportMomentumProjection(projected_planes); // resizes
+    t_export+=usecond();
+
+    /////////////////////////////////
+    // Reduce across MPI ranks
+    /////////////////////////////////
+    int nd = grid->Nd();
+    int gt = grid->GlobalDimensions()[nd-1];
+    int lt = grid->LocalDimensions()[nd-1];
+    projected_gdata.resize(gt*nmom);
+    for(int t=0;t<gt*nmom;t++){ // global Nt array with zeroes for stuff not on this node
+      projected_gdata[t]=Zero();
+    }
+    for(int t=0;t<lt;t++){
+    for(int m=0;m<nmom;m++){
+      int st = grid->LocalStarts()[nd-1];
+      projected_gdata[t+st + gt*m] = projected_planes[t+lt*m];
+    }}
+    t_allreduce-=usecond();
+    grid->GlobalSumVector((scalar *)&projected_gdata[0],gt*nmom*words);
+    t_allreduce+=usecond();
+
+    std::cout << GridLogPerformance<<" MomentumProject t_import  "<<t_import<<"us"<<std::endl;
+    std::cout << GridLogPerformance<<" MomentumProject t_export  "<<t_export<<"us"<<std::endl;
+    std::cout << GridLogPerformance<<" MomentumProject t_gemm    "<<t_gemm<<"us"<<std::endl;
+    std::cout << GridLogPerformance<<" MomentumProject t_reduce  "<<t_allreduce<<"us"<<std::endl;
+
+  }
+};
+
+NAMESPACE_END(Grid);

Grid/algorithms/deflation/Deflation.h

@@ -69,8 +69,8 @@ public:
   DeflatedGuesser(const std::vector<Field> & _evec, const std::vector<RealD> & _eval, const unsigned int _N)
   : evec(_evec), eval(_eval), N(_N)
   {
-    assert(evec.size()==eval.size());
-    assert(N <= evec.size());
+    GRID_ASSERT(evec.size()==eval.size());
+    GRID_ASSERT(N <= evec.size());
   } 
 
   virtual void operator()(const Field &src,Field &guess) {
@@ -141,11 +141,10 @@ public:
     }
     //postprocessing
     std::cout << GridLogMessage << "Start BlockPromote for loop" << std::endl;
-    for (int j=0;j<Nsrc;j++)
-    {
-    std::cout << GridLogMessage << "BlockProject iter: " << j << std::endl;
-    blockPromote(guess_coarse[j],guess[j],subspace);
-    guess[j].Checkerboard() = src[j].Checkerboard();
+    for (int j=0;j<Nsrc;j++) {
+      std::cout << GridLogMessage << "BlockProject iter: " << j << std::endl;
+      blockPromote(guess_coarse[j],guess[j],subspace);
+      guess[j].Checkerboard() = src[j].Checkerboard();
     }
   };
 

Grid/algorithms/deflation/MultiRHSBlockCGLinalg.h

@@ -160,7 +160,7 @@ public:
     uint64_t words;
 
     nrhs = X.size();
-    assert(X.size()==Y.size());
+    GRID_ASSERT(X.size()==Y.size());
     conformable(X[0],Y[0]);
 
     grid  = X[0].Grid();
@@ -259,7 +259,7 @@ public:
     uint64_t words;
 
     nrhs = X.size();
-    assert(X.size()==Y.size());
+    GRID_ASSERT(X.size()==Y.size());
     conformable(X[0],Y[0]);
 
     grid  = X[0].Grid();
@@ -267,7 +267,7 @@ public:
     vol   = grid->oSites()/rd0;
     words = rd0*sizeof(vector_object)/sizeof(scalar);
     int64_t vw = vol * words;
-    assert(vw == grid->lSites()*sizeof(scalar_object)/sizeof(scalar));
+    GRID_ASSERT(vw == grid->lSites()*sizeof(scalar_object)/sizeof(scalar));
 
     RealD t0 = usecond();
     BLAS_X.resize(nrhs * vw); // cost free if size doesn't change

Grid/algorithms/deflation/MultiRHSBlockProject.h

@@ -61,6 +61,17 @@ public:
   uint64_t coarse_vol;
   uint64_t words;
 
+  ////////////////////////////////////////////////////////////////////////////
+  // Blocking geometry in full local coordinates. Addressing in lSites rather
+  // than (lane,oSite) lets the fine and coarse spaces carry different SIMD
+  // layouts, so an unvectorised coarse space may block a vectorised fine one.
+  ////////////////////////////////////////////////////////////////////////////
+  Coordinate fine_ldimensions;
+  Coordinate coarse_ldimensions;
+  Coordinate block_ldimensions;
+  Coordinate fine_simd;
+  Coordinate coarse_simd;
+
   // Row major layout "C" order:
   // BLAS_V[coarse_vol][nbasis][block_vol][words]
   // BLAS_F[coarse_vol][nrhs][block_vol][words]
@@ -120,9 +131,26 @@ public:
     fine_vol   = fine_grid->lSites();
     coarse_vol = coarse_grid->lSites();
     block_vol = fine_vol/coarse_vol;
-    
+
     words = sizeof(scalar_object)/sizeof(scalar);
 
+    int nd = coarse_grid->_ndimension;
+    GRID_ASSERT(fine_grid->_ndimension == nd);
+
+    fine_ldimensions.resize(nd);
+    coarse_ldimensions.resize(nd);
+    block_ldimensions.resize(nd);
+    fine_simd   = fine_grid->_simd_layout;
+    coarse_simd = coarse_grid->_simd_layout;
+    for(int d=0;d<nd;d++){
+      GRID_ASSERT(fine_grid->_processors[d] == coarse_grid->_processors[d]);
+      fine_ldimensions  [d] = fine_grid->_rdimensions  [d]*fine_grid->_simd_layout  [d];
+      coarse_ldimensions[d] = coarse_grid->_rdimensions[d]*coarse_grid->_simd_layout[d];
+      block_ldimensions [d] = fine_ldimensions[d]/coarse_ldimensions[d];
+      GRID_ASSERT(block_ldimensions[d]*coarse_ldimensions[d] == fine_ldimensions[d]);
+    }
+    GRID_ASSERT(block_vol == fine_vol/coarse_vol);
+
     BLAS_V.resize (fine_vol * words * nbasis );
   }
   void ImportFineGridVectors(std::vector <Field > &vecs, deviceVector<scalar> &blas)
@@ -131,24 +159,18 @@ public:
     typedef typename Field::vector_object vobj;
     //    std::cout << GridLogMessage <<" BlockProjector importing "<<nvec<< " fine grid vectors" <<std::endl;
 
-    assert(vecs[0].Grid()==fine_grid);
-
-    subdivides(coarse_grid,fine_grid); // require they map
+    GRID_ASSERT(vecs[0].Grid()==fine_grid);
 
     int _ndimension = coarse_grid->_ndimension;
-    assert(block_vol == fine_grid->oSites() / coarse_grid->oSites());
-    
-    Coordinate  block_r      (_ndimension);
-    for(int d=0 ; d<_ndimension;d++){
-      block_r[d] = fine_grid->_rdimensions[d] / coarse_grid->_rdimensions[d];
-    }
 
     uint64_t sz = blas.size();
 
     acceleratorMemSet(&blas[0],0,blas.size()*sizeof(scalar));
 
     Coordinate fine_rdimensions = fine_grid->_rdimensions;
-    Coordinate coarse_rdimensions = coarse_grid->_rdimensions;
+    Coordinate coarse_l = coarse_ldimensions;
+    Coordinate block_l  = block_ldimensions;
+    Coordinate fsimd    = fine_simd;
     int64_t bv= block_vol;
     for(int v=0;v<vecs.size();v++){
 
@@ -162,9 +184,7 @@ public:
 
       // loop over fine sites
       const int Nsimd = vobj::Nsimd();
-      //      std::cout << "sz "<<sz<<std::endl;
-      //      std::cout << "prod "<<Nsimd * coarse_grid->oSites() * block_vol * nvec * words<<std::endl;
-      assert(sz == Nsimd * coarse_grid->oSites() * block_vol * nvec * words);
+      GRID_ASSERT(sz == coarse_vol * block_vol * nvec * words);
       uint64_t lwords= words; // local variable for copy in to GPU
       accelerator_for(sf,osites,Nsimd,{
 #ifdef GRID_SIMT
@@ -175,30 +195,30 @@ public:
 #endif
 	  // One thread per fine site
 	  Coordinate coor_f(_ndimension);
+	  Coordinate coor_l(_ndimension);
 	  Coordinate coor_b(_ndimension);
 	  Coordinate coor_c(_ndimension);
 
-	  // Fine site to fine coor
+	  // Fine (oSite,lane) to full local coor
 	  Lexicographic::CoorFromIndex(coor_f,sf,fine_rdimensions);
+	  Lexicographic::CoorFromIndex(coor_l,lane,fsimd);
+	  for(int d=0;d<_ndimension;d++) coor_f[d] += fine_rdimensions[d]*coor_l[d];
+
+	  for(int d=0;d<_ndimension;d++) coor_b[d] = coor_f[d]%block_l[d];
+	  for(int d=0;d<_ndimension;d++) coor_c[d] = coor_f[d]/block_l[d];
 
-	  for(int d=0;d<_ndimension;d++) coor_b[d] = coor_f[d]%block_r[d];
-	  for(int d=0;d<_ndimension;d++) coor_c[d] = coor_f[d]/block_r[d];
-	  
 	  int sc;// coarse site
 	  int sb;// block site
-	  Lexicographic::IndexFromCoor(coor_c,sc,coarse_rdimensions);
-	  Lexicographic::IndexFromCoor(coor_b,sb,block_r);
+	  Lexicographic::IndexFromCoor(coor_c,sc,coarse_l);
+	  Lexicographic::IndexFromCoor(coor_b,sb,block_l);
 
           scalar_object data = extractLane(lane,fineData[sf]);
 
-	  // BLAS layout address calculation
-	  // words * block_vol * nbasis x coarse_vol
-	  // coarse oSite x block vole x lanes
-	  int64_t site = (lane*osites + sc*bv)*nvec
-   	               + v*bv
+	  // BLAS_F[coarse_vol][nvec][block_vol][words]
+	  int64_t site = (sc*nvec + v)*bv
 	               + sb;
 
-	  //	  assert(site*lwords<sz);
+	  //	  GRID_ASSERT(site*lwords<sz);
 
 	  scalar_object * ptr = (scalar_object *)&blasData_p[site*lwords];
 
@@ -213,25 +233,332 @@ public:
       //      std::cout << " BlockProjector imported vector"<<v<<std::endl;
     }
   }
+  ////////////////////////////////////////////////////////////////////////////
+  // Import direct from multiRHS fine order, avoiding the unpack to a vector
+  // of single RHS fields.
+  //
+  //   fine mrhs grid  : rhs is dimension 0, undistributed, unvectorised
+  //   Grid  order     : F[fine_vol][nrhs][words]
+  //   BLAS  order     : BLAS_F[lane][coarse_vol][nrhs][block_vol][words]
+  //
+  // The gather of block_vol from fine_vol, and the transpose of nrhs against
+  // block_vol, are the irreducible part: this is not the identity.
+  ////////////////////////////////////////////////////////////////////////////
+  void ImportFineGridMrhsVectors(Field &vec_mrhs, deviceVector<scalar> &blas)
+  {
+    typedef typename Field::vector_object vobj;
+
+    GridBase *fine_mrhs_grid = vec_mrhs.Grid();
+    int _ndimension = coarse_grid->_ndimension;
+
+    GRID_ASSERT(fine_mrhs_grid->_ndimension == _ndimension+1);
+    GRID_ASSERT(fine_mrhs_grid->_simd_layout[0] == 1);
+    GRID_ASSERT(fine_mrhs_grid->_processors[0] == 1);
+    for(int d=0;d<_ndimension;d++){
+      GRID_ASSERT(fine_mrhs_grid->_rdimensions[d+1] == fine_grid->_rdimensions[d]);
+      GRID_ASSERT(fine_mrhs_grid->_simd_layout[d+1] == fine_grid->_simd_layout[d]);
+    }
+    int nvec = fine_mrhs_grid->_rdimensions[0];   // nrhs
+
+    uint64_t sz = blas.size();
+    acceleratorMemSet(&blas[0],0,blas.size()*sizeof(scalar));
+
+    Coordinate fine_mrhs_rdimensions = fine_mrhs_grid->_rdimensions;
+    Coordinate fine_rdimensions      = fine_grid->_rdimensions;
+    Coordinate coarse_l = coarse_ldimensions;
+    Coordinate block_l  = block_ldimensions;
+    Coordinate fsimd    = fine_simd;
+    int64_t bv= block_vol;
+
+    autoView( fineData   , vec_mrhs, AcceleratorRead);
+    auto blasData_p  = &blas[0];
+    auto fineData_p  = &fineData[0];
+
+    int64_t osites     = fine_grid->oSites();       // D dimensional
+    int64_t osites_hi  = fine_mrhs_grid->oSites();  // nvec * osites
+
+    const int Nsimd = vobj::Nsimd();
+    GRID_ASSERT(sz == coarse_vol * block_vol * nvec * words);
+    uint64_t lwords= words;
+    int64_t lnvec = nvec;
+
+    accelerator_for(sfr,osites_hi,Nsimd,{
+#ifdef GRID_SIMT
+      {
+	int lane=acceleratorSIMTlane(Nsimd); // buffer lane
+#else
+	for(int lane=0;lane<Nsimd;lane++) {
+#endif
+	  Coordinate coor_hi(_ndimension+1);
+	  Coordinate coor_f(_ndimension);
+	  Coordinate coor_l(_ndimension);
+	  Coordinate coor_b(_ndimension);
+	  Coordinate coor_c(_ndimension);
+
+	  // rhs is dimension 0 of the D+1 grid
+	  Lexicographic::CoorFromIndex(coor_hi,sfr,fine_mrhs_rdimensions);
+	  int v = coor_hi[0];
+	  for(int d=0;d<_ndimension;d++) coor_f[d] = coor_hi[d+1];
+
+	  Lexicographic::CoorFromIndex(coor_l,lane,fsimd);
+	  for(int d=0;d<_ndimension;d++) coor_f[d] += fine_rdimensions[d]*coor_l[d];
+
+	  for(int d=0;d<_ndimension;d++) coor_b[d] = coor_f[d]%block_l[d];
+	  for(int d=0;d<_ndimension;d++) coor_c[d] = coor_f[d]/block_l[d];
+
+	  int sc;// coarse site
+	  int sb;// block site
+	  Lexicographic::IndexFromCoor(coor_c,sc,coarse_l);
+	  Lexicographic::IndexFromCoor(coor_b,sb,block_l);
+
+	  scalar_object data = extractLane(lane,fineData[sfr]);
+
+	  int64_t site = (sc*lnvec + v)*bv
+	               + sb;
+
+	  scalar_object * ptr = (scalar_object *)&blasData_p[site*lwords];
+	  *ptr = data;
+#ifdef GRID_SIMT
+      }
+#else
+      }
+#endif
+    });
+  }
+
+  ////////////////////////////////////////////////////////////////////////////
+  // Export direct to multiRHS coarse order.
+  //
+  //   Grid  order : C[coarse_vol][nrhs][nbasis]
+  //   BLAS  order : BLAS_C[lane][coarse_vol][nrhs][nbasis]
+  //
+  // At Nsimd()==1 these are the same sequence of addresses.
+  ////////////////////////////////////////////////////////////////////////////
+  template<class vobj>
+  void ExportCoarseGridMrhsVectors(Lattice<vobj> &vec_mrhs, deviceVector<scalar> &blas)
+  {
+    typedef typename vobj::scalar_object coarse_scalar_object;
+
+    GridBase *coarse_mrhs_grid = vec_mrhs.Grid();
+    int _ndimension = coarse_grid->_ndimension;
+
+    GRID_ASSERT(coarse_mrhs_grid->_ndimension == _ndimension+1);
+    GRID_ASSERT(coarse_mrhs_grid->_simd_layout[0] == 1);
+    GRID_ASSERT(coarse_mrhs_grid->_processors[0] == 1);
+    for(int d=0;d<_ndimension;d++){
+      GRID_ASSERT(coarse_mrhs_grid->_rdimensions[d+1] == coarse_grid->_rdimensions[d]);
+      GRID_ASSERT(coarse_mrhs_grid->_simd_layout[d+1] == coarse_grid->_simd_layout[d]);
+    }
+    int nvec = coarse_mrhs_grid->_rdimensions[0];   // nrhs
+
+    Coordinate coarse_mrhs_rdimensions = coarse_mrhs_grid->_rdimensions;
+    Coordinate coarse_rdimensions      = coarse_grid->_rdimensions;
+    Coordinate coarse_l = coarse_ldimensions;
+    Coordinate csimd    = coarse_simd;
+
+    autoView( coarseData   , vec_mrhs, AcceleratorWrite);
+    auto blasData_p    = &blas[0];
+    auto coarseData_p  = &coarseData[0];
+
+    int64_t osites    = coarse_grid->oSites();       // D dimensional
+    int64_t osites_hi = coarse_mrhs_grid->oSites();  // nvec * osites
+
+    const int Nsimd = vobj::Nsimd();
+    uint64_t cwords=sizeof(typename vobj::scalar_object)/sizeof(scalar);
+    GRID_ASSERT(cwords==nbasis);
+    int64_t lnvec = nvec;
+
+    accelerator_for(scr,osites_hi,Nsimd,{
+#ifdef GRID_SIMT
+      {
+	int lane=acceleratorSIMTlane(Nsimd); // buffer lane
+#else
+	for(int lane=0;lane<Nsimd;lane++) {
+#endif
+	  Coordinate coor_hi(_ndimension+1);
+	  Coordinate coor_l(_ndimension);
+	  Coordinate coor_c(_ndimension);
+
+	  Lexicographic::CoorFromIndex(coor_hi,scr,coarse_mrhs_rdimensions);
+	  int v = coor_hi[0];
+	  for(int d=0;d<_ndimension;d++) coor_c[d] = coor_hi[d+1];
+
+	  Lexicographic::CoorFromIndex(coor_l,lane,csimd);
+	  for(int d=0;d<_ndimension;d++) coor_c[d] += coarse_rdimensions[d]*coor_l[d];
+
+	  int sc;
+	  Lexicographic::IndexFromCoor(coor_c,sc,coarse_l);
+
+	  int64_t blas_site = (sc*lnvec + v)*cwords;
+	  coarse_scalar_object * ptr = (coarse_scalar_object *)&blasData_p[blas_site];
+	  coarse_scalar_object data = *ptr;
+	  insertLane(lane,coarseData[scr],data);
+#ifdef GRID_SIMT
+      }
+#else
+      }
+#endif
+    });
+  }
+
+  ////////////////////////////////////////////////////////////////////////////
+  // Reverse directions: BLAS_F -> fine mrhs field, coarse mrhs field -> BLAS_C
+  ////////////////////////////////////////////////////////////////////////////
+  void ExportFineGridMrhsVectors(Field &vec_mrhs, deviceVector<scalar> &blas)
+  {
+    typedef typename Field::vector_object vobj;
+
+    GridBase *fine_mrhs_grid = vec_mrhs.Grid();
+    int _ndimension = coarse_grid->_ndimension;
+
+    GRID_ASSERT(fine_mrhs_grid->_ndimension == _ndimension+1);
+    GRID_ASSERT(fine_mrhs_grid->_simd_layout[0] == 1);
+    for(int d=0;d<_ndimension;d++){
+      GRID_ASSERT(fine_mrhs_grid->_rdimensions[d+1] == fine_grid->_rdimensions[d]);
+      GRID_ASSERT(fine_mrhs_grid->_simd_layout[d+1] == fine_grid->_simd_layout[d]);
+    }
+    int nvec = fine_mrhs_grid->_rdimensions[0];
+
+    Coordinate fine_mrhs_rdimensions = fine_mrhs_grid->_rdimensions;
+    Coordinate fine_rdimensions      = fine_grid->_rdimensions;
+    Coordinate coarse_l = coarse_ldimensions;
+    Coordinate block_l  = block_ldimensions;
+    Coordinate fsimd    = fine_simd;
+    int64_t bv= block_vol;
+
+    autoView( fineData   , vec_mrhs, AcceleratorWrite);
+    auto blasData_p  = &blas[0];
+    auto fineData_p  = &fineData[0];
+
+    int64_t osites     = fine_grid->oSites();
+    int64_t osites_hi  = fine_mrhs_grid->oSites();
+
+    const int Nsimd = vobj::Nsimd();
+    uint64_t lwords= words;
+    int64_t lnvec = nvec;
+
+    accelerator_for(sfr,osites_hi,Nsimd,{
+#ifdef GRID_SIMT
+      {
+	int lane=acceleratorSIMTlane(Nsimd);
+#else
+	for(int lane=0;lane<Nsimd;lane++) {
+#endif
+	  Coordinate coor_hi(_ndimension+1);
+	  Coordinate coor_f(_ndimension);
+	  Coordinate coor_l(_ndimension);
+	  Coordinate coor_b(_ndimension);
+	  Coordinate coor_c(_ndimension);
+
+	  Lexicographic::CoorFromIndex(coor_hi,sfr,fine_mrhs_rdimensions);
+	  int v = coor_hi[0];
+	  for(int d=0;d<_ndimension;d++) coor_f[d] = coor_hi[d+1];
+
+	  Lexicographic::CoorFromIndex(coor_l,lane,fsimd);
+	  for(int d=0;d<_ndimension;d++) coor_f[d] += fine_rdimensions[d]*coor_l[d];
+
+	  for(int d=0;d<_ndimension;d++) coor_b[d] = coor_f[d]%block_l[d];
+	  for(int d=0;d<_ndimension;d++) coor_c[d] = coor_f[d]/block_l[d];
+
+	  int sc,sb;
+	  Lexicographic::IndexFromCoor(coor_c,sc,coarse_l);
+	  Lexicographic::IndexFromCoor(coor_b,sb,block_l);
+
+	  int64_t site = (sc*lnvec + v)*bv
+	               + sb;
+
+	  scalar_object * ptr = (scalar_object *)&blasData_p[site*lwords];
+	  scalar_object data = *ptr;
+	  insertLane(lane,fineData[sfr],data);
+#ifdef GRID_SIMT
+      }
+#else
+      }
+#endif
+    });
+  }
+
+  template<class vobj>
+  void ImportCoarseGridMrhsVectors(Lattice<vobj> &vec_mrhs, deviceVector<scalar> &blas)
+  {
+    typedef typename vobj::scalar_object coarse_scalar_object;
+
+    GridBase *coarse_mrhs_grid = vec_mrhs.Grid();
+    int _ndimension = coarse_grid->_ndimension;
+
+    GRID_ASSERT(coarse_mrhs_grid->_ndimension == _ndimension+1);
+    GRID_ASSERT(coarse_mrhs_grid->_simd_layout[0] == 1);
+    for(int d=0;d<_ndimension;d++){
+      GRID_ASSERT(coarse_mrhs_grid->_rdimensions[d+1] == coarse_grid->_rdimensions[d]);
+      GRID_ASSERT(coarse_mrhs_grid->_simd_layout[d+1] == coarse_grid->_simd_layout[d]);
+    }
+    int nvec = coarse_mrhs_grid->_rdimensions[0];
+
+    Coordinate coarse_mrhs_rdimensions = coarse_mrhs_grid->_rdimensions;
+    Coordinate coarse_rdimensions      = coarse_grid->_rdimensions;
+    Coordinate coarse_l = coarse_ldimensions;
+    Coordinate csimd    = coarse_simd;
+
+    autoView( coarseData   , vec_mrhs, AcceleratorRead);
+    auto blasData_p    = &blas[0];
+    auto coarseData_p  = &coarseData[0];
+
+    int64_t osites    = coarse_grid->oSites();
+    int64_t osites_hi = coarse_mrhs_grid->oSites();
+
+    const int Nsimd = vobj::Nsimd();
+    uint64_t cwords=sizeof(typename vobj::scalar_object)/sizeof(scalar);
+    GRID_ASSERT(cwords==nbasis);
+    int64_t lnvec = nvec;
+
+    accelerator_for(scr,osites_hi,Nsimd,{
+#ifdef GRID_SIMT
+      {
+	int lane=acceleratorSIMTlane(Nsimd);
+#else
+	for(int lane=0;lane<Nsimd;lane++) {
+#endif
+	  Coordinate coor_hi(_ndimension+1);
+	  Coordinate coor_l(_ndimension);
+	  Coordinate coor_c(_ndimension);
+
+	  Lexicographic::CoorFromIndex(coor_hi,scr,coarse_mrhs_rdimensions);
+	  int v = coor_hi[0];
+	  for(int d=0;d<_ndimension;d++) coor_c[d] = coor_hi[d+1];
+
+	  Lexicographic::CoorFromIndex(coor_l,lane,csimd);
+	  for(int d=0;d<_ndimension;d++) coor_c[d] += coarse_rdimensions[d]*coor_l[d];
+
+	  int sc;
+	  Lexicographic::IndexFromCoor(coor_c,sc,coarse_l);
+
+	  coarse_scalar_object data = extractLane(lane,coarseData[scr]);
+
+	  int64_t blas_site = (sc*lnvec + v)*cwords;
+	  coarse_scalar_object * ptr = (coarse_scalar_object *)&blasData_p[blas_site];
+	  *ptr = data;
+#ifdef GRID_SIMT
+      }
+#else
+      }
+#endif
+    });
+  }
+
   void ExportFineGridVectors(std::vector <Field> &vecs, deviceVector<scalar> &blas)
   {
     typedef typename Field::vector_object vobj;
 
     int nvec = vecs.size();
 
-    assert(vecs[0].Grid()==fine_grid);
-
-    subdivides(coarse_grid,fine_grid); // require they map
+    GRID_ASSERT(vecs[0].Grid()==fine_grid);
 
     int _ndimension = coarse_grid->_ndimension;
-    assert(block_vol == fine_grid->oSites() / coarse_grid->oSites());
-    
-    Coordinate  block_r      (_ndimension);
-    for(int d=0 ; d<_ndimension;d++){
-      block_r[d] = fine_grid->_rdimensions[d] / coarse_grid->_rdimensions[d];
-    }
+
     Coordinate fine_rdimensions = fine_grid->_rdimensions;
-    Coordinate coarse_rdimensions = coarse_grid->_rdimensions;
+    Coordinate coarse_l = coarse_ldimensions;
+    Coordinate block_l  = block_ldimensions;
+    Coordinate fsimd    = fine_simd;
 
     //    std::cout << " export fine Blas norm "<<blasNorm2(blas)<<std::endl;
 
@@ -259,23 +586,24 @@ public:
 #endif
 	  // One thread per fine site
 	  Coordinate coor_f(_ndimension);
+	  Coordinate coor_l(_ndimension);
 	  Coordinate coor_b(_ndimension);
 	  Coordinate coor_c(_ndimension);
 
 	  Lexicographic::CoorFromIndex(coor_f,sf,fine_rdimensions);
+	  Lexicographic::CoorFromIndex(coor_l,lane,fsimd);
+	  for(int d=0;d<_ndimension;d++) coor_f[d] += fine_rdimensions[d]*coor_l[d];
+
+	  for(int d=0;d<_ndimension;d++) coor_b[d] = coor_f[d]%block_l[d];
+	  for(int d=0;d<_ndimension;d++) coor_c[d] = coor_f[d]/block_l[d];
 
-	  for(int d=0;d<_ndimension;d++) coor_b[d] = coor_f[d]%block_r[d];
-	  for(int d=0;d<_ndimension;d++) coor_c[d] = coor_f[d]/block_r[d];
-	  
 	  int sc;
 	  int sb;
-	  Lexicographic::IndexFromCoor(coor_c,sc,coarse_rdimensions);
-	  Lexicographic::IndexFromCoor(coor_b,sb,block_r);
+	  Lexicographic::IndexFromCoor(coor_c,sc,coarse_l);
+	  Lexicographic::IndexFromCoor(coor_b,sb,block_l);
 
-	  // BLAS layout address calculation
-	  // words * block_vol * nbasis x coarse_vol 	  
-	  int64_t site = (lane*osites + sc*bv)*nvec
-   	               + v*bv
+	  // BLAS_F[coarse_vol][nvec][block_vol][words]
+	  int64_t site = (sc*nvec + v)*bv
 	               + sb;
 
 	  scalar_object * ptr = (scalar_object *)&blasData_p[site*lwords];
@@ -299,14 +627,16 @@ public:
 
     //    std::cout << " BlockProjector importing "<<nvec<< " coarse grid vectors" <<std::endl;
 
-    assert(vecs[0].Grid()==coarse_grid);
+    GRID_ASSERT(vecs[0].Grid()==coarse_grid);
 
     int _ndimension = coarse_grid->_ndimension;
 
     uint64_t sz = blas.size();
 
     Coordinate coarse_rdimensions = coarse_grid->_rdimensions;
-    
+    Coordinate coarse_l = coarse_ldimensions;
+    Coordinate csimd    = coarse_simd;
+
     for(int v=0;v<vecs.size();v++){
 
       //      std::cout << " BlockProjector importing coarse vector"<<v<<" "<<norm2(vecs[v])<<std::endl;
@@ -320,7 +650,7 @@ public:
       // loop over fine sites
       const int Nsimd = vobj::Nsimd();
       uint64_t cwords=sizeof(typename vobj::scalar_object)/sizeof(scalar);
-      assert(cwords==nbasis);
+      GRID_ASSERT(cwords==nbasis);
       
       accelerator_for(sc,osites,Nsimd,{
 #ifdef GRID_SIMT
@@ -330,8 +660,17 @@ public:
 	  for(int lane=0;lane<Nsimd;lane++) {
 #endif
            // C_br per site
-	    int64_t blas_site = (lane*osites + sc)*nvec*cwords + v*cwords;
-	    
+	    Coordinate coor_c(_ndimension);
+	    Coordinate coor_l(_ndimension);
+	    Lexicographic::CoorFromIndex(coor_c,sc,coarse_rdimensions);
+	    Lexicographic::CoorFromIndex(coor_l,lane,csimd);
+	    for(int d=0;d<_ndimension;d++) coor_c[d] += coarse_rdimensions[d]*coor_l[d];
+
+	    int scl;
+	    Lexicographic::IndexFromCoor(coor_c,scl,coarse_l);
+
+	    int64_t blas_site = (scl*nvec + v)*cwords;
+
 	    coarse_scalar_object data = extractLane(lane,coarseData[sc]);
 
 	    coarse_scalar_object * ptr = (coarse_scalar_object *)&blasData_p[blas_site];
@@ -353,14 +692,16 @@ public:
     typedef typename vobj::scalar_object coarse_scalar_object;
     //    std::cout << GridLogMessage<<" BlockProjector exporting "<<nvec<< " coarse grid vectors" <<std::endl;
 
-    assert(vecs[0].Grid()==coarse_grid);
+    GRID_ASSERT(vecs[0].Grid()==coarse_grid);
 
     int _ndimension = coarse_grid->_ndimension;
     
     uint64_t sz = blas.size();
 
     Coordinate coarse_rdimensions = coarse_grid->_rdimensions;
-    
+    Coordinate coarse_l = coarse_ldimensions;
+    Coordinate csimd    = coarse_simd;
+
     //    std::cout << " export coarsee Blas norm "<<blasNorm2(blas)<<std::endl;
     for(int v=0;v<vecs.size();v++){
 
@@ -375,7 +716,7 @@ public:
       // loop over fine sites
       const int Nsimd = vobj::Nsimd();
       uint64_t cwords=sizeof(typename vobj::scalar_object)/sizeof(scalar);
-      assert(cwords==nbasis);
+      GRID_ASSERT(cwords==nbasis);
       
       accelerator_for(sc,osites,Nsimd,{
 	  // Wrap in a macro "FOR_ALL_LANES(lane,{ ... });
@@ -385,7 +726,16 @@ public:
 #else
 	  for(int lane=0;lane<Nsimd;lane++) {
 #endif
-	    int64_t blas_site = (lane*osites + sc)*nvec*cwords + v*cwords;
+	    Coordinate coor_c(_ndimension);
+	    Coordinate coor_l(_ndimension);
+	    Lexicographic::CoorFromIndex(coor_c,sc,coarse_rdimensions);
+	    Lexicographic::CoorFromIndex(coor_l,lane,csimd);
+	    for(int d=0;d<_ndimension;d++) coor_c[d] += coarse_rdimensions[d]*coor_l[d];
+
+	    int scl;
+	    Lexicographic::IndexFromCoor(coor_c,scl,coarse_l);
+
+	    int64_t blas_site = (scl*nvec + v)*cwords;
 	    coarse_scalar_object * ptr = (coarse_scalar_object *)&blasData_p[blas_site];
 	    coarse_scalar_object data = *ptr;
 	    insertLane(lane,coarseData[sc],data);
@@ -409,7 +759,7 @@ public:
     int nrhs=fine.size();
     int _nbasis = sizeof(typename cobj::scalar_object)/sizeof(scalar);
     //    std::cout << "blockProject nbasis " <<nbasis<<" " << _nbasis<<std::endl;
-    assert(nbasis==_nbasis);
+    GRID_ASSERT(nbasis==_nbasis);
     
     BLAS_F.resize (fine_vol * words * nrhs );
     BLAS_C.resize (coarse_vol * nbasis * nrhs );
@@ -464,7 +814,7 @@ public:
   {
     int nrhs=fine.size();
     int _nbasis = sizeof(typename cobj::scalar_object)/sizeof(scalar);
-    assert(nbasis==_nbasis);
+    GRID_ASSERT(nbasis==_nbasis);
     
     BLAS_F.resize (fine_vol * words * nrhs );
     BLAS_C.resize (coarse_vol * nbasis * nrhs );
@@ -508,6 +858,168 @@ public:
     ExportFineGridVectors(fine, BLAS_F);
     //    std::cout << " exported "<<std::endl;
   }
+
+  ////////////////////////////////////////////////////////////////////////////
+  // multiRHS ordered interfaces.  The GEMM is identical; only the import and
+  // export differ, so those are the whole of the layout question.
+  ////////////////////////////////////////////////////////////////////////////
+  template<class cobj>
+  void blockProject(Field &fine_mrhs,Lattice<cobj> &coarse_mrhs)
+  {
+    int nrhs = fine_mrhs.Grid()->_rdimensions[0];
+    int _nbasis = sizeof(typename cobj::scalar_object)/sizeof(scalar);
+    GRID_ASSERT(nbasis==_nbasis);
+    GRID_ASSERT(coarse_mrhs.Grid()->_rdimensions[0]==nrhs);
+
+    BLAS_F.resize (fine_vol * words * nrhs );
+    BLAS_C.resize (coarse_vol * nbasis * nrhs );
+
+    ImportFineGridMrhsVectors(fine_mrhs,BLAS_F);
+    ProjectBLAS(nrhs);
+    ExportCoarseGridMrhsVectors(coarse_mrhs,BLAS_C);
+  }
+
+  template<class cobj>
+  void blockPromote(Field &fine_mrhs,Lattice<cobj> &coarse_mrhs)
+  {
+    int nrhs = fine_mrhs.Grid()->_rdimensions[0];
+    int _nbasis = sizeof(typename cobj::scalar_object)/sizeof(scalar);
+    GRID_ASSERT(nbasis==_nbasis);
+    GRID_ASSERT(coarse_mrhs.Grid()->_rdimensions[0]==nrhs);
+
+    BLAS_F.resize (fine_vol * words * nrhs );
+    BLAS_C.resize (coarse_vol * nbasis * nrhs );
+
+    ImportCoarseGridMrhsVectors(coarse_mrhs,BLAS_C);
+    PromoteBLAS(nrhs);
+    ExportFineGridMrhsVectors(fine_mrhs,BLAS_F);
+  }
+
+  ////////////////////////////////////////////////////////////////////////////
+  // Mixed orderings. A single RHS fine operator produces a vector of fine
+  // fields with no packing; the coarse side is still wanted in mrhs order.
+  ////////////////////////////////////////////////////////////////////////////
+  template<class cobj>
+  void blockProject(std::vector<Field> &fine,Lattice<cobj> &coarse_mrhs)
+  {
+    int nrhs = fine.size();
+    int _nbasis = sizeof(typename cobj::scalar_object)/sizeof(scalar);
+    GRID_ASSERT(nbasis==_nbasis);
+    GRID_ASSERT(coarse_mrhs.Grid()->_rdimensions[0]==nrhs);
+
+    BLAS_F.resize (fine_vol * words * nrhs );
+    BLAS_C.resize (coarse_vol * nbasis * nrhs );
+
+    ImportFineGridVectors(fine,BLAS_F);
+    ProjectBLAS(nrhs);
+    ExportCoarseGridMrhsVectors(coarse_mrhs,BLAS_C);
+  }
+
+  template<class cobj>
+  void blockProject(Field &fine_mrhs,std::vector< Lattice<cobj> > &coarse)
+  {
+    int nrhs = fine_mrhs.Grid()->_rdimensions[0];
+    int _nbasis = sizeof(typename cobj::scalar_object)/sizeof(scalar);
+    GRID_ASSERT(nbasis==_nbasis);
+    GRID_ASSERT(coarse.size()==nrhs);
+
+    BLAS_F.resize (fine_vol * words * nrhs );
+    BLAS_C.resize (coarse_vol * nbasis * nrhs );
+
+    ImportFineGridMrhsVectors(fine_mrhs,BLAS_F);
+    ProjectBLAS(nrhs);
+    ExportCoarseGridVectors(coarse,BLAS_C);
+  }
+
+  template<class cobj>
+  void blockPromote(std::vector<Field> &fine,Lattice<cobj> &coarse_mrhs)
+  {
+    int nrhs = fine.size();
+    int _nbasis = sizeof(typename cobj::scalar_object)/sizeof(scalar);
+    GRID_ASSERT(nbasis==_nbasis);
+    GRID_ASSERT(coarse_mrhs.Grid()->_rdimensions[0]==nrhs);
+
+    BLAS_F.resize (fine_vol * words * nrhs );
+    BLAS_C.resize (coarse_vol * nbasis * nrhs );
+
+    ImportCoarseGridMrhsVectors(coarse_mrhs,BLAS_C);
+    PromoteBLAS(nrhs);
+    ExportFineGridVectors(fine,BLAS_F);
+  }
+
+  template<class cobj>
+  void blockPromote(Field &fine_mrhs,std::vector< Lattice<cobj> > &coarse)
+  {
+    int nrhs = fine_mrhs.Grid()->_rdimensions[0];
+    int _nbasis = sizeof(typename cobj::scalar_object)/sizeof(scalar);
+    GRID_ASSERT(nbasis==_nbasis);
+    GRID_ASSERT(coarse.size()==nrhs);
+
+    BLAS_F.resize (fine_vol * words * nrhs );
+    BLAS_C.resize (coarse_vol * nbasis * nrhs );
+
+    ImportCoarseGridVectors(coarse,BLAS_C);
+    PromoteBLAS(nrhs);
+    ExportFineGridMrhsVectors(fine_mrhs,BLAS_F);
+  }
+
+  ////////////////////////////////////////////////////////////////////////////
+  // Pointer tables and the batched GEMM, shared by both orderings
+  ////////////////////////////////////////////////////////////////////////////
+  void BLASPointers(int nrhs,
+		    deviceVector<scalar *> &Vd,
+		    deviceVector<scalar *> &Fd,
+		    deviceVector<scalar *> &Cd)
+  {
+    for(int c=0;c<coarse_vol;c++){
+      scalar * Vh = & BLAS_V[c*nbasis*block_vol*words];
+      scalar * Fh = & BLAS_F[c*nrhs*block_vol*words];
+      scalar * Ch = & BLAS_C[c*nrhs*nbasis];
+      acceleratorPut(Vd[c],Vh);
+      acceleratorPut(Fd[c],Fh);
+      acceleratorPut(Cd[c],Ch);
+    }
+  }
+
+  // C_br = V^dag F
+  void ProjectBLAS(int nrhs)
+  {
+    deviceVector<scalar *> Vd(coarse_vol);
+    deviceVector<scalar *> Fd(coarse_vol);
+    deviceVector<scalar *> Cd(coarse_vol);
+    BLASPointers(nrhs,Vd,Fd,Cd);
+
+    GridBLAS BLAS;
+    int64_t vw = block_vol * words;
+    BLAS.gemmBatched(GridBLAS_OP_C,GridBLAS_OP_N,
+		     nbasis,nrhs,vw,
+		     scalar(1.0),
+		     Vd,
+		     Fd,
+		     scalar(0.0),  // wipe out C
+		     Cd);
+    BLAS.synchronise();
+  }
+
+  // F_xr = Vxb Cbr
+  void PromoteBLAS(int nrhs)
+  {
+    deviceVector<scalar *> Vd(coarse_vol);
+    deviceVector<scalar *> Fd(coarse_vol);
+    deviceVector<scalar *> Cd(coarse_vol);
+    BLASPointers(nrhs,Vd,Fd,Cd);
+
+    GridBLAS BLAS;
+    int64_t vw = block_vol * words;
+    BLAS.gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N,
+		     vw,nrhs,nbasis,
+		     scalar(1.0),
+		     Vd,
+		     Cd,
+		     scalar(0.0),  // wipe out F
+		     Fd);
+    BLAS.synchronise();
+  }
 };
 
 NAMESPACE_END(Grid);

Grid/algorithms/deflation/MultiRHSDeflation.h

@@ -98,7 +98,7 @@ public:
   void ImportEigenVector(Field &evec,RealD &_eval, int ev)
   {
     //    std::cout << " ev " <<ev<<" eval "<<_eval<< std::endl;
-    assert(ev<eval.size());
+    GRID_ASSERT(ev<eval.size());
     eval[ev] = _eval;
 
     int64_t offset = ev*vol*words;
@@ -113,7 +113,7 @@ public:
   // Could use to import a batch of eigenvectors
   void ImportEigenBasis(std::vector<Field> &evec,std::vector<RealD> &_eval, int _ev0, int _nev)
   {
-    assert(_ev0+_nev<=evec.size());
+    GRID_ASSERT(_ev0+_nev<=evec.size());
 
     Allocate(_nev,evec[0].Grid());
     
@@ -126,8 +126,8 @@ public:
   void DeflateSources(std::vector<Field> &source,std::vector<Field> & guess)
   {
     int nrhs = source.size();
-    assert(source.size()==guess.size());
-    assert(grid == guess[0].Grid());
+    GRID_ASSERT(source.size()==guess.size());
+    GRID_ASSERT(grid == guess[0].Grid());
     conformable(guess[0],source[0]);
 
     int64_t vw = vol * words;
@@ -189,7 +189,7 @@ public:
 		     Cd);
     BLAS.synchronise();
 
-    assert(BLAS_C.size()==nev*nrhs);
+    GRID_ASSERT(BLAS_C.size()==nev*nrhs);
 
     std::vector<scalar> HOST_C(BLAS_C.size());      // nrhs . nev -- the coefficients 
     acceleratorCopyFromDevice(&BLAS_C[0],&HOST_C[0],BLAS_C.size()*sizeof(scalar));

Grid/algorithms/iterative/AdefGeneric.h

@@ -53,7 +53,22 @@ class TwoLevelCG : public LinearFunction<Field>
   // Fine operator, Smoother, CoarseSolver
   LinearOperatorBase<Field>   &_FineLinop;
   LinearFunction<Field>   &_Smoother;
-  
+
+  GridStopWatch ProjectTimer;
+  GridStopWatch PromoteTimer;
+  GridStopWatch CoarseTimer;
+  GridStopWatch SmoothTimer;
+  GridStopWatch MatrixTimer;
+  GridStopWatch M3Timer;
+  GridStopWatch LinalgTimer;
+
+  int64_t M3Calls;
+  int64_t SmoothCalls;
+  int64_t MatrixCalls;
+  int64_t ProjectCalls;
+  int64_t CoarseCalls;
+  int64_t PromoteCalls;
+
   // more most opertor functions
   TwoLevelCG(RealD tol,
 	     Integer maxit,
@@ -103,12 +118,20 @@ class TwoLevelCG : public LinearFunction<Field>
     RealD tn;
     
     GridStopWatch HDCGTimer;
+    ProjectTimer.Reset();
+    PromoteTimer.Reset();
+    CoarseTimer.Reset();
+    SmoothTimer.Reset();
+    MatrixTimer.Reset();
+    M3Timer.Reset();
+    LinalgTimer.Reset();
+    M3Calls = SmoothCalls = MatrixCalls = ProjectCalls = CoarseCalls = PromoteCalls = 0;
     HDCGTimer.Start();
     //////////////////////////
     // x0 = Vstart -- possibly modify guess
     //////////////////////////
     Vstart(x,src);
-    
+
     // r0 = b -A x0
     _FineLinop.HermOp(x,mmp[0]);
     axpy (r, -1.0,mmp[0], src);    // Recomputes r=src-Ax0
@@ -145,33 +168,40 @@ class TwoLevelCG : public LinearFunction<Field>
       int peri_kp = (k+1) % mmax;
 
       rtz=rtzp;
+      M3Timer.Start();
       d= PcgM3(p[peri_k],mmp[peri_k]);
+      M3Timer.Stop();
+      M3Calls++;
       a = rtz/d;
-    
+
       // Memorise this
       pAp[peri_k] = d;
-      
+
+      LinalgTimer.Start();
       axpy(x,a,p[peri_k],x);
       RealD rn = axpy_norm(r,-a,mmp[peri_k],r);
+      LinalgTimer.Stop();
 
       // Compute z = M x
       PcgM1(r,z);
-      
+
       {
 	RealD n1,n2;
 	n1=norm2(r);
 	n2=norm2(z);
 	std::cout << GridLogMessage<<"HDCG::fPcg iteration "<<k<<" : vector r,z "<<n1<<" "<<n2<<"\n";
       }
+      LinalgTimer.Start();
       rtzp =real(innerProduct(r,z));
+      LinalgTimer.Stop();
       std::cout << GridLogMessage<<"HDCG::fPcg iteration "<<k<<" : inner rtzp "<<rtzp<<"\n";
 
       //    PcgM2(z,p[0]);
       PcgM2(z,mu); // ADEF-2 this is identity. Axpy possible to eliminate
-      
+
       p[peri_kp]=mu;
 
-      // Standard search direction  p -> z + b p    
+      // Standard search direction  p -> z + b p
       b = (rtzp)/rtz;
       
       int northog;
@@ -202,8 +232,25 @@ class TwoLevelCG : public LinearFunction<Field>
       if ( rn <= rsq ) { 
 
 	HDCGTimer.Stop();
-	std::cout<<GridLogMessage<<"HDCG: fPcg converged in "<<k<<" iterations and "<<HDCGTimer.Elapsed()<<std::endl;;
-	
+	std::cout<<GridLogMessage<<"HDCG: fPcg converged in "<<k<<" iterations and "<<HDCGTimer.Elapsed()<<std::endl;
+	std::cout<<GridLogMessage<<"HDCG: fPcg breakdown"<<std::endl;
+	auto mspc = [](GridStopWatch &sw, int64_t n) -> double {
+	  return (n > 0) ? sw.useconds() * 1e-3 / n : 0.0;
+	};
+	std::cout<<GridLogMessage<<"HDCG: fPcg M3 (fine MVM)  "<<M3Timer.Elapsed()
+		 <<"  "<<M3Calls<<" calls  "<<mspc(M3Timer,M3Calls)<<" ms/call"<<std::endl;
+	std::cout<<GridLogMessage<<"HDCG: fPcg linalg         "<<LinalgTimer.Elapsed()<<std::endl;
+	std::cout<<GridLogMessage<<"HDCG: fPcg smoother       "<<SmoothTimer.Elapsed()
+		 <<"  "<<SmoothCalls<<" calls  "<<mspc(SmoothTimer,SmoothCalls)<<" ms/call"<<std::endl;
+	std::cout<<GridLogMessage<<"HDCG: fPcg matrix (in M1) "<<MatrixTimer.Elapsed()
+		 <<"  "<<MatrixCalls<<" calls  "<<mspc(MatrixTimer,MatrixCalls)<<" ms/call"<<std::endl;
+	std::cout<<GridLogMessage<<"HDCG: fPcg project        "<<ProjectTimer.Elapsed()
+		 <<"  "<<ProjectCalls<<" calls  "<<mspc(ProjectTimer,ProjectCalls)<<" ms/call"<<std::endl;
+	std::cout<<GridLogMessage<<"HDCG: fPcg coarse         "<<CoarseTimer.Elapsed()
+		 <<"  "<<CoarseCalls<<" calls  "<<mspc(CoarseTimer,CoarseCalls)<<" ms/call"<<std::endl;
+	std::cout<<GridLogMessage<<"HDCG: fPcg promote        "<<PromoteTimer.Elapsed()
+		 <<"  "<<PromoteCalls<<" calls  "<<mspc(PromoteTimer,PromoteCalls)<<" ms/call"<<std::endl;
+
 	_FineLinop.HermOp(x,mmp[0]);			  
 	axpy(tmp,-1.0,src,mmp[0]);
 	
@@ -270,7 +317,7 @@ class TwoLevelCG : public LinearFunction<Field>
     std::vector<RealD> src_nrm(nrhs);
     for(int rhs=0;rhs<nrhs;rhs++) {
       src_nrm[rhs]=norm2(src[rhs]);
-      assert(src_nrm[rhs]!=0.0);
+      GRID_ASSERT(src_nrm[rhs]!=0.0);
     }
     std::vector<RealD> tn(nrhs);
 
@@ -475,35 +522,29 @@ class TwoLevelADEF2 : public TwoLevelCG<Field>
     CoarseField PleftProj(this->coarsegrid);
     CoarseField PleftMss_proj(this->coarsegrid);
 
-    GridStopWatch SmootherTimer;
-    GridStopWatch MatrixTimer;
-    SmootherTimer.Start();
+    this->SmoothTimer.Start();
     this->_Smoother(in,Min);
-    SmootherTimer.Stop();
+    this->SmoothTimer.Stop();
+    this->SmoothCalls++;
 
-    MatrixTimer.Start();
+    this->MatrixTimer.Start();
     this->_FineLinop.HermOp(Min,out);
-    MatrixTimer.Stop();
+    this->MatrixTimer.Stop();
+    this->MatrixCalls++;
     axpy(tmp,-1.0,out,in);          // tmp  = in - A Min
 
-    GridStopWatch ProjTimer;
-    GridStopWatch CoarseTimer;
-    GridStopWatch PromTimer;
-    ProjTimer.Start();
-    this->_Aggregates.ProjectToSubspace(PleftProj,tmp);     
-    ProjTimer.Stop();
-    CoarseTimer.Start();
+    this->ProjectTimer.Start();
+    this->_Aggregates.ProjectToSubspace(PleftProj,tmp);
+    this->ProjectTimer.Stop();
+    this->ProjectCalls++;
+    this->CoarseTimer.Start();
     this->_CoarseSolver(PleftProj,PleftMss_proj); // Ass^{-1} [in - A Min]_s
-    CoarseTimer.Stop();
-    PromTimer.Start();
-    this->_Aggregates.PromoteFromSubspace(PleftMss_proj,tmp);// tmp = Q[in - A Min]  
-    PromTimer.Stop();
-    std::cout << GridLogPerformance << "PcgM1 breakdown "<<std::endl;
-    std::cout << GridLogPerformance << "\tSmoother   " << SmootherTimer.Elapsed() <<std::endl;
-    std::cout << GridLogPerformance << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
-    std::cout << GridLogPerformance << "\tProj       " << ProjTimer.Elapsed() <<std::endl;
-    std::cout << GridLogPerformance << "\tCoarse     " << CoarseTimer.Elapsed() <<std::endl;
-    std::cout << GridLogPerformance << "\tProm       " << PromTimer.Elapsed() <<std::endl;
+    this->CoarseTimer.Stop();
+    this->CoarseCalls++;
+    this->PromoteTimer.Start();
+    this->_Aggregates.PromoteFromSubspace(PleftMss_proj,tmp);// tmp = Q[in - A Min]
+    this->PromoteTimer.Stop();
+    this->PromoteCalls++;
 
     axpy(out,1.0,Min,tmp); // Min+tmp
   }

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);
   }
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -161,7 +161,7 @@ class TwoLevelCGmrhs
     ////////////////////////////////////////////
     std::vector<RealD> ssq(nrhs);
     for(int rhs=0;rhs<nrhs;rhs++){
-      ssq[rhs]=norm2(src[rhs]); assert(ssq[rhs]!=0.0);
+      ssq[rhs]=norm2(src[rhs]); GRID_ASSERT(ssq[rhs]!=0.0);
     }      
 
     ///////////////////////////
@@ -382,7 +382,7 @@ class TwoLevelCGmrhs
     }
     HDCGTimer.Stop();
     std::cout<<GridLogMessage<<"HDCG: PrecBlockCGrQ not converged "<<HDCGTimer.Elapsed()<<std::endl;
-    assert(0);
+    GRID_ASSERT(0);
   }
 
   virtual void SolveSingleSystem (std::vector<Field> &src, std::vector<Field> &x)
@@ -415,7 +415,7 @@ class TwoLevelCGmrhs
     std::vector<RealD> src_nrm(nrhs);
     for(int rhs=0;rhs<nrhs;rhs++) {
       src_nrm[rhs]=norm2(src[rhs]);
-      assert(src_nrm[rhs]!=0.0);
+      GRID_ASSERT(src_nrm[rhs]!=0.0);
     }
     std::vector<RealD> tn(nrhs);
 

Grid/algorithms/iterative/BiCGSTAB.h

@@ -47,7 +47,7 @@ class BiCGSTAB : public OperatorFunction<Field>
   public:
     using OperatorFunction<Field>::operator();
     
-    bool ErrorOnNoConverge;  // throw an assert when the CG fails to converge.
+    bool ErrorOnNoConverge;  // throw an GRID_ASSERT when the CG fails to converge.
                              // Defaults true.
     RealD Tolerance;
     Integer MaxIterations;
@@ -77,7 +77,7 @@ class BiCGSTAB : public OperatorFunction<Field>
 
       // Initial residual computation & set up
       RealD guess = norm2(psi);
-      assert(std::isnan(guess) == 0);
+      GRID_ASSERT(std::isnan(guess) == 0);
     
       Linop.Op(psi, v);
       b = norm2(v);
@@ -214,7 +214,7 @@ class BiCGSTAB : public OperatorFunction<Field>
           std::cout << GridLogMessage << "\tAxpyNorm   " << AxpyNormTimer.Elapsed() << std::endl;
           std::cout << GridLogMessage << "\tLinearComb " << LinearCombTimer.Elapsed() << std::endl;
 
-          if(ErrorOnNoConverge){ assert(true_residual / Tolerance < 10000.0); }
+          if(ErrorOnNoConverge){ GRID_ASSERT(true_residual / Tolerance < 10000.0); }
 
           IterationsToComplete = k;	
 
@@ -224,7 +224,7 @@ class BiCGSTAB : public OperatorFunction<Field>
       
       std::cout << GridLogMessage << "BiCGSTAB did NOT converge" << std::endl;
 
-      if(ErrorOnNoConverge){ assert(0); }
+      if(ErrorOnNoConverge){ GRID_ASSERT(0); }
       IterationsToComplete = k;
     }
 };

Grid/algorithms/iterative/BlockConjugateGradient.h

@@ -98,7 +98,7 @@ class BlockConjugateGradient : public OperatorFunction<Field> {
   int Nblock;
 
   BlockCGtype CGtype;
-  bool ErrorOnNoConverge;  // throw an assert when the CG fails to converge.
+  bool ErrorOnNoConverge;  // throw an GRID_ASSERT when the CG fails to converge.
                            // Defaults true.
   RealD Tolerance;
   Integer MaxIterations;
@@ -201,7 +201,7 @@ void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
   } else if (CGtype == CGmultiRHS ) {
     CGmultiRHSsolve(Linop,Src,Psi);
   } else {
-    assert(0);
+    GRID_ASSERT(0);
   }
 }
 virtual void operator()(LinearOperatorBase<Field> &Linop, const std::vector<Field> &Src, std::vector<Field> &Psi) 
@@ -209,7 +209,7 @@ virtual void operator()(LinearOperatorBase<Field> &Linop, const std::vector<Fiel
   if ( CGtype == BlockCGrQVec ) {
     BlockCGrQsolveVec(Linop,Src,Psi);
   } else {
-    assert(0);
+    GRID_ASSERT(0);
   }
 }
 
@@ -259,10 +259,10 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
   for(int b=0;b<Nblock;b++) std::cout << "src["<<b<<"]" << ssq[b] <<std::endl;
 
   sliceNorm(residuals,B,Orthog);
-  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
+  for(int b=0;b<Nblock;b++){ GRID_ASSERT(std::isnan(residuals[b])==0); }
 
   sliceNorm(residuals,X,Orthog);
-  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
+  for(int b=0;b<Nblock;b++){ GRID_ASSERT(std::isnan(residuals[b])==0); }
 
   /************************************************************************
    * Block conjugate gradient rQ (Sebastien Birk Thesis, after Dubrulle 2001)
@@ -402,7 +402,7 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
   std::cout << GridLogMessage << "BlockConjugateGradient(rQ) did NOT converge "<<k<<" / "<<MaxIterations
 	    <<" residual "<< std::sqrt(max_resid)<< std::endl;
 
-  if (ErrorOnNoConverge) assert(0);
+  if (ErrorOnNoConverge) GRID_ASSERT(0);
   IterationsToComplete = k;
 }
 //////////////////////////////////////////////////////////////////////////
@@ -438,10 +438,10 @@ void CGmultiRHSsolve(LinearOperatorBase<Field> &Linop, const Field &Src, Field &
   for(int b=0;b<Nblock;b++) sssum+=ssq[b];
 
   sliceNorm(residuals,Src,Orthog);
-  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
+  for(int b=0;b<Nblock;b++){ GRID_ASSERT(std::isnan(residuals[b])==0); }
 
   sliceNorm(residuals,Psi,Orthog);
-  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
+  for(int b=0;b<Nblock;b++){ GRID_ASSERT(std::isnan(residuals[b])==0); }
 
   // Initial search dir is guess
   Linop.HermOp(Psi, AP);
@@ -540,7 +540,7 @@ void CGmultiRHSsolve(LinearOperatorBase<Field> &Linop, const Field &Src, Field &
   }
   std::cout << GridLogMessage << "MultiRHSConjugateGradient did NOT converge" << std::endl;
 
-  if (ErrorOnNoConverge) assert(0);
+  if (ErrorOnNoConverge) GRID_ASSERT(0);
   IterationsToComplete = k;
 }
 
@@ -554,7 +554,7 @@ void CGmultiRHSsolve(LinearOperatorBase<Field> &Linop, const Field &Src, Field &
 void BlockCGrQsolveVec(LinearOperatorBase<Field> &Linop, const std::vector<Field> &B, std::vector<Field> &X) 
 {
   Nblock = B.size();
-  assert(Nblock == X.size());
+  GRID_ASSERT(Nblock == X.size());
 
   std::cout<<GridLogMessage<<" Block Conjugate Gradient Vec rQ : Nblock "<<Nblock<<std::endl;
 
@@ -594,10 +594,10 @@ void BlockCGrQsolveVec(LinearOperatorBase<Field> &Linop, const std::vector<Field
   for(int b=0;b<Nblock;b++) sssum+=ssq[b];
 
   for(int b=0;b<Nblock;b++){ residuals[b] = norm2(B[b]);}
-  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
+  for(int b=0;b<Nblock;b++){ GRID_ASSERT(std::isnan(residuals[b])==0); }
 
   for(int b=0;b<Nblock;b++){ residuals[b] = norm2(X[b]);}
-  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
+  for(int b=0;b<Nblock;b++){ GRID_ASSERT(std::isnan(residuals[b])==0); }
 
   /************************************************************************
    * Block conjugate gradient rQ (Sebastien Birk Thesis, after Dubrulle 2001)
@@ -731,7 +731,7 @@ void BlockCGrQsolveVec(LinearOperatorBase<Field> &Linop, const std::vector<Field
   }
   std::cout << GridLogMessage << "BlockConjugateGradient(rQ) did NOT converge" << std::endl;
 
-  if (ErrorOnNoConverge) assert(0);
+  if (ErrorOnNoConverge) GRID_ASSERT(0);
   IterationsToComplete = k;
 }
 

Grid/algorithms/iterative/CommunicationAvoidingGeneralisedMinimalResidual.h

@@ -36,7 +36,7 @@ class CommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<
  public:
   using OperatorFunction<Field>::operator();
 
-  bool ErrorOnNoConverge; // Throw an assert when CAGMRES fails to converge,
+  bool ErrorOnNoConverge; // Throw an GRID_ASSERT when CAGMRES fails to converge,
                           // defaults to true
 
   RealD   Tolerance;
@@ -82,7 +82,7 @@ class CommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<
     conformable(psi, src);
 
     RealD guess = norm2(psi);
-    assert(std::isnan(guess) == 0);
+    GRID_ASSERT(std::isnan(guess) == 0);
 
     RealD cp;
     RealD ssq = norm2(src);
@@ -137,7 +137,7 @@ class CommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<
     std::cout << GridLogMessage << "CommunicationAvoidingGeneralisedMinimalResidual did NOT converge" << std::endl;
 
     if (ErrorOnNoConverge)
-      assert(0);
+      GRID_ASSERT(0);
   }
 
   RealD outerLoopBody(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi, RealD rsq) {
@@ -185,7 +185,7 @@ class CommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<
       }
     }
 
-    assert(0); // Never reached
+    GRID_ASSERT(0); // Never reached
     return cp;
   }
 

Grid/algorithms/iterative/ConjugateGradient.h

@@ -45,7 +45,7 @@ public:
 
   using OperatorFunction<Field>::operator();
   
-  bool ErrorOnNoConverge;  // throw an assert when the CG fails to converge.
+  bool ErrorOnNoConverge;  // throw an GRID_ASSERT when the CG fails to converge.
                            // Defaults true.
   RealD Tolerance;
   Integer MaxIterations;
@@ -94,7 +94,7 @@ public:
     ssq = norm2(src);
     RealD guess = norm2(psi);
     NormTimer.Stop();
-    assert(std::isnan(guess) == 0);
+    GRID_ASSERT(std::isnan(guess) == 0);
     AssignTimer.Start();
     if ( guess == 0.0 ) {
       r = src;
@@ -212,17 +212,21 @@ public:
 		  << "\tTarget " << Tolerance << std::endl;
 
 	//	std::cout << GridLogMessage << "\tPreamble   " << PreambleTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tSolver Elapsed    " << SolverTimer.Elapsed() <<std::endl;
-        std::cout << GridLogPerformance << "Time breakdown "<<std::endl;
-	std::cout << GridLogPerformance << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
-	std::cout << GridLogPerformance << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
-	std::cout << GridLogPerformance << "\t\tInner      " << InnerTimer.Elapsed() <<std::endl;
-	std::cout << GridLogPerformance << "\t\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
-	std::cout << GridLogPerformance << "\t\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tPreamble   " << PreambleTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tConstruct  " << ConstructTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tNorm       " << NormTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tAssign     " << AssignTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tSolver     " << SolverTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "Solver breakdown "<<std::endl;
+	std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\t\tInner      " << InnerTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\t\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\t\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
 
 	std::cout << GridLogDebug << "\tMobius flop rate " << DwfFlops/ usecs<< " Gflops " <<std::endl;
 
-        if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
+        if (ErrorOnNoConverge) GRID_ASSERT(true_residual / Tolerance < 10000.0);
 
 	IterationsToComplete = k;	
 	TrueResidual = true_residual;
@@ -251,7 +255,7 @@ public:
     std::cout << GridLogPerformance << "\t\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
     std::cout << GridLogPerformance << "\t\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
 
-    if (ErrorOnNoConverge) assert(0);
+    if (ErrorOnNoConverge) GRID_ASSERT(0);
     IterationsToComplete = k;
 
   }

Grid/algorithms/iterative/ConjugateGradientMixedPrec.h

@@ -116,14 +116,14 @@ NAMESPACE_BEGIN(Grid);
       //Compute double precision rsd and also new RHS vector.
       Linop_d.HermOp(sol_d, tmp_d);
       RealD norm = axpy_norm(src_d, -1., tmp_d, src_d_in); //src_d is residual vector
-      
+      std::cout<<GridLogMessage<<" rsd norm "<<norm<<std::endl;
       std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Outer iteration " <<outer_iter<<" residual "<< norm<< " target "<< stop<<std::endl;
 
       if(norm < OuterLoopNormMult * stop){
 	std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Outer iteration converged on iteration " <<outer_iter <<std::endl;
 	break;
       }
-      while(norm * inner_tol * inner_tol < stop) inner_tol *= 2;  // inner_tol = sqrt(stop/norm) ??
+      while(norm * inner_tol * inner_tol < stop*1.01) inner_tol *= 2;  // inner_tol = sqrt(stop/norm) ??
 
       PrecChangeTimer.Start();
       precisionChange(src_f, src_d, pc_wk_dp_to_sp);

Grid/algorithms/iterative/ConjugateGradientMixedPrecBatched.h

@@ -77,7 +77,7 @@ public:
   }
 
   void operator() (const std::vector<FieldD> &src_d_in, std::vector<FieldD> &sol_d){
-    assert(src_d_in.size() == sol_d.size());
+    GRID_ASSERT(src_d_in.size() == sol_d.size());
     int NBatch = src_d_in.size();
 
     std::cout << GridLogMessage << "NBatch = " << NBatch << std::endl;

Grid/algorithms/iterative/ConjugateGradientMultiShift.h

@@ -98,15 +98,15 @@ public:
     std::vector<RealD> alpha(nshift,1.0);
     std::vector<Field>   ps(nshift,grid);// Search directions
 
-    assert(psi.size()==nshift);
-    assert(mass.size()==nshift);
-    assert(mresidual.size()==nshift);
+    GRID_ASSERT(psi.size()==nshift);
+    GRID_ASSERT(mass.size()==nshift);
+    GRID_ASSERT(mresidual.size()==nshift);
   
-    // dynamic sized arrays on stack; 2d is a pain with vector
-    RealD  bs[nshift];
-    RealD  rsq[nshift];
-    RealD  z[nshift][2];
-    int     converged[nshift];
+    // remove dynamic sized arrays on stack; 2d is a pain with vector
+    std::vector<RealD>  bs(nshift);
+    std::vector<RealD>  rsq(nshift);
+    std::vector<std::array<RealD,2> >  z(nshift);
+    std::vector<int>     converged(nshift);
   
     const int       primary =0;
   
@@ -122,7 +122,7 @@ public:
   
     // Check lightest mass
     for(int s=0;s<nshift;s++){
-      assert( mass[s]>= mass[primary] );
+      GRID_ASSERT( mass[s]>= mass[primary] );
       converged[s]=0;
     }
   
@@ -338,7 +338,7 @@ public:
     }
     // ugly hack
     std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;
-    //  assert(0);
+    //  GRID_ASSERT(0);
   }
 
 };

Grid/algorithms/iterative/ConjugateGradientMultiShiftCleanup.h

@@ -118,16 +118,16 @@ public:
     FieldF r_f(SinglePrecGrid);
     FieldD mmp_d(DoublePrecGrid);
 
-    assert(psi_d.size()==nshift);
-    assert(mass.size()==nshift);
-    assert(mresidual.size()==nshift);
+    GRID_ASSERT(psi_d.size()==nshift);
+    GRID_ASSERT(mass.size()==nshift);
+    GRID_ASSERT(mresidual.size()==nshift);
   
     // dynamic sized arrays on stack; 2d is a pain with vector
-    RealD  bs[nshift];
-    RealD  rsq[nshift];
-    RealD  rsqf[nshift];
-    RealD  z[nshift][2];
-    int     converged[nshift];
+    std::vector<RealD>  bs(nshift);
+    std::vector<RealD>  rsq(nshift);
+    std::vector<RealD>  rsqf(nshift);
+    std::vector<std::array<RealD,2> >  z(nshift);
+    std::vector<int>     converged(nshift);
   
     const int       primary =0;
   
@@ -141,7 +141,7 @@ public:
 
     // Check lightest mass
     for(int s=0;s<nshift;s++){
-      assert( mass[s]>= mass[primary] );
+      GRID_ASSERT( mass[s]>= mass[primary] );
       converged[s]=0;
     }
   
@@ -179,7 +179,7 @@ public:
     Linop_d.HermOpAndNorm(p_d,mmp_d,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
     tmp_d = tmp_d - mmp_d;
     std::cout << " Testing operators match "<<norm2(mmp_d)<<" f "<<norm2(mmp_f)<<" diff "<< norm2(tmp_d)<<std::endl;
-    //    assert(norm2(tmp_d)< 1.0e-4);
+    //    GRID_ASSERT(norm2(tmp_d)< 1.0e-4);
 
     axpy(mmp_d,mass[0],p_d,mmp_d);
     RealD rn = norm2(p_d);
@@ -365,7 +365,7 @@ public:
    
     }
     std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;
-    assert(0);
+    GRID_ASSERT(0);
   }
 
 };

Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h

@@ -48,12 +48,12 @@ public:
 
   ShiftedLinop(LinearOperatorBase<Field> &_linop_base, RealD _shift): linop_base(_linop_base), shift(_shift){}
 
-  void OpDiag (const Field &in, Field &out){ assert(0); }
-  void OpDir  (const Field &in, Field &out,int dir,int disp){ assert(0); }
-  void OpDirAll  (const Field &in, std::vector<Field> &out){ assert(0); }
+  void OpDiag (const Field &in, Field &out){ GRID_ASSERT(0); }
+  void OpDir  (const Field &in, Field &out,int dir,int disp){ GRID_ASSERT(0); }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){ GRID_ASSERT(0); }
   
-  void Op     (const Field &in, Field &out){ assert(0); }
-  void AdjOp  (const Field &in, Field &out){ assert(0); }
+  void Op     (const Field &in, Field &out){ GRID_ASSERT(0); }
+  void AdjOp  (const Field &in, Field &out){ GRID_ASSERT(0); }
 
   void HermOp(const Field &in, Field &out){
     linop_base.HermOp(in, out);
@@ -151,16 +151,16 @@ public:
     FieldD r_d(DoublePrecGrid);
     FieldD mmp_d(DoublePrecGrid);
 
-    assert(psi_d.size()==nshift);
-    assert(mass.size()==nshift);
-    assert(mresidual.size()==nshift);
+    GRID_ASSERT(psi_d.size()==nshift);
+    GRID_ASSERT(mass.size()==nshift);
+    GRID_ASSERT(mresidual.size()==nshift);
   
     // dynamic sized arrays on stack; 2d is a pain with vector
-    RealD  bs[nshift];
-    RealD  rsq[nshift];
-    RealD  rsqf[nshift];
-    RealD  z[nshift][2];
-    int     converged[nshift];
+    std::vector<RealD>  bs(nshift);
+    std::vector<RealD>  rsq(nshift);
+    std::vector<RealD>  rsqf(nshift);
+    std::vector<std::array<RealD,2> >  z(nshift);
+    std::vector<int>     converged(nshift);
   
     const int       primary =0;
   
@@ -174,7 +174,7 @@ public:
 
     // Check lightest mass
     for(int s=0;s<nshift;s++){
-      assert( mass[s]>= mass[primary] );
+      GRID_ASSERT( mass[s]>= mass[primary] );
       converged[s]=0;
     }
   
@@ -211,7 +211,7 @@ public:
     Linop_d.HermOpAndNorm(p_d,mmp_d,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
     tmp_d = tmp_d - mmp_d;
     std::cout << " Testing operators match "<<norm2(mmp_d)<<" f "<<norm2(mmp_f)<<" diff "<< norm2(tmp_d)<<std::endl;
-    assert(norm2(tmp_d)< 1.0);
+    GRID_ASSERT(norm2(tmp_d)< 1.0);
 
     axpy(mmp_d,mass[0],p_d,mmp_d);
     RealD rn = norm2(p_d);
@@ -408,7 +408,7 @@ public:
    
     }
     std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;
-    assert(0);
+    GRID_ASSERT(0);
   }
 
 };

Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h

@@ -35,7 +35,7 @@ template<class FieldD,class FieldF,
 	 typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
 class ConjugateGradientReliableUpdate : public LinearFunction<FieldD> {
 public:
-  bool ErrorOnNoConverge;  // throw an assert when the CG fails to converge.
+  bool ErrorOnNoConverge;  // throw an GRID_ASSERT when the CG fails to converge.
   // Defaults true.
   RealD Tolerance;
   Integer MaxIterations;
@@ -66,7 +66,7 @@ public:
       DoFinalCleanup(true),
       Linop_fallback(NULL)
   {
-    assert(Delta > 0. && Delta < 1. && "Expect  0 < Delta < 1");
+    GRID_ASSERT(Delta > 0. && Delta < 1. && "Expect  0 < Delta < 1");
   };
 
   void setFallbackLinop(LinearOperatorBase<FieldF> &_Linop_fallback, const RealD _fallback_transition_tol){
@@ -90,7 +90,7 @@ public:
 
     // Initial residual computation & set up
     RealD guess = norm2(psi);
-    assert(std::isnan(guess) == 0);
+    GRID_ASSERT(std::isnan(guess) == 0);
     
     Linop_d.HermOpAndNorm(psi, mmp, d, b);
     
@@ -217,7 +217,7 @@ public:
 	  CG(Linop_d,src,psi);
 	  IterationsToCleanup = CG.IterationsToComplete;
 	}
-	else if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
+	else if (ErrorOnNoConverge) GRID_ASSERT(true_residual / Tolerance < 10000.0);
 
 	std::cout << GridLogMessage << "ConjugateGradientReliableUpdate complete.\n";
 	return;
@@ -263,7 +263,7 @@ public:
     std::cout << GridLogMessage << "ConjugateGradientReliableUpdate did NOT converge"
 	      << std::endl;
       
-    if (ErrorOnNoConverge) assert(0);
+    if (ErrorOnNoConverge) GRID_ASSERT(0);
     IterationsToComplete = k;
     ReliableUpdatesPerformed = l;      
   }    

Grid/algorithms/iterative/ConjugateResidual.h

@@ -106,7 +106,7 @@ public:
     }
 
     std::cout<<GridLogMessage<<"ConjugateResidual did NOT converge"<<std::endl;
-    assert(0);
+    GRID_ASSERT(0);
   }
 };
 NAMESPACE_END(Grid);

Grid/algorithms/iterative/FlexibleCommunicationAvoidingGeneralisedMinimalResidual.h

@@ -36,7 +36,7 @@ class FlexibleCommunicationAvoidingGeneralisedMinimalResidual : public OperatorF
  public:
   using OperatorFunction<Field>::operator();
 
-  bool ErrorOnNoConverge; // Throw an assert when FCAGMRES fails to converge,
+  bool ErrorOnNoConverge; // Throw an GRID_ASSERT when FCAGMRES fails to converge,
                           // defaults to true
 
   RealD   Tolerance;
@@ -87,7 +87,7 @@ class FlexibleCommunicationAvoidingGeneralisedMinimalResidual : public OperatorF
     conformable(psi, src);
 
     RealD guess = norm2(psi);
-    assert(std::isnan(guess) == 0);
+    GRID_ASSERT(std::isnan(guess) == 0);
 
     RealD cp;
     RealD ssq = norm2(src);
@@ -144,7 +144,7 @@ class FlexibleCommunicationAvoidingGeneralisedMinimalResidual : public OperatorF
     std::cout << GridLogMessage << "FlexibleCommunicationAvoidingGeneralisedMinimalResidual did NOT converge" << std::endl;
 
     if (ErrorOnNoConverge)
-      assert(0);
+      GRID_ASSERT(0);
   }
 
   RealD outerLoopBody(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi, RealD rsq) {
@@ -191,7 +191,7 @@ class FlexibleCommunicationAvoidingGeneralisedMinimalResidual : public OperatorF
       }
     }
 
-    assert(0); // Never reached
+    GRID_ASSERT(0); // Never reached
     return cp;
   }
 

Grid/algorithms/iterative/FlexibleGeneralisedMinimalResidual.h

@@ -36,7 +36,7 @@ class FlexibleGeneralisedMinimalResidual : public OperatorFunction<Field> {
  public:
   using OperatorFunction<Field>::operator();
 
-  bool ErrorOnNoConverge; // Throw an assert when FGMRES fails to converge,
+  bool ErrorOnNoConverge; // Throw an GRID_ASSERT when FGMRES fails to converge,
                           // defaults to true
 
   RealD   Tolerance;
@@ -85,7 +85,7 @@ class FlexibleGeneralisedMinimalResidual : public OperatorFunction<Field> {
     conformable(psi, src);
 
     RealD guess = norm2(psi);
-    assert(std::isnan(guess) == 0);
+    GRID_ASSERT(std::isnan(guess) == 0);
 
     RealD cp;
     RealD ssq = norm2(src);
@@ -142,7 +142,7 @@ class FlexibleGeneralisedMinimalResidual : public OperatorFunction<Field> {
     std::cout << GridLogMessage << "FlexibleGeneralisedMinimalResidual did NOT converge" << std::endl;
 
     if (ErrorOnNoConverge)
-      assert(0);
+      GRID_ASSERT(0);
   }
 
   RealD outerLoopBody(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi, RealD rsq) {
@@ -189,7 +189,7 @@ class FlexibleGeneralisedMinimalResidual : public OperatorFunction<Field> {
       }
     }
 
-    assert(0); // Never reached
+    GRID_ASSERT(0); // Never reached
     return cp;
   }
 

Grid/algorithms/iterative/GeneralisedMinimalResidual.h

@@ -36,7 +36,7 @@ class GeneralisedMinimalResidual : public OperatorFunction<Field> {
  public:
   using OperatorFunction<Field>::operator();
 
-  bool ErrorOnNoConverge; // Throw an assert when GMRES fails to converge,
+  bool ErrorOnNoConverge; // Throw an GRID_ASSERT when GMRES fails to converge,
                           // defaults to true
 
   RealD   Tolerance;
@@ -80,7 +80,7 @@ class GeneralisedMinimalResidual : public OperatorFunction<Field> {
     conformable(psi, src);
 
     RealD guess = norm2(psi);
-    assert(std::isnan(guess) == 0);
+    GRID_ASSERT(std::isnan(guess) == 0);
 
     RealD cp;
     RealD ssq = norm2(src);
@@ -135,7 +135,7 @@ class GeneralisedMinimalResidual : public OperatorFunction<Field> {
     std::cout << GridLogMessage << "GeneralisedMinimalResidual did NOT converge" << std::endl;
 
     if (ErrorOnNoConverge)
-      assert(0);
+      GRID_ASSERT(0);
   }
 
   RealD outerLoopBody(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi, RealD rsq) {
@@ -181,7 +181,7 @@ class GeneralisedMinimalResidual : public OperatorFunction<Field> {
       }
     }
 
-    assert(0); // Never reached
+    GRID_ASSERT(0); // Never reached
     return cp;
   }
 

Grid/algorithms/iterative/ImplicitlyRestartedBlockLanczos.h

@@ -175,7 +175,7 @@ public:
       eresid(_eresid),  MaxIter(_MaxIter),
       diagonalisation(_diagonalisation),split_test(0),
       Nevec_acc(_Nu)
-  { assert( (Nk%Nu==0) && (Nm%Nu==0) ); };
+  { GRID_ASSERT( (Nk%Nu==0) && (Nm%Nu==0) ); };
 
   ////////////////////////////////
   // Helpers
@@ -206,7 +206,7 @@ public:
           Glog<<"orthogonalize after: "<<j<<" of "<<k<<" "<< ip <<std::endl;
       }
     }
-    assert(normalize(w,if_print) != 0);
+    GRID_ASSERT(normalize(w,if_print) != 0);
   }
   void reorthogonalize(Field& w, std::vector<Field>& evec, int k)
   {
@@ -225,7 +225,7 @@ public:
       w[i] = w[i] - ip * evec[j];
     }}
     for(int i=0; i<_Nu; ++i)
-    assert(normalize(w[i],if_print) !=0);
+    GRID_ASSERT(normalize(w[i],if_print) !=0);
   }
 
 
@@ -244,7 +244,7 @@ public:
     const uint64_t sites = grid->lSites();
 
     int Nbatch = R/Nevec_acc;
-    assert( R%Nevec_acc == 0 );
+    GRID_ASSERT( R%Nevec_acc == 0 );
 //    Glog << "nBatch, Nevec_acc, R, Nu = " 
 //         << Nbatch << "," << Nevec_acc << "," << R << "," << Nu << std::endl;
     
@@ -302,7 +302,7 @@ public:
       }
     }
     for (int i=0; i<Nu; ++i) {
-      assert(normalize(w[i],do_print)!=0);
+      GRID_ASSERT(normalize(w[i],do_print)!=0);
     }
     
     Glog << "cuBLAS Zgemm done"<< std::endl;
@@ -374,8 +374,8 @@ cudaStat = cudaMallocManaged((void **)&evec_acc, Nevec_acc*sites*12*sizeof(CUDA_
   {
     std::string fname = std::string(cname+"::calc_irbl()"); 
     GridBase *grid = evec[0].Grid();
-    assert(grid == src[0].Grid());
-    assert( Nu = src.size() );
+    GRID_ASSERT(grid == src[0].Grid());
+    GRID_ASSERT( Nu = src.size() );
     
     Glog << std::string(74,'*') << std::endl;
     Glog << fname + " starting iteration 0 /  "<< MaxIter<< std::endl;
@@ -396,7 +396,7 @@ cudaStat = cudaMallocManaged((void **)&evec_acc, Nevec_acc*sites*12*sizeof(CUDA_
     }
     Glog << std::string(74,'*') << std::endl;
     
-    assert(Nm == evec.size() && Nm == eval.size());
+    GRID_ASSERT(Nm == evec.size() && Nm == eval.size());
 
     std::vector<std::vector<ComplexD>> lmd(Nu,std::vector<ComplexD>(Nm,0.0));  
     std::vector<std::vector<ComplexD>> lme(Nu,std::vector<ComplexD>(Nm,0.0));  
@@ -579,8 +579,8 @@ cudaStat = cudaMallocManaged((void **)&evec_acc, Nevec_acc*sites*12*sizeof(CUDA_
   {
     std::string fname = std::string(cname+"::calc_rbl()"); 
     GridBase *grid = evec[0].Grid();
-    assert(grid == src[0].Grid());
-    assert( Nu = src.size() );
+    GRID_ASSERT(grid == src[0].Grid());
+    GRID_ASSERT( Nu = src.size() );
 
     int Np = (Nm-Nk);
     if (Np > 0 && MaxIter > 1) Np /= MaxIter;
@@ -607,7 +607,7 @@ cudaStat = cudaMallocManaged((void **)&evec_acc, Nevec_acc*sites*12*sizeof(CUDA_
     }
     Glog << std::string(74,'*') << std::endl;
     
-    assert(Nm == evec.size() && Nm == eval.size());
+    GRID_ASSERT(Nm == evec.size() && Nm == eval.size());
 	
     std::vector<std::vector<ComplexD>> lmd(Nu,std::vector<ComplexD>(Nm,0.0));  
     std::vector<std::vector<ComplexD>> lme(Nu,std::vector<ComplexD>(Nm,0.0));  
@@ -785,7 +785,7 @@ private:
     
     int Nu = w.size();
     int Nm = evec.size();
-    assert( b < Nm/Nu );
+    GRID_ASSERT( b < Nm/Nu );
 //    GridCartesian *grid = evec[0]._grid;
     
     // converts block index to full indicies for an interval [L,R)
@@ -796,7 +796,7 @@ private:
 
     Glog << "Using split grid"<< std::endl;
 //   LatticeGaugeField s_Umu(SGrid);
-   assert((Nu%mrhs)==0);
+   GRID_ASSERT((Nu%mrhs)==0);
    std::vector<Field>   in(mrhs,f_grid);
      
     Field s_in(sf_grid);
@@ -906,7 +906,7 @@ if(split_test){
     
     for (int u=0; u<Nu; ++u) {
 //      Glog << "norm2(w[" << u << "])= "<< norm2(w[u]) << std::endl;
-      assert (!isnan(norm2(w[u])));
+      GRID_ASSERT (!isnan(norm2(w[u])));
       for (int k=L+u; k<R; ++k) {
         Glog <<" In block "<< b << "," <<" beta[" << u << "," << k-L << "] = " << lme[u][k] << std::endl;
       }
@@ -929,8 +929,8 @@ if(split_test){
 			 Eigen::MatrixXcd & Qt, // Nm x Nm
 			 GridBase *grid)
   {
-    assert( Nk%Nu == 0 && Nm%Nu == 0 );
-    assert( Nk <= Nm );
+    GRID_ASSERT( Nk%Nu == 0 && Nm%Nu == 0 );
+    GRID_ASSERT( Nk <= Nm );
     Eigen::MatrixXcd BlockTriDiag = Eigen::MatrixXcd::Zero(Nk,Nk);
     
     for ( int u=0; u<Nu; ++u ) {
@@ -970,8 +970,8 @@ if(split_test){
 			 GridBase *grid)
   {
     Glog << "diagonalize_lapack: Nu= "<<Nu<<" Nk= "<<Nk<<" Nm= "<<std::endl;
-    assert( Nk%Nu == 0 && Nm%Nu == 0 );
-    assert( Nk <= Nm );
+    GRID_ASSERT( Nk%Nu == 0 && Nm%Nu == 0 );
+    GRID_ASSERT( Nk <= Nm );
     Eigen::MatrixXcd BlockTriDiag = Eigen::MatrixXcd::Zero(Nk,Nk);
     
     for ( int u=0; u<Nu; ++u ) {
@@ -1119,7 +1119,7 @@ if (1){
       diagonalize_lapack(eval,lmd,lme,Nu,Nk,Nm,Qt,grid);
 #endif
     } else { 
-      assert(0);
+      GRID_ASSERT(0);
     }
   }
   
@@ -1131,8 +1131,8 @@ if (1){
          Eigen::MatrixXcd& M)
   {
     //Glog << "unpackHermitBlockTriDiagMatToEigen() begin" << '\n'; 
-    assert( Nk%Nu == 0 && Nm%Nu == 0 );
-    assert( Nk <= Nm );
+    GRID_ASSERT( Nk%Nu == 0 && Nm%Nu == 0 );
+    GRID_ASSERT( Nk <= Nm );
     M = Eigen::MatrixXcd::Zero(Nk,Nk);
     
     // rearrange 
@@ -1159,8 +1159,8 @@ if (1){
          Eigen::MatrixXcd& M)
   {
     //Glog << "packHermitBlockTriDiagMatfromEigen() begin" << '\n'; 
-    assert( Nk%Nu == 0 && Nm%Nu == 0 );
-    assert( Nk <= Nm );
+    GRID_ASSERT( Nk%Nu == 0 && Nm%Nu == 0 );
+    GRID_ASSERT( Nk <= Nm );
     
     // rearrange 
     for ( int u=0; u<Nu; ++u ) {

Grid/algorithms/iterative/ImplicitlyRestartedBlockLanczosCoarse.h

@@ -121,7 +121,7 @@ public:
       eresid(_eresid),  MaxIter(_MaxIter),
       diagonalisation(_diagonalisation),
       Nevec_acc(_Nu)
-  { assert( (Nk%Nu==0) && (Nm%Nu==0) ); };
+  { GRID_ASSERT( (Nk%Nu==0) && (Nm%Nu==0) ); };
 
   ////////////////////////////////
   // Helpers
@@ -151,7 +151,7 @@ public:
           Glog<<"orthogonalize after: "<<j<<" of "<<k<<" "<< ip <<std::endl;
       }
     }
-    assert(normalize(w,if_print) != 0);
+    GRID_ASSERT(normalize(w,if_print) != 0);
   }
   void reorthogonalize(Field& w, std::vector<Field>& evec, int k)
   {
@@ -169,7 +169,7 @@ public:
       w[i] = w[i] - ip * evec[j];
     }}
     for(int i=0; i<_Nu; ++i)
-    assert(normalize(w[i],if_print) !=0);
+    GRID_ASSERT(normalize(w[i],if_print) !=0);
   }
   
   void orthogonalize_blockhead(Field& w, std::vector<Field>& evec, int k, int Nu)
@@ -205,8 +205,8 @@ public:
   {
     std::string fname = std::string(cname+"::calc_irbl()"); 
     GridBase *grid = evec[0].Grid();
-    assert(grid == src[0].Grid());
-    assert( Nu = src.size() );
+    GRID_ASSERT(grid == src[0].Grid());
+    GRID_ASSERT( Nu = src.size() );
     
     Glog << std::string(74,'*') << std::endl;
     Glog << fname + " starting iteration 0 /  "<< MaxIter<< std::endl;
@@ -227,7 +227,7 @@ public:
     }
     Glog << std::string(74,'*') << std::endl;
     
-    assert(Nm == evec.size() && Nm == eval.size());
+    GRID_ASSERT(Nm == evec.size() && Nm == eval.size());
 
     std::vector<std::vector<ComplexD>> lmd(Nu,std::vector<ComplexD>(Nm,0.0));  
     std::vector<std::vector<ComplexD>> lme(Nu,std::vector<ComplexD>(Nm,0.0));  
@@ -413,8 +413,8 @@ public:
   {
     std::string fname = std::string(cname+"::calc_rbl()"); 
     GridBase *grid = evec[0].Grid();
-    assert(grid == src[0].Grid());
-    assert( Nu = src.size() );
+    GRID_ASSERT(grid == src[0].Grid());
+    GRID_ASSERT( Nu = src.size() );
 
     int Np = (Nm-Nk);
     if (Np > 0 && MaxIter > 1) Np /= MaxIter;
@@ -441,7 +441,7 @@ public:
     }
     Glog << std::string(74,'*') << std::endl;
     
-    assert(Nm == evec.size() && Nm == eval.size());
+    GRID_ASSERT(Nm == evec.size() && Nm == eval.size());
 	
     std::vector<std::vector<ComplexD>> lmd(Nu,std::vector<ComplexD>(Nm,0.0));  
     std::vector<std::vector<ComplexD>> lme(Nu,std::vector<ComplexD>(Nm,0.0));  
@@ -622,7 +622,7 @@ private:
     
     int Nu = w.size();
     int Nm = evec.size();
-    assert( b < Nm/Nu );
+    GRID_ASSERT( b < Nm/Nu );
     
     // converts block index to full indicies for an interval [L,R)
     int L = Nu*b;
@@ -630,7 +630,7 @@ private:
 
     Real beta;
 
-    assert((Nu%mrhs)==0);
+    GRID_ASSERT((Nu%mrhs)==0);
     std::vector<Field>   in(mrhs,f_grid);
     std::vector<Field>   out(mrhs,f_grid);
 
@@ -711,7 +711,7 @@ private:
     
     for (int u=0; u<Nu; ++u) {
       //      Glog << "norm2(w[" << u << "])= "<< norm2(w[u]) << std::endl;
-      assert (!isnan(norm2(w[u])));
+      GRID_ASSERT (!isnan(norm2(w[u])));
       for (int k=L+u; k<R; ++k) {
 	//        Glog <<" In block "<< b << "," <<" beta[" << u << "," << k-L << "] = " << lme[u][k] << std::endl;
       }
@@ -734,8 +734,8 @@ private:
 			 Eigen::MatrixXcd & Qt, // Nm x Nm
 			 GridBase *grid)
   {
-    assert( Nk%Nu == 0 && Nm%Nu == 0 );
-    assert( Nk <= Nm );
+    GRID_ASSERT( Nk%Nu == 0 && Nm%Nu == 0 );
+    GRID_ASSERT( Nk <= Nm );
     Eigen::MatrixXcd BlockTriDiag = Eigen::MatrixXcd::Zero(Nk,Nk);
     
     for ( int u=0; u<Nu; ++u ) {
@@ -775,8 +775,8 @@ private:
 			 GridBase *grid)
   {
     Glog << "diagonalize_lapack: Nu= "<<Nu<<" Nk= "<<Nk<<" Nm= "<<std::endl;
-    assert( Nk%Nu == 0 && Nm%Nu == 0 );
-    assert( Nk <= Nm );
+    GRID_ASSERT( Nk%Nu == 0 && Nm%Nu == 0 );
+    GRID_ASSERT( Nk <= Nm );
     Eigen::MatrixXcd BlockTriDiag = Eigen::MatrixXcd::Zero(Nk,Nk);
     
     for ( int u=0; u<Nu; ++u ) {
@@ -924,7 +924,7 @@ if (1){
       diagonalize_lapack(eval,lmd,lme,Nu,Nk,Nm,Qt,grid);
 #endif
     } else { 
-      assert(0);
+      GRID_ASSERT(0);
     }
   }
   
@@ -936,8 +936,8 @@ if (1){
          Eigen::MatrixXcd& M)
   {
     //    Glog << "unpackHermitBlockTriDiagMatToEigen() begin" << '\n'; 
-    assert( Nk%Nu == 0 && Nm%Nu == 0 );
-    assert( Nk <= Nm );
+    GRID_ASSERT( Nk%Nu == 0 && Nm%Nu == 0 );
+    GRID_ASSERT( Nk <= Nm );
     M = Eigen::MatrixXcd::Zero(Nk,Nk);
     
     // rearrange 
@@ -964,8 +964,8 @@ if (1){
          Eigen::MatrixXcd& M)
   {
     //    Glog << "packHermitBlockTriDiagMatfromEigen() begin" << '\n'; 
-    assert( Nk%Nu == 0 && Nm%Nu == 0 );
-    assert( Nk <= Nm );
+    GRID_ASSERT( Nk%Nu == 0 && Nm%Nu == 0 );
+    GRID_ASSERT( Nk <= Nm );
     
     // rearrange 
     for ( int u=0; u<Nu; ++u ) {

Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h

@@ -211,7 +211,7 @@ until convergence
   void calc(std::vector<RealD>& eval, std::vector<Field>& evec,  const Field& src, int& Nconv, bool reverse=false)
   {
     GridBase *grid = src.Grid();
-    assert(grid == evec[0].Grid());
+    GRID_ASSERT(grid == evec[0].Grid());
     
     //    GridLogIRL.TimingMode(1);
     std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
@@ -231,7 +231,7 @@ until convergence
     }
     std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
 	
-    assert(Nm <= evec.size() && Nm <= eval.size());
+    GRID_ASSERT(Nm <= evec.size() && Nm <= eval.size());
     
     // quickly get an idea of the largest eigenvalue to more properly normalize the residuum
     RealD evalMaxApprox = 0.0;
@@ -245,9 +245,10 @@ until convergence
 	_HermOp(src_n,tmp);
 	//	std::cout << GridLogMessage<< tmp<<std::endl; exit(0);
 	//	std::cout << GridLogIRL << " _HermOp " << norm2(tmp) << std::endl;
-	RealD vnum = real(innerProduct(src_n,tmp)); // HermOp.
+//	RealD vnum = real(innerProduct(src_n,tmp)); // HermOp.
+	RealD vnum = real(innerProduct(tmp,tmp)); // HermOp^2.
 	RealD vden = norm2(src_n);
-	RealD na = vnum/vden;
+	RealD na = std::sqrt(vnum/vden);
 	if (fabs(evalMaxApprox/na - 1.0) < 0.0001)
 	  i=_MAX_ITER_IRL_MEVAPP_;
 	evalMaxApprox = na;
@@ -255,6 +256,7 @@ until convergence
 	src_n = tmp;
       }
     }
+    std::cout << GridLogIRL << " Final evalMaxApprox  " << evalMaxApprox << std::endl;
 	
     std::vector<RealD> lme(Nm);  
     std::vector<RealD> lme2(Nm);
@@ -335,7 +337,7 @@ until convergence
       }
       std::cout<<GridLogIRL <<"QR decomposed "<<std::endl;
 
-      assert(k2<Nm);      assert(k2<Nm);      assert(k1>0);
+      GRID_ASSERT(k2<Nm);      GRID_ASSERT(k2<Nm);      GRID_ASSERT(k1>0);
 
       basisRotate(evec,Qt,k1-1,k2+1,0,Nm,Nm); /// big constraint on the basis
       std::cout<<GridLogIRL <<"basisRotated  by Qt *"<<k1-1<<","<<k2+1<<")"<<std::endl;
@@ -461,7 +463,7 @@ until convergence
   {
     std::cout<<GridLogDebug << "Lanczos step " <<k<<std::endl;
     const RealD tiny = 1.0e-20;
-    assert( k< Nm );
+    GRID_ASSERT( k< Nm );
 
     GridStopWatch gsw_op,gsw_o;
 
@@ -595,7 +597,7 @@ until convergence
     }  else if ( diagonalisation == IRLdiagonaliseWithEigen ) { 
       diagonalize_Eigen(lmd,lme,Nk,Nm,Qt,grid);
     } else { 
-      assert(0);
+      GRID_ASSERT(0);
     }
   }
 
@@ -685,7 +687,7 @@ void diagonalize_lapack(std::vector<RealD>& lmd,
     }
   }
 #else 
-  assert(0);
+  GRID_ASSERT(0);
 #endif
 }
 

Grid/algorithms/iterative/LocalCoherenceLanczos.h

@@ -80,7 +80,7 @@ public:
   ProjectedHermOp(LinearOperatorBase<FineField>& linop, std::vector<FineField> & _subspace) : 
     _Linop(linop), subspace(_subspace)
   {  
-    assert(subspace.size() >0);
+    GRID_ASSERT(subspace.size() >0);
   };
 
   void operator()(const CoarseField& in, CoarseField& out) {
@@ -346,12 +346,12 @@ public:
 
   void testFine(RealD resid) 
   {
-    assert(evals_fine.size() == nbasis);
-    assert(subspace.size() == nbasis);
+    GRID_ASSERT(evals_fine.size() == nbasis);
+    GRID_ASSERT(subspace.size() == nbasis);
     PlainHermOp<FineField>    Op(_FineOp);
     ImplicitlyRestartedLanczosHermOpTester<FineField> SimpleTester(Op);
     for(int k=0;k<nbasis;k++){
-      assert(SimpleTester.ReconstructEval(k,resid,subspace[k],evals_fine[k],1.0)==1);
+      GRID_ASSERT(SimpleTester.ReconstructEval(k,resid,subspace[k],evals_fine[k],1.0)==1);
     }
   }
 
@@ -359,8 +359,8 @@ public:
   //hence the smoother can be tuned after running the coarse Lanczos by using a different smoother here
   void testCoarse(RealD resid,ChebyParams cheby_smooth,RealD relax) 
   {
-    assert(evals_fine.size() == nbasis);
-    assert(subspace.size() == nbasis);
+    GRID_ASSERT(evals_fine.size() == nbasis);
+    GRID_ASSERT(subspace.size() == nbasis);
     //////////////////////////////////////////////////////////////////////////////////////////////////
     // create a smoother and see if we can get a cheap convergence test and smooth inside the IRL
     //////////////////////////////////////////////////////////////////////////////////////////////////
@@ -380,7 +380,7 @@ public:
   void calcFine(ChebyParams cheby_parms,int Nstop,int Nk,int Nm,RealD resid, 
 		RealD MaxIt, RealD betastp, int MinRes)
   {
-    assert(nbasis<=Nm);
+    GRID_ASSERT(nbasis<=Nm);
     Chebyshev<FineField>      Cheby(cheby_parms);
     FunctionHermOp<FineField> ChebyOp(Cheby,_FineOp);
     PlainHermOp<FineField>    Op(_FineOp);
@@ -400,8 +400,8 @@ public:
     IRL.calc(evals_fine,subspace,src,Nconv,false);
     
     // Shrink down to number saved
-    assert(Nstop>=nbasis);
-    assert(Nconv>=nbasis);
+    GRID_ASSERT(Nstop>=nbasis);
+    GRID_ASSERT(Nconv>=nbasis);
     evals_fine.resize(nbasis);
     subspace.resize(nbasis,_FineGrid);
   }
@@ -433,7 +433,7 @@ public:
     ImplicitlyRestartedLanczos<CoarseField> IRL(ChebyOp,ChebyOp,ChebySmoothTester,Nstop,Nk,Nm,resid,MaxIt,betastp,MinRes);
     int Nconv=0;
     IRL.calc(evals_coarse,evec_coarse,src,Nconv,false);
-    assert(Nconv>=Nstop);
+    GRID_ASSERT(Nconv>=Nstop);
     evals_coarse.resize(Nstop);
     evec_coarse.resize (Nstop,_CoarseGrid);
     for (int i=0;i<Nstop;i++){

Grid/algorithms/iterative/MinimalResidual.h

@@ -35,7 +35,7 @@ template<class Field> class MinimalResidual : public OperatorFunction<Field> {
  public:
   using OperatorFunction<Field>::operator();
 
-  bool ErrorOnNoConverge; // throw an assert when the MR fails to converge.
+  bool ErrorOnNoConverge; // throw an GRID_ASSERT when the MR fails to converge.
                           // Defaults true.
   RealD   Tolerance;
   Integer MaxIterations;
@@ -59,7 +59,7 @@ template<class Field> class MinimalResidual : public OperatorFunction<Field> {
 
     // Initial residual computation & set up
     RealD guess = norm2(psi);
-    assert(std::isnan(guess) == 0);
+    GRID_ASSERT(std::isnan(guess) == 0);
 
     RealD ssq = norm2(src);
     RealD rsq = Tolerance * Tolerance * ssq;
@@ -136,7 +136,7 @@ template<class Field> class MinimalResidual : public OperatorFunction<Field> {
         std::cout << GridLogMessage << "MR Time elapsed: Linalg  " << LinalgTimer.Elapsed() << std::endl;
 
         if (ErrorOnNoConverge)
-          assert(true_residual / Tolerance < 10000.0);
+          GRID_ASSERT(true_residual / Tolerance < 10000.0);
 
         IterationsToComplete = k;
 
@@ -148,7 +148,7 @@ template<class Field> class MinimalResidual : public OperatorFunction<Field> {
               << std::endl;
 
     if (ErrorOnNoConverge)
-      assert(0);
+      GRID_ASSERT(0);
 
     IterationsToComplete = k;
   }

Grid/algorithms/iterative/MixedPrecisionFlexibleGeneralisedMinimalResidual.h

@@ -37,7 +37,7 @@ class MixedPrecisionFlexibleGeneralisedMinimalResidual : public OperatorFunction
 
   using OperatorFunction<FieldD>::operator();
 
-  bool ErrorOnNoConverge; // Throw an assert when MPFGMRES fails to converge,
+  bool ErrorOnNoConverge; // Throw an GRID_ASSERT when MPFGMRES fails to converge,
                           // defaults to true
 
   RealD   Tolerance;
@@ -91,7 +91,7 @@ class MixedPrecisionFlexibleGeneralisedMinimalResidual : public OperatorFunction
     conformable(psi, src);
 
     RealD guess = norm2(psi);
-    assert(std::isnan(guess) == 0);
+    GRID_ASSERT(std::isnan(guess) == 0);
 
     RealD cp;
     RealD ssq = norm2(src);
@@ -150,7 +150,7 @@ class MixedPrecisionFlexibleGeneralisedMinimalResidual : public OperatorFunction
     std::cout << GridLogMessage << "MPFGMRES did NOT converge" << std::endl;
 
     if (ErrorOnNoConverge)
-      assert(0);
+      GRID_ASSERT(0);
   }
 
   RealD outerLoopBody(LinearOperatorBase<FieldD> &LinOp, const FieldD &src, FieldD &psi, RealD rsq) {
@@ -197,7 +197,7 @@ class MixedPrecisionFlexibleGeneralisedMinimalResidual : public OperatorFunction
       }
     }
 
-    assert(0); // Never reached
+    GRID_ASSERT(0); // Never reached
     return cp;
   }
 

Grid/algorithms/iterative/PrecConjugateResidual.h

@@ -112,7 +112,7 @@ public:
     }
 
     std::cout<<GridLogMessage<<"PrecConjugateResidual did NOT converge"<<std::endl;
-    assert(0);
+    GRID_ASSERT(0);
   }
 };
 NAMESPACE_END(Grid);

Grid/algorithms/iterative/PrecGeneralisedConjugateResidual.h

@@ -118,7 +118,7 @@ public:
 
     }
     GCRLogLevel<<"Variable Preconditioned GCR did not converge"<<std::endl;
-    //    assert(0);
+    //    GRID_ASSERT(0);
   }
 
   RealD GCRnStep(const Field &src, Field &psi,RealD rsq){
@@ -221,7 +221,7 @@ public:
       int northog = ((kp)>(mmax-1))?(mmax-1):(kp);  // if more than mmax done, we orthog all mmax history.
       for(int back=0;back<northog;back++){
 
-	int peri_back=(k-back)%mmax;   	  assert((k-back)>=0);
+	int peri_back=(k-back)%mmax;   	  GRID_ASSERT((k-back)>=0);
 
 	b=-real(innerProduct(q[peri_back],Az))/qq[peri_back];
 	p[peri_kp]=p[peri_kp]+b*p[peri_back];
@@ -231,9 +231,10 @@ public:
       qq[peri_kp]=norm2(q[peri_kp]); // could use axpy_norm
       LinalgTimer.Stop();
     }
-    assert(0); // never reached
+    GRID_ASSERT(0); // never reached
     return cp;
   }
 };
 NAMESPACE_END(Grid);
+#undef GCRLogLevel
 #endif

Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h

@@ -38,13 +38,14 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 ///////////////////////////////////////////////////////////////////////////////////////////////////////
 NAMESPACE_BEGIN(Grid);
 
-#define GCRLogLevel std::cout << GridLogMessage <<std::string(level,'\t')<< " Level "<<level<<" " 
+#define GCRLogLevel std::cout << GridLogMessage <<std::string(level,'\t')<< name<<" "
 
 template<class Field>
 class PrecGeneralisedConjugateResidualNonHermitian : public LinearFunction<Field> {
 public:                                                
   using LinearFunction<Field>::operator();
   RealD   Tolerance;
+  RealD   SSQ;
   Integer MaxIterations;
   int verbose;
   int mmax;
@@ -54,11 +55,18 @@ public:
   GridStopWatch PrecTimer;
   GridStopWatch MatTimer;
   GridStopWatch LinalgTimer;
+  std::string name;
+  int ZeroGuess  = 0;  // caller contract: guess is always zero => first-cycle r0 = src, skip the apply
+  int FirstCycle = 0;
 
   LinearFunction<Field>     &Preconditioner;
   LinearOperatorBase<Field> &Linop;
 
-  void Level(int lv) { level=lv; };
+  void Name(std::string _name) { name = _name; };
+
+  void Level(int n) { Name("Level " + std::to_string(n)); level = n; }
+
+  void SetZeroGuess(int z) { ZeroGuess = z; };
 
   PrecGeneralisedConjugateResidualNonHermitian(RealD tol,Integer maxit,LinearOperatorBase<Field> &_Linop,LinearFunction<Field> &Prec,int _mmax,int _nstep) : 
     Tolerance(tol), 
@@ -67,16 +75,17 @@ public:
     Preconditioner(Prec),
     mmax(_mmax),
     nstep(_nstep)
-  { 
-    level=1;
+  {
+    Level(1);
     verbose=1;
   };
 
   void operator() (const Field &src, Field &psi){
 
-    psi=Zero();
+    //    psi=Zero();
     RealD cp, ssq,rsq;
     ssq=norm2(src);
+    SSQ=ssq;
     rsq=Tolerance*Tolerance*ssq;
       
     Field r(src.Grid());
@@ -89,11 +98,12 @@ public:
     SolverTimer.Start();
 
     steps=0;
+    FirstCycle=1;
     for(int k=0;k<MaxIterations;k++){
 
       cp=GCRnStep(src,psi,rsq);
 
-      GCRLogLevel <<"PGCR("<<mmax<<","<<nstep<<") "<< steps <<" steps cp = "<<cp<<" target "<<rsq <<std::endl;
+      GCRLogLevel <<"PGCR("<<mmax<<","<<nstep<<") "<< steps <<" steps cp = "<<sqrt(cp/ssq)<<" target "<<sqrt(rsq/ssq) <<std::endl;
 
       if(cp<rsq) {
 
@@ -113,7 +123,7 @@ public:
 
     }
     GCRLogLevel<<"Variable Preconditioned GCR did not converge"<<std::endl;
-    //    assert(0);
+    //    GRID_ASSERT(0);
   }
 
   RealD GCRnStep(const Field &src, Field &psi,RealD rsq){
@@ -142,21 +152,25 @@ public:
     GCRLogLevel<< "PGCR nStep("<<nstep<<")"<<std::endl;
 
     //////////////////////////////////
-    // initial guess x0 is taken as nonzero.
-    // r0=src-A x0 = src
+    // r0 = src - A x0.  ZeroGuess: on the first cycle x0==0 by caller
+    // contract (enforced here), so r0 = src exactly; skip the apply.
+    // Restart cycles (psi!=0) always do the full computation.
     //////////////////////////////////
-    MatTimer.Start();
-    Linop.Op(psi,Az);
-    //    zAz = innerProduct(Az,psi);
-    zAAz= norm2(Az);
-    MatTimer.Stop();
-    
+    if (ZeroGuess && FirstCycle) {
+      psi = Zero();
+      LinalgTimer.Start();
+      r = src;
+      LinalgTimer.Stop();
+    } else {
+      MatTimer.Start();
+      Linop.Op(psi,Az);
+      MatTimer.Stop();
+      LinalgTimer.Start();
+      r=src-Az;
+      LinalgTimer.Stop();
+    }
+    FirstCycle=0;
 
-    LinalgTimer.Start();
-    r=src-Az;
-    LinalgTimer.Stop();
-    GCRLogLevel<< "PGCR true residual r = src - A psi   "<<norm2(r) <<std::endl;
-    
     /////////////////////
     // p = Prec(r)
     /////////////////////
@@ -181,6 +195,7 @@ public:
     
     cp =norm2(r);
     LinalgTimer.Stop();
+    GCRLogLevel<< "PGCR true residual "<< sqrt(cp/SSQ)     <<std::endl;
 
     for(int k=0;k<nstep;k++){
 
@@ -199,13 +214,12 @@ public:
       cp = axpy_norm(r,-a,q[peri_k],r);
       LinalgTimer.Stop();
 
-      GCRLogLevel<< "PGCR step["<<steps<<"]  resid " << cp << " target " <<rsq<<std::endl; 
+      GCRLogLevel<< "PGCR step["<<steps<<"]  resid " << sqrt(cp/SSQ)<<std::endl;
 
       if((k==nstep-1)||(cp<rsq)){
 	return cp;
       }
 
-
       PrecTimer.Start();
       Preconditioner(r,z);// solve Az = r
       PrecTimer.Stop();
@@ -224,7 +238,7 @@ public:
       int northog = ((kp)>(mmax-1))?(mmax-1):(kp);  // if more than mmax done, we orthog all mmax history.
       for(int back=0;back<northog;back++){
 
-	int peri_back=(k-back)%mmax;   	  assert((k-back)>=0);
+	int peri_back=(k-back)%mmax;   	  GRID_ASSERT((k-back)>=0);
 
 	b=-real(innerProduct(q[peri_back],Az))/qq[peri_back];
 	p[peri_kp]=p[peri_kp]+b*p[peri_back];
@@ -234,9 +248,11 @@ public:
       qq[peri_kp]=norm2(q[peri_kp]); // could use axpy_norm
       LinalgTimer.Stop();
     }
-    assert(0); // never reached
+    GRID_ASSERT(0); // never reached
     return cp;
   }
 };
 NAMESPACE_END(Grid);
+
+#undef GCRLogLevel
 #endif

Grid/algorithms/iterative/QuasiMinimalResidual.h

@@ -79,7 +79,7 @@ class QuasiMinimalResidual : public OperatorFunction<Field> {
 
     LinOp.Op(x,r); r = b - r;
 
-    assert(normb> 0.0);
+    GRID_ASSERT(normb> 0.0);
 
     resid = norm2(r)/normb;
     if (resid <= Tolerance) {
@@ -105,8 +105,8 @@ class QuasiMinimalResidual : public OperatorFunction<Field> {
     for (int i = 1; i <= MaxIterations; i++) {
 
       // Breakdown tests
-      assert( rho != 0.0);
-      assert( xi  != 0.0);
+      GRID_ASSERT( rho != 0.0);
+      GRID_ASSERT( xi  != 0.0);
 
       v = (1. / rho) * v_tld;
       y = (1. / rho) * y;
@@ -134,10 +134,10 @@ class QuasiMinimalResidual : public OperatorFunction<Field> {
       ep=Zep.real();
       std::cout << "Zep "<<Zep <<std::endl;
       // Complex Audit
-      assert(abs(ep)>0);
+      GRID_ASSERT(abs(ep)>0);
 
       beta = ep / delta;
-      assert(abs(beta)>0);
+      GRID_ASSERT(abs(beta)>0);
 
       v_tld = p_tld - beta * v;
       y = v_tld;
@@ -158,7 +158,7 @@ class QuasiMinimalResidual : public OperatorFunction<Field> {
       std::cout << "theta "<<theta<<std::endl;
       std::cout << "gamma "<<gamma<<std::endl;
 
-      assert(abs(gamma)> 0.0);
+      GRID_ASSERT(abs(gamma)> 0.0);
 
       eta = -eta * rho_1 * gamma* gamma / (beta * gamma_1 * gamma_1);
 
@@ -178,7 +178,7 @@ class QuasiMinimalResidual : public OperatorFunction<Field> {
       }
       std::cout << "Iteration "<<i<<" resid " << resid<<std::endl;
     }
-    assert(0);
+    GRID_ASSERT(0);
     return;                            // no convergence
   }
 #else

Grid/algorithms/iterative/SchurRedBlack.h

@@ -327,9 +327,9 @@ namespace Grid {
       /////////////////////////////////////////////////////
       // src_o = (source_o - Moe MeeInv source_e)
       /////////////////////////////////////////////////////
-      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.Checkerboard() ==Even);
-      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.Checkerboard() ==Odd);     
-      tmp=src_o-Mtmp;                  assert(  tmp.Checkerboard() ==Odd);     
+      _Matrix.MooeeInv(src_e,tmp);     GRID_ASSERT(  tmp.Checkerboard() ==Even);
+      _Matrix.Meooe   (tmp,Mtmp);      GRID_ASSERT( Mtmp.Checkerboard() ==Odd);     
+      tmp=src_o-Mtmp;                  GRID_ASSERT(  tmp.Checkerboard() ==Odd);     
 
       _Matrix.Mooee(tmp,src_o); // Extra factor of "m" in source from dumb choice of matrix norm.
     }
@@ -347,17 +347,17 @@ namespace Grid {
       ///////////////////////////////////////////////////
       // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
       ///////////////////////////////////////////////////
-      _Matrix.Meooe(sol_o,tmp);        assert(  tmp.Checkerboard()   ==Even);
-      src_e = src_e-tmp;               assert(  src_e.Checkerboard() ==Even);
-      _Matrix.MooeeInv(src_e,sol_e);   assert(  sol_e.Checkerboard() ==Even);
+      _Matrix.Meooe(sol_o,tmp);        GRID_ASSERT(  tmp.Checkerboard()   ==Even);
+      src_e = src_e-tmp;               GRID_ASSERT(  src_e.Checkerboard() ==Even);
+      _Matrix.MooeeInv(src_e,sol_e);   GRID_ASSERT(  sol_e.Checkerboard() ==Even);
      
-      setCheckerboard(sol,sol_e); assert(  sol_e.Checkerboard() ==Even);
-      setCheckerboard(sol,sol_o); assert(  sol_o.Checkerboard() ==Odd );
+      setCheckerboard(sol,sol_e); GRID_ASSERT(  sol_e.Checkerboard() ==Even);
+      setCheckerboard(sol,sol_o); GRID_ASSERT(  sol_o.Checkerboard() ==Odd );
     }
     virtual void RedBlackSolve   (Matrix & _Matrix,const Field &src_o, Field &sol_o)
     {
       SchurStaggeredOperator<Matrix,Field> _HermOpEO(_Matrix);
-      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o);  assert(sol_o.Checkerboard()==Odd);
+      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o);  GRID_ASSERT(sol_o.Checkerboard()==Odd);
     };
     virtual void RedBlackSolve   (Matrix & _Matrix,const std::vector<Field> &src_o,  std::vector<Field> &sol_o)
     {
@@ -396,13 +396,13 @@ namespace Grid {
       /////////////////////////////////////////////////////
       // src_o = Mdag * (source_o - Moe MeeInv source_e)
       /////////////////////////////////////////////////////
-      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.Checkerboard() ==Even);
-      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.Checkerboard() ==Odd);     
-      tmp=src_o-Mtmp;                  assert(  tmp.Checkerboard() ==Odd);     
+      _Matrix.MooeeInv(src_e,tmp);     GRID_ASSERT(  tmp.Checkerboard() ==Even);
+      _Matrix.Meooe   (tmp,Mtmp);      GRID_ASSERT( Mtmp.Checkerboard() ==Odd);     
+      tmp=src_o-Mtmp;                  GRID_ASSERT(  tmp.Checkerboard() ==Odd);     
 
       // get the right MpcDag
       SchurDiagMooeeOperator<Matrix,Field> _HermOpEO(_Matrix);
-      _HermOpEO.MpcDag(tmp,src_o);     assert(src_o.Checkerboard() ==Odd);       
+      _HermOpEO.MpcDag(tmp,src_o);     GRID_ASSERT(src_o.Checkerboard() ==Odd);       
 
     }
     virtual void RedBlackSolution(Matrix & _Matrix,const Field &sol_o, const Field &src_e,Field &sol)
@@ -416,17 +416,17 @@ namespace Grid {
       ///////////////////////////////////////////////////
       // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
       ///////////////////////////////////////////////////
-      _Matrix.Meooe(sol_o,tmp);          assert(  tmp.Checkerboard()   ==Even);
-      src_e_i = src_e-tmp;               assert(  src_e_i.Checkerboard() ==Even);
-      _Matrix.MooeeInv(src_e_i,sol_e);   assert(  sol_e.Checkerboard() ==Even);
+      _Matrix.Meooe(sol_o,tmp);          GRID_ASSERT(  tmp.Checkerboard()   ==Even);
+      src_e_i = src_e-tmp;               GRID_ASSERT(  src_e_i.Checkerboard() ==Even);
+      _Matrix.MooeeInv(src_e_i,sol_e);   GRID_ASSERT(  sol_e.Checkerboard() ==Even);
      
-      setCheckerboard(sol,sol_e); assert(  sol_e.Checkerboard() ==Even);
-      setCheckerboard(sol,sol_o); assert(  sol_o.Checkerboard() ==Odd );
+      setCheckerboard(sol,sol_e); GRID_ASSERT(  sol_e.Checkerboard() ==Even);
+      setCheckerboard(sol,sol_o); GRID_ASSERT(  sol_o.Checkerboard() ==Odd );
     }
     virtual void RedBlackSolve   (Matrix & _Matrix,const Field &src_o, Field &sol_o)
     {
       SchurDiagMooeeOperator<Matrix,Field> _HermOpEO(_Matrix);
-      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o);  assert(sol_o.Checkerboard()==Odd);
+      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o);  GRID_ASSERT(sol_o.Checkerboard()==Odd);
     };
     virtual void RedBlackSolve   (Matrix & _Matrix,const std::vector<Field> &src_o,  std::vector<Field> &sol_o)
     {
@@ -461,9 +461,9 @@ namespace Grid {
         /////////////////////////////////////////////////////
         // src_o = Mdag * (source_o - Moe MeeInv source_e)
         /////////////////////////////////////////////////////
-        _Matrix.MooeeInv(src_e, tmp);   assert(   tmp.Checkerboard() == Even );
-        _Matrix.Meooe   (tmp, Mtmp);    assert(  Mtmp.Checkerboard() == Odd  );     
-        src_o -= Mtmp;                  assert( src_o.Checkerboard() == Odd  );     
+        _Matrix.MooeeInv(src_e, tmp);   GRID_ASSERT(   tmp.Checkerboard() == Even );
+        _Matrix.Meooe   (tmp, Mtmp);    GRID_ASSERT(  Mtmp.Checkerboard() == Odd  );     
+        src_o -= Mtmp;                  GRID_ASSERT( src_o.Checkerboard() == Odd  );     
       }
       
       virtual void RedBlackSolution(Matrix& _Matrix, const Field& sol_o, const Field& src_e, Field& sol)
@@ -478,18 +478,18 @@ namespace Grid {
         ///////////////////////////////////////////////////
         // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
         ///////////////////////////////////////////////////
-        _Matrix.Meooe(sol_o, tmp);         assert(     tmp.Checkerboard() == Even );
-        src_e_i = src_e - tmp;             assert( src_e_i.Checkerboard() == Even );
-        _Matrix.MooeeInv(src_e_i, sol_e);  assert(   sol_e.Checkerboard() == Even );
+        _Matrix.Meooe(sol_o, tmp);         GRID_ASSERT(     tmp.Checkerboard() == Even );
+        src_e_i = src_e - tmp;             GRID_ASSERT( src_e_i.Checkerboard() == Even );
+        _Matrix.MooeeInv(src_e_i, sol_e);  GRID_ASSERT(   sol_e.Checkerboard() == Even );
        
-        setCheckerboard(sol, sol_e); assert( sol_e.Checkerboard() == Even );
-        setCheckerboard(sol, sol_o); assert( sol_o.Checkerboard() == Odd  );
+        setCheckerboard(sol, sol_e); GRID_ASSERT( sol_e.Checkerboard() == Even );
+        setCheckerboard(sol, sol_o); GRID_ASSERT( sol_o.Checkerboard() == Odd  );
       }
 
       virtual void RedBlackSolve(Matrix& _Matrix, const Field& src_o, Field& sol_o)
       {
         NonHermitianSchurDiagMooeeOperator<Matrix,Field> _OpEO(_Matrix);
-        this->_HermitianRBSolver(_OpEO, src_o, sol_o);  assert(sol_o.Checkerboard() == Odd);
+        this->_HermitianRBSolver(_OpEO, src_o, sol_o);  GRID_ASSERT(sol_o.Checkerboard() == Odd);
       }
 
       virtual void RedBlackSolve(Matrix& _Matrix, const std::vector<Field>& src_o, std::vector<Field>& sol_o)
@@ -539,13 +539,13 @@ namespace Grid {
       /////////////////////////////////////////////////////
       // src_o = Mpcdag *MooeeInv * (source_o - Moe MeeInv source_e)
       /////////////////////////////////////////////////////
-      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.Checkerboard() ==Even);
-      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.Checkerboard() ==Odd);     
+      _Matrix.MooeeInv(src_e,tmp);     GRID_ASSERT(  tmp.Checkerboard() ==Even);
+      _Matrix.Meooe   (tmp,Mtmp);      GRID_ASSERT( Mtmp.Checkerboard() ==Odd);     
       Mtmp=src_o-Mtmp;                 
-      _Matrix.MooeeInv(Mtmp,tmp);      assert( tmp.Checkerboard() ==Odd);     
+      _Matrix.MooeeInv(Mtmp,tmp);      GRID_ASSERT( tmp.Checkerboard() ==Odd);     
       
       // get the right MpcDag
-      _HermOpEO.MpcDag(tmp,src_o);     assert(src_o.Checkerboard() ==Odd);       
+      _HermOpEO.MpcDag(tmp,src_o);     GRID_ASSERT(src_o.Checkerboard() ==Odd);       
     }
 
     virtual void RedBlackSolution(Matrix & _Matrix,const Field &sol_o, const Field &src_e,Field &sol)
@@ -560,12 +560,12 @@ namespace Grid {
       ///////////////////////////////////////////////////
       // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
       ///////////////////////////////////////////////////
-      _Matrix.Meooe(sol_o,tmp);    assert(  tmp.Checkerboard()   ==Even);
-      tmp = src_e-tmp;             assert(  src_e.Checkerboard() ==Even);
-      _Matrix.MooeeInv(tmp,sol_e); assert(  sol_e.Checkerboard() ==Even);
+      _Matrix.Meooe(sol_o,tmp);    GRID_ASSERT(  tmp.Checkerboard()   ==Even);
+      tmp = src_e-tmp;             GRID_ASSERT(  src_e.Checkerboard() ==Even);
+      _Matrix.MooeeInv(tmp,sol_e); GRID_ASSERT(  sol_e.Checkerboard() ==Even);
      
-      setCheckerboard(sol,sol_e);  assert(  sol_e.Checkerboard() ==Even);
-      setCheckerboard(sol,sol_o);  assert(  sol_o.Checkerboard() ==Odd );
+      setCheckerboard(sol,sol_e);  GRID_ASSERT(  sol_e.Checkerboard() ==Even);
+      setCheckerboard(sol,sol_o);  GRID_ASSERT(  sol_o.Checkerboard() ==Odd );
     };
 
     virtual void RedBlackSolve   (Matrix & _Matrix,const Field &src_o, Field &sol_o)
@@ -612,12 +612,12 @@ namespace Grid {
       /////////////////////////////////////////////////////
       // src_o = Mdag * (source_o - Moe MeeInv source_e)
       /////////////////////////////////////////////////////
-      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.Checkerboard() ==Even);
-      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.Checkerboard() ==Odd);     
-      tmp=src_o-Mtmp;                  assert(  tmp.Checkerboard() ==Odd);     
+      _Matrix.MooeeInv(src_e,tmp);     GRID_ASSERT(  tmp.Checkerboard() ==Even);
+      _Matrix.Meooe   (tmp,Mtmp);      GRID_ASSERT( Mtmp.Checkerboard() ==Odd);     
+      tmp=src_o-Mtmp;                  GRID_ASSERT(  tmp.Checkerboard() ==Odd);     
 
       // get the right MpcDag
-      _HermOpEO.MpcDag(tmp,src_o);     assert(src_o.Checkerboard() ==Odd);       
+      _HermOpEO.MpcDag(tmp,src_o);     GRID_ASSERT(src_o.Checkerboard() ==Odd);       
     }
 
     virtual void RedBlackSolution(Matrix & _Matrix,const Field &sol_o, const Field &src_e,Field &sol)
@@ -638,12 +638,12 @@ namespace Grid {
       ///////////////////////////////////////////////////
       // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
       ///////////////////////////////////////////////////
-      _Matrix.Meooe(sol_o_i,tmp);    assert(  tmp.Checkerboard()   ==Even);
-      tmp = src_e-tmp;               assert(  src_e.Checkerboard() ==Even);
-      _Matrix.MooeeInv(tmp,sol_e);   assert(  sol_e.Checkerboard() ==Even);
+      _Matrix.Meooe(sol_o_i,tmp);    GRID_ASSERT(  tmp.Checkerboard()   ==Even);
+      tmp = src_e-tmp;               GRID_ASSERT(  src_e.Checkerboard() ==Even);
+      _Matrix.MooeeInv(tmp,sol_e);   GRID_ASSERT(  sol_e.Checkerboard() ==Even);
      
-      setCheckerboard(sol,sol_e);    assert(  sol_e.Checkerboard() ==Even);
-      setCheckerboard(sol,sol_o_i);  assert(  sol_o_i.Checkerboard() ==Odd );
+      setCheckerboard(sol,sol_e);    GRID_ASSERT(  sol_e.Checkerboard() ==Even);
+      setCheckerboard(sol,sol_o_i);  GRID_ASSERT(  sol_o_i.Checkerboard() ==Odd );
     };
 
     virtual void RedBlackSolve   (Matrix & _Matrix,const Field &src_o, Field &sol_o)
@@ -684,9 +684,9 @@ namespace Grid {
         /////////////////////////////////////////////////////
         // src_o = Mdag * (source_o - Moe MeeInv source_e)
         /////////////////////////////////////////////////////
-        _Matrix.MooeeInv(src_e, tmp);   assert(   tmp.Checkerboard() == Even );
-        _Matrix.Meooe   (tmp, Mtmp);    assert(  Mtmp.Checkerboard() == Odd  );     
-        src_o -= Mtmp;                  assert( src_o.Checkerboard() == Odd  );     
+        _Matrix.MooeeInv(src_e, tmp);   GRID_ASSERT(   tmp.Checkerboard() == Even );
+        _Matrix.Meooe   (tmp, Mtmp);    GRID_ASSERT(  Mtmp.Checkerboard() == Odd  );     
+        src_o -= Mtmp;                  GRID_ASSERT( src_o.Checkerboard() == Odd  );     
       }
 
       virtual void RedBlackSolution(Matrix& _Matrix, const Field& sol_o, const Field& src_e, Field& sol)
@@ -707,12 +707,12 @@ namespace Grid {
         ///////////////////////////////////////////////////
         // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
         ///////////////////////////////////////////////////
-        _Matrix.Meooe(sol_o_i, tmp);    assert(   tmp.Checkerboard() == Even );
-        tmp = src_e - tmp;              assert( src_e.Checkerboard() == Even );
-        _Matrix.MooeeInv(tmp, sol_e);   assert( sol_e.Checkerboard() == Even );
+        _Matrix.Meooe(sol_o_i, tmp);    GRID_ASSERT(   tmp.Checkerboard() == Even );
+        tmp = src_e - tmp;              GRID_ASSERT( src_e.Checkerboard() == Even );
+        _Matrix.MooeeInv(tmp, sol_e);   GRID_ASSERT( sol_e.Checkerboard() == Even );
        
-        setCheckerboard(sol, sol_e);    assert(   sol_e.Checkerboard() == Even );
-        setCheckerboard(sol, sol_o_i);  assert( sol_o_i.Checkerboard() == Odd  );
+        setCheckerboard(sol, sol_e);    GRID_ASSERT(   sol_e.Checkerboard() == Even );
+        setCheckerboard(sol, sol_o_i);  GRID_ASSERT( sol_o_i.Checkerboard() == Odd  );
       };
 
       virtual void RedBlackSolve(Matrix& _Matrix, const Field& src_o, Field& sol_o)

Grid/algorithms/multigrid/Aggregates.h

@@ -30,6 +30,8 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 /*  END LEGAL */
 #pragma once
 
+#include <Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h>
+
 NAMESPACE_BEGIN(Grid);
 
 inline RealD AggregatePowerLaw(RealD x)
@@ -64,7 +66,21 @@ public:
   {
   };
   
-  
+  void GlobalOrthonormalise(void)
+  {
+    // Normalise all vectors
+    for(int i=0;i<nbasis; i++){
+      RealD scale = std::pow(norm2(subspace[i]),-0.5); 
+      subspace[i] = subspace[i]*scale;
+    }
+    for(int i=0;i<nbasis; i++){
+      for(int j=0;j<i; j++){
+	basisOrthogonalize(subspace,subspace[i],j);
+      }
+      RealD scale = std::pow(norm2(subspace[i]),-0.5); 
+      subspace[i] = subspace[i]*scale;
+    }
+  }  
   void Orthogonalise(void){
     CoarseScalar InnerProd(CoarseGrid); 
     //    std::cout << GridLogMessage <<" Block Gramm-Schmidt pass 1"<<std::endl;
@@ -95,7 +111,7 @@ public:
 
     RealD scale;
 
-    ConjugateGradient<FineField> CG(1.0e-2,100,false);
+    ConjugateGradient<FineField> CG(1.0e-4,2000,false);
     FineField noise(FineGrid);
     FineField Mn(FineGrid);
 
@@ -108,14 +124,16 @@ public:
       
       hermop.Op(noise,Mn); std::cout<<GridLogMessage << "noise   ["<<b<<"] <n|MdagM|n> "<<norm2(Mn)<<std::endl;
 
-      for(int i=0;i<1;i++){
+      for(int i=0;i<2;i++){
 
 	CG(hermop,noise,subspace[b]);
 
 	noise = subspace[b];
 	scale = std::pow(norm2(noise),-0.5); 
 	noise=noise*scale;
-
+	
+	hermop.Op(noise,Mn); std::cout<<GridLogMessage << "intermediate["<<i<<"] <i|MdagM|i> "<<norm2(Mn)<<std::endl;
+	
       }
 
       hermop.Op(noise,Mn); std::cout<<GridLogMessage << "filtered["<<b<<"] <f|MdagM|f> "<<norm2(Mn)<<std::endl;
@@ -124,6 +142,60 @@ public:
     }
   }
 
+  virtual void CreateSubspaceGCR(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &DiracOp,int nn=nbasis)
+  {
+    RealD scale;
+
+    TrivialPrecon<FineField> simple_fine;
+    //    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(0.0005,30,DiracOp,simple_fine,20,20);
+    PrecGeneralisedConjugateResidualNonHermitian<FineField> GCR(0.0005,30,DiracOp,simple_fine,10,10);
+    FineField noise(FineGrid);
+    FineField src(FineGrid);
+    FineField guess(FineGrid);
+    FineField Mn(FineGrid);
+
+    for(int b=0;b<nn;b++){
+      
+      subspace[b] = Zero();
+      gaussian(RNG,noise);
+      scale = std::pow(norm2(noise),-0.5); 
+      noise=noise*scale;
+      
+      DiracOp.Op(noise,Mn); std::cout<<GridLogMessage << "noise   ["<<b<<"] <n|Op|n> "<<innerProduct(noise,Mn)<<std::endl;
+
+      for(int i=0;i<3;i++){
+	//  void operator() (const Field &src, Field &psi){
+#if 1
+	if (i==0)std::cout << GridLogMessage << " inverting on noise "<<std::endl;
+	src = noise;
+	guess=Zero();
+	GCR(src,guess);
+	subspace[b] = guess;
+#else
+	if (i==0)std::cout << GridLogMessage << " inverting on zero "<<std::endl;
+	src=Zero();
+	guess = noise;
+	GCR(src,guess);
+	subspace[b] = guess;
+#endif
+	noise = subspace[b];
+	scale = std::pow(norm2(noise),-0.5); 
+	noise=noise*scale;
+	
+	DiracOp.Op(noise,Mn); std::cout<<GridLogMessage << "intermediate["<<i<<"] <f|Op|f> "<<innerProduct(noise,Mn)<<" <f|OpDagOp|f>"<<norm2(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;
+
+    }
+    GlobalOrthonormalise();
+  }
+
   ////////////////////////////////////////////////////////////////////////////////////////////////
   // World of possibilities here. But have tried quite a lot of experiments (250+ jobs run on Summit)
   // and this is the best I found
@@ -160,14 +232,21 @@ public:
 
     int b =0;
     {
+      ComplexD ip;
       // Filter
       Chebyshev<FineField> Cheb(lo,hi,orderfilter);
       Cheb(hermop,noise,Mn);
       // normalise
       scale = std::pow(norm2(Mn),-0.5); 	Mn=Mn*scale;
       subspace[b]   = Mn;
-      hermop.Op(Mn,tmp); 
-      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+
+      hermop.Op(Mn,tmp);
+      ip= innerProduct(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|Op|n> "<<norm2(tmp)<<" "<<ip<<std::endl;
+
+      hermop.AdjOp(Mn,tmp); 
+      ip = innerProduct(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|AdjOp|n> "<<norm2(tmp)<<" "<<ip<<std::endl;
       b++;
     }
 
@@ -213,8 +292,18 @@ public:
 	  Mn=*Tnp;
 	  scale = std::pow(norm2(Mn),-0.5);         Mn=Mn*scale;
 	  subspace[b] = Mn;
-	  hermop.Op(Mn,tmp); 
-	  std::cout<<GridLogMessage << n<<" filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+
+
+	  ComplexD ip;
+
+	  hermop.Op(Mn,tmp);
+	  ip= innerProduct(Mn,tmp); 
+	  std::cout<<GridLogMessage << "filt ["<<b<<"] <n|Op|n> "<<norm2(tmp)<<" "<<ip<<std::endl;
+
+	  hermop.AdjOp(Mn,tmp); 
+	  ip = innerProduct(Mn,tmp); 
+	  std::cout<<GridLogMessage << "filt ["<<b<<"] <n|AdjOp|n> "<<norm2(tmp)<<" "<<ip<<std::endl;
+	  
 	  b++;
 	}
 
@@ -226,7 +315,7 @@ public:
 	  
       }
     }
-    assert(b==nn);
+    GRID_ASSERT(b==nn);
   }
 
 

Grid/algorithms/multigrid/CoarsenedMatrix.h

@@ -99,7 +99,7 @@ public:
   CoarseMatrix AselfInvEven;
   CoarseMatrix AselfInvOdd;
 
-  Vector<RealD> dag_factor;
+  deviceVector<RealD> dag_factor;
 
   ///////////////////////
   // Interface
@@ -124,9 +124,13 @@ public:
     int npoint = geom.npoint;
     typedef LatticeView<Cobj> Aview;
       
-    Vector<Aview> AcceleratorViewContainer;
+    deviceVector<Aview> AcceleratorViewContainer(geom.npoint);
+    hostVector<Aview>   hAcceleratorViewContainer(geom.npoint);
   
-    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer.push_back(A[p].View(AcceleratorRead));
+    for(int p=0;p<geom.npoint;p++) {
+      hAcceleratorViewContainer[p] = A[p].View(AcceleratorRead);
+      acceleratorPut(AcceleratorViewContainer[p],hAcceleratorViewContainer[p]);
+    }
     Aview *Aview_p = & AcceleratorViewContainer[0];
 
     const int Nsimd = CComplex::Nsimd();
@@ -161,7 +165,7 @@ public:
       coalescedWrite(out_v[ss](b),res);
       });
 
-    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer[p].ViewClose();
+    for(int p=0;p<geom.npoint;p++) hAcceleratorViewContainer[p].ViewClose();
   };
 
   void Mdag (const CoarseVector &in, CoarseVector &out)
@@ -190,9 +194,14 @@ public:
     int npoint = geom.npoint;
     typedef LatticeView<Cobj> Aview;
 
-    Vector<Aview> AcceleratorViewContainer;
 
-    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer.push_back(A[p].View(AcceleratorRead));
+    deviceVector<Aview> AcceleratorViewContainer(geom.npoint);
+    hostVector<Aview>   hAcceleratorViewContainer(geom.npoint);
+  
+    for(int p=0;p<geom.npoint;p++) {
+      hAcceleratorViewContainer[p] = A[p].View(AcceleratorRead);
+      acceleratorPut(AcceleratorViewContainer[p],hAcceleratorViewContainer[p]);
+    }
     Aview *Aview_p = & AcceleratorViewContainer[0];
 
     const int Nsimd = CComplex::Nsimd();
@@ -201,10 +210,10 @@ public:
 
     int osites=Grid()->oSites();
 
-    Vector<int> points(geom.npoint, 0);
-    for(int p=0; p<geom.npoint; p++)
-      points[p] = geom.points_dagger[p];
-
+    deviceVector<int> points(geom.npoint);
+    for(int p=0; p<geom.npoint; p++) { 
+      acceleratorPut(points[p],geom.points_dagger[p]);
+    }
     auto points_p = &points[0];
 
     RealD* dag_factor_p = &dag_factor[0];
@@ -236,7 +245,7 @@ public:
       coalescedWrite(out_v[ss](b),res);
       });
 
-    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer[p].ViewClose();
+    for(int p=0;p<geom.npoint;p++) hAcceleratorViewContainer[p].ViewClose();
   }
 
   void MdirComms(const CoarseVector &in)
@@ -251,8 +260,14 @@ public:
     out.Checkerboard() = in.Checkerboard();
 
     typedef LatticeView<Cobj> Aview;
-    Vector<Aview> AcceleratorViewContainer;
-    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer.push_back(A[p].View(AcceleratorRead));
+
+    deviceVector<Aview> AcceleratorViewContainer(geom.npoint);
+    hostVector<Aview>   hAcceleratorViewContainer(geom.npoint);
+  
+    for(int p=0;p<geom.npoint;p++) {
+      hAcceleratorViewContainer[p] = A[p].View(AcceleratorRead);
+      acceleratorPut(AcceleratorViewContainer[p],hAcceleratorViewContainer[p]);
+    }
     Aview *Aview_p = & AcceleratorViewContainer[0];
 
     autoView( out_v , out, AcceleratorWrite);
@@ -285,7 +300,7 @@ public:
       }
       coalescedWrite(out_v[ss](b),res);
     });
-    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer[p].ViewClose();
+    for(int p=0;p<geom.npoint;p++) hAcceleratorViewContainer[p].ViewClose();
   }
   void MdirAll(const CoarseVector &in,std::vector<CoarseVector> &out)
   {
@@ -294,7 +309,7 @@ public:
     if ((out.size()!=ndir)&&(out.size()!=ndir+1)) { 
       std::cout <<"MdirAll out size "<< out.size()<<std::endl;
       std::cout <<"MdirAll ndir "<< ndir<<std::endl;
-      assert(0);
+      GRID_ASSERT(0);
     }
     for(int p=0;p<ndir;p++){
       MdirCalc(in,out[p],p);
@@ -358,7 +373,7 @@ public:
     conformable(in.Grid(), _cbgrid);    // verifies half grid
     conformable(in.Grid(), out.Grid()); // drops the cb check
 
-    assert(in.Checkerboard() == Even);
+    GRID_ASSERT(in.Checkerboard() == Even);
     out.Checkerboard() = Odd;
 
     DhopInternal(StencilEven, Aodd, in, out, dag);
@@ -368,7 +383,7 @@ public:
     conformable(in.Grid(), _cbgrid);    // verifies half grid
     conformable(in.Grid(), out.Grid()); // drops the cb check
 
-    assert(in.Checkerboard() == Odd);
+    GRID_ASSERT(in.Checkerboard() == Odd);
     out.Checkerboard() = Even;
 
     DhopInternal(StencilOdd, Aeven, in, out, dag);
@@ -376,7 +391,7 @@ public:
 
   void MooeeInternal(const CoarseVector &in, CoarseVector &out, int dag, int inv) {
     out.Checkerboard() = in.Checkerboard();
-    assert(in.Checkerboard() == Odd || in.Checkerboard() == Even);
+    GRID_ASSERT(in.Checkerboard() == Odd || in.Checkerboard() == Even);
 
     CoarseMatrix *Aself = nullptr;
     if(in.Grid()->_isCheckerBoarded) {
@@ -391,7 +406,7 @@ public:
       Aself = (inv) ? &AselfInv : &A[geom.npoint-1];
       DselfInternal(Stencil, *Aself, in, out, dag);
     }
-    assert(Aself != nullptr);
+    GRID_ASSERT(Aself != nullptr);
   }
 
   void DselfInternal(CartesianStencil<siteVector,siteVector,DefaultImplParams> &st, CoarseMatrix &a,
@@ -469,14 +484,20 @@ public:
 
     // determine in what order we need the points
     int npoint = geom.npoint-1;
-    Vector<int> points(npoint, 0);
-    for(int p=0; p<npoint; p++)
-      points[p] = (dag && !hermitian) ? geom.points_dagger[p] : p;
-
+    deviceVector<int> points(npoint);
+    for(int p=0; p<npoint; p++) {
+      int val = (dag && !hermitian) ? geom.points_dagger[p] : p;
+      acceleratorPut(points[p], val);
+    }
     auto points_p = &points[0];
 
-    Vector<Aview> AcceleratorViewContainer;
-    for(int p=0;p<npoint;p++) AcceleratorViewContainer.push_back(a[p].View(AcceleratorRead));
+    deviceVector<Aview> AcceleratorViewContainer(geom.npoint);
+    hostVector<Aview>   hAcceleratorViewContainer(geom.npoint);
+  
+    for(int p=0;p<geom.npoint;p++) {
+      hAcceleratorViewContainer[p] = a[p].View(AcceleratorRead);
+      acceleratorPut(AcceleratorViewContainer[p],hAcceleratorViewContainer[p]);
+    }
     Aview *Aview_p = & AcceleratorViewContainer[0];
 
     const int Nsimd = CComplex::Nsimd();
@@ -539,7 +560,7 @@ public:
       });
     }
 
-    for(int p=0;p<npoint;p++) AcceleratorViewContainer[p].ViewClose();
+    for(int p=0;p<npoint;p++) hAcceleratorViewContainer[p].ViewClose();
   }
   
   CoarsenedMatrix(GridCartesian &CoarseGrid, int hermitian_=0) 	:
@@ -590,11 +611,13 @@ public:
     }
 
     // GPU readable prefactor
+    std::vector<RealD> h_dag_factor(nbasis*nbasis);
     thread_for(i, nbasis*nbasis, {
       int j = i/nbasis;
       int k = i%nbasis;
-      dag_factor[i] = dag_factor_eigen(j, k);
+      h_dag_factor[i] = dag_factor_eigen(j, k);
     });
+    acceleratorCopyToDevice(&h_dag_factor[0],&dag_factor[0],dag_factor.size()*sizeof(RealD));
   }
 
   void CoarsenOperator(GridBase *FineGrid,LinearOperatorBase<Lattice<Fobj> > &linop,
@@ -674,7 +697,7 @@ public:
     evenmask = where(mod(bcb,2)==(Integer)0,one,zero);
     oddmask  = one-evenmask;
 
-    assert(self_stencil!=-1);
+    GRID_ASSERT(self_stencil!=-1);
 
     for(int i=0;i<nbasis;i++){
 

Grid/algorithms/multigrid/GeneralCoarsenedMatrix.h

@@ -31,6 +31,7 @@ Author: Peter Boyle <pboyle@bnl.gov>
 
 #include <Grid/lattice/PaddedCell.h>
 #include <Grid/stencil/GeneralLocalStencil.h>
+#include <Grid/algorithms/deflation/MultiRHSBlockProject.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -66,6 +67,10 @@ public:
   std::vector<CoarseMatrix> _Adag;
   std::vector<CoarseVector> MultTemporaries;
 
+  int64_t MultCalls;
+  double  MultFlopsAccum;
+  double  MultUsecAccum;
+
   ///////////////////////
   // Interface
   ///////////////////////
@@ -99,24 +104,25 @@ public:
 	}
       }
     }
-    assert(nfound==geom.npoint);
+    GRID_ASSERT(nfound==geom.npoint);
     ExchangeCoarseLinks();
   }
   */
   
-  GeneralCoarsenedMatrix(NonLocalStencilGeometry &_geom,GridBase *FineGrid, GridCartesian * CoarseGrid)
+  GeneralCoarsenedMatrix(NonLocalStencilGeometry &_geom,GridBase *FineGrid, GridCartesian * CoarseGrid,int _herm=1)
     : geom(_geom),
       _FineGrid(FineGrid),
       _CoarseGrid(CoarseGrid),
-      hermitian(1),
+      hermitian(_herm),
       Cell(_geom.Depth(),_CoarseGrid),
-      Stencil(Cell.grids.back(),geom.shifts)
+      Stencil(Cell.grids.back(),geom.shifts),
+      MultCalls(0), MultFlopsAccum(0.0), MultUsecAccum(0.0)
   {
     {
       int npoint = _geom.npoint;
     }
     _A.resize(geom.npoint,CoarseGrid);
-    //    _Adag.resize(geom.npoint,CoarseGrid);
+    if ( !hermitian ) _Adag.resize(geom.npoint,CoarseGrid);
   }
   void M (const CoarseVector &in, CoarseVector &out)
   {
@@ -124,10 +130,10 @@ public:
   }
   void Mdag (const CoarseVector &in, CoarseVector &out)
   {
-    assert(hermitian);
-    Mult(_A,in,out);
-    //    if ( hermitian ) M(in,out);
-    //    else Mult(_Adag,in,out);
+    if(hermitian)
+      Mult(_A,in,out);
+    else
+      Mult(_Adag,in,out);
   }
   void Mult (std::vector<CoarseMatrix> &A,const CoarseVector &in, CoarseVector &out)
   {
@@ -227,29 +233,28 @@ public:
     text+=usecond();
     ttot+=usecond();
     
-    std::cout << GridLogPerformance<<"Coarse 1rhs Mult Aviews "<<tviews<<" us"<<std::endl;
-    std::cout << GridLogPerformance<<"Coarse Mult exch "<<texch<<" us"<<std::endl;
-    std::cout << GridLogPerformance<<"Coarse Mult mult "<<tmult<<" us"<<std::endl;
-    std::cout << GridLogPerformance<<" of which mult2  "<<tmult2<<" us"<<std::endl;
-    std::cout << GridLogPerformance<<"Coarse Mult ext  "<<text<<" us"<<std::endl;
-    std::cout << GridLogPerformance<<"Coarse Mult temps "<<ttemps<<" us"<<std::endl;
-    std::cout << GridLogPerformance<<"Coarse Mult copy  "<<tcopy<<" us"<<std::endl;
-    std::cout << GridLogPerformance<<"Coarse Mult tot  "<<ttot<<" us"<<std::endl;
-    //    std::cout << GridLogPerformance<<std::endl;
-    std::cout << GridLogPerformance<<"Coarse Kernel flops "<< flops<<std::endl;
-    std::cout << GridLogPerformance<<"Coarse Kernel flop/s "<< flops/tmult<<" mflop/s"<<std::endl;
-    std::cout << GridLogPerformance<<"Coarse Kernel bytes/s "<< bytes/tmult<<" MB/s"<<std::endl;
-    std::cout << GridLogPerformance<<"Coarse overall flops/s "<< flops/ttot<<" mflop/s"<<std::endl;
-    std::cout << GridLogPerformance<<"Coarse total bytes   "<< bytes/1e6<<" MB"<<std::endl;
+    MultCalls++;
+    MultFlopsAccum += flops;
+    MultUsecAccum  += ttot;
+    std::cout << GridLogPerformance
+	      << "Coarse Mult call " << MultCalls
+	      << "  tot " << ttot << " us"
+	      << "  kernel " << tmult << " us"
+	      << "  kernel " << flops/tmult*1e-3 << " GFlop/s"
+	      << "  overall " << MultFlopsAccum/MultUsecAccum*1e-3 << " GFlop/s (cumul)"
+	      << "  bw " << bytes/tmult*1e-3 << " GB/s"
+	      << std::endl;
 
   };
   
   void PopulateAdag(void)
   {
+#if 0
+    // Serial global peek/poke reference implementation
     for(int64_t bidx=0;bidx<CoarseGrid()->gSites() ;bidx++){
       Coordinate bcoor;
       CoarseGrid()->GlobalIndexToGlobalCoor(bidx,bcoor);
-      
+
       for(int p=0;p<geom.npoint;p++){
 	Coordinate scoor = bcoor;
 	for(int mu=0;mu<bcoor.size();mu++){
@@ -262,6 +267,36 @@ public:
 	pokeSite(adj(link),_Adag[pp],bcoor);
       }
     }
+#else
+    // Parallel: _Adag[pp](x) = adj( _A[p](x + s_pp) ),  pp = Reverse(p), s_pp = -s_p.
+    // The neighbour fetch reuses the same padded-cell + stencil machinery as Mult,
+    // reading one matrix element per coalesced access so no whole site matrix
+    // (230KB at nbasis=60) ever lands on a GPU thread stack (HIP limit 128KB).
+    // Halo sites compute garbage neighbours; Cell.Extract discards them.
+    // Must run on the unpadded _A, i.e. before ExchangeCoarseLinks.
+    const int Nsimd = CComplex::Nsimd();
+    for(int p=0;p<geom.npoint;p++){
+      int pp = geom.Reverse(p);
+      CoarseMatrix Apad = Cell.ExchangePeriodic(_A[p]);
+      CoarseMatrix Dpad(Apad.Grid());
+      int64_t osites = Apad.Grid()->oSites();
+      {
+	autoView( Apad_v   , Apad, AcceleratorRead);
+	autoView( Dpad_v   , Dpad, AcceleratorWriteDiscard);
+	autoView( Stencil_v, Stencil, AcceleratorRead);
+	accelerator_for(sj, osites*nbasis, Nsimd, {
+	    int32_t ss = sj/nbasis;
+	    int32_t j  = sj%nbasis;
+	    auto SE = Stencil_v.GetEntry(pp,ss);
+	    for(int i=0;i<nbasis;i++){
+	      auto z = coalescedReadGeneralPermute(Apad_v[SE->_offset](i,j),SE->_permute,Nd);
+	      coalescedWrite(Dpad_v[ss](j,i),conjugate(z));
+	    }
+	});
+      }
+      _Adag[pp] = Cell.Extract(Dpad);
+    }
+#endif
   }
   /////////////////////////////////////////////////////////////
   // 
@@ -417,10 +452,19 @@ public:
 	int osites=CoarseGrid()->oSites();
 	autoView( A_v  , _A[k], AcceleratorWrite);
 	autoView( FT_v  , FT[k], AcceleratorRead);
-	accelerator_for(sss, osites, 1, {
+	accelerator_for(sss, osites, nbasis, {
+#ifdef GRID_SIMT
+	    int j = acceleratorSIMTlane(nbasis);
+	    A_v[sss](i,j) = FT_v[sss](j);
+#else
+	    // CPU build: acceleratorSIMTlane()==0 -- an un-looped SIMT tensor
+	    // index writes ONLY j=0 and silently drops the other nbasis-1
+	    // columns (caught by Test_schur_dense_coarse import certificate,
+	    // 2026-08-14).  Loop explicitly.
 	    for(int j=0;j<nbasis;j++){
 	      A_v[sss](i,j) = FT_v[sss](j);
 	    }
+#endif
         });
       }
       tinv+=usecond();
@@ -428,8 +472,8 @@ public:
 
     // Only needed if nonhermitian
     if ( ! hermitian ) {
-      //      std::cout << GridLogMessage<<"PopulateAdag  "<<std::endl;
-      //      PopulateAdag();
+      std::cout << GridLogMessage<<"PopulateAdag  "<<std::endl;
+      PopulateAdag();
     }
 
     // Need to write something to populate Adag from A
@@ -441,8 +485,20 @@ public:
     std::cout << GridLogMessage<<"CoarsenOperator inv    "<<tinv<<" us"<<std::endl;
   }
 #else
+  //////////////////////////////////////////////////////////////////////
+  // Galerkin projection of matrix
+  //////////////////////////////////////////////////////////////////////
   void CoarsenOperator(LinearOperatorBase<Lattice<Fobj> > &linop,
 		       Aggregation<Fobj,CComplex,nbasis> & Subspace)
+  {
+    CoarsenOperator(linop,Subspace,Subspace);
+  }
+  //////////////////////////////////////////////////////////////////////
+  // Petrov - Galerkin projection of matrix
+  //////////////////////////////////////////////////////////////////////
+  void CoarsenOperator(LinearOperatorBase<Lattice<Fobj> > &linop,
+		       Aggregation<Fobj,CComplex,nbasis> & U,
+		       Aggregation<Fobj,CComplex,nbasis> & V)
   {
     std::cout << GridLogMessage<< "GeneralCoarsenMatrix "<< std::endl;
     GridBase *grid = FineGrid();
@@ -458,11 +514,9 @@ public:
     // Orthogonalise the subblocks over the basis
     /////////////////////////////////////////////////////////////
     CoarseScalar InnerProd(CoarseGrid()); 
-    blockOrthogonalise(InnerProd,Subspace.subspace);
+    blockOrthogonalise(InnerProd,V.subspace);
+    blockOrthogonalise(InnerProd,U.subspace);
 
-    //    for(int s=0;s<Subspace.subspace.size();s++){
-      //      std::cout << " subspace norm "<<norm2(Subspace.subspace[s])<<std::endl;
-    //    }
     const int npoint = geom.npoint;
       
     Coordinate clatt = CoarseGrid()->GlobalDimensions();
@@ -507,13 +561,9 @@ public:
     // Now compute the matrix elements of linop between the orthonormal
     // set of vectors.
     ///////////////////////////////////////////////////////////////////////
-    FineField phaV(grid); // Phased block basis vector
-    FineField MphaV(grid);// Matrix applied
     std::vector<FineComplexField> phaF(npoint,grid);
     std::vector<CoarseComplexField> pha(npoint,CoarseGrid());
-    
-    CoarseVector coarseInner(CoarseGrid());
-    
+
     typedef typename CComplex::scalar_type SComplex;
     FineComplexField one(grid); one=SComplex(1.0);
     FineComplexField zz(grid); zz = Zero();
@@ -532,37 +582,52 @@ public:
       pha[p]  =exp(pha[p]*ci);
 
       blockZAXPY(phaF[p],pha[p],one,zz);
-      
+
     }
     tphase+=usecond();
-    
-    std::vector<CoarseVector> ComputeProj(npoint,CoarseGrid());
-    std::vector<CoarseVector>          FT(npoint,CoarseGrid());
+
+    // Import basis into BLAS layout once; blockProject then reads it once per
+    // basis vector rather than once per (i,p) as in scalar blockProject.
+    // Process all npoint in a single batch.
+    MultiRHSBlockProject<FineField> Projector;
+    Projector.Allocate(nbasis, grid, CoarseGrid());
+    Projector.ImportBasis(U.subspace);
+
+    std::vector<FineField>    phaV_batch(npoint, grid);
+    std::vector<FineField>    MphaV_batch(npoint, grid);
+    std::vector<CoarseVector> proj_batch(npoint, CoarseGrid());
+    std::vector<CoarseVector> ComputeProj(npoint, CoarseGrid());
+    std::vector<CoarseVector>          FT(npoint, CoarseGrid());
+
+    // Pre-allocate BLAS_F and BLAS_C to avoid repeated hipMalloc/hipFree of
+    // ~5.6 GB per blockProject call, which hangs on ROCm for large allocations.
+    Projector.BLAS_F.resize(Projector.fine_vol * Projector.words * npoint);
+    Projector.BLAS_C.resize(Projector.coarse_vol * nbasis * npoint);
+
     for(int i=0;i<nbasis;i++){// Loop over basis vectors
+      accelerator_barrier(); // ensure prior iteration's async writes are retired
       std::cout << GridLogMessage<< "CoarsenMatrixColoured vec "<<i<<"/"<<nbasis<< std::endl;
-      for(int p=0;p<npoint;p++){ // Loop over momenta in npoint
-	tphaseBZ-=usecond();
-	phaV = phaF[p]*Subspace.subspace[i];
-	tphaseBZ+=usecond();
 
-	/////////////////////////////////////////////////////////////////////
-	// Multiple phased subspace vector by matrix and project to subspace
-	// Remove local bulk phase to leave relative phases
-	/////////////////////////////////////////////////////////////////////
-	tmat-=usecond();
-	linop.Op(phaV,MphaV);
-	tmat+=usecond();
-	//	std::cout << i << " " <<p << " MphaV "<<norm2(MphaV)<<" "<<norm2(phaV)<<std::endl;
+      tphaseBZ-=usecond();
+      for(int p=0;p<npoint;p++)
+	phaV_batch[p] = phaF[p] * V.subspace[i];
+      tphaseBZ+=usecond();
+      std::cout << GridLogMessage<< "CoarsenMatrixColoured vec "<<i<<" phaseBZ done"<< std::endl;
 
-	tproj-=usecond();
-	blockProject(coarseInner,MphaV,Subspace.subspace);
-	coarseInner = conjugate(pha[p]) * coarseInner;
+      tmat-=usecond();
+      for(int p=0;p<npoint;p++)
+	linop.Op(phaV_batch[p], MphaV_batch[p]);
+      tmat+=usecond();
+      std::cout << GridLogMessage<< "CoarsenMatrixColoured vec "<<i<<" mat done"<< std::endl;
 
-	ComputeProj[p] = coarseInner;
-	tproj+=usecond();
-	//	std::cout << i << " " <<p << " ComputeProj "<<norm2(ComputeProj[p])<<std::endl;
-
-      }
+      // One batched GEMM reads BLAS_V once for all npoint vectors.
+      tproj-=usecond();
+      Projector.blockProject(MphaV_batch, proj_batch);
+      std::cout << GridLogMessage<< "CoarsenMatrixColoured vec "<<i<<" blockProject done"<< std::endl;
+      for(int p=0;p<npoint;p++)
+	ComputeProj[p] = conjugate(pha[p]) * proj_batch[p];
+      tproj+=usecond();
+      std::cout << GridLogMessage<< "CoarsenMatrixColoured vec "<<i<<" proj done"<< std::endl;
 
       tinv-=usecond();
       for(int k=0;k<npoint;k++){
@@ -570,14 +635,23 @@ public:
 	for(int l=0;l<npoint;l++){
 	  FT[k]= FT[k]+ invMkl(l,k)*ComputeProj[l];
 	}
-      
+
 	int osites=CoarseGrid()->oSites();
 	autoView( A_v  , _A[k], AcceleratorWrite);
 	autoView( FT_v  , FT[k], AcceleratorRead);
-	accelerator_for(sss, osites, 1, {
+	accelerator_for(sss, osites, nbasis, {
+#ifdef GRID_SIMT
+	    int j = acceleratorSIMTlane(nbasis);
+	    A_v[sss](i,j) = FT_v[sss](j);
+#else
+	    // CPU build: acceleratorSIMTlane()==0 -- an un-looped SIMT tensor
+	    // index writes ONLY j=0 and silently drops the other nbasis-1
+	    // columns (caught by Test_schur_dense_coarse import certificate,
+	    // 2026-08-14).  Loop explicitly.
 	    for(int j=0;j<nbasis;j++){
 	      A_v[sss](i,j) = FT_v[sss](j);
 	    }
+#endif
         });
       }
       tinv+=usecond();
@@ -585,13 +659,13 @@ public:
 
     // Only needed if nonhermitian
     if ( ! hermitian ) {
-      //      std::cout << GridLogMessage<<"PopulateAdag  "<<std::endl;
-      //      PopulateAdag();
+      std::cout << GridLogMessage<<"PopulateAdag  "<<std::endl;
+      PopulateAdag();
     }
 
-    for(int p=0;p<geom.npoint;p++){
-      std::cout << " _A["<<p<<"] "<<norm2(_A[p])<<std::endl;
-    }
+    //    for(int p=0;p<geom.npoint;p++){
+    //      std::cout << " _A["<<p<<"] "<<norm2(_A[p])<<std::endl;
+    //    }
 
     // Need to write something to populate Adag from A
     ExchangeCoarseLinks();
@@ -606,10 +680,10 @@ public:
   void ExchangeCoarseLinks(void){
     for(int p=0;p<geom.npoint;p++){
       _A[p] = Cell.ExchangePeriodic(_A[p]);
-      //      _Adag[p]= Cell.ExchangePeriodic(_Adag[p]);
+      if ( !hermitian ) _Adag[p]= Cell.ExchangePeriodic(_Adag[p]);
     }
   }
-  virtual  void Mdiag    (const Field &in, Field &out){ assert(0);};
+  virtual  void Mdiag    (const Field &in, Field &out){ GRID_ASSERT(0);};
   virtual  void Mdir     (const Field &in, Field &out,int dir, int disp){assert(0);};
   virtual  void MdirAll  (const Field &in, std::vector<Field> &out){assert(0);};
 };

Grid/algorithms/multigrid/GeneralCoarsenedMatrixMultiRHS.h

@@ -80,12 +80,12 @@ public:
   // Can be used to do I/O on the operator matrices externally
   void SetMatrix (int p,CoarseMatrix & A)
   {
-    assert(A.size()==geom_srhs.npoint);
+    GRID_ASSERT(A.size()==geom_srhs.npoint);
     GridtoBLAS(A[p],BLAS_A[p]);
   }
   void GetMatrix (int p,CoarseMatrix & A)
   {
-    assert(A.size()==geom_srhs.npoint);
+    GRID_ASSERT(A.size()==geom_srhs.npoint);
     BLAStoGrid(A[p],BLAS_A[p]);
   }
   void CopyMatrix (GeneralCoarseOp &_Op)
@@ -178,14 +178,14 @@ public:
 	for(int32_t point = 0 ; point < geom.npoint; point++){
 	  int i=s*orhs*geom.npoint+point;
  	  int32_t nbr = Stencil._entries[i]._offset*CComplex::Nsimd(); // oSite -> lSite
-	  assert(nbr<BLAS_B.size());
+	  GRID_ASSERT(nbr<BLAS_B.size());
 	  ComplexD * ptr = (ComplexD *)&BLAS_B[nbr];
 	  acceleratorPut(BLAS_BP[point][j],ptr); // neighbour indexing in ghost zone volume
 	}
 	j++;
       }
     }
-    assert(j==unpadded_sites);
+    GRID_ASSERT(j==unpadded_sites);
   }
   template<class vobj> void GridtoBLAS(const Lattice<vobj> &from,deviceVector<typename vobj::scalar_object> &to)
   {
@@ -194,7 +194,7 @@ public:
   typedef typename vobj::vector_type vector_type;
 
   GridBase *Fg = from.Grid();
-  assert(!Fg->_isCheckerBoarded);
+  GRID_ASSERT(!Fg->_isCheckerBoarded);
   int nd = Fg->_ndimension;
 
   to.resize(Fg->lSites());
@@ -241,10 +241,10 @@ public:
   typedef typename vobj::vector_type vector_type;
 
   GridBase *Tg = grid.Grid();
-  assert(!Tg->_isCheckerBoarded);
+  GRID_ASSERT(!Tg->_isCheckerBoarded);
   int nd = Tg->_ndimension;
   
-  assert(in.size()==Tg->lSites());
+  GRID_ASSERT(in.size()==Tg->lSites());
 
   Coordinate LocalLatt = Tg->LocalDimensions();
   size_t nsite = 1;
@@ -669,7 +669,7 @@ Grid : Message : 328.193436 s : CoarsenOperator mat    122213270 us
     const int Nsimd = CComplex::Nsimd();
 
     int64_t nrhs  =pin.Grid()->GlobalDimensions()[0];
-    assert(nrhs>=1);
+    GRID_ASSERT(nrhs>=1);
 
     RealD flops,bytes;
     int64_t osites=in.Grid()->oSites(); // unpadded
@@ -721,7 +721,7 @@ Grid : Message : 328.193436 s : CoarsenOperator mat    122213270 us
     //    std::cout << GridLogMessage<<"Coarse overall flops/s "<< flops/t_tot<<" mflop/s"<<std::endl;
     //    std::cout << GridLogMessage<<"Coarse total bytes   "<< bytes/1e6<<" MB"<<std::endl;
   };
-  virtual  void Mdiag    (const Field &in, Field &out){ assert(0);};
+  virtual  void Mdiag    (const Field &in, Field &out){ GRID_ASSERT(0);};
   virtual  void Mdir     (const Field &in, Field &out,int dir, int disp){assert(0);};
   virtual  void MdirAll  (const Field &in, std::vector<Field> &out){assert(0);};
 };

Grid/algorithms/multigrid/GeneralCoarsenedMatrixMultiRHSV2.h

@@ -0,0 +1,851 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/GeneralCoarsenedMatrixMultiRHS.h
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <pboyle@bnl.gov>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#pragma once
+
+
+NAMESPACE_BEGIN(Grid);
+
+
+// Fine Object == (per site) type of fine field
+// nbasis      == number of deflation vectors
+template<class Fobj,class CComplex,int nbasis>
+class MultiGeneralCoarsenedOperatorV2 : public SparseMatrixBase<Lattice<iVector<CComplex,nbasis > > >  {
+public:
+  typedef typename CComplex::scalar_object SComplex;
+  typedef GeneralCoarsenedMatrix<Fobj,CComplex,nbasis> GeneralCoarseOp;
+  typedef MultiGeneralCoarsenedOperatorV2<Fobj,CComplex,nbasis> MultiGeneralCoarseOp;
+
+  typedef iVector<CComplex,nbasis >           siteVector;
+  typedef iMatrix<CComplex,nbasis >           siteMatrix;
+  typedef iVector<SComplex,nbasis >           calcVector;
+  typedef iMatrix<SComplex,nbasis >           calcMatrix;
+  typedef Lattice<iScalar<CComplex> >         CoarseComplexField;
+  typedef Lattice<siteVector>                 CoarseVector;
+  typedef Lattice<iMatrix<CComplex,nbasis > > CoarseMatrix;
+  typedef iMatrix<CComplex,nbasis >  Cobj;
+  typedef iVector<CComplex,nbasis >  Cvec;
+  typedef Lattice< CComplex >   CoarseScalar; // used for inner products on fine field
+  typedef Lattice<Fobj >        FineField;
+  typedef CoarseVector Field;
+
+  // Block operations on the fine vectors carry the fine layout, which need
+  // not be the coarse one
+  typedef decltype(innerProduct(Fobj(),Fobj())) FineInner;
+  typedef Lattice<FineInner>                    FineComplexField;
+  typedef Lattice<FineInner>                    BlockComplexField;
+
+  ////////////////////
+  // Data members
+  //
+  // Nrhs independent: the D dimensional coarse grid, the geometry, the padded
+  // cell that supplies the stencil grid, the stencil, and the matrix elements.
+  //
+  // Nrhs dependent: the D+1 grid, its padded cell, and the BLAS B/C buffers
+  // with their pointer tables. Owned by SetNRHS().
+  ////////////////////
+  GridCartesian *       _CoarseGrid;        // D dimensional
+  NonLocalStencilGeometry geom;
+  NonLocalStencilGeometry geom_srhs;
+  PaddedCell CellD;                         // D dimensional, supplies stencil grid
+  GeneralLocalStencil Stencil;              // D dimensional
+
+  int                   _Nrhs;
+  GridCartesian *       _CoarseGridMulti;   // D+1 dimensional, SetNRHS
+  PaddedCell *          CellMulti;          // D+1 dimensional, SetNRHS
+
+  deviceVector<calcVector> BLAS_B;
+  deviceVector<calcVector> BLAS_C;
+  std::vector<deviceVector<calcMatrix> > BLAS_A;
+
+  std::vector<deviceVector<ComplexD *> > BLAS_AP;
+  std::vector<deviceVector<ComplexD *> > BLAS_BP;
+  deviceVector<ComplexD *>               BLAS_CP;
+
+  ///////////////////////
+  // Interface
+  ///////////////////////
+  GridBase      * Grid(void)           { CheckGridSet(); return _CoarseGridMulti; };
+  GridCartesian * CoarseGrid(void)     { CheckGridSet(); return _CoarseGridMulti; };
+  GridCartesian * CoarseGridD(void)    { return _CoarseGrid; };        // lower dimensional grid
+  int             Nrhs(void)           { CheckGridSet(); return _Nrhs; };
+
+  void CheckGridSet(void)
+  {
+    if ( _CoarseGridMulti == nullptr ) {
+      std::cout << GridLogError
+		<< "MultiGeneralCoarsenedOperatorV2: the multiRHS grid has not been set."
+		<< std::endl;
+      std::cout << GridLogError
+		<< "  Call SetGrid(CoarseGridMulti) with the D+1 dimensional grid your"
+		<< std::endl;
+      std::cout << GridLogError
+		<< "  coarse vectors live on, before Grid(), Nrhs() or M()."
+		<< std::endl;
+      GRID_ASSERT(_CoarseGridMulti != nullptr);
+    }
+  }
+
+  // Can be used to do I/O on the operator matrices externally
+  void SetMatrix (int p,CoarseMatrix & A)
+  {
+    GRID_ASSERT(A.size()==geom_srhs.npoint);
+    GridtoBLAS(A[p],BLAS_A[p]);
+  }
+  void GetMatrix (int p,CoarseMatrix & A)
+  {
+    GRID_ASSERT(A.size()==geom_srhs.npoint);
+    BLAStoGrid(A[p],BLAS_A[p]);
+  }
+  void CopyMatrix (GeneralCoarseOp &_Op)
+  {
+    for(int p=0;p<geom.npoint;p++){
+      auto Aup = _Op.Cell.Extract(_Op._A[p]);
+      //Unpadded
+      GridtoBLAS(Aup,BLAS_A[p]);
+    }
+  }
+  /*
+  void CheckMatrix (GeneralCoarseOp &_Op)
+  {
+    std::cout <<"************* Checking the little direc operator mRHS"<<std::endl;
+    for(int p=0;p<geom.npoint;p++){
+      //Unpadded
+      auto Aup = _Op.Cell.Extract(_Op._A[p]);
+      auto Ack = Aup;
+      BLAStoGrid(Ack,BLAS_A[p]);
+      std::cout << p<<" Ack "<<norm2(Ack)<<std::endl;
+      std::cout << p<<" Aup "<<norm2(Aup)<<std::endl;
+    }
+    std::cout <<"************* "<<std::endl;
+  }
+  */
+  
+  ///////////////////////////////////////////////////////////////////////////
+  // Constructor takes the D dimensional coarse grid. Everything built here
+  // is independent of Nrhs, in particular the matrix elements, which must
+  // survive a change of Nrhs untouched.
+  ///////////////////////////////////////////////////////////////////////////
+  MultiGeneralCoarsenedOperatorV2(NonLocalStencilGeometry &_geom,GridCartesian *CoarseGrid) :
+    _CoarseGrid(CoarseGrid),
+    geom_srhs(_geom),
+    geom(CoarseGrid,_geom.hops,_geom.skip),
+    CellD(geom.Depth(),CoarseGrid),
+    Stencil(CellD.grids.back(),geom.shifts), // D dimensional padded cell stencil
+    _Nrhs(-1),
+    _CoarseGridMulti(nullptr),
+    CellMulti(nullptr)
+  {
+    int32_t unpadded_sites = _CoarseGrid->lSites();
+
+    /////////////////////////////////////////////////
+    // Matrix elements and their pointer table
+    /////////////////////////////////////////////////
+    BLAS_A.resize(geom.npoint);
+    BLAS_AP.resize(geom.npoint);
+    for(int p=0;p<geom.npoint;p++){
+      BLAS_A[p].resize (unpadded_sites); // no ghost zone, npoint elements
+      BLAS_AP[p].resize(unpadded_sites);
+    }
+
+    // Site identity mapping for A
+    for(int p=0;p<geom.npoint;p++){
+      for(int ss=0;ss<unpadded_sites;ss++){
+	ComplexD *ptr = (ComplexD *)&BLAS_A[p][ss];
+	acceleratorPut(BLAS_AP[p][ss],ptr);
+      }
+    }
+  }
+
+  virtual ~MultiGeneralCoarsenedOperatorV2()
+  {
+    ReleaseGrid();
+  }
+
+  ///////////////////////////////////////////////////////////////////////////
+  // Free everything SetGrid allocated. The D+1 grid is borrowed from the
+  // caller and is never deleted here. Safe to call repeatedly and before
+  // the destructor.
+  ///////////////////////////////////////////////////////////////////////////
+  void ReleaseGrid(void)
+  {
+    if ( CellMulti != nullptr ) { delete CellMulti; CellMulti = nullptr; }
+
+    _CoarseGridMulti = nullptr;  // borrowed, not owned
+    _Nrhs            = -1;
+
+    BLAS_B.resize(0);
+    BLAS_C.resize(0);
+    for(int p=0;p<BLAS_BP.size();p++){
+      BLAS_BP[p].resize(0);
+    }
+    BLAS_BP.resize(0);
+    BLAS_CP.resize(0);
+  }
+
+  ///////////////////////////////////////////////////////////////////////////
+  // Everything that depends on Nrhs. Idempotent; lazy called on demand.
+  //
+  // The stencil lives on the D dimensional padded grid. PaddedCell only pads
+  // a dimension when it is distributed, and the rhs direction never is, so
+  // the D+1 padded grid is exactly Nrhs copies of the D dimensional padded
+  // grid with rhs innermost. The neighbour offset therefore carries an Nrhs
+  // factor, in the same way the Nsimd factor is carried.
+  ///////////////////////////////////////////////////////////////////////////
+  void SetGrid(GridCartesian *CoarseGridMulti)
+  {
+    GRID_ASSERT(CoarseGridMulti != nullptr);
+
+    if ( CoarseGridMulti == _CoarseGridMulti ) return; // idempotent on identity
+
+    ReleaseGrid();
+
+    /////////////////////////////////////////////////
+    // The D+1 grid is supplied and owned by the caller. Two operators over
+    // the same coarse space must share one grid object or their fields will
+    // not conform, so this is never manufactured internally.
+    /////////////////////////////////////////////////
+    int nd = _CoarseGrid->_ndimension;
+
+    GRID_ASSERT(CoarseGridMulti->_ndimension == nd+1);
+    GRID_ASSERT(CoarseGridMulti->_processors[0] == 1);   // rhs is not distributed
+    for(int d=0;d<nd;d++){
+      GRID_ASSERT(CoarseGridMulti->_fdimensions[d+1] == _CoarseGrid->_fdimensions[d]);
+      GRID_ASSERT(CoarseGridMulti->_processors [d+1] == _CoarseGrid->_processors [d]);
+      GRID_ASSERT(CoarseGridMulti->_simd_layout[d+1] == _CoarseGrid->_simd_layout[d]);
+    }
+
+    _CoarseGridMulti = CoarseGridMulti;
+    _Nrhs            = CoarseGridMulti->_fdimensions[0];
+    GRID_ASSERT(_Nrhs>=1);
+
+    int nrhs = _Nrhs;
+
+    CellMulti = new PaddedCell(geom.Depth(),_CoarseGridMulti);
+
+    int32_t padded_sites   = CellD.grids.back()->lSites(); // D dimensional
+    int32_t unpadded_sites = _CoarseGrid->lSites();        // D dimensional
+
+    // The neighbour offset multiplication by nrhs is exact only if the D+1
+    // padded volume is nrhs copies of the D dimensional one. Check it.
+    GRID_ASSERT(CellMulti->grids.back()->lSites() == nrhs*padded_sites);
+    GRID_ASSERT(_CoarseGridMulti->lSites()        == nrhs*unpadded_sites);
+
+    /////////////////////////////////////////////////
+    // Device data vector storage
+    /////////////////////////////////////////////////
+    BLAS_B.resize(nrhs *padded_sites);   // includes ghost zone
+    BLAS_C.resize(nrhs *unpadded_sites); // no ghost zone
+    BLAS_BP.resize(geom.npoint);
+    for(int p=0;p<geom.npoint;p++){
+      BLAS_BP[p].resize(unpadded_sites);
+    }
+    BLAS_CP.resize(unpadded_sites);
+
+    // Site identity mapping for C
+    for(int ss=0;ss<unpadded_sites;ss++){
+      ComplexD *ptr = (ComplexD *)&BLAS_C[ss*nrhs];
+      acceleratorPut(BLAS_CP[ss],ptr);
+    }
+
+    // Neighbour table is more complicated
+    int32_t j=0; // Interior point counter (unpadded)
+    for(int32_t s=0;s<padded_sites;s++){ // D volume, padded
+      int ghost_zone=0;
+      for(int32_t point = 0 ; point < geom.npoint; point++){
+	int i=s*geom.npoint+point;
+	if( Stencil._entries[i]._wrap ) { // stencil is indexed by the oSite of the D dim grid
+	  ghost_zone=1; // If general stencil wrapped in any direction, wrap=1
+	}
+      }
+
+      if( ghost_zone==0) {
+	for(int32_t point = 0 ; point < geom.npoint; point++){
+	  int i=s*geom.npoint+point;
+ 	  int32_t nbr = Stencil._entries[i]._offset*CComplex::Nsimd(); // oSite -> lSite, D dim
+	  nbr = nbr*nrhs;                                              // D -> D+1, rhs innermost
+	  GRID_ASSERT(nbr<BLAS_B.size());
+	  ComplexD * ptr = (ComplexD *)&BLAS_B[nbr];
+	  acceleratorPut(BLAS_BP[point][j],ptr); // neighbour indexing in ghost zone volume
+	}
+	j++;
+      }
+    }
+    GRID_ASSERT(j==unpadded_sites);
+  }
+  template<class vobj> void GridtoBLAS(const Lattice<vobj> &from,deviceVector<typename vobj::scalar_object> &to)
+  {
+  typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::vector_type vector_type;
+
+  GridBase *Fg = from.Grid();
+  GRID_ASSERT(!Fg->_isCheckerBoarded);
+  int nd = Fg->_ndimension;
+
+  to.resize(Fg->lSites());
+
+  Coordinate LocalLatt = Fg->LocalDimensions();
+  size_t nsite = 1;
+  for(int i=0;i<nd;i++) nsite *= LocalLatt[i];
+
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  // do the index calc on the GPU
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  Coordinate f_ostride = Fg->_ostride;
+  Coordinate f_istride = Fg->_istride;
+  Coordinate f_rdimensions = Fg->_rdimensions;
+
+  autoView(from_v,from,AcceleratorRead);
+  auto to_v = &to[0];
+
+  const int words=sizeof(vobj)/sizeof(vector_type);
+  accelerator_for(idx,nsite,1,{
+      
+      Coordinate from_coor, base;
+      Lexicographic::CoorFromIndex(base,idx,LocalLatt);
+      for(int i=0;i<nd;i++){
+	from_coor[i] = base[i];
+      }
+      int from_oidx = 0; for(int d=0;d<nd;d++) from_oidx+=f_ostride[d]*(from_coor[d]%f_rdimensions[d]);
+      int from_lane = 0; for(int d=0;d<nd;d++) from_lane+=f_istride[d]*(from_coor[d]/f_rdimensions[d]);
+
+      const vector_type* from = (const vector_type *)&from_v[from_oidx];
+      scalar_type* to = (scalar_type *)&to_v[idx];
+      
+      scalar_type stmp;
+      for(int w=0;w<words;w++){
+	stmp = getlane(from[w], from_lane);
+	to[w] = stmp;
+      }
+    });
+  }    
+  template<class vobj> void BLAStoGrid(Lattice<vobj> &grid,deviceVector<typename vobj::scalar_object> &in)
+  {
+  typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::vector_type vector_type;
+
+  GridBase *Tg = grid.Grid();
+  GRID_ASSERT(!Tg->_isCheckerBoarded);
+  int nd = Tg->_ndimension;
+  
+  GRID_ASSERT(in.size()==Tg->lSites());
+
+  Coordinate LocalLatt = Tg->LocalDimensions();
+  size_t nsite = 1;
+  for(int i=0;i<nd;i++) nsite *= LocalLatt[i];
+
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  // do the index calc on the GPU
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  Coordinate t_ostride = Tg->_ostride;
+  Coordinate t_istride = Tg->_istride;
+  Coordinate t_rdimensions = Tg->_rdimensions;
+
+  autoView(to_v,grid,AcceleratorWrite);
+  auto from_v = &in[0];
+
+  const int words=sizeof(vobj)/sizeof(vector_type);
+  accelerator_for(idx,nsite,1,{
+      
+      Coordinate to_coor, base;
+      Lexicographic::CoorFromIndex(base,idx,LocalLatt);
+      for(int i=0;i<nd;i++){
+	to_coor[i] = base[i];
+      }
+      int to_oidx = 0; for(int d=0;d<nd;d++) to_oidx+=t_ostride[d]*(to_coor[d]%t_rdimensions[d]);
+      int to_lane = 0; for(int d=0;d<nd;d++) to_lane+=t_istride[d]*(to_coor[d]/t_rdimensions[d]);
+
+      vector_type* to = (vector_type *)&to_v[to_oidx];
+      scalar_type* from = (scalar_type *)&from_v[idx];
+      
+      scalar_type stmp;
+      for(int w=0;w<words;w++){
+	stmp=from[w];
+	putlane(to[w], stmp, to_lane);
+      }
+    });
+  }
+  ///////////////////////////////////////////////////////////////////////////
+  // Shared by both CoarsenOperator variants
+  //
+  //     conj(pha[block]) proj[k (which mom)][j (basis vec cpt)][block]
+  //       =  \sum_{l in ball}  e^{i q_k . delta_l} < phi_{block,j} | MdagM | phi_{(block+delta_l),i} >
+  //       =  \sum_{l in ball} e^{iqk.delta_l} A_ji^{b.b+l}
+  //       = M_{kl} A_ji^{b.b+l}
+  //
+  //     Where q_k = delta_k . (2*M_PI/global_nb[mu])
+  //     Then A{ji}^{b,b+l} = M^{-1}_{lm} ComputeProj_{m,b,i,j}
+  ///////////////////////////////////////////////////////////////////////////
+  void CoarsenFourierMatrix(GridBase *CoarseGrid,Eigen::MatrixXcd &invMkl)
+  {
+    const int npoint = geom_srhs.npoint;
+    Coordinate clatt = CoarseGrid->GlobalDimensions();
+    int Nd = CoarseGrid->Nd();
+
+    Eigen::MatrixXcd Mkl = Eigen::MatrixXcd::Zero(npoint,npoint);
+    ComplexD ci(0.0,1.0);
+    for(int k=0;k<npoint;k++){ // Loop over momenta
+      for(int l=0;l<npoint;l++){ // Loop over nbr relative
+	ComplexD phase(0.0,0.0);
+	for(int mu=0;mu<Nd;mu++){
+	  RealD TwoPiL =  M_PI * 2.0/ clatt[mu];
+	  phase=phase+TwoPiL*geom_srhs.shifts[k][mu]*geom_srhs.shifts[l][mu];
+	}
+	phase=exp(phase*ci);
+	Mkl(k,l) = phase;
+      }
+    }
+    invMkl = Mkl.inverse();
+  }
+
+  ///////////////////////////////////////////////////////////////////////////
+  // blockOrthogonalise and blockZAXPY are block operations on the fine
+  // vectors, using a coarse shaped field only as an index set. They need a
+  // grid carrying the fine SIMD layout, which the coarse space no longer
+  // does. Constructed local to the caller so it cannot be mistaken for the
+  // coarse grid.
+  ///////////////////////////////////////////////////////////////////////////
+  void CoarsenBlockGridLayout(GridBase *grid,GridBase *CoarseGrid,
+			      Coordinate &latt,Coordinate &simd,Coordinate &mpi)
+  {
+    int nd = CoarseGrid->_ndimension;
+    latt.resize(nd); simd.resize(nd); mpi.resize(nd);
+    for(int d=0;d<nd;d++){
+      latt[d] = CoarseGrid->_fdimensions[d];
+      simd[d] = grid->_simd_layout[d];
+      mpi [d] = CoarseGrid->_processors[d];
+    }
+  }
+
+  // D+1 coarse grid holding the batch, rhs innermost and unvectorised
+  void CoarsenBatchGridLayout(GridBase *CoarseGrid,int batch,
+			      Coordinate &latt,Coordinate &simd,Coordinate &mpi)
+  {
+    latt.resize(1,batch); simd.resize(1,1); mpi.resize(1,1);
+    latt[0]=batch; simd[0]=1; mpi[0]=1;
+    for(int d=0;d<CoarseGrid->_ndimension;d++){
+      latt.push_back(CoarseGrid->_fdimensions[d]);
+      simd.push_back(CoarseGrid->_simd_layout[d]);
+      mpi .push_back(CoarseGrid->_processors[d]);
+    }
+  }
+
+  ///////////////////////////////////////////////////////////////////////////
+  // The Fourier inverse needs the phase in the coarse layout and the basis
+  // phasing needs it in the fine layout; each is built from its own
+  // coordinates rather than transferred.
+  ///////////////////////////////////////////////////////////////////////////
+  void CoarsenPhases(GridBase *grid,GridBase *CoarseGrid,GridCartesian *BlockGrid,
+		     std::vector<CoarseComplexField> &pha,
+		     std::vector<FineComplexField> &phaF)
+  {
+    const int npoint = geom_srhs.npoint;
+    Coordinate clatt = CoarseGrid->GlobalDimensions();
+    int Nd = CoarseGrid->Nd();
+    ComplexD ci(0.0,1.0);
+
+    typedef typename CComplex::scalar_type SComplex;
+    FineComplexField one(grid); one=SComplex(1.0);
+    FineComplexField zz(grid);  zz = Zero();
+    BlockComplexField pha_blk (BlockGrid);
+    BlockComplexField blk_coor(BlockGrid);
+
+    for(int p=0;p<npoint;p++){ // Loop over momenta in npoint
+      CoarseComplexField coor(CoarseGrid);
+      pha[p] =Zero();
+      pha_blk=Zero();
+      for(int mu=0;mu<Nd;mu++){
+	RealD TwoPiL =  M_PI * 2.0/ clatt[mu];
+	LatticeCoordinate(coor,mu);
+	pha[p]  = pha[p]  + (TwoPiL * geom_srhs.shifts[p][mu]) * coor;
+	LatticeCoordinate(blk_coor,mu);
+	pha_blk = pha_blk + (TwoPiL * geom_srhs.shifts[p][mu]) * blk_coor;
+      }
+      pha[p] =exp(pha[p] *ci);
+      pha_blk=exp(pha_blk*ci);
+
+      blockZAXPY(phaF[p],pha_blk,one,zz);
+    }
+  }
+
+  ///////////////////////////////////////////////////////////////////////////
+  // Remove the bulk phase from the batch of coarse projections and
+  // accumulate the Fourier inverse into A. Both variants reach here with
+  // TmpProj in the same batch coarse order, so this is shared verbatim.
+  ///////////////////////////////////////////////////////////////////////////
+  void CoarsenAccumulate(int p,int i0,int nbv,int batch,
+			 Eigen::MatrixXcd &invMkl,
+			 std::vector<CoarseComplexField> &pha,
+			 CoarseComplexField &phaB,
+			 CoarseVector &TmpProj,
+			 std::vector<CoarseMatrix> &_A,
+			 GridBase *CoarseGrid)
+  {
+    typedef typename CComplex::scalar_type SComplex;
+    const int npoint = geom_srhs.npoint;
+
+    for(int b=0;b<batch;b++) InsertSliceFast(pha[p],phaB,b,0);
+    TmpProj = conjugate(phaB)*TmpProj;
+
+    int osites=CoarseGrid->oSites();
+    for(int k=0;k<npoint;k++){
+      SComplex sc(invMkl(p,k).real(),invMkl(p,k).imag());
+      CComplex coef(sc);
+      autoView( A_v  , _A[k], AcceleratorWrite);
+      autoView( TP_v , TmpProj, AcceleratorRead);
+      accelerator_for(sss, osites, 1, {
+	  for(int b=0;b<nbv;b++){
+	    for(int j=0;j<nbasis;j++){
+	      A_v[sss](i0+b,j) = A_v[sss](i0+b,j) + coef*TP_v[b+batch*sss](j);
+	    }
+	  }
+      });
+    }
+  }
+
+  void CoarsenReport(RealD tphase,RealD tphaseBZ,RealD tslice,
+		     RealD tmat,RealD tproj,RealD tinv)
+  {
+    std::cout << GridLogMessage<<"CoarsenOperator phase  "<<tphase<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator phaseBZ "<<tphaseBZ<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator slice  "<<tslice <<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator mat    "<<tmat <<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator proj   "<<tproj<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator inv    "<<tinv<<" us"<<std::endl;
+  }
+
+  ///////////////////////////////////////////////////////////////////////////
+  // Coarsen a NATIVELY multiRHS fine operator.
+  //
+  // linop acts on the D+1 dimensional fine grid FineGridMulti, with the batch
+  // of phased basis vectors carried in the rhs direction. A single RHS
+  // operator can be promoted with MrhsPromotedOperator, but that pays an
+  // ExtractSlice/InsertSlice pair per rhs; prefer the single RHS variant
+  // below in that case.
+  ///////////////////////////////////////////////////////////////////////////
+  void CoarsenOperator(LinearOperatorBase<Lattice<Fobj> > &linop,
+		       GridCartesian *FineGridMulti,
+		       std::vector<FineField> &Subspace,
+		       GridBase *CoarseGrid)
+  {
+    RealD tproj=0.0, tmat=0.0, tphase=0.0, tphaseBZ=0.0, tslice=0.0, tinv=0.0;
+
+    std::cout << GridLogMessage<< "GeneralCoarsenMatrixMrhs (multiRHS fine operator)"<< std::endl;
+
+    GRID_ASSERT(Subspace.size()==nbasis);
+    GridBase *grid = Subspace[0].Grid();
+
+    GRID_ASSERT(FineGridMulti->_ndimension == grid->_ndimension+1);
+    GRID_ASSERT(FineGridMulti->_processors[0] == 1);
+    for(int d=0;d<grid->_ndimension;d++){
+      GRID_ASSERT(FineGridMulti->_fdimensions[d+1] == grid->_fdimensions[d]);
+      GRID_ASSERT(FineGridMulti->_processors [d+1] == grid->_processors [d]);
+    }
+    int batch = FineGridMulti->_fdimensions[0];
+
+    Coordinate blatt,bsimd,bmpi;
+    CoarsenBlockGridLayout(grid,CoarseGrid,blatt,bsimd,bmpi);
+    GridCartesian BlockGrid(blatt,bsimd,bmpi);
+
+    BlockComplexField InnerProd(&BlockGrid);
+    blockOrthogonalise(InnerProd,Subspace);
+
+    MultiRHSBlockProject<Lattice<Fobj> >    Projector;
+    Projector.Allocate(nbasis,grid,CoarseGrid);
+    Projector.ImportBasis(Subspace);
+
+    const int npoint = geom_srhs.npoint;
+
+    Eigen::MatrixXcd invMkl;
+    CoarsenFourierMatrix(CoarseGrid,invMkl);
+
+    FineField phaV(grid);
+    std::vector<FineComplexField>   phaF(npoint,grid);
+    std::vector<CoarseComplexField> pha (npoint,CoarseGrid);
+
+    tphase=-usecond();
+    CoarsenPhases(grid,CoarseGrid,&BlockGrid,pha,phaF);
+    tphase+=usecond();
+
+    std::vector<CoarseMatrix> _A;
+    _A.resize(npoint,CoarseGrid);
+    for(int k=0;k<npoint;k++) _A[k] = Zero();
+
+    Coordinate cmlatt,cmsimd,cmmpi;
+    CoarsenBatchGridLayout(CoarseGrid,batch,cmlatt,cmsimd,cmmpi);
+    GridCartesian CoarseBatchGrid(cmlatt,cmsimd,cmmpi);
+
+    CoarseVector       TmpProj(&CoarseBatchGrid);
+    CoarseComplexField phaB(&CoarseBatchGrid);
+
+    FineField hi_in (FineGridMulti);
+    FineField hi_out(FineGridMulti);
+    FineField zzF(grid); zzF = Zero();
+
+    for(int i0=0;i0<nbasis;i0+=batch){ // Loop over batches of basis vectors
+
+      int nbv = MIN(batch,nbasis-i0);
+      std::cout << GridLogMessage<< "CoarsenMatrixColoured vec "<<i0<<"/"<<nbasis<< std::endl;
+
+      for(int p=0;p<npoint;p++){ // Loop over momenta
+
+	// One phase, applied to the whole batch. Tail slices are zeroed so
+	// the operator never sees undefined data.
+	for(int b=0;b<nbv;b++){
+	  tphaseBZ-=usecond();
+	  phaV = phaF[p]*Subspace[i0+b];
+	  tphaseBZ+=usecond();
+	  tslice-=usecond();
+	  InsertSliceFast(phaV,hi_in,b,0);
+	  tslice+=usecond();
+	}
+	tslice-=usecond();
+	for(int b=nbv;b<batch;b++){
+	  InsertSliceFast(zzF,hi_in,b,0);
+	}
+	tslice+=usecond();
+
+	tmat-=usecond();
+	linop.Op(hi_in,hi_out);
+	tmat+=usecond();
+
+	tproj-=usecond();
+	Projector.blockProject(hi_out,TmpProj);
+	tproj+=usecond();
+
+	tinv-=usecond();
+	CoarsenAccumulate(p,i0,nbv,batch,invMkl,pha,phaB,TmpProj,_A,CoarseGrid);
+	tinv+=usecond();
+      }
+    }
+
+    for(int p=0;p<npoint;p++){
+      GridtoBLAS(_A[p],BLAS_A[p]);
+    }
+    CoarsenReport(tphase,tphaseBZ,tslice,tmat,tproj,tinv);
+  }
+
+  ///////////////////////////////////////////////////////////////////////////
+  // Coarsen a SINGLE RHS fine operator.
+  //
+  // No multiRHS packing: the operator is applied once per phased basis
+  // vector and the batch is assembled on the coarse side by the mixed
+  // blockProject, which takes a vector of fine fields and writes the batch
+  // coarse field the accumulate expects. Only nbv applications per momentum,
+  // so a batch that does not divide nbasis wastes nothing, and the live fine
+  // storage is batch fields rather than two D+1 fields of extent batch.
+  ///////////////////////////////////////////////////////////////////////////
+  void CoarsenOperator(LinearOperatorBase<Lattice<Fobj> > &linop,
+		       std::vector<FineField> &Subspace,
+		       GridBase *CoarseGrid,
+		       int batch)
+  {
+    RealD tproj=0.0, tmat=0.0, tphase=0.0, tphaseBZ=0.0, tslice=0.0, tinv=0.0;
+
+    std::cout << GridLogMessage<< "GeneralCoarsenMatrixMrhs (single RHS fine operator)"<< std::endl;
+
+    GRID_ASSERT(Subspace.size()==nbasis);
+    GRID_ASSERT(batch>=1);
+    GridBase *grid = Subspace[0].Grid();
+
+    Coordinate blatt,bsimd,bmpi;
+    CoarsenBlockGridLayout(grid,CoarseGrid,blatt,bsimd,bmpi);
+    GridCartesian BlockGrid(blatt,bsimd,bmpi);
+
+    BlockComplexField InnerProd(&BlockGrid);
+    blockOrthogonalise(InnerProd,Subspace);
+
+    MultiRHSBlockProject<Lattice<Fobj> >    Projector;
+    Projector.Allocate(nbasis,grid,CoarseGrid);
+    Projector.ImportBasis(Subspace);
+
+    const int npoint = geom_srhs.npoint;
+
+    Eigen::MatrixXcd invMkl;
+    CoarsenFourierMatrix(CoarseGrid,invMkl);
+
+    FineField phaV(grid);
+    std::vector<FineComplexField>   phaF(npoint,grid);
+    std::vector<CoarseComplexField> pha (npoint,CoarseGrid);
+
+    tphase=-usecond();
+    CoarsenPhases(grid,CoarseGrid,&BlockGrid,pha,phaF);
+    tphase+=usecond();
+
+    std::vector<CoarseMatrix> _A;
+    _A.resize(npoint,CoarseGrid);
+    for(int k=0;k<npoint;k++) _A[k] = Zero();
+
+    Coordinate cmlatt,cmsimd,cmmpi;
+    CoarsenBatchGridLayout(CoarseGrid,batch,cmlatt,cmsimd,cmmpi);
+    GridCartesian CoarseBatchGrid(cmlatt,cmsimd,cmmpi);
+
+    CoarseVector       TmpProj(&CoarseBatchGrid);
+    CoarseComplexField phaB(&CoarseBatchGrid);
+
+    std::vector<FineField> MphaV(batch,grid);
+
+    for(int i0=0;i0<nbasis;i0+=batch){ // Loop over batches of basis vectors
+
+      int nbv = MIN(batch,nbasis-i0);
+      std::cout << GridLogMessage<< "CoarsenMatrixColoured vec "<<i0<<"/"<<nbasis<< std::endl;
+
+      for(int p=0;p<npoint;p++){ // Loop over momenta
+
+	for(int b=0;b<nbv;b++){
+	  tphaseBZ-=usecond();
+	  phaV = phaF[p]*Subspace[i0+b];
+	  tphaseBZ+=usecond();
+	  tmat-=usecond();
+	  linop.Op(phaV,MphaV[b]);
+	  tmat+=usecond();
+	}
+	// The accumulate reads only the first nbv slices, but the projector
+	// sees the whole vector, so the tail must not be undefined.
+	for(int b=nbv;b<batch;b++) MphaV[b] = Zero();
+
+	tproj-=usecond();
+	Projector.blockProject(MphaV,TmpProj);
+	tproj+=usecond();
+
+	tinv-=usecond();
+	CoarsenAccumulate(p,i0,nbv,batch,invMkl,pha,phaB,TmpProj,_A,CoarseGrid);
+	tinv+=usecond();
+      }
+    }
+
+    for(int p=0;p<npoint;p++){
+      GridtoBLAS(_A[p],BLAS_A[p]);
+    }
+    CoarsenReport(tphase,tphaseBZ,tslice,tmat,tproj,tinv);
+  }
+  void Mdag(const CoarseVector &in, CoarseVector &out)
+  {
+    this->M(in,out);
+  }
+  void M (const CoarseVector &in, CoarseVector &out)
+  {
+    //    std::cout << GridLogMessage << "New Mrhs coarse"<<std::endl;
+    conformable(CoarseGrid(),in.Grid());
+    conformable(in.Grid(),out.Grid());
+    out.Checkerboard() = in.Checkerboard();
+
+    RealD t_tot;
+    RealD t_exch;
+    RealD t_GtoB;
+    RealD t_BtoG;
+    RealD t_mult;
+
+    CheckGridSet();
+    if ( in.Grid() != _CoarseGridMulti ) {
+      std::cout << GridLogError
+		<< "MultiGeneralCoarsenedOperatorV2::M called with a field on a"
+		<< std::endl;
+      std::cout << GridLogError
+		<< "  different grid object from the one given to SetGrid(). Two"
+		<< std::endl;
+      std::cout << GridLogError
+		<< "  grids of identical shape do not conform; share one object."
+		<< std::endl;
+      GRID_ASSERT(in.Grid() == _CoarseGridMulti);
+    }
+
+    t_tot=-usecond();
+    CoarseVector tin=in;
+    t_exch=-usecond();
+    CoarseVector pin = CellMulti->ExchangePeriodic(tin); //padded input
+    t_exch+=usecond();
+
+    CoarseVector pout(pin.Grid());
+
+    int npoint = geom.npoint;
+    typedef calcMatrix* Aview;
+    typedef LatticeView<Cvec> Vview;
+      
+    const int Nsimd = CComplex::Nsimd();
+
+    int64_t nrhs  =pin.Grid()->GlobalDimensions()[0];
+    GRID_ASSERT(nrhs>=1);
+
+    RealD flops,bytes;
+    int64_t osites=in.Grid()->oSites(); // unpadded
+    int64_t unpadded_vol = CoarseGrid()->lSites()/nrhs;
+    
+    flops = 1.0* npoint * nbasis * nbasis * 8.0 * osites * CComplex::Nsimd();
+    bytes = 1.0*osites*sizeof(siteMatrix)*npoint/pin.Grid()->GlobalDimensions()[0]
+          + 2.0*osites*sizeof(siteVector)*npoint;
+    
+
+    t_GtoB=-usecond();
+    GridtoBLAS(pin,BLAS_B);
+    t_GtoB+=usecond();
+
+    GridBLAS BLAS;
+
+    t_mult=-usecond();
+    for(int p=0;p<geom.npoint;p++){
+      RealD c = 1.0;
+      if (p==0) c = 0.0;
+      ComplexD beta(c);
+
+      BLAS.gemmBatched(nbasis,nrhs,nbasis,
+		       ComplexD(1.0),
+		       BLAS_AP[p], 
+		       BLAS_BP[p], 
+		       ComplexD(c), 
+		       BLAS_CP);
+    }
+    BLAS.synchronise();
+    t_mult+=usecond();
+
+    t_BtoG=-usecond();
+    BLAStoGrid(out,BLAS_C);
+    t_BtoG+=usecond();
+    t_tot+=usecond();
+    /*
+    std::cout << GridLogMessage << "New Mrhs coarse DONE "<<std::endl;
+    std::cout << GridLogMessage<<"Coarse Mult exch "<<t_exch<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"Coarse Mult mult "<<t_mult<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"Coarse Mult GtoB  "<<t_GtoB<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"Coarse Mult BtoG  "<<t_BtoG<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"Coarse Mult tot  "<<t_tot<<" us"<<std::endl;
+    */
+    //    std::cout << GridLogMessage<<std::endl;
+    //    std::cout << GridLogMessage<<"Coarse Kernel flops "<< flops<<std::endl;
+    //    std::cout << GridLogMessage<<"Coarse Kernel flop/s "<< flops/t_mult<<" mflop/s"<<std::endl;
+    //    std::cout << GridLogMessage<<"Coarse Kernel bytes/s "<< bytes/t_mult/1000<<" GB/s"<<std::endl;
+    //    std::cout << GridLogMessage<<"Coarse overall flops/s "<< flops/t_tot<<" mflop/s"<<std::endl;
+    //    std::cout << GridLogMessage<<"Coarse total bytes   "<< bytes/1e6<<" MB"<<std::endl;
+  };
+  virtual  void Mdiag    (const Field &in, Field &out){ GRID_ASSERT(0);};
+  virtual  void Mdir     (const Field &in, Field &out,int dir, int disp){assert(0);};
+  virtual  void MdirAll  (const Field &in, std::vector<Field> &out){assert(0);};
+};
+  
+NAMESPACE_END(Grid);

Grid/algorithms/multigrid/Geometry.h

@@ -67,8 +67,8 @@ public:
   }
 
   int point(int dir, int disp) {
-    assert(disp == -1 || disp == 0 || disp == 1);
-    assert(base+0 <= dir && dir < base+4);
+    GRID_ASSERT(disp == -1 || disp == 0 || disp == 1);
+    GRID_ASSERT(base+0 <= dir && dir < base+4);
 
     // directions faster index = new indexing
     // 4d (base = 0):
@@ -131,7 +131,7 @@ public:
 	return p;
       }
     }
-    assert(0);
+    GRID_ASSERT(0);
     return -1;
   }
   void BuildShifts(void)

Grid/algorithms/multigrid/MrhsPromotedOperator.h

@@ -0,0 +1,161 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: ./lib/algorithms/multigrid/MrhsPromotedOperator.h
+
+    Copyright (C) 2026
+
+Author: Peter Boyle <pboyle@bnl.gov>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+///////////////////////////////////////////////////////////////////////////////
+// Present a D dimensional operator as a D+1 dimensional operator with Nrhs in
+// dimension 0.  Field type is unchanged; only the Grid differs, so this is a
+// LinearOperatorBase and callers need no template parameter: a native mrhs
+// operator derives from the same base and substitutes without a call site
+// change.
+//
+// Slices in and out around the wrapped operator.  No arithmetic beyond the
+// wrapped call, but one ExtractSliceFast/InsertSliceFast pair per right hand
+// side: data motion, not work.
+//
+// AdjOp is carried so that A^dag may be coarsened as a separate coarse
+// operator when needed, rather than doubling coarse storage.
+///////////////////////////////////////////////////////////////////////////////
+template<class Field>
+class MrhsPromotedOperator : public LinearOperatorBase<Field>
+{
+private:
+
+  LinearOperatorBase<Field> &_LinOp;
+  GridBase                  *_LowGrid;
+  int                        _Nrhs;
+
+public:
+
+  MrhsPromotedOperator(LinearOperatorBase<Field> &LinOp,GridBase *LowGrid,int Nrhs)
+    : _LinOp(LinOp), _LowGrid(LowGrid), _Nrhs(Nrhs)
+  {
+    GRID_ASSERT(_Nrhs>=1);
+  }
+
+  GridBase *LowGrid(void) { return _LowGrid; }
+  int       Nrhs(void)    { return _Nrhs;    }
+
+  // Reset on each call; retrieve and accumulate in the caller
+  RealD tslice;
+  RealD top;
+
+  void OpDiag (const Field &in, Field &out)
+  {
+    SliceLoop(in,out,[&](Field &i,Field &o){ _LinOp.OpDiag(i,o); });
+  }
+
+  void Op     (const Field &in, Field &out)
+  {
+    SliceLoop(in,out,[&](Field &i,Field &o){ _LinOp.Op(i,o); });
+  }
+
+  void AdjOp  (const Field &in, Field &out)
+  {
+    SliceLoop(in,out,[&](Field &i,Field &o){ _LinOp.AdjOp(i,o); });
+  }
+
+  void HermOp (const Field &in, Field &out)
+  {
+    SliceLoop(in,out,[&](Field &i,Field &o){ _LinOp.HermOp(i,o); });
+  }
+
+  void OpDir  (const Field &in, Field &out,int dir,int disp)
+  {
+    SliceLoop(in,out,[&](Field &i,Field &o){ _LinOp.OpDir(i,o,dir,disp); });
+  }
+
+  //////////////////////////////////////////////////////////////////
+  // Norms of the D+1 field are the sums over slices
+  //////////////////////////////////////////////////////////////////
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2)
+  {
+    Conformable(in,out);
+    Field lo_in (_LowGrid);
+    Field lo_out(_LowGrid);
+    n1=0.0;
+    n2=0.0;
+    for(int r=0;r<_Nrhs;r++){
+      RealD r1,r2;
+      ExtractSliceFast(lo_in,in,r,0);
+      _LinOp.HermOpAndNorm(lo_in,lo_out,r1,r2);
+      InsertSliceFast(lo_out,out,r,0);
+      n1=n1+r1;
+      n2=n2+r2;
+    }
+  }
+
+  void OpDirAll(const Field &in, std::vector<Field> &out)
+  {
+    int npoint = out.size();
+    Field lo_in(_LowGrid);
+    std::vector<Field> lo_out(npoint,_LowGrid);
+    for(int r=0;r<_Nrhs;r++){
+      ExtractSliceFast(lo_in,in,r,0);
+      _LinOp.OpDirAll(lo_in,lo_out);
+      for(int p=0;p<npoint;p++){
+        InsertSliceFast(lo_out[p],out[p],r,0);
+      }
+    }
+  }
+
+private:
+
+  void Conformable(const Field &in,const Field &out)
+  {
+    conformable(in.Grid(),out.Grid());
+    GRID_ASSERT(in.Grid()->_ndimension == _LowGrid->_ndimension+1);
+    GRID_ASSERT(in.Grid()->_fdimensions[0] == _Nrhs);
+  }
+
+  template<class Kernel>
+  void SliceLoop(const Field &in,Field &out,Kernel K)
+  {
+    Conformable(in,out);
+    Field lo_in (_LowGrid);
+    Field lo_out(_LowGrid);
+    tslice=0.0;
+    top=0.0;
+    for(int r=0;r<_Nrhs;r++){
+      tslice-=usecond();
+      ExtractSliceFast(lo_in,in,r,0);
+      tslice+=usecond();
+      top-=usecond();
+      K(lo_in,lo_out);
+      top+=usecond();
+      tslice-=usecond();
+      InsertSliceFast(lo_out,out,r,0);
+      tslice+=usecond();
+    }
+  }
+
+};
+
+NAMESPACE_END(Grid);

Grid/algorithms/multigrid/MultiGrid.h

@@ -32,3 +32,5 @@ Author: Peter Boyle <pboyle@bnl.gov>
 #include <Grid/algorithms/multigrid/CoarsenedMatrix.h>
 #include <Grid/algorithms/multigrid/GeneralCoarsenedMatrix.h>
 #include <Grid/algorithms/multigrid/GeneralCoarsenedMatrixMultiRHS.h>
+#include <Grid/algorithms/multigrid/GeneralCoarsenedMatrixMultiRHSV2.h>
+#include <Grid/algorithms/multigrid/MrhsPromotedOperator.h>

Grid/algorithms/multigrid/RecursiveSchurInverse.h

@@ -0,0 +1,617 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: RecursiveSchurInverse.h
+
+    Copyright (C) 2026
+
+Author: Peter Boyle <pboyle@bnl.gov>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#pragma once
+
+#include <Grid/algorithms/blas/BatchedBlas.h>
+#include <Grid/algorithms/blas/BatchedInverse.h>
+
+NAMESPACE_BEGIN(Grid);
+
+///////////////////////////////////////////////////////////////////////////////
+// RecursiveSchurInverse: distributed dense inversion by recursive Schur
+// complement over a binary rank-range tree.
+//
+// Contract: rank r owns global rows [rowStart[r], rowStart[r+1]) of an
+// N x N matrix in rank-major ordering, and receives its rows of the
+// inverse in the same layout.  Consumes GridBase collectives, GridBLAS
+// and GridBLASInverse only; Eigen reference backends permit CPU unit
+// testing under mpirun (Test_schur_inverse).
+//
+// Arithmetic is fp64 throughout; the caller rounds once into fp32 storage.
+//
+// Execution: SPMD full-tree walk.  Every rank makes the identical call
+// sequence; data participation is ownership-gated; all collectives are
+// world-wide, so no deadlock surface exists.
+//
+// Storage: BlockRows is column-major, ld = rows; element (i,j) at
+// data[i + j*ld]; a column window is the contiguous slice at data[col0*ld].
+///////////////////////////////////////////////////////////////////////////////
+
+///////////////////////////////////////////////////////////////////////////////
+// My rows of a distributed dense matrix: rows x cols, column major, ld = rows.
+///////////////////////////////////////////////////////////////////////////////
+class BlockRows
+{
+public:
+  deviceVector<ComplexD> data;
+  int64_t                rows;
+  int64_t                cols;
+  int64_t                ld;
+
+  BlockRows()
+  {
+    rows = 0;
+    cols = 0;
+    ld   = 0;
+  }
+  void Resize(int64_t r, int64_t c)
+  {
+    rows = r;
+    cols = c;
+    ld   = r;
+    data.resize((uint64_t)r*c);
+  }
+  ComplexD *ColumnWindow(int64_t col0)
+  {
+    GRID_ASSERT( col0 >= 0 );
+    GRID_ASSERT( col0 <= cols );
+    return &data[(uint64_t)col0*ld];
+  }
+};
+
+class RecursiveSchurInverse
+{
+public:
+  GridBase             *grid;       // world collectives only
+  int64_t               N;          // global matrix dimension
+  int                   P;          // ranks
+  int                   me;         // this rank
+  std::vector<int64_t>  rowStart;   // P+1 entries: rank-major row ownership
+  int64_t               myRow0;
+  int64_t               myNrows;
+  int64_t               panelBytes; // gather panel budget (DENSE_PANEL_BYTES)
+
+  GridBLAS              BLAS;
+  GridBLASInverse       INV;
+
+  // Growth telemetry: one entry per merge node, walk order
+  std::vector<double>   telNormB;      // ||B||_F = ||A11inv A12||_F
+  std::vector<double>   telSratio;     // ||S||_F / ||A22||_F
+  double                telLeafMaxInv; // max |(leaf inverse)_ij| over leaves
+
+  // Phase timers/counters, reported by ReportTelemetry
+  double                tMemset;       // device panel zero-fill
+  double                tDeposit;      // owner rows -> panel (device kernel)
+  double                tAllreduce;    // GlobalSumVector on device panels
+  double                tGemm;         // strided gemm + synchronise
+  double                tLeaf;         // leaf inversions
+  double                tARmin;        // fastest single panel collective
+  double                tARmax;        // slowest single panel collective
+  uint64_t              bytesAllreduce;
+  uint64_t              nAllreduce;    // panel collectives
+  uint64_t              nGatherGemm;   // GatherGemm calls
+
+  // Persistent grow-only device panel; assembly and collectives are
+  // device-resident.  Device builds require GPU-aware MPI.
+  deviceVector<ComplexD> dPanelBuf;
+
+  ///////////////////////////////////////////////////////////////////////////
+  // Ownership-table validation: a proper partition of [0,N).
+  // Static and communicator-free so synthetic tables unit-test directly.
+  ///////////////////////////////////////////////////////////////////////////
+  static void CheckRowStart(const std::vector<int64_t> &table, int64_t N)
+  {
+    int P = (int)table.size() - 1;
+    GRID_ASSERT( P >= 1 );
+    GRID_ASSERT( table[0] == 0 );
+    GRID_ASSERT( table[P] == N );
+    for(int r=0; r<P; r++)
+    {
+      GRID_ASSERT( table[r+1] >= table[r] );  // zero-row ranks permitted
+    }
+  }
+
+  ///////////////////////////////////////////////////////////////////////////
+  // Build the ownership table from each rank's local row count: zero-fill
+  // allgather (the standing comms idiom) then prefix sum.  Every rank
+  // returns the identical table.
+  ///////////////////////////////////////////////////////////////////////////
+  static std::vector<int64_t> MakeRowStart(GridBase *g, int64_t myNrows)
+  {
+    int P  = g->ProcessorCount();
+    int me = g->ThisRank();
+
+    std::vector<uint64_t> counts(P, 0);
+    counts[me] = (uint64_t)myNrows;
+    g->GlobalSumVector(&counts[0], P);
+
+    std::vector<int64_t> table(P+1);
+    table[0] = 0;
+    for(int r=0; r<P; r++)
+    {
+      table[r+1] = table[r] + (int64_t)counts[r];
+    }
+    CheckRowStart(table, table[P]);
+    return table;
+  }
+
+  RecursiveSchurInverse(GridBase *g,
+                        int64_t N_,
+                        std::vector<int64_t> &rowStart_,
+                        int64_t panelBytes_)
+  {
+    grid       = g;
+    N          = N_;
+    P          = g->ProcessorCount();
+    me         = g->ThisRank();
+    rowStart   = rowStart_;
+    panelBytes = panelBytes_;
+
+    GRID_ASSERT( (int)rowStart.size() == P+1 );
+    CheckRowStart(rowStart, N);
+
+    myRow0  = rowStart[me];
+    myNrows = rowStart[me+1] - rowStart[me];
+
+    telLeafMaxInv = 0.0;
+  }
+
+  ///////////////////////////////////////////////////////////////////////////
+  // The communication primitive.
+  //
+  //   C(:, colC : colC+widthB) <-   beta  * C(:, colC : colC+widthB)
+  //                               + alpha * A(:, colA : colA+widthA) * Bsub
+  //
+  // Bsub is the widthA x widthB sub-block of a row-distributed operand
+  // owned by ranks [rB0, rB1): owner r contributes its rows of
+  // B(:, colB : colB+widthB) at sub-block row offset
+  // rowStart[r] - rowStart[rB0], gathered in panelBytes row-chunks by
+  // device zero-fill + deposit kernel + GlobalSumVector.
+  //
+  // Every rank calls; non-owners of B add zeros; ranks with A.rows == 0
+  // skip local compute but make every collective call.  Column offsets
+  // are local buffer offsets -- non-participants pass 0.
+  ///////////////////////////////////////////////////////////////////////////
+  void GatherGemm(ComplexD alpha,
+                  BlockRows &A, int64_t colA, int64_t widthA,
+                  int rB0, int rB1,
+                  BlockRows &B, int64_t colB, int64_t widthB,
+                  ComplexD beta,
+                  BlockRows &C, int64_t colC)
+  {
+    GRID_ASSERT( rB0 >= 0 );
+    GRID_ASSERT( rB1 >  rB0 );
+    GRID_ASSERT( rB1 <= P );
+
+    int64_t k = rowStart[rB1] - rowStart[rB0];
+    int64_t m = A.rows;
+    int64_t n = widthB;
+    GRID_ASSERT( widthA == k );
+    GRID_ASSERT( n >= 1 );
+
+    int owner = ( me >= rB0 ) && ( me < rB1 ) && ( B.rows > 0 );
+    int64_t myOff = 0;
+    if ( owner )
+    {
+      myOff = rowStart[me] - rowStart[rB0];
+    }
+
+    if ( m > 0 )
+    {
+      GRID_ASSERT( colA + widthA <= A.cols );
+      GRID_ASSERT( colC + widthB <= C.cols );
+      GRID_ASSERT( C.rows == m );
+    }
+
+    nGatherGemm++;
+    GRID_TRACE("GatherGemm");
+
+    if ( owner )
+    {
+      GRID_ASSERT( colB + widthB <= B.cols );
+    }
+
+    int64_t kc = panelBytes / ( (int64_t)sizeof(ComplexD) * n );
+    if ( kc < 1 ) kc = 1;
+    if ( kc > k ) kc = k;
+    GRID_ASSERT( kc*n < 2147483647L );   // GlobalSumVector count is int
+
+    deviceVector<ComplexD> &dPanel = dPanelBuf;
+    if ( dPanel.size() < (uint64_t)kc*n ) dPanel.resize((uint64_t)kc*n);
+    deviceVector<ComplexD*> ap(1);
+    deviceVector<ComplexD*> bp(1);
+    deviceVector<ComplexD*> cp(1);
+    std::vector<ComplexD*>  ptr(1);
+
+    for(int64_t k0=0; k0<k; k0+=kc)
+    {
+      int64_t kchunk = std::min(kc, k-k0);
+
+      // Zeroing must complete before MPI reads the panel; it is the sole
+      // producer on non-owner ranks.
+      tMemset -= usecond();
+      acceleratorMemSet(&dPanel[0], 0, (uint64_t)kchunk*n*sizeof(ComplexD));
+      tMemset += usecond();
+      if ( owner )
+      {
+        int64_t i0 = std::max(k0,        myOff);
+        int64_t i1 = std::min(k0+kchunk, myOff+B.rows);
+        if ( i1 > i0 )
+        {
+          // Deposit my rows: strided block copy, len rows x n columns
+          int64_t   len   = i1-i0;
+          int64_t   brows = B.rows;
+          int64_t   dof   = i0-k0;
+          int64_t   sof   = i0-myOff;
+          ComplexD *src   = B.ColumnWindow(colB);
+          ComplexD *dst   = &dPanel[0];
+          tDeposit -= usecond();
+          accelerator_for(idx, (uint64_t)(len*n), 1, {
+            int64_t j = idx / len;
+            int64_t i = idx % len;
+            dst[(uint64_t)(dof + i + j*kchunk)] = src[(uint64_t)(sof + i + j*brows)];
+          });
+          tDeposit += usecond();
+        }
+      }
+      double tar = -usecond();
+      grid->GlobalSumVector(&dPanel[0], (int)(kchunk*n));
+      tar += usecond();
+      tAllreduce += tar;
+      tARmin = std::min(tARmin, tar);
+      tARmax = std::max(tARmax, tar);
+      bytesAllreduce += (uint64_t)kchunk*n*sizeof(ComplexD);
+      nAllreduce++;
+
+      if ( m > 0 )
+      {
+        ComplexD beta_use = ( k0==0 ) ? beta : ComplexD(1.0,0.0);
+
+        ptr[0] = A.ColumnWindow(colA + k0);
+        acceleratorCopyToDevice(&ptr[0], &ap[0], sizeof(ComplexD*));
+        ptr[0] = &dPanel[0];
+        acceleratorCopyToDevice(&ptr[0], &bp[0], sizeof(ComplexD*));
+        ptr[0] = C.ColumnWindow(colC);
+        acceleratorCopyToDevice(&ptr[0], &cp[0], sizeof(ComplexD*));
+
+        tGemm -= usecond();
+        BLAS.gemmBatched(GridBLAS_OP_N, GridBLAS_OP_N,
+                         (int)m, (int)n, (int)kchunk,
+                         alpha,    ap, (int)A.ld,
+                                   bp, (int)kchunk,
+                         beta_use, cp, (int)C.ld);
+        BLAS.synchronise();
+        tGemm += usecond();
+      }
+    }
+  }
+
+  ///////////////////////////////////////////////////////////////////////////
+  // Local Frobenius norm-squared of a full-height column window.
+  // NO comms; callers GlobalSum the result.  Host staging, setup-scale.
+  ///////////////////////////////////////////////////////////////////////////
+  double FrobNorm2Local(BlockRows &X, int64_t col0, int64_t w)
+  {
+    if ( X.rows == 0 ) return 0.0;
+    GRID_ASSERT( col0 + w <= X.cols );
+    uint64_t len = (uint64_t)X.rows*w;
+    std::vector<ComplexD> h(len);
+    acceleratorCopyFromDevice(X.ColumnWindow(col0), &h[0], len*sizeof(ComplexD));
+    // Member real()/imag(): portable across std::complex (CPU) and
+    // thrust::complex (HIP), where std::norm does not resolve.
+    double s = 0.0;
+    for(uint64_t i=0; i<len; i++)
+    {
+      double re = h[i].real();
+      double im = h[i].imag();
+      s += re*re + im*im;
+    }
+    return s;
+  }
+
+  ///////////////////////////////////////////////////////////////////////////
+  // dst(:, dcol0 : dcol0+w) = - src(:, 0:w).  Both operands have ld == rows
+  // so full-height windows are contiguous: flat elementwise device copy.
+  ///////////////////////////////////////////////////////////////////////////
+  void NegateCopy(BlockRows &src, BlockRows &dst, int64_t dcol0, int64_t w)
+  {
+    GRID_ASSERT( src.rows == dst.rows );
+    GRID_ASSERT( w <= src.cols );
+    GRID_ASSERT( dcol0 + w <= dst.cols );
+    if ( src.rows == 0 ) return;
+    uint64_t len = (uint64_t)src.rows*w;
+    ComplexD *s = &src.data[0];
+    ComplexD *d = dst.ColumnWindow(dcol0);
+    accelerator_for(i, len, 1, {
+      d[i] = -s[i];
+    });
+  }
+
+  ///////////////////////////////////////////////////////////////////////////
+  // Leaf inversion: local, in place on the contiguous diagonal window.
+  // No collectives.
+  ///////////////////////////////////////////////////////////////////////////
+  void LeafInvert(int64_t col0, int64_t width, BlockRows &Arows)
+  {
+    GRID_TRACE("SchurLeaf");
+    GRID_ASSERT( width == Arows.rows );
+    GRID_ASSERT( col0 + width <= Arows.cols );
+    int64_t  w   = width;
+    uint64_t len = (uint64_t)w*w;
+    tLeaf -= usecond();
+
+    deviceVector<ComplexD*> bp(1);
+    std::vector<ComplexD*>  ptr(1);
+    ptr[0] = Arows.ColumnWindow(col0);
+    acceleratorCopyToDevice(&ptr[0], &bp[0], sizeof(ComplexD*));
+    INV.inverseBatched(w, bp);
+
+    // Telemetry: max |element| of the leaf inverse
+    {
+      std::vector<ComplexD> h(len);
+      acceleratorCopyFromDevice(Arows.ColumnWindow(col0), &h[0], len*sizeof(ComplexD));
+      double mx = 0.0;
+      for(uint64_t i=0; i<len; i++)
+      {
+        double re = h[i].real();
+        double im = h[i].imag();
+        mx = std::max(mx, re*re + im*im);
+      }
+      telLeafMaxInv = std::max(telLeafMaxInv, std::sqrt(mx));
+    }
+    tLeaf += usecond();
+  }
+
+  ///////////////////////////////////////////////////////////////////////////
+  // The recursion (plan 3.5 / 4B.3).  Inverts the diagonal block of the
+  // rank-major matrix spanned by ranks [r0, r1), living in every member
+  // rank's column window [col0, col0+width) -- IN PLACE.
+  //
+  // SPMD: every rank calls with IDENTICAL (r0, r1, width) and its own
+  // local (col0, Arows); ranks outside [r0, r1) participate in the
+  // collectives only (dummy operands, zero contributions).  The collective
+  // sequence -- 5 GatherGemm calls + 3 scalar GlobalSums per merge node --
+  // is identical on every rank by construction.
+  //
+  //   I = [r0, mid)   J = [mid, r1)   widths WI, WJ
+  //   1. recurse I:      A11 -> A11inv
+  //   2. B = A11inv.A12                     (I rows)
+  //   3. C = A21.A11inv                     (J rows)
+  //   4. S = A22 - A21.B    in place        (J rows)   [alpha=-1, beta=1]
+  //   5. recurse J:      S -> Sinv
+  //   6. T = Sinv.C                         (J rows)
+  //   7. U = B.Sinv                         (I rows)
+  //   8. X11 = A11inv + U.C  in place       (I rows)   [beta=1]
+  //   9. X12 = -U, X21 = -T  local negates; X22 = Sinv already in place
+  ///////////////////////////////////////////////////////////////////////////
+  void SchurNode(int r0, int r1, int64_t col0, int64_t width, BlockRows &Arows)
+  {
+    int span = r1 - r0;
+    GRID_ASSERT( span >= 1 );
+    GRID_ASSERT( width == rowStart[r1] - rowStart[r0] );
+
+    if ( span == 1 )
+    {
+      if ( ( me == r0 ) && ( myNrows > 0 ) )
+      {
+        LeafInvert(col0, width, Arows);
+      }
+      return;
+    }
+
+    int     mid = ( r0 + r1 ) / 2;
+    int64_t WI  = rowStart[mid] - rowStart[r0];
+    int64_t WJ  = rowStart[r1]  - rowStart[mid];
+
+    // Zero-width child ranges (all ranks of a half owning no rows) are a
+    // KNOWN LIMITATION: fail loudly rather than divide mysteriously.
+    GRID_ASSERT( WI > 0 );
+    GRID_ASSERT( WJ > 0 );
+
+    int inI = ( me >= r0  ) && ( me < mid );
+    int inJ = ( me >= mid ) && ( me < r1  );
+
+    ComplexD one  ( 1.0,0.0);
+    ComplexD mone (-1.0,0.0);
+    ComplexD zero ( 0.0,0.0);
+
+    BlockRows dummy;
+
+    // 1. A11 -> A11inv
+    SchurNode(r0, mid, col0, WI, Arows);
+
+    // 2. B = A11inv . A12   (I rows; gather A12 from I owners)
+    BlockRows Bbuf;
+    if ( inI ) Bbuf.Resize(myNrows, WJ);
+    {
+      BlockRows &Aop = inI ? Arows : dummy;
+      BlockRows &Cop = inI ? Bbuf  : dummy;
+      int64_t    cA  = inI ? col0  : 0;
+      GatherGemm(one, Aop, cA, WI,
+                 r0, mid,
+                 Arows, col0+WI, WJ,
+                 zero, Cop, 0);
+    }
+    double nB = FrobNorm2Local(Bbuf, 0, inI ? WJ : 0);
+    grid->GlobalSumVector(&nB, 1);
+    telNormB.push_back(std::sqrt(nB));
+
+    // 3. C = A21 . A11inv   (J rows; gather A11inv from I owners)
+    BlockRows Cbuf;
+    if ( inJ ) Cbuf.Resize(myNrows, WI);
+    {
+      BlockRows &Aop = inJ ? Arows : dummy;
+      BlockRows &Cop = inJ ? Cbuf  : dummy;
+      int64_t    cA  = inJ ? col0  : 0;
+      GatherGemm(one, Aop, cA, WI,
+                 r0, mid,
+                 Arows, col0, WI,
+                 zero, Cop, 0);
+    }
+
+    // 4. S = A22 - A21 . B  in place on my A22 window (J rows)
+    double nA22 = FrobNorm2Local( inJ ? Arows : dummy, inJ ? col0+WI : 0, inJ ? WJ : 0 );
+    grid->GlobalSumVector(&nA22, 1);
+    {
+      BlockRows &Aop = inJ ? Arows   : dummy;
+      BlockRows &Cop = inJ ? Arows   : dummy;
+      int64_t    cA  = inJ ? col0    : 0;
+      int64_t    cC  = inJ ? col0+WI : 0;
+      GatherGemm(mone, Aop, cA, WI,
+                 r0, mid,
+                 Bbuf, 0, WJ,
+                 one, Cop, cC);
+    }
+    double nS = FrobNorm2Local( inJ ? Arows : dummy, inJ ? col0+WI : 0, inJ ? WJ : 0 );
+    grid->GlobalSumVector(&nS, 1);
+    telSratio.push_back( std::sqrt(nS) / ( std::sqrt(nA22) + 1.0e-300 ) );
+
+    // 5. S -> Sinv
+    SchurNode(mid, r1, col0+WI, WJ, Arows);
+
+    // 6. T = Sinv . C       (J rows; gather C from J owners)
+    BlockRows Tbuf;
+    if ( inJ ) Tbuf.Resize(myNrows, WI);
+    {
+      BlockRows &Aop = inJ ? Arows   : dummy;
+      BlockRows &Cop = inJ ? Tbuf    : dummy;
+      int64_t    cA  = inJ ? col0+WI : 0;
+      GatherGemm(one, Aop, cA, WJ,
+                 mid, r1,
+                 Cbuf, 0, WI,
+                 zero, Cop, 0);
+    }
+
+    // 7. U = B . Sinv       (I rows; gather Sinv from J owners)
+    BlockRows Ubuf;
+    if ( inI ) Ubuf.Resize(myNrows, WJ);
+    {
+      BlockRows &Aop = inI ? Bbuf : dummy;
+      BlockRows &Cop = inI ? Ubuf : dummy;
+      GatherGemm(one, Aop, 0, WJ,
+                 mid, r1,
+                 Arows, col0+WI, WJ,
+                 zero, Cop, 0);
+    }
+
+    // 8. X11 = A11inv + U . C  in place (I rows; gather C from J owners)
+    {
+      BlockRows &Aop = inI ? Ubuf  : dummy;
+      BlockRows &Cop = inI ? Arows : dummy;
+      int64_t    cC  = inI ? col0  : 0;
+      GatherGemm(one, Aop, 0, WJ,
+                 mid, r1,
+                 Cbuf, 0, WI,
+                 one, Cop, cC);
+    }
+
+    // 9. Off-diagonal signs, local
+    if ( inI ) NegateCopy(Ubuf, Arows, col0+WI, WJ);
+    if ( inJ ) NegateCopy(Tbuf, Arows, col0,    WI);
+  }
+
+  ///////////////////////////////////////////////////////////////////////////
+  // PUBLIC ENTRY.  Arows: my rows of the rank-major N x N matrix (fp64).
+  // On exit Arows holds my rows of the inverse, still fp64; the caller
+  // owns the single terminal rounding into its fp32 apply storage.
+  ///////////////////////////////////////////////////////////////////////////
+  void Invert(BlockRows &Arows)
+  {
+    GRID_ASSERT( Arows.rows == myNrows );
+    GRID_ASSERT( Arows.cols == N );
+
+    telNormB.resize(0);
+    telSratio.resize(0);
+    telLeafMaxInv = 0.0;
+
+    tMemset        = 0.0;
+    tDeposit       = 0.0;
+    tAllreduce     = 0.0;
+    tGemm          = 0.0;
+    tLeaf          = 0.0;
+    tARmin         = 1.0e30;
+    tARmax         = 0.0;
+    bytesAllreduce = 0;
+    nAllreduce     = 0;
+    nGatherGemm    = 0;
+
+    SchurNode(0, P, 0, N, Arows);
+
+    RealD mx = telLeafMaxInv;
+    grid->GlobalMax(mx);
+    telLeafMaxInv = mx;
+  }
+
+  // All telemetry values are globally reduced or boss-local; safe to
+  // stream on every rank (Grid quiesces stdout to the boss unless
+  // --debug-stdout).  NOTE: tAllreduce INCLUDES wait/imbalance -- a rank
+  // arriving early books its wait here; the min/max spread across ranks
+  // separates true wire time (~min) from skew (max-min).
+  void ReportTelemetry(void)
+  {
+    for(uint64_t i=0; i<telNormB.size(); i++)
+    {
+      std::cout << GridLogPerformance
+                << "SchurNode " << i
+                << " ||B||_F "       << telNormB[i]
+                << " ||S||/||A22|| " << telSratio[i]
+                << std::endl;
+    }
+    std::cout << GridLogPerformance
+              << "Schur leaves max|Ainv| " << telLeafMaxInv
+              << std::endl;
+
+    RealD armax =  tAllreduce;
+    RealD armin = -tAllreduce;
+    grid->GlobalMax(armax);
+    grid->GlobalMax(armin);
+    armin = -armin;
+
+    std::cout << GridLogMessage << "Schur phases (boss rank, seconds):"
+              << "  memset "    << tMemset/1.0e6
+              << "  deposit "   << tDeposit/1.0e6
+              << "  allreduce " << tAllreduce/1.0e6
+              << "  gemm "      << tGemm/1.0e6
+              << "  leaf "      << tLeaf/1.0e6
+              << std::endl;
+    std::cout << GridLogMessage << "Schur comms:"
+              << "  GatherGemm calls " << nGatherGemm
+              << "  panel allreduces " << nAllreduce
+              << "  allreduce GB "     << bytesAllreduce/1024./1024./1024.
+              << "  allreduce s min/max over ranks " << armin/1.0e6
+              << " / " << armax/1.0e6
+              << std::endl;
+    std::cout << GridLogMessage << "Schur comms per-call (boss):"
+              << "  min " << (nAllreduce ? tARmin/1.0e3 : 0.0) << " ms"
+              << "  avg " << (nAllreduce ? tAllreduce/nAllreduce/1.0e3 : 0.0) << " ms"
+              << "  max " << tARmax/1.0e3 << " ms"
+              << std::endl;
+  }
+};
+
+NAMESPACE_END(Grid);

Grid/allocator/AlignedAllocator.h

@@ -57,7 +57,7 @@ public:
     if ( (_Tp*)ptr == (_Tp *) NULL ) {
       printf("Grid CPU Allocator got NULL for %lu bytes\n",(unsigned long) bytes );
     }
-    assert( ( (_Tp*)ptr != (_Tp *)NULL ) );
+    GRID_ASSERT( ( (_Tp*)ptr != (_Tp *)NULL ) );
     return ptr;
   }
 
@@ -69,7 +69,7 @@ public:
   }
 
   // FIXME: hack for the copy constructor: it must be avoided to avoid single thread loop
-  void construct(pointer __p, const _Tp& __val) { assert(0);};
+  void construct(pointer __p, const _Tp& __val) { };
   void construct(pointer __p) { };
   void destroy(pointer __p) { };
 };
@@ -106,7 +106,7 @@ public:
     if ( (_Tp*)ptr == (_Tp *) NULL ) {
       printf("Grid Shared Allocator got NULL for %lu bytes\n",(unsigned long) bytes );
     }
-    assert( ( (_Tp*)ptr != (_Tp *)NULL ) );
+    GRID_ASSERT( ( (_Tp*)ptr != (_Tp *)NULL ) );
     return ptr;
   }
 
@@ -154,7 +154,7 @@ public:
     if ( (_Tp*)ptr == (_Tp *) NULL ) {
       printf("Grid Device Allocator got NULL for %lu bytes\n",(unsigned long) bytes );
     }
-    assert( ( (_Tp*)ptr != (_Tp *)NULL ) );
+    GRID_ASSERT( ( (_Tp*)ptr != (_Tp *)NULL ) );
     return ptr;
   }
 
@@ -174,19 +174,10 @@ template<typename _Tp>  inline bool operator!=(const devAllocator<_Tp>&, const d
 ////////////////////////////////////////////////////////////////////////////////
 // Template typedefs
 ////////////////////////////////////////////////////////////////////////////////
-#ifdef ACCELERATOR_CSHIFT
-// Cshift on device
-template<class T> using cshiftAllocator = devAllocator<T>;
-#else
-// Cshift on host
-template<class T> using cshiftAllocator = std::allocator<T>;
-#endif
-
-template<class T> using Vector        = std::vector<T,uvmAllocator<T> >;           
-template<class T> using stencilVector = std::vector<T,alignedAllocator<T> >;           
-template<class T> using commVector    = std::vector<T,devAllocator<T> >;
-template<class T> using deviceVector  = std::vector<T,devAllocator<T> >;
-template<class T> using cshiftVector  = std::vector<T,cshiftAllocator<T> >;
+template<class T> using hostVector          = std::vector<T,alignedAllocator<T> >;           // Needs autoview
+template<class T> using Vector              = std::vector<T,uvmAllocator<T> >;               // Really want to deprecate
+template<class T> using uvmVector           = std::vector<T,uvmAllocator<T> >;               // auto migrating page
+template<class T> using deviceVector        = std::vector<T,devAllocator<T> >;               // device vector
 
 /*
 template<class T> class vecView
@@ -197,8 +188,9 @@ template<class T> class vecView
   ViewMode mode;
   void * cpu_ptr;
  public:
+  // Rvalue accessor
   accelerator_inline T & operator[](size_t i) const { return this->data[i]; };
-  vecView(std::vector<T> &refer_to_me,ViewMode _mode)
+  vecView(Vector<T> &refer_to_me,ViewMode _mode)
   {
     cpu_ptr = &refer_to_me[0];
     size = refer_to_me.size();
@@ -214,22 +206,12 @@ template<class T> class vecView
   }
 };
 
-template<class T> vecView<T> VectorView(std::vector<T> &vec,ViewMode _mode)
+template<class T> vecView<T> VectorView(Vector<T> &vec,ViewMode _mode)
 {
   vecView<T> ret(vec,_mode); // does the open
   return ret;                // must be closed
 }
 
-// Little autoscope assister
-template<class View> 
-class VectorViewCloser
-{
-  View v;  // Take a copy of view and call view close when I go out of scope automatically
- public:
-  VectorViewCloser(View &_v) : v(_v) {};
-  ~VectorViewCloser() { auto ptr = v.cpu_ptr; v.ViewClose();  MemoryManager::NotifyDeletion(ptr);}
-};
-
 #define autoVecView(v_v,v,mode)					\
   auto v_v = VectorView(v,mode);				\
   ViewCloser<decltype(v_v)> _autoView##v_v(v_v);

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();
@@ -292,7 +290,7 @@ void *MemoryManager::Insert(void *ptr,size_t bytes,int type)
 void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim, uint64_t &cacheBytes) 
 {
 #ifdef GRID_OMP
-  assert(omp_in_parallel()==0);
+  GRID_ASSERT(omp_in_parallel()==0);
 #endif 
 
   if (ncache == 0) return ptr;
@@ -345,7 +343,7 @@ void *MemoryManager::Lookup(size_t bytes,int type)
 void *MemoryManager::Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache,uint64_t & cacheBytes) 
 {
 #ifdef GRID_OMP
-  assert(omp_in_parallel()==0);
+  GRID_ASSERT(omp_in_parallel()==0);
 #endif 
   for(int e=0;e<ncache;e++){
     if ( entries[e].valid && ( entries[e].bytes == bytes ) ) {

Grid/allocator/MemoryManager.h

@@ -113,7 +113,7 @@ private:
   static void *Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim,uint64_t &cbytes) ;
   static void *Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache,uint64_t &cbytes) ;
 
- public:
+public:  
   static void PrintBytes(void);
   static void Audit(std::string s);
   static void Init(void);
@@ -215,6 +215,7 @@ private:
   static void NotifyDeletion(void * CpuPtr);
   static void Print(void);
   static void PrintAll(void);
+  static void EvictAll(void);
   static void PrintState( void* CpuPtr);
   static int   isOpen   (void* CpuPtr);
   static void  ViewClose(void* CpuPtr,ViewMode mode);

Grid/allocator/MemoryManagerCache.cc

@@ -1,16 +1,15 @@
 #include <Grid/GridCore.h>
 #ifndef GRID_UVM
 
-#warning "Using explicit device memory copies"
 NAMESPACE_BEGIN(Grid);
 
 #define MAXLINE 512
 static char print_buffer [ MAXLINE ];
 
-#define mprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer;
-#define dprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogDebug << print_buffer;
+#define mprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer << std::endl;
+#define dprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogDebug  << print_buffer << std::endl;
 //#define dprintf(...) 
-
+//#define mprintf(...) 
 
 ////////////////////////////////////////////////////////////
 // For caching copies of data on device
@@ -51,12 +50,12 @@ int   MemoryManager::EntryPresent(uint64_t CpuPtr)
 {
   if(AccViewTable.empty()) return 0;
 
-  auto count = AccViewTable.count(CpuPtr);  assert((count==0)||(count==1));
+  auto count = AccViewTable.count(CpuPtr);  GRID_ASSERT((count==0)||(count==1));
   return count;
 }
 void  MemoryManager::EntryCreate(uint64_t CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint)
 {
-  assert(!EntryPresent(CpuPtr));
+  GRID_ASSERT(!EntryPresent(CpuPtr));
   AcceleratorViewEntry AccCache;
   AccCache.CpuPtr = CpuPtr;
   AccCache.AccPtr = (uint64_t)NULL;
@@ -70,9 +69,9 @@ void  MemoryManager::EntryCreate(uint64_t CpuPtr,size_t bytes,ViewMode mode,View
 }
 MemoryManager::AccViewTableIterator MemoryManager::EntryLookup(uint64_t CpuPtr)
 {
-  assert(EntryPresent(CpuPtr));
+  GRID_ASSERT(EntryPresent(CpuPtr));
   auto AccCacheIterator = AccViewTable.find(CpuPtr);
-  assert(AccCacheIterator!=AccViewTable.end());
+  GRID_ASSERT(AccCacheIterator!=AccViewTable.end());
   return AccCacheIterator;
 }
 void MemoryManager::EntryErase(uint64_t CpuPtr)
@@ -80,9 +79,28 @@ void MemoryManager::EntryErase(uint64_t CpuPtr)
   auto AccCache = EntryLookup(CpuPtr);
   AccViewTable.erase(CpuPtr);
 }
+/////////////////////////////////////////////////////////////////////////////////
+// LRU membership invariant:
+//
+//    LRU_valid == 1   <=>   AccPtr != NULL  &&  accLock == 0  &&  cpuLock == 0
+//
+// i.e. the LRU queue contains exactly the device-resident, completely unlocked
+// entries -- the evictable set.  Membership is maintained EAGERLY at the lock
+// 0<->1 edges, O(1) via the stored LRU_entry iterator:
+//
+//    AcceleratorViewOpen   lock 0->1 : LRUremove   (gated on LRU_valid)
+//    AcceleratorViewClose  accLock->0: LRUinsert   (AccPtr necessarily exists)
+//    CpuViewOpen           lock 0->1 : LRUremove   (gated on LRU_valid)
+//    CpuViewClose          cpuLock->0: LRUinsert   (iff AccPtr exists)
+//    Evict/AccDiscard                : LRUremove   (frees the device copy)
+//
+// Consequences: victims taken from LRU.back() are evictable by construction;
+// Evict() on a locked entry is an invariant violation (asserted), and the
+// eviction loops (EvictVictims/EvictAll) cannot spin.
+/////////////////////////////////////////////////////////////////////////////////
 void  MemoryManager::LRUinsert(AcceleratorViewEntry &AccCache)
 {
-  assert(AccCache.LRU_valid==0);
+  GRID_ASSERT(AccCache.LRU_valid==0);
   if (AccCache.transient) { 
     LRU.push_back(AccCache.CpuPtr);
     AccCache.LRU_entry = --LRU.end();
@@ -95,7 +113,7 @@ void  MemoryManager::LRUinsert(AcceleratorViewEntry &AccCache)
 }
 void  MemoryManager::LRUremove(AcceleratorViewEntry &AccCache)
 {
-  assert(AccCache.LRU_valid==1);
+  GRID_ASSERT(AccCache.LRU_valid==1);
   LRU.erase(AccCache.LRU_entry);
   AccCache.LRU_valid = 0;
   DeviceLRUBytes-=AccCache.bytes;
@@ -109,19 +127,19 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
   // Remove from Accelerator, remove entry, without flush
   // Cannot be locked. If allocated Must be in LRU pool.
   ///////////////////////////////////////////////////////////
-  assert(AccCache.state!=Empty);
+  GRID_ASSERT(AccCache.state!=Empty);
   
-  dprintf("MemoryManager: Discard(%lx) %lx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
-  assert(AccCache.accLock==0);
-  assert(AccCache.cpuLock==0);
-  assert(AccCache.CpuPtr!=(uint64_t)NULL);
+  dprintf("MemoryManager: Discard(%lx) %lx",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
+  GRID_ASSERT(AccCache.accLock==0);
+  GRID_ASSERT(AccCache.cpuLock==0);
+  GRID_ASSERT(AccCache.CpuPtr!=(uint64_t)NULL);
   if(AccCache.AccPtr) {
     AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
     DeviceDestroy++;
     DeviceBytes   -=AccCache.bytes;
     LRUremove(AccCache);
     AccCache.AccPtr=(uint64_t) NULL;
-    dprintf("MemoryManager: Free(%lx) LRU %ld Total %ld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
+    dprintf("MemoryManager: Free(%lx) LRU %ld Total %ld",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
   }
   uint64_t CpuPtr = AccCache.CpuPtr;
   EntryErase(CpuPtr);
@@ -131,21 +149,21 @@ void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
 {
   ///////////////////////////////////////////////////////////////////////////
   // Make CPU consistent, remove from Accelerator, remove from LRU, LEAVE CPU only entry
-  // Cannot be acclocked. If allocated must be in LRU pool.
+  // Cannot be locked. If allocated must be in LRU pool.
   //
-  // Nov 2022... Felix issue: Allocating two CpuPtrs, can have an entry in LRU-q with CPUlock.
-  //                          and require to evict the AccPtr copy. Eviction was a mistake in CpuViewOpen
-  //                          but there is a weakness where CpuLock entries are attempted for erase
-  //                          Take these OUT LRU queue when CPU locked?
-  //                          Cannot take out the table as cpuLock data is important.
+  // (Historical: a Nov 2022 incident (two CpuPtrs; eviction called from
+  //  CpuViewOpen -- since excised) could present a cpuLocked entry here, and
+  //  silent-return guards were added.  The LRU membership invariant (see
+  //  LRUinsert) now excludes ALL locked entries from the queue eagerly at the
+  //  lock edges, so a locked victim is an invariant violation: asserted.)
   ///////////////////////////////////////////////////////////////////////////
-  assert(AccCache.state!=Empty);
+  GRID_ASSERT(AccCache.state!=Empty);
   
-  mprintf("MemoryManager: Evict CpuPtr %lx AccPtr %lx cpuLock %ld accLock %ld\n",
+  mprintf("MemoryManager: Evict CpuPtr %lx AccPtr %lx cpuLock %ld accLock %ld",
 	  (uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr,
 	  (uint64_t)AccCache.cpuLock,(uint64_t)AccCache.accLock); 
-  if (AccCache.accLock!=0) return;
-  if (AccCache.cpuLock!=0) return;
+  GRID_ASSERT(AccCache.accLock==0);
+  GRID_ASSERT(AccCache.cpuLock==0);
   if(AccCache.state==AccDirty) {
     Flush(AccCache);
   }
@@ -155,7 +173,7 @@ void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
     AccCache.AccPtr=(uint64_t)NULL;
     AccCache.state=CpuDirty; // CPU primary now
     DeviceBytes   -=AccCache.bytes;
-    dprintf("MemoryManager: Free(AccPtr %lx) footprint now %ld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);  
+    dprintf("MemoryManager: Free(AccPtr %lx) footprint now %ld ",(uint64_t)AccCache.AccPtr,DeviceBytes);  
   }
   //  uint64_t CpuPtr = AccCache.CpuPtr;
   DeviceEvictions++;
@@ -163,28 +181,30 @@ void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
 }
 void MemoryManager::Flush(AcceleratorViewEntry &AccCache)
 {
-  assert(AccCache.state==AccDirty);
-  assert(AccCache.cpuLock==0);
-  assert(AccCache.accLock==0);
-  assert(AccCache.AccPtr!=(uint64_t)NULL);
-  assert(AccCache.CpuPtr!=(uint64_t)NULL);
+  GRID_ASSERT(AccCache.state==AccDirty);
+  GRID_ASSERT(AccCache.cpuLock==0);
+  GRID_ASSERT(AccCache.accLock==0);
+  GRID_ASSERT(AccCache.AccPtr!=(uint64_t)NULL);
+  GRID_ASSERT(AccCache.CpuPtr!=(uint64_t)NULL);
   acceleratorCopyFromDevice((void *)AccCache.AccPtr,(void *)AccCache.CpuPtr,AccCache.bytes);
-  mprintf("MemoryManager: acceleratorCopyFromDevice Flush AccPtr %lx -> CpuPtr %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
+  mprintf("MemoryManager: acceleratorCopyFromDevice Flush size %ld AccPtr %lx -> CpuPtr %lx",(uint64_t)AccCache.bytes,(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   DeviceToHostBytes+=AccCache.bytes;
   DeviceToHostXfer++;
   AccCache.state=Consistent;
 }
 void MemoryManager::Clone(AcceleratorViewEntry &AccCache)
 {
-  assert(AccCache.state==CpuDirty);
-  assert(AccCache.cpuLock==0);
-  assert(AccCache.accLock==0);
-  assert(AccCache.CpuPtr!=(uint64_t)NULL);
+  GRID_ASSERT(AccCache.state==CpuDirty);
+  GRID_ASSERT(AccCache.cpuLock==0);
+  GRID_ASSERT(AccCache.accLock==0);
+  GRID_ASSERT(AccCache.CpuPtr!=(uint64_t)NULL);
   if(AccCache.AccPtr==(uint64_t)NULL){
     AccCache.AccPtr=(uint64_t)AcceleratorAllocate(AccCache.bytes);
     DeviceBytes+=AccCache.bytes;
   }
-  mprintf("MemoryManager: acceleratorCopyToDevice   Clone AccPtr %lx <- CpuPtr %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
+  mprintf("MemoryManager: acceleratorCopyToDevice   Clone size %ld AccPtr %lx <- CpuPtr %lx",
+	  (uint64_t)AccCache.bytes,
+	  (uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   acceleratorCopyToDevice((void *)AccCache.CpuPtr,(void *)AccCache.AccPtr,AccCache.bytes);
   HostToDeviceBytes+=AccCache.bytes;
   HostToDeviceXfer++;
@@ -193,10 +213,10 @@ void MemoryManager::Clone(AcceleratorViewEntry &AccCache)
 
 void MemoryManager::CpuDiscard(AcceleratorViewEntry &AccCache)
 {
-  assert(AccCache.state!=Empty);
-  assert(AccCache.cpuLock==0);
-  assert(AccCache.accLock==0);
-  assert(AccCache.CpuPtr!=(uint64_t)NULL);
+  GRID_ASSERT(AccCache.state!=Empty);
+  GRID_ASSERT(AccCache.cpuLock==0);
+  GRID_ASSERT(AccCache.accLock==0);
+  GRID_ASSERT(AccCache.CpuPtr!=(uint64_t)NULL);
   if(AccCache.AccPtr==(uint64_t)NULL){
     AccCache.AccPtr=(uint64_t)AcceleratorAllocate(AccCache.bytes);
     DeviceBytes+=AccCache.bytes;
@@ -210,33 +230,49 @@ void MemoryManager::CpuDiscard(AcceleratorViewEntry &AccCache)
 void MemoryManager::ViewClose(void* Ptr,ViewMode mode)
 {
   if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){
-    dprintf("AcceleratorViewClose %lx\n",(uint64_t)Ptr);
+    dprintf("AcceleratorViewClose %lx",(uint64_t)Ptr);
     AcceleratorViewClose((uint64_t)Ptr);
   } else if( (mode==CpuRead)||(mode==CpuWrite)){
     CpuViewClose((uint64_t)Ptr);
   } else { 
-    assert(0);
+    GRID_ASSERT(0);
   }
 }
 void *MemoryManager::ViewOpen(void* _CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint)
 {
   uint64_t CpuPtr = (uint64_t)_CpuPtr;
   if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){
-    dprintf("AcceleratorViewOpen %lx\n",(uint64_t)CpuPtr);
+    dprintf("AcceleratorViewOpen %lx",(uint64_t)CpuPtr);
     return (void *) AcceleratorViewOpen(CpuPtr,bytes,mode,hint);
   } else if( (mode==CpuRead)||(mode==CpuWrite)){
     return (void *)CpuViewOpen(CpuPtr,bytes,mode,hint);
   } else { 
-    assert(0);
+    GRID_ASSERT(0);
     return NULL;
   }
 }
 void  MemoryManager::EvictVictims(uint64_t bytes)
 {
-  assert(bytes<DeviceMaxBytes);
+  if(bytes>=DeviceMaxBytes) {
+    printf("EvictVictims bytes %ld DeviceMaxBytes %ld\n",bytes,DeviceMaxBytes);
+  }
+  GRID_ASSERT(bytes<DeviceMaxBytes);
   while(bytes+DeviceLRUBytes > DeviceMaxBytes){
     if ( DeviceLRUBytes > 0){
-      assert(LRU.size()>0);
+      GRID_ASSERT(LRU.size()>0);
+      uint64_t victim = LRU.back(); // From the LRU
+      auto AccCacheIterator = EntryLookup(victim);
+      auto & AccCache = AccCacheIterator->second;
+      Evict(AccCache);
+    } else {
+      return;
+    }
+  }
+}
+void  MemoryManager::EvictAll(void)
+{
+  while(LRU.size()>0){
+    if ( DeviceLRUBytes > 0){
       uint64_t victim = LRU.back(); // From the LRU
       auto AccCacheIterator = EntryLookup(victim);
       auto & AccCache = AccCacheIterator->second;
@@ -260,19 +296,19 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
   if (!AccCache.AccPtr) {
     EvictVictims(bytes); 
   } 
-  assert((mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard));
+  GRID_ASSERT((mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard));
 
-  assert(AccCache.cpuLock==0);  // Programming error
+  GRID_ASSERT(AccCache.cpuLock==0);  // Programming error
 
   if(AccCache.state!=Empty) {
-    dprintf("ViewOpen found entry %lx %lx : %ld %ld accLock %ld\n",
+    dprintf("ViewOpen found entry %lx %lx : sizes %ld %ld accLock %ld",
 		    (uint64_t)AccCache.CpuPtr,
 		    (uint64_t)CpuPtr,
 		    (uint64_t)AccCache.bytes,
 	            (uint64_t)bytes,
 		    (uint64_t)AccCache.accLock);
-    assert(AccCache.CpuPtr == CpuPtr);
-    assert(AccCache.bytes  ==bytes);
+    GRID_ASSERT(AccCache.CpuPtr == CpuPtr);
+    GRID_ASSERT(AccCache.bytes  ==bytes);
   }
 /*
  *  State transitions and actions
@@ -289,7 +325,7 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
  *  AccWrite AccDirty   AccDirty       -        - 
  */
   if(AccCache.state==Empty) {
-    assert(AccCache.LRU_valid==0);
+    GRID_ASSERT(AccCache.LRU_valid==0);
     AccCache.CpuPtr = CpuPtr;
     AccCache.AccPtr = (uint64_t)NULL;
     AccCache.bytes  = bytes;
@@ -305,7 +341,7 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
       AccCache.state  = Consistent; // Empty + AccRead => Consistent
     }
     AccCache.accLock= 1;
-    dprintf("Copied Empty entry into device accLock= %d\n",AccCache.accLock);
+    dprintf("Copied Empty entry into device accLock= %d",AccCache.accLock);
   } else if(AccCache.state==CpuDirty ){
     if(mode==AcceleratorWriteDiscard) {
       CpuDiscard(AccCache);
@@ -318,30 +354,30 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
       AccCache.state  = Consistent; // CpuDirty + AccRead => Consistent
     }
     AccCache.accLock++;
-    dprintf("CpuDirty entry into device ++accLock= %d\n",AccCache.accLock);
+    dprintf("CpuDirty entry into device ++accLock= %d",AccCache.accLock);
   } else if(AccCache.state==Consistent) {
     if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard))
       AccCache.state  = AccDirty;   // Consistent + AcceleratorWrite=> AccDirty
     else
       AccCache.state  = Consistent; // Consistent + AccRead => Consistent
     AccCache.accLock++;
-    dprintf("Consistent entry into device ++accLock= %d\n",AccCache.accLock);
+    dprintf("Consistent entry into device ++accLock= %d",AccCache.accLock);
   } else if(AccCache.state==AccDirty) {
     if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard))
       AccCache.state  = AccDirty; // AccDirty + AcceleratorWrite=> AccDirty
     else
       AccCache.state  = AccDirty; // AccDirty + AccRead => AccDirty
     AccCache.accLock++;
-    dprintf("AccDirty entry ++accLock= %d\n",AccCache.accLock);
+    dprintf("AccDirty entry ++accLock= %d",AccCache.accLock);
   } else {
-    assert(0);
+    GRID_ASSERT(0);
   }
 
-  assert(AccCache.accLock>0);
+  GRID_ASSERT(AccCache.accLock>0);
   // If view is opened on device must remove from LRU
   if(AccCache.LRU_valid==1){
     // must possibly remove from LRU as now locked on GPU
-    dprintf("AccCache entry removed from LRU \n");
+    dprintf("AccCache entry removed from LRU ");
     LRUremove(AccCache);
   }
 
@@ -358,16 +394,16 @@ void MemoryManager::AcceleratorViewClose(uint64_t CpuPtr)
   auto AccCacheIterator = EntryLookup(CpuPtr);
   auto & AccCache = AccCacheIterator->second;
 
-  assert(AccCache.cpuLock==0);
-  assert(AccCache.accLock>0);
+  GRID_ASSERT(AccCache.cpuLock==0);
+  GRID_ASSERT(AccCache.accLock>0);
 
   AccCache.accLock--;
   // Move to LRU queue if not locked and close on device
   if(AccCache.accLock==0) {
-    dprintf("AccleratorViewClose %lx AccLock decremented to %ld move to LRU queue\n",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
+    dprintf("AccleratorViewClose %lx AccLock decremented to %ld move to LRU queue",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
     LRUinsert(AccCache);
   } else {
-    dprintf("AccleratorViewClose %lx AccLock decremented to %ld\n",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
+    dprintf("AccleratorViewClose %lx AccLock decremented to %ld",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
   }
 }
 void MemoryManager::CpuViewClose(uint64_t CpuPtr)
@@ -375,10 +411,17 @@ void MemoryManager::CpuViewClose(uint64_t CpuPtr)
   auto AccCacheIterator = EntryLookup(CpuPtr);
   auto & AccCache = AccCacheIterator->second;
 
-  assert(AccCache.cpuLock>0);
-  assert(AccCache.accLock==0);
+  GRID_ASSERT(AccCache.cpuLock>0);
+  GRID_ASSERT(AccCache.accLock==0);
 
   AccCache.cpuLock--;
+  // Return to LRU queue when fully unlocked -- mirrors AcceleratorViewClose.
+  // Asymmetry vs the Acc side: a device copy need not exist for a host view;
+  // only device-resident entries belong in the (evictable) LRU queue.
+  if( (AccCache.cpuLock==0) && (AccCache.AccPtr!=(uint64_t)NULL) ) {
+    dprintf("CpuViewClose %lx cpuLock decremented to zero, move to LRU queue",(uint64_t)CpuPtr);
+    LRUinsert(AccCache);
+  }
 }
 /*
  *  Action  State   StateNext         Flush    Clone
@@ -409,12 +452,12 @@ uint64_t MemoryManager::CpuViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,V
   //    EvictVictims(bytes);
   //  }
 
-  assert((mode==CpuRead)||(mode==CpuWrite));
-  assert(AccCache.accLock==0);  // Programming error
+  GRID_ASSERT((mode==CpuRead)||(mode==CpuWrite));
+  GRID_ASSERT(AccCache.accLock==0);  // Programming error
 
   if(AccCache.state!=Empty) {
-    assert(AccCache.CpuPtr == CpuPtr);
-    assert(AccCache.bytes==bytes);
+    GRID_ASSERT(AccCache.CpuPtr == CpuPtr);
+    GRID_ASSERT(AccCache.bytes==bytes);
   }
 
   if(AccCache.state==Empty) {
@@ -429,20 +472,28 @@ uint64_t MemoryManager::CpuViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,V
     AccCache.state = CpuDirty; // CpuDirty +CpuRead/CpuWrite => CpuDirty
     AccCache.cpuLock++;
   } else if(AccCache.state==Consistent) {
-    assert(AccCache.AccPtr != (uint64_t)NULL);
+    GRID_ASSERT(AccCache.AccPtr != (uint64_t)NULL);
     if(mode==CpuWrite)
       AccCache.state = CpuDirty;   // Consistent +CpuWrite => CpuDirty
     else 
       AccCache.state = Consistent; // Consistent +CpuRead  => Consistent
     AccCache.cpuLock++;
   } else if(AccCache.state==AccDirty) {
-    assert(AccCache.AccPtr != (uint64_t)NULL);
+    GRID_ASSERT(AccCache.AccPtr != (uint64_t)NULL);
     Flush(AccCache);
     if(mode==CpuWrite) AccCache.state = CpuDirty;   // AccDirty +CpuWrite => CpuDirty, Flush
     else            AccCache.state = Consistent; // AccDirty +CpuRead  => Consistent, Flush
     AccCache.cpuLock++;
   } else {
-    assert(0); // should be unreachable
+    GRID_ASSERT(0); // should be unreachable
+  }
+
+  GRID_ASSERT(AccCache.cpuLock>0);
+  // If view is opened on host must remove from LRU -- mirrors AcceleratorViewOpen.
+  // LRU_valid==1 here implies this is the 0->1 lock edge of a device-resident entry.
+  if(AccCache.LRU_valid==1){
+    dprintf("CpuViewOpen: entry removed from LRU ");
+    LRUremove(AccCache);
   }
 
   AccCache.transient= transient? EvictNext : 0;
@@ -524,12 +575,12 @@ void MemoryManager::Audit(std::string s)
   std::cout << " Memory Manager::Audit() from "<<s<<std::endl;
   for(auto it=LRU.begin();it!=LRU.end();it++){
     uint64_t cpuPtr = *it;
-    assert(EntryPresent(cpuPtr));
+    GRID_ASSERT(EntryPresent(cpuPtr));
     auto AccCacheIterator = EntryLookup(cpuPtr);
     auto & AccCache = AccCacheIterator->second;
     LruBytes2+=AccCache.bytes;
-    assert(AccCache.LRU_valid==1);
-    assert(AccCache.LRU_entry==it);
+    GRID_ASSERT(AccCache.LRU_valid==1);
+    GRID_ASSERT(AccCache.LRU_entry==it);
   }
   std::cout << " Memory Manager::Audit() LRU queue matches table entries "<<std::endl;
 
@@ -548,7 +599,7 @@ void MemoryManager::Audit(std::string s)
     if( AccCache.LRU_valid ) LruCnt++;
     
     if ( AccCache.cpuLock || AccCache.accLock ) {
-      assert(AccCache.LRU_valid==0);
+      GRID_ASSERT(AccCache.LRU_valid==0);
 
       std::cout << GridLogError << s<< "\n\t 0x"<<std::hex<<AccCache.CpuPtr<<std::dec
 		<< "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
@@ -557,16 +608,16 @@ void MemoryManager::Audit(std::string s)
 		<< "\t LRUvalid " << AccCache.LRU_valid<<std::endl;
     }
 
-    assert( AccCache.cpuLock== 0 ) ;
-    assert( AccCache.accLock== 0 ) ;
+    GRID_ASSERT( AccCache.cpuLock== 0 ) ;
+    GRID_ASSERT( AccCache.accLock== 0 ) ;
   }
   std::cout << " Memory Manager::Audit() no locked table entries "<<std::endl;
-  assert(LruBytes1==LruBytes2);
-  assert(LruBytes1==DeviceLRUBytes);
+  GRID_ASSERT(LruBytes1==LruBytes2);
+  GRID_ASSERT(LruBytes1==DeviceLRUBytes);
   std::cout << " Memory Manager::Audit() evictable bytes matches sum over table "<<std::endl;
-  assert(AccBytes==DeviceBytes);
+  GRID_ASSERT(AccBytes==DeviceBytes);
   std::cout << " Memory Manager::Audit() device bytes matches sum over table "<<std::endl;
-  assert(LruCnt == LRU.size());
+  GRID_ASSERT(LruCnt == LRU.size());
   std::cout << " Memory Manager::Audit() LRU entry count matches "<<std::endl;
 
 }

Grid/allocator/MemoryStats.cc

@@ -10,16 +10,16 @@ void check_huge_pages(void *Buf,uint64_t BYTES)
 {
 #ifdef __linux__
   int fd = open("/proc/self/pagemap", O_RDONLY);
-  assert(fd >= 0);
+  GRID_ASSERT(fd >= 0);
   const int page_size = 4096;
   uint64_t virt_pfn = (uint64_t)Buf / page_size;
   off_t offset = sizeof(uint64_t) * virt_pfn;
   uint64_t npages = (BYTES + page_size-1) / page_size;
-  uint64_t pagedata[npages];
+  std::vector<uint64_t> pagedata(npages);
   uint64_t ret = lseek(fd, offset, SEEK_SET);
-  assert(ret == offset);
-  ret = ::read(fd, pagedata, sizeof(uint64_t)*npages);
-  assert(ret == sizeof(uint64_t) * npages);
+  GRID_ASSERT(ret == offset);
+  ret = ::read(fd, &pagedata[0], sizeof(uint64_t)*npages);
+  GRID_ASSERT(ret == sizeof(uint64_t) * npages);
   int nhugepages = npages / 512;
   int n4ktotal, nnothuge;
   n4ktotal = 0;

Grid/cartesian/Cartesian_base.h

@@ -165,7 +165,7 @@ public:
     //
     if ( _simd_layout[dimension] > 2 ) { 
       for(int d=0;d<_ndimension;d++){
-	if ( d != dimension ) assert ( (_simd_layout[d]==1)  );
+	if ( d != dimension ) GRID_ASSERT ( (_simd_layout[d]==1)  );
       }
       permute_type = RotateBit; // How to specify distance; this is not just direction.
       return permute_type;
@@ -187,7 +187,7 @@ public:
   inline int64_t gSites(void) const { return (int64_t)_isites*(int64_t)_osites*(int64_t)_Nprocessors; }; 
   inline int Nd    (void) const { return _ndimension;};
 
-  inline const Coordinate LocalStarts(void)             { return _lstart;    };
+  inline const Coordinate &LocalStarts(void)            { return _lstart;    };
   inline const Coordinate &FullDimensions(void)         { return _fdimensions;};
   inline const Coordinate &GlobalDimensions(void)       { return _gdimensions;};
   inline const Coordinate &LocalDimensions(void)        { return _ldimensions;};
@@ -216,11 +216,11 @@ public:
   // Global addressing
   ////////////////////////////////////////////////////////////////
   void GlobalIndexToGlobalCoor(int64_t gidx,Coordinate &gcoor){
-    assert(gidx< gSites());
+    GRID_ASSERT(gidx< gSites());
     Lexicographic::CoorFromIndex(gcoor,gidx,_gdimensions);
   }
   void LocalIndexToLocalCoor(int lidx,Coordinate &lcoor){
-    assert(lidx<lSites());
+    GRID_ASSERT(lidx<lSites());
     Lexicographic::CoorFromIndex(lcoor,lidx,_ldimensions);
   }
   void GlobalCoorToGlobalIndex(const Coordinate & gcoor,int64_t & gidx){

Grid/cartesian/Cartesian_full.h

@@ -128,10 +128,10 @@ public:
         // Use a reduced simd grid
         _ldimensions[d] = _gdimensions[d] / _processors[d]; //local dimensions
         //std::cout << _ldimensions[d] << "  " << _gdimensions[d] << "  " << _processors[d] << std::endl;
-        assert(_ldimensions[d] * _processors[d] == _gdimensions[d]);
+        GRID_ASSERT(_ldimensions[d] * _processors[d] == _gdimensions[d]);
 
         _rdimensions[d] = _ldimensions[d] / _simd_layout[d]; //overdecomposition
-        assert(_rdimensions[d] * _simd_layout[d] == _ldimensions[d]);
+        GRID_ASSERT(_rdimensions[d] * _simd_layout[d] == _ldimensions[d]);
 
         _lstart[d] = _processor_coor[d] * _ldimensions[d];
         _lend[d] = _processor_coor[d] * _ldimensions[d] + _ldimensions[d] - 1;

Grid/cartesian/Cartesian_red_black.h

@@ -67,7 +67,7 @@ public:
   }
   virtual int CheckerBoard(const Coordinate &site){
     int linear=0;
-    assert(site.size()==_ndimension);
+    GRID_ASSERT(site.size()==_ndimension);
     for(int d=0;d<_ndimension;d++){ 
       if(_checker_dim_mask[d])
 	linear=linear+site[d];
@@ -160,11 +160,11 @@ public:
 
       _isCheckerBoarded = true;
     _checker_dim = checker_dim;
-    assert(checker_dim_mask[checker_dim] == 1);
+    GRID_ASSERT(checker_dim_mask[checker_dim] == 1);
     _ndimension = dimensions.size();
-    assert(checker_dim_mask.size() == _ndimension);
-    assert(processor_grid.size() == _ndimension);
-    assert(simd_layout.size() == _ndimension);
+    GRID_ASSERT(checker_dim_mask.size() == _ndimension);
+    GRID_ASSERT(processor_grid.size() == _ndimension);
+    GRID_ASSERT(simd_layout.size() == _ndimension);
 
     _fdimensions.resize(_ndimension);
     _gdimensions.resize(_ndimension);
@@ -190,20 +190,20 @@ public:
 
         if (d == _checker_dim)
 	  {
-	    assert((_gdimensions[d] & 0x1) == 0);
+	    GRID_ASSERT((_gdimensions[d] & 0x1) == 0);
 	    _gdimensions[d] = _gdimensions[d] / 2; // Remove a checkerboard
 	    _gsites /= 2;
 	  }
         _ldimensions[d] = _gdimensions[d] / _processors[d];
-        assert(_ldimensions[d] * _processors[d] == _gdimensions[d]);
+        GRID_ASSERT(_ldimensions[d] * _processors[d] == _gdimensions[d]);
         _lstart[d] = _processor_coor[d] * _ldimensions[d];
         _lend[d] = _processor_coor[d] * _ldimensions[d] + _ldimensions[d] - 1;
 
         // Use a reduced simd grid
         _simd_layout[d] = simd_layout[d];
         _rdimensions[d] = _ldimensions[d] / _simd_layout[d]; // this is not checking if this is integer
-        assert(_rdimensions[d] * _simd_layout[d] == _ldimensions[d]);
-        assert(_rdimensions[d] > 0);
+        GRID_ASSERT(_rdimensions[d] * _simd_layout[d] == _ldimensions[d]);
+        GRID_ASSERT(_rdimensions[d] > 0);
 
         // all elements of a simd vector must have same checkerboard.
         // If Ls vectorised, this must still be the case; e.g. dwf rb5d

Grid/communicator/Communicator_base.cc

@@ -57,18 +57,29 @@ int                      CartesianCommunicator::ProcessorCount(void)    { return
 // very VERY rarely (Log, serial RNG) we need world without a grid
 ////////////////////////////////////////////////////////////////////////////////
 
+#ifdef USE_GRID_REDUCTION
+void CartesianCommunicator::GlobalSum(ComplexF &c)
+{
+  GlobalSumP2P(c);
+}
+void CartesianCommunicator::GlobalSum(ComplexD &c)
+{
+  GlobalSumP2P(c);
+}
+#else
 void CartesianCommunicator::GlobalSum(ComplexF &c)
 {
   GlobalSumVector((float *)&c,2);
 }
-void CartesianCommunicator::GlobalSumVector(ComplexF *c,int N)
-{
-  GlobalSumVector((float *)c,2*N);
-}
 void CartesianCommunicator::GlobalSum(ComplexD &c)
 {
   GlobalSumVector((double *)&c,2);
 }
+#endif
+void CartesianCommunicator::GlobalSumVector(ComplexF *c,int N)
+{
+  GlobalSumVector((float *)c,2*N);
+}
 void CartesianCommunicator::GlobalSumVector(ComplexD *c,int N)
 {
   GlobalSumVector((double *)c,2*N);

Grid/communicator/Communicator_base.h

@@ -33,6 +33,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 ///////////////////////////////////
 #include <Grid/communicator/SharedMemory.h>
 
+#define NVLINK_GET
+
 NAMESPACE_BEGIN(Grid);
 
 extern bool Stencil_force_mpi ;
@@ -106,7 +108,7 @@ public:
   // very VERY rarely (Log, serial RNG) we need world without a grid
   ////////////////////////////////////////////////////////////////////////////////
   static int  RankWorld(void) ;
-  static void BroadcastWorld(int root,void* data, int bytes);
+  static void BroadcastWorld(int root,void* data, uint64_t bytes);
   static void BarrierWorld(void);
   
   ////////////////////////////////////////////////////////////
@@ -136,7 +138,7 @@ public:
     for(int d=0;d<_ndimension;d++){
       column.resize(_processors[d]);
       column[0] = accum;
-      std::vector<CommsRequest_t> list;
+      std::vector<MpiCommsRequest_t> list;
       for(int p=1;p<_processors[d];p++){
 	ShiftedRanks(d,p,source,dest);
 	SendToRecvFromBegin(list,
@@ -147,7 +149,8 @@ public:
 			    sizeof(obj),d*100+p);
 
       }
-      CommsComplete(list);
+      if (!list.empty()) // avoid triggering GRID_ASSERT in comms == none
+	CommsComplete(list);
       for(int p=1;p<_processors[d];p++){
 	accum = accum + column[p];
       }
@@ -166,32 +169,44 @@ public:
   ////////////////////////////////////////////////////////////
   // Face exchange, buffer swap in translational invariant way
   ////////////////////////////////////////////////////////////
-  void CommsComplete(std::vector<CommsRequest_t> &list);
-  void SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+  void CommsComplete(std::vector<MpiCommsRequest_t> &list);
+  void SendToRecvFromBegin(std::vector<MpiCommsRequest_t> &list,
 			   void *xmit,
 			   int dest,
 			   void *recv,
 			   int from,
-			   int bytes,int dir);
+			   uint64_t bytes,int dir);
   
   void SendToRecvFrom(void *xmit,
 		      int xmit_to_rank,
 		      void *recv,
 		      int recv_from_rank,
-		      int bytes);
+		      uint64_t bytes);
   
+  int IsOffNode(int rank);
   double StencilSendToRecvFrom(void *xmit,
 			       int xmit_to_rank,int do_xmit,
 			       void *recv,
 			       int recv_from_rank,int do_recv,
-			       int bytes,int dir);
+			       uint64_t bytes,int dir);
+
+  double StencilSendToRecvFromPrepare(std::vector<CommsRequest_t> &list,
+				      void *xmit,
+				      int xmit_to_rank,int do_xmit,
+				      void *recv,
+				      int recv_from_rank,int do_recv,
+				      uint64_t xbytes,uint64_t rbytes,int dir);
+
+  // Could do a PollHtoD and have a CommsMerge dependence
+  void StencilSendToRecvFromPollDtoH (std::vector<CommsRequest_t> &list);
+  void StencilSendToRecvFromPollIRecv(std::vector<CommsRequest_t> &list);
 
   double StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
-				    void *xmit,
+				    void *xmit,void *xmit_comp,
 				    int xmit_to_rank,int do_xmit,
-				    void *recv,
+				    void *recv,void *recv_comp,
 				    int recv_from_rank,int do_recv,
-				    int xbytes,int rbytes,int dir);
+				    uint64_t xbytes,uint64_t rbytes,int dir);
   
   
   void StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int i);
@@ -205,24 +220,34 @@ public:
   ////////////////////////////////////////////////////////////
   // Broadcast a buffer and composite larger
   ////////////////////////////////////////////////////////////
-  void Broadcast(int root,void* data, int bytes);
+  void Broadcast(int root,void* data, uint64_t bytes);
 
   ////////////////////////////////////////////////////////////
   // All2All down one dimension
   ////////////////////////////////////////////////////////////
   template<class T> void AllToAll(int dim,std::vector<T> &in, std::vector<T> &out){
-    assert(dim>=0);
-    assert(dim<_ndimension);
-    assert(in.size()==out.size());
+    GRID_ASSERT(dim>=0);
+    GRID_ASSERT(dim<_ndimension);
+    GRID_ASSERT(in.size()==out.size());
     int numnode = _processors[dim];
     uint64_t bytes=sizeof(T);
     uint64_t words=in.size()/numnode;
-    assert(numnode * words == in.size());
-    assert(words < (1ULL<<31));
+    GRID_ASSERT(numnode * words == in.size());
+    GRID_ASSERT(words < (1ULL<<31));
     AllToAll(dim,(void *)&in[0],(void *)&out[0],words,bytes);
   }
   void AllToAll(int dim  ,void *in,void *out,uint64_t words,uint64_t bytes);
   void AllToAll(void  *in,void *out,uint64_t words         ,uint64_t bytes);
+
+  ////////////////////////////////////////////////////////////////////////////
+  // 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

@@ -28,9 +28,17 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/GridCore.h>
 #include <Grid/communicator/SharedMemory.h>
 
+void GridAbort(void) { MPI_Abort(MPI_COMM_WORLD,SIGABRT); }
+extern void * Grid_backtrace_buffer[_NBACKTRACE];
+
 NAMESPACE_BEGIN(Grid);
 
+
 Grid_MPI_Comm       CartesianCommunicator::communicator_world;
+#ifdef GRID_CHECKSUM_COMMS
+uint64_t checksum_index = 1;
+#endif
+
 
 ////////////////////////////////////////////
 // First initialise of comms system
@@ -55,11 +63,11 @@ void CartesianCommunicator::Init(int *argc, char ***argv)
 #endif
     //If only 1 comms thread we require any threading mode other than SINGLE, but for multiple comms threads we need MULTIPLE
     if( (nCommThreads == 1) && (provided == MPI_THREAD_SINGLE) ) {
-      assert(0);
+      GRID_ASSERT(0);
     }
 
     if( (nCommThreads > 1) && (provided != MPI_THREAD_MULTIPLE) ) {
-      assert(0);
+      GRID_ASSERT(0);
     }
   }
 
@@ -80,20 +88,20 @@ void CartesianCommunicator::Init(int *argc, char ***argv)
 void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
 {
   int ierr=MPI_Cart_shift(communicator,dim,shift,&source,&dest);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
 }
 int CartesianCommunicator::RankFromProcessorCoor(Coordinate &coor)
 {
   int rank;
   int ierr=MPI_Cart_rank  (communicator, &coor[0], &rank);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
   return rank;
 }
 void  CartesianCommunicator::ProcessorCoorFromRank(int rank, Coordinate &coor)
 {
   coor.resize(_ndimension);
   int ierr=MPI_Cart_coords  (communicator, rank, _ndimension,&coor[0]);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
 }
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -120,8 +128,8 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
 //////////////////////////////////
 CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank)
 {
-  _ndimension = processors.size();  assert(_ndimension>=1);
-  int parent_ndimension = parent._ndimension; assert(_ndimension >= parent._ndimension);
+  _ndimension = processors.size();  GRID_ASSERT(_ndimension>=1);
+  int parent_ndimension = parent._ndimension; GRID_ASSERT(_ndimension >= parent._ndimension);
   Coordinate parent_processor_coor(_ndimension,0);
   Coordinate parent_processors    (_ndimension,1);
   Coordinate shm_processors       (_ndimension,1);
@@ -145,7 +153,7 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
     childsize *= processors[d];
   }
   int Nchild = Nparent/childsize;
-  assert (childsize * Nchild == Nparent);
+  GRID_ASSERT (childsize * Nchild == Nparent);
 
   Coordinate ccoor(_ndimension); // coor within subcommunicator
   Coordinate scoor(_ndimension); // coor of split within parent
@@ -171,12 +179,12 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
     // Split the communicator
     ////////////////////////////////////////////////////////////////
     int ierr= MPI_Comm_split(parent.communicator,srank,crank,&comm_split);
-    assert(ierr==0);
+    GRID_ASSERT(ierr==0);
 
   } else {
     srank = 0;
     int ierr = MPI_Comm_dup (parent.communicator,&comm_split);
-    assert(ierr==0);
+    GRID_ASSERT(ierr==0);
   }
 
   //////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -201,7 +209,7 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
     }
   }
   for(int d=0;d<processors.size();d++){
-    assert(_processor_coor[d] == ccoor[d] );
+    GRID_ASSERT(_processor_coor[d] == ccoor[d] );
   }
 }
 
@@ -243,7 +251,7 @@ void CartesianCommunicator::InitFromMPICommunicator(const Coordinate &processors
   for(int i=0;i<_ndimension*2;i++){
     MPI_Comm_dup(communicator,&communicator_halo[i]);
   }
-  assert(Size==_Nprocessors);
+  GRID_ASSERT(Size==_Nprocessors);
 }
 
 CartesianCommunicator::~CartesianCommunicator()
@@ -257,82 +265,103 @@ CartesianCommunicator::~CartesianCommunicator()
     }
   }
 }
-void CartesianCommunicator::GlobalSum(uint32_t &u){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalSum(uint64_t &u){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalSumVector(uint64_t* u,int N){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,u,N,MPI_UINT64_T,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalXOR(uint32_t &u){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_BXOR,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalXOR(uint64_t &u){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_BXOR,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalMax(float &f)
-{
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_MAX,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalMax(double &d)
-{
-  int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_MAX,communicator);
-  assert(ierr==0);
-}
+#ifdef USE_GRID_REDUCTION
 void CartesianCommunicator::GlobalSum(float &f){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalSumVector(float *f,int N)
-{
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,f,N,MPI_FLOAT,MPI_SUM,communicator);
-  assert(ierr==0);
+  FlightRecorder::StepLog("GlobalSumP2P");
+  CartesianCommunicator::GlobalSumP2P(f);
 }
 void CartesianCommunicator::GlobalSum(double &d)
 {
+  FlightRecorder::StepLog("GlobalSumP2P");
+  CartesianCommunicator::GlobalSumP2P(d);
+}
+#else
+void CartesianCommunicator::GlobalSum(float &f){
+  FlightRecorder::StepLog("AllReduce float");
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
+  GRID_ASSERT(ierr==0);
+}
+void CartesianCommunicator::GlobalSum(double &d)
+{
+  FlightRecorder::StepLog("AllReduce double");
   int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_SUM,communicator);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
+}
+#endif
+void CartesianCommunicator::GlobalSum(uint32_t &u){
+  FlightRecorder::StepLog("AllReduce uint32_t");
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_SUM,communicator);
+  GRID_ASSERT(ierr==0);
+}
+void CartesianCommunicator::GlobalSum(uint64_t &u){
+  FlightRecorder::StepLog("AllReduce uint64_t");
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_SUM,communicator);
+  GRID_ASSERT(ierr==0);
+}
+void CartesianCommunicator::GlobalSumVector(uint64_t* u,int N){
+  FlightRecorder::StepLog("AllReduceVector");
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,u,N,MPI_UINT64_T,MPI_SUM,communicator);
+  GRID_ASSERT(ierr==0);
+}
+void CartesianCommunicator::GlobalXOR(uint32_t &u){
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_BXOR,communicator);
+  GRID_ASSERT(ierr==0);
+}
+void CartesianCommunicator::GlobalXOR(uint64_t &u){
+  FlightRecorder::StepLog("GlobalXOR");
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_BXOR,communicator);
+  GRID_ASSERT(ierr==0);
+}
+void CartesianCommunicator::GlobalMax(float &f)
+{
+  FlightRecorder::StepLog("GlobalMax");
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_MAX,communicator);
+  GRID_ASSERT(ierr==0);
+}
+void CartesianCommunicator::GlobalMax(double &d)
+{
+  FlightRecorder::StepLog("GlobalMax");
+  int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_MAX,communicator);
+  GRID_ASSERT(ierr==0);
+}
+void CartesianCommunicator::GlobalSumVector(float *f,int N)
+{
+  FlightRecorder::StepLog("GlobalSumVector(float *)");
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,f,N,MPI_FLOAT,MPI_SUM,communicator);
+  GRID_ASSERT(ierr==0);
 }
 void CartesianCommunicator::GlobalSumVector(double *d,int N)
 {
+  FlightRecorder::StepLog("GlobalSumVector(double *)");
   int ierr = MPI_Allreduce(MPI_IN_PLACE,d,N,MPI_DOUBLE,MPI_SUM,communicator);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
 }
 
-void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+void CartesianCommunicator::SendToRecvFromBegin(std::vector<MpiCommsRequest_t> &list,
 						void *xmit,
 						int dest,
 						void *recv,
 						int from,
-						int bytes,int dir)
+						uint64_t bytes,int dir)
 {
   MPI_Request xrq;
   MPI_Request rrq;
 
-  assert(dest != _processor);
-  assert(from != _processor);
-
+  GRID_ASSERT(dest != _processor);
+  GRID_ASSERT(from != _processor);
   int tag;
 
   tag= dir+from*32;
-  int ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator,&rrq);
-  assert(ierr==0);
+  int ierr=MPI_Irecv(recv,(int)( bytes/sizeof(int32_t)), MPI_INT32_T,from,tag,communicator,&rrq);
+  GRID_ASSERT(ierr==0);
   list.push_back(rrq);
   
   tag= dir+_processor*32;
-  ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator,&xrq);
-  assert(ierr==0);
+  ierr =MPI_Isend(xmit,(int)(bytes/sizeof(int32_t)), MPI_INT32_T,dest,tag,communicator,&xrq);
+  GRID_ASSERT(ierr==0);
   list.push_back(xrq);
 }
-void CartesianCommunicator::CommsComplete(std::vector<CommsRequest_t> &list)
+void CartesianCommunicator::CommsComplete(std::vector<MpiCommsRequest_t> &list)
 {
   int nreq=list.size();
 
@@ -340,7 +369,7 @@ void CartesianCommunicator::CommsComplete(std::vector<CommsRequest_t> &list)
 
   std::vector<MPI_Status> status(nreq);
   int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
   list.resize(0);
 }
 
@@ -349,50 +378,63 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 					   int dest,
 					   void *recv,
 					   int from,
-					   int bytes)
+					   uint64_t bytes)
 {
-  std::vector<CommsRequest_t> reqs(0);
-  unsigned long  xcrc = crc32(0L, Z_NULL, 0);
-  unsigned long  rcrc = crc32(0L, Z_NULL, 0);
+  std::vector<MpiCommsRequest_t> reqs(0);
 
   int myrank = _processor;
   int ierr;
 
   // Enforce no UVM in comms, device or host OK
-  assert(acceleratorIsCommunicable(xmit));
-  assert(acceleratorIsCommunicable(recv));
+  GRID_ASSERT(acceleratorIsCommunicable(xmit));
+  GRID_ASSERT(acceleratorIsCommunicable(recv));
 
   // Give the CPU to MPI immediately; can use threads to overlap optionally
   //  printf("proc %d SendToRecvFrom %d bytes Sendrecv \n",_processor,bytes);
-  ierr=MPI_Sendrecv(xmit,bytes,MPI_CHAR,dest,myrank,
-		    recv,bytes,MPI_CHAR,from, from,
+  ierr=MPI_Sendrecv(xmit,(int)(bytes/sizeof(int32_t)),MPI_INT32_T,dest,myrank,
+		    recv,(int)(bytes/sizeof(int32_t)),MPI_INT32_T,from, from,
 		    communicator,MPI_STATUS_IGNORE);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
 
-  //  xcrc = crc32(xcrc,(unsigned char *)xmit,bytes);
-  //  rcrc = crc32(rcrc,(unsigned char *)recv,bytes);
-  //  printf("proc %d SendToRecvFrom %d bytes xcrc %lx rcrc %lx\n",_processor,bytes,xcrc,rcrc); fflush
 }
 // Basic Halo comms primitive
 double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
 						     int dest, int dox,
 						     void *recv,
 						     int from, int dor,
-						     int bytes,int dir)
+						     uint64_t bytes,int dir)
 {
   std::vector<CommsRequest_t> list;
-  double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,dox,recv,from,dor,bytes,bytes,dir);
+  double offbytes = StencilSendToRecvFromPrepare(list,xmit,dest,dox,recv,from,dor,bytes,bytes,dir);
+  offbytes       += StencilSendToRecvFromBegin(list,xmit,xmit,dest,dox,recv,recv,from,dor,bytes,bytes,dir);
   StencilSendToRecvFromComplete(list,dir);
   return offbytes;
 }
+int CartesianCommunicator::IsOffNode(int rank)
+{
+  int grank = ShmRanks[rank];
+  if ( grank == MPI_UNDEFINED ) return true;
+  else return false;
+}
 
-#undef NVLINK_GET // Define to use get instead of put DMA
+#ifdef ACCELERATOR_AWARE_MPI
+void CartesianCommunicator::StencilSendToRecvFromPollIRecv(std::vector<CommsRequest_t> &list) {};
+void CartesianCommunicator::StencilSendToRecvFromPollDtoH(std::vector<CommsRequest_t> &list) {};
+double CartesianCommunicator::StencilSendToRecvFromPrepare(std::vector<CommsRequest_t> &list,
+							   void *xmit,
+							   int dest,int dox,
+							   void *recv,
+							   int from,int dor,
+							   uint64_t xbytes,uint64_t rbytes,int dir)
+{
+  return 0.0; // Do nothing -- no preparation required
+}
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
-							 void *xmit,
+							 void *xmit,void *xmit_comp,
 							 int dest,int dox,
-							 void *recv,
+							 void *recv,void *recv_comp,
 							 int from,int dor,
-							 int xbytes,int rbytes,int dir)
+							 uint64_t xbytes,uint64_t rbytes,int dir)
 {
   int ncomm  =communicator_halo.size();
   int commdir=dir%ncomm;
@@ -405,62 +447,431 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
   int gfrom = ShmRanks[from];
   int gme   = ShmRanks[_processor];
 
-  assert(dest != _processor);
-  assert(from != _processor);
-  assert(gme  == ShmRank);
+  GRID_ASSERT(dest != _processor);
+  GRID_ASSERT(from != _processor);
+  GRID_ASSERT(gme  == ShmRank);
   double off_node_bytes=0.0;
   int tag;
-
+  
   if ( dor ) {
     if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
       tag= dir+from*32;
-      ierr=MPI_Irecv(recv, rbytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
-      assert(ierr==0);
+      //      std::cout << " StencilSendToRecvFrom "<<dir<<" MPI_Irecv "<<std::hex<<recv<<std::dec<<std::endl;
+      ierr=MPI_Irecv(recv_comp,(int)(rbytes/sizeof(int32_t)), MPI_INT32_T,from,tag,communicator_halo[commdir],&rrq);
+      GRID_ASSERT(ierr==0);
       list.push_back(rrq);
       off_node_bytes+=rbytes;
     }
 #ifdef NVLINK_GET
+    else { 
       void *shm = (void *) this->ShmBufferTranslate(from,xmit);
-      assert(shm!=NULL);
+      GRID_ASSERT(shm!=NULL);
+      //      std::cout << " StencilSendToRecvFrom "<<dir<<" CopyDeviceToDevice recv "<<std::hex<<recv<<" remote "<<shm <<std::dec<<std::endl;
       acceleratorCopyDeviceToDeviceAsynch(shm,recv,rbytes);
+    }
 #endif
   }
-  
+  // This is a NVLINK PUT  
   if (dox) {
-    //  rcrc = crc32(rcrc,(unsigned char *)recv,bytes);
     if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
       tag= dir+_processor*32;
-      ierr =MPI_Isend(xmit, xbytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
-      assert(ierr==0);
+      ierr =MPI_Isend(xmit_comp,(int)(xbytes/sizeof(int32_t)), MPI_INT32_T,dest,tag,communicator_halo[commdir],&xrq);
+      GRID_ASSERT(ierr==0);
       list.push_back(xrq);
       off_node_bytes+=xbytes;
     } else {
 #ifndef NVLINK_GET
       void *shm = (void *) this->ShmBufferTranslate(dest,recv);
-      assert(shm!=NULL);
+      GRID_ASSERT(shm!=NULL);
       acceleratorCopyDeviceToDeviceAsynch(xmit,shm,xbytes);
 #endif
+    }
+  }
+  return off_node_bytes;
+}
+
+void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
+{
+  int nreq=list.size();
+  /*finishes Get/Put*/
+  acceleratorCopySynchronise();
+
+  if (nreq==0) return;
+  std::vector<MPI_Status> status(nreq);
+  int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
+  GRID_ASSERT(ierr==0);
+  list.resize(0);
+  this->StencilBarrier(); 
+}
+
+#else /* NOT     ... ACCELERATOR_AWARE_MPI */
+///////////////////////////////////////////
+// Pipeline mode through host memory
+///////////////////////////////////////////
+  /*
+   * In prepare (phase 1):
+   * PHASE 1: (prepare)
+   * - post MPI receive buffers asynch
+   * - post device - host send buffer transfer asynch
+   * PHASE 2: (Begin)
+   * - complete all copies
+   * - post MPI send asynch
+   * - post device - device transfers
+   * PHASE 3: (Complete)
+   * - MPI_waitall
+   * - host-device transfers
+   *
+   *********************************
+   * NB could split this further:
+   *--------------------------------
+   * PHASE 1: (Prepare)
+   * - post MPI receive buffers asynch
+   * - post device - host send buffer transfer asynch
+   * PHASE 2: (BeginInterNode)
+   * - complete all copies 
+   * - post MPI send asynch
+   * PHASE 3: (BeginIntraNode)
+   * - post device - device transfers
+   * PHASE 4: (Complete)
+   * - MPI_waitall
+   * - host-device transfers asynch
+   * - (complete all copies) 
+   */
+double CartesianCommunicator::StencilSendToRecvFromPrepare(std::vector<CommsRequest_t> &list,
+							   void *xmit,
+							   int dest,int dox,
+							   void *recv,
+							   int from,int dor,
+							   uint64_t xbytes,uint64_t rbytes,int dir)
+{
+/*
+ * Bring sequence from Stencil.h down to lower level.
+ * Assume using XeLink is ok
+ */  
+  int ncomm  =communicator_halo.size();
+  int commdir=dir%ncomm;
+
+  MPI_Request xrq;
+  MPI_Request rrq;
+
+  int ierr;
+  int gdest = ShmRanks[dest];
+  int gfrom = ShmRanks[from];
+  int gme   = ShmRanks[_processor];
+
+  GRID_ASSERT(dest != _processor);
+  GRID_ASSERT(from != _processor);
+  GRID_ASSERT(gme  == ShmRank);
+  double off_node_bytes=0.0;
+  int tag;
+
+  void * host_recv = NULL;
+  void * host_xmit = NULL;
+
+  /*
+   * PHASE 1: (Prepare)
+   * - post MPI receive buffers asynch
+   * - post device - host send buffer transfer asynch
+   */
+  
+#ifdef GRID_CHECKSUM_COMMS
+  rbytes += 8;
+  xbytes += 8;
+#endif
+
+  if ( dor ) {
+    if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
+      tag= dir+from*32;
+      host_recv = this->HostBufferMalloc(rbytes);
+      ierr=MPI_Irecv(host_recv,(int)(rbytes/sizeof(int32_t)), MPI_INT32_T,from,tag,communicator_halo[commdir],&rrq);
+      GRID_ASSERT(ierr==0);
+      CommsRequest_t srq;
+      srq.PacketType = InterNodeRecv;
+      srq.bytes      = rbytes;
+      srq.req        = rrq;
+      srq.host_buf   = host_recv;
+      srq.device_buf = recv;
+      srq.tag        = tag;
+      list.push_back(srq);
+      off_node_bytes+=rbytes;
+    }
+  }
+  
+  if (dox) {
+    if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
+
+      tag= dir+_processor*32;
+
+      host_xmit = this->HostBufferMalloc(xbytes);
+      CommsRequest_t srq;
+
+#ifdef GRID_CHECKSUM_COMMS
+      uint64_t xbytes_data = xbytes - 8;
+      srq.ev = acceleratorCopyFromDeviceAsynch(xmit, host_xmit,xbytes_data); // Make this Asynch
+      GRID_ASSERT(xbytes % 8 == 0);
+      // flip one bit so that a zero buffer is not consistent
+      uint64_t xsum = checksum_gpu((uint64_t*)xmit, xbytes_data / 8) ^ (checksum_index + 1 + 1000 * tag); 
+      *(uint64_t*)(((char*)host_xmit) + xbytes_data) = xsum;
+#else
+      srq.ev = acceleratorCopyFromDeviceAsynch(xmit, host_xmit,xbytes); // Make this Asynch
+#endif
       
+      //      ierr =MPI_Isend(host_xmit, xbytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
+      //      GRID_ASSERT(ierr==0);
+      //      off_node_bytes+=xbytes;
+
+      srq.PacketType = InterNodeXmit;
+      srq.bytes      = xbytes;
+      //      srq.req        = xrq;
+      srq.host_buf   = host_xmit;
+      srq.device_buf = xmit;
+      srq.tag        = tag;
+      srq.dest       = dest;
+      srq.commdir    = commdir;
+      list.push_back(srq);
     }
   }
 
   return off_node_bytes;
 }
+/*
+ * In the interest of better pipelining, poll for completion on each DtoH and 
+ * start MPI_ISend in the meantime
+ */
+void CartesianCommunicator::StencilSendToRecvFromPollIRecv(std::vector<CommsRequest_t> &list)
+{
+  int pending = 0;
+  do {
+
+    pending = 0;
+
+    for(int idx = 0; idx<list.size();idx++){
+
+      if ( list[idx].PacketType==InterNodeRecv ) {
+
+	int flag = 0;
+	MPI_Status status;
+	int ierr = MPI_Test(&list[idx].req,&flag,&status);
+	assert(ierr==0);
+
+	if ( flag ) {
+	  //	  std::cout << " PollIrecv "<<idx<<" flag "<<flag<<std::endl;
+#ifdef GRID_CHECKSUM_COMMS
+ 	  acceleratorCopyToDeviceAsynch(list[idx].host_buf,list[idx].device_buf,list[idx].bytes - 8);
+#else
+	  acceleratorCopyToDeviceAsynch(list[idx].host_buf,list[idx].device_buf,list[idx].bytes);
+#endif
+	  list[idx].PacketType=InterNodeReceiveHtoD;
+	} else {
+	  pending ++;
+	}
+      }
+    }
+    //    std::cout << " PollIrecv "<<pending<<" pending requests"<<std::endl;
+  } while ( pending );
+  
+}
+void CartesianCommunicator::StencilSendToRecvFromPollDtoH(std::vector<CommsRequest_t> &list)
+{
+  int pending = 0;
+  do {
+
+    pending = 0;
+
+    for(int idx = 0; idx<list.size();idx++){
+
+      if ( list[idx].PacketType==InterNodeXmit ) {
+
+	if ( acceleratorEventIsComplete(list[idx].ev) ) {
+
+	  void *host_xmit = list[idx].host_buf;
+	  uint64_t xbytes = list[idx].bytes;
+	  int dest        = list[idx].dest;
+	  int tag         = list[idx].tag;
+	  int commdir     = list[idx].commdir;
+	  ///////////////////
+	  // Send packet
+	  ///////////////////
+
+	  //	  std::cout << " DtoH is complete for index "<<idx<<" calling MPI_Isend "<<std::endl;
+	  
+	  MPI_Request xrq;
+	  int ierr =MPI_Isend(host_xmit, (int)(xbytes/sizeof(int32_t)), MPI_INT32_T,dest,tag,communicator_halo[commdir],&xrq);
+	  GRID_ASSERT(ierr==0);
+
+	  list[idx].req        = xrq; // Update the MPI request in the list
+
+	  list[idx].PacketType=InterNodeXmitISend;
+
+	} else {
+	  // not done, so return to polling loop
+	  pending++;
+	}
+      }
+    }
+  } while (pending);
+}  
+
+double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+							 void *xmit,void *xmit_comp,
+							 int dest,int dox,
+							 void *recv,void *recv_comp,
+							 int from,int dor,
+							 uint64_t xbytes,uint64_t rbytes,int dir)
+{
+  int ncomm  =communicator_halo.size();
+  int commdir=dir%ncomm;
+
+  MPI_Request xrq;
+  MPI_Request rrq;
+
+  int ierr;
+  int gdest = ShmRanks[dest];
+  int gfrom = ShmRanks[from];
+  int gme   = ShmRanks[_processor];
+
+  GRID_ASSERT(dest != _processor);
+  GRID_ASSERT(from != _processor);
+  GRID_ASSERT(gme  == ShmRank);
+  double off_node_bytes=0.0;
+  int tag;
+
+  void * host_xmit = NULL;
+
+  ////////////////////////////////
+  // Receives already posted
+  // Copies already started
+  ////////////////////////////////
+  /*  
+   * PHASE 2: (Begin)
+   * - complete all copies
+   * - post MPI send asynch
+   */
+#ifdef NVLINK_GET
+  if ( dor ) {
+
+    if ( ! ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) ) {
+      // Intranode
+      void *shm = (void *) this->ShmBufferTranslate(from,xmit);
+      GRID_ASSERT(shm!=NULL);
+
+      CommsRequest_t srq;
+
+      srq.ev = acceleratorCopyDeviceToDeviceAsynch(shm,recv,rbytes);
+
+      srq.PacketType = IntraNodeRecv;
+      srq.bytes      = xbytes;
+      //      srq.req        = xrq;
+      srq.host_buf   = NULL;
+      srq.device_buf = xmit;
+      srq.tag        = -1;
+      srq.dest       = dest;
+      srq.commdir    = dir;
+      list.push_back(srq);
+    }
+  }  
+#else
+  if (dox) {
+
+    if ( !( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) ) {
+      // Intranode
+      void *shm = (void *) this->ShmBufferTranslate(dest,recv);
+      GRID_ASSERT(shm!=NULL);
+
+      CommsRequest_t srq;
+      
+      srq.ev = acceleratorCopyDeviceToDeviceAsynch(xmit,shm,xbytes);
+
+      srq.PacketType = IntraNodeXmit;
+      srq.bytes      = xbytes;
+      //      srq.req        = xrq;
+      srq.host_buf   = NULL;
+      srq.device_buf = xmit;
+      srq.tag        = -1;
+      srq.dest       = dest;
+      srq.commdir    = dir;
+      list.push_back(srq);
+      
+    }
+  }
+#endif
+  return off_node_bytes;
+}
 void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
 {
-  int nreq=list.size();
+  acceleratorCopySynchronise(); // Complete all pending copy transfers D2D
 
-  acceleratorCopySynchronise();
+  std::vector<MPI_Status> status;
+  std::vector<MPI_Request> MpiRequests;
+    
+  for(int r=0;r<list.size();r++){
+    // Must check each Send buf is clear to reuse
+    if ( list[r].PacketType == InterNodeXmitISend ) MpiRequests.push_back(list[r].req);
+    //    if ( list[r].PacketType == InterNodeRecv ) MpiRequests.push_back(list[r].req); // Already "Test" passed
+  }
 
-  if (nreq==0) return;
+  int nreq=MpiRequests.size();
 
-  std::vector<MPI_Status> status(nreq);
-  int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
-  assert(ierr==0);
-  list.resize(0);
+  if (nreq>0) {
+    status.resize(MpiRequests.size());
+    int ierr = MPI_Waitall(MpiRequests.size(),&MpiRequests[0],&status[0]); // Sends are guaranteed in order. No harm in not completing.
+    GRID_ASSERT(ierr==0);
+  }
+  
+  //  for(int r=0;r<nreq;r++){
+  //    if ( list[r].PacketType==InterNodeRecv ) {
+  //      acceleratorCopyToDeviceAsynch(list[r].host_buf,list[r].device_buf,list[r].bytes);
+  //    }
+  //  }
+#ifdef GRID_CHECKSUM_COMMS
+  for(int r=0;r<list.size();r++){
+    if ( list[r].PacketType == InterNodeReceiveHtoD ) {
+      uint64_t rbytes_data = list[r].bytes - 8;
+      uint64_t expected_cs = *(uint64_t*)(((char*)list[r].host_buf) + rbytes_data);
+      uint64_t computed_cs = checksum_gpu((uint64_t*)list[r].device_buf, rbytes_data / 8) ^ (checksum_index + 1 + 1000 * list[r].tag); //
+      if (expected_cs != computed_cs) {
+	// TODO: error message, backtrace, quit
+
+	fprintf(stderr, "GRID_CHECKSUM_COMMS error:\n");
+	fprintf(stderr, " processor = %d\n", (int)_processor);
+	for(int d=0;d<_processors.size();d++)
+	  fprintf(stderr, " processor_coord[%d] = %d\n", d, _processor_coor[d]);
+	fprintf(stderr, " hostname: %s\n", GridHostname());
+	fprintf(stderr, " expected_cs: %ld\n", expected_cs);
+	fprintf(stderr, " computed_cs: %ld\n", computed_cs);
+	fprintf(stderr, " dest: %d\n", list[r].dest);
+	fprintf(stderr, " tag: %d\n", list[r].tag);
+	fprintf(stderr, " commdir: %d\n", list[r].commdir);
+	fprintf(stderr, " bytes: %ld\n", (uint64_t)list[r].bytes);
+
+	fflush(stderr);
+
+	// backtrace
+	int symbols = backtrace(Grid_backtrace_buffer,_NBACKTRACE);
+	backtrace_symbols_fd(Grid_backtrace_buffer,symbols, 2);
+
+	exit(1);
+      }
+    }
+  }
+
+  checksum_index += 1;
+#endif
+  
+  list.resize(0);               // Delete the list
+  this->HostBufferFreeAll();    // Clean up the buffer allocs
+#ifndef NVLINK_GET
+  this->StencilBarrier(); // if PUT must check our nbrs have filled our receive buffers.
+#endif   
 }
+#endif
+////////////////////////////////////////////
+// END PIPELINE MODE / NO CUDA AWARE MPI
+////////////////////////////////////////////
+
 void CartesianCommunicator::StencilBarrier(void)
 {
+  FlightRecorder::StepLog("NodeBarrier");
   MPI_Barrier  (ShmComm);
 }
 //void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
@@ -468,17 +879,19 @@ void CartesianCommunicator::StencilBarrier(void)
 //}
 void CartesianCommunicator::Barrier(void)
 {
+  FlightRecorder::StepLog("GridBarrier");
   int ierr = MPI_Barrier(communicator);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
 }
-void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
+void CartesianCommunicator::Broadcast(int root,void* data,uint64_t bytes)
 {
+  FlightRecorder::StepLog("Broadcast");
   int ierr=MPI_Bcast(data,
-		     bytes,
+		     (int)bytes,
 		     MPI_BYTE,
 		     root,
 		     communicator);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
 }
 int CartesianCommunicator::RankWorld(void){
   int r;
@@ -486,23 +899,25 @@ int CartesianCommunicator::RankWorld(void){
   return r;
 }
 void CartesianCommunicator::BarrierWorld(void){
+  FlightRecorder::StepLog("BarrierWorld");
   int ierr = MPI_Barrier(communicator_world);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
 }
-void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
+void CartesianCommunicator::BroadcastWorld(int root,void* data, uint64_t bytes)
 {
+  FlightRecorder::StepLog("BroadcastWorld");
   int ierr= MPI_Bcast(data,
-		      bytes,
+		      (int)bytes,
 		      MPI_BYTE,
 		      root,
 		      communicator_world);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
 }
 
 void CartesianCommunicator::AllToAll(int dim,void  *in,void *out,uint64_t words,uint64_t bytes)
 {
   Coordinate row(_ndimension,1);
-  assert(dim>=0 && dim<_ndimension);
+  GRID_ASSERT(dim>=0 && dim<_ndimension);
 
   //  Split the communicator
   row[dim] = _processors[dim];
@@ -513,6 +928,7 @@ void CartesianCommunicator::AllToAll(int dim,void  *in,void *out,uint64_t words,
 }
 void CartesianCommunicator::AllToAll(void  *in,void *out,uint64_t words,uint64_t bytes)
 {
+  FlightRecorder::StepLog("AllToAll");
   // MPI is a pain and uses "int" arguments
   // 64*64*64*128*16 == 500Million elements of data.
   // When 24*4 bytes multiples get 50x 10^9 >>> 2x10^9 Y2K bug.
@@ -522,12 +938,32 @@ void CartesianCommunicator::AllToAll(void  *in,void *out,uint64_t words,uint64_t
   int ibytes;
   iwords = words;
   ibytes = bytes;
-  assert(words == iwords); // safe to cast to int ?
-  assert(bytes == ibytes); // safe to cast to int ?
+  GRID_ASSERT(words == iwords); // safe to cast to int ?
+  GRID_ASSERT(bytes == ibytes); // safe to cast to int ?
   MPI_Type_contiguous(ibytes,MPI_BYTE,&object);
   MPI_Type_commit(&object);
   MPI_Alltoall(in,iwords,object,out,iwords,object,communicator);
   MPI_Type_free(&object);
 }
+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

@@ -27,6 +27,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 /*  END LEGAL */
 #include <Grid/GridCore.h>
 
+void GridAbort(void) { abort(); }
+
 NAMESPACE_BEGIN(Grid);
 
 ///////////////////////////////////////////////////////////////////////////////////////////////////
@@ -34,6 +36,7 @@ NAMESPACE_BEGIN(Grid);
 ///////////////////////////////////////////////////////////////////////////////////////////////////
 Grid_MPI_Comm       CartesianCommunicator::communicator_world;
 
+
 void CartesianCommunicator::Init(int *argc, char *** arv)
 {
   GlobalSharedMemory::Init(communicator_world);
@@ -54,14 +57,14 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
 {
   _shm_processors = Coordinate(processors.size(),1);
   _processors = processors;
-  _ndimension = processors.size();  assert(_ndimension>=1);
+  _ndimension = processors.size();  GRID_ASSERT(_ndimension>=1);
   _processor_coor.resize(_ndimension);
   
   // Require 1^N processor grid for fake
   _Nprocessors=1;
   _processor = 0;
   for(int d=0;d<_ndimension;d++) {
-    assert(_processors[d]==1);
+    GRID_ASSERT(_processors[d]==1);
     _processor_coor[d] = 0;
   }
   SetCommunicator(communicator_world);
@@ -87,19 +90,19 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 					   int dest,
 					   void *recv,
 					   int from,
-					   int bytes)
+					   uint64_t bytes)
 {
-  assert(0);
+  GRID_ASSERT(0);
 }
-void CartesianCommunicator::CommsComplete(std::vector<CommsRequest_t> &list){ assert(0);}
+void CartesianCommunicator::CommsComplete(std::vector<CommsRequest_t> &list){ GRID_ASSERT(list.size()==0);}
 void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 						void *xmit,
 						int dest,
 						void *recv,
 						int from,
-						int bytes,int dir)
+						uint64_t bytes,int dir)
 {
-  assert(0);
+  GRID_ASSERT(0);
 }
 
 void CartesianCommunicator::AllToAll(int dim,void  *in,void *out,uint64_t words,uint64_t bytes)
@@ -110,11 +113,22 @@ 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){}
-void CartesianCommunicator::Broadcast(int root,void* data, int bytes) {}
-void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes) { }
+void CartesianCommunicator::Broadcast(int root,void* data, uint64_t bytes) {}
+void CartesianCommunicator::BroadcastWorld(int root,void* data, uint64_t bytes) { }
 void CartesianCommunicator::BarrierWorld(void) { }
 int  CartesianCommunicator::RankFromProcessorCoor(Coordinate &coor) {  return 0;}
 void CartesianCommunicator::ProcessorCoorFromRank(int rank, Coordinate &coor){  coor = _processor_coor; }
@@ -124,20 +138,33 @@ void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest
   dest=0;
 }
 
+int CartesianCommunicator::IsOffNode(int rank) { return false; }
+
 double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
 						     int xmit_to_rank,int dox,
 						     void *recv,
 						     int recv_from_rank,int dor,
-						     int bytes, int dir)
+						     uint64_t bytes, int dir)
 {
   return 2.0*bytes;
 }
+void CartesianCommunicator::StencilSendToRecvFromPollIRecv(std::vector<CommsRequest_t> &list) {};
+void CartesianCommunicator::StencilSendToRecvFromPollDtoH(std::vector<CommsRequest_t> &list) {};
+double CartesianCommunicator::StencilSendToRecvFromPrepare(std::vector<CommsRequest_t> &list,
+							   void *xmit,
+							   int xmit_to_rank,int dox,
+							   void *recv,
+							   int recv_from_rank,int dor,
+							   uint64_t xbytes,uint64_t rbytes, int dir)
+{
+  return 0.0;
+}
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
-							 void *xmit,
+							 void *xmit, void *xmit_comp,
 							 int xmit_to_rank,int dox,
-							 void *recv,
+							 void *recv, void *recv_comp,
 							 int recv_from_rank,int dor,
-							 int xbytes,int rbytes, int dir)
+							 uint64_t xbytes,uint64_t rbytes, int dir)
 {
   return xbytes+rbytes;
 }

Grid/communicator/SharedMemory.cc

@@ -58,8 +58,8 @@ int                 GlobalSharedMemory::WorldNode;
 
 void GlobalSharedMemory::SharedMemoryFree(void)
 {
-  assert(_ShmAlloc);
-  assert(_ShmAllocBytes>0);
+  GRID_ASSERT(_ShmAlloc);
+  GRID_ASSERT(_ShmAllocBytes>0);
   for(int r=0;r<WorldShmSize;r++){
     munmap(WorldShmCommBufs[r],_ShmAllocBytes);
   }
@@ -80,7 +80,7 @@ void *SharedMemory::HostBufferMalloc(size_t bytes){
     std::cout<< " Current alloc is " << (bytes/(1024*1024)) <<"MB"<<std::endl;
     std::cout<< " Current bytes is " << (host_heap_bytes/(1024*1024)) <<"MB"<<std::endl;
     std::cout<< " Current heap  is " << (host_heap_size/(1024*1024)) <<"MB"<<std::endl;
-    assert(host_heap_bytes<host_heap_size);
+    GRID_ASSERT(host_heap_bytes<host_heap_size);
   }
   return ptr;
 }
@@ -100,7 +100,7 @@ void *SharedMemory::ShmBufferMalloc(size_t bytes){
     std::cout<< " Current alloc is " << (bytes/(1024*1024)) <<"MB"<<std::endl;
     std::cout<< " Current bytes is " << (heap_bytes/(1024*1024)) <<"MB"<<std::endl;
     std::cout<< " Current heap  is " << (heap_size/(1024*1024)) <<"MB"<<std::endl;
-    assert(heap_bytes<heap_size);
+    GRID_ASSERT(heap_bytes<heap_size);
   }
   //std::cerr << "ShmBufferMalloc "<<std::hex<< ptr<<" - "<<((uint64_t)ptr+bytes)<<std::dec<<std::endl;
   return ptr;
@@ -127,13 +127,13 @@ void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmD
   if ( str ) {
     std::vector<int> IntShmDims;
     GridCmdOptionIntVector(std::string(str),IntShmDims);
-    assert(IntShmDims.size() == WorldDims.size());
+    GRID_ASSERT(IntShmDims.size() == WorldDims.size());
     long ShmSize = 1;
     for (int dim=0;dim<WorldDims.size();dim++) {
       ShmSize *= (ShmDims[dim] = IntShmDims[dim]);
-      assert(divides(ShmDims[dim],WorldDims[dim]));
+      GRID_ASSERT(divides(ShmDims[dim],WorldDims[dim]));
     }
-    assert(ShmSize == WorldShmSize);
+    GRID_ASSERT(ShmSize == WorldShmSize);
     return;
   }
   

Grid/communicator/SharedMemory.h

@@ -46,8 +46,40 @@ NAMESPACE_BEGIN(Grid);
 
 #if defined (GRID_COMMS_MPI3) 
 typedef MPI_Comm    Grid_MPI_Comm;
+typedef MPI_Request MpiCommsRequest_t;
+#ifdef ACCELERATOR_AWARE_MPI
 typedef MPI_Request CommsRequest_t;
+#else
+/*
+ * Enable state transitions as each packet flows.
+ */
+enum PacketType_t {
+  FaceGather,
+  InterNodeXmit,
+  InterNodeRecv,
+  IntraNodeXmit,
+  IntraNodeRecv,
+  InterNodeXmitISend,
+  InterNodeReceiveHtoD
+};
+/*
+ *Package arguments needed for various actions along packet flow
+ */
+typedef struct {
+  PacketType_t PacketType;
+  void *host_buf;
+  void *device_buf;
+  int dest;
+  int tag;
+  int commdir;
+  unsigned long bytes;
+  acceleratorEvent_t ev;
+  MpiCommsRequest_t req;
+} CommsRequest_t;
+#endif
+
 #else 
+typedef int MpiCommsRequest_t;
 typedef int CommsRequest_t;
 typedef int Grid_MPI_Comm;
 #endif
@@ -105,7 +137,7 @@ public:
   ///////////////////////////////////////////////////
   static void SharedMemoryAllocate(uint64_t bytes, int flags);
   static void SharedMemoryFree(void);
-  static void SharedMemoryCopy(void *dest,void *src,size_t bytes);
+  //  static void SharedMemoryCopy(void *dest,void *src,size_t bytes);
   static void SharedMemoryZero(void *dest,size_t bytes);
 
 };

Grid/communicator/SharedMemoryMPI.cc

@@ -42,6 +42,7 @@ Author: Christoph Lehner <christoph@lhnr.de>
 #ifdef ACCELERATOR_AWARE_MPI
 #define GRID_SYCL_LEVEL_ZERO_IPC
 #define SHM_SOCKETS
+#else
 #endif 
 #include <syscall.h>
 #endif
@@ -66,7 +67,7 @@ public:
   {
     int errnum;
 
-    sock = socket(AF_UNIX, SOCK_DGRAM, 0);  assert(sock>0);
+    sock = socket(AF_UNIX, SOCK_DGRAM, 0);  GRID_ASSERT(sock>0);
 
     struct sockaddr_un sa_un = { 0 };
     sa_un.sun_family = AF_UNIX;
@@ -157,7 +158,7 @@ public:
 /*Construct from an MPI communicator*/
 void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
 {
-  assert(_ShmSetup==0);
+  GRID_ASSERT(_ShmSetup==0);
   WorldComm = comm;
   MPI_Comm_rank(WorldComm,&WorldRank);
   MPI_Comm_size(WorldComm,&WorldSize);
@@ -183,7 +184,7 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
 
   // WorldNodes
   WorldNodes = WorldSize/WorldShmSize;
-  assert( (WorldNodes * WorldShmSize) == WorldSize );
+  GRID_ASSERT( (WorldNodes * WorldShmSize) == WorldSize );
 
 
   // FIXME: Check all WorldShmSize are the same ?
@@ -208,7 +209,7 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
   MyGroup.resize(WorldShmSize);
   for(int rank=0;rank<WorldSize;rank++){
     if(WorldShmRanks[rank]!=MPI_UNDEFINED){
-      assert(g<WorldShmSize);
+      GRID_ASSERT(g<WorldShmSize);
       MyGroup[g++] = rank;
     }
   }
@@ -224,7 +225,7 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
   // global sum leaders over comm world
   ///////////////////////////////////////////////////////////////////
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&leaders_1hot[0],WorldSize,MPI_INT,MPI_SUM,WorldComm);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
 
   ///////////////////////////////////////////////////////////////////
   // find the group leaders world rank
@@ -245,7 +246,7 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
       WorldNode=g;
     }
   }
-  assert(WorldNode!=-1);
+  GRID_ASSERT(WorldNode!=-1);
   _ShmSetup=1;
 }
 // Gray encode support 
@@ -287,7 +288,7 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processo
   // Assert power of two shm_size.
   ////////////////////////////////////////////////////////////////
   int log2size = Log2Size(WorldShmSize,MAXLOG2RANKSPERNODE);
-  assert(log2size != -1);
+  GRID_ASSERT(log2size != -1);
 
   ////////////////////////////////////////////////////////////////
   // Identify the hypercube coordinate of this node using hostname
@@ -308,7 +309,7 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processo
   // Parse ICE-XA hostname to get hypercube location
   gethostname(name,namelen);
   int nscan = sscanf(name,"r%di%dn%d",&R,&I,&N) ;
-  assert(nscan==3);
+  GRID_ASSERT(nscan==3);
 
   int nlo = N%9;
   int nhi = N/9;
@@ -332,8 +333,8 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processo
   //////////////////////////////////////////////////////////////////
   MPI_Bcast(&rootcoor, sizeof(rootcoor), MPI_BYTE, 0, WorldComm); 
   hypercoor=hypercoor-rootcoor;
-  assert(hypercoor<WorldSize);
-  assert(hypercoor>=0);
+  GRID_ASSERT(hypercoor<WorldSize);
+  GRID_ASSERT(hypercoor>=0);
 
   //////////////////////////////////////
   // Printing
@@ -381,7 +382,7 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processo
   for(int i=0;i<ndimension;i++){
     Nprocessors*=processors[i];
   }
-  assert(WorldSize==Nprocessors);
+  GRID_ASSERT(WorldSize==Nprocessors);
 
   ////////////////////////////////////////////////////////////////
   // Establish mapping between lexico physics coord and WorldRank
@@ -400,7 +401,7 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processo
   // Build the new communicator
   /////////////////////////////////////////////////////////////////
   int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
 }
 void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
 {
@@ -430,7 +431,8 @@ void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &proce
   for(int i=0;i<ndimension;i++){
     Nprocessors*=processors[i];
   }
-  assert(WorldSize==Nprocessors);
+  //  std::cerr << " WorldSize "<<WorldSize << " Nprocessors "<<Nprocessors<<" "<<processors<<std::endl; 
+  GRID_ASSERT(WorldSize==Nprocessors);
 
   ////////////////////////////////////////////////////////////////
   // Establish mapping between lexico physics coord and WorldRank
@@ -446,7 +448,7 @@ void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &proce
   // Build the new communicator
   /////////////////////////////////////////////////////////////////
   int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
 }
 ////////////////////////////////////////////////////////////////////////////////////////////
 // SHMGET
@@ -455,8 +457,8 @@ void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &proce
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
   std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " shmget implementation "<<std::endl;
-  assert(_ShmSetup==1);
-  assert(_ShmAlloc==0);
+  GRID_ASSERT(_ShmSetup==1);
+  GRID_ASSERT(_ShmAlloc==0);
 
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
   // allocate the shared windows for our group
@@ -517,8 +519,8 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
   void * ShmCommBuf ; 
-  assert(_ShmSetup==1);
-  assert(_ShmAlloc==0);
+  GRID_ASSERT(_ShmSetup==1);
+  GRID_ASSERT(_ShmAlloc==0);
 
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
   // allocate the pointer array for shared windows for our group
@@ -537,19 +539,22 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   // Each MPI rank should allocate our own buffer
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
 #ifndef ACCELERATOR_AWARE_MPI
-  HostCommBuf= malloc(bytes);
+  // printf("Host buffer allocate for GPU non-aware MPI\n");
+  HostCommBuf= malloc(bytes); /// CHANGE THIS TO malloc_host
 #endif  
   ShmCommBuf = acceleratorAllocDevice(bytes);
   if (ShmCommBuf == (void *)NULL ) {
-    std::cerr << " SharedMemoryMPI.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;
+    std::cerr << "SharedMemoryMPI.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;
     exit(EXIT_FAILURE);  
   }
   if ( WorldRank == 0 ){
-    std::cout << WorldRank << Mheader " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
+    std::cout << Mheader " acceleratorAllocDevice "<< bytes 
 	      << "bytes at "<< std::hex<< ShmCommBuf << " - "<<(bytes-1+(uint64_t)ShmCommBuf) <<std::dec<<" for comms buffers " <<std::endl;
   }
   SharedMemoryZero(ShmCommBuf,bytes);
-  std::cout<< "Setting up IPC"<<std::endl;
+  if ( WorldRank == 0 ){
+    std::cout<< Mheader "Setting up IPC"<<std::endl;
+  }
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
   // Loop over ranks/gpu's on our node
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -569,8 +574,8 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #ifdef GRID_SYCL_LEVEL_ZERO_IPC
     typedef struct { int fd; pid_t pid ; ze_ipc_mem_handle_t ze; } clone_mem_t;
 
-    auto zeDevice    = cl::sycl::get_native<cl::sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_device());
-    auto zeContext   = cl::sycl::get_native<cl::sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_context());
+    auto zeDevice    = sycl::get_native<sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_device());
+    auto zeContext   = sycl::get_native<sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_context());
       
     ze_ipc_mem_handle_t ihandle;
     clone_mem_t handle;
@@ -580,8 +585,6 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
       if ( err != ZE_RESULT_SUCCESS ) {
 	std::cerr << "SharedMemoryMPI.cc zeMemGetIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
 	exit(EXIT_FAILURE);
-      } else {
-	std::cout << "SharedMemoryMPI.cc zeMemGetIpcHandle succeeded for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
       }
       memcpy((void *)&handle.fd,(void *)&ihandle,sizeof(int));
       handle.pid = getpid();
@@ -626,7 +629,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 			 MPI_BYTE,
 			 r,
 			 WorldShmComm);
-      assert(ierr==0);
+      GRID_ASSERT(ierr==0);
     }
     
     ///////////////////////////////////////////////////////////////
@@ -640,12 +643,12 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #ifdef SHM_SOCKETS
       myfd=UnixSockets::RecvFileDescriptor();
 #else
-      std::cout<<"mapping seeking remote pid/fd "
-	       <<handle.pid<<"/"
-	       <<handle.fd<<std::endl;
+      //      std::cout<<"mapping seeking remote pid/fd "
+      //	       <<handle.pid<<"/"
+      //	       <<handle.fd<<std::endl;
 
       int pidfd = syscall(SYS_pidfd_open,handle.pid,0);
-      std::cout<<"Using IpcHandle pidfd "<<pidfd<<"\n";
+      //      std::cout<<"Using IpcHandle pidfd "<<pidfd<<"\n";
       //      int myfd  = syscall(SYS_pidfd_getfd,pidfd,handle.fd,0);
       myfd  = syscall(438,pidfd,handle.fd,0);
       int err_t = errno;
@@ -655,7 +658,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 	assert(0);
       }
 #endif
-      std::cout<<"Using IpcHandle mapped remote pid "<<handle.pid <<" FD "<<handle.fd <<" to myfd "<<myfd<<"\n";
+      //      std::cout<<"Using IpcHandle mapped remote pid "<<handle.pid <<" FD "<<handle.fd <<" to myfd "<<myfd<<"\n";
       memcpy((void *)&ihandle,(void *)&handle.ze,sizeof(ihandle));
       memcpy((void *)&ihandle,(void *)&myfd,sizeof(int));
 
@@ -664,11 +667,8 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 	std::cerr << "SharedMemoryMPI.cc "<<zeContext<<" "<<zeDevice<<std::endl;
 	std::cerr << "SharedMemoryMPI.cc zeMemOpenIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl; 
 	exit(EXIT_FAILURE);
-      } else {
-	std::cout << "SharedMemoryMPI.cc zeMemOpenIpcHandle succeeded for rank "<<r<<std::endl;
-	std::cout << "SharedMemoryMPI.cc zeMemOpenIpcHandle pointer is "<<std::hex<<thisBuf<<std::dec<<std::endl;
       }
-      assert(thisBuf!=nullptr);
+      GRID_ASSERT(thisBuf!=nullptr);
     }
 #endif
 #ifdef GRID_CUDA
@@ -709,8 +709,8 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
   std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " MMAP implementation "<< GRID_SHM_PATH <<std::endl;
-  assert(_ShmSetup==1);
-  assert(_ShmAlloc==0);
+  GRID_ASSERT(_ShmSetup==1);
+  GRID_ASSERT(_ShmAlloc==0);
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
   // allocate the shared windows for our group
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -740,13 +740,14 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     void *ptr = (void *) mmap(NULL, bytes, PROT_READ | PROT_WRITE, mmap_flag,fd, 0); 
     if ( ptr == (void *)MAP_FAILED ) {    
       printf("mmap %s failed\n",shm_name);
-      perror("failed mmap");      assert(0);    
+      perror("failed mmap");      GRID_ASSERT(0);    
     }
-    assert(((uint64_t)ptr&0x3F)==0);
+    GRID_ASSERT(((uint64_t)ptr&0x3F)==0);
     close(fd);
     WorldShmCommBufs[r] =ptr;
     //    std::cout << Mheader "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
   }
+  std::cout<< Mheader " Intra-node IPC setup is complete "<<std::endl;
   _ShmAlloc=1;
   _ShmAllocBytes  = bytes;
 };
@@ -756,8 +757,8 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
   std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " MMAP anonymous implementation "<<std::endl;
-  assert(_ShmSetup==1);
-  assert(_ShmAlloc==0);
+  GRID_ASSERT(_ShmSetup==1);
+  GRID_ASSERT(_ShmAlloc==0);
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
   // allocate the shared windows for our group
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -768,7 +769,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   // Hugetlbf and others map filesystems as mappable huge pages
   ////////////////////////////////////////////////////////////////////////////////////////////
   char shm_name [NAME_MAX];
-  assert(WorldShmSize == 1);
+  GRID_ASSERT(WorldShmSize == 1);
   for(int r=0;r<WorldShmSize;r++){
     
     int fd=-1;
@@ -782,9 +783,9 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     void *ptr = (void *) mmap(NULL, bytes, PROT_READ | PROT_WRITE, mmap_flag,fd, 0); 
     if ( ptr == (void *)MAP_FAILED ) {    
       printf("mmap %s failed\n",shm_name);
-      perror("failed mmap");      assert(0);    
+      perror("failed mmap");      GRID_ASSERT(0);    
     }
-    assert(((uint64_t)ptr&0x3F)==0);
+    GRID_ASSERT(((uint64_t)ptr&0x3F)==0);
     close(fd);
     WorldShmCommBufs[r] =ptr;
     //    std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
@@ -803,8 +804,8 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 { 
   std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " SHMOPEN implementation "<<std::endl;
-  assert(_ShmSetup==1);
-  assert(_ShmAlloc==0); 
+  GRID_ASSERT(_ShmSetup==1);
+  GRID_ASSERT(_ShmAlloc==0); 
   MPI_Barrier(WorldShmComm);
   WorldShmCommBufs.resize(WorldShmSize);
 
@@ -835,7 +836,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 	perror("failed mmap");     
 	assert(0);    
       }
-      assert(((uint64_t)ptr&0x3F)==0);
+      GRID_ASSERT(((uint64_t)ptr&0x3F)==0);
       
       WorldShmCommBufs[r] =ptr;
       close(fd);
@@ -856,8 +857,8 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
       if ( fd<0 ) {	perror("failed shm_open");	assert(0);      }
       
       void * ptr =  mmap(NULL,size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
-      if ( ptr == MAP_FAILED ) {       perror("failed mmap");      assert(0);    }
-      assert(((uint64_t)ptr&0x3F)==0);
+      if ( ptr == MAP_FAILED ) {       perror("failed mmap");      GRID_ASSERT(0);    }
+      GRID_ASSERT(((uint64_t)ptr&0x3F)==0);
       WorldShmCommBufs[r] =ptr;
 
       close(fd);
@@ -880,14 +881,14 @@ void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes)
   bzero(dest,bytes);
 #endif
 }
-void GlobalSharedMemory::SharedMemoryCopy(void *dest,void *src,size_t bytes)
-{
-#if defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)
-  acceleratorCopyToDevice(src,dest,bytes);
-#else   
-  bcopy(src,dest,bytes);
-#endif
-}
+//void GlobalSharedMemory::SharedMemoryCopy(void *dest,void *src,size_t bytes)
+//{
+//#if defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)
+//  acceleratorCopyToDevice(src,dest,bytes);
+//#else   
+//  bcopy(src,dest,bytes);
+//#endif
+//}
 ////////////////////////////////////////////////////////
 // Global shared functionality finished
 // Now move to per communicator functionality
@@ -914,7 +915,7 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
   //////////////////////////////////////////////////////////////////////
   // Map ShmRank to WorldShmRank and use the right buffer
   //////////////////////////////////////////////////////////////////////
-  assert (GlobalSharedMemory::ShmAlloc()==1);
+  GRID_ASSERT (GlobalSharedMemory::ShmAlloc()==1);
   heap_size = GlobalSharedMemory::ShmAllocBytes();
   for(int r=0;r<ShmSize;r++){
 
@@ -923,6 +924,7 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
     MPI_Allreduce(MPI_IN_PLACE,&wsr,1,MPI_UINT32_T,MPI_SUM,ShmComm);
 
     ShmCommBufs[r] = GlobalSharedMemory::WorldShmCommBufs[wsr];
+    //    std::cerr << " SetCommunicator rank "<<r<<" comm "<<ShmCommBufs[r] <<std::endl;
   }
   ShmBufferFreeAll();
 
@@ -953,7 +955,7 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
   }
 #endif
 
-  //SharedMemoryTest();
+  //  SharedMemoryTest();
 }
 //////////////////////////////////////////////////////////////////
 // On node barrier
@@ -975,19 +977,18 @@ void SharedMemory::SharedMemoryTest(void)
        check[0]=GlobalSharedMemory::WorldNode;
        check[1]=r;
        check[2]=magic;
-       GlobalSharedMemory::SharedMemoryCopy( ShmCommBufs[r], check, 3*sizeof(uint64_t));
+       acceleratorCopyToDevice(check,ShmCommBufs[r],3*sizeof(uint64_t));
     }
   }
   ShmBarrier();
   for(uint64_t r=0;r<ShmSize;r++){
-    ShmBarrier();
-    GlobalSharedMemory::SharedMemoryCopy(check,ShmCommBufs[r], 3*sizeof(uint64_t));
-    ShmBarrier();
-    assert(check[0]==GlobalSharedMemory::WorldNode);
-    assert(check[1]==r);
-    assert(check[2]==magic);
-    ShmBarrier();
+    acceleratorCopyFromDevice(ShmCommBufs[r],check,3*sizeof(uint64_t));
+    GRID_ASSERT(check[0]==GlobalSharedMemory::WorldNode);
+    GRID_ASSERT(check[1]==r);
+    GRID_ASSERT(check[2]==magic);
   }
+  ShmBarrier();
+  std::cout << GridLogDebug << " SharedMemoryTest has passed "<<std::endl;
 }
 
 void *SharedMemory::ShmBuffer(int rank)
@@ -1002,12 +1003,14 @@ void *SharedMemory::ShmBuffer(int rank)
 void *SharedMemory::ShmBufferTranslate(int rank,void * local_p)
 {
   int gpeer = ShmRanks[rank];
-  assert(gpeer!=ShmRank); // never send to self
+  GRID_ASSERT(gpeer!=ShmRank); // never send to self
+  //  std::cout << "ShmBufferTranslate for rank " << rank<<" peer "<<gpeer<<std::endl;
   if (gpeer == MPI_UNDEFINED){
     return NULL;
   } else { 
     uint64_t offset = (uint64_t)local_p - (uint64_t)ShmCommBufs[ShmRank];
     uint64_t remote = (uint64_t)ShmCommBufs[gpeer]+offset;
+    //    std::cout << "ShmBufferTranslate : local,offset,remote "<<std::hex<<local_p<<" "<<offset<<" "<<remote<<std::dec<<std::endl;
     return (void *) remote;
   }
 }

Grid/communicator/SharedMemoryNone.cc

@@ -34,7 +34,7 @@ NAMESPACE_BEGIN(Grid);
 /*Construct from an MPI communicator*/
 void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
 {
-  assert(_ShmSetup==0);
+  GRID_ASSERT(_ShmSetup==0);
   WorldComm = 0;
   WorldRank = 0;
   WorldSize = 1;
@@ -62,8 +62,8 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
   std::cout << header "SharedMemoryAllocate "<< bytes<< " GPU implementation "<<std::endl;
   void * ShmCommBuf ; 
-  assert(_ShmSetup==1);
-  assert(_ShmAlloc==0);
+  GRID_ASSERT(_ShmSetup==1);
+  GRID_ASSERT(_ShmAlloc==0);
 
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
   // Each MPI rank should allocate our own buffer
@@ -92,8 +92,8 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
   void * ShmCommBuf ; 
-  assert(_ShmSetup==1);
-  assert(_ShmAlloc==0);
+  GRID_ASSERT(_ShmSetup==1);
+  GRID_ASSERT(_ShmAlloc==0);
   int mmap_flag =0;
 #ifdef MAP_ANONYMOUS
   mmap_flag = mmap_flag| MAP_SHARED | MAP_ANONYMOUS;
@@ -122,17 +122,17 @@ void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes)
 {
   acceleratorMemSet(dest,0,bytes);
 }
-void GlobalSharedMemory::SharedMemoryCopy(void *dest,void *src,size_t bytes)
-{
-  acceleratorCopyToDevice(src,dest,bytes);
-}
+//void GlobalSharedMemory::SharedMemoryCopy(void *dest,void *src,size_t bytes)
+//{
+//  acceleratorCopyToDevice(src,dest,bytes);
+//}
 ////////////////////////////////////////////////////////
 // Global shared functionality finished
 // Now move to per communicator functionality
 ////////////////////////////////////////////////////////
 void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
 {
-  assert(GlobalSharedMemory::ShmAlloc()==1);
+  GRID_ASSERT(GlobalSharedMemory::ShmAlloc()==1);
   ShmRanks.resize(1);
   ShmCommBufs.resize(1);
   ShmRanks[0] = 0;

Grid/cshift/Cshift.h

@@ -51,7 +51,6 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #endif 
 
 NAMESPACE_BEGIN(Grid);
-
 template<class Expression,typename std::enable_if<is_lattice_expr<Expression>::value,void>::type * = nullptr> 
 auto Cshift(const Expression &expr,int dim,int shift)  -> decltype(closure(expr)) 
 {

Grid/cshift/Cshift_common.h

@@ -30,12 +30,11 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 NAMESPACE_BEGIN(Grid);
 
 extern std::vector<std::pair<int,int> > Cshift_table; 
-extern commVector<std::pair<int,int> > Cshift_table_device; 
+extern deviceVector<std::pair<int,int> > Cshift_table_device; 
 
 inline std::pair<int,int> *MapCshiftTable(void)
 {
   // GPU version
-#ifdef ACCELERATOR_CSHIFT    
   uint64_t sz=Cshift_table.size();
   if (Cshift_table_device.size()!=sz )    {
     Cshift_table_device.resize(sz);
@@ -45,16 +44,13 @@ inline std::pair<int,int> *MapCshiftTable(void)
 			  sizeof(Cshift_table[0])*sz);
 
   return &Cshift_table_device[0];
-#else 
-  return &Cshift_table[0];
-#endif
   // CPU version use identify map
 }
 ///////////////////////////////////////////////////////////////////
 // Gather for when there is no need to SIMD split 
 ///////////////////////////////////////////////////////////////////
 template<class vobj> void 
-Gather_plane_simple (const Lattice<vobj> &rhs,cshiftVector<vobj> &buffer,int dimension,int plane,int cbmask, int off=0)
+Gather_plane_simple (const Lattice<vobj> &rhs,deviceVector<vobj> &buffer,int dimension,int plane,int cbmask, int off=0)
 {
   int rd = rhs.Grid()->_rdimensions[dimension];
 
@@ -94,17 +90,10 @@ Gather_plane_simple (const Lattice<vobj> &rhs,cshiftVector<vobj> &buffer,int dim
   {
     auto buffer_p = & buffer[0];
     auto table = MapCshiftTable();
-#ifdef ACCELERATOR_CSHIFT
     autoView(rhs_v , rhs, AcceleratorRead);
     accelerator_for(i,ent,vobj::Nsimd(),{
 	coalescedWrite(buffer_p[table[i].first],coalescedRead(rhs_v[table[i].second]));
     });
-#else
-    autoView(rhs_v , rhs, CpuRead);
-    thread_for(i,ent,{
-      buffer_p[table[i].first]=rhs_v[table[i].second];
-    });
-#endif
   }
 }
 
@@ -129,7 +118,6 @@ Gather_plane_extract(const Lattice<vobj> &rhs,
   int n1=rhs.Grid()->_slice_stride[dimension];
 
   if ( cbmask ==0x3){
-#ifdef ACCELERATOR_CSHIFT
     autoView(rhs_v , rhs, AcceleratorRead);
     accelerator_for(nn,e1*e2,1,{
 	int n = nn%e1;
@@ -140,21 +128,10 @@ Gather_plane_extract(const Lattice<vobj> &rhs,
 	vobj temp =rhs_v[so+o+b];
 	extract<vobj>(temp,pointers,offset);
       });
-#else
-    autoView(rhs_v , rhs, CpuRead);
-    thread_for2d(n,e1,b,e2,{
-	int o      =   n*n1;
-	int offset = b+n*e2;
-	
-	vobj temp =rhs_v[so+o+b];
-	extract<vobj>(temp,pointers,offset);
-      });
-#endif
   } else { 
     Coordinate rdim=rhs.Grid()->_rdimensions;
     Coordinate cdm =rhs.Grid()->_checker_dim_mask;
     std::cout << " Dense packed buffer WARNING " <<std::endl; // Does this get called twice once for each cb?
-#ifdef ACCELERATOR_CSHIFT    
     autoView(rhs_v , rhs, AcceleratorRead);
     accelerator_for(nn,e1*e2,1,{
 	int n = nn%e1;
@@ -175,33 +152,13 @@ Gather_plane_extract(const Lattice<vobj> &rhs,
 	  extract<vobj>(temp,pointers,offset);
 	}
       });
-#else
-    autoView(rhs_v , rhs, CpuRead);
-    thread_for2d(n,e1,b,e2,{
-
-	Coordinate coor;
-
-	int o=n*n1;
-	int oindex = o+b;
-
-       	int cb = RedBlackCheckerBoardFromOindex(oindex, rdim, cdm);
-
-	int ocb=1<<cb;
-	int offset = b+n*e2;
-
-	if ( ocb & cbmask ) {
-	  vobj temp =rhs_v[so+o+b];
-	  extract<vobj>(temp,pointers,offset);
-	}
-      });
-#endif
   }
 }
 
 //////////////////////////////////////////////////////
 // Scatter for when there is no need to SIMD split
 //////////////////////////////////////////////////////
-template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,cshiftVector<vobj> &buffer, int dimension,int plane,int cbmask)
+template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,deviceVector<vobj> &buffer, int dimension,int plane,int cbmask)
 {
   int rd = rhs.Grid()->_rdimensions[dimension];
 
@@ -245,17 +202,10 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,cshiftVector<
   {
     auto buffer_p = & buffer[0];
     auto table = MapCshiftTable();
-#ifdef ACCELERATOR_CSHIFT    
-    autoView( rhs_v, rhs, AcceleratorWrite);
+    autoView( rhs_v, rhs, AcceleratorWriteDiscard);
     accelerator_for(i,ent,vobj::Nsimd(),{
 	coalescedWrite(rhs_v[table[i].first],coalescedRead(buffer_p[table[i].second]));
     });
-#else
-    autoView( rhs_v, rhs, CpuWrite);
-    thread_for(i,ent,{
-      rhs_v[table[i].first]=buffer_p[table[i].second];
-    });
-#endif
   }
 }
 
@@ -278,8 +228,7 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerA
   if(cbmask ==0x3 ) {
     int _slice_stride = rhs.Grid()->_slice_stride[dimension];
     int _slice_block = rhs.Grid()->_slice_block[dimension];
-#ifdef ACCELERATOR_CSHIFT    
-    autoView( rhs_v , rhs, AcceleratorWrite);
+    autoView( rhs_v , rhs, AcceleratorWriteDiscard);
     accelerator_for(nn,e1*e2,1,{
 	int n = nn%e1;
 	int b = nn/e1;
@@ -287,21 +236,13 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerA
 	int offset = b+n*_slice_block;
 	merge(rhs_v[so+o+b],pointers,offset);
       });
-#else
-    autoView( rhs_v , rhs, CpuWrite);
-    thread_for2d(n,e1,b,e2,{
-	int o      = n*_slice_stride;
-	int offset = b+n*_slice_block;
-	merge(rhs_v[so+o+b],pointers,offset);
-    });
-#endif
   } else { 
 
     // Case of SIMD split AND checker dim cannot currently be hit, except in 
     // Test_cshift_red_black code.
-    std::cout << "Scatter_plane merge assert(0); think this is buggy FIXME "<< std::endl;// think this is buggy FIXME
+    std::cout << "Scatter_plane merge GRID_ASSERT(0); think this is buggy FIXME "<< std::endl;// think this is buggy FIXME
     std::cout<<" Unthreaded warning -- buffer is not densely packed ??"<<std::endl;
-    assert(0); // This will fail if hit on GPU
+    GRID_ASSERT(0); // This will fail if hit on GPU
     autoView( rhs_v, rhs, CpuWrite);
     for(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
@@ -360,19 +301,11 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs
 
   {
     auto table = MapCshiftTable();
-#ifdef ACCELERATOR_CSHIFT    
     autoView(rhs_v , rhs, AcceleratorRead);
-    autoView(lhs_v , lhs, AcceleratorWrite);
+    autoView(lhs_v , lhs, AcceleratorWriteDiscard);
     accelerator_for(i,ent,vobj::Nsimd(),{
       coalescedWrite(lhs_v[table[i].first],coalescedRead(rhs_v[table[i].second]));
     });
-#else
-    autoView(rhs_v , rhs, CpuRead);
-    autoView(lhs_v , lhs, CpuWrite);
-    thread_for(i,ent,{
-      lhs_v[table[i].first]=rhs_v[table[i].second];
-    });
-#endif
   }
 }
 
@@ -412,19 +345,11 @@ template<class vobj> void Copy_plane_permute(Lattice<vobj>& lhs,const Lattice<vo
 
   {
     auto table = MapCshiftTable();
-#ifdef ACCELERATOR_CSHIFT    
     autoView( rhs_v, rhs, AcceleratorRead);
     autoView( lhs_v, lhs, AcceleratorWrite);
     accelerator_for(i,ent,1,{
       permute(lhs_v[table[i].first],rhs_v[table[i].second],permute_type);
     });
-#else
-    autoView( rhs_v, rhs, CpuRead);
-    autoView( lhs_v, lhs, CpuWrite);
-    thread_for(i,ent,{
-      permute(lhs_v[table[i].first],rhs_v[table[i].second],permute_type);
-    });
-#endif
   }
 }
 

Grid/cshift/Cshift_mpi.h

@@ -29,9 +29,13 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #ifndef _GRID_CSHIFT_MPI_H_
 #define _GRID_CSHIFT_MPI_H_
 
-
 NAMESPACE_BEGIN(Grid); 
 
+#ifdef GRID_CHECKSUM_COMMS
+extern uint64_t checksum_index;
+#endif
+
+const int Cshift_verbose=0;
 template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension,int shift)
 {
   typedef typename vobj::vector_type vector_type;
@@ -45,6 +49,20 @@ template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension
   // Map to always positive shift modulo global full dimension.
   shift = (shift+fd)%fd;
 
+  if( shift ==0 ) {
+    ret = rhs;
+    return ret;
+  }
+  //
+  // Potential easy fast cases:
+  // Shift is a multiple of the local lattice extent.
+  // Then need only to shift whole subvolumes
+  int L = rhs.Grid()->_ldimensions[dimension];
+  if ( (shift%L )==0 && !rhs.Grid()->CheckerBoarded(dimension) ) {
+    Cshift_simple(ret,rhs,dimension,shift);
+    return ret;
+  }
+  
   ret.Checkerboard() = rhs.Grid()->CheckerBoardDestination(rhs.Checkerboard(),shift,dimension);
         
   // the permute type
@@ -55,20 +73,69 @@ template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension
   RealD t1,t0;
   t0=usecond();
   if ( !comm_dim ) {
-    //std::cout << "CSHIFT: Cshift_local" <<std::endl;
+    //    std::cout << "CSHIFT: Cshift_local" <<std::endl;
     Cshift_local(ret,rhs,dimension,shift); // Handles checkerboarding
   } else if ( splice_dim ) {
-    //std::cout << "CSHIFT: Cshift_comms_simd call - splice_dim = " << splice_dim << " shift " << shift << " dimension = " << dimension << std::endl;
+    //    std::cout << "CSHIFT: Cshift_comms_simd call - splice_dim = " << splice_dim << " shift " << shift << " dimension = " << dimension << std::endl;
     Cshift_comms_simd(ret,rhs,dimension,shift);
   } else {
-    //std::cout << "CSHIFT: Cshift_comms" <<std::endl;
+    //    std::cout << "CSHIFT: Cshift_comms" <<std::endl;
     Cshift_comms(ret,rhs,dimension,shift);
   }
   t1=usecond();
-  //  std::cout << GridLogPerformance << "Cshift took "<< (t1-t0)/1e3 << " ms"<<std::endl;
+  if(Cshift_verbose) std::cout << GridLogPerformance << "Cshift took "<< (t1-t0)/1e3 << " ms"<<std::endl;
   return ret;
 }
 
+template<class vobj> void Cshift_simple(Lattice<vobj>& ret,const Lattice<vobj> &rhs,int dimension,int shift)
+{
+  GridBase *grid=rhs.Grid();
+  int comm_proc, xmit_to_rank, recv_from_rank;
+  
+  int fd              = rhs.Grid()->_fdimensions[dimension];
+  int rd              = rhs.Grid()->_rdimensions[dimension];
+  int ld              = rhs.Grid()->_ldimensions[dimension];
+  int pd              = rhs.Grid()->_processors[dimension];
+  int simd_layout     = rhs.Grid()->_simd_layout[dimension];
+  int comm_dim        = rhs.Grid()->_processors[dimension] >1 ;
+
+  comm_proc = ((shift)/ld)%pd;
+
+  grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
+  if(comm_dim) {
+
+    int64_t bytes = sizeof(vobj) * grid->oSites();
+
+    autoView(rhs_v , rhs, AcceleratorRead);
+    autoView(ret_v , ret, AcceleratorWrite);
+    void *send_buf  = (void *)&rhs_v[0];
+    void *recv_buf  = (void *)&ret_v[0];
+
+#ifdef ACCELERATOR_AWARE_MPI
+    grid->SendToRecvFrom(send_buf,
+			 xmit_to_rank,
+			 recv_buf,
+			 recv_from_rank,
+			 bytes);
+#else
+    static hostVector<vobj> hrhs; hrhs.resize(grid->oSites());
+    static hostVector<vobj> hret; hret.resize(grid->oSites());
+
+    void *hsend_buf = (void *)&hrhs[0];
+    void *hrecv_buf = (void *)&hret[0];
+
+    acceleratorCopyFromDevice(send_buf,hsend_buf,bytes);
+
+    grid->SendToRecvFrom(hsend_buf,
+			 xmit_to_rank,
+			 hrecv_buf,
+			 recv_from_rank,
+			 bytes);
+
+    acceleratorCopyToDevice(hrecv_buf,recv_buf,bytes);
+#endif
+  }
+}
 template<class vobj> void Cshift_comms(Lattice<vobj>& ret,const Lattice<vobj> &rhs,int dimension,int shift)
 {
   int sshift[2];
@@ -94,18 +161,16 @@ template<class vobj> void Cshift_comms_simd(Lattice<vobj>& ret,const Lattice<vob
   sshift[0] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Even);
   sshift[1] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Odd);
 
-  //std::cout << "Cshift_comms_simd dim "<<dimension<<"cb "<<rhs.checkerboard<<"shift "<<shift<<" sshift " << sshift[0]<<" "<<sshift[1]<<std::endl;
+  //  std::cout << "Cshift_comms_simd dim "<<dimension<<"cb "<<rhs.Checkerboard()<<"shift "<<shift<<" sshift " << sshift[0]<<" "<<sshift[1]<<std::endl;
   if ( sshift[0] == sshift[1] ) {
-    //std::cout << "Single pass Cshift_comms" <<std::endl;
+    //    std::cout << "Single pass Cshift_comms" <<std::endl;
     Cshift_comms_simd(ret,rhs,dimension,shift,0x3);
   } else {
-    //std::cout << "Two pass Cshift_comms" <<std::endl;
+    //    std::cout << "Two pass Cshift_comms" <<std::endl;
     Cshift_comms_simd(ret,rhs,dimension,shift,0x1);// if checkerboard is unfavourable take two passes
     Cshift_comms_simd(ret,rhs,dimension,shift,0x2);// both with block stride loop iteration
   }
 }
-#define ACCELERATOR_CSHIFT_NO_COPY
-#ifdef ACCELERATOR_CSHIFT_NO_COPY
 template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
 {
   typedef typename vobj::vector_type vector_type;
@@ -119,15 +184,20 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
   int pd              = rhs.Grid()->_processors[dimension];
   int simd_layout     = rhs.Grid()->_simd_layout[dimension];
   int comm_dim        = rhs.Grid()->_processors[dimension] >1 ;
-  assert(simd_layout==1);
-  assert(comm_dim==1);
-  assert(shift>=0);
-  assert(shift<fd);
+  GRID_ASSERT(simd_layout==1);
+  GRID_ASSERT(comm_dim==1);
+  GRID_ASSERT(shift>=0);
+  GRID_ASSERT(shift<fd);
   
   int buffer_size = rhs.Grid()->_slice_nblock[dimension]*rhs.Grid()->_slice_block[dimension];
-  static cshiftVector<vobj> send_buf; send_buf.resize(buffer_size);
-  static cshiftVector<vobj> recv_buf; recv_buf.resize(buffer_size);
-    
+  static deviceVector<vobj> send_buf; send_buf.resize(buffer_size);
+  static deviceVector<vobj> recv_buf; recv_buf.resize(buffer_size);
+#ifndef ACCELERATOR_AWARE_MPI
+  int pad = (8 + sizeof(vobj) - 1) / sizeof(vobj);
+  static hostVector<vobj> hsend_buf; hsend_buf.resize(buffer_size+pad);
+  static hostVector<vobj> hrecv_buf; hrecv_buf.resize(buffer_size+pad);
+#endif
+  
   int cb= (cbmask==0x2)? Odd : Even;
   int sshift= rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
   RealD tcopy=0.0;
@@ -141,9 +211,11 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
     int comm_proc = ((x+sshift)/rd)%pd;
     
     if (comm_proc==0) {
+      FlightRecorder::StepLog("Cshift_Copy_plane");
       tcopy-=usecond();
       Copy_plane(ret,rhs,dimension,x,sx,cbmask); 
       tcopy+=usecond();
+      FlightRecorder::StepLog("Cshift_Copy_plane_complete");
     } else {
 
       int words = buffer_size;
@@ -151,39 +223,84 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
 
       int bytes = words * sizeof(vobj);
 
+      FlightRecorder::StepLog("Cshift_Gather_plane");
       tgather-=usecond();
       Gather_plane_simple (rhs,send_buf,dimension,sx,cbmask);
       tgather+=usecond();
+      FlightRecorder::StepLog("Cshift_Gather_plane_complete");
 
       //      int rank           = grid->_processor;
       int recv_from_rank;
       int xmit_to_rank;
+
       grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
       
       tcomms-=usecond();
-      //      grid->Barrier();
+      grid->Barrier();
 
+      FlightRecorder::StepLog("Cshift_SendRecv");
+#ifdef ACCELERATOR_AWARE_MPI
       grid->SendToRecvFrom((void *)&send_buf[0],
 			   xmit_to_rank,
 			   (void *)&recv_buf[0],
 			   recv_from_rank,
 			   bytes);
+#else
+      // bouncy bouncy
+      acceleratorCopyFromDevice(&send_buf[0],&hsend_buf[0],bytes);
+
+#ifdef GRID_CHECKSUM_COMMS
+      GRID_ASSERT(bytes % 8 == 0);
+      checksum_index++;
+      uint64_t xsum = checksum_gpu((uint64_t*)&send_buf[0], bytes / 8) ^ (1 + checksum_index);
+      *(uint64_t*)(((char*)&hsend_buf[0]) + bytes) = xsum;
+      bytes += 8;
+#endif
+
+      grid->SendToRecvFrom((void *)&hsend_buf[0],
+			   xmit_to_rank,
+			   (void *)&hrecv_buf[0],
+			   recv_from_rank,
+			   bytes);
+
+#ifdef GRID_CHECKSUM_COMMS
+      bytes -= 8;
+      acceleratorCopyToDevice(&hrecv_buf[0],&recv_buf[0],bytes);
+      uint64_t expected_cs = *(uint64_t*)(((char*)&hrecv_buf[0]) + bytes);
+      uint64_t computed_cs = checksum_gpu((uint64_t*)&recv_buf[0], bytes / 8) ^ (1 + checksum_index);
+      std::cout << GridLogComms<< " Cshift: "
+		<<" dim"<<dimension
+		<<" shift "<<shift
+		<< " rank "<< grid->ThisRank()
+		<<" Coor "<<grid->ThisProcessorCoor()
+		<<" send "<<xsum<<" to   "<<xmit_to_rank
+		<<" recv "<<computed_cs<<" from "<<recv_from_rank
+		<<std::endl;
+      GRID_ASSERT(expected_cs == computed_cs);
+#else
+      acceleratorCopyToDevice(&hrecv_buf[0],&recv_buf[0],bytes);
+#endif
+
+#endif
+      FlightRecorder::StepLog("Cshift_SendRecv_complete");
+
       xbytes+=bytes;
-      //      grid->Barrier();
+      grid->Barrier();
       tcomms+=usecond();
+      FlightRecorder::StepLog("Cshift_barrier_complete");
 
       tscatter-=usecond();
       Scatter_plane_simple (ret,recv_buf,dimension,x,cbmask);
       tscatter+=usecond();
     }
   }
-  /*
-  std::cout << GridLogPerformance << " Cshift copy    "<<tcopy/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift gather  "<<tgather/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift scatter "<<tscatter/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift comm    "<<tcomms/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s "<<2*xbytes<< " Bytes "<<std::endl;
-  */
+  if (Cshift_verbose){
+    std::cout << GridLogPerformance << " Cshift copy    "<<tcopy/1e3<<" ms"<<std::endl;
+    std::cout << GridLogPerformance << " Cshift gather  "<<tgather/1e3<<" ms"<<std::endl;
+    std::cout << GridLogPerformance << " Cshift scatter "<<tscatter/1e3<<" ms"<<std::endl;
+    std::cout << GridLogPerformance << " Cshift comm    "<<tcomms/1e3<<" ms"<<std::endl;
+    std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s "<<2*xbytes<< " Bytes "<<std::endl;
+  }
 }
 
 template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
@@ -201,14 +318,14 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
   int simd_layout     = grid->_simd_layout[dimension];
   int comm_dim        = grid->_processors[dimension] >1 ;
 
-  //std::cout << "Cshift_comms_simd dim "<< dimension << " fd "<<fd<<" rd "<<rd
-  //    << " ld "<<ld<<" pd " << pd<<" simd_layout "<<simd_layout 
-  //    << " comm_dim " << comm_dim << " cbmask " << cbmask <<std::endl;
+  //  std::cout << "Cshift_comms_simd dim "<< dimension << " fd "<<fd<<" rd "<<rd
+  //	    << " ld "<<ld<<" pd " << pd<<" simd_layout "<<simd_layout 
+  //	    << " comm_dim " << comm_dim << " cbmask " << cbmask <<std::endl;
 
-  assert(comm_dim==1);
-  assert(simd_layout==2);
-  assert(shift>=0);
-  assert(shift<fd);
+  GRID_ASSERT(comm_dim==1);
+  GRID_ASSERT(simd_layout==2);
+  GRID_ASSERT(shift>=0);
+  GRID_ASSERT(shift<fd);
 
   RealD tcopy=0.0;
   RealD tgather=0.0;
@@ -224,16 +341,28 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
   int buffer_size = grid->_slice_nblock[dimension]*grid->_slice_block[dimension];
   //  int words = sizeof(vobj)/sizeof(vector_type);
 
-  static std::vector<cshiftVector<scalar_object> >  send_buf_extract; send_buf_extract.resize(Nsimd);
-  static std::vector<cshiftVector<scalar_object> >  recv_buf_extract; recv_buf_extract.resize(Nsimd);
+  static std::vector<deviceVector<scalar_object> >  send_buf_extract; send_buf_extract.resize(Nsimd);
+  static std::vector<deviceVector<scalar_object> >  recv_buf_extract; recv_buf_extract.resize(Nsimd);
   scalar_object *  recv_buf_extract_mpi;
   scalar_object *  send_buf_extract_mpi;
- 
+
   for(int s=0;s<Nsimd;s++){
     send_buf_extract[s].resize(buffer_size);
     recv_buf_extract[s].resize(buffer_size);
   }
+#ifndef ACCELERATOR_AWARE_MPI
+#ifdef GRID_CHECKSUM_COMMS
+  buffer_size += (8 + sizeof(vobj) - 1) / sizeof(vobj);
+#endif
 
+  static hostVector<vobj> hsend_buf; hsend_buf.resize(buffer_size);
+  static hostVector<vobj> hrecv_buf; hrecv_buf.resize(buffer_size);
+
+#ifdef GRID_CHECKSUM_COMMS
+  buffer_size -= (8 + sizeof(vobj) - 1) / sizeof(vobj);
+#endif
+#endif
+  
   int bytes = buffer_size*sizeof(scalar_object);
 
   ExtractPointerArray<scalar_object>  pointers(Nsimd); // 
@@ -275,272 +404,81 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
 
       if (nbr_ic) nbr_lane|=inner_bit;
 
-      assert (sx == nbr_ox);
+      GRID_ASSERT (sx == nbr_ox);
 
       if(nbr_proc){
 	grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank); 
 
 	tcomms-=usecond();
-	//	grid->Barrier();
+	grid->Barrier();
 
 	send_buf_extract_mpi = &send_buf_extract[nbr_lane][0];
 	recv_buf_extract_mpi = &recv_buf_extract[i][0];
+#ifdef ACCELERATOR_AWARE_MPI
 	grid->SendToRecvFrom((void *)send_buf_extract_mpi,
 			     xmit_to_rank,
 			     (void *)recv_buf_extract_mpi,
 			     recv_from_rank,
 			     bytes);
-
-	xbytes+=bytes;
-	//	grid->Barrier();
-	tcomms+=usecond();
-
-	rpointers[i] = &recv_buf_extract[i][0];
-      } else { 
-	rpointers[i] = &send_buf_extract[nbr_lane][0];
-      }
-
-    }
-    tscatter-=usecond();
-    Scatter_plane_merge(ret,rpointers,dimension,x,cbmask);
-    tscatter+=usecond();
-  }
-  /*
-  std::cout << GridLogPerformance << " Cshift (s) copy    "<<tcopy/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift (s) gather  "<<tgather/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift (s) scatter "<<tscatter/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift (s) comm    "<<tcomms/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s "<<2*xbytes<< " Bytes "<<std::endl;
-  */
-}
-#else 
-template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
-{
-  typedef typename vobj::vector_type vector_type;
-  typedef typename vobj::scalar_type scalar_type;
-
-  GridBase *grid=rhs.Grid();
-  Lattice<vobj> temp(rhs.Grid());
-
-  int fd              = rhs.Grid()->_fdimensions[dimension];
-  int rd              = rhs.Grid()->_rdimensions[dimension];
-  int pd              = rhs.Grid()->_processors[dimension];
-  int simd_layout     = rhs.Grid()->_simd_layout[dimension];
-  int comm_dim        = rhs.Grid()->_processors[dimension] >1 ;
-  assert(simd_layout==1);
-  assert(comm_dim==1);
-  assert(shift>=0);
-  assert(shift<fd);
-  RealD tcopy=0.0;
-  RealD tgather=0.0;
-  RealD tscatter=0.0;
-  RealD tcomms=0.0;
-  uint64_t xbytes=0;
-  
-  int buffer_size = rhs.Grid()->_slice_nblock[dimension]*rhs.Grid()->_slice_block[dimension];
-  static cshiftVector<vobj> send_buf_v; send_buf_v.resize(buffer_size);
-  static cshiftVector<vobj> recv_buf_v; recv_buf_v.resize(buffer_size);
-  vobj *send_buf;
-  vobj *recv_buf;
-  {
-    grid->ShmBufferFreeAll();
-    size_t bytes = buffer_size*sizeof(vobj);
-    send_buf=(vobj *)grid->ShmBufferMalloc(bytes);
-    recv_buf=(vobj *)grid->ShmBufferMalloc(bytes);
-  }
-    
-  int cb= (cbmask==0x2)? Odd : Even;
-  int sshift= rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
-
-  for(int x=0;x<rd;x++){       
-
-    int sx        =  (x+sshift)%rd;
-    int comm_proc = ((x+sshift)/rd)%pd;
-    
-    if (comm_proc==0) {
-
-      tcopy-=usecond();
-      Copy_plane(ret,rhs,dimension,x,sx,cbmask); 
-      tcopy+=usecond();
-
-    } else {
-
-      int words = buffer_size;
-      if (cbmask != 0x3) words=words>>1;
-
-      int bytes = words * sizeof(vobj);
-
-      tgather-=usecond();
-      Gather_plane_simple (rhs,send_buf_v,dimension,sx,cbmask);
-      tgather+=usecond();
-
-      //      int rank           = grid->_processor;
-      int recv_from_rank;
-      int xmit_to_rank;
-      grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
-
-
-      tcomms-=usecond();
-      //      grid->Barrier();
-
-      acceleratorCopyDeviceToDevice((void *)&send_buf_v[0],(void *)&send_buf[0],bytes);
-      grid->SendToRecvFrom((void *)&send_buf[0],
-			   xmit_to_rank,
-			   (void *)&recv_buf[0],
-			   recv_from_rank,
-			   bytes);
-      xbytes+=bytes;
-      acceleratorCopyDeviceToDevice((void *)&recv_buf[0],(void *)&recv_buf_v[0],bytes);
-
-      //      grid->Barrier();
-      tcomms+=usecond();
-
-      tscatter-=usecond();
-      Scatter_plane_simple (ret,recv_buf_v,dimension,x,cbmask);
-      tscatter+=usecond();
-    }
-  }
-  /*
-  std::cout << GridLogPerformance << " Cshift copy    "<<tcopy/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift gather  "<<tgather/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift scatter "<<tscatter/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift comm    "<<tcomms/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s "<<2*xbytes<< " Bytes "<<std::endl;
-  */
-}
-
-template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
-{
-  GridBase *grid=rhs.Grid();
-  const int Nsimd = grid->Nsimd();
-  typedef typename vobj::vector_type vector_type;
-  typedef typename vobj::scalar_object scalar_object;
-  typedef typename vobj::scalar_type scalar_type;
-   
-  int fd = grid->_fdimensions[dimension];
-  int rd = grid->_rdimensions[dimension];
-  int ld = grid->_ldimensions[dimension];
-  int pd = grid->_processors[dimension];
-  int simd_layout     = grid->_simd_layout[dimension];
-  int comm_dim        = grid->_processors[dimension] >1 ;
-
-  //std::cout << "Cshift_comms_simd dim "<< dimension << " fd "<<fd<<" rd "<<rd
-  //    << " ld "<<ld<<" pd " << pd<<" simd_layout "<<simd_layout 
-  //    << " comm_dim " << comm_dim << " cbmask " << cbmask <<std::endl;
-
-  assert(comm_dim==1);
-  assert(simd_layout==2);
-  assert(shift>=0);
-  assert(shift<fd);
-  RealD tcopy=0.0;
-  RealD tgather=0.0;
-  RealD tscatter=0.0;
-  RealD tcomms=0.0;
-  uint64_t xbytes=0;
-
-  int permute_type=grid->PermuteType(dimension);
-
-  ///////////////////////////////////////////////
-  // Simd direction uses an extract/merge pair
-  ///////////////////////////////////////////////
-  int buffer_size = grid->_slice_nblock[dimension]*grid->_slice_block[dimension];
-  //  int words = sizeof(vobj)/sizeof(vector_type);
-
-  static std::vector<cshiftVector<scalar_object> >  send_buf_extract; send_buf_extract.resize(Nsimd);
-  static std::vector<cshiftVector<scalar_object> >  recv_buf_extract; recv_buf_extract.resize(Nsimd);
-  scalar_object *  recv_buf_extract_mpi;
-  scalar_object *  send_buf_extract_mpi;
-  {
-    size_t bytes = sizeof(scalar_object)*buffer_size;
-    grid->ShmBufferFreeAll();
-    send_buf_extract_mpi = (scalar_object *)grid->ShmBufferMalloc(bytes);
-    recv_buf_extract_mpi = (scalar_object *)grid->ShmBufferMalloc(bytes);
-  }
-  for(int s=0;s<Nsimd;s++){
-    send_buf_extract[s].resize(buffer_size);
-    recv_buf_extract[s].resize(buffer_size);
-  }
-
-  int bytes = buffer_size*sizeof(scalar_object);
-
-  ExtractPointerArray<scalar_object>  pointers(Nsimd); // 
-  ExtractPointerArray<scalar_object> rpointers(Nsimd); // received pointers
-
-  ///////////////////////////////////////////
-  // Work out what to send where
-  ///////////////////////////////////////////
-  int cb    = (cbmask==0x2)? Odd : Even;
-  int sshift= grid->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
-
-  // loop over outer coord planes orthog to dim
-  for(int x=0;x<rd;x++){       
-
-    // FIXME call local permute copy if none are offnode.
-    for(int i=0;i<Nsimd;i++){       
-      pointers[i] = &send_buf_extract[i][0];
-    }
-    tgather-=usecond();
-    int sx   = (x+sshift)%rd;
-    Gather_plane_extract(rhs,pointers,dimension,sx,cbmask);
-    tgather+=usecond();
-
-    for(int i=0;i<Nsimd;i++){
-      
-      int inner_bit = (Nsimd>>(permute_type+1));
-      int ic= (i&inner_bit)? 1:0;
-
-      int my_coor          = rd*ic + x;
-      int nbr_coor         = my_coor+sshift;
-      int nbr_proc = ((nbr_coor)/ld) % pd;// relative shift in processors
-
-      int nbr_ic   = (nbr_coor%ld)/rd;    // inner coord of peer
-      int nbr_ox   = (nbr_coor%rd);       // outer coord of peer
-      int nbr_lane = (i&(~inner_bit));
-
-      int recv_from_rank;
-      int xmit_to_rank;
-
-      if (nbr_ic) nbr_lane|=inner_bit;
-
-      assert (sx == nbr_ox);
-
-      if(nbr_proc){
-	grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank); 
-
-	tcomms-=usecond();
-	//	grid->Barrier();
-
-	acceleratorCopyDeviceToDevice((void *)&send_buf_extract[nbr_lane][0],(void *)send_buf_extract_mpi,bytes);
-	grid->SendToRecvFrom((void *)send_buf_extract_mpi,
-			     xmit_to_rank,
-			     (void *)recv_buf_extract_mpi,
-			     recv_from_rank,
-			     bytes);
-	acceleratorCopyDeviceToDevice((void *)recv_buf_extract_mpi,(void *)&recv_buf_extract[i][0],bytes);
-	xbytes+=bytes;
-
-	//	grid->Barrier();
-	tcomms+=usecond();
-	rpointers[i] = &recv_buf_extract[i][0];
-      } else { 
-	rpointers[i] = &send_buf_extract[nbr_lane][0];
-      }
-
-    }
-    tscatter-=usecond();
-    Scatter_plane_merge(ret,rpointers,dimension,x,cbmask);
-    tscatter+=usecond();
-
-  }
-  /*
-  std::cout << GridLogPerformance << " Cshift (s) copy    "<<tcopy/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift (s) gather  "<<tgather/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift (s) scatter "<<tscatter/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift (s) comm    "<<tcomms/1e3<<" ms"<<std::endl;
-  std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s"<<std::endl;
-  */
-}
+#else
+      // bouncy bouncy
+	acceleratorCopyFromDevice((void *)send_buf_extract_mpi,(void *)&hsend_buf[0],bytes);
+#ifdef GRID_CHECKSUM_COMMS
+	assert(bytes % 8 == 0);
+	checksum_index++;
+	uint64_t xsum = checksum_gpu((uint64_t*)send_buf_extract_mpi, bytes / 8) ^ (1 + checksum_index);
+	*(uint64_t*)(((char*)&hsend_buf[0]) + bytes) = xsum;
+	bytes += 8;
 #endif
+	grid->SendToRecvFrom((void *)&hsend_buf[0],
+			     xmit_to_rank,
+			     (void *)&hrecv_buf[0],
+			     recv_from_rank,
+			     bytes);
+#ifdef GRID_CHECKSUM_COMMS
+	bytes -= 8;
+	acceleratorCopyToDevice((void *)&hrecv_buf[0],(void *)recv_buf_extract_mpi,bytes);
+	uint64_t expected_cs = *(uint64_t*)(((char*)&hrecv_buf[0]) + bytes);
+	uint64_t computed_cs = checksum_gpu((uint64_t*)recv_buf_extract_mpi, bytes / 8) ^ (1 + checksum_index);
+
+	std::cout << GridLogComms<< " Cshift_comms_simd: "
+		<<" dim"<<dimension
+		<<" shift "<<shift
+		<< " rank "<< grid->ThisRank()
+		<<" Coor "<<grid->ThisProcessorCoor()
+		<<" send "<<xsum<<" to   "<<xmit_to_rank
+		<<" recv "<<computed_cs<<" from "<<recv_from_rank
+		<<std::endl;
+	assert(expected_cs == computed_cs);
+#else
+	acceleratorCopyToDevice((void *)&hrecv_buf[0],(void *)recv_buf_extract_mpi,bytes);
+#endif
+
+#endif
+
+	xbytes+=bytes;
+	grid->Barrier();
+	tcomms+=usecond();
+
+	rpointers[i] = &recv_buf_extract[i][0];
+      } else { 
+	rpointers[i] = &send_buf_extract[nbr_lane][0];
+      }
+
+    }
+    tscatter-=usecond();
+    Scatter_plane_merge(ret,rpointers,dimension,x,cbmask);
+    tscatter+=usecond();
+  }
+  if(Cshift_verbose){
+    std::cout << GridLogPerformance << " Cshift (s) copy    "<<tcopy/1e3<<" ms"<<std::endl;
+    std::cout << GridLogPerformance << " Cshift (s) gather  "<<tgather/1e3<<" ms"<<std::endl;
+    std::cout << GridLogPerformance << " Cshift (s) scatter "<<tscatter/1e3<<" ms"<<std::endl;
+    std::cout << GridLogPerformance << " Cshift (s) comm    "<<tcomms/1e3<<" ms"<<std::endl;
+    std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s "<<2*xbytes<< " Bytes "<<std::endl;
+  }
+}
+
 NAMESPACE_END(Grid); 
 
 #endif

Grid/cshift/Cshift_table.cc

@@ -1,5 +1,5 @@
 #include <Grid/GridCore.h>       
 NAMESPACE_BEGIN(Grid);
 std::vector<std::pair<int,int> > Cshift_table; 
-commVector<std::pair<int,int> > Cshift_table_device; 
+deviceVector<std::pair<int,int> > Cshift_table_device; 
 NAMESPACE_END(Grid);

Grid/json/json.hpp

@@ -21895,7 +21895,7 @@ inline void swap(nlohmann::NLOHMANN_BASIC_JSON_TPL& j1, nlohmann::NLOHMANN_BASIC
 /// @brief user-defined string literal for JSON values
 /// @sa https://json.nlohmann.me/api/basic_json/operator_literal_json/
 JSON_HEDLEY_NON_NULL(1)
-inline nlohmann::json operator "" _json(const char* s, std::size_t n)
+inline nlohmann::json operator ""_json(const char* s, std::size_t n)
 {
     return nlohmann::json::parse(s, s + n);
 }
@@ -21903,7 +21903,7 @@ inline nlohmann::json operator "" _json(const char* s, std::size_t n)
 /// @brief user-defined string literal for JSON pointer
 /// @sa https://json.nlohmann.me/api/basic_json/operator_literal_json_pointer/
 JSON_HEDLEY_NON_NULL(1)
-inline nlohmann::json::json_pointer operator "" _json_pointer(const char* s, std::size_t n)
+inline nlohmann::json::json_pointer operator ""_json_pointer(const char* s, std::size_t n)
 {
     return nlohmann::json::json_pointer(std::string(s, n));
 }

Grid/lattice/Lattice_ET.h

@@ -74,7 +74,7 @@ accelerator_inline vobj predicatedWhere(const iobj &predicate,
 
   extract(iftrue, truevals);
   extract(iffalse, falsevals);
-  extract<vInteger, Integer>(TensorRemove(predicate), mask);
+  extract(TensorRemove(predicate), mask);
 
   for (int s = 0; s < Nsimd; s++) {
     if (mask[s]) falsevals[s] = truevals[s];
@@ -245,7 +245,7 @@ template <class T1,typename std::enable_if<is_lattice<T1>::value, T1>::type * =
 inline void CBFromExpression(int &cb, const T1 &lat)  // Lattice leaf
 {
   if ((cb == Odd) || (cb == Even)) {
-    assert(cb == lat.Checkerboard());
+    GRID_ASSERT(cb == lat.Checkerboard());
   }
   cb = lat.Checkerboard();
 }

Grid/lattice/Lattice_arith.h

@@ -257,17 +257,30 @@ void axpby(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice
   });
 }
 
+#define FAST_AXPY_NORM
 template<class sobj,class vobj> inline
 RealD axpy_norm(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y)
 {
   GRID_TRACE("axpy_norm");
-    return axpy_norm_fast(ret,a,x,y);
+#ifdef FAST_AXPY_NORM
+  return axpy_norm_fast(ret,a,x,y);
+#else
+  ret = a*x+y;
+  RealD nn=norm2(ret);
+  return nn;
+#endif
 }
 template<class sobj,class vobj> inline
 RealD axpby_norm(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y)
 {
   GRID_TRACE("axpby_norm");
-    return axpby_norm_fast(ret,a,b,x,y);
+#ifdef FAST_AXPY_NORM
+  return axpby_norm_fast(ret,a,b,x,y);
+#else
+  ret = a*x+b*y;
+  RealD nn=norm2(ret);
+  return nn;
+#endif
 }
 
 /// Trace product

Grid/lattice/Lattice_base.h

@@ -120,12 +120,12 @@ public:
     GRID_TRACE("ExpressionTemplateEval");
     GridBase *egrid(nullptr);
     GridFromExpression(egrid,expr);
-    assert(egrid!=nullptr);
+    GRID_ASSERT(egrid!=nullptr);
     conformable(this->_grid,egrid);
 
     int cb=-1;
     CBFromExpression(cb,expr);
-    assert( (cb==Odd) || (cb==Even));
+    GRID_ASSERT( (cb==Odd) || (cb==Even));
     this->checkerboard=cb;
     
     auto exprCopy = expr;
@@ -144,12 +144,12 @@ public:
     GRID_TRACE("ExpressionTemplateEval");
     GridBase *egrid(nullptr);
     GridFromExpression(egrid,expr);
-    assert(egrid!=nullptr);
+    GRID_ASSERT(egrid!=nullptr);
     conformable(this->_grid,egrid);
 
     int cb=-1;
     CBFromExpression(cb,expr);
-    assert( (cb==Odd) || (cb==Even));
+    GRID_ASSERT( (cb==Odd) || (cb==Even));
     this->checkerboard=cb;
 
     auto exprCopy = expr;
@@ -168,12 +168,12 @@ public:
     GRID_TRACE("ExpressionTemplateEval");
     GridBase *egrid(nullptr);
     GridFromExpression(egrid,expr);
-    assert(egrid!=nullptr);
+    GRID_ASSERT(egrid!=nullptr);
     conformable(this->_grid,egrid);
 
     int cb=-1;
     CBFromExpression(cb,expr);
-    assert( (cb==Odd) || (cb==Even));
+    GRID_ASSERT( (cb==Odd) || (cb==Even));
     this->checkerboard=cb;
     auto exprCopy = expr;
     ExpressionViewOpen(exprCopy);
@@ -191,11 +191,11 @@ public:
   Lattice(const LatticeUnaryExpression<Op,T1> & expr) {
     this->_grid = nullptr;
     GridFromExpression(this->_grid,expr);
-    assert(this->_grid!=nullptr);
+    GRID_ASSERT(this->_grid!=nullptr);
 
     int cb=-1;
     CBFromExpression(cb,expr);
-    assert( (cb==Odd) || (cb==Even));
+    GRID_ASSERT( (cb==Odd) || (cb==Even));
     this->checkerboard=cb;
 
     resize(this->_grid->oSites());
@@ -206,11 +206,11 @@ public:
   Lattice(const LatticeBinaryExpression<Op,T1,T2> & expr) {
     this->_grid = nullptr;
     GridFromExpression(this->_grid,expr);
-    assert(this->_grid!=nullptr);
+    GRID_ASSERT(this->_grid!=nullptr);
 
     int cb=-1;
     CBFromExpression(cb,expr);
-    assert( (cb==Odd) || (cb==Even));
+    GRID_ASSERT( (cb==Odd) || (cb==Even));
     this->checkerboard=cb;
 
     resize(this->_grid->oSites());
@@ -221,11 +221,11 @@ public:
   Lattice(const LatticeTrinaryExpression<Op,T1,T2,T3> & expr) {
     this->_grid = nullptr;
     GridFromExpression(this->_grid,expr);
-    assert(this->_grid!=nullptr);
+    GRID_ASSERT(this->_grid!=nullptr);
 
     int cb=-1;
     CBFromExpression(cb,expr);
-    assert( (cb==Odd) || (cb==Even));
+    GRID_ASSERT( (cb==Odd) || (cb==Even));
     this->checkerboard=cb;
 
     resize(this->_grid->oSites());
@@ -236,10 +236,13 @@ public:
   template<class sobj> inline Lattice<vobj> & operator = (const sobj & r){
     vobj vtmp;
     vtmp = r;
-#if 0
+#if 1
+    deviceVector<vobj> vvtmp(1);
+    acceleratorPut(vvtmp[0],vtmp);
+    vobj *vvtmp_p = & vvtmp[0];
     auto me  = View(AcceleratorWrite);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
-	auto stmp=coalescedRead(vtmp);
+	auto stmp=coalescedRead(*vvtmp_p);
 	coalescedWrite(me[ss],stmp);
     });
 #else    
@@ -261,7 +264,7 @@ public:
   Lattice(GridBase *grid,ViewMode mode=AcceleratorWriteDiscard) { 
     this->_grid = grid;
     resize(this->_grid->oSites());
-    assert((((uint64_t)&this->_odata[0])&0xF) ==0);
+    GRID_ASSERT((((uint64_t)&this->_odata[0])&0xF) ==0);
     this->checkerboard=0;
     SetViewMode(mode);
   }
@@ -286,7 +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;
@@ -327,7 +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_basis.h

@@ -53,36 +53,19 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
   typedef decltype(basis[0]) Field;
   typedef decltype(basis[0].View(AcceleratorRead)) View;
 
-  Vector<View> basis_v; basis_v.reserve(basis.size());
-  typedef typename std::remove_reference<decltype(basis_v[0][0])>::type vobj;
+  hostVector<View>  h_basis_v(basis.size());
+  deviceVector<View> d_basis_v(basis.size());
+  typedef typename std::remove_reference<decltype(h_basis_v[0][0])>::type vobj;
   typedef typename std::remove_reference<decltype(Qt(0,0))>::type Coeff_t;
+
   GridBase* grid = basis[0].Grid();
       
   for(int k=0;k<basis.size();k++){
-    basis_v.push_back(basis[k].View(AcceleratorWrite));
+    h_basis_v[k] = basis[k].View(AcceleratorWrite);
+    acceleratorPut(d_basis_v[k],h_basis_v[k]);
   }
 
-#if ( !(defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)) )
-  int max_threads = thread_max();
-  Vector < vobj > Bt(Nm * max_threads);
-  thread_region
-    {
-      vobj* B = &Bt[Nm * thread_num()];
-      thread_for_in_region(ss, grid->oSites(),{
-	  for(int j=j0; j<j1; ++j) B[j]=0.;
-      
-	  for(int j=j0; j<j1; ++j){
-	    for(int k=k0; k<k1; ++k){
-	      B[j] +=Qt(j,k) * basis_v[k][ss];
-	    }
-	  }
-	  for(int j=j0; j<j1; ++j){
-	    basis_v[j][ss] = B[j];
-	  }
-	});
-    }
-#else
-  View *basis_vp = &basis_v[0];
+  View *basis_vp = &d_basis_v[0];
 
   int nrot = j1-j0;
   if (!nrot) // edge case not handled gracefully by Cuda
@@ -91,17 +74,19 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
   uint64_t oSites   =grid->oSites();
   uint64_t siteBlock=(grid->oSites()+nrot-1)/nrot; // Maximum 1 additional vector overhead
 
-  Vector <vobj> Bt(siteBlock * nrot); 
+  deviceVector <vobj> Bt(siteBlock * nrot); 
   auto Bp=&Bt[0];
 
   // GPU readable copy of matrix
-  Vector<Coeff_t> Qt_jv(Nm*Nm);
+  hostVector<Coeff_t> h_Qt_jv(Nm*Nm);
+  deviceVector<Coeff_t> Qt_jv(Nm*Nm);
   Coeff_t *Qt_p = & Qt_jv[0];
   thread_for(i,Nm*Nm,{
       int j = i/Nm;
       int k = i%Nm;
-      Qt_p[i]=Qt(j,k);
+      h_Qt_jv[i]=Qt(j,k);
   });
+  acceleratorCopyToDevice(&h_Qt_jv[0],Qt_p,Nm*Nm*sizeof(Coeff_t));
 
   // Block the loop to keep storage footprint down
   for(uint64_t s=0;s<oSites;s+=siteBlock){
@@ -137,9 +122,8 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
 	coalescedWrite(basis_vp[jj][sss],coalescedRead(Bp[ss*nrot+j]));
       });
   }
-#endif
 
-  for(int k=0;k<basis.size();k++) basis_v[k].ViewClose();
+  for(int k=0;k<basis.size();k++) h_basis_v[k].ViewClose();
 }
 
 // Extract a single rotated vector
@@ -152,16 +136,19 @@ void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,in
 
   result.Checkerboard() = basis[0].Checkerboard();
 
-  Vector<View> basis_v; basis_v.reserve(basis.size());
+  hostVector<View>  h_basis_v(basis.size());
+  deviceVector<View> d_basis_v(basis.size());
   for(int k=0;k<basis.size();k++){
-    basis_v.push_back(basis[k].View(AcceleratorRead));
+    h_basis_v[k]=basis[k].View(AcceleratorRead);
+    acceleratorPut(d_basis_v[k],h_basis_v[k]);
   }
-  vobj zz=Zero();
-  Vector<double> Qt_jv(Nm);
-  double * Qt_j = & Qt_jv[0];
-  for(int k=0;k<Nm;++k) Qt_j[k]=Qt(j,k);
 
-  auto basis_vp=& basis_v[0];
+  vobj zz=Zero();
+  deviceVector<double> Qt_jv(Nm);
+  double * Qt_j = & Qt_jv[0];
+  for(int k=0;k<Nm;++k) acceleratorPut(Qt_j[k],Qt(j,k));
+
+  auto basis_vp=& d_basis_v[0];
   autoView(result_v,result,AcceleratorWrite);
   accelerator_for(ss, grid->oSites(),vobj::Nsimd(),{
     vobj zzz=Zero();
@@ -171,7 +158,7 @@ void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,in
     }
     coalescedWrite(result_v[ss], B);
   });
-  for(int k=0;k<basis.size();k++) basis_v[k].ViewClose();
+  for(int k=0;k<basis.size();k++) h_basis_v[k].ViewClose();
 }
 
 template<class Field>
@@ -179,9 +166,9 @@ void basisReorderInPlace(std::vector<Field> &_v,std::vector<RealD>& sort_vals, s
 {
   int vlen = idx.size();
 
-  assert(vlen>=1);
-  assert(vlen<=sort_vals.size());
-  assert(vlen<=_v.size());
+  GRID_ASSERT(vlen>=1);
+  GRID_ASSERT(vlen<=sort_vals.size());
+  GRID_ASSERT(vlen<=_v.size());
 
   for (size_t i=0;i<vlen;i++) {
 
@@ -199,7 +186,7 @@ void basisReorderInPlace(std::vector<Field> &_v,std::vector<RealD>& sort_vals, s
 	if (idx[j]==i)
 	  break;
 
-      assert(idx[i] > i);     assert(j!=idx.size());      assert(idx[j]==i);
+      GRID_ASSERT(idx[i] > i);     GRID_ASSERT(j!=idx.size());      GRID_ASSERT(idx[j]==i);
 
       swap(_v[i],_v[idx[i]]); // should use vector move constructor, no data copy
       std::swap(sort_vals[i],sort_vals[idx[i]]);
@@ -237,7 +224,7 @@ void basisSortInPlace(std::vector<Field> & _v,std::vector<RealD>& sort_vals, boo
 template<class Field>
 void basisDeflate(const std::vector<Field> &_v,const std::vector<RealD>& eval,const Field& src_orig,Field& result) {
   result = Zero();
-  assert(_v.size()==eval.size());
+  GRID_ASSERT(_v.size()==eval.size());
   int N = (int)_v.size();
   for (int i=0;i<N;i++) {
     Field& tmp = _v[i];

Grid/lattice/Lattice_comparison.h

@@ -40,16 +40,16 @@ NAMESPACE_BEGIN(Grid);
 //Query supporting logical &&, ||, 
 //////////////////////////////////////////////////////////////////////////
 
-typedef iScalar<vInteger> vPredicate ;
+template<class vobj> using vPredicate = iScalar<IntegerPredicate<vobj> > ;
 
 //////////////////////////////////////////////////////////////////////////
 // compare lattice to lattice
 //////////////////////////////////////////////////////////////////////////
 
 template<class vfunctor,class lobj,class robj>  
-inline Lattice<vPredicate> LLComparison(vfunctor op,const Lattice<lobj> &lhs,const Lattice<robj> &rhs)
+inline Lattice<vPredicate<lobj> > LLComparison(vfunctor op,const Lattice<lobj> &lhs,const Lattice<robj> &rhs)
 {
-  Lattice<vPredicate> ret(rhs.Grid());
+  Lattice<vPredicate<lobj> > ret(rhs.Grid());
   autoView( lhs_v, lhs, CpuRead);
   autoView( rhs_v, rhs, CpuRead);
   autoView( ret_v, ret, CpuWrite);
@@ -62,9 +62,9 @@ inline Lattice<vPredicate> LLComparison(vfunctor op,const Lattice<lobj> &lhs,con
 // compare lattice to scalar
 //////////////////////////////////////////////////////////////////////////
 template<class vfunctor,class lobj,class robj> 
-inline Lattice<vPredicate> LSComparison(vfunctor op,const Lattice<lobj> &lhs,const robj &rhs)
+inline Lattice<vPredicate<lobj> > LSComparison(vfunctor op,const Lattice<lobj> &lhs,const robj &rhs)
 {
-  Lattice<vPredicate> ret(lhs.Grid());
+  Lattice<vPredicate<lobj> > ret(lhs.Grid());
   autoView( lhs_v, lhs, CpuRead);
   autoView( ret_v, ret, CpuWrite);
   thread_for( ss, lhs_v.size(), {
@@ -76,9 +76,9 @@ inline Lattice<vPredicate> LSComparison(vfunctor op,const Lattice<lobj> &lhs,con
 // compare scalar to lattice
 //////////////////////////////////////////////////////////////////////////
 template<class vfunctor,class lobj,class robj> 
-inline Lattice<vPredicate> SLComparison(vfunctor op,const lobj &lhs,const Lattice<robj> &rhs)
+inline Lattice<vPredicate<robj> > SLComparison(vfunctor op,const lobj &lhs,const Lattice<robj> &rhs)
 {
-  Lattice<vPredicate> ret(rhs.Grid());
+  Lattice<vPredicate<robj> > ret(rhs.Grid());
   autoView( rhs_v, rhs, CpuRead);
   autoView( ret_v, ret, CpuWrite);
   thread_for( ss, rhs_v.size(), {
@@ -92,87 +92,87 @@ inline Lattice<vPredicate> SLComparison(vfunctor op,const lobj &lhs,const Lattic
 //////////////////////////////////////////////////////////////////////////
 // Less than
 template<class lobj,class robj>
-inline Lattice<vPredicate> operator < (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
+inline Lattice<vPredicate<lobj> > operator < (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
   return LLComparison(vlt<lobj,robj>(),lhs,rhs);
 }
 template<class lobj,class robj>
-inline Lattice<vPredicate> operator < (const Lattice<lobj> & lhs, const robj & rhs) {
+inline Lattice<vPredicate<lobj> > operator < (const Lattice<lobj> & lhs, const robj & rhs) {
   return LSComparison(vlt<lobj,robj>(),lhs,rhs);
 }
 template<class lobj,class robj>
-inline Lattice<vPredicate> operator < (const lobj & lhs, const Lattice<robj> & rhs) {
+inline Lattice<vPredicate<robj> > operator < (const lobj & lhs, const Lattice<robj> & rhs) {
   return SLComparison(vlt<lobj,robj>(),lhs,rhs);
 }
   
 // Less than equal
 template<class lobj,class robj>
-inline Lattice<vPredicate> operator <= (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
+inline Lattice<vPredicate<lobj> > operator <= (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
   return LLComparison(vle<lobj,robj>(),lhs,rhs);
 }
 template<class lobj,class robj>
-inline Lattice<vPredicate> operator <= (const Lattice<lobj> & lhs, const robj & rhs) {
+inline Lattice<vPredicate<lobj> > operator <= (const Lattice<lobj> & lhs, const robj & rhs) {
   return LSComparison(vle<lobj,robj>(),lhs,rhs);
 }
 template<class lobj,class robj>
-inline Lattice<vPredicate> operator <= (const lobj & lhs, const Lattice<robj> & rhs) {
+inline Lattice<vPredicate<robj> > operator <= (const lobj & lhs, const Lattice<robj> & rhs) {
   return SLComparison(vle<lobj,robj>(),lhs,rhs);
 }
   
 // Greater than 
 template<class lobj,class robj>
-inline Lattice<vPredicate> operator > (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
+inline Lattice<vPredicate<lobj> > operator > (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
   return LLComparison(vgt<lobj,robj>(),lhs,rhs);
 }
 template<class lobj,class robj>
-inline Lattice<vPredicate> operator > (const Lattice<lobj> & lhs, const robj & rhs) {
+inline Lattice<vPredicate<lobj> > operator > (const Lattice<lobj> & lhs, const robj & rhs) {
   return LSComparison(vgt<lobj,robj>(),lhs,rhs);
 }
 template<class lobj,class robj>
-inline Lattice<vPredicate> operator > (const lobj & lhs, const Lattice<robj> & rhs) {
+inline Lattice<vPredicate<robj> > operator > (const lobj & lhs, const Lattice<robj> & rhs) {
   return SLComparison(vgt<lobj,robj>(),lhs,rhs);
 }
   
   
 // Greater than equal
 template<class lobj,class robj>
-inline Lattice<vPredicate> operator >= (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
+inline Lattice<vPredicate<lobj> > operator >= (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
   return LLComparison(vge<lobj,robj>(),lhs,rhs);
 }
 template<class lobj,class robj>
-inline Lattice<vPredicate> operator >= (const Lattice<lobj> & lhs, const robj & rhs) {
+inline Lattice<vPredicate<lobj> > operator >= (const Lattice<lobj> & lhs, const robj & rhs) {
   return LSComparison(vge<lobj,robj>(),lhs,rhs);
 }
 template<class lobj,class robj>
-inline Lattice<vPredicate> operator >= (const lobj & lhs, const Lattice<robj> & rhs) {
+inline Lattice<vPredicate<robj> > operator >= (const lobj & lhs, const Lattice<robj> & rhs) {
   return SLComparison(vge<lobj,robj>(),lhs,rhs);
 }
    
 // equal
 template<class lobj,class robj>
-inline Lattice<vPredicate> operator == (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
+inline Lattice<vPredicate<lobj> > operator == (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
   return LLComparison(veq<lobj,robj>(),lhs,rhs);
 }
 template<class lobj,class robj>
-inline Lattice<vPredicate> operator == (const Lattice<lobj> & lhs, const robj & rhs) {
+inline Lattice<vPredicate<lobj> > operator == (const Lattice<lobj> & lhs, const robj & rhs) {
   return LSComparison(veq<lobj,robj>(),lhs,rhs);
 }
 template<class lobj,class robj>
-inline Lattice<vPredicate> operator == (const lobj & lhs, const Lattice<robj> & rhs) {
+inline Lattice<vPredicate<robj> > operator == (const lobj & lhs, const Lattice<robj> & rhs) {
   return SLComparison(veq<lobj,robj>(),lhs,rhs);
 }
    
    
 // not equal
 template<class lobj,class robj>
-inline Lattice<vPredicate> operator != (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
+inline Lattice<vPredicate<lobj> > operator != (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
   return LLComparison(vne<lobj,robj>(),lhs,rhs);
 }
 template<class lobj,class robj>
-inline Lattice<vPredicate> operator != (const Lattice<lobj> & lhs, const robj & rhs) {
+inline Lattice<vPredicate<lobj> > operator != (const Lattice<lobj> & lhs, const robj & rhs) {
   return LSComparison(vne<lobj,robj>(),lhs,rhs);
 }
 template<class lobj,class robj>
-inline Lattice<vPredicate> operator != (const lobj & lhs, const Lattice<robj> & rhs) {
+inline Lattice<vPredicate<robj> > operator != (const lobj & lhs, const Lattice<robj> & rhs) {
   return SLComparison(vne<lobj,robj>(),lhs,rhs);
 }
 NAMESPACE_END(Grid);

Grid/lattice/Lattice_comparison_utils.h

@@ -39,47 +39,53 @@ NAMESPACE_BEGIN(Grid);
   //
   // Should guard this with isGridTensor<> enable if?
   /////////////////////////////////////////
+  //
+  // Predicate type matching the operand's vectorisation
+  //
+  template<class obj> using IntegerPredicate =
+    typename GridTypeMapper<typename GridTypeMapper<obj>::vector_type>::Integerified;
+
   //
   // Generic list of functors
   //
   template<class lobj,class robj> class veq {
   public:
-    accelerator vInteger operator()(const lobj &lhs, const robj &rhs)
+    accelerator IntegerPredicate<lobj> operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) == (rhs);
     }
   };
   template<class lobj,class robj> class vne {
   public:
-    accelerator vInteger operator()(const lobj &lhs, const robj &rhs)
+    accelerator IntegerPredicate<lobj> operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) != (rhs);
     }
   };
   template<class lobj,class robj> class vlt {
   public:
-    accelerator vInteger operator()(const lobj &lhs, const robj &rhs)
+    accelerator IntegerPredicate<lobj> operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) < (rhs);
     }
   };
   template<class lobj,class robj> class vle {
   public:
-    accelerator vInteger operator()(const lobj &lhs, const robj &rhs)
+    accelerator IntegerPredicate<lobj> operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) <= (rhs);
     }
   };
   template<class lobj,class robj> class vgt {
   public:
-    accelerator vInteger operator()(const lobj &lhs, const robj &rhs)
+    accelerator IntegerPredicate<lobj> operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) > (rhs);
     }
   };
   template<class lobj,class robj> class vge {
     public:
-    accelerator vInteger operator()(const lobj &lhs, const robj &rhs)
+    accelerator IntegerPredicate<lobj> operator()(const lobj &lhs, const robj &rhs)
     { 
       return (lhs) >= (rhs);
     }
@@ -133,78 +139,78 @@ NAMESPACE_BEGIN(Grid);
   // Integer and real get extra relational functions.
   //////////////////////////////////////////////////////////////////////////////////////////////////////
   template<class sfunctor, class vsimd,IfNotComplex<vsimd> = 0> 
-    accelerator_inline vInteger Comparison(sfunctor sop,const vsimd & lhs, const vsimd & rhs)
+    accelerator_inline IntegerPredicate<vsimd> Comparison(sfunctor sop,const vsimd & lhs, const vsimd & rhs)
     {
       typedef typename vsimd::scalar_type scalar;
       ExtractBuffer<scalar> vlhs(vsimd::Nsimd());   // Use functors to reduce this to single implementation
       ExtractBuffer<scalar> vrhs(vsimd::Nsimd());
       ExtractBuffer<Integer> vpred(vsimd::Nsimd());
-      vInteger ret;
+      IntegerPredicate<vsimd> ret;
       extract<vsimd,scalar>(lhs,vlhs);
       extract<vsimd,scalar>(rhs,vrhs);
       for(int s=0;s<vsimd::Nsimd();s++){
 	vpred[s] = sop(vlhs[s],vrhs[s]);
       }
-      merge<vInteger,Integer>(ret,vpred);
+      merge<IntegerPredicate<vsimd>,Integer>(ret,vpred);
       return ret;
     }
 
   template<class sfunctor, class vsimd,IfNotComplex<vsimd> = 0> 
-    accelerator_inline vInteger Comparison(sfunctor sop,const vsimd & lhs, const typename vsimd::scalar_type & rhs)
+    accelerator_inline IntegerPredicate<vsimd> Comparison(sfunctor sop,const vsimd & lhs, const typename vsimd::scalar_type & rhs)
     {
       typedef typename vsimd::scalar_type scalar;
       ExtractBuffer<scalar> vlhs(vsimd::Nsimd());   // Use functors to reduce this to single implementation
       ExtractBuffer<Integer> vpred(vsimd::Nsimd());
-      vInteger ret;
+      IntegerPredicate<vsimd> ret;
       extract<vsimd,scalar>(lhs,vlhs);
       for(int s=0;s<vsimd::Nsimd();s++){
 	vpred[s] = sop(vlhs[s],rhs);
       }
-      merge<vInteger,Integer>(ret,vpred);
+      merge<IntegerPredicate<vsimd>,Integer>(ret,vpred);
       return ret;
     }
 
   template<class sfunctor, class vsimd,IfNotComplex<vsimd> = 0> 
-    accelerator_inline vInteger Comparison(sfunctor sop,const typename vsimd::scalar_type & lhs, const vsimd & rhs)
+    accelerator_inline IntegerPredicate<vsimd> Comparison(sfunctor sop,const typename vsimd::scalar_type & lhs, const vsimd & rhs)
     {
       typedef typename vsimd::scalar_type scalar;
       ExtractBuffer<scalar> vrhs(vsimd::Nsimd());   // Use functors to reduce this to single implementation
       ExtractBuffer<Integer> vpred(vsimd::Nsimd());
-      vInteger ret;
+      IntegerPredicate<vsimd> ret;
       extract<vsimd,scalar>(rhs,vrhs);
       for(int s=0;s<vsimd::Nsimd();s++){
 	vpred[s] = sop(lhs,vrhs[s]);
       }
-      merge<vInteger,Integer>(ret,vpred);
+      merge<IntegerPredicate<vsimd>,Integer>(ret,vpred);
       return ret;
     }
 
 #define DECLARE_RELATIONAL_EQ(op,functor) \
   template<class vsimd,IfSimd<vsimd> = 0>\
-    accelerator_inline vInteger operator op (const vsimd & lhs, const vsimd & rhs)\
+    accelerator_inline IntegerPredicate<vsimd> operator op (const vsimd & lhs, const vsimd & rhs)\
     {\
       typedef typename vsimd::scalar_type scalar;\
       return Comparison(functor<scalar,scalar>(),lhs,rhs);\
     }\
   template<class vsimd,IfSimd<vsimd> = 0>\
-    accelerator_inline vInteger operator op (const vsimd & lhs, const typename vsimd::scalar_type & rhs) \
+    accelerator_inline IntegerPredicate<vsimd> operator op (const vsimd & lhs, const typename vsimd::scalar_type & rhs) \
     {\
       typedef typename vsimd::scalar_type scalar;\
       return Comparison(functor<scalar,scalar>(),lhs,rhs);\
     }\
   template<class vsimd,IfSimd<vsimd> = 0>\
-    accelerator_inline vInteger operator op (const typename vsimd::scalar_type & lhs, const vsimd & rhs) \
+    accelerator_inline IntegerPredicate<vsimd> operator op (const typename vsimd::scalar_type & lhs, const vsimd & rhs) \
     {\
       typedef typename vsimd::scalar_type scalar;\
       return Comparison(functor<scalar,scalar>(),lhs,rhs);\
     }\
   template<class vsimd>\
-    accelerator_inline vInteger operator op(const iScalar<vsimd> &lhs,const typename vsimd::scalar_type &rhs) \
+    accelerator_inline IntegerPredicate<vsimd> operator op(const iScalar<vsimd> &lhs,const typename vsimd::scalar_type &rhs) \
     {									\
       return lhs._internal op rhs;					\
     }									\
   template<class vsimd>\
-    accelerator_inline vInteger operator op(const typename vsimd::scalar_type &lhs,const iScalar<vsimd> &rhs) \
+    accelerator_inline IntegerPredicate<vsimd> operator op(const typename vsimd::scalar_type &lhs,const iScalar<vsimd> &rhs) \
     {									\
       return lhs op rhs._internal;					\
     }									\
@@ -212,7 +218,7 @@ NAMESPACE_BEGIN(Grid);
 #define DECLARE_RELATIONAL(op,functor) \
   DECLARE_RELATIONAL_EQ(op,functor)    \
   template<class vsimd>\
-    accelerator_inline vInteger operator op(const iScalar<vsimd> &lhs,const iScalar<vsimd> &rhs)\
+    accelerator_inline IntegerPredicate<vsimd> operator op(const iScalar<vsimd> &lhs,const iScalar<vsimd> &rhs)\
     {									\
       return lhs._internal op rhs._internal;				\
     }									

Grid/lattice/Lattice_conformable.h

@@ -32,8 +32,8 @@ NAMESPACE_BEGIN(Grid);
 
 template<class obj1,class obj2> void conformable(const Lattice<obj1> &lhs,const Lattice<obj2> &rhs)
 {
-  assert(lhs.Grid() == rhs.Grid());
-  assert(lhs.Checkerboard() == rhs.Checkerboard());
+  GRID_ASSERT(lhs.Grid() == rhs.Grid());
+  GRID_ASSERT(lhs.Checkerboard() == rhs.Checkerboard());
 }
 
 NAMESPACE_END(Grid);

Grid/lattice/Lattice_matrix_reduction.h

@@ -42,7 +42,7 @@ static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice
   //  Lattice<vobj> Xslice(SliceGrid);
   //  Lattice<vobj> Rslice(SliceGrid);
 
-  assert( FullGrid->_simd_layout[Orthog]==1);
+  GRID_ASSERT( FullGrid->_simd_layout[Orthog]==1);
 
   //FIXME package in a convenient iterator
   //Should loop over a plane orthogonal to direction "Orthog"
@@ -86,7 +86,7 @@ static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<
   int Nblock = X.Grid()->GlobalDimensions()[Orthog];
 
   GridBase *FullGrid  = X.Grid();
-  assert( FullGrid->_simd_layout[Orthog]==1);
+  GRID_ASSERT( FullGrid->_simd_layout[Orthog]==1);
 
   //FIXME package in a convenient iterator
   //Should loop over a plane orthogonal to direction "Orthog"
@@ -140,7 +140,7 @@ static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj>
   
   mat = Eigen::MatrixXcd::Zero(Nblock,Nblock);
 
-  assert( FullGrid->_simd_layout[Orthog]==1);
+  GRID_ASSERT( FullGrid->_simd_layout[Orthog]==1);
   //  int nh =  FullGrid->_ndimension;
   //  int nl = SliceGrid->_ndimension;
   //  int nl = nh-1;

Grid/lattice/Lattice_peekpoke.h

@@ -98,8 +98,8 @@ void pokeSite(const sobj &s,Lattice<vobj> &l,const Coordinate &site){
 
   int Nsimd = grid->Nsimd();
 
-  assert( l.Checkerboard()== l.Grid()->CheckerBoard(site));
-  assert( sizeof(sobj)*Nsimd == sizeof(vobj));
+  GRID_ASSERT( l.Checkerboard()== l.Grid()->CheckerBoard(site));
+  GRID_ASSERT( sizeof(sobj)*Nsimd == sizeof(vobj));
 
   int rank,odx,idx;
   // Optional to broadcast from node 0.
@@ -135,7 +135,7 @@ void peekSite(sobj &s,const Lattice<vobj> &l,const Coordinate &site){
 
   int Nsimd = grid->Nsimd();
 
-  assert( l.Checkerboard() == l.Grid()->CheckerBoard(site));
+  GRID_ASSERT( l.Checkerboard() == l.Grid()->CheckerBoard(site));
 
   int rank,odx,idx;
   grid->GlobalCoorToRankIndex(rank,odx,idx,site);
@@ -159,14 +159,14 @@ template<class vobj,class sobj>
 inline void peekLocalSite(sobj &s,const LatticeView<vobj> &l,Coordinate &site)
 {
   GridBase *grid = l.getGrid();
-  assert(l.mode==CpuRead);
+  GRID_ASSERT(l.mode==CpuRead);
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   int Nsimd = grid->Nsimd();
 
-  //  assert( l.Checkerboard()== grid->CheckerBoard(site));
-  assert( sizeof(sobj)*Nsimd == sizeof(vobj));
+  //  GRID_ASSERT( l.Checkerboard()== grid->CheckerBoard(site));
+  GRID_ASSERT( sizeof(sobj)*Nsimd == sizeof(vobj));
 
   static const int words=sizeof(vobj)/sizeof(vector_type);
   int odx,idx;
@@ -195,15 +195,15 @@ template<class vobj,class sobj>
 inline void pokeLocalSite(const sobj &s,LatticeView<vobj> &l,Coordinate &site)
 {
   GridBase *grid=l.getGrid();
-  assert(l.mode==CpuWrite);
+  GRID_ASSERT(l.mode==CpuWrite);
 
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   int Nsimd = grid->Nsimd();
 
-  //  assert( l.Checkerboard()== grid->CheckerBoard(site));
-  assert( sizeof(sobj)*Nsimd == sizeof(vobj));
+  //  GRID_ASSERT( l.Checkerboard()== grid->CheckerBoard(site));
+  GRID_ASSERT( sizeof(sobj)*Nsimd == sizeof(vobj));
 
   static const int words=sizeof(vobj)/sizeof(vector_type);
   int odx,idx;

Grid/lattice/Lattice_reduction.h

@@ -46,7 +46,7 @@ inline typename vobj::scalar_object sum_cpu(const vobj *arg, Integer osites)
   //  const int Nsimd = vobj::Nsimd();
   const int nthread = GridThread::GetThreads();
 
-  Vector<sobj> sumarray(nthread);
+  std::vector<sobj> sumarray(nthread);
   for(int i=0;i<nthread;i++){
     sumarray[i]=Zero();
   }
@@ -75,7 +75,7 @@ inline typename vobj::scalar_objectD sumD_cpu(const vobj *arg, Integer osites)
 
   const int nthread = GridThread::GetThreads();
 
-  Vector<sobj> sumarray(nthread);
+  std::vector<sobj> sumarray(nthread);
   for(int i=0;i<nthread;i++){
     sumarray[i]=Zero();
   }
@@ -290,23 +290,45 @@ template<class vobj>
 inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right) {
   GridBase *grid = left.Grid();
 
+  bool ok;
 #ifdef GRID_SYCL
-  uint64_t csum=0;
-  if ( FlightRecorder::LoggingMode != FlightRecorder::LoggingModeNone)
-  {
-    // Hack
-    // Fast integer xor checksum. Can also be used in comms now.
-    autoView(l_v,left,AcceleratorRead);
-    Integer words = left.Grid()->oSites()*sizeof(vobj)/sizeof(uint64_t);
-    uint64_t *base= (uint64_t *)&l_v[0];
-    csum=svm_xor(base,words);
-  }
-  FlightRecorder::CsumLog(csum);
+  //  uint64_t csum=0;
+  //  uint64_t csum2=0;
+  //  if ( FlightRecorder::LoggingMode != FlightRecorder::LoggingModeNone)
+  //  {
+  // Hack
+  // Fast integer xor checksum. Can also be used in comms now.
+  //    autoView(l_v,left,AcceleratorRead);
+  //    Integer words = left.Grid()->oSites()*sizeof(vobj)/sizeof(uint64_t);
+  //    uint64_t *base= (uint64_t *)&l_v[0];
+  //    csum=svm_xor(base,words);
+  //    ok = FlightRecorder::CsumLog(csum);
+  //    if ( !ok ) {
+  //      csum2=svm_xor(base,words);
+  //      std::cerr<< " Bad CSUM " << std::hex<< csum << " recomputed as "<<csum2<<std::dec<<std::endl;
+  //    } else {
+  //      csum2=svm_xor(base,words);
+  //      std::cerr<< " ok CSUM " << std::hex<< csum << " recomputed as "<<csum2<<std::dec<<std::endl;
+  //    }
+  //    GRID_ASSERT(ok);
+  // }
 #endif
+  FlightRecorder::StepLog("rank inner product");
   ComplexD nrm = rankInnerProduct(left,right);
+  //  ComplexD nrmck=nrm;
   RealD local = real(nrm);
-  FlightRecorder::NormLog(real(nrm)); 
-  grid->GlobalSum(nrm);
+  ok = FlightRecorder::NormLog(real(nrm));
+  if ( !ok ) {
+    ComplexD nrm2 = rankInnerProduct(left,right);
+    RealD local2 = real(nrm2);
+    std::cerr<< " Bad NORM " << local << " recomputed as "<<local2<<std::endl;
+    GRID_ASSERT(ok);
+  }
+  FlightRecorder::StepLog("Start global sum");
+  grid->GlobalSumP2P(nrm);
+  //  grid->GlobalSum(nrm);
+  FlightRecorder::StepLog("Finished global sum");
+  //  std::cout << " norm "<< nrm << " p2p norm "<<nrmck<<std::endl;
   FlightRecorder::ReductionLog(local,real(nrm)); 
   return nrm;
 }
@@ -343,18 +365,6 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
   autoView( x_v, x, AcceleratorRead);
   autoView( y_v, y, AcceleratorRead);
   autoView( z_v, z, AcceleratorWrite);
-#if 0
-  typedef decltype(innerProductD(x_v[0],y_v[0])) inner_t;
-  Vector<inner_t> inner_tmp(sites);
-  auto inner_tmp_v = &inner_tmp[0];
-
-  accelerator_for( ss, sites, nsimd,{
-      auto tmp = a*x_v(ss)+b*y_v(ss);
-      coalescedWrite(inner_tmp_v[ss],innerProductD(tmp,tmp));
-      coalescedWrite(z_v[ss],tmp);
-  });
-  nrm = real(TensorRemove(sum(inner_tmp_v,sites)));
-#else
   typedef decltype(innerProduct(x_v[0],y_v[0])) inner_t;
   deviceVector<inner_t> inner_tmp;
   inner_tmp.resize(sites);
@@ -365,9 +375,13 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
       coalescedWrite(inner_tmp_v[ss],innerProduct(tmp,tmp));
       coalescedWrite(z_v[ss],tmp);
   });
+  bool ok;
   nrm = real(TensorRemove(sumD(inner_tmp_v,sites)));
-#endif
+  ok = FlightRecorder::NormLog(real(nrm));
+  GRID_ASSERT(ok);
+  RealD local = real(nrm);
   grid->GlobalSum(nrm);
+  FlightRecorder::ReductionLog(local,real(nrm));
   return nrm; 
 }
  
@@ -377,7 +391,7 @@ innerProductNorm(ComplexD& ip, RealD &nrm, const Lattice<vobj> &left,const Latti
   conformable(left,right);
 
   typedef typename vobj::vector_typeD vector_type;
-  Vector<ComplexD> tmp(2);
+  std::vector<ComplexD> tmp(2);
 
   GridBase *grid = left.Grid();
 
@@ -387,8 +401,8 @@ innerProductNorm(ComplexD& ip, RealD &nrm, const Lattice<vobj> &left,const Latti
   // GPU
   typedef decltype(innerProductD(vobj(),vobj())) inner_t;
   typedef decltype(innerProductD(vobj(),vobj())) norm_t;
-  Vector<inner_t> inner_tmp(sites);
-  Vector<norm_t>  norm_tmp(sites);
+  deviceVector<inner_t> inner_tmp(sites);
+  deviceVector<norm_t>  norm_tmp(sites);
   auto inner_tmp_v = &inner_tmp[0];
   auto norm_tmp_v = &norm_tmp[0];
   {
@@ -438,7 +452,9 @@ inline auto sum(const LatticeTrinaryExpression<Op,T1,T2,T3> & expr)
 // sliceSum, sliceInnerProduct, sliceAxpy, sliceNorm etc...
 //////////////////////////////////////////////////////////////////////////////////////////////////////////////
 
-template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<typename vobj::scalar_object> &result,int orthogdim)
+template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,
+					  std::vector<typename vobj::scalar_object> &result,
+					  int orthogdim)
 {
   ///////////////////////////////////////////////////////
   // FIXME precision promoted summation
@@ -448,20 +464,20 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
   typedef typename vobj::scalar_object sobj;
   typedef typename vobj::scalar_object::scalar_type scalar_type;
   GridBase  *grid = Data.Grid();
-  assert(grid!=NULL);
+  GRID_ASSERT(grid!=NULL);
 
   const int    Nd = grid->_ndimension;
   const int Nsimd = grid->Nsimd();
 
-  assert(orthogdim >= 0);
-  assert(orthogdim < Nd);
+  GRID_ASSERT(orthogdim >= 0);
+  GRID_ASSERT(orthogdim < Nd);
 
   int fd=grid->_fdimensions[orthogdim];
   int ld=grid->_ldimensions[orthogdim];
   int rd=grid->_rdimensions[orthogdim];
 
-  Vector<vobj> lvSum(rd); // will locally sum vectors first
-  Vector<sobj> lsSum(ld,Zero());                    // sum across these down to scalars
+  std::vector<vobj> lvSum(rd); // will locally sum vectors first
+  std::vector<sobj> lsSum(ld,Zero());                    // sum across these down to scalars
   ExtractBuffer<sobj> extracted(Nsimd);                  // splitting the SIMD
 
   result.resize(fd); // And then global sum to return the same vector to every node 
@@ -509,6 +525,8 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
   scalar_type * ptr = (scalar_type *) &result[0];
   int words = fd*sizeof(sobj)/sizeof(scalar_type);
   grid->GlobalSumVector(ptr, words);
+  //  std::cout << GridLogMessage << " sliceSum local"<<t_sum<<" us, host+mpi "<<t_rest<<std::endl;
+  
 }
 template<class vobj> inline
 std::vector<typename vobj::scalar_object> 
@@ -539,21 +557,21 @@ static void sliceInnerProductVector( std::vector<ComplexD> & result, const Latti
   typedef typename vobj::vector_type   vector_type;
   typedef typename vobj::scalar_type   scalar_type;
   GridBase  *grid = lhs.Grid();
-  assert(grid!=NULL);
+  GRID_ASSERT(grid!=NULL);
   conformable(grid,rhs.Grid());
 
   const int    Nd = grid->_ndimension;
   const int Nsimd = grid->Nsimd();
 
-  assert(orthogdim >= 0);
-  assert(orthogdim < Nd);
+  GRID_ASSERT(orthogdim >= 0);
+  GRID_ASSERT(orthogdim < Nd);
 
   int fd=grid->_fdimensions[orthogdim];
   int ld=grid->_ldimensions[orthogdim];
   int rd=grid->_rdimensions[orthogdim];
 
-  Vector<vector_type> lvSum(rd); // will locally sum vectors first
-  Vector<scalar_type > lsSum(ld,scalar_type(0.0));                    // sum across these down to scalars
+  std::vector<vector_type> lvSum(rd); // will locally sum vectors first
+  std::vector<scalar_type > lsSum(ld,scalar_type(0.0));                    // sum across these down to scalars
   ExtractBuffer<iScalar<scalar_type> > extracted(Nsimd);   // splitting the SIMD  
 
   result.resize(fd); // And then global sum to return the same vector to every node for IO to file

Grid/lattice/Lattice_reduction_gpu.h

@@ -197,65 +197,205 @@ __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;
 
   Integer numThreads, numBlocks;
   int ok = getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
-  assert(ok);
+  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
-#undef UVM_BLOCK_BUFFER  
-#ifndef UVM_BLOCK_BUFFER  
-  commVector<sobj> buffer(numBlocks);
+  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));
-#else
-  Vector<sobj> buffer(numBlocks);
-  sobj *buffer_v = &buffer[0];
-  sobj result;
-  reduceKernel<<< numBlocks, numThreads, smemSize, computeStream >>>(lat, buffer_v, size);
-  accelerator_barrier();
-  result = *buffer_v;
+#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);
 
-  Vector<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;
 }
 
@@ -298,5 +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

@@ -4,35 +4,29 @@ NAMESPACE_BEGIN(Grid);
 // Possibly promote to double and sum
 /////////////////////////////////////////////////////////////////////////////////////////////////////////
 
+
 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;
-  static Vector<sobj> mysum;
-  mysum.resize(1);
-  sobj *mysum_p = & mysum[0];
-  sobj identity; zeroit(identity);
-  mysum[0] = identity;
-  sobj ret ; 
 
-  Integer nsimd= vobj::Nsimd();
-
-  const cl::sycl::property_list PropList ({ cl::sycl::property::reduction::initialize_to_identity() });
-  theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
-    auto Reduction = cl::sycl::reduction(mysum_p,identity,std::plus<>(),PropList);
-     cgh.parallel_for(cl::sycl::range<1>{osites},
-		      Reduction,
-		      [=] (cl::sycl::id<1> item, auto &sum) {
-      auto osite   = item[0];
-      sum +=Reduce(lat[osite]);
-     });
-   });
-  theGridAccelerator->wait();
-  ret = mysum[0];
-  //  free(mysum,*theGridAccelerator);
-  sobjD dret; convertType(dret,ret);
-  return dret;
+  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>{(size_t)osites},
+                       Reduction,
+                       [=](sycl::id<1> item, auto &sum) {
+                         sobj s = Reduce(lat[item[0]]);
+                         sobjD sd; sd = s;
+                         sum += sd;
+                       });
+    });
+  }
+  return ret;
 }
 
 template <class vobj>
@@ -76,59 +70,41 @@ inline typename vobj::scalar_object sum_gpu_large(const vobj *lat, Integer osite
 
 template<class Word> Word svm_xor(Word *vec,uint64_t L)
 {
-  Word xorResult; xorResult = 0;
-  static Vector<Word> d_sum;
-  d_sum.resize(1);
-  Word *d_sum_p=&d_sum[0];
   Word identity;  identity=0;
-  d_sum[0] = identity;
-  const cl::sycl::property_list PropList ({ cl::sycl::property::reduction::initialize_to_identity() });
-  theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
-    auto Reduction = cl::sycl::reduction(d_sum_p,identity,std::bit_xor<>(),PropList);
-     cgh.parallel_for(cl::sycl::range<1>{L},
-		      Reduction,
-		      [=] (cl::sycl::id<1> index, auto &sum) {
-	 sum^=vec[index];
-     });
-   });
+  Word ret = 0;
+  { 
+    sycl::buffer<Word, 1> abuff(&ret, {1});
+    theGridAccelerator->submit([&](sycl::handler &cgh) {
+      auto Reduction = sycl::reduction(abuff,cgh,identity,std::bit_xor<>());
+      cgh.parallel_for(sycl::range<1>{L},
+                      Reduction,
+                      [=] (sycl::id<1> index, auto &sum) {
+                        sum ^=vec[index];
+                      });
+    });
+  }
+  theGridAccelerator->wait();
+  return ret;
+}
+template<class Word> Word checksum_gpu(Word *vec,uint64_t L)
+{
+  Word identity;  identity=0;
+  Word ret = 0;
+  { 
+    sycl::buffer<Word, 1> abuff(&ret, {1});
+    theGridAccelerator->submit([&](sycl::handler &cgh) {
+      auto Reduction = sycl::reduction(abuff,cgh,identity,std::bit_xor<>());
+      cgh.parallel_for(sycl::range<1>{L},
+                       Reduction,
+                       [=] (sycl::id<1> index, auto &sum) {
+			 auto l = index % 61;
+                         sum ^= vec[index]<<l | vec[index]>>(64-l);
+                       });
+    });
+  }
   theGridAccelerator->wait();
-  Word ret = d_sum[0];
-  //  free(d_sum,*theGridAccelerator);
   return ret;
 }
 
 NAMESPACE_END(Grid);
 
-/*
-
-template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu_repack(const vobj *lat, Integer osites)
-{
-  typedef typename vobj::vector_type  vector;
-  typedef typename vobj::scalar_type  scalar;
-
-  typedef typename vobj::scalar_typeD scalarD;
-  typedef typename vobj::scalar_objectD sobjD;
-
-  sobjD ret;
-  scalarD *ret_p = (scalarD *)&ret;
-  
-  const int nsimd = vobj::Nsimd();
-  const int words = sizeof(vobj)/sizeof(vector);
-
-  Vector<scalar> buffer(osites*nsimd);
-  scalar *buf = &buffer[0];
-  vector *dat = (vector *)lat;
-
-  for(int w=0;w<words;w++) {
-
-    accelerator_for(ss,osites,nsimd,{
-	int lane = acceleratorSIMTlane(nsimd);
-	buf[ss*nsimd+lane] = dat[ss*words+w].getlane(lane);
-    });
-    //Precision change at this point is to late to gain precision
-    ret_p[w] = svm_reduce(buf,nsimd*osites);
-  }
-  return ret;
-}
-*/

Grid/lattice/Lattice_rng.h

@@ -53,10 +53,10 @@ inline int RNGfillable(GridBase *coarse,GridBase *fine)
 
   // trivially extended in higher dims, with locality guaranteeing RNG state is local to node
   int lowerdims   = fine->_ndimension - coarse->_ndimension;
-  assert(lowerdims >= 0);
+  GRID_ASSERT(lowerdims >= 0);
   for(int d=0;d<lowerdims;d++){
-    assert(fine->_simd_layout[d]==1);
-    assert(fine->_processors[d]==1);
+    GRID_ASSERT(fine->_simd_layout[d]==1);
+    GRID_ASSERT(fine->_processors[d]==1);
   }
 
   int multiplicity=1;
@@ -66,9 +66,9 @@ inline int RNGfillable(GridBase *coarse,GridBase *fine)
   // local and global volumes subdivide cleanly after SIMDization
   for(int d=0;d<rngdims;d++){
     int fd= d+lowerdims;
-    assert(coarse->_processors[d]  == fine->_processors[fd]);
-    assert(coarse->_simd_layout[d] == fine->_simd_layout[fd]);
-    assert(((fine->_rdimensions[fd] / coarse->_rdimensions[d])* coarse->_rdimensions[d])==fine->_rdimensions[fd]); 
+    GRID_ASSERT(coarse->_processors[d]  == fine->_processors[fd]);
+    GRID_ASSERT(coarse->_simd_layout[d] == fine->_simd_layout[fd]);
+    GRID_ASSERT(((fine->_rdimensions[fd] / coarse->_rdimensions[d])* coarse->_rdimensions[d])==fine->_rdimensions[fd]); 
 
     multiplicity = multiplicity *fine->_rdimensions[fd] / coarse->_rdimensions[d]; 
   }
@@ -83,18 +83,18 @@ inline int RNGfillable_general(GridBase *coarse,GridBase *fine)
   int rngdims = coarse->_ndimension;
     
   // trivially extended in higher dims, with locality guaranteeing RNG state is local to node
-  int lowerdims   = fine->_ndimension - coarse->_ndimension;  assert(lowerdims >= 0);
+  int lowerdims   = fine->_ndimension - coarse->_ndimension;  GRID_ASSERT(lowerdims >= 0);
   // assumes that the higher dimensions are not using more processors
   // all further divisions are local
-  for(int d=0;d<lowerdims;d++) assert(fine->_processors[d]==1);
-  for(int d=0;d<rngdims;d++) assert(coarse->_processors[d] == fine->_processors[d+lowerdims]);
+  for(int d=0;d<lowerdims;d++) GRID_ASSERT(fine->_processors[d]==1);
+  for(int d=0;d<rngdims;d++) GRID_ASSERT(coarse->_processors[d] == fine->_processors[d+lowerdims]);
 
   // then divide the number of local sites
   // check that the total number of sims agree, meanse the iSites are the same
-  assert(fine->Nsimd() == coarse->Nsimd());
+  GRID_ASSERT(fine->Nsimd() == coarse->Nsimd());
 
   // check that the two grids divide cleanly
-  assert( (fine->lSites() / coarse->lSites() ) * coarse->lSites() == fine->lSites() );
+  GRID_ASSERT( (fine->lSites() / coarse->lSites() ) * coarse->lSites() == fine->lSites() );
 
   return fine->lSites() / coarse->lSites();
 }
@@ -177,7 +177,7 @@ public:
 
     skip = skip<<shift;
 
-    assert((skip >> shift)==site); // check for overflow
+    GRID_ASSERT((skip >> shift)==site); // check for overflow
 
     eng.discard(skip);
 #else
@@ -218,7 +218,7 @@ public:
     GetState(saved,_generators[gen]);
   }
   void SetState(std::vector<RngStateType> & saved,RngEngine &eng){
-    assert(saved.size()==RngStateCount);
+    GRID_ASSERT(saved.size()==RngStateCount);
     std::stringstream ss;
     for(int i=0;i<RngStateCount;i++){
       ss<< saved[i]<<" ";

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
@@ -21,9 +20,18 @@ NAMESPACE_BEGIN(Grid);
 
 
 #if defined(GRID_CUDA) || defined(GRID_HIP)
-template<class vobj> inline void sliceSumReduction_cub_small(const vobj *Data, Vector<vobj> &lvSum, const int rd, const int e1, const int e2, const int stride, const int ostride, const int Nsimd) {
+template<class vobj>
+inline void sliceSumReduction_cub_small(const vobj *Data,
+					std::vector<vobj> &lvSum,
+					const int rd,
+					const int e1,
+					const int e2,
+					const int stride,
+					const int ostride,
+					const int Nsimd)
+{
   size_t subvol_size = e1*e2;
-  commVector<vobj> reduction_buffer(rd*subvol_size);
+  deviceVector<vobj> reduction_buffer(rd*subvol_size);
   auto rb_p = &reduction_buffer[0];
   vobj zero_init;
   zeroit(zero_init);
@@ -46,10 +54,15 @@ template<class vobj> inline void sliceSumReduction_cub_small(const vobj *Data, V
   d_offsets = static_cast<int*>(acceleratorAllocDevice((rd+1)*sizeof(int)));
   
   //copy offsets to device
-  acceleratorCopyToDeviceAsync(&offsets[0],d_offsets,sizeof(int)*(rd+1),computeStream);
+  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);
@@ -73,13 +86,15 @@ template<class vobj> inline void sliceSumReduction_cub_small(const vobj *Data, V
   });
   
   //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
   
-  acceleratorCopyFromDeviceAsync(d_out,&lvSum[0],rd*sizeof(vobj),computeStream);
+  acceleratorCopyFromDeviceAsynch(d_out,&lvSum[0],rd*sizeof(vobj),computeStream);
   
   //sync after copy
   accelerator_barrier();
@@ -94,7 +109,15 @@ template<class vobj> inline void sliceSumReduction_cub_small(const vobj *Data, V
 
 
 #if defined(GRID_SYCL)
-template<class vobj> inline void sliceSumReduction_sycl_small(const vobj *Data, Vector <vobj> &lvSum, const int  &rd, const int &e1, const int &e2, const int &stride, const int &ostride, const int &Nsimd)
+template<class vobj>
+inline void sliceSumReduction_sycl_small(const vobj *Data,
+					 std::vector <vobj> &lvSum,
+					 const int  &rd,
+					 const int &e1,
+					 const int &e2,
+					 const int &stride,
+					 const int &ostride,
+					 const int &Nsimd)
 {
   size_t subvol_size = e1*e2;
 
@@ -105,7 +128,7 @@ template<class vobj> inline void sliceSumReduction_sycl_small(const vobj *Data,
     mysum[r] = vobj_zero; 
   }
 
-  commVector<vobj> reduction_buffer(rd*subvol_size);    
+  deviceVector<vobj> reduction_buffer(rd*subvol_size);    
 
   auto rb_p = &reduction_buffer[0];
 
@@ -124,11 +147,11 @@ template<class vobj> inline void sliceSumReduction_sycl_small(const vobj *Data,
   });
 
   for (int r = 0; r < rd; r++) {
-      theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
-          auto Reduction = cl::sycl::reduction(&mysum[r],std::plus<>());
-          cgh.parallel_for(cl::sycl::range<1>{subvol_size},
+      theGridAccelerator->submit([&](sycl::handler &cgh) {
+          auto Reduction = sycl::reduction(&mysum[r],std::plus<>());
+          cgh.parallel_for(sycl::range<1>{subvol_size},
           Reduction,
-          [=](cl::sycl::id<1> item, auto &sum) {
+          [=](sycl::id<1> item, auto &sum) {
               auto s = item[0];
               sum += rb_p[r*subvol_size+s];
           });
@@ -144,14 +167,23 @@ template<class vobj> inline void sliceSumReduction_sycl_small(const vobj *Data,
 }
 #endif
 
-template<class vobj> inline void sliceSumReduction_large(const vobj *Data, Vector<vobj> &lvSum, const int rd, const int e1, const int e2, const int stride, const int ostride, const int Nsimd) {
+template<class vobj>
+inline void sliceSumReduction_large(const vobj *Data,
+				    std::vector<vobj> &lvSum,
+				    const int rd,
+				    const int e1,
+				    const int e2,
+				    const int stride,
+				    const int ostride,
+				    const int Nsimd)
+{
   typedef typename vobj::vector_type vector;
   const int words = sizeof(vobj)/sizeof(vector);
   const int osites = rd*e1*e2;
-  commVector<vector>buffer(osites);
+  deviceVector<vector>buffer(osites);
   vector *dat = (vector *)Data;
   vector *buf = &buffer[0];
-  Vector<vector> lvSum_small(rd);
+  std::vector<vector> lvSum_small(rd);
   vector *lvSum_ptr = (vector *)&lvSum[0];
 
   for (int w = 0; w < words; w++) {
@@ -168,13 +200,18 @@ template<class vobj> inline void sliceSumReduction_large(const vobj *Data, Vecto
     for (int r = 0; r < rd; r++) {
       lvSum_ptr[w+words*r]=lvSum_small[r];
     }
-
   }
-
-  
 }
 
-template<class vobj> inline void sliceSumReduction_gpu(const Lattice<vobj> &Data, Vector<vobj> &lvSum, const int rd, const int e1, const int e2, const int stride, const int ostride, const int Nsimd)
+template<class vobj>
+inline void sliceSumReduction_gpu(const Lattice<vobj> &Data,
+				  std::vector<vobj> &lvSum,
+				  const int rd,
+				  const int e1,
+				  const int e2,
+				  const int stride,
+				  const int ostride,
+				  const int Nsimd)
 {
   autoView(Data_v, Data, AcceleratorRead); //reduction libraries cannot deal with large vobjs so we split into small/large case.
     if constexpr (sizeof(vobj) <= 256) { 
@@ -192,7 +229,15 @@ template<class vobj> inline void sliceSumReduction_gpu(const Lattice<vobj> &Data
 }
 
 
-template<class vobj> inline void sliceSumReduction_cpu(const Lattice<vobj> &Data, Vector<vobj> &lvSum, const int &rd, const int &e1, const int &e2, const int &stride, const int &ostride, const int &Nsimd)
+template<class vobj>
+inline void sliceSumReduction_cpu(const Lattice<vobj> &Data,
+				  std::vector<vobj> &lvSum,
+				  const int &rd,
+				  const int &e1,
+				  const int &e2,
+				  const int &stride,
+				  const int &ostride,
+				  const int &Nsimd)
 {
   // sum over reduced dimension planes, breaking out orthog dir
   // Parallel over orthog direction
@@ -208,16 +253,20 @@ template<class vobj> inline void sliceSumReduction_cpu(const Lattice<vobj> &Data
   });
 }
 
-template<class vobj> inline void sliceSumReduction(const Lattice<vobj> &Data, Vector<vobj> &lvSum, const int &rd, const int &e1, const int &e2, const int &stride, const int &ostride, const int &Nsimd) 
+template<class vobj> inline void sliceSumReduction(const Lattice<vobj> &Data,
+						   std::vector<vobj> &lvSum,
+						   const int &rd,
+						   const int &e1,
+						   const int &e2,
+						   const int &stride,
+						   const int &ostride,
+						   const int &Nsimd) 
 {
-  #if defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)
-  
+#if defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)
   sliceSumReduction_gpu(Data, lvSum, rd, e1, e2, stride, ostride, Nsimd);
-  
-  #else
+#else
   sliceSumReduction_cpu(Data, lvSum, rd, e1, e2, stride, ostride, Nsimd);
-
-  #endif
+#endif
 }
 
 

Grid/lattice/Lattice_trace.h

@@ -124,6 +124,68 @@ Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > Inverse(const Lattice<iScala
   return ret;
 }
 
+template<int N>
+Lattice<iMatrix<iScalar<iScalar<iScalar<vComplexD> > > , N> > Inverse(const Lattice<iMatrix<iScalar<iScalar<iScalar<vComplexD> > >, N> > &Umu)
+{
+  GridBase *grid=Umu.Grid();
+  auto lvol = grid->lSites();
+  Lattice<iMatrix<iScalar<iScalar<iScalar<vComplexD> > >, N > > ret(grid);
+  
+  autoView(Umu_v,Umu,CpuRead);
+  autoView(ret_v,ret,CpuWrite);
+  thread_for(site,lvol,{
+    Eigen::MatrixXcd EigenU = Eigen::MatrixXcd::Zero(N,N);
+    Coordinate lcoor;
+    grid->LocalIndexToLocalCoor(site, lcoor);
+    iMatrix<iScalar<iScalar<iScalar<ComplexD> > >, N > Us;
+    iMatrix<iScalar<iScalar<iScalar<ComplexD> > >, N > Ui;
+    peekLocalSite(Us, Umu_v, lcoor);
+    for(int i=0;i<N;i++){
+      for(int j=0;j<N;j++){
+	EigenU(i,j) = Us(i,j)()()();
+      }}
+    Eigen::MatrixXcd EigenUinv = EigenU.inverse();
+    for(int i=0;i<N;i++){
+      for(int j=0;j<N;j++){
+	Ui(i,j)()()() = EigenUinv(i,j);
+      }}
+    pokeLocalSite(Ui,ret_v,lcoor);
+  });
+  return ret;
+}
+
+template<int N>
+Lattice<iMatrix<iScalar<iScalar<iScalar<iScalar<vComplexD> > > > , N> > Inverse(const Lattice<iMatrix<iScalar<iScalar<iScalar<iScalar<vComplexD> > > >, N> > &Umu)
+{
+  GridBase *grid=Umu.Grid();
+  auto lvol = grid->lSites();
+  Lattice<iMatrix<iScalar<iScalar<iScalar<iScalar<vComplexD> > > >, N > > ret(grid);
+  
+  autoView(Umu_v,Umu,CpuRead);
+  autoView(ret_v,ret,CpuWrite);
+  thread_for(site,lvol,{
+    Eigen::MatrixXcd EigenU = Eigen::MatrixXcd::Zero(N,N);
+    Coordinate lcoor;
+    grid->LocalIndexToLocalCoor(site, lcoor);
+    iMatrix<iScalar<iScalar<iScalar<iScalar<ComplexD> > > >, N > Us;
+    iMatrix<iScalar<iScalar<iScalar<iScalar<ComplexD> > > >, N > Ui;
+    peekLocalSite(Us, Umu_v, lcoor);
+    for(int i=0;i<N;i++){
+      for(int j=0;j<N;j++){
+	EigenU(i,j) = Us(i,j)()()()();
+      }}
+    Eigen::MatrixXcd EigenUinv = EigenU.inverse();
+    for(int i=0;i<N;i++){
+      for(int j=0;j<N;j++){
+	Ui(i,j)()()()() = EigenUinv(i,j);
+      }}
+    pokeLocalSite(Ui,ret_v,lcoor);
+  });
+  return ret;
+}
+
+
+
 
 NAMESPACE_END(Grid);
 #endif

Grid/lattice/Lattice_transfer.h

@@ -31,15 +31,15 @@ NAMESPACE_BEGIN(Grid);
 
 inline void subdivides(GridBase *coarse,GridBase *fine)
 {
-  assert(coarse->_ndimension == fine->_ndimension);
+  GRID_ASSERT(coarse->_ndimension == fine->_ndimension);
 
   int _ndimension = coarse->_ndimension;
 
   // local and global volumes subdivide cleanly after SIMDization
   for(int d=0;d<_ndimension;d++){
-    assert(coarse->_processors[d]  == fine->_processors[d]);
-    assert(coarse->_simd_layout[d] == fine->_simd_layout[d]);
-    assert((fine->_rdimensions[d] / coarse->_rdimensions[d])* coarse->_rdimensions[d]==fine->_rdimensions[d]); 
+    GRID_ASSERT(coarse->_processors[d]  == fine->_processors[d]);
+    GRID_ASSERT(coarse->_simd_layout[d] == fine->_simd_layout[d]);
+    GRID_ASSERT((fine->_rdimensions[d] / coarse->_rdimensions[d])* coarse->_rdimensions[d]==fine->_rdimensions[d]); 
   }
 }
 
@@ -309,7 +309,7 @@ inline void batchBlockProject(std::vector<Lattice<iVector<CComplex,nbasis>>> &co
                                const VLattice &Basis)
 {
   int NBatch = fineData.size();
-  assert(coarseData.size() == NBatch);
+  GRID_ASSERT(coarseData.size() == NBatch);
 
   GridBase * fine  = fineData[0].Grid();
   GridBase * coarse= coarseData[0].Grid();
@@ -344,7 +344,7 @@ template<class vobj,class vobj2,class CComplex>
   GridBase * coarse= coarseA.Grid();
 
   fineZ.Checkerboard()=fineX.Checkerboard();
-  assert(fineX.Checkerboard()==fineY.Checkerboard());
+  GRID_ASSERT(fineX.Checkerboard()==fineY.Checkerboard());
   subdivides(coarse,fine); // require they map
   conformable(fineX,fineY);
   conformable(fineX,fineZ);
@@ -356,7 +356,7 @@ template<class vobj,class vobj2,class CComplex>
   // FIXME merge with subdivide checking routine as this is redundant
   for(int d=0 ; d<_ndimension;d++){
     block_r[d] = fine->_rdimensions[d] / coarse->_rdimensions[d];
-    assert(block_r[d]*coarse->_rdimensions[d]==fine->_rdimensions[d]);
+    GRID_ASSERT(block_r[d]*coarse->_rdimensions[d]==fine->_rdimensions[d]);
   }
 
   autoView( fineZ_  , fineZ, AcceleratorWrite);
@@ -613,7 +613,7 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
   int  _ndimension = coarse->_ndimension;
 
   // checks
-  assert( nbasis == Basis.size() );
+  GRID_ASSERT( nbasis == Basis.size() );
   subdivides(coarse,fine); 
   for(int i=0;i<nbasis;i++){
     conformable(Basis[i].Grid(),fine);
@@ -687,7 +687,7 @@ inline void batchBlockPromote(const std::vector<Lattice<iVector<CComplex,nbasis>
                                const VLattice &Basis)
 {
   int NBatch = coarseData.size();
-  assert(fineData.size() == NBatch);
+  GRID_ASSERT(fineData.size() == NBatch);
 
   GridBase * fine   = fineData[0].Grid();
   GridBase * coarse = coarseData[0].Grid();
@@ -715,12 +715,12 @@ void localConvert(const Lattice<vobj> &in,Lattice<vvobj> &out)
   int ni = ig->_ndimension;
   int no = og->_ndimension;
 
-  assert(ni == no);
+  GRID_ASSERT(ni == no);
 
   for(int d=0;d<no;d++){
-    assert(ig->_processors[d]  == og->_processors[d]);
-    assert(ig->_ldimensions[d] == og->_ldimensions[d]);
-    assert(ig->lSites() == og->lSites());
+    GRID_ASSERT(ig->_processors[d]  == og->_processors[d]);
+    GRID_ASSERT(ig->_ldimensions[d] == og->_ldimensions[d]);
+    GRID_ASSERT(ig->lSites() == og->lSites());
   }
 
   autoView(in_v,in,CpuRead);
@@ -752,16 +752,16 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
 
   GridBase *Fg = From.Grid();
   GridBase *Tg = To.Grid();
-  assert(!Fg->_isCheckerBoarded);
-  assert(!Tg->_isCheckerBoarded);
+  GRID_ASSERT(!Fg->_isCheckerBoarded);
+  GRID_ASSERT(!Tg->_isCheckerBoarded);
   int Nsimd = Fg->Nsimd();
   int nF = Fg->_ndimension;
   int nT = Tg->_ndimension;
   int nd = nF;
-  assert(nF == nT);
+  GRID_ASSERT(nF == nT);
 
   for(int d=0;d<nd;d++){
-    assert(Fg->_processors[d]  == Tg->_processors[d]);
+    GRID_ASSERT(Fg->_processors[d]  == Tg->_processors[d]);
   }
 
   ///////////////////////////////////////////////////////////
@@ -821,12 +821,12 @@ void InsertSliceFast(const Lattice<vobj> &From,Lattice<vobj> & To,int slice, int
   //////////////////////////////////////////////////////////////////////////////////////////
   GridBase *Fg = From.Grid();
   GridBase *Tg = To.Grid();
-  assert(!Fg->_isCheckerBoarded);
-  assert(!Tg->_isCheckerBoarded);
+  GRID_ASSERT(!Fg->_isCheckerBoarded);
+  GRID_ASSERT(!Tg->_isCheckerBoarded);
   int Nsimd = Fg->Nsimd();
   int nF = Fg->_ndimension;
   int nT = Tg->_ndimension;
-  assert(nF+1 == nT);
+  GRID_ASSERT(nF+1 == nT);
 
   ///////////////////////////////////////////////////////////
   // do the index calc on the GPU
@@ -890,12 +890,12 @@ void ExtractSliceFast(Lattice<vobj> &To,const Lattice<vobj> & From,int slice, in
   //////////////////////////////////////////////////////////////////////////////////////////
   GridBase *Fg = From.Grid();
   GridBase *Tg = To.Grid();
-  assert(!Fg->_isCheckerBoarded);
-  assert(!Tg->_isCheckerBoarded);
+  GRID_ASSERT(!Fg->_isCheckerBoarded);
+  GRID_ASSERT(!Tg->_isCheckerBoarded);
   int Nsimd = Fg->Nsimd();
   int nF = Fg->_ndimension;
   int nT = Tg->_ndimension;
-  assert(nT+1 == nF);
+  GRID_ASSERT(nT+1 == nF);
 
   ///////////////////////////////////////////////////////////
   // do the index calc on the GPU
@@ -955,16 +955,16 @@ void InsertSlice(const Lattice<vobj> &lowDim,Lattice<vobj> & higherDim,int slice
   int nl = lg->_ndimension;
   int nh = hg->_ndimension;
 
-  assert(nl+1 == nh);
-  assert(orthog<nh);
-  assert(orthog>=0);
-  assert(hg->_processors[orthog]==1);
+  GRID_ASSERT(nl+1 == nh);
+  GRID_ASSERT(orthog<nh);
+  GRID_ASSERT(orthog>=0);
+  GRID_ASSERT(hg->_processors[orthog]==1);
 
   int dl; dl = 0;
   for(int d=0;d<nh;d++){
     if ( d != orthog) {
-      assert(lg->_processors[dl]  == hg->_processors[d]);
-      assert(lg->_ldimensions[dl] == hg->_ldimensions[d]);
+      GRID_ASSERT(lg->_processors[dl]  == hg->_processors[d]);
+      GRID_ASSERT(lg->_ldimensions[dl] == hg->_ldimensions[d]);
       dl++;
     }
   }
@@ -1005,17 +1005,17 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
   int nl = lg->_ndimension;
   int nh = hg->_ndimension;
 
-  assert(nl+1 == nh);
-  assert(orthog<nh);
-  assert(orthog>=0);
-  assert(hg->_processors[orthog]==1);
+  GRID_ASSERT(nl+1 == nh);
+  GRID_ASSERT(orthog<nh);
+  GRID_ASSERT(orthog>=0);
+  GRID_ASSERT(hg->_processors[orthog]==1);
   lowDim.Checkerboard() = higherDim.Checkerboard();
 
   int dl; dl = 0;
   for(int d=0;d<nh;d++){
     if ( d != orthog) {
-      assert(lg->_processors[dl]  == hg->_processors[d]);
-      assert(lg->_ldimensions[dl] == hg->_ldimensions[d]);
+      GRID_ASSERT(lg->_processors[dl]  == hg->_processors[d]);
+      GRID_ASSERT(lg->_ldimensions[dl] == hg->_ldimensions[d]);
       dl++;
     }
   }
@@ -1056,14 +1056,14 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
   int nl = lg->_ndimension;
   int nh = hg->_ndimension;
 
-  assert(nl == nh);
-  assert(orthog<nh);
-  assert(orthog>=0);
+  GRID_ASSERT(nl == nh);
+  GRID_ASSERT(orthog<nh);
+  GRID_ASSERT(orthog>=0);
 
   for(int d=0;d<nh;d++){
     if ( d!=orthog ) {
-      assert(lg->_processors[d]  == hg->_processors[d]);
-      assert(lg->_ldimensions[d] == hg->_ldimensions[d]);
+      GRID_ASSERT(lg->_processors[d]  == hg->_processors[d]);
+      GRID_ASSERT(lg->_ldimensions[d] == hg->_ldimensions[d]);
     }
   }
   Coordinate sz = lg->_ldimensions;
@@ -1093,7 +1093,7 @@ void Replicate(const Lattice<vobj> &coarse,Lattice<vobj> & fine)
 
   subdivides(cg,fg); 
 
-  assert(cg->_ndimension==fg->_ndimension);
+  GRID_ASSERT(cg->_ndimension==fg->_ndimension);
 
   Coordinate ratio(cg->_ndimension);
 
@@ -1157,7 +1157,7 @@ unvectorizeToLexOrdArray(std::vector<sobj> &out, const Lattice<vobj> &in)
 
       int lex;
       Lexicographic::IndexFromCoor(lcoor, lex, in_grid->_ldimensions);
-      assert(lex < out.size());
+      GRID_ASSERT(lex < out.size());
       out_ptrs[lane] = &out[lex];
     }
     
@@ -1221,7 +1221,7 @@ vectorizeFromLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out)
   typedef typename vobj::vector_type vtype;
   
   GridBase* grid = out.Grid();
-  assert(in.size()==grid->lSites());
+  GRID_ASSERT(in.size()==grid->lSites());
   
   const int ndim     = grid->Nd();
   constexpr int nsimd    = vtype::Nsimd();
@@ -1268,7 +1268,7 @@ vectorizeFromRevLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out)
   typedef typename vobj::vector_type vtype;
   
   GridBase* grid = out._grid;
-  assert(in.size()==grid->lSites());
+  GRID_ASSERT(in.size()==grid->lSites());
   
   int ndim     = grid->Nd();
   int nsimd    = vtype::Nsimd();
@@ -1329,9 +1329,9 @@ void precisionChangeFast(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
 template<class VobjOut, class VobjIn>
 void precisionChangeOrig(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
 {
-  assert(out.Grid()->Nd() == in.Grid()->Nd());
+  GRID_ASSERT(out.Grid()->Nd() == in.Grid()->Nd());
   for(int d=0;d<out.Grid()->Nd();d++){
-    assert(out.Grid()->FullDimensions()[d] == in.Grid()->FullDimensions()[d]);
+    GRID_ASSERT(out.Grid()->FullDimensions()[d] == in.Grid()->FullDimensions()[d]);
   }
   out.Checkerboard() = in.Checkerboard();
   GridBase *in_grid=in.Grid();
@@ -1382,9 +1382,9 @@ class precisionChangeWorkspace{
 public:
   precisionChangeWorkspace(GridBase *out_grid, GridBase *in_grid): _out_grid(out_grid), _in_grid(in_grid){
     //Build a map between the sites and lanes of the output field and the input field as we cannot use the Grids on the device
-    assert(out_grid->Nd() == in_grid->Nd());
+    GRID_ASSERT(out_grid->Nd() == in_grid->Nd());
     for(int d=0;d<out_grid->Nd();d++){
-      assert(out_grid->FullDimensions()[d] == in_grid->FullDimensions()[d]);
+      GRID_ASSERT(out_grid->FullDimensions()[d] == in_grid->FullDimensions()[d]);
     }
     int Nsimd_out = out_grid->Nsimd();
 
@@ -1549,7 +1549,7 @@ void Grid_split(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
 
   int full_vecs   = full.size();
 
-  assert(full_vecs>=1);
+  GRID_ASSERT(full_vecs>=1);
 
   GridBase * full_grid = full[0].Grid();
   GridBase *split_grid = split.Grid();
@@ -1567,18 +1567,18 @@ void Grid_split(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
   //////////////////////////////
   // Checks
   //////////////////////////////
-  assert(full_grid->_ndimension==split_grid->_ndimension);
+  GRID_ASSERT(full_grid->_ndimension==split_grid->_ndimension);
   for(int n=0;n<full_vecs;n++){
-    assert(full[n].Checkerboard() == cb);
+    GRID_ASSERT(full[n].Checkerboard() == cb);
     for(int d=0;d<ndim;d++){
-      assert(full[n].Grid()->_gdimensions[d]==split.Grid()->_gdimensions[d]);
-      assert(full[n].Grid()->_fdimensions[d]==split.Grid()->_fdimensions[d]);
+      GRID_ASSERT(full[n].Grid()->_gdimensions[d]==split.Grid()->_gdimensions[d]);
+      GRID_ASSERT(full[n].Grid()->_fdimensions[d]==split.Grid()->_fdimensions[d]);
     }
   }
 
   int   nvector   =full_nproc/split_nproc; 
-  assert(nvector*split_nproc==full_nproc);
-  assert(nvector == full_vecs);
+  GRID_ASSERT(nvector*split_nproc==full_nproc);
+  GRID_ASSERT(nvector == full_vecs);
 
   Coordinate ratio(ndim);
   for(int d=0;d<ndim;d++){
@@ -1622,7 +1622,7 @@ void Grid_split(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
 
       int fvol   = lsites;
       
-      int chunk  = (nvec*fvol)/sP;          assert(chunk*sP == nvec*fvol);
+      int chunk  = (nvec*fvol)/sP;          GRID_ASSERT(chunk*sP == nvec*fvol);
 
       // Loop over reordered data post A2A
       thread_for(c, chunk, {
@@ -1675,7 +1675,7 @@ void Grid_unsplit(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
 
   int full_vecs   = full.size();
 
-  assert(full_vecs>=1);
+  GRID_ASSERT(full_vecs>=1);
 
   GridBase * full_grid = full[0].Grid();
   GridBase *split_grid = split.Grid();
@@ -1693,18 +1693,18 @@ void Grid_unsplit(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
   //////////////////////////////
   // Checks
   //////////////////////////////
-  assert(full_grid->_ndimension==split_grid->_ndimension);
+  GRID_ASSERT(full_grid->_ndimension==split_grid->_ndimension);
   for(int n=0;n<full_vecs;n++){
-    assert(full[n].Checkerboard() == cb);
+    GRID_ASSERT(full[n].Checkerboard() == cb);
     for(int d=0;d<ndim;d++){
-      assert(full[n].Grid()->_gdimensions[d]==split.Grid()->_gdimensions[d]);
-      assert(full[n].Grid()->_fdimensions[d]==split.Grid()->_fdimensions[d]);
+      GRID_ASSERT(full[n].Grid()->_gdimensions[d]==split.Grid()->_gdimensions[d]);
+      GRID_ASSERT(full[n].Grid()->_fdimensions[d]==split.Grid()->_fdimensions[d]);
     }
   }
 
   int   nvector   =full_nproc/split_nproc; 
-  assert(nvector*split_nproc==full_nproc);
-  assert(nvector == full_vecs);
+  GRID_ASSERT(nvector*split_nproc==full_nproc);
+  GRID_ASSERT(nvector == full_vecs);
 
   Coordinate ratio(ndim);
   for(int d=0;d<ndim;d++){
@@ -1740,7 +1740,7 @@ void Grid_unsplit(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
       auto lsites= rsites/M;                // Decreases rsites by M
       
       int fvol   = lsites;
-      int chunk  = (nvec*fvol)/sP;          assert(chunk*sP == nvec*fvol);
+      int chunk  = (nvec*fvol)/sP;          GRID_ASSERT(chunk*sP == nvec*fvol);
 	
       {
 	// Loop over reordered data post A2A

Grid/lattice/Lattice_view.h

@@ -106,6 +106,47 @@ public:
   }
 
 };
+
+#ifdef GRID_LOG_VIEWS
+// Little autoscope assister
+template<class View> 
+class ViewCloser
+{
+  View v;  // Take a copy of view and call view close when I go out of scope automatically
+  const char* filename; int line, mode;
+public:
+  ViewCloser(View &_v, const char* _filename, int _line, int _mode) :
+    v(_v), filename(_filename), line(_line), mode(_mode) {
+    
+    switch (mode){
+    case AcceleratorRead:
+    case AcceleratorWrite:
+    case CpuRead:
+    case CpuWrite:
+      ViewLogger::LogOpen(filename, line, 1, mode, &v[0], v.size() * sizeof(v[0]));
+      break;
+    } 
+    
+  };
+  ~ViewCloser() {
+    
+    switch (mode) {
+    case AcceleratorWriteDiscard:
+    case AcceleratorWrite:
+    case CpuWrite:
+      ViewLogger::LogClose(filename, line, -1, mode, &v[0], v.size() * sizeof(v[0]));
+      break;
+    }
+    
+    v.ViewClose();
+  }
+};
+
+#define autoView(l_v,l,mode)				\
+	  auto l_v = l.View(mode);			\
+	  ViewCloser<decltype(l_v)> _autoView##l_v(l_v,__FILE__,__LINE__,mode);
+
+#else
 // Little autoscope assister
 template<class View> 
 class ViewCloser
@@ -119,6 +160,7 @@ class ViewCloser
 #define autoView(l_v,l,mode)				\
 	  auto l_v = l.View(mode);			\
 	  ViewCloser<decltype(l_v)> _autoView##l_v(l_v);
+#endif
 
 /////////////////////////////////////////////////////////////////////////////////////////
 // Lattice expression types used by ET to assemble the AST

Grid/lattice/PaddedCell.h

@@ -54,7 +54,7 @@ struct CshiftImplGauge: public CshiftImplBase<typename Gimpl::GaugeLinkField::ve
  *
  */
 
-template<class vobj> inline void ScatterSlice(const cshiftVector<vobj> &buf,
+template<class vobj> inline void ScatterSlice(const deviceVector<vobj> &buf,
 					      Lattice<vobj> &lat,
 					      int x,
 					      int dim,
@@ -82,10 +82,10 @@ template<class vobj> inline void ScatterSlice(const cshiftVector<vobj> &buf,
 
   int rNsimd = 1; for(int d=0;d<Nd;d++) rNsimd*=rsimd[d];
   int rNsimda= Nsimd/simd[dim]; // should be equal
-  assert(rNsimda==rNsimd);
+  GRID_ASSERT(rNsimda==rNsimd);
   int face_ovol=block*nblock;
 
-  //  assert(buf.size()==face_ovol*rNsimd);
+  //  GRID_ASSERT(buf.size()==face_ovol*rNsimd);
 
   /*This will work GPU ONLY unless rNsimd is put in the lexico index*/
   //Let's make it work on GPU and then make a special accelerator_for that
@@ -140,7 +140,7 @@ template<class vobj> inline void ScatterSlice(const cshiftVector<vobj> &buf,
   });
 }
 
-template<class vobj> inline void GatherSlice(cshiftVector<vobj> &buf,
+template<class vobj> inline void GatherSlice(deviceVector<vobj> &buf,
 					     const Lattice<vobj> &lat,
 					     int x,
 					     int dim,
@@ -172,7 +172,7 @@ template<class vobj> inline void GatherSlice(cshiftVector<vobj> &buf,
   
   int face_ovol=block*nblock;
 
-  //  assert(buf.size()==face_ovol*rNsimd);
+  //  GRID_ASSERT(buf.size()==face_ovol*rNsimd);
 
   /*This will work GPU ONLY unless rNsimd is put in the lexico index*/
   //Let's make it work on GPU and then make a special accelerator_for that
@@ -247,7 +247,7 @@ public:
     Coordinate local     =unpadded_grid->LocalDimensions();
     Coordinate procs     =unpadded_grid->ProcessorGrid();
     for(int d=0;d<dims;d++){
-      if ( procs[d] > 1 ) assert(local[d]>=depth);
+      if ( procs[d] > 1 ) GRID_ASSERT(local[d]>=depth);
     }
   }
   void DeleteGrids(void)
@@ -448,9 +448,9 @@ public:
     int nld   = to.Grid()->_ldimensions[dimension];
     const int Nsimd = vobj::Nsimd();
 
-    assert(depth<=lds[dimension]); // A must be on neighbouring node
-    assert(depth>0);   // A caller bug if zero
-    assert(ld+2*depth==nld);
+    GRID_ASSERT(depth<=lds[dimension]); // A must be on neighbouring node
+    GRID_ASSERT(depth>0);   // A caller bug if zero
+    GRID_ASSERT(ld+2*depth==nld);
     ////////////////////////////////////////////////////////////////////////////
     // Face size and byte calculations
     ////////////////////////////////////////////////////////////////////////////
@@ -460,15 +460,21 @@ public:
     }
     buffer_size = buffer_size  / Nsimd;
     int rNsimd = Nsimd / simd[dimension];
-    assert( buffer_size == from.Grid()->_slice_nblock[dimension]*from.Grid()->_slice_block[dimension] / simd[dimension]);
+    GRID_ASSERT( buffer_size == from.Grid()->_slice_nblock[dimension]*from.Grid()->_slice_block[dimension] / simd[dimension]);
 
-    static cshiftVector<vobj> send_buf; 
-    static cshiftVector<vobj> recv_buf;
+    static deviceVector<vobj> send_buf; 
+    static deviceVector<vobj> recv_buf;
     send_buf.resize(buffer_size*2*depth);    
     recv_buf.resize(buffer_size*2*depth);
+#ifndef ACCELERATOR_AWARE_MPI
+    static hostVector<vobj> hsend_buf; 
+    static hostVector<vobj> hrecv_buf;
+    hsend_buf.resize(buffer_size*2*depth);    
+    hrecv_buf.resize(buffer_size*2*depth);
+#endif    
 
-    std::vector<CommsRequest_t> fwd_req;   
-    std::vector<CommsRequest_t> bwd_req;   
+    std::vector<MpiCommsRequest_t> fwd_req;   
+    std::vector<MpiCommsRequest_t> bwd_req;   
 
     int words = buffer_size;
     int bytes = words * sizeof(vobj);
@@ -495,9 +501,16 @@ public:
       t_gather+=usecond()-t;
 
       t=usecond();
+#ifdef ACCELERATOR_AWARE_MPI
       grid->SendToRecvFromBegin(fwd_req,
 				(void *)&send_buf[d*buffer_size], xmit_to_rank,
 				(void *)&recv_buf[d*buffer_size], recv_from_rank, bytes, tag);
+#else
+      acceleratorCopyFromDevice(&send_buf[d*buffer_size],&hsend_buf[d*buffer_size],bytes);
+      grid->SendToRecvFromBegin(fwd_req,
+				(void *)&hsend_buf[d*buffer_size], xmit_to_rank,
+				(void *)&hrecv_buf[d*buffer_size], recv_from_rank, bytes, tag);
+#endif
       t_comms+=usecond()-t;
      }
     for ( int d=0;d < depth ; d ++ ) {
@@ -508,9 +521,16 @@ public:
       t_gather+= usecond() - t;
 
       t=usecond();
+#ifdef ACCELERATOR_AWARE_MPI
       grid->SendToRecvFromBegin(bwd_req,
 				(void *)&send_buf[(d+depth)*buffer_size], recv_from_rank,
 				(void *)&recv_buf[(d+depth)*buffer_size], xmit_to_rank, bytes,tag);
+#else
+      acceleratorCopyFromDevice(&send_buf[(d+depth)*buffer_size],&hsend_buf[(d+depth)*buffer_size],bytes);
+      grid->SendToRecvFromBegin(bwd_req,
+				(void *)&hsend_buf[(d+depth)*buffer_size], recv_from_rank,
+				(void *)&hrecv_buf[(d+depth)*buffer_size], xmit_to_rank, bytes,tag);
+#endif      
       t_comms+=usecond()-t;
     }
 
@@ -533,8 +553,13 @@ public:
 
     t=usecond();
     grid->CommsComplete(fwd_req);
+#ifndef ACCELERATOR_AWARE_MPI
+    for ( int d=0;d < depth ; d ++ ) {
+      acceleratorCopyToDevice(&hrecv_buf[d*buffer_size],&recv_buf[d*buffer_size],bytes);
+    }
+#endif
     t_comms+= usecond() - t;
-
+    
     t=usecond();
     for ( int d=0;d < depth ; d ++ ) {
       ScatterSlice(recv_buf,to,nld-depth+d,dimension,plane*buffer_size); plane++;
@@ -543,6 +568,11 @@ public:
 
     t=usecond();
     grid->CommsComplete(bwd_req);
+#ifndef ACCELERATOR_AWARE_MPI
+    for ( int d=0;d < depth ; d ++ ) {
+      acceleratorCopyToDevice(&hrecv_buf[(d+depth)*buffer_size],&recv_buf[(d+depth)*buffer_size],bytes);
+    }
+#endif
     t_comms+= usecond() - t;
     
     t=usecond();

Grid/log/Log.cc

@@ -69,6 +69,7 @@ GridLogger GridLogMemory (1, "Memory", GridLogColours, "NORMAL");
 GridLogger GridLogTracing(1, "Tracing", GridLogColours, "NORMAL");
 GridLogger GridLogDebug  (1, "Debug", GridLogColours, "PURPLE");
 GridLogger GridLogPerformance(1, "Performance", GridLogColours, "GREEN");
+GridLogger GridLogComms      (1, "Comms",  GridLogColours, "BLUE");
 GridLogger GridLogDslash     (1, "Dslash", GridLogColours, "BLUE");
 GridLogger GridLogIterative  (1, "Iterative", GridLogColours, "BLUE");
 GridLogger GridLogIntegrator (1, "Integrator", GridLogColours, "BLUE");
@@ -84,6 +85,7 @@ void GridLogConfigure(std::vector<std::string> &logstreams) {
   GridLogDebug.Active(0);
   GridLogPerformance.Active(0);
   GridLogDslash.Active(0);
+  GridLogComms.Active(0);
   GridLogIntegrator.Active(1);
   GridLogColours.Active(0);
   GridLogHMC.Active(1);
@@ -97,6 +99,7 @@ void GridLogConfigure(std::vector<std::string> &logstreams) {
     if (logstreams[i] == std::string("Debug"))       GridLogDebug.Active(1);
     if (logstreams[i] == std::string("Performance")) GridLogPerformance.Active(1);
     if (logstreams[i] == std::string("Dslash"))      GridLogDslash.Active(1);
+    if (logstreams[i] == std::string("Comms"))       GridLogComms.Active(1);
     if (logstreams[i] == std::string("NoIntegrator"))GridLogIntegrator.Active(0);
     if (logstreams[i] == std::string("NoHMC"))       GridLogHMC.Active(0);
     if (logstreams[i] == std::string("Colours"))     GridLogColours.Active(1);

Grid/log/Log.h

@@ -33,10 +33,6 @@
 #ifndef GRID_LOG_H
 #define GRID_LOG_H
 
-#ifdef HAVE_EXECINFO_H
-#include <execinfo.h>
-#endif
-
 NAMESPACE_BEGIN(Grid);
 
 //////////////////////////////////////////////////////////////////////////////////////////////////
@@ -180,6 +176,7 @@ extern GridLogger GridLogError;
 extern GridLogger GridLogWarning;
 extern GridLogger GridLogMessage;
 extern GridLogger GridLogDebug;
+extern GridLogger GridLogComms;
 extern GridLogger GridLogPerformance;
 extern GridLogger GridLogDslash;
 extern GridLogger GridLogIterative;
@@ -226,8 +223,6 @@ inline void Grid_pass(Args&&... args) {
     std::cout << "\033[32m" << GridLogMessage << msg << "\033[0m" << std::endl;
 }
 
-#define _NBACKTRACE (256)
-extern void * Grid_backtrace_buffer[_NBACKTRACE];
 
 #define BACKTRACEFILE() {						\
     char string[20];							\