Fix zeroing

.github/workflows/main.yml

@@ -2,6 +2,8 @@
 name: CI
 
 on:
+  push:
+    branches: [ master, cuda ]
   pull_request:
     branches: [ master, cuda ]
 

README.org

@@ -69,10 +69,10 @@ And then you can see the =configure= options
 ../../configure --help
 #+end_src
 
-** Benchmarks
+** Benches
 
 The script =tools/configure-benches.sh= can be used to create
-a couple of configurations for benchmarks:
+a couple of configurations for benches:
 
 #+begin_src sh :exports results :results verbatim org   :results verbatim drawer replace output 
 awk '/begin +doc/,/end +doc/ { print $NL }' tools/configure-benches.sh |
@@ -87,8 +87,49 @@ sed "s/^# //; s/^# *$//; /^$/d"
   and without computing slices.
 - only-dgemm ::
   This only runs the computation part that involves dgemms.
-- slices-on-gpu-only-dgemm ::
+- cuda-only-dgemm ::
+  This is the naive CUDA implementation compiling only the dgemm parts
+  of the compute.
+- cuda-slices-on-gpu-only-dgemm ::
   This configuration tests that slices reside completely on the gpu
   and it should use a CUDA aware MPI implementation.
   It also only uses the routines that involve dgemm.
 :end:
+
+In order to generate the benches just create a suitable directory for it
+
+#+begin_src sh :eval no
+mkdir -p build/benches
+cd buid/benches
+../../tools/configure-benches.sh CXX=g++ ...
+#+end_src
+
+and you will get a Makefile together with several project folders.
+You can either configure all projects with =make all= or
+then go in each folder.
+
+Notice that you can give a path for ctf for all of them by doing
+#+begin_src sh :eval no
+../../tools/configure-benches.sh --with-ctf=/absolute/path/to/ctf
+#+end_src
+
+* Running benches
+
+** Main benchmark
+
+The main benchmark gets built in =bench/atrip= and is used to run an
+atrip run with random tensors.
+
+A common run of this script will be the following
+
+#+begin_src sh
+bench/atrip \
+  --no 100 \
+  --nv 1000 \
+  --mod 1 \
+  --% 0 \
+  --dist group \
+  --nocheckpoint \
+  --max-iterations 1000
+#+end_src
+

bench/main.cxx

@@ -5,18 +5,20 @@
 #include <CLI11.hpp>
 
 #define _print_size(what, size)                 \
+  do {                                          \
     if (rank == 0) {                            \
       std::cout << #what                        \
                 << " => "                       \
                 << (double)size * elem_to_gb    \
                 << "GB"                         \
                 << std::endl;                   \
-  }
+    }                                           \
+  } while (0)
 
 int main(int argc, char** argv) {
   MPI_Init(&argc, &argv);
 
-  size_t checkpoint_it;
+  size_t checkpoint_it, max_iterations;
   int no(10), nv(100), itMod(-1), percentageMod(10);
   float checkpoint_percentage;
   bool
@@ -30,6 +32,9 @@ int main(int argc, char** argv) {
   app.add_option("--no", no, "Occupied orbitals");
   app.add_option("--nv", nv, "Virtual orbitals");
   app.add_option("--mod", itMod, "Iteration modifier");
+  app.add_option("--max-iterations",
+                 max_iterations,
+                 "Maximum number of iterations to run");
   app.add_flag("--keep-vppph", keepVppph, "Do not delete Vppph");
   app.add_flag("--nochrono", nochrono, "Do not print chrono");
   app.add_flag("--rank-round-robin", rankRoundRobin, "Do rank round robin");
@@ -45,6 +50,19 @@ int main(int argc, char** argv) {
                  checkpoint_percentage,
                  "Percentage for checkpoints");
 
+  // Optional tensor files
+  std::string
+    ei_path, ea_path,
+    Tph_path, Tpphh_path,
+    Vpphh_path, Vhhhp_path, Vppph_path;
+  app.add_option("--ei", ei_path, "Path for ei");
+  app.add_option("--ea", ea_path, "Path for ea");
+  app.add_option("--Tpphh", Tpphh_path, "Path for Tpphh");
+  app.add_option("--Tph", Tph_path, "Path for Tph");
+  app.add_option("--Vpphh", Vpphh_path, "Path for Vpphh");
+  app.add_option("--Vhhhp", Vhhhp_path, "Path for Vhhhp");
+  app.add_option("--Vppph", Vppph_path, "Path for Vppph");
+
 #if defined(HAVE_CUDA)
   size_t ooo_threads = 0, ooo_blocks = 0;
   app.add_option("--ooo-blocks",
@@ -148,37 +166,64 @@ int main(int argc, char** argv) {
   }
 
 
-  std::vector<int> symmetries(4, NS)
+  std::vector<int>
-                 , vo({nv, no})
+    symmetries(4, NS),
-                 , vvoo({nv, nv, no, no})
+    vo({nv, no}),
-                 , ooov({no, no, no, nv})
+    vvoo({nv, nv, no, no}),
-                 , vvvo({nv, nv, nv, no})
+    ooov({no, no, no, nv}),
-                 ;
+    vvvo({nv, nv, nv, no});
 
   CTF::Tensor<double>
-      ei(1, ooov.data(), symmetries.data(), world)
+    ei(1, ooov.data(), symmetries.data(), world),
-    , ea(1, vo.data(), symmetries.data(), world)
+    ea(1, vo.data(), symmetries.data(), world),
-    , Tph(2, vo.data(), symmetries.data(), world)
+    Tph(2, vo.data(), symmetries.data(), world),
-    , Tpphh(4, vvoo.data(), symmetries.data(), world)
+    Tpphh(4, vvoo.data(), symmetries.data(), world),
-    , Vpphh(4, vvoo.data(), symmetries.data(), world)
+    Vpphh(4, vvoo.data(), symmetries.data(), world),
-    , Vhhhp(4, ooov.data(), symmetries.data(), world)
+    Vhhhp(4, ooov.data(), symmetries.data(), world);
-    ;
 
   // initialize deletable tensors in heap
   auto Vppph
     = new CTF::Tensor<double>(4, vvvo.data(), symmetries.data(), world);
 
-  _print_size(Vabci, no*nv*nv*nv)
+  _print_size(Vabci, no*nv*nv*nv);
-  _print_size(Vabij, no*no*nv*nv)
+  _print_size(Vabij, no*no*nv*nv);
-  _print_size(Vijka, no*no*no*nv)
+  _print_size(Vijka, no*no*no*nv);
 
+  if (ei_path.size()) {
+    ei.read_dense_from_file(ei_path.c_str());
+  } else {
     ei.fill_random(-40.0, -2);
+  }
+  if (ea_path.size()) {
+    ea.read_dense_from_file(ea_path.c_str());
+  } else {
     ea.fill_random(2, 50);
+  }
+  if (Tpphh_path.size()) {
+    Tpphh.read_dense_from_file(Tpphh_path.c_str());
+  } else {
     Tpphh.fill_random(0, 1);
+  }
+  if (Tph_path.size()) {
+    Tph.read_dense_from_file(Tph_path.c_str());
+  } else {
     Tph.fill_random(0, 1);
+  }
+  if (Vpphh_path.size()) {
+    Vpphh.read_dense_from_file(Vpphh_path.c_str());
+  } else {
     Vpphh.fill_random(0, 1);
+  }
+  if (Vhhhp_path.size()) {
+    Vhhhp.read_dense_from_file(Vhhhp_path.c_str());
+  } else {
     Vhhhp.fill_random(0, 1);
+  }
+  if (Vppph_path.size()) {
+    Vppph->read_dense_from_file(Vppph_path.c_str());
+  } else {
     Vppph->fill_random(0, 1);
+  }
 
   atrip::Atrip::init(MPI_COMM_WORLD);
   const auto in
@@ -199,6 +244,7 @@ int main(int argc, char** argv) {
        .with_iterationMod(itMod)
        .with_percentageMod(percentageMod)
        .with_tuplesDistribution(tuplesDistribution)
+       .with_maxIterations(max_iterations)
        // checkpoint options
        .with_checkpointAtEveryIteration(checkpoint_it)
        .with_checkpointAtPercentage(checkpoint_percentage)

configure.ac

@@ -164,8 +164,7 @@ AC_TYPE_SIZE_T
 dnl -----------------------------------------------------------------------
 dnl CHECK CTF
 if test xYES = x${BUILD_CTF}; then
-  AC_MSG_WARN([Sorry, building CTF not supported yet provide a build path
+  AC_MSG_WARN([You will have to do make ctf before building the project.])
-                      with --with-ctf=path/to/ctf/installation])
 else
   CPPFLAGS="$CPPFLAGS -I${LIBCTF_CPATH}"
   LDFLAGS="$LDFLAGS -L${LIBCTF_LD_LIBRARY_PATH} -lctf"

etc/env/raven/cuda

@@ -0,0 +1,56 @@
+mods=(
+  cuda/11.6
+  intel/19.1.2
+  mkl/2020.4
+  impi/2019.8
+  autoconf/2.69
+  automake/1.15
+  libtool/2.4.6
+)
+
+
+module purge
+module load ${mods[@]}
+LIB_PATH="${CUDA_HOME}/lib64"
+export CUDA_ROOT=${CUDA_HOME}
+export CUDA_LDFLAGS="-L${LIB_PATH} -lcuda -L${LIB_PATH} -lcudart -L${LIB_PATH} -lcublas"
+export CUDA_CXXFLAGS="-I${CUDA_HOME}/include"
+
+export LD_LIBRARY_PATH="${MKL_HOME}/lib/intel64_lin:${LD_LIBRARY_PATH}"
+
+BLAS_STATIC_PATH="$MKL_HOME/lib/intel64/libmkl_intel_lp64.a"
+
+ls ${LIB_PATH}/libcublas.so
+ls ${LIB_PATH}/libcudart.so
+
+cat <<EOF
+
+////////////////////////////////////////////////////////////////////////////////
+                                    info
+////////////////////////////////////////////////////////////////////////////////
+
+
+MKL_HOME         = $MKL_HOME
+BLAS_STATIC_PATH = $BLAS_STATIC_PATH
+
+CUDA_ROOT      = ${CUDA_HOME}
+CUDA_LDFLAGS   = "-L${LIB_PATH} -lcuda -L${LIB_PATH} -lcudart -L${LIB_PATH} -lcublas"
+CUDA_CXXFLAGS  = "-I${CUDA_HOME}/include"
+
+
+
+Consider now runnng the following
+
+../configure \\
+  --enable-cuda \\
+  --disable-slice \\
+  --with-blas="-L\$MKL_HOME/lib/intel64/ -lmkl_intel_lp64 -mkl" \\
+  CXX=mpiicpc \\
+  CC=mpiicc \\
+  MPICXX=mpiicpc
+
+
+EOF
+	
+
+return

include/atrip/Atrip.hpp

@@ -86,7 +86,7 @@ namespace atrip {
       ADD_ATTRIBUTE(bool, rankRoundRobin, false)
       ADD_ATTRIBUTE(bool, chrono, false)
       ADD_ATTRIBUTE(bool, barrier, false)
-      ADD_ATTRIBUTE(int, maxIterations, 0)
+      ADD_ATTRIBUTE(size_t, maxIterations, 0)
       ADD_ATTRIBUTE(int, iterationMod, -1)
       ADD_ATTRIBUTE(int, percentageMod, -1)
       ADD_ATTRIBUTE(TuplesDistribution, tuplesDistribution, NAIVE)

include/atrip/CUDA.hpp

@@ -11,11 +11,22 @@
 #if defined(HAVE_CUDA) && defined(__CUDACC__)
 #  define __MAYBE_GLOBAL__ __global__
 #  define __MAYBE_DEVICE__ __device__
+#  define __MAYBE_HOST__ __host__
+#  define __INLINE__ __inline__
 #else
 #  define __MAYBE_GLOBAL__
 #  define __MAYBE_DEVICE__
+#  define __MAYBE_HOST__
+#  define __INLINE__ inline
 #endif
 
+#if defined(HAVE_CUDA)
+#define ACC_FUNCALL(fname, i, j, ...) fname<<<(i), (j)>>>(__VA_ARGS__)
+#else
+#define ACC_FUNCALL(fname, i, j, ...) fname(__VA_ARGS__)
+#endif /*  defined(HAVE_CUDA) */
+
+
 #define _CHECK_CUDA_SUCCESS(message, ...)                               \
   do {                                                                  \
     CUresult result = __VA_ARGS__;                                      \

include/atrip/Equations.hpp

@@ -23,6 +23,8 @@
 #include<thrust/device_vector.h>
 #endif
 
+#include<atrip/CUDA.hpp>
+
 
 namespace atrip {
 using ABCTuple = std::array<size_t, 3>;
@@ -32,21 +34,25 @@ using ABCTuples = std::vector<ABCTuple>;
 
 // [[file:~/cuda/atrip/atrip.org::*Energy][Energy:1]]
 template <typename F=double>
-double getEnergyDistinct
+__MAYBE_GLOBAL__
+void getEnergyDistinct
   ( F const epsabc
   , size_t const No
   , F* const epsi
   , F* const Tijk
   , F* const Zijk
+  , double* energy
   );
 
 template <typename F=double>
-double getEnergySame
+__MAYBE_GLOBAL__
+void getEnergySame
   ( F const epsabc
   , size_t const No
   , F* const epsi
   , F* const Tijk
   , F* const Zijk
+  , double* energy
   );
 // Energy:1 ends here
 
@@ -97,6 +103,11 @@ void singlesContribution
     // -- TIJK
     // , DataPtr<F> Tijk
     , DataFieldType<F>* Tijk_
+#if defined(HAVE_CUDA)
+     // -- tmp buffers
+    , DataFieldType<F>* _t_buffer
+    , DataFieldType<F>* _vhhh 
+#endif
     );
 // Doubles contribution:1 ends here
 

include/atrip/Operations.hpp

@@ -0,0 +1,171 @@
+// Copyright 2022 Alejandro Gallo
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef OPERATIONS_HPP_
+#define OPERATIONS_HPP_
+
+#include <atrip/CUDA.hpp>
+#include <atrip/Types.hpp>
+#include <atrip/Complex.hpp>
+
+namespace atrip {
+namespace acc {
+
+  // cuda kernels
+
+  template <typename F>
+  __MAYBE_GLOBAL__
+  void zeroing(F* a, size_t n) {
+    F zero = {0};
+    for (size_t i = 0; i < n; i++) {
+      a[i] = zero;
+    }
+  }
+
+  ////
+  template <typename F>
+  __MAYBE_DEVICE__ __MAYBE_HOST__ __INLINE__
+  F maybeConjugateScalar(const F &a) { return a; }
+
+#if defined(HAVE_CUDA)
+  template <>
+  __MAYBE_DEVICE__ __MAYBE_HOST__ __INLINE__
+  cuDoubleComplex maybeConjugateScalar(const cuDoubleComplex &a) {
+    return {a.x, -a.y};
+  }
+#endif /*  defined(HAVE_CUDA) */
+
+  template <typename F>
+  __MAYBE_GLOBAL__
+  void maybeConjugate(F* to, F* from, size_t n) {
+    for (size_t i = 0; i < n; ++i) {
+      to[i] = maybeConjugateScalar<F>(from[i]);
+    }
+  }
+
+
+  template <typename F>
+  __MAYBE_DEVICE__ __MAYBE_HOST__
+  void reorder(F* to, F* from, size_t size, size_t I, size_t J, size_t K) {
+    size_t idx = 0;
+    const size_t IDX = I + J*size + K*size*size;
+    for (size_t k = 0; k < size; k++)
+    for (size_t j = 0; j < size; j++)
+    for (size_t i = 0; i < size; i++, idx++)
+      to[idx] += from[IDX];
+  }
+
+  // Multiplication operation
+  //////////////////////////////////////////////////////////////////////////////
+
+  template <typename F>
+  __MAYBE_DEVICE__ __MAYBE_HOST__ __INLINE__
+  F prod(const F &a, const F &b) { return a * b; }
+
+#if defined(HAVE_CUDA)
+  template <>
+  __MAYBE_DEVICE__ __MAYBE_HOST__ __INLINE__
+  cuDoubleComplex prod(const cuDoubleComplex &a, const cuDoubleComplex &b) {
+    return cuCmul(a, b);
+  }
+#endif /*  defined(HAVE_CUDA) */
+
+  // Division operation
+  //////////////////////////////////////////////////////////////////////////////
+
+  template <typename F>
+  __MAYBE_DEVICE__ __MAYBE_HOST__ __INLINE__
+  F div(const F &a, const F &b) { return a / b; }
+
+#if defined(HAVE_CUDA)
+  template <>
+  __MAYBE_DEVICE__ __MAYBE_HOST__ __INLINE__
+  cuDoubleComplex div(const cuDoubleComplex &a, const cuDoubleComplex &b) {
+    return cuCdiv(a, b);
+  }
+#endif /*  defined(HAVE_CUDA) */
+
+  // Real part
+  //////////////////////////////////////////////////////////////////////////////
+
+  template <typename F>
+  __MAYBE_HOST__ __INLINE__
+  double real(F &a) { return std::real(a); }
+
+  template <>
+  __MAYBE_DEVICE__ __MAYBE_HOST__ __INLINE__
+  double real(double &a) {
+    return a;
+  }
+
+#if defined(HAVE_CUDA)
+  template <>
+  __MAYBE_DEVICE__ __MAYBE_HOST__ __INLINE__
+  double real(cuDoubleComplex &a) {
+    return cuCreal(a);
+  }
+#endif /*  defined(HAVE_CUDA) */
+
+  // Substraction operator
+  //////////////////////////////////////////////////////////////////////////////
+
+  template <typename F>
+  __MAYBE_DEVICE__ __MAYBE_HOST__ __INLINE__
+  F sub(const F &a, const F &b) { return a - b; }
+
+#if defined(HAVE_CUDA)
+  template <>
+  __MAYBE_DEVICE__ __MAYBE_HOST__ __INLINE__
+  cuDoubleComplex sub(const cuDoubleComplex &a,
+                               const cuDoubleComplex &b) {
+    return cuCsub(a, b);
+  }
+#endif /*  defined(HAVE_CUDA) */
+
+  // Addition operator
+  //////////////////////////////////////////////////////////////////////////////
+
+  template <typename F>
+  __MAYBE_DEVICE__ __MAYBE_HOST__ __INLINE__
+  F add(const F &a, const F &b) { return a + b; }
+
+#if defined(HAVE_CUDA)
+  template <>
+  __MAYBE_DEVICE__ __MAYBE_HOST__ __INLINE__
+  cuDoubleComplex add(const cuDoubleComplex &a, const cuDoubleComplex &b) {
+    return cuCadd(a, b);
+  }
+#endif /*  defined(HAVE_CUDA) */
+
+  // Sum in place operator
+  //////////////////////////////////////////////////////////////////////////////
+
+  template <typename F>
+  __MAYBE_DEVICE__ __MAYBE_HOST__
+  void sum_in_place(F* to, const F* from) { *to += *from; }
+
+#if defined(HAVE_CUDA)
+  template <>
+  __MAYBE_DEVICE__ __MAYBE_HOST__
+  void sum_in_place(cuDoubleComplex* to, const cuDoubleComplex* from) {
+    to->x += from->x;
+    to->y += from->y;
+  }
+#endif /*  defined(HAVE_CUDA) */
+
+
+}  // namespace acc
+}  // namespace atrip
+
+#endif

include/atrip/Slice.hpp

@@ -352,7 +352,7 @@ Info info;
 
 // [[file:~/cuda/atrip/atrip.org::*Attributes][Attributes:2]]
 DataPtr<F> data;
-#if defined(HAVE_CUDA)
+#if defined(HAVE_CUDA) && !defined (ATRIP_SOURCES_IN_GPU)
     F* mpi_data;
 #endif
 // Attributes:2 ends here
@@ -456,7 +456,7 @@ void unwrapAndMarkReady() {
       if (errorCode != MPI_SUCCESS)
         throw "Atrip: Unexpected error MPI ERROR";
 
-#if defined(HAVE_CUDA)
+#if defined(HAVE_CUDA) && !defined(ATRIP_SOURCES_IN_GPU)
       // copy the retrieved mpi data to the device
       WITH_CHRONO("cuda:memcpy",
                   _CHECK_CUDA_SUCCESS("copying mpi data to device",
@@ -488,7 +488,7 @@ void unwrapAndMarkReady() {
 Slice(size_t size_)
       : info({})
       , data(DataNullPtr)
-#if defined(HAVE_CUDA)
+#if defined(HAVE_CUDA) && !defined(ATRIP_SOURCES_IN_GPU)
       , mpi_data(nullptr)
 #endif
       , size(size_)

include/atrip/SliceUnion.hpp

@@ -405,6 +405,7 @@ template <typename F=double>
               , sliceSize(std::accumulate(sliceLength.begin(),
                                           sliceLength.end(),
                                           1UL, std::multiplies<size_t>()))
+
 #if defined(ATRIP_SOURCES_IN_GPU)
               , sources(rankMap.nSources())
 #else
@@ -417,6 +418,23 @@ template <typename F=double>
     { // constructor begin
 
       LOG(0,"Atrip") << "INIT SliceUnion: " << name << "\n";
+        printf("sliceSize %d, number of slices %d\n\n\n", sliceSize, sources.size());
+
+#if defined(ATRIP_SOURCES_IN_GPU)
+      for (auto& ptr: sources) {
+        const CUresult sourceError =
+          cuMemAlloc(&ptr, sizeof(F) * sliceSize);
+        if (ptr == 0UL) {
+          throw "UNSUFICCIENT MEMORY ON THE GRAPHIC CARD FOR SOURCES";
+        }
+        if (sourceError != CUDA_SUCCESS) {
+          std::stringstream s;
+          s << "Error allocating memory for sources "
+            << "code " << sourceError << "\n";
+          throw s.str();
+        }
+      }
+#endif
 
       for (auto& ptr: sliceBuffers) {
 #if defined(HAVE_CUDA)
@@ -445,6 +463,34 @@ template <typename F=double>
                      std::inserter(freePointers, freePointers.begin()),
                      [](DataPtr<F> ptr) { return ptr; });
 
+#if defined(HAVE_CUDA)
+      LOG(1,"Atrip") << "warming communication up " << slices.size() << "\n";
+      WITH_CHRONO("cuda:warmup",
+                  int nRanks=Atrip::np, requestCount=0;
+                  int nSends=sliceBuffers.size()*nRanks;
+                  MPI_Request *requests = (MPI_Request*) malloc(nSends*2 * sizeof(MPI_Request));
+                  MPI_Status *statuses = (MPI_Status*) malloc(nSends*2 * sizeof(MPI_Status));
+                  for (int sliceId=0; sliceId<sliceBuffers.size(); sliceId++){
+                    for (int rankId=0; rankId<nRanks; rankId++){
+                      MPI_Isend((void*)SOURCES_DATA(sources[0]),
+                                sliceSize,
+                                traits::mpi::datatypeOf<F>(),
+                                rankId,
+                                100,
+                                universe,
+                                &requests[requestCount++]);
+                      MPI_Irecv((void*)sliceBuffers[sliceId],
+                                sliceSize,
+                                traits::mpi::datatypeOf<F>(),
+                                rankId,
+                                100,
+                                universe,
+                                &requests[requestCount++]);
+                    }
+                  }
+                  MPI_Waitall(nSends*2, requests, statuses);
+                  )
+#endif
 
 
       LOG(1,"Atrip") << "#slices " << slices.size() << "\n";
@@ -527,12 +573,11 @@ template <typename F=double>
       if (slice.info.state == Slice<F>::Fetch) { // if-1
         // TODO: do it through the slice class
         slice.info.state = Slice<F>::Dispatched;
-#if defined(HAVE_CUDA)
+#if defined(HAVE_CUDA) && defined(ATRIP_SOURCES_IN_GPU)
-#  if !defined(ATRIP_CUDA_AWARE_MPI) && defined(ATRIP_SOURCES_IN_GPU)
+#  if !defined(ATRIP_CUDA_AWARE_MPI) 
 #    error "You need CUDA aware MPI to have slices on the GPU"
 #  endif
-        slice.mpi_data = (F*)malloc(sizeof(F) * slice.size);
+        MPI_Irecv((void*)slice.data,
-        MPI_Irecv(slice.mpi_data,
 #else
         MPI_Irecv(slice.data,
 #endif

src/atrip/Atrip.cxx

@@ -202,7 +202,7 @@ Atrip::Output Atrip::run(Atrip::Input<F> const& in) {
   _CHECK_CUDA_SUCCESS("Zijk",
                       cuMemAlloc(&Zijk, sizeof(F) * No * No * No));
 #else
-  std::vector<F> &Tai = _Tai, &epsi = _epsi, &epsa = _epsa;
+  DataPtr<F> Tai = _Tai.data(), epsi = _epsi.data(), epsa = _epsa.data();
   Zijk = (DataFieldType<F>*)malloc(No*No*No * sizeof(DataFieldType<F>));
   Tijk = (DataFieldType<F>*)malloc(No*No*No * sizeof(DataFieldType<F>));
 #endif
@@ -258,6 +258,25 @@ Atrip::Output Atrip::run(Atrip::Input<F> const& in) {
   // all tensors
   std::vector< SliceUnion<F>* > unions = {&taphh, &hhha, &abph, &abhh, &tabhh};
 
+#ifdef HAVE_CUDA
+    // TODO: free buffers
+    DataFieldType<F>* _t_buffer;
+    DataFieldType<F>* _vhhh;
+    WITH_CHRONO("double:cuda:alloc",
+    _CHECK_CUDA_SUCCESS("Allocating _t_buffer",
+                        cuMemAlloc((CUdeviceptr*)&_t_buffer,
+                                   No*No*No * sizeof(DataFieldType<F>)));
+    _CHECK_CUDA_SUCCESS("Allocating _vhhh",
+                        cuMemAlloc((CUdeviceptr*)&_vhhh,
+                                   No*No*No * sizeof(DataFieldType<F>)));
+                )
+    //const size_t
+     // bs = Atrip::kernelDimensions.ooo.blocks,
+      //ths = Atrip::kernelDimensions.ooo.threads;
+    //cuda::zeroing<<<bs, ths>>>((DataFieldType<F>*)_t_buffer, NoNoNo);
+    //cuda::zeroing<<<bs, ths>>>((DataFieldType<F>*)_vhhh, NoNoNo);
+#endif
+
   // get tuples for the current rank
   TuplesDistribution *distribution;
 
@@ -639,7 +658,14 @@ Atrip::Output Atrip::run(Atrip::Input<F> const& in) {
                                          tabhh.unwrapSlice(Slice<F>::AC, abc),
                                          tabhh.unwrapSlice(Slice<F>::BC, abc),
                                          // -- TIJK
-                                         (DataFieldType<F>*)Tijk);
+                                         (DataFieldType<F>*)Tijk
+#if defined(HAVE_CUDA)
+                                         // -- tmp buffers
+                                         ,(DataFieldType<F>*)_t_buffer
+                                         ,(DataFieldType<F>*)_vhhh
+#endif
+                                         );
+
                   WITH_RANK << iteration << "-th doubles done\n";
       ))
     }
@@ -667,7 +693,7 @@ Atrip::Output Atrip::run(Atrip::Input<F> const& in) {
                                       (DataFieldType<F>*)Tai,
 #else
       singlesContribution<F>(No, Nv, abc[0], abc[1], abc[2],
-                             Tai.data(),
+                             Tai,
 #endif
                              (DataFieldType<F>*)abhh.unwrapSlice(Slice<F>::AB,
                                                                  abc),
@@ -683,31 +709,71 @@ Atrip::Output Atrip::run(Atrip::Input<F> const& in) {
     // COMPUTE ENERGY %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% {{{1
 #if defined(ATRIP_ONLY_DGEMM)
     if (false)
-#endif
+#endif /* defined(ATRIP_ONLY_DGEMM) */
     if (!isFakeTuple(i)) {
-      double tupleEnergy(0.);
+#if defined(HAVE_CUDA)
+      double *tupleEnergy;
+      cuMemAlloc((DataPtr<double>*)&tupleEnergy, sizeof(double));
+#else
+      double _tupleEnergy(0.);
+      double *tupleEnergy = &_tupleEnergy;
+#endif /* defined(HAVE_CUDA) */
 
       int distinct(0);
       if (abc[0] == abc[1]) distinct++;
       if (abc[1] == abc[2]) distinct--;
-      const F epsabc(_epsa[abc[0]] + _epsa[abc[1]] + _epsa[abc[2]]);
+      const double
+        epsabc = std::real(_epsa[abc[0]] + _epsa[abc[1]] + _epsa[abc[2]]);
+
+      DataFieldType<F> _epsabc{epsabc};
 
-      // LOG(0, "AtripCUDA") <<  "doing energy " << i << "distinct " << distinct << "\n";
       WITH_CHRONO("energy",
-/*
+                  if ( distinct == 0) {
-    TODO: think about how to do this on the GPU in the best way possible
+                    ACC_FUNCALL(getEnergyDistinct<DataFieldType<F>>,
-        if ( distinct == 0)
+                                1, 1, // for cuda
-          tupleEnergy = getEnergyDistinct<F>(epsabc, No, (F*)epsi, (F*)Tijk, (F*)Zijk);
+                                _epsabc,
-        else
+                                No,
-          tupleEnergy = getEnergySame<F>(epsabc, No, (F*)epsi, (F*)Tijk, (F*)Zijk);
+#if defined(HAVE_CUDA)
-*/
+                                (DataFieldType<F>*)epsi,
-      )
+                                (DataFieldType<F>*)Tijk,
+                                (DataFieldType<F>*)Zijk,
+#else
+                                epsi,
+                                Tijk,
+                                Zijk,
+#endif
+                                tupleEnergy);
+                  } else {
+                    ACC_FUNCALL(getEnergySame<DataFieldType<F>>,
+                                1, 1, // for cuda
+                                _epsabc,
+                                No,
+#if defined(HAVE_CUDA)
+                                (DataFieldType<F>*)epsi,
+                                (DataFieldType<F>*)Tijk,
+                                (DataFieldType<F>*)Zijk,
+#else
+                                epsi,
+                                Tijk,
+                                Zijk,
+#endif
+                                tupleEnergy);
+                  })
+
+#if defined(HAVE_CUDA)
+        double host_tuple_energy;
+        cuMemcpyDtoH((void*)&host_tuple_energy,
+                     (DataPtr<double>)tupleEnergy,
+                     sizeof(double));
+#else
+        double host_tuple_energy = *tupleEnergy;
+#endif /* defined(HAVE_CUDA) */
 
 #if defined(HAVE_OCD) || defined(ATRIP_PRINT_TUPLES)
-      tupleEnergies[abc] = tupleEnergy;
+      tupleEnergies[abc] = host_tuple_energy;
 #endif
 
-      energy += tupleEnergy;
+      energy += host_tuple_energy;
 
     }
 
@@ -773,6 +839,8 @@ Atrip::Output Atrip::run(Atrip::Input<F> const& in) {
     Atrip::chrono["iterations"].stop();
     // ITERATION END %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%{{{1
 
+    if (in.maxIterations != 0 && i >= in.maxIterations) break;
+
   }
     // END OF MAIN LOOP
 

src/atrip/Equations.cxx

@@ -16,96 +16,13 @@
 #include<atrip/Equations.hpp>
 
 #include<atrip/CUDA.hpp>
+#include<atrip/Operations.hpp>
 
 namespace atrip {
 // Prolog:2 ends here
 
 
 
-#ifdef HAVE_CUDA
-namespace cuda {
-
-  // cuda kernels
-
-  template <typename F>
-  __global__
-  void zeroing(F* a, size_t n) {
-    F zero = {0};
-    for (size_t i = 0; i < n; i++) {
-      a[i] = zero;
-    }
-  }
-
-  ////
-  template <typename F>
-  __device__
-  F maybeConjugateScalar(const F a);
-
-  template <>
-  __device__
-  double maybeConjugateScalar(const double a) { return a; }
-
-  template <>
-  __device__
-  cuDoubleComplex
-  maybeConjugateScalar(const cuDoubleComplex a) {
-    return {a.x, -a.y};
-  }
-
-  template <typename F>
-  __global__
-  void maybeConjugate(F* to, F* from, size_t n) {
-    for (size_t i = 0; i < n; ++i) {
-      to[i] = maybeConjugateScalar<F>(from[i]);
-    }
-  }
-
-
-  template <typename F>
-  __global__
-  void reorder(F* to, F* from, size_t size, size_t I, size_t J, size_t K) {
-    size_t idx = 0;
-    const size_t IDX = I + J*size + K*size*size;
-    for (size_t k = 0; k < size; k++)
-    for (size_t j = 0; j < size; j++)
-    for (size_t i = 0; i < size; i++, idx++)
-      to[idx] += from[IDX];
-  }
-
-  // I mean, really CUDA... really!?
-  template <typename F>
-  __device__
-  F multiply(const F &a, const F &b);
-  template <>
-  __device__
-  double multiply(const double &a, const double &b) { return a * b; }
-
-  template <>
-  __device__
-  cuDoubleComplex multiply(const cuDoubleComplex &a, const cuDoubleComplex &b) {
-    return
-      {a.x * b.x - a.y * b.y,
-       a.x * b.y + a.y * b.x};
-  }
-
-  template <typename F>
-  __device__
-  void sum_in_place(F* to, const F* from);
-
-  template <>
-  __device__
-  void sum_in_place(double* to, const double *from) { *to += *from; }
-
-  template <>
-  __device__
-  void sum_in_place(cuDoubleComplex* to, const cuDoubleComplex* from) {
-    to->x += from->x;
-    to->y += from->y;
-  }
-
-};
-#endif
-
 #if defined(HAVE_CUDA)
 #define FOR_K()                                             \
   for (size_t kmin = blockIdx.x * blockDim.x + threadIdx.x, \
@@ -133,7 +50,7 @@ namespace cuda {
     _REORDER_BODY_(__VA_ARGS__)                 \
       }
 #if defined(HAVE_CUDA)
-#define GO(__TO, __FROM) cuda::sum_in_place<F>(&__TO, &__FROM);
+#define GO(__TO, __FROM) acc::sum_in_place<F>(&__TO, &__FROM);
 #else
 #define GO(__TO, __FROM) __TO += __FROM;
 #endif
@@ -179,162 +96,205 @@ namespace cuda {
 #undef _IJK_
 #undef GO
 
+#if defined(HAVE_CUDA)
+#  define MIN(a, b) min((a), (b))
+#else
+#  define MIN(a, b) std::min((a), (b))
+#endif
+
 
 // [[file:~/cuda/atrip/atrip.org::*Energy][Energy:2]]
 template <typename F>
-double getEnergyDistinct
+__MAYBE_GLOBAL__
+void getEnergyDistinct
   ( F const epsabc
   , size_t const No
   , F* const epsi
   , F* const Tijk
   , F* const Zijk
+  , double* energy
   ) {
   constexpr size_t blockSize=16;
-  F energy(0.);
+  F _energy = {0.};
   for (size_t kk=0; kk<No; kk+=blockSize){
-    const size_t kend( std::min(No, kk+blockSize) );
+    const size_t kend( MIN(No, kk+blockSize) );
     for (size_t jj(kk); jj<No; jj+=blockSize){
-      const size_t jend( std::min( No, jj+blockSize) );
+      const size_t jend( MIN( No, jj+blockSize) );
       for (size_t ii(jj); ii<No; ii+=blockSize){
-        const size_t iend( std::min( No, ii+blockSize) );
+        const size_t iend( MIN( No, ii+blockSize) );
         for (size_t k(kk); k < kend; k++){
           const F ek(epsi[k]);
           const size_t jstart = jj > k ? jj : k;
           for (size_t j(jstart); j < jend; j++){
             F const ej(epsi[j]);
-            F const facjk = j == k ? F(0.5) : F(1.0);
+            F const facjk = j == k ? F{0.5} : F{1.0};
             size_t istart = ii > j ? ii : j;
             for (size_t i(istart); i < iend; i++){
               const F
                   ei(epsi[i])
-                , facij = i == j ? F(0.5) : F(1.0)
+                , facij = i == j ? F{0.5} : F{1.0}
-                , denominator(epsabc - ei - ej - ek)
+                , eijk(acc::add(acc::add(ei, ej), ek))
+                , denominator(acc::sub(epsabc, eijk))
                 , U(Zijk[i + No*j + No*No*k])
                 , V(Zijk[i + No*k + No*No*j])
                 , W(Zijk[j + No*i + No*No*k])
                 , X(Zijk[j + No*k + No*No*i])
                 , Y(Zijk[k + No*i + No*No*j])
                 , Z(Zijk[k + No*j + No*No*i])
-                , A(maybeConjugate<F>(Tijk[i + No*j + No*No*k]))
+                , A(acc::maybeConjugateScalar(Tijk[i + No*j + No*No*k]))
-                , B(maybeConjugate<F>(Tijk[i + No*k + No*No*j]))
+                , B(acc::maybeConjugateScalar(Tijk[i + No*k + No*No*j]))
-                , C(maybeConjugate<F>(Tijk[j + No*i + No*No*k]))
+                , C(acc::maybeConjugateScalar(Tijk[j + No*i + No*No*k]))
-                , D(maybeConjugate<F>(Tijk[j + No*k + No*No*i]))
+                , D(acc::maybeConjugateScalar(Tijk[j + No*k + No*No*i]))
-                , E(maybeConjugate<F>(Tijk[k + No*i + No*No*j]))
+                , E(acc::maybeConjugateScalar(Tijk[k + No*i + No*No*j]))
-                , _F(maybeConjugate<F>(Tijk[k + No*j + No*No*i]))
+                , _F(acc::maybeConjugateScalar(Tijk[k + No*j + No*No*i]))
-                , value
+                , AU = acc::prod(A, U)
-                  = 3.0 * ( A * U
+                , BV = acc::prod(B, V)
-                            + B * V
+                , CW = acc::prod(C, W)
-                            + C * W
+                , DX = acc::prod(D, X)
-                            + D * X
+                , EY = acc::prod(E, Y)
-                            + E * Y
+                , FZ = acc::prod(_F, Z)
-                            + _F * Z )
+                , UXY = acc::add(U, acc::add(X, Y))
-                 + ( ( U + X + Y )
+                , VWZ = acc::add(V, acc::add(W, Z))
-                   - 2.0 * ( V + W + Z )
+                , ADE = acc::add(A, acc::add(D, E))
-                   ) * ( A + D + E )
+                , BCF = acc::add(B, acc::add(C, _F))
-                 + ( ( V + W + Z )
+                // I just might as well write this in CL
-                   - 2.0 * ( U + X + Y )
+                , _first =  acc::add(AU,
-                   ) * ( B + C + _F )
+                                     acc::add(BV,
+                                              acc::add(CW,
+                                                       acc::add(DX,
+                                                                acc::add(EY, FZ)))))
+                , _second = acc::prod(acc::sub(UXY,
+                                               acc::prod(F{-2.0}, VWZ)),
+                                      ADE)
+                , _third = acc::prod(acc::sub(VWZ,
+                                              acc::prod(F{-2.0}, UXY)),
+                                     BCF)
+                , value = acc::add(acc::prod(F{3.0}, _first),
+                                   acc::add(_second,
+                                            _third))
+                , _loop_energy = acc::prod(acc::prod(F{2.0}, value),
+                                           acc::div(acc::prod(facjk, facij),
+                                                    denominator))
                 ;
-              energy += 2.0 * value / denominator * facjk * facij;
+              acc::sum_in_place(&_energy, &_loop_energy);
             } // i
           } // j
         } // k
       } // ii
     } // jj
   } // kk
-  return std::real(energy);
+  const double real_part = acc::real(_energy);
+  acc::sum_in_place(energy, &real_part);
 }
 
 
 template <typename F>
-double getEnergySame
+__MAYBE_GLOBAL__
+void getEnergySame
   ( F const epsabc
   , size_t const No
   , F* const epsi
   , F* const Tijk
   , F* const Zijk
+  , double* energy
   ) {
   constexpr size_t blockSize = 16;
-  F energy = F(0.);
+  F _energy = F{0.};
   for (size_t kk=0; kk<No; kk+=blockSize){
-    const size_t kend( std::min( kk+blockSize, No) );
+    const size_t kend( MIN( kk+blockSize, No) );
     for (size_t jj(kk); jj<No; jj+=blockSize){
-      const size_t jend( std::min( jj+blockSize, No) );
+      const size_t jend( MIN( jj+blockSize, No) );
       for (size_t ii(jj); ii<No; ii+=blockSize){
-        const size_t iend( std::min( ii+blockSize, No) );
+        const size_t iend( MIN( ii+blockSize, No) );
         for (size_t k(kk); k < kend; k++){
           const F ek(epsi[k]);
           const size_t jstart = jj > k ? jj : k;
           for(size_t j(jstart); j < jend; j++){
-            const F facjk( j == k ? F(0.5) : F(1.0));
+            const F facjk( j == k ? F{0.5} : F{1.0});
             const F ej(epsi[j]);
             const size_t istart = ii > j ? ii : j;
             for(size_t i(istart); i < iend; i++){
               const F
                 ei(epsi[i])
-              , facij ( i==j ? F(0.5) : F(1.0))
+              , facij ( i==j ? F{0.5} : F{1.0})
-              , denominator(epsabc - ei - ej - ek)
+              , eijk(acc::add(acc::add(ei, ej), ek))
+              , denominator(acc::sub(epsabc, eijk))
               , U(Zijk[i + No*j + No*No*k])
               , V(Zijk[j + No*k + No*No*i])
               , W(Zijk[k + No*i + No*No*j])
-              , A(maybeConjugate<F>(Tijk[i + No*j + No*No*k]))
+              , A(acc::maybeConjugateScalar(Tijk[i + No*j + No*No*k]))
-              , B(maybeConjugate<F>(Tijk[j + No*k + No*No*i]))
+              , B(acc::maybeConjugateScalar(Tijk[j + No*k + No*No*i]))
-              , C(maybeConjugate<F>(Tijk[k + No*i + No*No*j]))
+              , C(acc::maybeConjugateScalar(Tijk[k + No*i + No*No*j]))
-              , value
+              , ABC = acc::add(A, acc::add(B, C))
-                = F(3.0) * ( A * U
+              , UVW = acc::add(U, acc::add(V, W))
-                           + B * V
+              , AU = acc::prod(A, U)
-                           + C * W
+              , BV = acc::prod(B, V)
-                           )
+              , CW = acc::prod(C, W)
-                - ( A + B + C ) * ( U + V + W )
+              , AU_and_BV_and_CW = acc::add(acc::add(AU, BV), CW)
+              , value = acc::sub(acc::prod(F{3.0}, AU_and_BV_and_CW),
+                                 acc::prod(ABC, UVW))
+              , _loop_energy = acc::prod(acc::prod(F{2.0}, value),
+                                         acc::div(acc::prod(facjk, facij),
+                                                  denominator))
               ;
-              energy += F(2.0) * value / denominator * facjk * facij;
+
+              acc::sum_in_place(&_energy, &_loop_energy);
             } // i
           } // j
         } // k
       } // ii
     } // jj
   } // kk
-  return std::real(energy);
+  const double real_part = acc::real(_energy);
+  acc::sum_in_place(energy, &real_part);
 }
 // Energy:2 ends here
 
 // [[file:~/cuda/atrip/atrip.org::*Energy][Energy:3]]
 // instantiate double
 template
-double getEnergyDistinct
+__MAYBE_GLOBAL__
-  ( double const epsabc
+void getEnergyDistinct
+  ( DataFieldType<double> const epsabc
   , size_t const No
-  , double* const epsi
+  , DataFieldType<double>* const epsi
-  , double* const Tijk
+  , DataFieldType<double>* const Tijk
-  , double* const Zijk
+  , DataFieldType<double>* const Zijk
+  , DataFieldType<double>* energy
   );
 
 template
-double getEnergySame
+__MAYBE_GLOBAL__
-  ( double const epsabc
+void getEnergySame
+  ( DataFieldType<double> const epsabc
   , size_t const No
-  , double* const epsi
+  , DataFieldType<double>* const epsi
-  , double* const Tijk
+  , DataFieldType<double>* const Tijk
-  , double* const Zijk
+  , DataFieldType<double>* const Zijk
+  , DataFieldType<double>* energy
   );
 
 // instantiate Complex
 template
-double getEnergyDistinct
+__MAYBE_GLOBAL__
-  ( Complex const epsabc
+void getEnergyDistinct
+  ( DataFieldType<Complex> const epsabc
   , size_t const No
-  , Complex* const epsi
+  , DataFieldType<Complex>* const epsi
-  , Complex* const Tijk
+  , DataFieldType<Complex>* const Tijk
-  , Complex* const Zijk
+  , DataFieldType<Complex>* const Zijk
+  , DataFieldType<double>* energy
   );
 
 template
-double getEnergySame
+__MAYBE_GLOBAL__
-  ( Complex const epsabc
+void getEnergySame
+  ( DataFieldType<Complex> const epsabc
   , size_t const No
-  , Complex* const epsi
+  , DataFieldType<Complex>* const epsi
-  , Complex* const Tijk
+  , DataFieldType<Complex>* const Tijk
-  , Complex* const Zijk
+  , DataFieldType<Complex>* const Zijk
+  , DataFieldType<double>* energy
   );
 // Energy:3 ends here
 
@@ -360,18 +320,26 @@ double getEnergySame
       const size_t ijk = i + j*No + k*NoNo;
 
 #ifdef HAVE_CUDA
+
 #define GO(__TPH, __VABIJ)                              \
-    {                                                                 \
+  do {                                                  \
-      const DataFieldType<F> product                                  \
+    const DataFieldType<F>                              \
-            = cuda::multiply<DataFieldType<F>>((__TPH), (__VABIJ));   \
+      product = acc::prod<DataFieldType<F>>((__TPH),    \
-      cuda::sum_in_place<DataFieldType<F>>(&Zijk[ijk], &product);     \
+                                            (__VABIJ)); \
-    }
+    acc::sum_in_place<DataFieldType<F>>(&Zijk[ijk],     \
+                                        &product);      \
+  } while (0)
+
 #else
-#  define GO(__TPH, __VABIJ) Zijk[ijk] += (__TPH) * (__VABIJ);
+
+#define GO(__TPH, __VABIJ) Zijk[ijk] += (__TPH) * (__VABIJ)
+
 #endif
-      GO(Tph[ a + i * Nv ], VBCij[ j + k * No ])
+
-      GO(Tph[ b + j * Nv ], VACij[ i + k * No ])
+      GO(Tph[ a + i * Nv ], VBCij[ j + k * No ]);
-      GO(Tph[ c + k * Nv ], VABij[ i + j * No ])
+      GO(Tph[ b + j * Nv ], VACij[ i + k * No ]);
+      GO(Tph[ c + k * Nv ], VABij[ i + j * No ]);
+
 #undef GO
     } // for loop j
   }
@@ -433,9 +401,15 @@ double getEnergySame
     // -- TIJK
     // , DataPtr<F> Tijk_
     , DataFieldType<F>* Tijk_
+#if defined(HAVE_CUDA)
+     // -- tmp buffers
+    , DataFieldType<F>* _t_buffer
+    , DataFieldType<F>* _vhhh 
+#endif
      ) {
-
+    const size_t a = abc[0], b = abc[1], c = abc[2]
-    const size_t NoNo = No*No;
+              , NoNo = No*No
+              ;
 
     DataFieldType<F>* Tijk = (DataFieldType<F>*)Tijk_;
 
@@ -480,7 +454,7 @@ double getEnergySame
                   )
 #define MAYBE_CONJ(_conj, _buffer)                                \
   do {                                                            \
-    cuda::maybeConjugate<<<                                       \
+    acc::maybeConjugate<<<                                        \
                                                                   \
                             Atrip::kernelDimensions.ooo.blocks,   \
                                                                   \
@@ -549,23 +523,23 @@ double getEnergySame
     F one{1.0}, m_one{-1.0}, zero{0.0};
     const size_t NoNoNo = No*NoNo;
 #ifdef HAVE_CUDA
-    DataFieldType<F>* _t_buffer;
+//    DataFieldType<F>* _t_buffer;
-    DataFieldType<F>* _vhhh;
+//    DataFieldType<F>* _vhhh;
-    WITH_CHRONO("double:cuda:alloc",
+//    WITH_CHRONO("double:cuda:alloc",
-    _CHECK_CUDA_SUCCESS("Allocating _t_buffer",
+//    _CHECK_CUDA_SUCCESS("Allocating _t_buffer",
-                        cuMemAlloc((CUdeviceptr*)&_t_buffer,
+//                        cuMemAlloc((CUdeviceptr*)&_t_buffer,
-                                   NoNoNo * sizeof(DataFieldType<F>)));
+//                                   NoNoNo * sizeof(DataFieldType<F>)));
-    _CHECK_CUDA_SUCCESS("Allocating _vhhh",
+//    _CHECK_CUDA_SUCCESS("Allocating _vhhh",
-                        cuMemAlloc((CUdeviceptr*)&_vhhh,
+//                        cuMemAlloc((CUdeviceptr*)&_vhhh,
-                                   NoNoNo * sizeof(DataFieldType<F>)));
+//                                   NoNoNo * sizeof(DataFieldType<F>)));
-                )
+//                )
+#if !defined(ATRIP_ONLY_DGEMM)
+    // we still have to zero this
     const size_t
       bs = Atrip::kernelDimensions.ooo.blocks,
       ths = Atrip::kernelDimensions.ooo.threads;
-
+    acc::zeroing<<<bs, ths>>>((DataFieldType<F>*)_t_buffer, NoNoNo);
-#if !defined(ATRIP_ONLY_DGEMM)
+    acc::zeroing<<<bs, ths>>>((DataFieldType<F>*)_vhhh, NoNoNo);
-    cuda::zeroing<<<bs, ths>>>((DataFieldType<F>*)_t_buffer, NoNoNo);
-    cuda::zeroing<<<bs, ths>>>((DataFieldType<F>*)_vhhh, NoNoNo);
 #endif
 
 #else
@@ -581,15 +555,17 @@ double getEnergySame
     // Set Tijk to zero
 #if defined(HAVE_CUDA) && !defined(ATRIP_ONLY_DGEMM)
     WITH_CHRONO("double:reorder",
-                cuda::zeroing<<<bs, ths>>>((DataFieldType<F>*)Tijk,
+                acc::zeroing<<<bs, ths>>>((DataFieldType<F>*)Tijk,
                                            NoNoNo);
                 )
-#else
+#endif
+
+#if !defined(HAVE_CUDA)
     WITH_CHRONO("double:reorder",
       for (size_t k = 0; k < NoNoNo; k++) {
         Tijk[k] = DataFieldType<F>{0.0};
        })
-#endif
+#endif /* !defined(HAVE_CUDA) */
 
 
 #if defined(ATRIP_ONLY_DGEMM)
@@ -597,7 +573,7 @@ double getEnergySame
 #undef REORDER
 #define MAYBE_CONJ(a, b) do {} while(0)
 #define REORDER(i, j, k) do {} while(0)
-#endif
+#endif /* defined(ATRIP_ONLY_DGEMM) */
 
     // HOLES
     WITH_CHRONO("doubles:holes",
@@ -681,16 +657,16 @@ double getEnergySame
 #ifdef HAVE_CUDA
     // we need to synchronize here since we need
     // the Tijk for next process in the pipeline
-    _CHECK_CUDA_SUCCESS("Synchronizing",
+    //_CHECK_CUDA_SUCCESS("Synchronizing",
-                        cuCtxSynchronize());
+    //                    cuCtxSynchronize());
-    _CHECK_CUDA_SUCCESS("Freeing _vhhh",
+    //_CHECK_CUDA_SUCCESS("Freeing _vhhh",
-                        cuMemFree((CUdeviceptr)_vhhh));
+    //                    cuMemFree((CUdeviceptr)_vhhh));
-    _CHECK_CUDA_SUCCESS("Freeing _t_buffer",
+    //_CHECK_CUDA_SUCCESS("Freeing _t_buffer",
-                        cuMemFree((CUdeviceptr)_t_buffer));
+    //                    cuMemFree((CUdeviceptr)_t_buffer));
 #else
     free(_vhhh);
     free(_t_buffer);
-#endif
+#endif  /* defined(HAVE_CUDA) */
   }
 
   #undef REORDER
@@ -741,7 +717,7 @@ double getEnergySame
       }
 
     }
-#endif
+#endif /* defined(ATRIP_USE_DGEMM) */
   }
 
 
@@ -773,6 +749,12 @@ double getEnergySame
     , DataPtr<double> const TBChh
     // -- TIJK
     , DataFieldType<double>* Tijk
+#if defined(HAVE_CUDA)
+     // -- tmp buffers
+    , DataFieldType<double>* _t_buffer
+    , DataFieldType<double>* _vhhh 
+#endif
+
     );
 
   template
@@ -801,6 +783,12 @@ double getEnergySame
     , DataPtr<Complex> const TBChh
     // -- TIJK
     , DataFieldType<Complex>* Tijk
+#if defined(HAVE_CUDA)
+     // -- tmp buffers
+    , DataFieldType<Complex>* _t_buffer
+    , DataFieldType<Complex>* _vhhh 
+#endif
+
     );
 // Doubles contribution:2 ends here
 

tools/configure-benches.sh

@@ -98,10 +98,27 @@ EOF
 create_config $tmp only-dgemm
 rm $tmp
 
-#
 # begin doc
 #
-# - slices-on-gpu-only-dgemm ::
+# - cuda-only-dgemm ::
+#   This is the naive CUDA implementation compiling only the dgemm parts
+#   of the compute.
+#
+# end doc
+
+tmp=`mktemp`
+cat <<EOF > $tmp
+--enable-cuda
+--enable-only-dgemm
+--disable-slice
+EOF
+
+create_config $tmp cuda-only-dgemm
+rm $tmp
+
+# begin doc
+#
+# - cuda-slices-on-gpu-only-dgemm ::
 #   This configuration tests that slices reside completely on the gpu
 #   and it should use a CUDA aware MPI implementation.
 #   It also only uses the routines that involve dgemm.
@@ -117,7 +134,7 @@ cat <<EOF > $tmp
 --disable-slice
 EOF
 
-create_config $tmp sources-in-gpu
+create_config $tmp cuda-slices-on-gpu-only-dgemm
 rm $tmp
 
 ############################################################