include/atrip/CUDA.hpp

@@ -11,22 +11,11 @@
 #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

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

@@ -1,171 +0,0 @@
-// 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) && !defined (ATRIP_SOURCES_IN_GPU)
+#if defined(HAVE_CUDA)
     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) && !defined(ATRIP_SOURCES_IN_GPU)
+#if defined(HAVE_CUDA)
       // 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) && !defined(ATRIP_SOURCES_IN_GPU)
+#if defined(HAVE_CUDA)
       , mpi_data(nullptr)
 #endif
       , size(size_)

include/atrip/SliceUnion.hpp

@@ -405,7 +405,6 @@ 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
@@ -418,23 +417,6 @@ 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)
@@ -463,34 +445,6 @@ 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";
@@ -573,11 +527,12 @@ 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) && defined(ATRIP_SOURCES_IN_GPU)
+#if defined(HAVE_CUDA)
-#  if !defined(ATRIP_CUDA_AWARE_MPI) 
+#  if !defined(ATRIP_CUDA_AWARE_MPI) && defined(ATRIP_SOURCES_IN_GPU)
 #    error "You need CUDA aware MPI to have slices on the GPU"
 #  endif
-        MPI_Irecv((void*)slice.data,
+        slice.mpi_data = (F*)malloc(sizeof(F) * slice.size);
+        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
-  DataPtr<F> Tai = _Tai.data(), epsi = _epsi.data(), epsa = _epsa.data();
+  std::vector<F> &Tai = _Tai, &epsi = _epsi, &epsa = _epsa;
   Zijk = (DataFieldType<F>*)malloc(No*No*No * sizeof(DataFieldType<F>));
   Tijk = (DataFieldType<F>*)malloc(No*No*No * sizeof(DataFieldType<F>));
 #endif
@@ -258,25 +258,6 @@ 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;
 
@@ -658,14 +639,7 @@ 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";
       ))
     }
@@ -693,7 +667,7 @@ Atrip::Output Atrip::run(Atrip::Input<F> const& in) {
                                       (DataFieldType<F>*)Tai,
 #else
       singlesContribution<F>(No, Nv, abc[0], abc[1], abc[2],
-                             Tai,
+                             Tai.data(),
 #endif
                              (DataFieldType<F>*)abhh.unwrapSlice(Slice<F>::AB,
                                                                  abc),
@@ -709,71 +683,31 @@ Atrip::Output Atrip::run(Atrip::Input<F> const& in) {
     // COMPUTE ENERGY %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% {{{1
 #if defined(ATRIP_ONLY_DGEMM)
     if (false)
-#endif /* defined(ATRIP_ONLY_DGEMM) */
+#endif
     if (!isFakeTuple(i)) {
-#if defined(HAVE_CUDA)
+      double tupleEnergy(0.);
-      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 double
+      const F epsabc(_epsa[abc[0]] + _epsa[abc[1]] + _epsa[abc[2]]);
-        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) {
+/*
-                    ACC_FUNCALL(getEnergyDistinct<DataFieldType<F>>,
+    TODO: think about how to do this on the GPU in the best way possible
-                                1, 1, // for cuda
+        if ( distinct == 0)
-                                _epsabc,
+          tupleEnergy = getEnergyDistinct<F>(epsabc, No, (F*)epsi, (F*)Tijk, (F*)Zijk);
-                                No,
+        else
-#if defined(HAVE_CUDA)
+          tupleEnergy = getEnergySame<F>(epsabc, No, (F*)epsi, (F*)Tijk, (F*)Zijk);
-                                (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] = host_tuple_energy;
+      tupleEnergies[abc] = tupleEnergy;
 #endif
 
-      energy += host_tuple_energy;
+      energy += tupleEnergy;
 
     }
 

src/atrip/Equations.cxx

@@ -16,13 +16,96 @@
 #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, \
@@ -50,7 +133,7 @@ namespace atrip {
     _REORDER_BODY_(__VA_ARGS__)                 \
       }
 #if defined(HAVE_CUDA)
-#define GO(__TO, __FROM) acc::sum_in_place<F>(&__TO, &__FROM);
+#define GO(__TO, __FROM) cuda::sum_in_place<F>(&__TO, &__FROM);
 #else
 #define GO(__TO, __FROM) __TO += __FROM;
 #endif
@@ -96,205 +179,162 @@ namespace atrip {
 #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>
-__MAYBE_GLOBAL__
+double getEnergyDistinct
-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( MIN(No, kk+blockSize) );
+    const size_t kend( std::min(No, kk+blockSize) );
     for (size_t jj(kk); jj<No; jj+=blockSize){
-      const size_t jend( MIN( No, jj+blockSize) );
+      const size_t jend( std::min( No, jj+blockSize) );
       for (size_t ii(jj); ii<No; ii+=blockSize){
-        const size_t iend( MIN( No, ii+blockSize) );
+        const size_t iend( std::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)
-                , eijk(acc::add(acc::add(ei, ej), ek))
+                , denominator(epsabc - 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(acc::maybeConjugateScalar(Tijk[i + No*j + No*No*k]))
+                , A(maybeConjugate<F>(Tijk[i + No*j + No*No*k]))
-                , B(acc::maybeConjugateScalar(Tijk[i + No*k + No*No*j]))
+                , B(maybeConjugate<F>(Tijk[i + No*k + No*No*j]))
-                , C(acc::maybeConjugateScalar(Tijk[j + No*i + No*No*k]))
+                , C(maybeConjugate<F>(Tijk[j + No*i + No*No*k]))
-                , D(acc::maybeConjugateScalar(Tijk[j + No*k + No*No*i]))
+                , D(maybeConjugate<F>(Tijk[j + No*k + No*No*i]))
-                , E(acc::maybeConjugateScalar(Tijk[k + No*i + No*No*j]))
+                , E(maybeConjugate<F>(Tijk[k + No*i + No*No*j]))
-                , _F(acc::maybeConjugateScalar(Tijk[k + No*j + No*No*i]))
+                , _F(maybeConjugate<F>(Tijk[k + No*j + No*No*i]))
-                , AU = acc::prod(A, U)
+                , value
-                , BV = acc::prod(B, V)
+                  = 3.0 * ( A * U
-                , CW = acc::prod(C, W)
+                            + B * V
-                , DX = acc::prod(D, X)
+                            + C * W
-                , EY = acc::prod(E, Y)
+                            + D * X
-                , FZ = acc::prod(_F, Z)
+                            + E * Y
-                , UXY = acc::add(U, acc::add(X, Y))
+                            + _F * Z )
-                , VWZ = acc::add(V, acc::add(W, Z))
+                 + ( ( U + X + Y )
-                , ADE = acc::add(A, acc::add(D, E))
+                   - 2.0 * ( V + W + Z )
-                , BCF = acc::add(B, acc::add(C, _F))
+                   ) * ( A + D + E )
-                // I just might as well write this in CL
+                 + ( ( V + W + Z )
-                , _first =  acc::add(AU,
+                   - 2.0 * ( U + X + Y )
-                                     acc::add(BV,
+                   ) * ( B + C + _F )
-                                              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))
                 ;
-              acc::sum_in_place(&_energy, &_loop_energy);
+              energy += 2.0 * value / denominator * facjk * facij;
             } // i
           } // j
         } // k
       } // ii
     } // jj
   } // kk
-  const double real_part = acc::real(_energy);
+  return std::real(energy);
-  acc::sum_in_place(energy, &real_part);
 }
 
 
 template <typename F>
-__MAYBE_GLOBAL__
+double getEnergySame
-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( MIN( kk+blockSize, No) );
+    const size_t kend( std::min( kk+blockSize, No) );
     for (size_t jj(kk); jj<No; jj+=blockSize){
-      const size_t jend( MIN( jj+blockSize, No) );
+      const size_t jend( std::min( jj+blockSize, No) );
       for (size_t ii(jj); ii<No; ii+=blockSize){
-        const size_t iend( MIN( ii+blockSize, No) );
+        const size_t iend( std::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))
-              , eijk(acc::add(acc::add(ei, ej), ek))
+              , denominator(epsabc - 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(acc::maybeConjugateScalar(Tijk[i + No*j + No*No*k]))
+              , A(maybeConjugate<F>(Tijk[i + No*j + No*No*k]))
-              , B(acc::maybeConjugateScalar(Tijk[j + No*k + No*No*i]))
+              , B(maybeConjugate<F>(Tijk[j + No*k + No*No*i]))
-              , C(acc::maybeConjugateScalar(Tijk[k + No*i + No*No*j]))
+              , C(maybeConjugate<F>(Tijk[k + No*i + No*No*j]))
-              , ABC = acc::add(A, acc::add(B, C))
+              , value
-              , UVW = acc::add(U, acc::add(V, W))
+                = F(3.0) * ( A * U
-              , AU = acc::prod(A, U)
+                           + B * V
-              , BV = acc::prod(B, V)
+                           + C * W
-              , CW = acc::prod(C, W)
+                           )
-              , AU_and_BV_and_CW = acc::add(acc::add(AU, BV), CW)
+                - ( A + B + C ) * ( U + V + W )
-              , 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
-  const double real_part = acc::real(_energy);
+  return std::real(energy);
-  acc::sum_in_place(energy, &real_part);
 }
 // Energy:2 ends here
 
 // [[file:~/cuda/atrip/atrip.org::*Energy][Energy:3]]
 // instantiate double
 template
-__MAYBE_GLOBAL__
+double getEnergyDistinct
-void getEnergyDistinct
+  ( double const epsabc
-  ( DataFieldType<double> const epsabc
   , size_t const No
-  , DataFieldType<double>* const epsi
+  , double* const epsi
-  , DataFieldType<double>* const Tijk
+  , double* const Tijk
-  , DataFieldType<double>* const Zijk
+  , double* const Zijk
-  , DataFieldType<double>* energy
   );
 
 template
-__MAYBE_GLOBAL__
+double getEnergySame
-void getEnergySame
+  ( double const epsabc
-  ( DataFieldType<double> const epsabc
   , size_t const No
-  , DataFieldType<double>* const epsi
+  , double* const epsi
-  , DataFieldType<double>* const Tijk
+  , double* const Tijk
-  , DataFieldType<double>* const Zijk
+  , double* const Zijk
-  , DataFieldType<double>* energy
   );
 
 // instantiate Complex
 template
-__MAYBE_GLOBAL__
+double getEnergyDistinct
-void getEnergyDistinct
+  ( Complex const epsabc
-  ( DataFieldType<Complex> const epsabc
   , size_t const No
-  , DataFieldType<Complex>* const epsi
+  , Complex* const epsi
-  , DataFieldType<Complex>* const Tijk
+  , Complex* const Tijk
-  , DataFieldType<Complex>* const Zijk
+  , Complex* const Zijk
-  , DataFieldType<double>* energy
   );
 
 template
-__MAYBE_GLOBAL__
+double getEnergySame
-void getEnergySame
+  ( Complex const epsabc
-  ( DataFieldType<Complex> const epsabc
   , size_t const No
-  , DataFieldType<Complex>* const epsi
+  , Complex* const epsi
-  , DataFieldType<Complex>* const Tijk
+  , Complex* const Tijk
-  , DataFieldType<Complex>* const Zijk
+  , Complex* const Zijk
-  , DataFieldType<double>* energy
   );
 // Energy:3 ends here
 
@@ -320,26 +360,18 @@ void getEnergySame
       const size_t ijk = i + j*No + k*NoNo;
 
 #ifdef HAVE_CUDA
-
+#  define GO(__TPH, __VABIJ)                                          \
-#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[ a + i * Nv ], VBCij[ j + k * No ]);
+      GO(Tph[ b + j * Nv ], VACij[ i + k * No ])
-      GO(Tph[ b + j * Nv ], VACij[ i + k * No ]);
+      GO(Tph[ c + k * Nv ], VABij[ i + j * No ])
-      GO(Tph[ c + k * Nv ], VABij[ i + j * No ]);
-
 #undef GO
     } // for loop j
   }
@@ -401,15 +433,9 @@ void getEnergySame
     // -- TIJK
     // , DataPtr<F> Tijk_
     , DataFieldType<F>* Tijk_
-#if defined(HAVE_CUDA)
+    ) {
-     // -- tmp buffers
+
-    , DataFieldType<F>* _t_buffer
+    const size_t NoNo = No*No;
-    , DataFieldType<F>* _vhhh 
-#endif
-     ) {
-    const size_t a = abc[0], b = abc[1], c = abc[2]
-              , NoNo = No*No
-              ;
 
     DataFieldType<F>* Tijk = (DataFieldType<F>*)Tijk_;
 
@@ -454,7 +480,7 @@ void getEnergySame
                   )
 #define MAYBE_CONJ(_conj, _buffer)                                \
   do {                                                            \
-    acc::maybeConjugate<<<                                        \
+    cuda::maybeConjugate<<<                                       \
                                                                   \
                             Atrip::kernelDimensions.ooo.blocks,   \
                                                                   \
@@ -523,23 +549,23 @@ void 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);
+
-    acc::zeroing<<<bs, ths>>>((DataFieldType<F>*)_vhhh, NoNoNo);
+#if !defined(ATRIP_ONLY_DGEMM)
+    cuda::zeroing<<<bs, ths>>>((DataFieldType<F>*)_t_buffer, NoNoNo);
+    cuda::zeroing<<<bs, ths>>>((DataFieldType<F>*)_vhhh, NoNoNo);
 #endif
 
 #else
@@ -555,17 +581,15 @@ void getEnergySame
     // Set Tijk to zero
 #if defined(HAVE_CUDA) && !defined(ATRIP_ONLY_DGEMM)
     WITH_CHRONO("double:reorder",
-                acc::zeroing<<<bs, ths>>>((DataFieldType<F>*)Tijk,
+                cuda::zeroing<<<bs, ths>>>((DataFieldType<F>*)Tijk,
                                            NoNoNo);
                 )
-#endif
+#else
-
-#if !defined(HAVE_CUDA)
     WITH_CHRONO("double:reorder",
       for (size_t k = 0; k < NoNoNo; k++) {
         Tijk[k] = DataFieldType<F>{0.0};
        })
-#endif /* !defined(HAVE_CUDA) */
+#endif
 
 
 #if defined(ATRIP_ONLY_DGEMM)
@@ -573,7 +597,7 @@ void getEnergySame
 #undef REORDER
 #define MAYBE_CONJ(a, b) do {} while(0)
 #define REORDER(i, j, k) do {} while(0)
-#endif /* defined(ATRIP_ONLY_DGEMM) */
+#endif
 
     // HOLES
     WITH_CHRONO("doubles:holes",
@@ -657,16 +681,16 @@ void 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  /* defined(HAVE_CUDA) */
+#endif
   }
 
   #undef REORDER
@@ -717,7 +741,7 @@ void getEnergySame
       }
 
     }
-#endif /* defined(ATRIP_USE_DGEMM) */
+#endif
   }
 
 
@@ -749,12 +773,6 @@ void getEnergySame
     , DataPtr<double> const TBChh
     // -- TIJK
     , DataFieldType<double>* Tijk
-#if defined(HAVE_CUDA)
-     // -- tmp buffers
-    , DataFieldType<double>* _t_buffer
-    , DataFieldType<double>* _vhhh 
-#endif
-
     );
 
   template
@@ -783,12 +801,6 @@ void 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