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__
-void getEnergyDistinct
+double 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__
-void getEnergySame
+double getEnergySame
   ( F const epsabc
   , size_t const No
   , F* const epsi
   , F* const Tijk
   , F* const Zijk
-  , double* energy
   );
 // Energy: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/SliceUnion.hpp

@@ -445,34 +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";

src/atrip/Atrip.cxx

@@ -683,59 +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;
-      cuMemAlloc((DataPtr<double>*)&tupleEnergy, sizeof(double));
-#elif
-      double _tupleEnergy(0.);
-      double *tupleEnergy = &_tupleEnergy;
-#endif /* defined(HAVE_CUDA) */
+      double tupleEnergy(0.);
 
       int distinct(0);
       if (abc[0] == abc[1]) distinct++;
       if (abc[1] == abc[2]) distinct--;
-      const double
-        epsabc = std::real(_epsa[abc[0]] + _epsa[abc[1]] + _epsa[abc[2]]);
-
-      DataFieldType<F> _epsabc{epsabc};
+      const F epsabc(_epsa[abc[0]] + _epsa[abc[1]] + _epsa[abc[2]]);
 
+      // LOG(0, "AtripCUDA") <<  "doing energy " << i << "distinct " << distinct << "\n";
       WITH_CHRONO("energy",
-                  if ( distinct == 0) {
-                    ACC_FUNCALL(getEnergyDistinct<DataFieldType<F>>,
-                                1, 1, // for cuda
-                                _epsabc,
-                                No,
-                                (DataFieldType<F>*)epsi,
-                                (DataFieldType<F>*)Tijk,
-                                (DataFieldType<F>*)Zijk,
-                                tupleEnergy);
-                  } else {
-                    ACC_FUNCALL(getEnergySame<DataFieldType<F>>,
-                                1, 1, // for cuda
-                                _epsabc,
-                                No,
-                                (DataFieldType<F>*)epsi,
-                                (DataFieldType<F>*)Tijk,
-                                (DataFieldType<F>*)Zijk,
-                                tupleEnergy);
-                  })
-
-#if defined(HAVE_CUDA)
-        double host_tuple_energy;
-        cuMemcpyDtoH((void*)&host_tuple_energy,
-                     (DataPtr<double>)tupleEnergy,
-                     sizeof(double));
-#elif
-        double host_tuple_energy = *tupleEnergy;
-#endif /* defined(HAVE_CUDA) */
+/*
+    TODO: think about how to do this on the GPU in the best way possible
+        if ( distinct == 0)
+          tupleEnergy = getEnergyDistinct<F>(epsabc, No, (F*)epsi, (F*)Tijk, (F*)Zijk);
+        else
+          tupleEnergy = getEnergySame<F>(epsabc, No, (F*)epsi, (F*)Tijk, (F*)Zijk);
+*/
+      )
 
 #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__
-void getEnergyDistinct
+double 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}
-                , eijk(acc::add(acc::add(ei, ej), ek))
-                , denominator(acc::sub(epsabc, eijk))
+                , facij = i == j ? F(0.5) : F(1.0)
+                , denominator(epsabc - ei - ej - ek)
                 , 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]))
-                , B(acc::maybeConjugateScalar(Tijk[i + No*k + No*No*j]))
-                , C(acc::maybeConjugateScalar(Tijk[j + No*i + No*No*k]))
-                , D(acc::maybeConjugateScalar(Tijk[j + No*k + No*No*i]))
-                , E(acc::maybeConjugateScalar(Tijk[k + No*i + No*No*j]))
-                , _F(acc::maybeConjugateScalar(Tijk[k + No*j + No*No*i]))
-                , AU = acc::prod(A, U)
-                , BV = acc::prod(B, V)
-                , CW = acc::prod(C, W)
-                , DX = acc::prod(D, X)
-                , EY = acc::prod(E, Y)
-                , FZ = acc::prod(_F, Z)
-                , UXY = acc::add(U, acc::add(X, Y))
-                , VWZ = acc::add(V, acc::add(W, Z))
-                , ADE = acc::add(A, acc::add(D, E))
-                , BCF = acc::add(B, acc::add(C, _F))
-                // I just might as well write this in CL
-                , _first =  acc::add(AU,
-                                     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))
+                , A(maybeConjugate<F>(Tijk[i + No*j + No*No*k]))
+                , B(maybeConjugate<F>(Tijk[i + No*k + No*No*j]))
+                , C(maybeConjugate<F>(Tijk[j + No*i + No*No*k]))
+                , D(maybeConjugate<F>(Tijk[j + No*k + No*No*i]))
+                , E(maybeConjugate<F>(Tijk[k + No*i + No*No*j]))
+                , _F(maybeConjugate<F>(Tijk[k + No*j + No*No*i]))
+                , value
+                  = 3.0 * ( A * U
+                            + B * V
+                            + C * W
+                            + D * X
+                            + E * Y
+                            + _F * Z )
+                 + ( ( U + X + Y )
+                   - 2.0 * ( V + W + Z )
+                   ) * ( A + D + E )
+                 + ( ( V + W + Z )
+                   - 2.0 * ( U + X + Y )
+                   ) * ( B + C + _F )
                 ;
-              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);
-  acc::sum_in_place(energy, &real_part);
+  return std::real(energy);
 }
 
 
 template <typename F>
-__MAYBE_GLOBAL__
-void getEnergySame
+double 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})
-              , eijk(acc::add(acc::add(ei, ej), ek))
-              , denominator(acc::sub(epsabc, eijk))
+              , facij ( i==j ? F(0.5) : F(1.0))
+              , denominator(epsabc - ei - ej - ek)
               , 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]))
-              , B(acc::maybeConjugateScalar(Tijk[j + No*k + No*No*i]))
-              , C(acc::maybeConjugateScalar(Tijk[k + No*i + No*No*j]))
-              , ABC = acc::add(A, acc::add(B, C))
-              , UVW = acc::add(U, acc::add(V, W))
-              , AU = acc::prod(A, U)
-              , BV = acc::prod(B, V)
-              , CW = acc::prod(C, 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))
+              , A(maybeConjugate<F>(Tijk[i + No*j + No*No*k]))
+              , B(maybeConjugate<F>(Tijk[j + No*k + No*No*i]))
+              , C(maybeConjugate<F>(Tijk[k + No*i + No*No*j]))
+              , value
+                = F(3.0) * ( A * U
+                           + B * V
+                           + C * W
+                           )
+                - ( A + B + C ) * ( U + V + W )
               ;
-
-              acc::sum_in_place(&_energy, &_loop_energy);
+              energy += F(2.0) * value / denominator * facjk * facij;
             } // i
           } // j
         } // k
       } // ii
     } // jj
   } // kk
-  const double real_part = acc::real(_energy);
-  acc::sum_in_place(energy, &real_part);
+  return std::real(energy);
 }
 // Energy:2 ends here
 
 // [[file:~/cuda/atrip/atrip.org::*Energy][Energy:3]]
 // instantiate double
 template
-__MAYBE_GLOBAL__
-void getEnergyDistinct
-  ( DataFieldType<double> const epsabc
+double getEnergyDistinct
+  ( double const epsabc
   , size_t const No
-  , DataFieldType<double>* const epsi
-  , DataFieldType<double>* const Tijk
-  , DataFieldType<double>* const Zijk
-  , DataFieldType<double>* energy
+  , double* const epsi
+  , double* const Tijk
+  , double* const Zijk
   );
 
 template
-__MAYBE_GLOBAL__
-void getEnergySame
-  ( DataFieldType<double> const epsabc
+double getEnergySame
+  ( double const epsabc
   , size_t const No
-  , DataFieldType<double>* const epsi
-  , DataFieldType<double>* const Tijk
-  , DataFieldType<double>* const Zijk
-  , DataFieldType<double>* energy
+  , double* const epsi
+  , double* const Tijk
+  , double* const Zijk
   );
 
 // instantiate Complex
 template
-__MAYBE_GLOBAL__
-void getEnergyDistinct
-  ( DataFieldType<Complex> const epsabc
+double getEnergyDistinct
+  ( Complex const epsabc
   , size_t const No
-  , DataFieldType<Complex>* const epsi
-  , DataFieldType<Complex>* const Tijk
-  , DataFieldType<Complex>* const Zijk
-  , DataFieldType<double>* energy
+  , Complex* const epsi
+  , Complex* const Tijk
+  , Complex* const Zijk
   );
 
 template
-__MAYBE_GLOBAL__
-void getEnergySame
-  ( DataFieldType<Complex> const epsabc
+double getEnergySame
+  ( Complex const epsabc
   , size_t const No
-  , DataFieldType<Complex>* const epsi
-  , DataFieldType<Complex>* const Tijk
-  , DataFieldType<Complex>* const Zijk
-  , DataFieldType<double>* energy
+  , Complex* const epsi
+  , Complex* const Tijk
+  , Complex* const Zijk
   );
 // 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)                              \
-  do {                                                  \
-    const DataFieldType<F>                              \
-      product = acc::prod<DataFieldType<F>>((__TPH),    \
-                                            (__VABIJ)); \
-    acc::sum_in_place<DataFieldType<F>>(&Zijk[ijk],     \
-                                        &product);      \
-  } while (0)
-
+#  define GO(__TPH, __VABIJ)                                          \
+    {                                                                 \
+      const DataFieldType<F> product                                  \
+            = cuda::multiply<DataFieldType<F>>((__TPH), (__VABIJ));   \
+      cuda::sum_in_place<DataFieldType<F>>(&Zijk[ijk], &product);     \
+    }
 #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[ c + k * Nv ], VABij[ i + j * No ]);
-
+      GO(Tph[ a + i * Nv ], VBCij[ j + k * No ])
+      GO(Tph[ b + j * Nv ], VACij[ i + k * No ])
+      GO(Tph[ c + k * Nv ], VABij[ i + j * No ])
 #undef GO
     } // for loop j
   }
@@ -448,7 +480,7 @@ void getEnergySame
                   )
 #define MAYBE_CONJ(_conj, _buffer)                                \
   do {                                                            \
-    acc::maybeConjugate<<<                                        \
+    cuda::maybeConjugate<<<                                       \
                                                                   \
                             Atrip::kernelDimensions.ooo.blocks,   \
                                                                   \
@@ -532,8 +564,8 @@ void getEnergySame
       ths = Atrip::kernelDimensions.ooo.threads;
 
 #if !defined(ATRIP_ONLY_DGEMM)
-    acc::zeroing<<<bs, ths>>>((DataFieldType<F>*)_t_buffer, NoNoNo);
-    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
@@ -549,7 +581,7 @@ 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);
                 )
 #else
@@ -557,7 +589,7 @@ void getEnergySame
       for (size_t k = 0; k < NoNoNo; k++) {
         Tijk[k] = DataFieldType<F>{0.0};
        })
-#endif /* defined(HAVE_CUDA) && !defined(ATRIP_ONLY_DGEMM) */
+#endif
 
 
 #if defined(ATRIP_ONLY_DGEMM)
@@ -565,7 +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",
@@ -658,7 +690,7 @@ void getEnergySame
 #else
     free(_vhhh);
     free(_t_buffer);
-#endif  /* defined(HAVE_CUDA) */
+#endif
   }
 
   #undef REORDER
@@ -709,7 +741,7 @@ void getEnergySame
       }
 
     }
-#endif /* defined(ATRIP_USE_DGEMM) */
+#endif
   }