Fix zeroing

.dir-locals.el

@@ -0,0 +1,22 @@
+;;; Directory Local Variables
+;;; For more information see (info "(emacs) Directory Variables")
+
+((c++-mode . ((outline-regexp . "// \\[\\[file:")
+              (eval . (let
+                          ((root
+                            (expand-file-name
+                             (project-root
+                              (project-current)))))
+                        (setq-local flycheck-gcc-include-path
+                                    (list
+                                     (format "%s/vendor/include/" root)
+                                     (format "%s/include/" root)
+                                     (format "%s/" root)
+                                     (format "%s/bench/" root)
+                                     (format "%s/build/main/" root)))
+                        (setq-local flycheck-clang-include-path
+                                    flycheck-gcc-include-path)))
+              (eval . (flycheck-mode))
+              (eval . (outline-minor-mode))
+              (indent-tabs-mode . nil)
+              (tab-width . 2))))

.gitignore

@@ -25,3 +25,6 @@ config.mk
 /atrip.html
 /TAGS
 /config.h.in
+/result
+/result-dev
+/vendor/

README.org

@@ -26,3 +26,110 @@ before the proper paper is released please contact me.
 
 In the mean time the code has been used in
 [[https://aip.scitation.org/doi/10.1063/5.0074936][this publication]] and can therefore been cited.
+
+* Building
+
+Atrip uses autotools to build the system.
+Autotools works by first creating a =configure= script from
+a =configure.ac= file.
+
+Atrip should be built out of source, this means that
+you have to create a build directory other that the root
+directory, for instance in the =build/tutorial= directory
+
+#+begin_src sh :exports code 
+mkdir -p build/tutorial/
+cd build/tutorial
+#+end_src
+
+First you have to build the =configure= script by doing
+
+#+begin_src sh :dir build/tutorial :exports code :results raw drawer
+../../bootstrap.sh
+#+end_src
+
+#+RESULTS:
+:results:
+
+  Creating configure script
+
+
+  Now you can build by doing
+
+  mkdir build
+  cd build
+  ../configure
+  make extern
+  make all
+
+:end:
+
+And then you can see the =configure= options
+#+begin_src sh :dir build/tutorial :results raw drawer :eval no 
+../../configure --help
+#+end_src
+
+** Benches
+
+The script =tools/configure-benches.sh= can be used to create
+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 $NL }' tools/configure-benches.sh |
+grep -v -e "begin \+doc" -e "end \+doc" |
+sed "s/^# //; s/^# *$//; /^$/d"
+#+end_src
+
+#+RESULTS:
+:results:
+- default ::
+  This configuration uses a CPU code with dgemm
+  and without computing slices.
+- only-dgemm ::
+  This only runs the computation part that involves dgemms.
+- 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/Makefile.am

@@ -1,25 +1,42 @@
 AUTOMAKE_OPTIONS = subdir-objects
 include $(top_srcdir)/atrip.mk
 
-AM_CPPFLAGS = -I$(top_srcdir)/include/ $(CTF_CPPFLAGS)
+AM_CPPFLAGS = -I$(top_srcdir)/include/ -I$(top_srcdir) $(CTF_CPPFLAGS)
 AM_LDFLAGS =   @LAPACK_LIBS@ @BLAS_LIBS@
 
-bin_PROGRAMS = test_main
-test_main_SOURCES = test_main.cxx
-
-
-test_main_LDADD  = \
-	$(top_builddir)/src/libatrip.a
-
 if WITH_BUILD_CTF
-test_main_LDADD += $(CTF_BUILD_PATH)/lib/libctf.a
+ATRIP_CTF = $(CTF_BUILD_PATH)/lib/libctf.a
 else
-test_main_LDADD += @LIBCTF_LD_LIBRARY_PATH@/libctf.a
+ATRIP_CTF = @LIBCTF_LD_LIBRARY_PATH@/libctf.a
+endif
+ATRIP_LIB = $(top_builddir)/src/libatrip.a $(ATRIP_CTF)
+
+bin_PROGRAMS =
+BENCHES_LDADD = $(ATRIP_LIB) $(ATRIP_CTF)
+
+
+##
+## main entry point and bench
+##
+bin_PROGRAMS += atrip
+atrip_SOURCES = main.cxx
+atrip_CPPFLAGS = $(AM_CPPFLAGS)
+atrip_LDADD = $(BENCHES_LDADD)
+
+
+if !WITH_CUDA
+##
+## tuples distribution
+##
+bin_PROGRAMS += tuples-distribution
+tuples_distribution_LDADD  = $(BENCHES_LDADD)
+tuples_distribution_SOURCES = tuples-distribution.cxx
 endif
 
+
 if WITH_CUDA
-test_main_CXXFLAGS = $(CUDA_CXXFLAGS)
-test_main_LDADD += $(CUDA_LDFLAGS)
+AM_CPPFLAGS += $(CUDA_CXXFLAGS)
+BENCHES_LDADD += $(CUDA_LDFLAGS)
 
 AM_CXXFLAGS = $(CUDA_CXXFLAGS)
 AM_LDFLAGS += $(CUDA_LDFLAGS)

bench/main.cxx

@@ -5,18 +5,20 @@
 #include <CLI11.hpp>
 
 #define _print_size(what, size)                 \
-  if (rank == 0) {                              \
-    std::cout << #what                          \
-              << " => "                         \
-              << (double)size * elem_to_gb      \
-              << "GB"                           \
-              << std::endl;                     \
-  }
+  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");
@@ -45,14 +50,27 @@ 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",
-		 ooo_blocks,
-		 "CUDA: Number of blocks per block for kernels going through ooo tensors");
+                 ooo_blocks,
+                 "CUDA: Number of blocks per block for kernels going through ooo tensors");
   app.add_option("--ooo-threads",
-		 ooo_threads,
-		 "CUDA: Number of threads per block for kernels going through ooo tensors");
+                 ooo_threads,
+                 "CUDA: Number of threads per block for kernels going through ooo tensors");
 #endif
 
   CLI11_PARSE(app, argc, argv);
@@ -148,37 +166,64 @@ int main(int argc, char** argv) {
   }
 
 
-  std::vector<int> symmetries(4, NS)
-                 , vo({nv, no})
-                 , vvoo({nv, nv, no, no})
-                 , ooov({no, no, no, nv})
-                 , vvvo({nv, nv, nv, no})
-                 ;
+  std::vector<int>
+    symmetries(4, NS),
+    vo({nv, no}),
+    vvoo({nv, nv, no, no}),
+    ooov({no, no, no, nv}),
+    vvvo({nv, nv, nv, no});
 
   CTF::Tensor<double>
-      ei(1, ooov.data(), symmetries.data(), world)
-    , ea(1, vo.data(), symmetries.data(), world)
-    , Tph(2, vo.data(), symmetries.data(), world)
-    , Tpphh(4, vvoo.data(), symmetries.data(), world)
-    , Vpphh(4, vvoo.data(), symmetries.data(), world)
-    , Vhhhp(4, ooov.data(), symmetries.data(), world)
-    ;
+    ei(1, ooov.data(), symmetries.data(), world),
+    ea(1, vo.data(), symmetries.data(), world),
+    Tph(2, vo.data(), symmetries.data(), world),
+    Tpphh(4, vvoo.data(), symmetries.data(), world),
+    Vpphh(4, vvoo.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(Vabij, no*no*nv*nv)
-  _print_size(Vijka, no*no*no*nv)
+  _print_size(Vabci, no*nv*nv*nv);
+  _print_size(Vabij, no*no*nv*nv);
+  _print_size(Vijka, no*no*no*nv);
 
-  ei.fill_random(-40.0, -2);
-  ea.fill_random(2, 50);
-  Tpphh.fill_random(0, 1);
-  Tph.fill_random(0, 1);
-  Vpphh.fill_random(0, 1);
-  Vhhhp.fill_random(0, 1);
-  Vppph->fill_random(0, 1);
+  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)

bench/tuples-distribution.cxx

@@ -0,0 +1,443 @@
+#include <iostream>
+#define ATRIP_DEBUG 2
+#include <atrip/Atrip.hpp>
+#include <atrip/Tuples.hpp>
+#include <atrip/Unions.hpp>
+#include <bench/CLI11.hpp>
+#include <bench/utils.hpp>
+
+using namespace atrip;
+
+using F = double;
+using Tr = CTF::Tensor<F>;
+
+#define INIT_DRY(name, ...)                                       \
+  do {                                                            \
+    std::vector<int64_t> lens = __VA_ARGS__;                      \
+    int i = -1;                                                   \
+    name.order = lens.size();                                     \
+    name.lens = (int64_t*)malloc(sizeof(int64_t) * lens.size());  \
+    name.sym = (int*)malloc(sizeof(int) * lens.size());           \
+    name.lens[++i] = lens[i]; name.lens[++i] = lens[i];           \
+    name.lens[++i] = lens[i]; name.lens[++i] = lens[i];           \
+    i = 0;                                                        \
+    name.sym[i++] = NS; name.sym[i++] = NS;                       \
+    name.sym[i++] = NS; name.sym[i++] = NS;                       \
+  } while (0)
+
+#define DEINIT_DRY(name)                        \
+  do {                                          \
+    name.order = 0;                             \
+    name.lens = NULL;                           \
+    name.sym = NULL;                            \
+  } while (0)
+
+using LocalDatabase = typename Slice<F>::LocalDatabase;
+using LocalDatabaseElement = typename Slice<F>::LocalDatabaseElement;
+
+LocalDatabase buildLocalDatabase(SliceUnion<F> &u,
+                                 ABCTuple const& abc) {
+  LocalDatabase result;
+
+  auto const needed = u.neededSlices(abc);
+
+  // BUILD THE DATABASE
+  // we need to loop over all sliceTypes that this TensorUnion
+  // is representing and find out how we will get the corresponding
+  // slice for the abc we are considering right now.
+  for (auto const& pair: needed) {
+    auto const type = pair.first;
+    auto const tuple = pair.second;
+    auto const from  = u.rankMap.find(abc, type);
+
+    {
+      // FIRST: look up if there is already a *Ready* slice matching what we
+      // need
+      auto const& it
+        = std::find_if(u.slices.begin(), u.slices.end(),
+                       [&tuple, &type](Slice<F> const& other) {
+                         return other.info.tuple == tuple
+                           && other.info.type == type
+                           // we only want another slice when it
+                           // has already ready-to-use data
+                           && other.isUnwrappable()
+                           ;
+                       });
+      if (it != u.slices.end()) {
+        // if we find this slice, it means that we don't have to do anything
+        result.push_back({u.name, it->info});
+        continue;
+      }
+    }
+
+    //
+    // Try to find a recyling possibility ie. find a slice with the same
+    // tuple and that has a valid data pointer.
+    //
+    auto const& recycleIt
+      = std::find_if(u.slices.begin(), u.slices.end(),
+                     [&tuple, &type](Slice<F> const& other) {
+                       return other.info.tuple == tuple
+                         && other.info.type != type
+                         && other.isRecyclable()
+                         ;
+                     });
+
+    //
+    // if we find this recylce, then we find a Blank slice
+    // (which should exist by construction :THINK)
+    //
+    if (recycleIt != u.slices.end()) {
+      auto& blank = Slice<F>::findOneByType(u.slices, Slice<F>::Blank);
+      // TODO: formalize this through a method to copy information
+      //       from another slice
+      blank.data = recycleIt->data;
+      blank.info.type = type;
+      blank.info.tuple = tuple;
+      blank.info.state = Slice<F>::Recycled;
+      blank.info.from = from;
+      blank.info.recycling = recycleIt->info.type;
+      result.push_back({u.name, blank.info});
+      WITH_RANK << "__db__: RECYCLING: n" << u.name
+                << " " << pretty_print(abc)
+                << " get " << pretty_print(blank.info)
+                << " from " << pretty_print(recycleIt->info)
+                << " ptr " << recycleIt->data
+                << "\n"
+                ;
+      continue;
+    }
+
+    // in this case we have to create a new slice
+    // this means that we should have a blank slice at our disposal
+    // and also the freePointers should have some elements inside,
+    // so we pop a data pointer from the freePointers container
+    {
+      auto& blank = Slice<F>::findOneByType(u.slices, Slice<F>::Blank);
+      blank.info.type = type;
+      blank.info.tuple = tuple;
+      blank.info.from = from;
+
+      // Handle self sufficiency
+      blank.info.state = Atrip::rank == from.rank
+                        ? Slice<F>::SelfSufficient
+                        : Slice<F>::Fetch
+                        ;
+      if (blank.info.state == Slice<F>::SelfSufficient) {
+        blank.data = (F*)0xBADA55;
+      } else {
+        blank.data = (F*)0xA55A55;
+      }
+
+      result.push_back({u.name, blank.info});
+      continue;
+    }
+
+  }
+
+  return result;
+
+}
+
+void clearUnusedSlicesForNext(SliceUnion<F> &u,
+                              ABCTuple const& abc) {
+  auto const needed = u.neededSlices(abc);
+
+  // CLEAN UP SLICES, FREE THE ONES THAT ARE NOT NEEDED ANYMORE
+  for (auto& slice: u.slices) {
+    // if the slice is free, then it was not used anyways
+    if (slice.isFree()) continue;
+
+
+    // try to find the slice in the needed slices list
+    auto const found
+      = std::find_if(needed.begin(), needed.end(),
+                      [&slice] (typename Slice<F>::Ty_x_Tu const& tytu) {
+                        return slice.info.tuple == tytu.second
+                            && slice.info.type == tytu.first
+                            ;
+                      });
+
+    // if we did not find slice in needed, then erase it
+    if (found == needed.end()) {
+
+      // allow to gc unwrapped and recycled, never Fetch,
+      // if we have a Fetch slice then something has gone very wrong.
+      if (!slice.isUnwrapped() && slice.info.state != Slice<F>::Recycled)
+        throw
+          std::domain_error(_FORMAT("Trying to garbage collect (%d, %d) "
+                                    " a non-unwrapped slice! ",
+                                    slice.info.type,
+                                    slice.info.state));
+
+      // it can be that our slice is ready, but it has some hanging
+      // references lying around in the form of a recycled slice.
+      // Of course if we need the recycled slice the next iteration
+      // this would be fatal, because we would then free the pointer
+      // of the slice and at some point in the future we would
+      // overwrite it. Therefore, we must check if slice has some
+      // references in slices and if so then
+      //
+      //  - we should mark those references as the original (since the data
+      //    pointer should be the same)
+      //
+      //  - we should make sure that the data pointer of slice
+      //    does not get freed.
+      //
+      if (slice.info.state == Slice<F>::Ready) {
+        WITH_OCD WITH_RANK
+          << "__gc__:" << "checking for data recycled dependencies\n";
+        auto recycled
+          = Slice<F>::hasRecycledReferencingToIt(u.slices, slice.info);
+        if (recycled.size()) {
+          Slice<F>* newReady = recycled[0];
+          WITH_OCD WITH_RANK
+            << "__gc__:" << "swaping recycled "
+            << pretty_print(newReady->info)
+            << " and "
+            << pretty_print(slice.info)
+            << "\n";
+          newReady->markReady();
+
+          for (size_t i = 1; i < recycled.size(); i++) {
+            auto newRecyled = recycled[i];
+            newRecyled->info.recycling = newReady->info.type;
+            WITH_OCD WITH_RANK
+              << "__gc__:" << "updating recycled "
+              << pretty_print(newRecyled->info)
+              << "\n";
+          }
+
+        }
+      }
+
+      slice.free();
+    }  // we did not find the slice
+
+  }
+}
+
+
+void unwrapSlice(Slice<F>::Type t, ABCTuple abc, SliceUnion<F> *u) {
+  auto& slice = Slice<F>::findByTypeAbc(u->slices, t, abc);
+  switch  (slice.info.state) {
+  case Slice<F>::Dispatched:
+    slice.markReady();
+    break;
+  case Slice<F>::Recycled:
+    unwrapSlice(t, abc, u);
+    break;
+  }
+}
+
+#define PRINT_VARIABLE(v)                                         \
+  do {                                                            \
+    if (!rank) std::cout << "# " << #v << ": " << v << std::endl; \
+  } while (0)
+
+int main(int argc, char** argv) {
+  MPI_Init(&argc, &argv);
+
+  int no(10), nv(100);
+  std::string tuplesDistributionString = "naive";
+
+  CLI::App app{"Main bench for atrip"};
+  app.add_option("--no", no, "Occupied orbitals");
+  app.add_option("--nv", nv, "Virtual orbitals");
+  app.add_option("--dist", tuplesDistributionString, "Which distribution");
+  CLI11_PARSE(app, argc, argv);
+
+  CTF::World world(argc, argv);
+  auto kaun = world.comm;
+  int rank, np;
+  MPI_Comm_rank(kaun, &rank);
+  MPI_Comm_size(kaun, &np);
+  Atrip::init(world.comm);
+
+
+
+  atrip::ABCTuples tuplesList;
+  atrip::TuplesDistribution *dist;
+  {
+    using namespace atrip;
+    if (tuplesDistributionString == "naive") {
+      dist = new NaiveDistribution();
+      tuplesList = dist->getTuples(nv, world.comm);
+    } else if (tuplesDistributionString == "group") {
+      dist = new group_and_sort::Distribution();
+      tuplesList = dist->getTuples(nv, world.comm);
+    } else {
+      std::cout << "--dist should be either naive or group\n";
+      exit(1);
+    }
+  }
+
+  double tuplesListGb
+    = tuplesList.size() * sizeof(tuplesList[0])
+    / 1024.0 / 1024.0 / 1024.0;
+
+  std::cout << "\n";
+  PRINT_VARIABLE(tuplesDistributionString);
+  PRINT_VARIABLE(np);
+  PRINT_VARIABLE(no);
+  PRINT_VARIABLE(nv);
+  PRINT_VARIABLE(tuplesList.size());
+  PRINT_VARIABLE(tuplesListGb);
+
+  // create a fake dry tensor
+  Tr t_abph, t_abhh, t_tabhh, t_taphh, t_hhha;
+  INIT_DRY(t_abph  , {nv, nv, nv, no});
+  INIT_DRY(t_abhh  , {nv, nv, no, no});
+  INIT_DRY(t_tabhh , {nv, nv, no, no});
+  INIT_DRY(t_taphh , {nv, nv, no, no});
+  INIT_DRY(t_hhha  , {no, no, no, nv});
+
+  ABPH<F> abph(t_abph, (size_t)no, (size_t)nv, (size_t)np, kaun, kaun);
+  ABHH<F> abhh(t_abhh, (size_t)no, (size_t)nv, (size_t)np, kaun, kaun);
+  TABHH<F> tabhh(t_tabhh, (size_t)no, (size_t)nv, (size_t)np, kaun, kaun);
+  TAPHH<F> taphh(t_taphh, (size_t)no, (size_t)nv, (size_t)np, kaun, kaun);
+  HHHA<F>  hhha(t_hhha, (size_t)no, (size_t)nv, (size_t)np, kaun, kaun);
+  std::vector< SliceUnion<F>* > unions = {&taphh, &hhha, &abph, &abhh, &tabhh};
+
+
+
+  using Database = typename Slice<F>::Database;
+  auto communicateDatabase
+    = [ &unions
+      , np
+      ] (ABCTuple const& abc, MPI_Comm const& c) -> Database {
+
+        WITH_CHRONO("db:comm:type:do",
+          auto MPI_LDB_ELEMENT = Slice<F>::mpi::localDatabaseElement();
+        )
+
+        WITH_CHRONO("db:comm:ldb",
+          typename Slice<F>::LocalDatabase ldb;
+          for (auto const& tensor: unions) {
+            auto const& tensorDb = buildLocalDatabase(*tensor, abc);
+            ldb.insert(ldb.end(), tensorDb.begin(), tensorDb.end());
+          }
+        )
+
+        Database db(np * ldb.size(), ldb[0]);
+
+        WITH_CHRONO("oneshot-db:comm:allgather",
+        WITH_CHRONO("db:comm:allgather",
+                    MPI_Allgather(ldb.data(),
+                                  /* ldb.size() * sizeof(typename
+                                     Slice<F>::LocalDatabaseElement) */
+                                  ldb.size(),
+                                  MPI_LDB_ELEMENT,
+                                  db.data(),
+                                  /* ldb.size() * sizeof(typename
+                                     Slice<F>::LocalDatabaseElement), */
+                                  ldb.size(),
+                                  MPI_LDB_ELEMENT,
+                                  c);
+        ))
+
+        WITH_CHRONO("db:comm:type:free", MPI_Type_free(&MPI_LDB_ELEMENT);)
+
+        return db;
+      };
+
+  auto doIOPhase
+    = [&unions, &rank, &np] (Database const& db,
+                             std::vector<LocalDatabaseElement> &to_send) {
+
+    const size_t localDBLength = db.size() / np;
+
+    size_t sendTag = 0
+         , recvTag = rank * localDBLength
+         ;
+
+    {
+      // At this point, we have already send to everyone that fits
+      auto const& begin = &db[rank * localDBLength]
+                , end   = begin + localDBLength
+                ;
+      for (auto it = begin; it != end; ++it) {
+        recvTag++;
+        auto const& el = *it;
+        auto& u = unionByName(unions, el.name);
+        auto& slice = Slice<F>::findByInfo(u.slices, el.info);
+        slice.markReady();
+        // u.receive(el.info, recvTag);
+
+      } // recv
+    }
+
+    // SEND PHASE =========================================================
+    for (size_t otherRank = 0; otherRank < np; otherRank++) {
+      auto const& begin = &db[otherRank * localDBLength]
+                , end = begin + localDBLength
+                ;
+      for (auto it = begin; it != end; ++it) {
+        sendTag++;
+        typename Slice<F>::LocalDatabaseElement const& el = *it;
+        if (el.info.from.rank != rank) continue;
+        auto& u = unionByName(unions, el.name);
+        if (el.info.state == Slice<F>::Fetch) {
+          to_send.push_back(el);
+        }
+        // u.send(otherRank, el, sendTag);
+
+      } // send phase
+
+    } // otherRank
+
+
+  };
+
+  std::vector<LocalDatabaseElement>
+    to_send;
+
+  for (size_t it = 0; it < tuplesList.size(); it++) {
+
+
+    const ABCTuple abc = dist->tupleIsFake(tuplesList[it])
+                       ? tuplesList[tuplesList.size() - 1]
+                       : tuplesList[it]
+                 ;
+
+    if (it > 0) {
+      for (auto const& u: unions) {
+        clearUnusedSlicesForNext(*u, abc);
+      }
+    }
+
+    const auto db = communicateDatabase(abc, kaun);
+    doIOPhase(db, to_send);
+
+    if (it % 1000 == 0)
+      std::cout << _FORMAT("%ld :it %ld  %f %% ∷ %ld ∷ %f GB\n",
+                           rank,
+                           it,
+                           100.0 * double(to_send.size()) / double(tuplesList.size()),
+                           to_send.size(),
+                           double(to_send.size()) * sizeof(to_send[0])
+                           / 1024.0 / 1024.0 / 1024.0);
+
+
+    for (auto const& u: unions) {
+      for (auto type: u->sliceTypes) {
+        unwrapSlice(type, abc, u);
+      }
+    }
+
+
+  }
+
+  std::cout << "=========================================================\n";
+  std::cout << "FINISHING, it will segfaulten, that's ok, don't even trip"
+            << std::endl;
+  MPI_Barrier(kaun);
+  DEINIT_DRY(t_abph);
+  DEINIT_DRY(t_abhh);
+  DEINIT_DRY(t_tabhh);
+  DEINIT_DRY(t_taphh);
+  DEINIT_DRY(t_hhha);
+
+  MPI_Finalize();
+  return 0;
+}

bench/utils.hpp

@@ -0,0 +1,12 @@
+#ifndef UTILS_HPP_
+#define UTILS_HPP_
+
+#define _FORMAT(_fmt, ...)                                    \
+  ([&] (void) -> std::string {                                \
+     int _sz = std::snprintf(nullptr, 0, _fmt, __VA_ARGS__);  \
+     std::vector<char>  _out(_sz  +  1);                      \
+     std::snprintf(&_out[0], _out.size(), _fmt, __VA_ARGS__); \
+     return std::string(_out.data());                         \
+   })()
+
+#endif

configure.ac

@@ -21,26 +21,6 @@ AC_ARG_ENABLE(shared,
                                      files (default=YES)]),
               [], [enable_shared=yes])
 
-AC_ARG_ENABLE(
-    [slice],
-    [AS_HELP_STRING(
-      [--disable-slice],
-      [Disable the step of slicing tensors for CTF, this is useful for example for benchmarking or testing.])],
-    [atrip_dont_slice=1
-     AC_DEFINE([ATRIP_DONT_SLICE],1,[Wether CTF will slice tensors or skip the step])
-    ],
-    [atrip_dont_slice=0]
-)
-
-AC_ARG_ENABLE(
-    [atrip_dgemm],
-    [AS_HELP_STRING(
-      [--disable-dgemm],
-      [Disable using dgemm for the doubles equations])],
-    [],
-    [AC_DEFINE([ATRIP_USE_DGEMM],1,[Use dgemm for the doubles equations])]
-)
-
 
 AC_ARG_ENABLE([docs],
               [AS_HELP_STRING([--enable-docs],
@@ -74,13 +54,53 @@ AC_ARG_VAR([NVCC], [Path to the nvidia cuda compiler.])
 AC_ARG_VAR([CUDA_LDFLAGS], [LDFLAGS to find libraries -lcuda, -lcudart, -lcublas.])
 AC_ARG_VAR([CUDA_CXXFLAGS], [CXXFLAGS to find the CUDA headers])
 
+dnl -----------------------------------------------------------------------
+dnl ATRIP CPP DEFINES
+dnl -----------------------------------------------------------------------
 
 AC_ARG_WITH([atrip-debug],
             [AS_HELP_STRING([--with-atrip-debug],
-                            [Debug level for atrip, possible values: 1, 2, 3, 4])],
+                            [Debug level for atrip, possible values:
+                                   1, 2, 3, 4])],
             [AC_DEFINE([ATRIP_DEBUG],[atrip-debug],[Atrip debug level])],
-            [AC_DEFINE([ATRIP_DEBUG],[1],[Atrip debug level])]
-            )
+            [AC_DEFINE([ATRIP_DEBUG],[1],[Atrip debug level])])
+
+
+AC_ARG_ENABLE([atrip_dgemm],
+              [AS_HELP_STRING([--disable-dgemm],
+                              [Disable using dgemm for the doubles equations])],
+              [],
+              [AC_DEFINE([ATRIP_USE_DGEMM],
+                         1,
+                         [Use dgemm for the doubles equations])])
+
+ATRIP_DEF([slice], [disable],
+          [ATRIP_DONT_SLICE],
+          [Disable the step of slicing tensors for CTF, this is useful
+          for example for benchmarking or testing.])
+
+ATRIP_DEF([only-dgemm], [enable],
+          [ATRIP_ONLY_DGEMM],
+          [Run only the parts of atrip that involve dgemm calls, this
+          is useful for benchmarking and testing the code, it is
+          intended for developers of Atrip.])
+
+ATRIP_DEF([naive-slow], [enable],
+          [ATRIP_NAIVE_SLOW],
+          [Run slow but correct code for the mapping of (iteration,
+          rank) to tuple of the naive tuple distribution.])
+
+ATRIP_DEF([sources-in-gpu], [enable],
+          [ATRIP_SOURCES_IN_GPU],
+          [When using CUDA, activate storing all sources (slices of
+          the input tensors) in the GPU. This means that a lot of GPUs
+          will be needed.])
+
+ATRIP_DEF([cuda-aware-mpi], [enable],
+          [ATRIP_CUDA_AWARE_MPI],
+          [When using MPI, assume support for CUDA aware mpi by the
+          given MPI implementation.])
+
 
 dnl -----------------------------------------------------------------------
 
@@ -144,8 +164,7 @@ AC_TYPE_SIZE_T
 dnl -----------------------------------------------------------------------
 dnl CHECK CTF
 if test xYES = x${BUILD_CTF}; then
-  AC_MSG_WARN([Sorry, building CTF not supported yet provide a build path
-                      with --with-ctf=path/to/ctf/installation])
+  AC_MSG_WARN([You will have to do make ctf before building the project.])
 else
   CPPFLAGS="$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

etc/m4/atrip-def.m4

@@ -0,0 +1,8 @@
+AC_DEFUN([ATRIP_DEF],
+         [AC_ARG_ENABLE([$1],
+                        [AS_HELP_STRING([--$2-$1],
+                                        [$4])],
+                        [AC_DEFINE([$3],
+                                   1,
+                                   [$4])])])
+

etc/nix/openblas.nix

@@ -20,6 +20,8 @@ in
 
 {
 
+  pkg = myopenblas;
+
   buildInputs = with pkgs; [
     myopenblas
     scalapack

etc/nix/vendor-shell.nix

@@ -0,0 +1,27 @@
+rec {
+
+  directory = "vendor";
+  src = ''
+
+    _add_vendor_cpath () {
+      export CPATH=$CPATH:$1
+      mkdir -p ${directory}/include
+      ln -frs $1/* ${directory}/include/
+    }
+
+    _add_vendor_lib () {
+      mkdir -p ${directory}/lib
+      ln -frs $1/* ${directory}/lib/
+    }
+
+  '';
+
+  cpath = path: ''
+    _add_vendor_cpath ${path}
+  '';
+
+  lib = path: ''
+    _add_vendor_lib ${path}
+  '';
+
+}

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/DatabaseCommunicator.hpp

@@ -0,0 +1,20 @@
+#pragma once
+#include <atrip/Utils.hpp>
+#include <atrip/Equations.hpp>
+#include <atrip/SliceUnion.hpp>
+#include <atrip/Unions.hpp>
+
+namespace atrip {
+
+  template <typename F>
+  using Unions = std::vector<SliceUnion<F>*>;
+
+  template <typename F>
+  typename Slice<F>::Database
+  naiveDatabase(Unions<F> &unions,
+                size_t nv,
+                size_t np,
+                size_t iteration,
+                MPI_Comm const& c);
+
+}  // namespace atrip

include/atrip/Debug.hpp

@@ -76,7 +76,7 @@
 
 // [[file:~/cuda/atrip/atrip.org::*Macros][Macros:2]]
 #ifndef LOG
-#define LOG(level, name) if (Atrip::rank == 0) std::cout << name << ": "
+#define LOG(level, name) if (atrip::Atrip::rank == 0) std::cout << name << ": "
 #endif
 // Macros:2 ends here
 

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

@@ -18,6 +18,12 @@
 #include <atrip/Slice.hpp>
 #include <atrip/RankMap.hpp>
 
+#if defined(ATRIP_SOURCES_IN_GPU)
+#  define SOURCES_DATA(s) (s)
+#else
+#  define SOURCES_DATA(s) (s).data()
+#endif
+
 namespace atrip {
 // Prolog:1 ends here
 
@@ -195,7 +201,7 @@ template <typename F=double>
                            ;
           if (blank.info.state == Slice<F>::SelfSufficient) {
 #if defined(HAVE_CUDA)
-            const size_t _size = sizeof(F) * sources[from.source].size();
+            const size_t _size = sizeof(F) * sliceSize;
             // TODO: this is code duplication with downstairs
             if (freePointers.size() == 0) {
               std::stringstream stream;
@@ -212,12 +218,12 @@ template <typename F=double>
             WITH_CHRONO("cuda:memcpy:self-sufficient",
                         _CHECK_CUDA_SUCCESS("copying mpi data to device",
                                             cuMemcpyHtoD(blank.data,
-                                                         (void*)sources[from.source].data(),
-                                                         sizeof(F) * sources[from.source].size()));
+                                                         (void*)SOURCES_DATA(sources[from.source]),
+                                                         sizeof(F) * sliceSize));
                         ))
 
 #else
-            blank.data = sources[from.source].data();
+            blank.data = SOURCES_DATA(sources[from.source]);
 #endif
           } else {
             if (freePointers.size() == 0) {
@@ -396,23 +402,44 @@ template <typename F=double>
               , world(child_world)
               , universe(global_world)
               , sliceLength(sliceLength_)
+              , sliceSize(std::accumulate(sliceLength.begin(),
+                                          sliceLength.end(),
+                                          1UL, std::multiplies<size_t>()))
+
+#if defined(ATRIP_SOURCES_IN_GPU)
+              , sources(rankMap.nSources())
+#else
               , sources(rankMap.nSources(),
-                        std::vector<F>
-                          (std::accumulate(sliceLength.begin(),
-                                           sliceLength.end(),
-                                           1UL, std::multiplies<size_t>())))
+                        std::vector<F>(sliceSize))
+#endif
               , name(name_)
               , sliceTypes(sliceTypes_)
               , sliceBuffers(nSliceBuffers)
-              //, slices(2 * sliceTypes.size(), Slice<F>{ sources[0].size() })
     { // 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)
         const CUresult error =
-          cuMemAlloc(&ptr, sizeof(F) * sources[0].size());
+          cuMemAlloc(&ptr, sizeof(F) * sliceSize);
         if (ptr == 0UL) {
           throw "UNSUFICCIENT MEMORY ON THE GRAPHIC CARD FOR FREE POINTERS";
         }
@@ -423,12 +450,12 @@ template <typename F=double>
           throw s.str();
         }
 #else
-        ptr = (DataPtr<F>)malloc(sizeof(F) * sources[0].size());
+        ptr = (DataPtr<F>)malloc(sizeof(F) * sliceSize);
 #endif
       }
 
       slices
-        = std::vector<Slice<F>>(2 * sliceTypes.size(), { sources[0].size() });
+        = std::vector<Slice<F>>(2 * sliceTypes.size(), { sliceSize });
       // TODO: think exactly    ^------------------- about this number
 
       // initialize the freePointers with the pointers to the buffers
@@ -436,17 +463,45 @@ 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";
       WITH_RANK << "#slices[0] " << slices[0].size << "\n";
       LOG(1,"Atrip") << "#sources " << sources.size() << "\n";
-      WITH_RANK << "#sources[0] " << sources[0].size() << "\n";
+      WITH_RANK << "#sources[0] " << sliceSize << "\n";
       WITH_RANK << "#freePointers " << freePointers.size() << "\n";
       LOG(1,"Atrip") << "#sliceBuffers " << sliceBuffers.size() << "\n";
       LOG(1,"Atrip") << "GB*" << np << " "
                            << double(sources.size() + sliceBuffers.size())
-                            * sources[0].size()
+                            * sliceSize
                             * 8 * np
                             / 1073741824.0
                            << "\n";
@@ -495,14 +550,13 @@ template <typename F=double>
       if (otherRank == info.from.rank)      sendData_p = false;
       if (!sendData_p) return;
 
-      MPI_Isend( sources[info.from.source].data()
-               , sources[info.from.source].size()
-               , traits::mpi::datatypeOf<F>()
-               , otherRank
-               , tag
-               , universe
-               , &request
-               );
+      MPI_Isend((void*)SOURCES_DATA(sources[info.from.source]),
+                sliceSize,
+                traits::mpi::datatypeOf<F>(),
+                otherRank,
+                tag,
+                universe,
+                &request);
       WITH_CRAZY_DEBUG
       WITH_RANK << "sent to " << otherRank << "\n";
 
@@ -516,25 +570,25 @@ template <typename F=double>
 
       if (Atrip::rank == info.from.rank) return;
 
-      if (slice.info.state == Slice<F>::Fetch) {
+      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)
-        slice.mpi_data = (F*)malloc(sizeof(F) * slice.size);
-        MPI_Irecv( slice.mpi_data
+#if defined(HAVE_CUDA) && defined(ATRIP_SOURCES_IN_GPU)
+#  if !defined(ATRIP_CUDA_AWARE_MPI) 
+#    error "You need CUDA aware MPI to have slices on the GPU"
+#  endif
+        MPI_Irecv((void*)slice.data,
 #else
-        MPI_Irecv( slice.data
+        MPI_Irecv(slice.data,
 #endif
-                 , slice.size
-                 , traits::mpi::datatypeOf<F>()
-                 , info.from.rank
-                 , tag
-                 , universe
-                 , &slice.request
-                //, MPI_STATUS_IGNORE
-                 );
-      }
-    }
+                  slice.size,
+                  traits::mpi::datatypeOf<F>(),
+                  info.from.rank,
+                  tag,
+                  universe,
+                  &slice.request);
+       } // if-1
+    } // receive
 
     void unwrapAll(ABCTuple const& abc) {
       for (auto type: sliceTypes) unwrapSlice(type, abc);
@@ -597,7 +651,12 @@ template <typename F=double>
     const MPI_Comm world;
     const MPI_Comm universe;
     const std::vector<size_t> sliceLength;
+    const size_t sliceSize;
+#if defined(ATRIP_SOURCES_IN_GPU)
+    std::vector< DataPtr<F> > sources;
+#else
     std::vector< std::vector<F> > sources;
+#endif
     std::vector< Slice<F> > slices;
     typename Slice<F>::Name name;
     const std::vector<typename Slice<F>::Type> sliceTypes;

include/atrip/Tuples.hpp

@@ -52,43 +52,7 @@ struct TuplesDistribution {
 // Distributing the tuples:1 ends here
 
 // [[file:~/cuda/atrip/atrip.org::*Node%20information][Node information:1]]
-std::vector<std::string> getNodeNames(MPI_Comm comm){
-  int rank, np;
-  MPI_Comm_rank(comm, &rank);
-  MPI_Comm_size(comm, &np);
-
-  std::vector<std::string> nodeList(np);
-  char nodeName[MPI_MAX_PROCESSOR_NAME];
-  char *nodeNames = (char*)malloc(np * MPI_MAX_PROCESSOR_NAME);
-  std::vector<int> nameLengths(np)
-                 , off(np)
-                 ;
-  int nameLength;
-  MPI_Get_processor_name(nodeName, &nameLength);
-  MPI_Allgather(&nameLength,
-                1,
-                MPI_INT,
-                nameLengths.data(),
-                1,
-                MPI_INT,
-                comm);
-  for (int i(1); i < np; i++)
-    off[i] = off[i-1] + nameLengths[i-1];
-  MPI_Allgatherv(nodeName,
-                 nameLengths[rank],
-                 MPI_BYTE,
-                 nodeNames,
-                 nameLengths.data(),
-                 off.data(),
-                 MPI_BYTE,
-                 comm);
-  for (int i(0); i < np; i++) {
-    std::string const s(&nodeNames[off[i]], nameLengths[i]);
-    nodeList[i] = s;
-  }
-  std::free(nodeNames);
-  return nodeList;
-}
+  std::vector<std::string> getNodeNames(MPI_Comm comm);
 // Node information:1 ends here
 
 // [[file:~/cuda/atrip/atrip.org::*Node%20information][Node information:2]]
@@ -100,118 +64,28 @@ struct RankInfo {
   const size_t ranksPerNode;
 };
 
-template <typename A>
-A unique(A const &xs) {
-  auto result = xs;
-  std::sort(std::begin(result), std::end(result));
-  auto const& last = std::unique(std::begin(result), std::end(result));
-  result.erase(last, std::end(result));
-  return result;
-}
-
 std::vector<RankInfo>
-getNodeInfos(std::vector<string> const& nodeNames) {
-  std::vector<RankInfo> result;
-  auto const uniqueNames = unique(nodeNames);
-  auto const index = [&uniqueNames](std::string const& s) {
-    auto const& it = std::find(uniqueNames.begin(), uniqueNames.end(), s);
-    return std::distance(uniqueNames.begin(), it);
-  };
-  std::vector<size_t> localRanks(uniqueNames.size(), 0);
-  size_t globalRank = 0;
-  for (auto const& name: nodeNames) {
-    const size_t nodeId = index(name);
-    result.push_back({name,
-                      nodeId,
-                      globalRank++,
-                      localRanks[nodeId]++,
-                      (size_t)
-                      std::count(nodeNames.begin(),
-                                 nodeNames.end(),
-                                 name)
-                      });
-  }
-  return result;
-}
+getNodeInfos(std::vector<string> const& nodeNames);
 
 struct ClusterInfo {
   const size_t nNodes, np, ranksPerNode;
   const std::vector<RankInfo> rankInfos;
 };
 
-ClusterInfo
-getClusterInfo(MPI_Comm comm) {
-  auto const names = getNodeNames(comm);
-  auto const rankInfos = getNodeInfos(names);
-
-  return ClusterInfo {
-    unique(names).size(),
-    names.size(),
-    rankInfos[0].ranksPerNode,
-    rankInfos
-  };
-
-}
+ClusterInfo getClusterInfo(MPI_Comm comm);
 // Node information:2 ends here
 
 // [[file:~/cuda/atrip/atrip.org::*Naive%20list][Naive list:1]]
-ABCTuples getTuplesList(size_t Nv, size_t rank, size_t np) {
-
-  const size_t
-    // total number of tuples for the problem
-       n = Nv * (Nv + 1) * (Nv + 2) / 6 - Nv
-
-    // all ranks should have the same number of tuples_per_rank
-    , tuples_per_rank = n / np + size_t(n % np != 0)
-
-    // start index for the global tuples list
-    , start = tuples_per_rank * rank
-
-    // end index for the global tuples list
-    , end = tuples_per_rank * (rank + 1)
-    ;
-
-  LOG(1,"Atrip") << "tuples_per_rank = " << tuples_per_rank << "\n";
-  WITH_RANK << "start, end = " << start << ", " << end << "\n";
-  ABCTuples result(tuples_per_rank, FAKE_TUPLE);
-
-  for (size_t a(0), r(0), g(0); a < Nv; a++)
-  for (size_t b(a);             b < Nv; b++)
-  for (size_t c(b);             c < Nv; c++){
-    if ( a == b && b == c ) continue;
-    if ( start <= g && g < end) result[r++] = {a, b, c};
-    g++;
-  }
-
-  return result;
-
-}
+ABCTuples getTuplesList(size_t Nv, size_t rank, size_t np);
 // Naive list:1 ends here
 
 // [[file:~/cuda/atrip/atrip.org::*Naive%20list][Naive list:2]]
-ABCTuples getAllTuplesList(const size_t Nv) {
-  const size_t n = Nv * (Nv + 1) * (Nv + 2) / 6 - Nv;
-  ABCTuples result(n);
-
-  for (size_t a(0), u(0); a < Nv; a++)
-  for (size_t b(a); b < Nv; b++)
-  for (size_t c(b); c < Nv; c++){
-    if ( a == b && b == c ) continue;
-    result[u++] = {a, b, c};
-  }
-
-  return result;
-}
+ABCTuples getAllTuplesList(const size_t Nv);
 // Naive list:2 ends here
 
 // [[file:~/cuda/atrip/atrip.org::*Naive%20list][Naive list:3]]
 struct NaiveDistribution : public TuplesDistribution {
-  ABCTuples getTuples(size_t Nv, MPI_Comm universe) override {
-    int rank, np;
-    MPI_Comm_rank(universe, &rank);
-    MPI_Comm_size(universe, &np);
-    return getTuplesList(Nv, (size_t)rank, (size_t)np);
-  }
+  ABCTuples getTuples(size_t Nv, MPI_Comm universe) override;
 };
 // Naive list:3 ends here
 
@@ -224,19 +98,12 @@ namespace group_and_sort {
 // Right now we distribute the slices in a round robin fashion
 // over the different nodes (NOTE: not mpi ranks but nodes)
 inline
-size_t isOnNode(size_t tuple, size_t nNodes) { return tuple % nNodes; }
+size_t isOnNode(size_t tuple, size_t nNodes);
 
 
 // return the node (or all nodes) where the elements of this
 // tuple are located
-std::vector<size_t> getTupleNodes(ABCTuple const& t, size_t nNodes) {
-  std::vector<size_t>
-    nTuple = { isOnNode(t[0], nNodes)
-             , isOnNode(t[1], nNodes)
-             , isOnNode(t[2], nNodes)
-             };
-  return unique(nTuple);
-}
+std::vector<size_t> getTupleNodes(ABCTuple const& t, size_t nNodes);
 
 struct Info {
   size_t nNodes;
@@ -245,302 +112,16 @@ struct Info {
 // Utils:1 ends here
 
 // [[file:~/cuda/atrip/atrip.org::*Distribution][Distribution:1]]
-ABCTuples specialDistribution(Info const& info, ABCTuples const& allTuples) {
-
-  ABCTuples nodeTuples;
-  size_t const nNodes(info.nNodes);
-
-  std::vector<ABCTuples>
-      container1d(nNodes)
-    , container2d(nNodes * nNodes)
-    , container3d(nNodes * nNodes * nNodes)
-    ;
-
-  WITH_DBG if (info.nodeId == 0)
-    std::cout << "\tGoing through all "
-              << allTuples.size()
-              << " tuples in "
-              << nNodes
-              << " nodes\n";
-
-  // build container-n-d's
-  for (auto const& t: allTuples) {
-    // one which node(s) are the tuple elements located...
-    // put them into the right container
-    auto const _nodes = getTupleNodes(t, nNodes);
-
-    switch (_nodes.size()) {
-      case 1:
-        container1d[_nodes[0]].push_back(t);
-        break;
-      case 2:
-        container2d[ _nodes[0]
-                   + _nodes[1] * nNodes
-                   ].push_back(t);
-        break;
-      case 3:
-        container3d[ _nodes[0]
-                   + _nodes[1] * nNodes
-                   + _nodes[2] * nNodes * nNodes
-                   ].push_back(t);
-        break;
-    }
-
-  }
-
-  WITH_DBG if (info.nodeId == 0)
-    std::cout << "\tBuilding 1-d containers\n";
-  // DISTRIBUTE 1-d containers
-  // every tuple which is only located at one node belongs to this node
-  {
-    auto const& _tuples = container1d[info.nodeId];
-    nodeTuples.resize(_tuples.size(), INVALID_TUPLE);
-    std::copy(_tuples.begin(), _tuples.end(), nodeTuples.begin());
-  }
-
-  WITH_DBG if (info.nodeId == 0)
-    std::cout << "\tBuilding 2-d containers\n";
-  // DISTRIBUTE 2-d containers
-  //the tuples which are located at two nodes are half/half given to these nodes
-  for (size_t yx = 0; yx < container2d.size(); yx++) {
-
-    auto const& _tuples = container2d[yx];
-      const
-    size_t idx = yx % nNodes
-         // remeber: yx = idy * nNodes + idx
-         , idy = yx / nNodes
-         , n_half = _tuples.size() / 2
-         , size = nodeTuples.size()
-         ;
-
-    size_t nbeg, nend;
-    if (info.nodeId == idx) {
-      nbeg = 0 * n_half;
-      nend = n_half;
-    } else if (info.nodeId == idy) {
-      nbeg = 1 * n_half;
-      nend = _tuples.size();
-    } else {
-      // either idx or idy is my node
-      continue;
-    }
-
-    size_t const nextra = nend - nbeg;
-    nodeTuples.resize(size + nextra, INVALID_TUPLE);
-    std::copy(_tuples.begin() + nbeg,
-              _tuples.begin() + nend,
-              nodeTuples.begin() + size);
-
-  }
-
-  WITH_DBG if (info.nodeId == 0)
-    std::cout << "\tBuilding 3-d containers\n";
-  // DISTRIBUTE 3-d containers
-  for (size_t zyx = 0; zyx < container3d.size(); zyx++) {
-    auto const& _tuples = container3d[zyx];
-
-      const
-    size_t idx = zyx % nNodes
-         , idy = (zyx / nNodes) % nNodes
-         // remember: zyx = idx + idy * nNodes + idz * nNodes^2
-         , idz = zyx / nNodes / nNodes
-         , n_third = _tuples.size() / 3
-         , size = nodeTuples.size()
-         ;
-
-    size_t nbeg, nend;
-    if (info.nodeId == idx) {
-      nbeg = 0 * n_third;
-      nend = 1 * n_third;
-    } else if (info.nodeId == idy) {
-      nbeg = 1 * n_third;
-      nend = 2 * n_third;
-    } else if (info.nodeId == idz) {
-      nbeg = 2 * n_third;
-      nend = _tuples.size();
-    } else {
-      // either idx or idy or idz is my node
-      continue;
-    }
-
-    size_t const nextra = nend - nbeg;
-    nodeTuples.resize(size + nextra, INVALID_TUPLE);
-    std::copy(_tuples.begin() + nbeg,
-              _tuples.begin() + nend,
-              nodeTuples.begin() + size);
-
-  }
-
-
-  WITH_DBG if (info.nodeId == 0) std::cout << "\tswapping tuples...\n";
-  /*
-   *  sort part of group-and-sort algorithm
-   *  every tuple on a given node is sorted in a way that
-   *  the 'home elements' are the fastest index.
-   *  1:yyy 2:yyn(x) 3:yny(x) 4:ynn(x) 5:nyy 6:nyn(x) 7:nny 8:nnn
-   */
-  for (auto &nt: nodeTuples){
-    if ( isOnNode(nt[0], nNodes) == info.nodeId ){ // 1234
-      if ( isOnNode(nt[2], nNodes) != info.nodeId ){ // 24
-        size_t const x(nt[0]);
-        nt[0] = nt[2];         // switch first and last
-        nt[2] = x;
-      }
-      else if ( isOnNode(nt[1], nNodes) != info.nodeId){ // 3
-        size_t const x(nt[0]);
-        nt[0] = nt[1];         // switch first two
-        nt[1] = x;
-      }
-    } else {
-      if ( isOnNode(nt[1], nNodes) == info.nodeId   // 56
-        && isOnNode(nt[2], nNodes) != info.nodeId
-        ) { // 6
-        size_t const x(nt[1]);
-        nt[1] = nt[2];         // switch last two
-        nt[2] = x;
-      }
-    }
-  }
-
-  WITH_DBG if (info.nodeId == 0) std::cout << "\tsorting list of tuples...\n";
-  //now we sort the list of tuples
-  std::sort(nodeTuples.begin(), nodeTuples.end());
-
-  WITH_DBG if (info.nodeId == 0) std::cout << "\trestoring tuples...\n";
-  // we bring the tuples abc back in the order a<b<c
-  for (auto &t: nodeTuples)  std::sort(t.begin(), t.end());
-
-#if ATRIP_DEBUG > 1
-  WITH_DBG if (info.nodeId == 0)
-  std::cout << "checking for validity of " << nodeTuples.size() << std::endl;
-  const bool anyInvalid
-    = std::any_of(nodeTuples.begin(),
-                  nodeTuples.end(),
-                  [](ABCTuple const& t) { return t == INVALID_TUPLE; });
-  if (anyInvalid) throw "Some tuple is invalid in group-and-sort algorithm";
-#endif
-
-  WITH_DBG if (info.nodeId == 0) std::cout << "\treturning tuples...\n";
-  return nodeTuples;
-
-}
+ABCTuples specialDistribution(Info const& info, ABCTuples const& allTuples);
 // Distribution:1 ends here
 
 // [[file:~/cuda/atrip/atrip.org::*Main][Main:1]]
-std::vector<ABCTuple> main(MPI_Comm universe, size_t Nv) {
-
-  int rank, np;
-  MPI_Comm_rank(universe, &rank);
-  MPI_Comm_size(universe, &np);
-
-  std::vector<ABCTuple> result;
-
-  auto const nodeNames(getNodeNames(universe));
-  size_t const nNodes = unique(nodeNames).size();
-  auto const nodeInfos = getNodeInfos(nodeNames);
-
-  // We want to construct a communicator which only contains of one
-  // element per node
-  bool const computeDistribution
-    = nodeInfos[rank].localRank == 0;
-
-  std::vector<ABCTuple>
-    nodeTuples
-      = computeDistribution
-      ? specialDistribution(Info{nNodes, nodeInfos[rank].nodeId},
-                            getAllTuplesList(Nv))
-      : std::vector<ABCTuple>()
-      ;
-
-  LOG(1,"Atrip") << "got nodeTuples\n";
-
-  // now we have to send the data from **one** rank on each node
-  // to all others ranks of this node
-    const
-  int color = nodeInfos[rank].nodeId
-    , key = nodeInfos[rank].localRank
-    ;
-
-
-  MPI_Comm INTRA_COMM;
-  MPI_Comm_split(universe, color, key, &INTRA_COMM);
-// Main:1 ends here
-
-// [[file:~/cuda/atrip/atrip.org::*Main][Main:2]]
-size_t const
-  tuplesPerRankLocal
-     = nodeTuples.size() / nodeInfos[rank].ranksPerNode
-     + size_t(nodeTuples.size() % nodeInfos[rank].ranksPerNode != 0)
-     ;
-
-size_t tuplesPerRankGlobal;
-
-MPI_Reduce(&tuplesPerRankLocal,
-           &tuplesPerRankGlobal,
-           1,
-           MPI_UINT64_T,
-           MPI_MAX,
-           0,
-           universe);
-
-MPI_Bcast(&tuplesPerRankGlobal,
-          1,
-          MPI_UINT64_T,
-          0,
-          universe);
-
-LOG(1,"Atrip") << "Tuples per rank: " << tuplesPerRankGlobal << "\n";
-LOG(1,"Atrip") << "ranks per node " << nodeInfos[rank].ranksPerNode << "\n";
-LOG(1,"Atrip") << "#nodes " << nNodes << "\n";
-// Main:2 ends here
-
-// [[file:~/cuda/atrip/atrip.org::*Main][Main:3]]
-size_t const totalTuples
-  = tuplesPerRankGlobal * nodeInfos[rank].ranksPerNode;
-
-if (computeDistribution) {
-  // pad with FAKE_TUPLEs
-  nodeTuples.insert(nodeTuples.end(),
-                    totalTuples - nodeTuples.size(),
-                    FAKE_TUPLE);
-}
-// Main:3 ends here
-
-// [[file:~/cuda/atrip/atrip.org::*Main][Main:4]]
-{
-  // construct mpi type for abctuple
-  MPI_Datatype MPI_ABCTUPLE;
-  MPI_Type_vector(nodeTuples[0].size(), 1, 1, MPI_UINT64_T, &MPI_ABCTUPLE);
-  MPI_Type_commit(&MPI_ABCTUPLE);
-
-  LOG(1,"Atrip") << "scattering tuples \n";
-
-  result.resize(tuplesPerRankGlobal);
-  MPI_Scatter(nodeTuples.data(),
-              tuplesPerRankGlobal,
-              MPI_ABCTUPLE,
-              result.data(),
-              tuplesPerRankGlobal,
-              MPI_ABCTUPLE,
-              0,
-              INTRA_COMM);
-
-  MPI_Type_free(&MPI_ABCTUPLE);
-
-}
-// Main:4 ends here
-
-// [[file:~/cuda/atrip/atrip.org::*Main][Main:5]]
-return result;
-
-}
+std::vector<ABCTuple> main(MPI_Comm universe, size_t Nv);
 // Main:5 ends here
 
 // [[file:~/cuda/atrip/atrip.org::*Interface][Interface:1]]
 struct Distribution : public TuplesDistribution {
-  ABCTuples getTuples(size_t Nv, MPI_Comm universe) override {
-    return main(universe, Nv);
-  }
+  ABCTuples getTuples(size_t Nv, MPI_Comm universe) override;
 };
 // Interface:1 ends here
 

include/atrip/Unions.hpp

@@ -19,8 +19,14 @@
 namespace atrip {
 
   template <typename F=double>
+  static
   void sliceIntoVector
-    ( std::vector<F> &v
+#if defined(ATRIP_SOURCES_IN_GPU)
+    ( DataPtr<F> &source
+#else
+    ( std::vector<F> &source
+#endif
+    , size_t sliceSize
     , CTF::Tensor<F> &toSlice
     , std::vector<int64_t> const low
     , std::vector<int64_t> const up
@@ -44,18 +50,30 @@ namespace atrip {
               << "\n";
 
 #ifndef ATRIP_DONT_SLICE
-    toSlice.slice( toSlice_.low.data()
-                 , toSlice_.up.data()
-                 , 0.0
-                 , origin
-                 , origin_.low.data()
-                 , origin_.up.data()
-                 , 1.0);
-    memcpy(v.data(), toSlice.data, sizeof(F) * v.size());
+    toSlice.slice(toSlice_.low.data(),
+                  toSlice_.up.data(),
+                  0.0,
+                  origin,
+                  origin_.low.data(),
+                  origin_.up.data(),
+                  1.0);
 #else
 #  pragma message("WARNING: COMPILING WITHOUT SLICING THE TENSORS")
 #endif
 
+#if defined(ATRIP_SOURCES_IN_GPU)
+    WITH_CHRONO("cuda:sources",
+                _CHECK_CUDA_SUCCESS("copying sources data to device",
+                                    cuMemcpyHtoD(source,
+                                                 toSlice.data,
+                                                 sliceSize));
+                )
+#else
+    memcpy(source.data(),
+           toSlice.data,
+           sizeof(F) * sliceSize);
+#endif
+
   }
 
 
@@ -80,16 +98,15 @@ namespace atrip {
 
     void sliceIntoBuffer(size_t it, CTF::Tensor<F> &to, CTF::Tensor<F> const& from) override
     {
-      const int Nv = this->sliceLength[0]
-              , No = this->sliceLength[1]
-              , a = this->rankMap.find({static_cast<size_t>(Atrip::rank), it});
-              ;
 
+      const int
+        Nv = this->sliceLength[0],
+        No = this->sliceLength[1],
+        a = this->rankMap.find({static_cast<size_t>(Atrip::rank), it});
 
-      sliceIntoVector<F>( this->sources[it]
-                        , to,   {0, 0, 0},    {Nv, No, No}
-                        , from, {a, 0, 0, 0}, {a+1, Nv, No, No}
-                        );
+      sliceIntoVector<F>(this->sources[it], this->sliceSize,
+                         to,   {0, 0, 0},    {Nv, No, No},
+                         from, {a, 0, 0, 0}, {a+1, Nv, No, No});
 
     }
 
@@ -118,14 +135,13 @@ namespace atrip {
     void sliceIntoBuffer(size_t it, CTF::Tensor<F> &to, CTF::Tensor<F> const& from) override
     {
 
-      const int No = this->sliceLength[0]
-              , a = this->rankMap.find({static_cast<size_t>(Atrip::rank), it})
-              ;
+      const int
+        No = this->sliceLength[0],
+        a = this->rankMap.find({static_cast<size_t>(Atrip::rank), it});
 
-      sliceIntoVector<F>( this->sources[it]
-                        , to,   {0, 0, 0},    {No, No, No}
-                        , from, {0, 0, 0, a}, {No, No, No, a+1}
-                        );
+      sliceIntoVector<F>(this->sources[it], this->sliceSize,
+                         to,   {0, 0, 0},    {No, No, No},
+                         from, {0, 0, 0, a}, {No, No, No, a+1});
 
     }
   };
@@ -153,18 +169,17 @@ namespace atrip {
 
     void sliceIntoBuffer(size_t it, CTF::Tensor<F> &to, CTF::Tensor<F> const& from) override {
 
-      const int Nv = this->sliceLength[0]
-              , No = this->sliceLength[1]
-              , el = this->rankMap.find({static_cast<size_t>(Atrip::rank), it})
-              , a = el % Nv
-              , b = el / Nv
-              ;
+      const int
+        Nv = this->sliceLength[0],
+        No = this->sliceLength[1],
+        el = this->rankMap.find({static_cast<size_t>(Atrip::rank), it}),
+        a = el % Nv,
+        b = el / Nv;
 
 
-      sliceIntoVector<F>( this->sources[it]
-                        , to,   {0, 0},       {Nv, No}
-                        , from, {a, b, 0, 0}, {a+1, b+1, Nv, No}
-                        );
+      sliceIntoVector<F>(this->sources[it], this->sliceSize,
+                         to,   {0, 0},       {Nv, No},
+                         from, {a, b, 0, 0}, {a+1, b+1, Nv, No});
 
     }
 
@@ -191,17 +206,17 @@ namespace atrip {
 
     void sliceIntoBuffer(size_t it, CTF::Tensor<F> &to, CTF::Tensor<F> const& from) override {
 
-      const int Nv = from.lens[0]
-              , No = this->sliceLength[1]
-              , el = this->rankMap.find({static_cast<size_t>(Atrip::rank), it})
-              , a = el % Nv
-              , b = el / Nv
-              ;
+      const int
+        Nv = from.lens[0],
+        No = this->sliceLength[1],
+        el = this->rankMap.find({static_cast<size_t>(Atrip::rank), it}),
+        a = el % Nv,
+        b = el / Nv;
 
-      sliceIntoVector<F>( this->sources[it]
-                        , to,   {0, 0},       {No, No}
-                        , from, {a, b, 0, 0}, {a+1, b+1, No, No}
-                        );
+
+      sliceIntoVector<F>(this->sources[it], this->sliceSize,
+                         to,   {0, 0},       {No, No},
+                         from, {a, b, 0, 0}, {a+1, b+1, No, No});
 
 
     }
@@ -231,17 +246,16 @@ namespace atrip {
     void sliceIntoBuffer(size_t it, CTF::Tensor<F> &to, CTF::Tensor<F> const& from) override {
       // TODO: maybe generalize this with ABHH
 
-      const int Nv = from.lens[0]
-              , No = this->sliceLength[1]
-              , el = this->rankMap.find({static_cast<size_t>(Atrip::rank), it})
-              , a = el % Nv
-              , b = el / Nv
-              ;
+      const int
+        Nv = from.lens[0],
+        No = this->sliceLength[1],
+        el = this->rankMap.find({static_cast<size_t>(Atrip::rank), it}),
+        a = el % Nv,
+        b = el / Nv;
 
-      sliceIntoVector<F>( this->sources[it]
-                        , to,   {0, 0},       {No, No}
-                        , from, {a, b, 0, 0}, {a+1, b+1, No, No}
-                        );
+      sliceIntoVector<F>(this->sources[it], this->sliceSize,
+                         to,   {0, 0},       {No, No},
+                         from, {a, b, 0, 0}, {a+1, b+1, No, No});
 
 
     }

misc/naive-tuples.lisp

@@ -0,0 +1,228 @@
+#+quicklisp
+(eval-when (:compile-toplevel :load-toplevel :execute)
+  (ql:quickload '(vgplot fiveam)))
+
+(defpackage :naive-tuples
+  (:use :cl :vgplot))
+(in-package :naive-tuples)
+
+(defun tuples-atrip (nv)
+  (declare (optimize (speed 3) (safety 0) (debug 0)))
+  (loop :for a :below nv
+        :append
+        (loop :for b :from a :below nv
+              :append
+              (loop :for c :from b :below nv
+                    :unless (= a b c)
+                      :collect (list a b c)))))
+
+(defun tuples-half (nv)
+  (declare (optimize (speed 3) (safety 0) (debug 0)))
+  (loop :for a :below nv
+        :append
+        (loop :for b :from a :below nv
+              :append
+              (loop :for c :from b :below nv
+                    :collect (list a b c)))))
+
+(defun tuples-all (nv)
+  (declare (optimize (speed 3) (safety 0) (debug 0)))
+  (loop :for a :below nv
+        :append
+        (loop :for b :below nv
+              :append
+              (loop :for c :below nv
+                    :collect (list a b c)))))
+
+(defun tuples-all-nth (i nv)
+  (declare (optimize (speed 3) (safety 0) (debug 0)))
+  (list (floor i (* nv nv))
+        (mod (floor i nv) nv)
+        (mod i nv)))
+
+
+(defparameter tups (tuples-all 10))
+
+(defun compare-all (l)
+  (declare (optimize (speed 3) (safety 0) (debug 0)))
+  (let* ((tups (tuples-all l)))
+    (loop for i below (length tups)
+          do (let* ((good (nth i tups))
+                    (bad (tuples-all-nth i l))
+                    (eq? (equal good bad)))
+               (unless eq?
+                 (print (list :|i| i
+                              :good good
+                              :bad bad)))))))
+
+
+;; (defun a-half (i nv)
+;;   (let ((divisor t)
+;;         (j i)
+;;         (total-blk 0))
+;;     (loop :for a :below nv
+;;           :unless (eq divisor 0)
+;;             :do (let ((blk (a-block a nv)))
+;;                   (multiple-value-bind (d r) (floor j blk)
+;;                     (declare (ignore r))
+;;                     (when (> d 0)
+;;                       (incf total-blk blk))
+;;                     (setq j (- j blk)
+;;                           divisor d)))
+;;           :else
+;;             :return (values (- a 1)
+;;                             i
+;;                             total-blk))))
+
+;; (defun b-half (i a nv a-block-sum)
+;;   "we have
+;;     \begin{equation}
+;;       i = \underbrace{B(a_0) +
+;;                       \cdots +
+;;                       B(a_{i-1})}_{\texttt{a-block-sum}}
+;;          + idx
+;;     \end{equation}
+;;     and with this we just have to divide.
+;;   "
+;;   (let ((bj (if (> a-block-sum 0)
+;;                 (mod i a-block-sum)
+;;                 i))
+;;         (total-blk 0))
+;;     (loop :for b :from a :below Nv
+;;           :with divisor = 1
+;;           :unless (eq divisor 0)
+;;             :do (let ((blk (+ (- nv a)
+;;                               #|because|# 1)))
+;;                   (incf total-blk blk)
+;;                   (if (> blk 0)
+;;                       (multiple-value-bind (d r) (floor bj blk)
+;;                         (declare (ignore r))
+;;                         (setq bj (- bj blk)
+;;                               divisor d))
+;;                       (setq divisor 0)))
+;;           :else
+;;             :return (values (- b 1)
+;;                             bj
+;;                             total-blk))))
+
+(defun a-block (a nv)
+  (declare (optimize (speed 3) (safety 0) (debug 0)))
+  (- (* (- nv 1) (- nv (- a 1)))
+     (- (floor (* (- nv 1) nv)
+               2)
+        (floor (* (- a 1) (- a 2))
+               2))))
+
+(defun a-block-sum (|t| nv)
+  (macrolet ((ssum (n) `(floor (* ,n (+ ,n 1))
+                               2))
+             (qsum (n) `(floor (* ,n
+                                  (+ ,n 1)
+                                  (+ 1 (* 2 ,n)))
+                               6)))
+    (let ((nv-1 (- nv 1))
+          (t+1 (+ |t| 1)))
+      (+ (* t+1 nv-1 nv)
+         (* nv-1 t+1)
+         (- (* nv-1
+               (ssum |t|)))
+         (- (* t+1
+               (ssum nv-1)))
+         (floor (- (qsum |t|)
+                   (* 3 (ssum |t|)))
+                2)
+         t+1))))
+
+(defun get-half (i nv &key from block)
+  (let ((divisor 1)
+        (j i)
+        (total-blk 0))
+    (loop :for α :from from :below nv
+          :unless (eq divisor 0)
+            :do (let ((blk (funcall block α nv)))
+                  (multiple-value-bind (d r) (floor j blk)
+                    (declare (ignore r))
+                    (when (> d 0)
+                      (incf total-blk blk)
+                      (setq j (- j blk)))
+                    (setq divisor d)))
+          :else
+            :return (values (- α 1)
+                            j
+                            total-blk))))
+
+(defun tuples-half-nth (i nv)
+  (declare (optimize (speed 3) (safety 0) (debug 0)))
+  (flet ((bc-block (x %nv)
+           (+ 1 (- %nv x))))
+    (multiple-value-bind (a aj blks) (get-half i nv :from 0 :block #'a-block)
+      (declare (ignore blks))
+      (multiple-value-bind (b bj blks) (get-half aj nv
+                                                 :from a
+                                                 :block #'bc-block)
+        (declare (ignore blks))
+        (multiple-value-bind (c cj blks) (get-half bj nv
+                                                   :from b
+                                                   :block #'bc-block)
+          (declare (ignore cj blks))
+          (print (list :idxs aj bj cj))
+          (list a b c))))))
+
+(defun a-block-atrip (a nv)
+  (declare (optimize (speed 3) (safety 0) (debug 0)))
+  (- (a-block a nv) 1))
+
+(defun a-block-sum-atrip (|t| nv)
+  (declare (optimize (speed 3) (safety 0) (debug 0)))
+  (- (a-block-sum |t| nv) (+ |t| 1)))
+
+(defun b-block-sum-atrip (a |t| nv)
+  (- (* nv
+        (1+ (- |t| a)))
+     (floor (- (* |t| (1+ |t|))
+               (* a (- a 1)))
+            2)
+     1))
+
+(defun nth-atrip (i nv)
+  (let ((sums (mapcar (lambda (s) (a-block-sum-atrip s nv))
+                      (loop :for j :below nv :collect j))))
+    (multiple-value-bind (a ablk)
+        (loop :for sum :in sums
+              :with a = -1
+              :with base = 0
+              :do (incf a)
+              :if (eq (floor i sum) 0)
+                :return (values a base)
+              :else
+                :do (setq base sum))
+      (multiple-value-bind (b bblk)
+          (let ((sums (mapcar (lambda (s)
+                                (+ ablk
+                                   #+nil(- nv s 1)
+                                   (b-block-sum-atrip a s nv)))
+                              (loop :for b :from a :below nv
+                                    :collect b))))
+            (loop :for sum :in sums
+                  :with b = (- a 1)
+                  :with base = ablk
+                  :do (incf b)
+                  :if (< i sum)
+                    :return (values b base)
+                  :else
+                    :do (progn
+                          ;; (print sums)
+                          (setq base sum))))
+        (list a b (+ b
+                     (- i bblk)
+                     (if (eq a b)
+                         1
+                         0)))))))
+
+(defun atrip-test (i nv)
+  (let ((tuples (tuples-atrip nv))
+        (cheaper (nth-atrip i nv)))
+    (values (nth i tuples)
+            cheaper
+            (print (equal (nth i tuples)
+                   cheaper)))))

shell.nix

@@ -12,6 +12,7 @@ let
   };
 
   openblas = import ./etc/nix/openblas.nix { inherit pkgs; };
+  vendor = import ./etc/nix/vendor-shell.nix;
 
   mkl-pkg = import ./etc/nix/mkl.nix { pkgs = unfree-pkgs; };
   cuda-pkg = if cuda then (import ./cuda.nix { pkgs = unfree-pkgs; }) else {};
@@ -57,14 +58,15 @@ pkgs.mkShell rec {
   buildInputs
     = with pkgs; [
 
+        gdb
         coreutils
-        git vim
+        git
+        vim
 
         openmpi
         llvmPackages.openmp
 
         binutils
-        emacs
         gfortran
 
         gnumake
@@ -84,6 +86,15 @@ pkgs.mkShell rec {
   shellHook
     =
     ''
+
+    ${vendor.src}
+
+    ${vendor.cpath "${pkgs.openmpi.out}/include"}
+    ${vendor.cpath "${openblas.pkg.dev}/include"}
+
+    ${vendor.lib "${pkgs.openmpi.out}/lib"}
+    ${vendor.lib "${openblas.pkg.out}/lib"}
+
     export OMPI_CXX=${CXX}
     export OMPI_CC=${CC}
     CXX=${CXX}

src/Makefile.am

@@ -7,7 +7,7 @@ AM_CPPFLAGS = $(CTF_CPPFLAGS)
 lib_LIBRARIES = libatrip.a
 
 libatrip_a_CPPFLAGS = -I$(top_srcdir)/include/
-libatrip_a_SOURCES = ./atrip/Blas.cxx
+libatrip_a_SOURCES = ./atrip/Blas.cxx ./atrip/Tuples.cxx ./atrip/DatabaseCommunicator.cxx
 NVCC_FILES = ./atrip/Equations.cxx ./atrip/Complex.cxx ./atrip/Atrip.cxx
 
 if WITH_CUDA

src/atrip/Atrip.cxx

@@ -21,6 +21,7 @@
 #include <atrip/SliceUnion.hpp>
 #include <atrip/Unions.hpp>
 #include <atrip/Checkpoint.hpp>
+#include <atrip/DatabaseCommunicator.hpp>
 
 using namespace atrip;
 #if defined(HAVE_CUDA)
@@ -201,7 +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
@@ -257,6 +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;
 
@@ -299,9 +319,23 @@ Atrip::Output Atrip::run(Atrip::Input<F> const& in) {
   using Database = typename Slice<F>::Database;
   auto communicateDatabase
     = [ &unions
+      , &in
+      , Nv
       , np
-      ] (ABCTuple const& abc, MPI_Comm const& c) -> Database {
+      ] (ABCTuple const& abc, MPI_Comm const& c, size_t iteration) -> Database {
 
+      if (in.tuplesDistribution == Atrip::Input<F>::TuplesDistribution::NAIVE) {
+
+        WITH_CHRONO("db:comm:naive",
+                    auto const& db = naiveDatabase<F>(unions,
+                                                      Nv,
+                                                      np,
+                                                      iteration,
+                                                      c);
+                    )
+        return db;
+
+      } else {
         WITH_CHRONO("db:comm:type:do",
           auto MPI_LDB_ELEMENT = Slice<F>::mpi::localDatabaseElement();
         )
@@ -334,6 +368,8 @@ Atrip::Output Atrip::run(Atrip::Input<F> const& in) {
         WITH_CHRONO("db:comm:type:free", MPI_Type_free(&MPI_LDB_ELEMENT);)
 
         return db;
+      }
+
       };
 
   auto doIOPhase
@@ -437,6 +473,7 @@ Atrip::Output Atrip::run(Atrip::Input<F> const& in) {
 
   // START MAIN LOOP ======================================================{{{1
 
+  MPI_Barrier(universe);
   double energy(0.);
   size_t first_iteration = 0;
   Checkpoint c;
@@ -564,7 +601,7 @@ Atrip::Output Atrip::run(Atrip::Input<F> const& in) {
     // COMM FIRST DATABASE ================================================{{{1
     if (i == first_iteration) {
       WITH_RANK << "__first__:first database ............ \n";
-      const auto db = communicateDatabase(abc, universe);
+      const auto db = communicateDatabase(abc, universe, i);
       WITH_RANK << "__first__:first database communicated \n";
       WITH_RANK << "__first__:first database io phase \n";
       doIOPhase(db);
@@ -579,7 +616,7 @@ Atrip::Output Atrip::run(Atrip::Input<F> const& in) {
     if (abcNext) {
       WITH_RANK << "__comm__:" << iteration << "th communicating database\n";
       WITH_CHRONO("db:comm",
-        const auto db = communicateDatabase(*abcNext, universe);
+        const auto db = communicateDatabase(*abcNext, universe, i);
       )
       WITH_CHRONO("db:io",
         doIOPhase(db);
@@ -621,14 +658,23 @@ 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";
       ))
     }
 
     // COMPUTE SINGLES %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% {{{1
     OCD_Barrier(universe);
+#if defined(ATRIP_ONLY_DGEMM)
     if (false)
+#endif
     if (!isFakeTuple(i)) {
       WITH_CHRONO("oneshot-unwrap",
       WITH_CHRONO("unwrap",
@@ -647,7 +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),
@@ -661,30 +707,73 @@ Atrip::Output Atrip::run(Atrip::Input<F> const& in) {
 
 
     // COMPUTE ENERGY %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% {{{1
+#if defined(ATRIP_ONLY_DGEMM)
+    if (false)
+#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",
-/*
-    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 ( distinct == 0) {
+                    ACC_FUNCALL(getEnergyDistinct<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);
+                  } 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;
 
     }
 
@@ -750,6 +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/DatabaseCommunicator.cxx

@@ -0,0 +1,303 @@
+#include <atrip/DatabaseCommunicator.hpp>
+#include <atrip/Complex.hpp>
+
+
+namespace atrip {
+
+#if defined(ATRIP_NAIVE_SLOW)
+  /*
+   * This function is really too slow, below are more performant
+   * functions to get tuples.
+   */
+  static
+  ABCTuples get_nth_naive_tuples(size_t Nv, size_t np, int64_t i) {
+
+    const size_t
+      // total number of tuples for the problem
+      n = Nv * (Nv + 1) * (Nv + 2) / 6 - Nv
+
+      // all ranks should have the same number of tuples_per_rank
+      , tuples_per_rank = n / np + size_t(n % np != 0)
+      ;
+
+
+    ABCTuples result(np);
+    if (i < 0) return result;
+    std::vector<size_t>
+      rank_indices(np, 0);
+
+    for (size_t a(0), g(0); a < Nv; a++)
+    for (size_t b(a);       b < Nv; b++)
+    for (size_t c(b);       c < Nv; c++){
+      if ( a == b && b == c ) continue;
+      for (size_t rank = 0; rank < np; rank++) {
+
+        const size_t
+          // start index for the global tuples list
+            start = tuples_per_rank * rank
+
+          // end index for the global tuples list
+          , end = tuples_per_rank * (rank + 1)
+          ;
+
+        if ( start <= g && g < end) {
+          if (rank_indices[rank] == i) {
+            result[rank] = {a, b, c};
+          }
+          rank_indices[rank] += 1;
+        }
+
+      }
+      g++;
+    }
+
+    return result;
+
+  }
+#endif
+
+  static
+  inline
+  size_t a_block_sum_atrip(int64_t T, int64_t nv) {
+    const int64_t nv_min_1 = nv - 1, t_plus_1 = T + 1;
+    return t_plus_1 * nv_min_1 * nv
+         + nv_min_1 * t_plus_1
+         - (nv_min_1 * (T * t_plus_1) / 2)
+         - (t_plus_1 * (nv_min_1 * nv) / 2)
+        // do not simplify this expression, only the addition of both parts
+        // is a pair integer, prepare to endure the consequences of
+        // simplifying otherwise
+         + (((T * t_plus_1 * (1 + 2 * T)) / 6) - 3 * ((T * t_plus_1) / 2)) / 2
+         ;
+  }
+
+  static
+  inline
+  int64_t b_block_sum_atrip (int64_t a, int64_t T, int64_t nv) {
+    return nv * ((T - a) + 1)
+         - (T * (T + 1) - a * (a - 1)) / 2
+         - 1;
+  }
+
+  static std::vector<size_t> a_sums;
+  static
+  inline
+  ABCTuple nth_atrip(size_t it, size_t nv) {
+
+    // build the sums if necessary
+    if (!a_sums.size()) {
+      a_sums.resize(nv);
+      for (size_t _i = 0; _i < nv; _i++) {
+        a_sums[_i] = a_block_sum_atrip(_i, nv);
+      }
+    }
+
+    int64_t a = -1, block_a = 0;
+    for (const auto& sum: a_sums) {
+      ++a;
+      if (sum > it) {
+        break;
+      } else {
+        block_a = sum;
+      }
+    }
+
+    // build the b_sums
+    std::vector<int64_t> b_sums(nv - a);
+    for (size_t t = a, i=0; t < nv; t++) {
+      b_sums[i++] = b_block_sum_atrip(a, t, nv);
+    }
+    int64_t b = a - 1, block_b = block_a;
+    for (const auto& sum: b_sums) {
+      ++b;
+      if (sum + block_a > it) {
+        break;
+      } else {
+        block_b = sum + block_a;
+      }
+    }
+
+
+    const int64_t
+      c = b + it - block_b + (a == b);
+
+    return {(size_t)a, (size_t)b, (size_t)c};
+
+  }
+
+  static
+  inline
+  ABCTuples nth_atrip_distributed(int64_t it, size_t nv, size_t np) {
+
+    // If we are getting the previous tuples in the first iteration,
+    // then just return an impossible tuple, different from the FAKE_TUPLE,
+    // because if FAKE_TUPLE is defined as {0,0,0} slices thereof
+    // are actually attainable.
+    //
+    if (it < 0) {
+      ABCTuples result(np, {nv, nv, nv});
+      return result;
+    }
+
+    ABCTuples result(np);
+
+    const size_t
+      // total number of tuples for the problem
+      n = nv * (nv + 1) * (nv + 2) / 6 - nv
+
+      // all ranks should have the same number of tuples_per_rank
+      , tuples_per_rank = n / np + size_t(n % np != 0)
+      ;
+
+
+    for (size_t rank = 0; rank < np; rank++) {
+      const size_t
+        global_iteration = tuples_per_rank * rank + it;
+      result[rank] = nth_atrip(global_iteration, nv);
+    }
+
+    return result;
+  }
+
+
+  template <typename F>
+  static
+  typename Slice<F>::LocalDatabase
+  build_local_database_fake(ABCTuple const& abc_prev,
+                            ABCTuple const& abc,
+                            size_t rank,
+                            SliceUnion<F>* u) {
+
+    typename Slice<F>::LocalDatabase result;
+
+    // vector of type x tuple
+    auto const needed = u->neededSlices(abc);
+    auto const needed_prev = u->neededSlices(abc_prev);
+
+    for (auto const& pair: needed) {
+      auto const type = pair.first;
+      auto const tuple = pair.second;
+      auto const from  = u->rankMap.find(abc, type);
+
+      // Try to find in the previously needed slices
+      // one that exactly matches the tuple.
+      // Not necessarily has to match the type.
+      //
+      // If we find it, then it means that the fake rank
+      // will mark it as recycled. This covers
+      // the finding of Ready slices and Recycled slices.
+      {
+        auto const& it
+          = std::find_if(needed_prev.begin(), needed_prev.end(),
+                         [&tuple, &type](typename Slice<F>::Ty_x_Tu const& o) {
+                           return o.second == tuple;
+                         });
+
+        if (it != needed_prev.end()) {
+          typename Slice<F>::Info info;
+          info.tuple = tuple;
+          info.type = type;
+          info.from = from;
+          info.state = Slice<F>::Recycled;
+          result.push_back({u->name, info});
+          continue;
+        }
+      }
+
+      {
+        typename Slice<F>::Info info;
+        info.type = type;
+        info.tuple = tuple;
+        info.from = from;
+
+        // Handle self sufficiency
+        info.state = rank == from.rank
+                   ? Slice<F>::SelfSufficient
+                   : Slice<F>::Fetch
+                   ;
+        result.push_back({u->name, info});
+        continue;
+      }
+
+    }
+
+    return result;
+
+  }
+
+
+
+  template <typename F>
+  typename Slice<F>::Database
+  naiveDatabase(Unions<F> &unions,
+                size_t nv,
+                size_t np,
+                size_t iteration,
+                MPI_Comm const& c) {
+
+    using Database = typename Slice<F>::Database;
+    Database db;
+
+#ifdef ATRIP_NAIVE_SLOW
+    WITH_CHRONO("db:comm:naive:tuples",
+                const auto tuples = get_nth_naive_tuples(nv,
+                                                         np,
+                                                         iteration);
+                const auto prev_tuples = get_nth_naive_tuples(nv,
+                                                              np,
+                                                              iteration - 1);
+                )
+#else
+    WITH_CHRONO("db:comm:naive:tuples",
+                const auto tuples = nth_atrip_distributed(iteration,
+                                                          nv,
+                                                          np);
+                const auto prev_tuples = nth_atrip_distributed(iteration - 1,
+                                                               nv,
+                                                               np);
+                )
+
+#endif
+
+    for (size_t rank = 0; rank < np; rank++) {
+      auto abc = tuples[rank];
+      typename Slice<F>::LocalDatabase ldb;
+
+      for (auto const& tensor: unions) {
+        if (rank == Atrip::rank) {
+          auto const& tensorDb = tensor->buildLocalDatabase(abc);
+          ldb.insert(ldb.end(), tensorDb.begin(), tensorDb.end());
+        } else {
+          auto const& tensorDb
+            = build_local_database_fake(prev_tuples[rank],
+                                        abc,
+                                        rank,
+                                        tensor);
+          ldb.insert(ldb.end(), tensorDb.begin(), tensorDb.end());
+        }
+      }
+
+      db.insert(db.end(), ldb.begin(), ldb.end());
+
+    }
+
+    return db;
+  }
+
+  template
+  typename Slice<double>::Database
+  naiveDatabase<double>(Unions<double> &unions,
+                size_t nv,
+                size_t np,
+                size_t iteration,
+                MPI_Comm const& c);
+
+  template
+  typename Slice<Complex>::Database
+  naiveDatabase<Complex>(Unions<Complex> &unions,
+                size_t nv,
+                size_t np,
+                size_t iteration,
+                MPI_Comm const& c);
+
+}  // namespace atrip

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
@@ -156,7 +73,6 @@ namespace cuda {
    * in order to have an argument in the signature of
    * the function  that helps the compiler know which
    * instantiation it should take.
-   * 
    */
   template <typename F, reordering_t R>
   struct reorder_proxy {};
@@ -180,162 +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)
-                , denominator(epsabc - ei - ej - ek)
+                , facij = i == j ? F{0.5} : F{1.0}
+                , 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]))
-                , 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 )
+                , 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))
                 ;
-              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))
-              , denominator(epsabc - ei - ej - ek)
+              , facij ( i==j ? F{0.5} : F{1.0})
+              , 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]))
-              , 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 )
+              , 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))
               ;
-              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
-  ( double const epsabc
+__MAYBE_GLOBAL__
+void getEnergyDistinct
+  ( DataFieldType<double> const epsabc
   , size_t const No
-  , double* const epsi
-  , double* const Tijk
-  , double* const Zijk
+  , DataFieldType<double>* const epsi
+  , DataFieldType<double>* const Tijk
+  , DataFieldType<double>* const Zijk
+  , DataFieldType<double>* energy
   );
 
 template
-double getEnergySame
-  ( double const epsabc
+__MAYBE_GLOBAL__
+void getEnergySame
+  ( DataFieldType<double> const epsabc
   , size_t const No
-  , double* const epsi
-  , double* const Tijk
-  , double* const Zijk
+  , DataFieldType<double>* const epsi
+  , DataFieldType<double>* const Tijk
+  , DataFieldType<double>* const Zijk
+  , DataFieldType<double>* energy
   );
 
 // instantiate Complex
 template
-double getEnergyDistinct
-  ( Complex const epsabc
+__MAYBE_GLOBAL__
+void getEnergyDistinct
+  ( DataFieldType<Complex> const epsabc
   , size_t const No
-  , Complex* const epsi
-  , Complex* const Tijk
-  , Complex* const Zijk
+  , DataFieldType<Complex>* const epsi
+  , DataFieldType<Complex>* const Tijk
+  , DataFieldType<Complex>* const Zijk
+  , DataFieldType<double>* energy
   );
 
 template
-double getEnergySame
-  ( Complex const epsabc
+__MAYBE_GLOBAL__
+void getEnergySame
+  ( DataFieldType<Complex> const epsabc
   , size_t const No
-  , Complex* const epsi
-  , Complex* const Tijk
-  , Complex* const Zijk
+  , DataFieldType<Complex>* const epsi
+  , DataFieldType<Complex>* const Tijk
+  , DataFieldType<Complex>* const Zijk
+  , DataFieldType<double>* energy
   );
 // Energy:3 ends here
 
@@ -361,18 +320,26 @@ double getEnergySame
       const size_t ijk = i + j*No + k*NoNo;
 
 #ifdef HAVE_CUDA
-#  define GO(__TPH, __VABIJ)                                          \
-    {                                                                 \
-      const DataFieldType<F> product                                  \
-            = cuda::multiply<DataFieldType<F>>((__TPH), (__VABIJ));   \
-      cuda::sum_in_place<DataFieldType<F>>(&Zijk[ijk], &product);     \
-    }
+
+#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)
+
 #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
   }
@@ -434,8 +401,12 @@ 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]
               , NoNo = No*No
               ;
@@ -444,14 +415,14 @@ double getEnergySame
 
 #if defined(ATRIP_USE_DGEMM)
 #if defined(HAVE_CUDA)
-#define REORDER(__II, __JJ, __KK)                           \
-  reorder<<<bs, ths>>>(reorder_proxy<                       \
-                       DataFieldType<F>,                    \
-                       __II ## __JJ ## __KK                 \
-                       >{},                                 \
-                       No,                                  \
-                       Tijk,                                \
-                       _t_buffer);
+#define REORDER(__II, __JJ, __KK)               \
+  reorder<<<bs, ths>>>(reorder_proxy<           \
+                       DataFieldType<F>,        \
+                       __II ## __JJ ## __KK     \
+                       >{},                     \
+                       No,                      \
+                       Tijk,                    \
+                       _t_buffer)
 #define DGEMM_PARTICLES(__A, __B)               \
   atrip::xgemm<F>("T",                          \
                   "N",                          \
@@ -481,11 +452,18 @@ double getEnergySame
                   _t_buffer,                    \
                   (int const*)&NoNo             \
                   )
-#define MAYBE_CONJ(_conj, _buffer)                                      \
-  cuda::maybeConjugate<<<                                               \
-    Atrip::kernelDimensions.ooo.blocks,                                 \
-    Atrip::kernelDimensions.ooo.threads                                 \
-    >>>((DataFieldType<F>*)_conj, (DataFieldType<F>*)_buffer, NoNoNo);
+#define MAYBE_CONJ(_conj, _buffer)                                \
+  do {                                                            \
+    acc::maybeConjugate<<<                                        \
+                                                                  \
+                            Atrip::kernelDimensions.ooo.blocks,   \
+                                                                  \
+                              Atrip::kernelDimensions.ooo.threads \
+                                                                  \
+                        >>>((DataFieldType<F>*)_conj,             \
+                            (DataFieldType<F>*)_buffer,           \
+                            NoNoNo);                              \
+  } while (0)
 
 
 // END CUDA ////////////////////////////////////////////////////////////////////
@@ -500,7 +478,9 @@ double getEnergySame
 #define REORDER(__II, __JJ, __KK)               \
   reorder(reorder_proxy<DataFieldType<F>,       \
           __II ## __JJ ## __KK >{},             \
-          No, Tijk, _t_buffer);
+          No,                                   \
+          Tijk,                                 \
+          _t_buffer)
 #define DGEMM_PARTICLES(__A, __B)               \
   atrip::xgemm<F>("T",                          \
                   "N",                          \
@@ -531,29 +511,37 @@ double getEnergySame
                   _t_buffer,                    \
                   (int const*)&NoNo             \
                   )
-#define MAYBE_CONJ(_conj, _buffer)                \
-  for (size_t __i = 0; __i < NoNoNo; ++__i)       \
-    _conj[__i] = maybeConjugate<F>(_buffer[__i]);
+#define MAYBE_CONJ(_conj, _buffer)              \
+  do {                                          \
+    for (size_t __i = 0; __i < NoNoNo; ++__i) { \
+      _conj[__i]                                \
+        = maybeConjugate<F>(_buffer[__i]);      \
+    }                                           \
+  } while (0)
 #endif
 
     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>* _vhhh;
-    WITH_CHRONO("double:cuda:alloc",
-    _CHECK_CUDA_SUCCESS("Allocating _t_buffer",
-                        cuMemAlloc((CUdeviceptr*)&_t_buffer,
-                                   NoNoNo * sizeof(DataFieldType<F>)));
-    _CHECK_CUDA_SUCCESS("Allocating _vhhh",
-                        cuMemAlloc((CUdeviceptr*)&_vhhh,
-                                   NoNoNo * sizeof(DataFieldType<F>)));
-                )
+//    DataFieldType<F>* _t_buffer;
+//    DataFieldType<F>* _vhhh;
+//    WITH_CHRONO("double:cuda:alloc",
+//    _CHECK_CUDA_SUCCESS("Allocating _t_buffer",
+//                        cuMemAlloc((CUdeviceptr*)&_t_buffer,
+//                                   NoNoNo * sizeof(DataFieldType<F>)));
+//    _CHECK_CUDA_SUCCESS("Allocating _vhhh",
+//                        cuMemAlloc((CUdeviceptr*)&_vhhh,
+//                                   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;
-    //cuda::zeroing<<<bs, ths>>>((DataFieldType<F>*)_t_buffer, NoNoNo);
-    //cuda::zeroing<<<bs, ths>>>((DataFieldType<F>*)_vhhh, NoNoNo);
+    acc::zeroing<<<bs, ths>>>((DataFieldType<F>*)_t_buffer, NoNoNo);
+    acc::zeroing<<<bs, ths>>>((DataFieldType<F>*)_vhhh, NoNoNo);
+#endif
+
 #else
     DataFieldType<F>* _t_buffer = (DataFieldType<F>*)malloc(NoNoNo * sizeof(F));
     DataFieldType<F>* _vhhh = (DataFieldType<F>*)malloc(NoNoNo * sizeof(F));
@@ -565,55 +553,65 @@ double getEnergySame
 #endif
 
     // Set Tijk to zero
-#ifdef HAVE_CUDA
+#if defined(HAVE_CUDA) && !defined(ATRIP_ONLY_DGEMM)
     WITH_CHRONO("double:reorder",
-                //cuda::zeroing<<<bs, ths>>>((DataFieldType<F>*)Tijk,
-                //NoNoNo);
+                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)
+#undef MAYBE_CONJ
+#undef REORDER
+#define MAYBE_CONJ(a, b) do {} while(0)
+#define REORDER(i, j, k) do {} while(0)
+#endif /* defined(ATRIP_ONLY_DGEMM) */
 
     // HOLES
     WITH_CHRONO("doubles:holes",
       {
         // VhhhC[i + k*No + L*NoNo] * TABhh[L + j*No]; H1
-        //MAYBE_CONJ(_vhhh, VhhhC)
+        MAYBE_CONJ(_vhhh, VhhhC);
         WITH_CHRONO("doubles:holes:1",
                     DGEMM_HOLES(_vhhh, TABhh, "N");
-                    //REORDER(I, K, J)
+                    REORDER(I, K, J);
         )
         // VhhhC[j + k*No + L*NoNo] * TABhh[i + L*No]; H0
         WITH_CHRONO("doubles:holes:2",
                     DGEMM_HOLES(_vhhh, TABhh, "T");
-                    //REORDER(J, K, I)
+                    REORDER(J, K, I);
         )
 
         // VhhhB[i + j*No + L*NoNo] * TAChh[L + k*No]; H5
-        //MAYBE_CONJ(_vhhh, VhhhB)
+          MAYBE_CONJ(_vhhh, VhhhB);
         WITH_CHRONO("doubles:holes:3",
                     DGEMM_HOLES(_vhhh, TAChh, "N");
-                    //REORDER(I, J, K)
+                    REORDER(I, J, K);
         )
         // VhhhB[k + j*No + L*NoNo] * TAChh[i + L*No]; H3
         WITH_CHRONO("doubles:holes:4",
                     DGEMM_HOLES(_vhhh, TAChh, "T");
-                    //REORDER(K, J, I)
+                    REORDER(K, J, I);
         )
 
         // VhhhA[j + i*No + L*NoNo] * TBChh[L + k*No]; H1
-        //MAYBE_CONJ(_vhhh, VhhhA)
+          MAYBE_CONJ(_vhhh, VhhhA);
         WITH_CHRONO("doubles:holes:5",
                     DGEMM_HOLES(_vhhh, TBChh, "N");
-                    //REORDER(J, I, K)
+                    REORDER(J, I, K);
         )
         // VhhhA[k + i*No + L*NoNo] * TBChh[j + L*No]; H4
         WITH_CHRONO("doubles:holes:6",
                     DGEMM_HOLES(_vhhh, TBChh, "T");
-                    //REORDER(K, I, J)
+                    REORDER(K, I, J);
         )
       }
     )
@@ -625,32 +623,32 @@ double getEnergySame
         // TAphh[E + i*Nv + j*NoNv] * VBCph[E + k*Nv]; P0
         WITH_CHRONO("doubles:particles:1",
                     DGEMM_PARTICLES(TAphh, VBCph);
-                    //REORDER(I, J, K)
+                    REORDER(I, J, K);
         )
         // TAphh[E + i*Nv + k*NoNv] * VCBph[E + j*Nv]; P3
         WITH_CHRONO("doubles:particles:2",
                     DGEMM_PARTICLES(TAphh, VCBph);
-                    //REORDER(I, K, J)
+                    REORDER(I, K, J);
         )
         // TCphh[E + k*Nv + i*NoNv] * VABph[E + j*Nv]; P5
         WITH_CHRONO("doubles:particles:3",
                     DGEMM_PARTICLES(TCphh, VABph);
-                    //REORDER(K, I, J)
+                    REORDER(K, I, J);
         )
         // TCphh[E + k*Nv + j*NoNv] * VBAph[E + i*Nv]; P2
         WITH_CHRONO("doubles:particles:4",
                     DGEMM_PARTICLES(TCphh, VBAph);
-                    //REORDER(K, J, I)
+                    REORDER(K, J, I);
         )
         // TBphh[E + j*Nv + i*NoNv] * VACph[E + k*Nv]; P1
         WITH_CHRONO("doubles:particles:5",
                     DGEMM_PARTICLES(TBphh, VACph);
-                    //REORDER(J, I, K)
+                    REORDER(J, I, K);
         )
         // TBphh[E + j*Nv + k*NoNv] * VCAph[E + i*Nv]; P4
         WITH_CHRONO("doubles:particles:6",
                     DGEMM_PARTICLES(TBphh, VCAph);
-                    //REORDER(J, K, I)
+                    REORDER(J, K, I);
         )
       }
     )
@@ -659,16 +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",
-                        cuCtxSynchronize());
-    _CHECK_CUDA_SUCCESS("Freeing _vhhh",
-                        cuMemFree((CUdeviceptr)_vhhh));
-    _CHECK_CUDA_SUCCESS("Freeing _t_buffer",
-                        cuMemFree((CUdeviceptr)_t_buffer));
+    //_CHECK_CUDA_SUCCESS("Synchronizing",
+    //                    cuCtxSynchronize());
+    //_CHECK_CUDA_SUCCESS("Freeing _vhhh",
+    //                    cuMemFree((CUdeviceptr)_vhhh));
+    //_CHECK_CUDA_SUCCESS("Freeing _t_buffer",
+    //                    cuMemFree((CUdeviceptr)_t_buffer));
 #else
     free(_vhhh);
     free(_t_buffer);
-#endif
+#endif  /* defined(HAVE_CUDA) */
   }
 
   #undef REORDER
@@ -719,7 +717,7 @@ double getEnergySame
       }
 
     }
-#endif
+#endif /* defined(ATRIP_USE_DGEMM) */
   }
 
 
@@ -751,6 +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
@@ -779,6 +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
 

src/atrip/Tuples.cxx

@@ -0,0 +1,464 @@
+#include <atrip/Tuples.hpp>
+#include <atrip/Atrip.hpp>
+
+namespace atrip {
+
+template <typename A>
+static A unique(A const &xs) {
+  auto result = xs;
+  std::sort(std::begin(result), std::end(result));
+  auto const& last = std::unique(std::begin(result), std::end(result));
+  result.erase(last, std::end(result));
+  return result;
+}
+
+
+std::vector<std::string> getNodeNames(MPI_Comm comm){
+  int rank, np;
+  MPI_Comm_rank(comm, &rank);
+  MPI_Comm_size(comm, &np);
+
+  std::vector<std::string> nodeList(np);
+  char nodeName[MPI_MAX_PROCESSOR_NAME];
+  char *nodeNames = (char*)malloc(np * MPI_MAX_PROCESSOR_NAME);
+  std::vector<int> nameLengths(np)
+                 , off(np)
+                 ;
+  int nameLength;
+  MPI_Get_processor_name(nodeName, &nameLength);
+  MPI_Allgather(&nameLength,
+                1,
+                MPI_INT,
+                nameLengths.data(),
+                1,
+                MPI_INT,
+                comm);
+  for (int i(1); i < np; i++)
+    off[i] = off[i-1] + nameLengths[i-1];
+  MPI_Allgatherv(nodeName,
+                 nameLengths[rank],
+                 MPI_BYTE,
+                 nodeNames,
+                 nameLengths.data(),
+                 off.data(),
+                 MPI_BYTE,
+                 comm);
+  for (int i(0); i < np; i++) {
+    std::string const s(&nodeNames[off[i]], nameLengths[i]);
+    nodeList[i] = s;
+  }
+  std::free(nodeNames);
+  return nodeList;
+}
+
+
+
+std::vector<RankInfo>
+getNodeInfos(std::vector<string> const& nodeNames) {
+  std::vector<RankInfo> result;
+  auto const uniqueNames = unique(nodeNames);
+  auto const index = [&uniqueNames](std::string const& s) {
+    auto const& it = std::find(uniqueNames.begin(), uniqueNames.end(), s);
+    return std::distance(uniqueNames.begin(), it);
+  };
+  std::vector<size_t> localRanks(uniqueNames.size(), 0);
+  size_t globalRank = 0;
+  for (auto const& name: nodeNames) {
+    const size_t nodeId = index(name);
+    result.push_back({name,
+                      nodeId,
+                      globalRank++,
+                      localRanks[nodeId]++,
+                      (size_t)
+                      std::count(nodeNames.begin(),
+                                 nodeNames.end(),
+                                 name)
+                      });
+  }
+  return result;
+}
+
+ClusterInfo
+getClusterInfo(MPI_Comm comm) {
+  auto const names = getNodeNames(comm);
+  auto const rankInfos = getNodeInfos(names);
+
+  return ClusterInfo {
+    unique(names).size(),
+    names.size(),
+    rankInfos[0].ranksPerNode,
+    rankInfos
+  };
+
+}
+
+
+
+ABCTuples getTuplesList(size_t Nv, size_t rank, size_t np) {
+
+  const size_t
+    // total number of tuples for the problem
+       n = Nv * (Nv + 1) * (Nv + 2) / 6 - Nv
+
+    // all ranks should have the same number of tuples_per_rank
+    , tuples_per_rank = n / np + size_t(n % np != 0)
+
+    // start index for the global tuples list
+    , start = tuples_per_rank * rank
+
+    // end index for the global tuples list
+    , end = tuples_per_rank * (rank + 1)
+    ;
+
+  LOG(1,"Atrip") << "tuples_per_rank = " << tuples_per_rank << "\n";
+  WITH_RANK << "start, end = " << start << ", " << end << "\n";
+  ABCTuples result(tuples_per_rank, FAKE_TUPLE);
+
+  for (size_t a(0), r(0), g(0); a < Nv; a++)
+  for (size_t b(a);             b < Nv; b++)
+  for (size_t c(b);             c < Nv; c++){
+    if ( a == b && b == c ) continue;
+    if ( start <= g && g < end) result[r++] = {a, b, c};
+    g++;
+  }
+
+  return result;
+
+}
+
+
+ABCTuples getAllTuplesList(const size_t Nv) {
+  const size_t n = Nv * (Nv + 1) * (Nv + 2) / 6 - Nv;
+  ABCTuples result(n);
+
+  for (size_t a(0), u(0); a < Nv; a++)
+  for (size_t b(a); b < Nv; b++)
+  for (size_t c(b); c < Nv; c++){
+    if ( a == b && b == c ) continue;
+    result[u++] = {a, b, c};
+  }
+
+  return result;
+}
+
+
+ABCTuples atrip::NaiveDistribution::getTuples(size_t Nv, MPI_Comm universe) {
+  int rank, np;
+  MPI_Comm_rank(universe, &rank);
+  MPI_Comm_size(universe, &np);
+  return getTuplesList(Nv, (size_t)rank, (size_t)np);
+}
+
+
+
+
+namespace group_and_sort {
+
+inline
+size_t isOnNode(size_t tuple, size_t nNodes) { return tuple % nNodes; }
+
+std::vector<size_t> getTupleNodes(ABCTuple const& t, size_t nNodes) {
+  std::vector<size_t>
+    nTuple = { isOnNode(t[0], nNodes)
+             , isOnNode(t[1], nNodes)
+             , isOnNode(t[2], nNodes)
+             };
+  return unique(nTuple);
+}
+
+
+ABCTuples specialDistribution(Info const& info, ABCTuples const& allTuples) {
+
+  ABCTuples nodeTuples;
+  size_t const nNodes(info.nNodes);
+
+  std::vector<ABCTuples>
+      container1d(nNodes)
+    , container2d(nNodes * nNodes)
+    , container3d(nNodes * nNodes * nNodes)
+    ;
+
+  WITH_DBG if (info.nodeId == 0)
+    std::cout << "\tGoing through all "
+              << allTuples.size()
+              << " tuples in "
+              << nNodes
+              << " nodes\n";
+
+  // build container-n-d's
+  for (auto const& t: allTuples) {
+    // one which node(s) are the tuple elements located...
+    // put them into the right container
+    auto const _nodes = getTupleNodes(t, nNodes);
+
+    switch (_nodes.size()) {
+      case 1:
+        container1d[_nodes[0]].push_back(t);
+        break;
+      case 2:
+        container2d[ _nodes[0]
+                   + _nodes[1] * nNodes
+                   ].push_back(t);
+        break;
+      case 3:
+        container3d[ _nodes[0]
+                   + _nodes[1] * nNodes
+                   + _nodes[2] * nNodes * nNodes
+                   ].push_back(t);
+        break;
+    }
+
+  }
+
+  WITH_DBG if (info.nodeId == 0)
+    std::cout << "\tBuilding 1-d containers\n";
+  // DISTRIBUTE 1-d containers
+  // every tuple which is only located at one node belongs to this node
+  {
+    auto const& _tuples = container1d[info.nodeId];
+    nodeTuples.resize(_tuples.size(), INVALID_TUPLE);
+    std::copy(_tuples.begin(), _tuples.end(), nodeTuples.begin());
+  }
+
+  WITH_DBG if (info.nodeId == 0)
+    std::cout << "\tBuilding 2-d containers\n";
+  // DISTRIBUTE 2-d containers
+  //the tuples which are located at two nodes are half/half given to these nodes
+  for (size_t yx = 0; yx < container2d.size(); yx++) {
+
+    auto const& _tuples = container2d[yx];
+      const
+    size_t idx = yx % nNodes
+         // remeber: yx = idy * nNodes + idx
+         , idy = yx / nNodes
+         , n_half = _tuples.size() / 2
+         , size = nodeTuples.size()
+         ;
+
+    size_t nbeg, nend;
+    if (info.nodeId == idx) {
+      nbeg = 0 * n_half;
+      nend = n_half;
+    } else if (info.nodeId == idy) {
+      nbeg = 1 * n_half;
+      nend = _tuples.size();
+    } else {
+      // either idx or idy is my node
+      continue;
+    }
+
+    size_t const nextra = nend - nbeg;
+    nodeTuples.resize(size + nextra, INVALID_TUPLE);
+    std::copy(_tuples.begin() + nbeg,
+              _tuples.begin() + nend,
+              nodeTuples.begin() + size);
+
+  }
+
+  WITH_DBG if (info.nodeId == 0)
+    std::cout << "\tBuilding 3-d containers\n";
+  // DISTRIBUTE 3-d containers
+  for (size_t zyx = 0; zyx < container3d.size(); zyx++) {
+    auto const& _tuples = container3d[zyx];
+
+      const
+    size_t idx = zyx % nNodes
+         , idy = (zyx / nNodes) % nNodes
+         // remember: zyx = idx + idy * nNodes + idz * nNodes^2
+         , idz = zyx / nNodes / nNodes
+         , n_third = _tuples.size() / 3
+         , size = nodeTuples.size()
+         ;
+
+    size_t nbeg, nend;
+    if (info.nodeId == idx) {
+      nbeg = 0 * n_third;
+      nend = 1 * n_third;
+    } else if (info.nodeId == idy) {
+      nbeg = 1 * n_third;
+      nend = 2 * n_third;
+    } else if (info.nodeId == idz) {
+      nbeg = 2 * n_third;
+      nend = _tuples.size();
+    } else {
+      // either idx or idy or idz is my node
+      continue;
+    }
+
+    size_t const nextra = nend - nbeg;
+    nodeTuples.resize(size + nextra, INVALID_TUPLE);
+    std::copy(_tuples.begin() + nbeg,
+              _tuples.begin() + nend,
+              nodeTuples.begin() + size);
+
+  }
+
+
+  WITH_DBG if (info.nodeId == 0) std::cout << "\tswapping tuples...\n";
+  /*
+   *  sort part of group-and-sort algorithm
+   *  every tuple on a given node is sorted in a way that
+   *  the 'home elements' are the fastest index.
+   *  1:yyy 2:yyn(x) 3:yny(x) 4:ynn(x) 5:nyy 6:nyn(x) 7:nny 8:nnn
+   */
+  for (auto &nt: nodeTuples){
+    if ( isOnNode(nt[0], nNodes) == info.nodeId ){ // 1234
+      if ( isOnNode(nt[2], nNodes) != info.nodeId ){ // 24
+        size_t const x(nt[0]);
+        nt[0] = nt[2];         // switch first and last
+        nt[2] = x;
+      }
+      else if ( isOnNode(nt[1], nNodes) != info.nodeId){ // 3
+        size_t const x(nt[0]);
+        nt[0] = nt[1];         // switch first two
+        nt[1] = x;
+      }
+    } else {
+      if ( isOnNode(nt[1], nNodes) == info.nodeId   // 56
+        && isOnNode(nt[2], nNodes) != info.nodeId
+        ) { // 6
+        size_t const x(nt[1]);
+        nt[1] = nt[2];         // switch last two
+        nt[2] = x;
+      }
+    }
+  }
+
+  WITH_DBG if (info.nodeId == 0) std::cout << "\tsorting list of tuples...\n";
+  //now we sort the list of tuples
+  std::sort(nodeTuples.begin(), nodeTuples.end());
+
+  WITH_DBG if (info.nodeId == 0) std::cout << "\trestoring tuples...\n";
+  // we bring the tuples abc back in the order a<b<c
+  for (auto &t: nodeTuples)  std::sort(t.begin(), t.end());
+
+#if ATRIP_DEBUG > 1
+  WITH_DBG if (info.nodeId == 0)
+  std::cout << "checking for validity of " << nodeTuples.size() << std::endl;
+  const bool anyInvalid
+    = std::any_of(nodeTuples.begin(),
+                  nodeTuples.end(),
+                  [](ABCTuple const& t) { return t == INVALID_TUPLE; });
+  if (anyInvalid) throw "Some tuple is invalid in group-and-sort algorithm";
+#endif
+
+  WITH_DBG if (info.nodeId == 0) std::cout << "\treturning tuples...\n";
+  return nodeTuples;
+
+}
+
+
+  std::vector<ABCTuple> main(MPI_Comm universe, size_t Nv) {
+
+    int rank, np;
+    MPI_Comm_rank(universe, &rank);
+    MPI_Comm_size(universe, &np);
+
+    std::vector<ABCTuple> result;
+
+    auto const nodeNames(getNodeNames(universe));
+    size_t const nNodes = unique(nodeNames).size();
+    auto const nodeInfos = getNodeInfos(nodeNames);
+
+    // We want to construct a communicator which only contains of one
+    // element per node
+    bool const computeDistribution
+      = nodeInfos[rank].localRank == 0;
+
+    std::vector<ABCTuple>
+      nodeTuples
+      = computeDistribution
+      ? specialDistribution(Info{nNodes, nodeInfos[rank].nodeId},
+                            getAllTuplesList(Nv))
+      : std::vector<ABCTuple>()
+      ;
+
+    LOG(1,"Atrip") << "got nodeTuples\n";
+
+    // now we have to send the data from **one** rank on each node
+    // to all others ranks of this node
+    const
+      int color = nodeInfos[rank].nodeId,
+      key = nodeInfos[rank].localRank
+      ;
+
+
+    MPI_Comm INTRA_COMM;
+    MPI_Comm_split(universe, color, key, &INTRA_COMM);
+    // Main:1 ends here
+
+    // [[file:~/cuda/atrip/atrip.org::*Main][Main:2]]
+    size_t const
+      tuplesPerRankLocal
+      = nodeTuples.size() / nodeInfos[rank].ranksPerNode
+      + size_t(nodeTuples.size() % nodeInfos[rank].ranksPerNode != 0)
+      ;
+
+    size_t tuplesPerRankGlobal;
+
+    MPI_Reduce(&tuplesPerRankLocal,
+               &tuplesPerRankGlobal,
+               1,
+               MPI_UINT64_T,
+               MPI_MAX,
+               0,
+               universe);
+
+    MPI_Bcast(&tuplesPerRankGlobal,
+              1,
+              MPI_UINT64_T,
+              0,
+              universe);
+
+    LOG(1,"Atrip") << "Tuples per rank: " << tuplesPerRankGlobal << "\n";
+    LOG(1,"Atrip") << "ranks per node " << nodeInfos[rank].ranksPerNode << "\n";
+    LOG(1,"Atrip") << "#nodes " << nNodes << "\n";
+    // Main:2 ends here
+
+    // [[file:~/cuda/atrip/atrip.org::*Main][Main:3]]
+    size_t const totalTuples
+      = tuplesPerRankGlobal * nodeInfos[rank].ranksPerNode;
+
+    if (computeDistribution) {
+      // pad with FAKE_TUPLEs
+      nodeTuples.insert(nodeTuples.end(),
+                        totalTuples - nodeTuples.size(),
+                        FAKE_TUPLE);
+    }
+    // Main:3 ends here
+
+    // [[file:~/cuda/atrip/atrip.org::*Main][Main:4]]
+    {
+      // construct mpi type for abctuple
+      MPI_Datatype MPI_ABCTUPLE;
+      MPI_Type_vector(nodeTuples[0].size(), 1, 1, MPI_UINT64_T, &MPI_ABCTUPLE);
+      MPI_Type_commit(&MPI_ABCTUPLE);
+
+      LOG(1,"Atrip") << "scattering tuples \n";
+
+      result.resize(tuplesPerRankGlobal);
+      MPI_Scatter(nodeTuples.data(),
+                  tuplesPerRankGlobal,
+                  MPI_ABCTUPLE,
+                  result.data(),
+                  tuplesPerRankGlobal,
+                  MPI_ABCTUPLE,
+                  0,
+                  INTRA_COMM);
+
+      MPI_Type_free(&MPI_ABCTUPLE);
+
+    }
+
+    return result;
+
+  }
+
+
+ABCTuples Distribution::getTuples(size_t Nv, MPI_Comm universe) {
+  return main(universe, Nv);
+}
+
+
+}  // namespace group_and_sort
+}  // namespace atrip

tools/configure-benches.sh

@@ -0,0 +1,183 @@
+#!/usr/bin/env bash
+# Copyright (C) 2022 by Alejandro Gallo <aamsgallo@gmail.com>
+
+set -eu
+
+flags=("${@}")
+PROJECTS=()
+
+############################################################
+#
+## Check root directory
+#
+root_project=$(git rev-parse --show-toplevel)
+configure=$root_project/configure
+if [[ $(basename $PWD) == $(basename $root_project) ]]; then
+    cat <<EOF
+
+ You are trying to build in the root directory, create a build folder
+ and then configure.
+
+                             mkdir build
+                             cd build
+                             $(readlink -f $0)
+
+EOF
+    exit 1
+fi
+
+[[ -f $configure ]] || {
+    cat <<EOF
+  No configure script at $configure create it with bootstrap.sh or
+
+    autoreconf -vif
+
+EOF
+    exit 1
+}
+
+############################################################
+#
+## Create configuration function
+#
+
+create_config () {
+    file=$1
+    name=$2
+    PROJECTS=(${PROJECTS[@]} "$name")
+    mkdir -p $name
+    cd $name
+    echo "> creating: $name"
+    cat <<SH > configure
+#!/usr/bin/env bash
+# creator: $0
+# date: $(date)
+
+$root_project/configure $(cat $file | paste -s) \\
+$(for word in "${flags[@]}"; do
+    printf "  \"%s\"" "$word";
+  done)
+
+
+exit 0
+SH
+    chmod +x configure
+    cd - > /dev/null
+}
+
+############################################################
+# begin doc
+#
+# - default ::
+#   This configuration uses a CPU code with dgemm
+#   and without computing slices.
+#
+# end doc
+
+tmp=`mktemp`
+cat <<EOF > $tmp
+--disable-slice
+EOF
+
+create_config $tmp default
+rm $tmp
+
+# begin doc
+#
+# - only-dgemm ::
+#   This only runs the computation part that involves dgemms.
+#
+# end doc
+
+tmp=`mktemp`
+cat <<EOF > $tmp
+--disable-slice
+--enable-only-dgemm
+EOF
+
+create_config $tmp only-dgemm
+rm $tmp
+
+# begin doc
+#
+# - 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.
+#
+# end doc
+
+tmp=`mktemp`
+cat <<EOF > $tmp
+--enable-cuda
+--enable-sources-in-gpu
+--enable-cuda-aware-mpi
+--enable-only-dgemm
+--disable-slice
+EOF
+
+create_config $tmp cuda-slices-on-gpu-only-dgemm
+rm $tmp
+
+############################################################
+#
+## Create makefile
+#
+
+cat <<MAKE > Makefile
+
+all: configure do
+do: configure
+
+configure: ${PROJECTS[@]/%/\/Makefile}
+
+%/Makefile: %/configure
+	cd \$* && ./configure
+
+do: ${PROJECTS[@]/%/\/src\/libatrip.a}
+
+
+%/src/libatrip.a:
+	cd \$* && \$(MAKE)
+
+
+.PHONY: configure do all
+MAKE
+
+cat <<EOF
+
+Now you can do
+
+  make all
+
+or go into one of the directories
+   ${PROJECTS[@]}
+and do
+  ./configure
+  make
+
+EOF
+
+## Emacs stuff
+# Local Variables:
+# eval: (outline-minor-mode)
+# outline-regexp: "############################################################"
+# End: