.dir-locals.el

@@ -1,22 +0,0 @@
-;;; 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))))

.github/workflows/main.yml

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

.gitignore

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

README.org

@@ -26,86 +26,3 @@ 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.
-- 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

bench/Makefile.am

@@ -1,42 +1,25 @@
 AUTOMAKE_OPTIONS = subdir-objects
 include $(top_srcdir)/atrip.mk
 
-AM_CPPFLAGS = -I$(top_srcdir)/include/ -I$(top_srcdir) $(CTF_CPPFLAGS)
+AM_CPPFLAGS = -I$(top_srcdir)/include/ $(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
-ATRIP_CTF = $(CTF_BUILD_PATH)/lib/libctf.a
+test_main_LDADD += $(CTF_BUILD_PATH)/lib/libctf.a
 else
-ATRIP_CTF = @LIBCTF_LD_LIBRARY_PATH@/libctf.a
+test_main_LDADD += @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
-AM_CPPFLAGS += $(CUDA_CXXFLAGS)
-BENCHES_LDADD += $(CUDA_LDFLAGS)
+test_main_CXXFLAGS = $(CUDA_CXXFLAGS)
+test_main_LDADD += $(CUDA_LDFLAGS)
 
 AM_CXXFLAGS = $(CUDA_CXXFLAGS)
 AM_LDFLAGS += $(CUDA_LDFLAGS)

bench/tuples-distribution.cxx

@@ -1,443 +0,0 @@
-#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

@@ -1,12 +0,0 @@
-#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

@@ -21,6 +21,26 @@ 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],
@@ -54,53 +74,13 @@ 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_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.])
-
+            [AC_DEFINE([ATRIP_DEBUG],[1],[Atrip debug level])]
+            )
 
 dnl -----------------------------------------------------------------------
 
@@ -164,7 +144,8 @@ AC_TYPE_SIZE_T
 dnl -----------------------------------------------------------------------
 dnl CHECK CTF
 if test xYES = x${BUILD_CTF}; then
-  AC_MSG_WARN([You will have to do make ctf before building the project.])
+  AC_MSG_WARN([Sorry, building CTF not supported yet provide a build path
+                      with --with-ctf=path/to/ctf/installation])
 else
   CPPFLAGS="$CPPFLAGS -I${LIBCTF_CPATH}"
   LDFLAGS="$LDFLAGS -L${LIBCTF_LD_LIBRARY_PATH} -lctf"

etc/m4/atrip-def.m4

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

etc/make/ctf.mk

@@ -16,7 +16,7 @@ $(CTF_SRC_PATH)/configure:
 
 # Here make sure that ctf does not builld with CUDA support
 # since it is broken anyways
-#
+# 
 # Also we patch the file kernel.h because it mostl
 # doesn't work when we try to include ctf in a CUDACC
 # compiler code.

etc/nix/openblas.nix

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

etc/nix/vendor-shell.nix

@@ -1,27 +0,0 @@
-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/DatabaseCommunicator.hpp

@@ -1,20 +0,0 @@
-#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::Atrip::rank == 0) std::cout << name << ": "
+#define LOG(level, name) if (Atrip::rank == 0) std::cout << name << ": "
 #endif
 // Macros:2 ends here
 

include/atrip/SliceUnion.hpp

@@ -18,12 +18,6 @@
 #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
 
@@ -201,7 +195,7 @@ template <typename F=double>
                            ;
           if (blank.info.state == Slice<F>::SelfSufficient) {
 #if defined(HAVE_CUDA)
-            const size_t _size = sizeof(F) * sliceSize;
+            const size_t _size = sizeof(F) * sources[from.source].size();
             // TODO: this is code duplication with downstairs
             if (freePointers.size() == 0) {
               std::stringstream stream;
@@ -218,12 +212,12 @@ template <typename F=double>
             WITH_CHRONO("cuda:memcpy:self-sufficient",
                         _CHECK_CUDA_SUCCESS("copying mpi data to device",
                                             cuMemcpyHtoD(blank.data,
-                                                         (void*)SOURCES_DATA(sources[from.source]),
-                                                         sizeof(F) * sliceSize));
+                                                         (void*)sources[from.source].data(),
+                                                         sizeof(F) * sources[from.source].size()));
                         ))
 
 #else
-            blank.data = SOURCES_DATA(sources[from.source]);
+            blank.data = sources[from.source].data();
 #endif
           } else {
             if (freePointers.size() == 0) {
@@ -402,18 +396,15 @@ 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>(sliceSize))
-#endif
+                        std::vector<F>
+                          (std::accumulate(sliceLength.begin(),
+                                           sliceLength.end(),
+                                           1UL, std::multiplies<size_t>())))
               , name(name_)
               , sliceTypes(sliceTypes_)
               , sliceBuffers(nSliceBuffers)
+              //, slices(2 * sliceTypes.size(), Slice<F>{ sources[0].size() })
     { // constructor begin
 
       LOG(0,"Atrip") << "INIT SliceUnion: " << name << "\n";
@@ -421,7 +412,7 @@ template <typename F=double>
       for (auto& ptr: sliceBuffers) {
 #if defined(HAVE_CUDA)
         const CUresult error =
-          cuMemAlloc(&ptr, sizeof(F) * sliceSize);
+          cuMemAlloc(&ptr, sizeof(F) * sources[0].size());
         if (ptr == 0UL) {
           throw "UNSUFICCIENT MEMORY ON THE GRAPHIC CARD FOR FREE POINTERS";
         }
@@ -432,12 +423,12 @@ template <typename F=double>
           throw s.str();
         }
 #else
-        ptr = (DataPtr<F>)malloc(sizeof(F) * sliceSize);
+        ptr = (DataPtr<F>)malloc(sizeof(F) * sources[0].size());
 #endif
       }
 
       slices
-        = std::vector<Slice<F>>(2 * sliceTypes.size(), { sliceSize });
+        = std::vector<Slice<F>>(2 * sliceTypes.size(), { sources[0].size() });
       // TODO: think exactly    ^------------------- about this number
 
       // initialize the freePointers with the pointers to the buffers
@@ -450,12 +441,12 @@ template <typename F=double>
       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] " << sliceSize << "\n";
+      WITH_RANK << "#sources[0] " << sources[0].size() << "\n";
       WITH_RANK << "#freePointers " << freePointers.size() << "\n";
       LOG(1,"Atrip") << "#sliceBuffers " << sliceBuffers.size() << "\n";
       LOG(1,"Atrip") << "GB*" << np << " "
                            << double(sources.size() + sliceBuffers.size())
-                            * sliceSize
+                            * sources[0].size()
                             * 8 * np
                             / 1073741824.0
                            << "\n";
@@ -504,13 +495,14 @@ template <typename F=double>
       if (otherRank == info.from.rank)      sendData_p = false;
       if (!sendData_p) return;
 
-      MPI_Isend((void*)SOURCES_DATA(sources[info.from.source]),
-                sliceSize,
-                traits::mpi::datatypeOf<F>(),
-                otherRank,
-                tag,
-                universe,
-                &request);
+      MPI_Isend( sources[info.from.source].data()
+               , sources[info.from.source].size()
+               , traits::mpi::datatypeOf<F>()
+               , otherRank
+               , tag
+               , universe
+               , &request
+               );
       WITH_CRAZY_DEBUG
       WITH_RANK << "sent to " << otherRank << "\n";
 
@@ -524,26 +516,25 @@ template <typename F=double>
 
       if (Atrip::rank == info.from.rank) return;
 
-      if (slice.info.state == Slice<F>::Fetch) { // if-1
+      if (slice.info.state == Slice<F>::Fetch) {
         // TODO: do it through the slice class
         slice.info.state = Slice<F>::Dispatched;
 #if defined(HAVE_CUDA)
-#  if !defined(ATRIP_CUDA_AWARE_MPI) && defined(ATRIP_SOURCES_IN_GPU)
-#    error "You need CUDA aware MPI to have slices on the GPU"
-#  endif
         slice.mpi_data = (F*)malloc(sizeof(F) * slice.size);
-        MPI_Irecv(slice.mpi_data,
+        MPI_Irecv( slice.mpi_data
 #else
-        MPI_Irecv(slice.data,
+        MPI_Irecv( slice.data
 #endif
-                  slice.size,
-                  traits::mpi::datatypeOf<F>(),
-                  info.from.rank,
-                  tag,
-                  universe,
-                  &slice.request);
-       } // if-1
-    } // receive
+                 , slice.size
+                 , traits::mpi::datatypeOf<F>()
+                 , info.from.rank
+                 , tag
+                 , universe
+                 , &slice.request
+                //, MPI_STATUS_IGNORE
+                 );
+      }
+    }
 
     void unwrapAll(ABCTuple const& abc) {
       for (auto type: sliceTypes) unwrapSlice(type, abc);
@@ -606,12 +597,7 @@ 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

@@ -52,7 +52,43 @@ 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);
+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;
+}
 // Node information:1 ends here
 
 // [[file:~/cuda/atrip/atrip.org::*Node%20information][Node information:2]]
@@ -64,28 +100,118 @@ 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);
+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;
+}
 
 struct ClusterInfo {
   const size_t nNodes, np, ranksPerNode;
   const std::vector<RankInfo> rankInfos;
 };
 
-ClusterInfo getClusterInfo(MPI_Comm comm);
+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
+  };
+
+}
 // 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);
+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;
+
+}
 // Naive list:1 ends here
 
 // [[file:~/cuda/atrip/atrip.org::*Naive%20list][Naive list:2]]
-ABCTuples getAllTuplesList(const size_t Nv);
+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;
+}
 // 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;
+  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);
+  }
 };
 // Naive list:3 ends here
 
@@ -98,12 +224,19 @@ 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);
+size_t isOnNode(size_t tuple, size_t nNodes) { return tuple % 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> 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);
+}
 
 struct Info {
   size_t nNodes;
@@ -112,16 +245,302 @@ struct Info {
 // Utils:1 ends here
 
 // [[file:~/cuda/atrip/atrip.org::*Distribution][Distribution:1]]
-ABCTuples specialDistribution(Info const& info, ABCTuples const& allTuples);
+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;
+
+}
 // Distribution:1 ends here
 
 // [[file:~/cuda/atrip/atrip.org::*Main][Main:1]]
-std::vector<ABCTuple> main(MPI_Comm universe, size_t Nv);
+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;
+
+}
 // 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;
+  ABCTuples getTuples(size_t Nv, MPI_Comm universe) override {
+    return main(universe, Nv);
+  }
 };
 // Interface:1 ends here
 

include/atrip/Unions.hpp

@@ -19,14 +19,8 @@
 namespace atrip {
 
   template <typename F=double>
-  static
   void sliceIntoVector
-#if defined(ATRIP_SOURCES_IN_GPU)
-    ( DataPtr<F> &source
-#else
-    ( std::vector<F> &source
-#endif
-    , size_t sliceSize
+    ( std::vector<F> &v
     , CTF::Tensor<F> &toSlice
     , std::vector<int64_t> const low
     , std::vector<int64_t> const up
@@ -50,30 +44,18 @@ 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);
+    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());
 #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
-
   }
 
 
@@ -98,15 +80,16 @@ 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], this->sliceSize,
-                         to,   {0, 0, 0},    {Nv, No, No},
-                         from, {a, 0, 0, 0}, {a+1, Nv, No, No});
+      sliceIntoVector<F>( this->sources[it]
+                        , to,   {0, 0, 0},    {Nv, No, No}
+                        , from, {a, 0, 0, 0}, {a+1, Nv, No, No}
+                        );
 
     }
 
@@ -135,13 +118,14 @@ 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], this->sliceSize,
-                         to,   {0, 0, 0},    {No, No, No},
-                         from, {0, 0, 0, a}, {No, No, No, a+1});
+      sliceIntoVector<F>( this->sources[it]
+                        , to,   {0, 0, 0},    {No, No, No}
+                        , from, {0, 0, 0, a}, {No, No, No, a+1}
+                        );
 
     }
   };
@@ -169,17 +153,18 @@ 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], this->sliceSize,
-                         to,   {0, 0},       {Nv, No},
-                         from, {a, b, 0, 0}, {a+1, b+1, Nv, No});
+      sliceIntoVector<F>( this->sources[it]
+                        , to,   {0, 0},       {Nv, No}
+                        , from, {a, b, 0, 0}, {a+1, b+1, Nv, No}
+                        );
 
     }
 
@@ -206,17 +191,17 @@ 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], this->sliceSize,
-                         to,   {0, 0},       {No, No},
-                         from, {a, b, 0, 0}, {a+1, b+1, No, No});
+      sliceIntoVector<F>( this->sources[it]
+                        , to,   {0, 0},       {No, No}
+                        , from, {a, b, 0, 0}, {a+1, b+1, No, No}
+                        );
 
 
     }
@@ -246,16 +231,17 @@ 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], this->sliceSize,
-                         to,   {0, 0},       {No, No},
-                         from, {a, b, 0, 0}, {a+1, b+1, No, No});
+      sliceIntoVector<F>( this->sources[it]
+                        , to,   {0, 0},       {No, No}
+                        , from, {a, b, 0, 0}, {a+1, b+1, No, No}
+                        );
 
 
     }

misc/naive-tuples.lisp

@@ -1,228 +0,0 @@
-#+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

@@ -1,5 +1,5 @@
 { compiler ? "gcc"
-, pkgs ? import <nixpkgs> {}
+, pkgs ? import <nixpkgs> {} 
 , mkl ? false
 , cuda ? false
 , docs ? true
@@ -11,8 +11,7 @@ let
     config.allowUnfree = true;
   };
 
-  openblas = import ./etc/nix/openblas.nix { inherit pkgs; };
-  vendor = import ./etc/nix/vendor-shell.nix;
+  openblas = import ./etc/nix/openblas.nix { inherit pkgs; }; 
 
   mkl-pkg = import ./etc/nix/mkl.nix { pkgs = unfree-pkgs; };
   cuda-pkg = if cuda then (import ./cuda.nix { pkgs = unfree-pkgs; }) else {};
@@ -58,15 +57,14 @@ pkgs.mkShell rec {
   buildInputs
     = with pkgs; [
 
-        gdb
         coreutils
-        git
-        vim
+        git vim
 
         openmpi
         llvmPackages.openmp
 
         binutils
+        emacs
         gfortran
 
         gnumake
@@ -86,15 +84,6 @@ 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 ./atrip/Tuples.cxx ./atrip/DatabaseCommunicator.cxx
+libatrip_a_SOURCES = ./atrip/Blas.cxx
 NVCC_FILES = ./atrip/Equations.cxx ./atrip/Complex.cxx ./atrip/Atrip.cxx
 
 if WITH_CUDA

src/atrip/Atrip.cxx

@@ -21,7 +21,6 @@
 #include <atrip/SliceUnion.hpp>
 #include <atrip/Unions.hpp>
 #include <atrip/Checkpoint.hpp>
-#include <atrip/DatabaseCommunicator.hpp>
 
 using namespace atrip;
 #if defined(HAVE_CUDA)
@@ -300,23 +299,9 @@ 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, size_t iteration) -> Database {
+      ] (ABCTuple const& abc, MPI_Comm const& c) -> 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();
         )
@@ -349,8 +334,6 @@ 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
@@ -454,7 +437,6 @@ 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;
@@ -582,7 +564,7 @@ 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, i);
+      const auto db = communicateDatabase(abc, universe);
       WITH_RANK << "__first__:first database communicated \n";
       WITH_RANK << "__first__:first database io phase \n";
       doIOPhase(db);
@@ -597,7 +579,7 @@ 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, i);
+        const auto db = communicateDatabase(*abcNext, universe);
       )
       WITH_CHRONO("db:io",
         doIOPhase(db);
@@ -646,9 +628,6 @@ Atrip::Output Atrip::run(Atrip::Input<F> const& in) {
 
     // COMPUTE SINGLES %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% {{{1
     OCD_Barrier(universe);
-#if defined(ATRIP_ONLY_DGEMM)
-    if (false)
-#endif
     if (!isFakeTuple(i)) {
       WITH_CHRONO("oneshot-unwrap",
       WITH_CHRONO("unwrap",
@@ -681,9 +660,6 @@ Atrip::Output Atrip::run(Atrip::Input<F> const& in) {
 
 
     // COMPUTE ENERGY %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% {{{1
-#if defined(ATRIP_ONLY_DGEMM)
-    if (false)
-#endif
     if (!isFakeTuple(i)) {
       double tupleEnergy(0.);
 

src/atrip/DatabaseCommunicator.cxx

@@ -1,303 +0,0 @@
-#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

@@ -151,11 +151,12 @@ namespace cuda {
      KJI
     };
 
-  /*
+  /* 
    * Please the c++ type checker and template creator
    * 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 {};
@@ -435,20 +436,22 @@ double getEnergySame
     , DataFieldType<F>* Tijk_
     ) {
 
-    const size_t NoNo = No*No;
+    const size_t a = abc[0], b = abc[1], c = abc[2]
+              , NoNo = No*No
+              ;
 
     DataFieldType<F>* Tijk = (DataFieldType<F>*)Tijk_;
 
 #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",                          \
@@ -478,18 +481,11 @@ double getEnergySame
                   _t_buffer,                    \
                   (int const*)&NoNo             \
                   )
-#define MAYBE_CONJ(_conj, _buffer)                                \
-  do {                                                            \
-    cuda::maybeConjugate<<<                                       \
-                                                                  \
-                            Atrip::kernelDimensions.ooo.blocks,   \
-                                                                  \
-                              Atrip::kernelDimensions.ooo.threads \
-                                                                  \
-                        >>>((DataFieldType<F>*)_conj,             \
-                            (DataFieldType<F>*)_buffer,           \
-                            NoNoNo);                              \
-  } while (0)
+#define MAYBE_CONJ(_conj, _buffer)                                      \
+  cuda::maybeConjugate<<<                                               \
+    Atrip::kernelDimensions.ooo.blocks,                                 \
+    Atrip::kernelDimensions.ooo.threads                                 \
+    >>>((DataFieldType<F>*)_conj, (DataFieldType<F>*)_buffer, NoNoNo);
 
 
 // END CUDA ////////////////////////////////////////////////////////////////////
@@ -504,9 +500,7 @@ 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",                          \
@@ -537,13 +531,9 @@ double getEnergySame
                   _t_buffer,                    \
                   (int const*)&NoNo             \
                   )
-#define MAYBE_CONJ(_conj, _buffer)              \
-  do {                                          \
-    for (size_t __i = 0; __i < NoNoNo; ++__i) { \
-      _conj[__i]                                \
-        = maybeConjugate<F>(_buffer[__i]);      \
-    }                                           \
-  } while (0)
+#define MAYBE_CONJ(_conj, _buffer)                \
+  for (size_t __i = 0; __i < NoNoNo; ++__i)       \
+    _conj[__i] = maybeConjugate<F>(_buffer[__i]);
 #endif
 
     F one{1.0}, m_one{-1.0}, zero{0.0};
@@ -562,12 +552,8 @@ double getEnergySame
     const size_t
       bs = Atrip::kernelDimensions.ooo.blocks,
       ths = Atrip::kernelDimensions.ooo.threads;
-
-#if !defined(ATRIP_ONLY_DGEMM)
     cuda::zeroing<<<bs, ths>>>((DataFieldType<F>*)_t_buffer, NoNoNo);
     cuda::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));
@@ -579,7 +565,7 @@ double getEnergySame
 #endif
 
     // Set Tijk to zero
-#if defined(HAVE_CUDA) && !defined(ATRIP_ONLY_DGEMM)
+#ifdef HAVE_CUDA
     WITH_CHRONO("double:reorder",
                 cuda::zeroing<<<bs, ths>>>((DataFieldType<F>*)Tijk,
                                            NoNoNo);
@@ -591,51 +577,43 @@ double getEnergySame
        })
 #endif
 
-
-#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
-
     // 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)
         )
       }
     )
@@ -647,32 +625,32 @@ 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)
         )
       }
     )

src/atrip/Tuples.cxx

@@ -1,464 +0,0 @@
-#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

@@ -1,166 +0,0 @@
-#!/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
-#
-# - 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 sources-in-gpu
-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: