Add preliminary openacc support, atrip bench not linking

.dir-locals.el

@@ -13,7 +13,9 @@
                                      (format "%s/include/" root)
                                      (format "%s/" root)
                                      (format "%s/bench/" root)
-                                     (format "%s/build/main/" 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)

.github/workflows/main.yml

@@ -18,6 +18,8 @@ jobs:
     strategy:
       matrix:
         compiler:
+          - gcc12
+          - gcc11
           - gcc11
           - gcc10
           - gcc9

README.org

@@ -26,3 +26,110 @@ before the proper paper is released please contact me.
 
 In the mean time the code has been used in
 [[https://aip.scitation.org/doi/10.1063/5.0074936][this publication]] and can therefore been cited.
+
+* Building
+
+Atrip uses autotools to build the system.
+Autotools works by first creating a =configure= script from
+a =configure.ac= file.
+
+Atrip should be built out of source, this means that
+you have to create a build directory other that the root
+directory, for instance in the =build/tutorial= directory
+
+#+begin_src sh :exports code 
+mkdir -p build/tutorial/
+cd build/tutorial
+#+end_src
+
+First you have to build the =configure= script by doing
+
+#+begin_src sh :dir build/tutorial :exports code :results raw drawer
+../../bootstrap.sh
+#+end_src
+
+#+RESULTS:
+:results:
+
+  Creating configure script
+
+
+  Now you can build by doing
+
+  mkdir build
+  cd build
+  ../configure
+  make extern
+  make all
+
+:end:
+
+And then you can see the =configure= options
+#+begin_src sh :dir build/tutorial :results raw drawer :eval no 
+../../configure --help
+#+end_src
+
+** Benches
+
+The script =tools/configure-benches.sh= can be used to create
+a couple of configurations for benches:
+
+#+begin_src sh :exports results :results verbatim org   :results verbatim drawer replace output 
+awk '/begin +doc/,/end +doc/ { print $NL }' tools/configure-benches.sh |
+grep -v -e "begin \+doc" -e "end \+doc" |
+sed "s/^# //; s/^# *$//; /^$/d"
+#+end_src
+
+#+RESULTS:
+:results:
+- default ::
+  This configuration uses a CPU code with dgemm
+  and without computing slices.
+- only-dgemm ::
+  This only runs the computation part that involves dgemms.
+- cuda-only-dgemm ::
+  This is the naive CUDA implementation compiling only the dgemm parts
+  of the compute.
+- cuda-slices-on-gpu-only-dgemm ::
+  This configuration tests that slices reside completely on the gpu
+  and it should use a CUDA aware MPI implementation.
+  It also only uses the routines that involve dgemm.
+:end:
+
+In order to generate the benches just create a suitable directory for it
+
+#+begin_src sh :eval no
+mkdir -p build/benches
+cd buid/benches
+../../tools/configure-benches.sh CXX=g++ ...
+#+end_src
+
+and you will get a Makefile together with several project folders.
+You can either configure all projects with =make all= or
+then go in each folder.
+
+Notice that you can give a path for ctf for all of them by doing
+#+begin_src sh :eval no
+../../tools/configure-benches.sh --with-ctf=/absolute/path/to/ctf
+#+end_src
+
+* Running benches
+
+** Main benchmark
+
+The main benchmark gets built in =bench/atrip= and is used to run an
+atrip run with random tensors.
+
+A common run of this script will be the following
+
+#+begin_src sh
+bench/atrip \
+  --no 100 \
+  --nv 1000 \
+  --mod 1 \
+  --% 0 \
+  --dist group \
+  --nocheckpoint \
+  --max-iterations 1000
+#+end_src
+

bench/Makefile.am

@@ -19,10 +19,22 @@ BENCHES_LDADD = $(ATRIP_LIB) $(ATRIP_CTF)
 ## main entry point and bench
 ##
 bin_PROGRAMS += atrip
-atrip_SOURCES = test_main.cxx
+atrip_SOURCES = main.cxx
 atrip_CPPFLAGS = $(AM_CPPFLAGS)
 atrip_LDADD = $(BENCHES_LDADD)
 
+atrip: main.cxx
+	$(NVCXX) -cuda     \
+		-x cu -I../      \
+		$(MPILIBS)       \
+		-I$(srcdir)/     \
+		$(AM_CPPFLAGS)   \
+		$(DEFS)          \
+		$(BENCHES_LDADD) \
+		$(AM_LDFLAGS)    \
+		$< -o $@
+
+endif
 
 if !WITH_CUDA
 ##

bench/main.cxx

@@ -5,18 +5,20 @@
 #include <CLI11.hpp>
 
 #define _print_size(what, size)                 \
-  if (rank == 0) {                              \
+  do {                                          \
-    std::cout << #what                          \
+    if (rank == 0) {                            \
-              << " => "                         \
+      std::cout << #what                        \
-              << (double)size * elem_to_gb      \
+                << " => "                       \
-              << "GB"                           \
+                << (double)size * elem_to_gb    \
-              << std::endl;                     \
+                << "GB"                         \
-  }
+                << std::endl;                   \
+    }                                           \
+  } while (0)
 
 int main(int argc, char** argv) {
   MPI_Init(&argc, &argv);
 
-  size_t checkpoint_it;
+  size_t checkpoint_it, max_iterations;
   int no(10), nv(100), itMod(-1), percentageMod(10);
   float checkpoint_percentage;
   bool
@@ -30,6 +32,9 @@ int main(int argc, char** argv) {
   app.add_option("--no", no, "Occupied orbitals");
   app.add_option("--nv", nv, "Virtual orbitals");
   app.add_option("--mod", itMod, "Iteration modifier");
+  app.add_option("--max-iterations",
+                 max_iterations,
+                 "Maximum number of iterations to run");
   app.add_flag("--keep-vppph", keepVppph, "Do not delete Vppph");
   app.add_flag("--nochrono", nochrono, "Do not print chrono");
   app.add_flag("--rank-round-robin", rankRoundRobin, "Do rank round robin");
@@ -45,14 +50,27 @@ int main(int argc, char** argv) {
                  checkpoint_percentage,
                  "Percentage for checkpoints");
 
+  // Optional tensor files
+  std::string
+    ei_path, ea_path,
+    Tph_path, Tpphh_path,
+    Vpphh_path, Vhhhp_path, Vppph_path;
+  app.add_option("--ei", ei_path, "Path for ei");
+  app.add_option("--ea", ea_path, "Path for ea");
+  app.add_option("--Tpphh", Tpphh_path, "Path for Tpphh");
+  app.add_option("--Tph", Tph_path, "Path for Tph");
+  app.add_option("--Vpphh", Vpphh_path, "Path for Vpphh");
+  app.add_option("--Vhhhp", Vhhhp_path, "Path for Vhhhp");
+  app.add_option("--Vppph", Vppph_path, "Path for Vppph");
+
 #if defined(HAVE_CUDA)
   size_t ooo_threads = 0, ooo_blocks = 0;
   app.add_option("--ooo-blocks",
-		 ooo_blocks,
+                 ooo_blocks,
-		 "CUDA: Number of blocks per block for kernels going through ooo tensors");
+                 "CUDA: Number of blocks per block for kernels going through ooo tensors");
   app.add_option("--ooo-threads",
-		 ooo_threads,
+                 ooo_threads,
-		 "CUDA: Number of threads per block for kernels going through ooo tensors");
+                 "CUDA: Number of threads per block for kernels going through ooo tensors");
 #endif
 
   CLI11_PARSE(app, argc, argv);
@@ -148,37 +166,64 @@ int main(int argc, char** argv) {
   }
 
 
-  std::vector<int> symmetries(4, NS)
+  std::vector<int>
-                 , vo({nv, no})
+    symmetries(4, NS),
-                 , vvoo({nv, nv, no, no})
+    vo({nv, no}),
-                 , ooov({no, no, no, nv})
+    vvoo({nv, nv, no, no}),
-                 , vvvo({nv, nv, nv, no})
+    ooov({no, no, no, nv}),
-                 ;
+    vvvo({nv, nv, nv, no});
 
   CTF::Tensor<double>
-      ei(1, ooov.data(), symmetries.data(), world)
+    ei(1, ooov.data(), symmetries.data(), world),
-    , ea(1, vo.data(), symmetries.data(), world)
+    ea(1, vo.data(), symmetries.data(), world),
-    , Tph(2, vo.data(), symmetries.data(), world)
+    Tph(2, vo.data(), symmetries.data(), world),
-    , Tpphh(4, vvoo.data(), symmetries.data(), world)
+    Tpphh(4, vvoo.data(), symmetries.data(), world),
-    , Vpphh(4, vvoo.data(), symmetries.data(), world)
+    Vpphh(4, vvoo.data(), symmetries.data(), world),
-    , Vhhhp(4, ooov.data(), symmetries.data(), world)
+    Vhhhp(4, ooov.data(), symmetries.data(), world);
-    ;
 
   // initialize deletable tensors in heap
   auto Vppph
     = new CTF::Tensor<double>(4, vvvo.data(), symmetries.data(), world);
 
-  _print_size(Vabci, no*nv*nv*nv)
+  _print_size(Vabci, no*nv*nv*nv);
-  _print_size(Vabij, no*no*nv*nv)
+  _print_size(Vabij, no*no*nv*nv);
-  _print_size(Vijka, no*no*no*nv)
+  _print_size(Vijka, no*no*no*nv);
 
-  ei.fill_random(-40.0, -2);
+  if (ei_path.size()) {
-  ea.fill_random(2, 50);
+    ei.read_dense_from_file(ei_path.c_str());
-  Tpphh.fill_random(0, 1);
+  } else {
-  Tph.fill_random(0, 1);
+    ei.fill_random(-40.0, -2);
-  Vpphh.fill_random(0, 1);
+  }
-  Vhhhp.fill_random(0, 1);
+  if (ea_path.size()) {
-  Vppph->fill_random(0, 1);
+    ea.read_dense_from_file(ea_path.c_str());
+  } else {
+    ea.fill_random(2, 50);
+  }
+  if (Tpphh_path.size()) {
+    Tpphh.read_dense_from_file(Tpphh_path.c_str());
+  } else {
+    Tpphh.fill_random(0, 1);
+  }
+  if (Tph_path.size()) {
+    Tph.read_dense_from_file(Tph_path.c_str());
+  } else {
+    Tph.fill_random(0, 1);
+  }
+  if (Vpphh_path.size()) {
+    Vpphh.read_dense_from_file(Vpphh_path.c_str());
+  } else {
+    Vpphh.fill_random(0, 1);
+  }
+  if (Vhhhp_path.size()) {
+    Vhhhp.read_dense_from_file(Vhhhp_path.c_str());
+  } else {
+    Vhhhp.fill_random(0, 1);
+  }
+  if (Vppph_path.size()) {
+    Vppph->read_dense_from_file(Vppph_path.c_str());
+  } else {
+    Vppph->fill_random(0, 1);
+  }
 
   atrip::Atrip::init(MPI_COMM_WORLD);
   const auto in
@@ -199,6 +244,7 @@ int main(int argc, char** argv) {
        .with_iterationMod(itMod)
        .with_percentageMod(percentageMod)
        .with_tuplesDistribution(tuplesDistribution)
+       .with_maxIterations(max_iterations)
        // checkpoint options
        .with_checkpointAtEveryIteration(checkpoint_it)
        .with_checkpointAtPercentage(checkpoint_percentage)

configure.ac

@@ -21,26 +21,6 @@ AC_ARG_ENABLE(shared,
                                      files (default=YES)]),
               [], [enable_shared=yes])
 
-AC_ARG_ENABLE(
-    [slice],
-    [AS_HELP_STRING(
-      [--disable-slice],
-      [Disable the step of slicing tensors for CTF, this is useful for example for benchmarking or testing.])],
-    [atrip_dont_slice=1
-     AC_DEFINE([ATRIP_DONT_SLICE],1,[Wether CTF will slice tensors or skip the step])
-    ],
-    [atrip_dont_slice=0]
-)
-
-AC_ARG_ENABLE(
-    [atrip_dgemm],
-    [AS_HELP_STRING(
-      [--disable-dgemm],
-      [Disable using dgemm for the doubles equations])],
-    [],
-    [AC_DEFINE([ATRIP_USE_DGEMM],1,[Use dgemm for the doubles equations])]
-)
-
 
 AC_ARG_ENABLE([docs],
               [AS_HELP_STRING([--enable-docs],
@@ -68,19 +48,60 @@ AM_CONDITIONAL([WITH_CLANG_CHECK], [test x${clang_check} = xYES])
 AC_ARG_ENABLE([cuda],
               [AS_HELP_STRING([--enable-cuda],
                               [Build with cuda])],
-                              [WITH_CUDA=yes],
+              [WITH_CUDA=yes
-                              [WITH_CUDA=no])
+               WITH_OPENACC=yes],
+              [WITH_CUDA=no])
 AC_ARG_VAR([NVCC], [Path to the nvidia cuda compiler.])
 AC_ARG_VAR([CUDA_LDFLAGS], [LDFLAGS to find libraries -lcuda, -lcudart, -lcublas.])
 AC_ARG_VAR([CUDA_CXXFLAGS], [CXXFLAGS to find the CUDA headers])
 
+dnl -----------------------------------------------------------------------
+dnl ATRIP CPP DEFINES
+dnl -----------------------------------------------------------------------
 
 AC_ARG_WITH([atrip-debug],
             [AS_HELP_STRING([--with-atrip-debug],
-                            [Debug level for atrip, possible values: 1, 2, 3, 4])],
+                            [Debug level for atrip, possible values:
+                                   1, 2, 3, 4])],
             [AC_DEFINE([ATRIP_DEBUG],[atrip-debug],[Atrip debug level])],
-            [AC_DEFINE([ATRIP_DEBUG],[1],[Atrip debug level])]
+            [AC_DEFINE([ATRIP_DEBUG],[1],[Atrip debug level])])
-            )
+
+
+AC_ARG_ENABLE([atrip_dgemm],
+              [AS_HELP_STRING([--disable-dgemm],
+                              [Disable using dgemm for the doubles equations])],
+              [],
+              [AC_DEFINE([ATRIP_USE_DGEMM],
+                         1,
+                         [Use dgemm for the doubles equations])])
+
+ATRIP_DEF([slice], [disable],
+          [ATRIP_DONT_SLICE],
+          [Disable the step of slicing tensors for CTF, this is useful
+          for example for benchmarking or testing.])
+
+ATRIP_DEF([only-dgemm], [enable],
+          [ATRIP_ONLY_DGEMM],
+          [Run only the parts of atrip that involve dgemm calls, this
+          is useful for benchmarking and testing the code, it is
+          intended for developers of Atrip.])
+
+ATRIP_DEF([naive-slow], [enable],
+          [ATRIP_NAIVE_SLOW],
+          [Run slow but correct code for the mapping of (iteration,
+          rank) to tuple of the naive tuple distribution.])
+
+ATRIP_DEF([sources-in-gpu], [enable],
+          [ATRIP_SOURCES_IN_GPU],
+          [When using CUDA, activate storing all sources (slices of
+          the input tensors) in the GPU. This means that a lot of GPUs
+          will be needed.])
+
+ATRIP_DEF([cuda-aware-mpi], [enable],
+          [ATRIP_CUDA_AWARE_MPI],
+          [When using MPI, assume support for CUDA aware mpi by the
+          given MPI implementation.])
+
 
 dnl -----------------------------------------------------------------------
 
@@ -144,8 +165,7 @@ AC_TYPE_SIZE_T
 dnl -----------------------------------------------------------------------
 dnl CHECK CTF
 if test xYES = x${BUILD_CTF}; then
-  AC_MSG_WARN([Sorry, building CTF not supported yet provide a build path
+  AC_MSG_WARN([You will have to do make ctf before building the project.])
-                      with --with-ctf=path/to/ctf/installation])
 else
   CPPFLAGS="$CPPFLAGS -I${LIBCTF_CPATH}"
   LDFLAGS="$LDFLAGS -L${LIBCTF_LD_LIBRARY_PATH} -lctf"
@@ -163,6 +183,8 @@ if test x${WITH_CUDA} = xyes; then
                      -----------------------
   ])
   AC_CHECK_PROGS([NVCC], [nvcc])
+  AC_CHECK_PROGS([NVCXX], [nvc++])
+  MPILIBS=$($MPICXX -show | awk '!($1="")')
   AC_SUBST([CUDA_LDFLAGS])
   AC_DEFINE([HAVE_CUDA],1,[Wether we are using CUDA])
   # TODO: make sure to find cuda and cudart
@@ -208,6 +230,7 @@ AC_MSG_RESULT([
                        ATRIP_LDFLAGS  = $ATRIP_LDFLAGS
                        BLAS = ${BLAS_LIBS}
                        LIBS = ${LIBS}
+                       MPILIBS = $MPILIBS
 ])
 
 AC_OUTPUT

etc/env/raven/cuda

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

etc/env/raven/cuda-openacc

@@ -0,0 +1,82 @@
+#!/usr/bin/env bash
+mods=(
+  #cuda/11.6
+  nvhpcsdk/22  # for openacc
+  gcc/12
+  openmpi
+  mkl/2020.4
+  autoconf/2.69
+  automake/1.15
+  libtool/2.4.6
+)
+
+
+module purge
+module load ${mods[@]}
+
+LIB_PATH="${NVHPC_CUDA_HOME}/lib64"
+export CUBLAS_LD_PATH="${NVHPC_ROOT}/math_libs/lib64/"
+export CUDA_ROOT=${CUDA_HOME}
+export CUDA_LDFLAGS="-L${LIB_PATH} -lcuda -L${LIB_PATH} -lcudart -L${CUBLAS_LD_PATH} -lcublas"
+export CUDA_CXXFLAGS="-I${CUDA_HOME}/include"
+
+export LD_LIBRARY_PATH="${MKL_HOME}/lib/intel64:${LD_LIBRARY_PATH}"
+
+MPILIBS=$(mpicxx -show | awk '!($1="")')
+export MPILIBS
+export MPINVCXX="nv++ ${MPILIBS}"
+
+ls ${CUBLAS_LD_PATH}/libcublas.so
+ls ${LIB_PATH}/libcudart.so
+
+#export OMPI_CC="nvc"
+#export OMPI_CXX="nvc++"
+
+BLAS_LDFLAGS="-L${PWD}/OpenBLAS-0.3.20/ -lopenblas"
+_openblas_make () {
+
+  [[ -d OpenBLAS-0.3.20/ ]] || {
+    wget https://github.com/xianyi/OpenBLAS/releases/download/v0.3.20/OpenBLAS-0.3.20.tar.gz
+    tar xvzf OpenBLAS-0.3.20.tar.gz
+    cd OpenBLAS-0.3.20/
+    make FC=gfortran CC=gcc USE_OPENMP=1 NUM_THREADS=72 TARGET=SKYLAKEX
+  } && {
+    echo "Openblas built"
+  }
+
+}
+( _openblas_make; )
+
+
+
+cat <<EOF
+
+////////////////////////////////////////////////////////////////////////////////
+                                    info
+////////////////////////////////////////////////////////////////////////////////
+
+
+MKL_HOME         = $MKL_HOME
+BLAS_STATIC_PATH = $BLAS_STATIC_PATH
+
+CUDA_ROOT      = ${CUDA_HOME}
+CUDA_LDFLAGS   = "-L${LIB_PATH} -lcuda -L${LIB_PATH} -lcudart -L${LIB_PATH} -lcublas"
+CUDA_CXXFLAGS  = "-I${CUDA_HOME}/include"
+
+
+
+Consider now runnng the following
+
+../../configure \\
+  --enable-cuda \\
+  --disable-slice \\
+  --with-blas="${BLAS_LDFLAGS}" \\
+  CXX="gcc" \\
+  NVCC="\$MPINVCXX" \\
+  MPICXX="mpicxx"
+
+
+EOF
+	
+
+return

etc/m4/atrip-def.m4

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

etc/m4/atrip_openacc.m4

@@ -0,0 +1,70 @@
+#   SYNOPSIS
+#
+#   ATRIP_OPENACC([ACTION-SUCCESS], [ACTION-FAILURE])
+#
+#   DESCRIPTION
+#
+#   Check whether the given the -fopenacc flag works with the current language's compiler
+#   or gives an error.
+#
+#   ACTION-SUCCESS/ACTION-FAILURE are shell commands to execute on
+#   success/failure.
+#
+#   LICENSE
+#
+#   Copyright (c) 2023 Alejandro Gallo <aamsgallo@gmail.com>
+#
+#   Copying and distribution of this file, with or without modification, are
+#   permitted in any medium without royalty provided the copyright notice
+#   and this notice are preserved.  This file is offered as-is, without any
+#   warranty.
+
+AC_DEFUN([ATRIP_OPENACC],
+[
+AC_MSG_CHECKING([that the compiler works with the -fopenacc])
+AC_COMPILE_IFELSE([AC_LANG_SOURCE([_ATRIP_OPENACC_SOURCE])],
+                  [
+                   $1
+                   AC_MSG_RESULT([yes])
+                   ],
+                  [
+                   $2
+                   AC_MSG_ERROR([no])
+                   ])
+])dnl DEFUN
+
+m4_define([_ATRIP_OPENACC_SOURCE], [[
+#include <stdio.h>
+#include <stdlib.h>
+#include <openacc.h>
+
+#define SIZE 10
+
+int main(int argc, char **argv) {
+  float matrix[SIZE * SIZE];
+  float result[SIZE * SIZE];
+
+  // Initialize the matrix with random values
+  for (int i = 0; i < SIZE * SIZE; i++) {
+    matrix[i] = rand() / (float)RAND_MAX;
+  }
+
+#pragma acc data                                            \
+        copy(matrix[0:SIZE * SIZE])           \
+        copyout(result[0:SIZE * SIZE])
+  {
+    // Calculate the matrix multiplication
+#pragma acc parallel loop collapse(2)
+    for (int i = 0; i < SIZE; i++) {
+      for (int j = 0; j < SIZE; j++) {
+        float sum = 0.0f;
+        for (int k = 0; k < SIZE; k++) {
+          sum += matrix[i * SIZE + k] * matrix[j * SIZE + k];
+        }
+        result[i * SIZE + j] = sum;
+      }
+    }
+  }
+  return 0;
+}
+]])

etc/m4/ax_cxx_compile_stdcxx.m4

@@ -43,7 +43,7 @@
 #   and this notice are preserved.  This file is offered as-is, without any
 #   warranty.
 
-#serial 14
+#serial 15
 
 dnl  This macro is based on the code from the AX_CXX_COMPILE_STDCXX_11 macro
 dnl  (serial version number 13).
@@ -189,7 +189,11 @@ m4_define([_AX_CXX_COMPILE_STDCXX_testbody_new_in_11], [[
 
 #error "This is not a C++ compiler"
 
-#elif __cplusplus < 201103L
+// MSVC always sets __cplusplus to 199711L in older versions; newer versions
+// only set it correctly if /Zc:__cplusplus is specified as well as a
+// /std:c++NN switch:
+// https://devblogs.microsoft.com/cppblog/msvc-now-correctly-reports-__cplusplus/
+#elif __cplusplus < 201103L && !defined _MSC_VER
 
 #error "This is not a C++11 compiler"
 
@@ -480,7 +484,7 @@ m4_define([_AX_CXX_COMPILE_STDCXX_testbody_new_in_14], [[
 
 #error "This is not a C++ compiler"
 
-#elif __cplusplus < 201402L
+#elif __cplusplus < 201402L && !defined _MSC_VER
 
 #error "This is not a C++14 compiler"
 
@@ -604,7 +608,7 @@ m4_define([_AX_CXX_COMPILE_STDCXX_testbody_new_in_17], [[
 
 #error "This is not a C++ compiler"
 
-#elif __cplusplus < 201703L
+#elif __cplusplus < 201703L && !defined _MSC_VER
 
 #error "This is not a C++17 compiler"
 
@@ -970,7 +974,7 @@ namespace cxx17
 
 }  // namespace cxx17
 
-#endif  // __cplusplus < 201703L
+#endif  // __cplusplus < 201703L && !defined _MSC_VER
 
 ]])
 
@@ -983,7 +987,7 @@ m4_define([_AX_CXX_COMPILE_STDCXX_testbody_new_in_20], [[
 
 #error "This is not a C++ compiler"
 
-#elif __cplusplus < 202002L
+#elif __cplusplus < 202002L && !defined _MSC_VER
 
 #error "This is not a C++20 compiler"
 
@@ -1000,6 +1004,6 @@ namespace cxx20
 
 }  // namespace cxx20
 
-#endif  // __cplusplus < 202002L
+#endif  // __cplusplus < 202002L && !defined _MSC_VER
 
 ]])

include/atrip/Atrip.hpp

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

include/atrip/SliceUnion.hpp

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

include/atrip/Unions.hpp

@@ -19,8 +19,14 @@
 namespace atrip {
 
   template <typename F=double>
+  static
   void sliceIntoVector
-    ( std::vector<F> &v
+#if defined(ATRIP_SOURCES_IN_GPU)
+    ( DataPtr<F> &source
+#else
+    ( std::vector<F> &source
+#endif
+    , size_t sliceSize
     , CTF::Tensor<F> &toSlice
     , std::vector<int64_t> const low
     , std::vector<int64_t> const up
@@ -44,18 +50,30 @@ namespace atrip {
               << "\n";
 
 #ifndef ATRIP_DONT_SLICE
-    toSlice.slice( toSlice_.low.data()
+    toSlice.slice(toSlice_.low.data(),
-                 , toSlice_.up.data()
+                  toSlice_.up.data(),
-                 , 0.0
+                  0.0,
-                 , origin
+                  origin,
-                 , origin_.low.data()
+                  origin_.low.data(),
-                 , origin_.up.data()
+                  origin_.up.data(),
-                 , 1.0);
+                  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
+
   }
 
 
@@ -80,16 +98,15 @@ namespace atrip {
 
     void sliceIntoBuffer(size_t it, CTF::Tensor<F> &to, CTF::Tensor<F> const& from) override
     {
-      const int Nv = this->sliceLength[0]
-              , No = this->sliceLength[1]
-              , a = this->rankMap.find({static_cast<size_t>(Atrip::rank), it});
-              ;
 
+      const int
+        Nv = this->sliceLength[0],
+        No = this->sliceLength[1],
+        a = this->rankMap.find({static_cast<size_t>(Atrip::rank), it});
 
-      sliceIntoVector<F>( this->sources[it]
+      sliceIntoVector<F>(this->sources[it], this->sliceSize,
-                        , to,   {0, 0, 0},    {Nv, No, No}
+                         to,   {0, 0, 0},    {Nv, No, No},
-                        , from, {a, 0, 0, 0}, {a+1, Nv, No, No}
+                         from, {a, 0, 0, 0}, {a+1, Nv, No, No});
-                        );
 
     }
 
@@ -118,14 +135,13 @@ namespace atrip {
     void sliceIntoBuffer(size_t it, CTF::Tensor<F> &to, CTF::Tensor<F> const& from) override
     {
 
-      const int No = this->sliceLength[0]
+      const int
-              , a = this->rankMap.find({static_cast<size_t>(Atrip::rank), it})
+        No = this->sliceLength[0],
-              ;
+        a = this->rankMap.find({static_cast<size_t>(Atrip::rank), it});
 
-      sliceIntoVector<F>( this->sources[it]
+      sliceIntoVector<F>(this->sources[it], this->sliceSize,
-                        , to,   {0, 0, 0},    {No, No, No}
+                         to,   {0, 0, 0},    {No, No, No},
-                        , from, {0, 0, 0, a}, {No, No, No, a+1}
+                         from, {0, 0, 0, a}, {No, No, No, a+1});
-                        );
 
     }
   };
@@ -153,18 +169,17 @@ namespace atrip {
 
     void sliceIntoBuffer(size_t it, CTF::Tensor<F> &to, CTF::Tensor<F> const& from) override {
 
-      const int Nv = this->sliceLength[0]
+      const int
-              , No = this->sliceLength[1]
+        Nv = this->sliceLength[0],
-              , el = this->rankMap.find({static_cast<size_t>(Atrip::rank), it})
+        No = this->sliceLength[1],
-              , a = el % Nv
+        el = this->rankMap.find({static_cast<size_t>(Atrip::rank), it}),
-              , b = el / Nv
+        a = el % Nv,
-              ;
+        b = el / Nv;
 
 
-      sliceIntoVector<F>( this->sources[it]
+      sliceIntoVector<F>(this->sources[it], this->sliceSize,
-                        , to,   {0, 0},       {Nv, No}
+                         to,   {0, 0},       {Nv, No},
-                        , from, {a, b, 0, 0}, {a+1, b+1, Nv, No}
+                         from, {a, b, 0, 0}, {a+1, b+1, Nv, No});
-                        );
 
     }
 
@@ -191,17 +206,17 @@ namespace atrip {
 
     void sliceIntoBuffer(size_t it, CTF::Tensor<F> &to, CTF::Tensor<F> const& from) override {
 
-      const int Nv = from.lens[0]
+      const int
-              , No = this->sliceLength[1]
+        Nv = from.lens[0],
-              , el = this->rankMap.find({static_cast<size_t>(Atrip::rank), it})
+        No = this->sliceLength[1],
-              , a = el % Nv
+        el = this->rankMap.find({static_cast<size_t>(Atrip::rank), it}),
-              , b = el / Nv
+        a = el % Nv,
-              ;
+        b = el / Nv;
 
-      sliceIntoVector<F>( this->sources[it]
+
-                        , to,   {0, 0},       {No, No}
+      sliceIntoVector<F>(this->sources[it], this->sliceSize,
-                        , from, {a, b, 0, 0}, {a+1, b+1, No, No}
+                         to,   {0, 0},       {No, No},
-                        );
+                         from, {a, b, 0, 0}, {a+1, b+1, No, No});
 
 
     }
@@ -231,17 +246,16 @@ namespace atrip {
     void sliceIntoBuffer(size_t it, CTF::Tensor<F> &to, CTF::Tensor<F> const& from) override {
       // TODO: maybe generalize this with ABHH
 
-      const int Nv = from.lens[0]
+      const int
-              , No = this->sliceLength[1]
+        Nv = from.lens[0],
-              , el = this->rankMap.find({static_cast<size_t>(Atrip::rank), it})
+        No = this->sliceLength[1],
-              , a = el % Nv
+        el = this->rankMap.find({static_cast<size_t>(Atrip::rank), it}),
-              , b = el / Nv
+        a = el % Nv,
-              ;
+        b = el / Nv;
 
-      sliceIntoVector<F>( this->sources[it]
+      sliceIntoVector<F>(this->sources[it], this->sliceSize,
-                        , to,   {0, 0},       {No, No}
+                         to,   {0, 0},       {No, No},
-                        , from, {a, b, 0, 0}, {a+1, b+1, No, No}
+                         from, {a, b, 0, 0}, {a+1, b+1, No, No});
-                        );
 
 
     }

misc/naive-tuples.lisp

@@ -1,3 +1,11 @@
+#+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
@@ -218,58 +226,3 @@
             cheaper
             (print (equal (nth i tuples)
                    cheaper)))))
-
-(let* ((l 101)
-       (tuples (tuples-atrip l)))
-  (loop :for a below l
-        :do (print (let ((s (a-block-atrip a l))
-                         (c (count-if (lambda (x) (eq (car x) a))
-                                      tuples)))
-                     (list :a a
-                           :size s
-                           :real c
-                           :? (eq c s))))))
-
-(ql:quickload 'vgplot)
-(import 'vgplot:plot)
-(import 'vgplot:replot)
-
-(let ((l 10))
-  (plot (mapcar (lambda (x) (getf x :size))
-                (loop :for a upto l
-                      collect (list :a a :size (a-block a l))))
-        "penis"))
-
-(let* ((l 50)
-       (tuples (tuples-half l)))
-  (loop :for a below l
-        :do (print (let ((s (a-block a l))
-                         (c (count-if (lambda (x) (eq (car x) a))
-                                      tuples)))
-                     (list :a a
-                           :size s
-                           :real c
-                           :? (eq c s))))))
-
-(defun range (from to) (loop for i :from from :to to collect i))
-
-(defun half-again (i nv)
-  (let ((a-block-list (let ((ll (mapcar (lambda (i) (a-block i nv))
-                                        (range 0 (- nv 1)))))
-                        (loop :for i :from 1 :to (length ll)
-                              :collect
-                              (reduce #'+
-                                      ll
-                                      :end i)))))
-    (loop :for blk :in a-block-list
-          :with a = 0
-          :with total-blk = 0
-          :if (eq 0 (floor i blk))
-            :do
-               (let ((i (mod i blk)))
-                 (print (list i (- i total-blk) blk a))
-                 (return))
-          :else
-            :do (progn
-                  (incf a)
-                  (setq total-blk blk)))))

src/Makefile.am

@@ -7,16 +7,17 @@ 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 =
-NVCC_FILES = ./atrip/Equations.cxx ./atrip/Complex.cxx ./atrip/Atrip.cxx
+NVCC_FILES  = ./atrip/Equations.cxx ./atrip/Complex.cxx ./atrip/Atrip.cxx
-
+NVCC_FILES += ./atrip/Blas.cxx ./atrip/Tuples.cxx ./atrip/DatabaseCommunicator.cxx
 if WITH_CUDA
 NVCC_OBJS = $(patsubst %.cxx,%.nvcc.o,$(NVCC_FILES))
 libatrip_a_CPPFLAGS += $(CUDA_CXXFLAGS)
 libatrip_a_DEPENDENCIES = $(NVCC_OBJS)
 libatrip_a_LIBADD = $(NVCC_OBJS)
 %.nvcc.o: %.cxx
-	$(NVCC) -c -x cu -ccbin="${MPICXX}" -I../ $(CPPFLAGS) $(CTF_CPPFLAGS) $(DEFS) $(libatrip_a_CPPFLAGS) $< -o $@
+	##$(NVCC) -c -x cu -ccbin="${MPICXX}" -I../ $(CPPFLAGS) $(CTF_CPPFLAGS) $(DEFS) $(libatrip_a_CPPFLAGS) $< -o $@
+	$(NVCXX) -cuda $(MPILIBS) -c -x cu -I../ $(CPPFLAGS) $(CTF_CPPFLAGS) $(DEFS) $(libatrip_a_CPPFLAGS) $< -o $@
 
 #./atrip/Equations.o: ./atrip/Equations.cxx
 #	$(NVCC) -c -I../ $(CPPFLAGS) $(libatrip_a_CPPFLAGS) $< -o $@

src/atrip/Atrip.cxx

@@ -646,6 +646,9 @@ 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",
@@ -678,6 +681,9 @@ 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.);
 
@@ -688,13 +694,10 @@ Atrip::Output Atrip::run(Atrip::Input<F> const& in) {
 
       // LOG(0, "AtripCUDA") <<  "doing energy " << i << "distinct " << distinct << "\n";
       WITH_CHRONO("energy",
-/*
-    TODO: think about how to do this on the GPU in the best way possible
         if ( distinct == 0)
           tupleEnergy = getEnergyDistinct<F>(epsabc, No, (F*)epsi, (F*)Tijk, (F*)Zijk);
         else
           tupleEnergy = getEnergySame<F>(epsabc, No, (F*)epsi, (F*)Tijk, (F*)Zijk);
-*/
       )
 
 #if defined(HAVE_OCD) || defined(ATRIP_PRINT_TUPLES)
@@ -767,6 +770,8 @@ Atrip::Output Atrip::run(Atrip::Input<F> const& in) {
     Atrip::chrono["iterations"].stop();
     // ITERATION END %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%{{{1
 
+    if (in.maxIterations != 0 && i >= in.maxIterations) break;
+
   }
     // END OF MAIN LOOP
 

src/atrip/DatabaseCommunicator.cxx

@@ -4,8 +4,10 @@
 
 namespace atrip {
 
-  /*   This function is really too slow, below are more performant
+#if defined(ATRIP_NAIVE_SLOW)
-   functions to get tuples.
+  /*
+   * 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) {
@@ -52,33 +54,26 @@ namespace atrip {
     return result;
 
   }
-
+#endif
-  static
-  inline
-  size_t a_block_atrip(size_t a, size_t nv) {
-    return (nv - 1) * (nv - (a - 1))
-         - ((nv - 1) * nv) / 2
-         + ((a - 1) * (a - 2)) / 2
-         - 1;
-  }
 
   static
   inline
   size_t a_block_sum_atrip(int64_t T, int64_t nv) {
-    int64_t nv1 = nv - 1, tplus1 = T + 1;
+    const int64_t nv_min_1 = nv - 1, t_plus_1 = T + 1;
-    return tplus1 * nv1 * nv
+    return t_plus_1 * nv_min_1 * nv
-         + nv1 * tplus1
+         + nv_min_1 * t_plus_1
-         - (nv1 * (T * (T + 1)) / 2)
+         - (nv_min_1 * (T * t_plus_1) / 2)
-         - (tplus1 * (nv1 * nv) / 2)
+         - (t_plus_1 * (nv_min_1 * nv) / 2)
-         + (((T * (T + 1) * (1 + 2 * T)) / 6) - 3 * ((T * (T + 1)) / 2)) / 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
          ;
-         // + tplus1;
   }
 
   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;
@@ -94,9 +89,6 @@ namespace atrip {
       a_sums.resize(nv);
       for (size_t _i = 0; _i < nv; _i++) {
         a_sums[_i] = a_block_sum_atrip(_i, nv);
-        /*
-        std::cout << Atrip::rank << ": " << _i << " " << a_sums[_i] << std::endl;
-        */
       }
     }
 
@@ -114,10 +106,6 @@ namespace atrip {
     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);
-      /*
-      std::cout << Atrip::rank << ": b-sum " << i-1 << " "
-                << ":a " << a << " :t " << t  << " = " << b_sums[i-1] << std::endl;
-      */
     }
     int64_t b = a - 1, block_b = block_a;
     for (const auto& sum: b_sums) {
@@ -141,6 +129,11 @@ namespace atrip {
   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;
@@ -160,9 +153,6 @@ namespace atrip {
     for (size_t rank = 0; rank < np; rank++) {
       const size_t
         global_iteration = tuples_per_rank * rank + it;
-      /*
-      std::cout << Atrip::rank << ":" << "global_bit " <<  global_iteration << "\n";
-      */
       result[rank] = nth_atrip(global_iteration, nv);
     }
 
@@ -248,38 +238,25 @@ namespace atrip {
     using Database = typename Slice<F>::Database;
     Database db;
 
-#ifdef NAIVE_SLOW
+#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,
-                                                              (int64_t)iteration - 1);
+                                                              iteration - 1);
                 )
 #else
     WITH_CHRONO("db:comm:naive:tuples",
-                const auto tuples = nth_atrip_distributed((int64_t)iteration,
+                const auto tuples = nth_atrip_distributed(iteration,
                                                           nv,
                                                           np);
-                const auto prev_tuples = nth_atrip_distributed((int64_t)iteration - 1,
+                const auto prev_tuples = nth_atrip_distributed(iteration - 1,
                                                                nv,
                                                                np);
                 )
 
-      if (false)
-      for (size_t rank = 0; rank < np; rank++) {
-        std::cout << Atrip::rank << ":"
-                  << " :tuples< " << rank << ">" << iteration
-                  << " :abc " << tuples[rank][0]
-                  << ", " << tuples[rank][1]
-                  << ", " << tuples[rank][2] << "\n";
-        std::cout << Atrip::rank << ":"
-                  << " :prev-tuples< " << rank << ">" << iteration
-                  << " :abc-prev " << prev_tuples[rank][0]
-                  << ", " << prev_tuples[rank][1]
-                  << ", " << prev_tuples[rank][2] << "\n";
-      }
 #endif
 
     for (size_t rank = 0; rank < np; rank++) {

src/atrip/Equations.cxx

@@ -151,12 +151,11 @@ 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 {};
@@ -183,15 +182,21 @@ namespace cuda {
 
 // [[file:~/cuda/atrip/atrip.org::*Energy][Energy:2]]
 template <typename F>
+  __MAYBE_DEVICE__
 double getEnergyDistinct
-  ( F const epsabc
+  (F const epsabc,
-  , size_t const No
+   size_t const No,
-  , F* const epsi
+   F* const epsi,
-  , F* const Tijk
+   F* const Tijk,
-  , F* const Zijk
+   F* const Zijk) {
-  ) {
   constexpr size_t blockSize=16;
   F energy(0.);
+#if defined(HAVE_CUDA)
+  #pragma acc kernels
+  for (size_t k(0); k < No; k++) {
+    for (size_t j(k); j < No; j++) {
+      for (size_t i(j); i < No; i++) {
+#else
   for (size_t kk=0; kk<No; kk+=blockSize){
     const size_t kend( std::min(No, kk+blockSize) );
     for (size_t jj(kk); jj<No; jj+=blockSize){
@@ -199,13 +204,14 @@ double getEnergyDistinct
       for (size_t ii(jj); ii<No; ii+=blockSize){
         const size_t iend( std::min( No, ii+blockSize) );
         for (size_t k(kk); k < kend; k++){
-          const F ek(epsi[k]);
           const size_t jstart = jj > k ? jj : k;
           for (size_t j(jstart); j < jend; j++){
-            F const ej(epsi[j]);
-            F const facjk = j == k ? F(0.5) : F(1.0);
             size_t istart = ii > j ? ii : j;
             for (size_t i(istart); i < iend; i++){
+#endif
+              const F ek(epsi[k]);
+              const F ej(epsi[j]);
+              const F facjk = j == k ? F(0.5) : F(1.0);
               const F
                   ei(epsi[i])
                 , facij = i == j ? F(0.5) : F(1.0)
@@ -240,14 +246,17 @@ double getEnergyDistinct
             } // i
           } // j
         } // k
+#if !defined(HAVE_CUDA)
       } // ii
     } // jj
   } // kk
+#endif
   return std::real(energy);
 }
 
 
 template <typename F>
+  __MAYBE_DEVICE__
 double getEnergySame
   ( F const epsabc
   , size_t const No
@@ -257,6 +266,12 @@ double getEnergySame
   ) {
   constexpr size_t blockSize = 16;
   F energy = F(0.);
+#if defined(HAVE_CUDA)
+  #pragma acc kernels
+  for (size_t k(0); k < No; k++) {
+    for (size_t j(k); j < No; j++) {
+      for (size_t i(j); i < No; i++) {
+#else
   for (size_t kk=0; kk<No; kk+=blockSize){
     const size_t kend( std::min( kk+blockSize, No) );
     for (size_t jj(kk); jj<No; jj+=blockSize){
@@ -264,13 +279,14 @@ double getEnergySame
       for (size_t ii(jj); ii<No; ii+=blockSize){
         const size_t iend( std::min( ii+blockSize, No) );
         for (size_t k(kk); k < kend; k++){
-          const F ek(epsi[k]);
           const size_t jstart = jj > k ? jj : k;
           for(size_t j(jstart); j < jend; j++){
-            const F facjk( j == k ? F(0.5) : F(1.0));
-            const F ej(epsi[j]);
             const size_t istart = ii > j ? ii : j;
             for(size_t i(istart); i < iend; i++){
+#endif
+              const F facjk( j == k ? F(0.5) : F(1.0));
+              const F ek(epsi[k]);
+              const F ej(epsi[j]);
               const F
                 ei(epsi[i])
               , facij ( i==j ? F(0.5) : F(1.0))
@@ -292,9 +308,11 @@ double getEnergySame
             } // i
           } // j
         } // k
+#if !defined(HAVE_CUDA)
       } // ii
     } // jj
   } // kk
+#endif
   return std::real(energy);
 }
 // Energy:2 ends here
@@ -436,22 +454,20 @@ double getEnergySame
     , DataFieldType<F>* Tijk_
     ) {
 
-    const size_t a = abc[0], b = abc[1], c = abc[2]
+    const size_t NoNo = No*No;
-              , NoNo = No*No
-              ;
 
     DataFieldType<F>* Tijk = (DataFieldType<F>*)Tijk_;
 
 #if defined(ATRIP_USE_DGEMM)
 #if defined(HAVE_CUDA)
-#define REORDER(__II, __JJ, __KK)                           \
+#define REORDER(__II, __JJ, __KK)               \
-  reorder<<<bs, ths>>>(reorder_proxy<                       \
+  reorder<<<bs, ths>>>(reorder_proxy<           \
-                       DataFieldType<F>,                    \
+                       DataFieldType<F>,        \
-                       __II ## __JJ ## __KK                 \
+                       __II ## __JJ ## __KK     \
-                       >{},                                 \
+                       >{},                     \
-                       No,                                  \
+                       No,                      \
-                       Tijk,                                \
+                       Tijk,                    \
-                       _t_buffer);
+                       _t_buffer)
 #define DGEMM_PARTICLES(__A, __B)               \
   atrip::xgemm<F>("T",                          \
                   "N",                          \
@@ -481,11 +497,18 @@ double getEnergySame
                   _t_buffer,                    \
                   (int const*)&NoNo             \
                   )
-#define MAYBE_CONJ(_conj, _buffer)                                      \
+#define MAYBE_CONJ(_conj, _buffer)                                \
-  cuda::maybeConjugate<<<                                               \
+  do {                                                            \
-    Atrip::kernelDimensions.ooo.blocks,                                 \
+    cuda::maybeConjugate<<<                                       \
-    Atrip::kernelDimensions.ooo.threads                                 \
+                                                                  \
-    >>>((DataFieldType<F>*)_conj, (DataFieldType<F>*)_buffer, NoNoNo);
+                            Atrip::kernelDimensions.ooo.blocks,   \
+                                                                  \
+                              Atrip::kernelDimensions.ooo.threads \
+                                                                  \
+                        >>>((DataFieldType<F>*)_conj,             \
+                            (DataFieldType<F>*)_buffer,           \
+                            NoNoNo);                              \
+  } while (0)
 
 
 // END CUDA ////////////////////////////////////////////////////////////////////
@@ -500,7 +523,9 @@ double getEnergySame
 #define REORDER(__II, __JJ, __KK)               \
   reorder(reorder_proxy<DataFieldType<F>,       \
           __II ## __JJ ## __KK >{},             \
-          No, Tijk, _t_buffer);
+          No,                                   \
+          Tijk,                                 \
+          _t_buffer)
 #define DGEMM_PARTICLES(__A, __B)               \
   atrip::xgemm<F>("T",                          \
                   "N",                          \
@@ -531,9 +556,13 @@ double getEnergySame
                   _t_buffer,                    \
                   (int const*)&NoNo             \
                   )
-#define MAYBE_CONJ(_conj, _buffer)                \
+#define MAYBE_CONJ(_conj, _buffer)              \
-  for (size_t __i = 0; __i < NoNoNo; ++__i)       \
+  do {                                          \
-    _conj[__i] = maybeConjugate<F>(_buffer[__i]);
+    for (size_t __i = 0; __i < NoNoNo; ++__i) { \
+      _conj[__i]                                \
+        = maybeConjugate<F>(_buffer[__i]);      \
+    }                                           \
+  } while (0)
 #endif
 
     F one{1.0}, m_one{-1.0}, zero{0.0};
@@ -552,8 +581,12 @@ 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));
@@ -565,7 +598,7 @@ double getEnergySame
 #endif
 
     // Set Tijk to zero
-#ifdef HAVE_CUDA
+#if defined(HAVE_CUDA) && !defined(ATRIP_ONLY_DGEMM)
     WITH_CHRONO("double:reorder",
                 cuda::zeroing<<<bs, ths>>>((DataFieldType<F>*)Tijk,
                                            NoNoNo);
@@ -577,43 +610,51 @@ 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);
         )
       }
     )
@@ -625,32 +666,32 @@ 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);
         )
       }
     )

tools/configure-benches.sh

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