add DWF benchmark

Grid/Benchmark_Grid.cpp

@@ -1,7 +1,7 @@
 /*
 Copyright © 2015 Peter Boyle <paboyle@ph.ed.ac.uk>
 Copyright © 2022 Antonin Portelli <antonin.portelli@me.com>
-Copyright © 2024 Simon Buerger <simon.buerger@rwth-aachen.de>
+Copyright © 2022 Simon Buerger <simon.buerger@rwth-aachen.de>
 
 This is a fork of Benchmark_ITT.cpp from Grid
 
@@ -29,43 +29,6 @@ int NN_global;
 
 nlohmann::json json_results;
 
-// NOTE: Grid::GridClock is just a typedef to
-// `std::chrono::high_resolution_clock`, but `Grid::usecond` rounds to
-// microseconds (no idea why, probably wasnt ever relevant before), so we need
-// our own wrapper here.
-double usecond_precise()
-{
-  using namespace std::chrono;
-  auto nsecs = duration_cast<nanoseconds>(GridClock::now() - Grid::theProgramStart);
-  return nsecs.count() * 1e-3;
-}
-
-std::vector<std::string> get_mpi_hostnames()
-{
-  int world_size;
-  MPI_Comm_size(MPI_COMM_WORLD, &world_size);
-
-  char hostname[MPI_MAX_PROCESSOR_NAME];
-  int name_len = 0;
-  MPI_Get_processor_name(hostname, &name_len);
-
-  // Allocate buffer to gather all hostnames
-  std::vector<char> all_hostnames(world_size * MPI_MAX_PROCESSOR_NAME);
-
-  // Use MPI_Allgather to gather all hostnames on all ranks
-  MPI_Allgather(hostname, MPI_MAX_PROCESSOR_NAME, MPI_CHAR, all_hostnames.data(),
-                MPI_MAX_PROCESSOR_NAME, MPI_CHAR, MPI_COMM_WORLD);
-
-  // Convert the gathered hostnames back into a vector of std::string
-  std::vector<std::string> hostname_list(world_size);
-  for (int i = 0; i < world_size; ++i)
-  {
-    hostname_list[i] = std::string(&all_hostnames[i * MPI_MAX_PROCESSOR_NAME]);
-  }
-
-  return hostname_list;
-}
-
 struct time_statistics
 {
   double mean;
@@ -110,8 +73,6 @@ class Benchmark
         {local[0] * mpi[0], local[1] * mpi[1], local[2] * mpi[2], local[3] * mpi[3]});
     GridCartesian *TmpGrid = SpaceTimeGrid::makeFourDimGrid(
         latt4, GridDefaultSimd(Nd, vComplex::Nsimd()), GridDefaultMpi());
-    Grid::Coordinate shm;
-    GlobalSharedMemory::GetShmDims(mpi, shm);
 
     uint64_t NP = TmpGrid->RankCount();
     uint64_t NN = TmpGrid->NodeCount();
@@ -124,9 +85,7 @@ class Benchmark
     std::cout << GridLogMessage << "* OpenMP threads : " << GridThread::GetThreads()
               << std::endl;
 
-    std::cout << GridLogMessage << "* MPI layout     : " << GridCmdVectorIntToString(mpi)
-              << std::endl;
-    std::cout << GridLogMessage << "* Shm layout     : " << GridCmdVectorIntToString(shm)
+    std::cout << GridLogMessage << "* MPI tasks      : " << GridCmdVectorIntToString(mpi)
               << std::endl;
 
     std::cout << GridLogMessage << "* vReal          : " << sizeof(vReal) * 8 << "bits ; "
@@ -159,7 +118,6 @@ class Benchmark
     for (unsigned int i = 0; i < mpi.size(); ++i)
     {
       tmp["mpi"].push_back(mpi[i]);
-      tmp["shm"].push_back(shm[i]);
     }
     tmp["ranks"] = NP;
     tmp["nodes"] = NN;
@@ -174,8 +132,6 @@ class Benchmark
 
     Coordinate simd_layout = GridDefaultSimd(Nd, vComplexD::Nsimd());
     Coordinate mpi_layout = GridDefaultMpi();
-    Coordinate shm_layout;
-    GlobalSharedMemory::GetShmDims(mpi_layout, shm_layout);
 
     for (int mu = 0; mu < Nd; mu++)
       if (mpi_layout[mu] > 1)
@@ -187,8 +143,8 @@ class Benchmark
     std::cout << GridLogMessage << "Benchmarking threaded STENCIL halo exchange in "
               << nmu << " dimensions" << std::endl;
     grid_small_sep();
-    grid_printf("%5s %5s %7s %15s %15s %15s %15s %15s\n", "L", "dir", "shm",
-                "payload (B)", "time (usec)", "rate (GB/s/node)", "std dev", "max");
+    grid_printf("%5s %5s %15s %15s %15s %15s %15s\n", "L", "dir", "payload (B)",
+                "time (usec)", "rate (GB/s/node)", "std dev", "max");
 
     for (int lat = 16; lat <= maxlat; lat += 8)
     {
@@ -217,80 +173,74 @@ class Benchmark
       for (int dir = 0; dir < 8; dir++)
       {
         int mu = dir % 4;
-        if (mpi_layout[mu] == 1) // skip directions that are not distributed
-          continue;
-        bool is_shm = mpi_layout[mu] == shm_layout[mu];
-        bool is_partial_shm = !is_shm && shm_layout[mu] != 1;
-
-        std::vector<double> times(Nloop);
-        for (int i = 0; i < NWARMUP; i++)
-        {
-          int xmit_to_rank;
-          int recv_from_rank;
-
-          if (dir == mu)
-          {
-            int comm_proc = 1;
-            Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
-          }
-          else
-          {
-            int comm_proc = mpi_layout[mu] - 1;
-            Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
-          }
-          Grid.SendToRecvFrom((void *)&xbuf[dir][0], xmit_to_rank, (void *)&rbuf[dir][0],
-                              recv_from_rank, bytes);
-        }
-        for (int i = 0; i < Nloop; i++)
+        if (mpi_layout[mu] > 1)
         {
 
-          dbytes = 0;
-          double start = usecond();
-          int xmit_to_rank;
-          int recv_from_rank;
-
-          if (dir == mu)
+          std::vector<double> times(Nloop);
+          for (int i = 0; i < NWARMUP; i++)
           {
-            int comm_proc = 1;
-            Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
-          }
-          else
-          {
-            int comm_proc = mpi_layout[mu] - 1;
-            Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
-          }
-          Grid.SendToRecvFrom((void *)&xbuf[dir][0], xmit_to_rank, (void *)&rbuf[dir][0],
-                              recv_from_rank, bytes);
-          dbytes += bytes;
+            int xmit_to_rank;
+            int recv_from_rank;
 
-          double stop = usecond();
-          t_time[i] = stop - start; // microseconds
+            if (dir == mu)
+            {
+              int comm_proc = 1;
+              Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
+            }
+            else
+            {
+              int comm_proc = mpi_layout[mu] - 1;
+              Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
+            }
+            Grid.SendToRecvFrom((void *)&xbuf[dir][0], xmit_to_rank,
+                                (void *)&rbuf[dir][0], recv_from_rank, bytes);
+          }
+          for (int i = 0; i < Nloop; i++)
+          {
+
+            dbytes = 0;
+            double start = usecond();
+            int xmit_to_rank;
+            int recv_from_rank;
+
+            if (dir == mu)
+            {
+              int comm_proc = 1;
+              Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
+            }
+            else
+            {
+              int comm_proc = mpi_layout[mu] - 1;
+              Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
+            }
+            Grid.SendToRecvFrom((void *)&xbuf[dir][0], xmit_to_rank,
+                                (void *)&rbuf[dir][0], recv_from_rank, bytes);
+            dbytes += bytes;
+
+            double stop = usecond();
+            t_time[i] = stop - start; // microseconds
+          }
+          timestat.statistics(t_time);
+
+          dbytes = dbytes * ppn;
+          double bidibytes = 2. * dbytes;
+          double rate = bidibytes / (timestat.mean / 1.e6) / 1024. / 1024. / 1024.;
+          double rate_err = rate * timestat.err / timestat.mean;
+          double rate_max = rate * timestat.mean / timestat.min;
+          grid_printf("%5d %5d %15d %15.2f %15.2f %15.1f %15.2f\n", lat, dir, bytes,
+                      timestat.mean, rate, rate_err, rate_max);
+          nlohmann::json tmp;
+          nlohmann::json tmp_rate;
+          tmp["L"] = lat;
+          tmp["dir"] = dir;
+          tmp["bytes"] = bytes;
+          tmp["time_usec"] = timestat.mean;
+          tmp_rate["mean"] = rate;
+          tmp_rate["error"] = rate_err;
+          tmp_rate["max"] = rate_max;
+          tmp["rate_GBps"] = tmp_rate;
+          json_results["comms"].push_back(tmp);
         }
-        timestat.statistics(t_time);
-
-        dbytes = dbytes * ppn;
-        double bidibytes = 2. * dbytes;
-        double rate = bidibytes / (timestat.mean / 1.e6) / 1024. / 1024. / 1024.;
-        double rate_err = rate * timestat.err / timestat.mean;
-        double rate_max = rate * timestat.mean / timestat.min;
-        grid_printf("%5d %5d %7s %15d %15.2f %15.2f %15.1f %15.2f\n", lat, dir,
-                    is_shm           ? "yes"
-                    : is_partial_shm ? "partial"
-                                     : "no",
-                    bytes, timestat.mean, rate, rate_err, rate_max);
-        nlohmann::json tmp;
-        nlohmann::json tmp_rate;
-        tmp["L"] = lat;
-        tmp["dir"] = dir;
-        tmp["shared_mem"] = is_shm;
-        tmp["partial_shared_mem"] = is_partial_shm;
-        tmp["bytes"] = bytes;
-        tmp["time_usec"] = timestat.mean;
-        tmp_rate["mean"] = rate;
-        tmp_rate["error"] = rate_err;
-        tmp_rate["max"] = rate_max;
-        tmp["rate_GBps"] = tmp_rate;
-        json_results["comms"].push_back(tmp);
       }
       for (int d = 0; d < 8; d++)
       {
@@ -301,170 +251,6 @@ class Benchmark
     return;
   }
 
-  static void Latency(void)
-  {
-    int Nwarmup = 100;
-    int Nloop = 300;
-
-    std::cout << GridLogMessage << "Benchmarking point-to-point latency" << std::endl;
-    grid_small_sep();
-    grid_printf("from to      mean(usec)           err           max\n");
-
-    int ranks;
-    int me;
-    MPI_Comm_size(MPI_COMM_WORLD, &ranks);
-    MPI_Comm_rank(MPI_COMM_WORLD, &me);
-
-    int bytes = 8;
-    void *buf_from = acceleratorAllocDevice(bytes);
-    void *buf_to = acceleratorAllocDevice(bytes);
-    nlohmann::json json_latency;
-    for (int from = 0; from < ranks; ++from)
-      for (int to = 0; to < ranks; ++to)
-      {
-        if (from == to)
-          continue;
-
-        std::vector<double> t_time(Nloop);
-        time_statistics timestat;
-        MPI_Status status;
-
-        for (int i = -Nwarmup; i < Nloop; ++i)
-        {
-          double start = usecond_precise();
-          if (from == me)
-          {
-            auto err = MPI_Send(buf_from, bytes, MPI_CHAR, to, 0, MPI_COMM_WORLD);
-            assert(err == MPI_SUCCESS);
-          }
-          if (to == me)
-          {
-            auto err =
-                MPI_Recv(buf_to, bytes, MPI_CHAR, from, 0, MPI_COMM_WORLD, &status);
-            assert(err == MPI_SUCCESS);
-          }
-          double stop = usecond_precise();
-          if (i >= 0)
-            t_time[i] = stop - start;
-        }
-        // important: only 'from' and 'to' have meaningful timings. we use
-        // 'from's.
-        MPI_Bcast(t_time.data(), Nloop, MPI_DOUBLE, from, MPI_COMM_WORLD);
-
-        timestat.statistics(t_time);
-        grid_printf("%2d %2d %15.4f %15.3f %15.4f\n", from, to, timestat.mean,
-                    timestat.err, timestat.max);
-        nlohmann::json tmp;
-        tmp["from"] = from;
-        tmp["to"] = to;
-        tmp["time_usec"] = timestat.mean;
-        tmp["time_usec_error"] = timestat.err;
-        tmp["time_usec_min"] = timestat.min;
-        tmp["time_usec_max"] = timestat.max;
-        tmp["time_usec_full"] = t_time;
-        json_latency.push_back(tmp);
-      }
-    json_results["latency"] = json_latency;
-
-    acceleratorFreeDevice(buf_from);
-    acceleratorFreeDevice(buf_to);
-  }
-
-  static void P2P(void)
-  {
-    // IMPORTANT: The P2P benchmark uses "MPI_COMM_WORLD" communicator, which is
-    // not the quite the same as Grid.communicator. Practically speaking, the
-    // latter one contains the same MPI-ranks but in a different order. Grid
-    // does this make sure it can exploit ranks with shared memory (i.e.
-    // multiple ranks on the same node) as best as possible.
-
-    // buffer-size to benchmark. This number is the same as the largest one used
-    // in the "Comms()" benchmark. ( L=48, Ls=12, double-prec-complex,
-    // half-color-spin-vector. ). Mostly an arbitrary choice, but nice to match
-    // it here
-    size_t bytes = 127401984;
-
-    int Nwarmup = 20;
-    int Nloop = 100;
-
-    std::cout << GridLogMessage << "Benchmarking point-to-point bandwidth" << std::endl;
-    grid_small_sep();
-    grid_printf("from to      mean(usec)           err           min           "
-                "bytes    rate (GiB/s)\n");
-
-    int ranks;
-    int me;
-    MPI_Comm_size(MPI_COMM_WORLD, &ranks);
-    MPI_Comm_rank(MPI_COMM_WORLD, &me);
-
-    void *buf_from = acceleratorAllocDevice(bytes);
-    void *buf_to = acceleratorAllocDevice(bytes);
-    nlohmann::json json_p2p;
-    for (int from = 0; from < ranks; ++from)
-      for (int to = 0; to < ranks; ++to)
-      {
-        if (from == to)
-          continue;
-
-        std::vector<double> t_time(Nloop);
-        time_statistics timestat;
-        MPI_Status status;
-
-        for (int i = -Nwarmup; i < Nloop; ++i)
-        {
-          double start = usecond_precise();
-          if (from == me)
-          {
-            auto err = MPI_Send(buf_from, bytes, MPI_CHAR, to, 0, MPI_COMM_WORLD);
-            assert(err == MPI_SUCCESS);
-          }
-          if (to == me)
-          {
-            auto err =
-                MPI_Recv(buf_to, bytes, MPI_CHAR, from, 0, MPI_COMM_WORLD, &status);
-            assert(err == MPI_SUCCESS);
-          }
-          double stop = usecond_precise();
-          if (i >= 0)
-            t_time[i] = stop - start;
-        }
-        // important: only 'from' and 'to' have meaningful timings. we use
-        // 'from's.
-        MPI_Bcast(t_time.data(), Nloop, MPI_DOUBLE, from, MPI_COMM_WORLD);
-
-        timestat.statistics(t_time);
-        double rate = bytes / (timestat.mean / 1.e6) / 1024. / 1024. / 1024.;
-        double rate_err = rate * timestat.err / timestat.mean;
-        double rate_max = rate * timestat.mean / timestat.min;
-        double rate_min = rate * timestat.mean / timestat.max;
-
-        grid_printf("%2d %2d %15.4f %15.3f %15.4f %15d %15.2f\n", from, to, timestat.mean,
-                    timestat.err, timestat.min, bytes, rate);
-
-        nlohmann::json tmp;
-        tmp["from"] = from;
-        tmp["to"] = to;
-        tmp["bytes"] = bytes;
-        tmp["time_usec"] = timestat.mean;
-        tmp["time_usec_error"] = timestat.err;
-        tmp["time_usec_min"] = timestat.min;
-        tmp["time_usec_max"] = timestat.max;
-        tmp["time_usec_full"] = t_time;
-        nlohmann::json tmp_rate;
-        tmp_rate["mean"] = rate;
-        tmp_rate["error"] = rate_err;
-        tmp_rate["max"] = rate_max;
-        tmp_rate["min"] = rate_min;
-        tmp["rate_GBps"] = tmp_rate;
-
-        json_p2p.push_back(tmp);
-      }
-    json_results["p2p"] = json_p2p;
-
-    acceleratorFreeDevice(buf_from);
-    acceleratorFreeDevice(buf_to);
-  }
-
   static void Memory(void)
   {
     const int Nvec = 8;
@@ -726,6 +512,8 @@ class Benchmark
 
         FGrid->Broadcast(0, &ncall, sizeof(ncall));
 
+        Dw.ZeroCounters();
+
         time_statistics timestat;
         std::vector<double> t_time(ncall);
         for (uint64_t i = 0; i < ncall; i++)
@@ -920,6 +708,7 @@ class Benchmark
         uint64_t ncall = 500;
 
         FGrid->Broadcast(0, &ncall, sizeof(ncall));
+        Ds.ZeroCounters();
 
         time_statistics timestat;
         std::vector<double> t_time(ncall);
@@ -987,47 +776,11 @@ int main(int argc, char **argv)
 {
   Grid_init(&argc, &argv);
 
-  int Ls = 1;
-  bool do_su4 = true;
-  bool do_memory = true;
-  bool do_comms = true;
-  bool do_flops = true;
-
-  // NOTE: these two take O((number of ranks)^2) time, which might be a lot, so they are
-  // off by default
-  bool do_latency = false;
-  bool do_p2p = false;
-
   std::string json_filename = ""; // empty indicates no json output
   for (int i = 0; i < argc; i++)
   {
-    auto arg = std::string(argv[i]);
-    if (arg == "--json-out")
+    if (std::string(argv[i]) == "--json-out")
       json_filename = argv[i + 1];
-    if (arg == "--benchmark-su4")
-      do_su4 = true;
-    if (arg == "--benchmark-memory")
-      do_memory = true;
-    if (arg == "--benchmark-comms")
-      do_comms = true;
-    if (arg == "--benchmark-flops")
-      do_flops = true;
-    if (arg == "--benchmark-latency")
-      do_latency = true;
-    if (arg == "--benchmark-p2p")
-      do_p2p = true;
-    if (arg == "--no-benchmark-su4")
-      do_su4 = false;
-    if (arg == "--no-benchmark-memory")
-      do_memory = false;
-    if (arg == "--no-benchmark-comms")
-      do_comms = false;
-    if (arg == "--no-benchmark-flops")
-      do_flops = false;
-    if (arg == "--no-benchmark-latency")
-      do_latency = false;
-    if (arg == "--no-benchmark-p2p")
-      do_p2p = false;
   }
 
   CartesianCommunicator::SetCommunicatorPolicy(
@@ -1039,6 +792,12 @@ int main(int argc, char **argv)
 #endif
   Benchmark::Decomposition();
 
+  int do_su4 = 1;
+  int do_memory = 1;
+  int do_comms = 1;
+  int do_flops = 1;
+  int Ls = 1;
+
   int sel = 4;
   std::vector<int> L_list({8, 12, 16, 24, 32});
   int selm1 = sel - 1;
@@ -1071,22 +830,6 @@ int main(int argc, char **argv)
     Benchmark::Comms();
   }
 
-  if (do_latency)
-  {
-    grid_big_sep();
-    std::cout << GridLogMessage << " Latency benchmark " << std::endl;
-    grid_big_sep();
-    Benchmark::Latency();
-  }
-
-  if (do_p2p)
-  {
-    grid_big_sep();
-    std::cout << GridLogMessage << " Point-To-Point benchmark " << std::endl;
-    grid_big_sep();
-    Benchmark::P2P();
-  }
-
   if (do_flops)
   {
     Ls = 1;
@@ -1146,8 +889,6 @@ int main(int argc, char **argv)
     json_results["flops"] = tmp_flops;
   }
 
-  json_results["hostnames"] = get_mpi_hostnames();
-
   if (!json_filename.empty())
   {
     std::cout << GridLogMessage << "writing benchmark results to " << json_filename

Grid/systems/tursa/spack-bootstrap.sh

@@ -4,13 +4,7 @@ set -euo pipefail
 
 gcc_spec='gcc@9.4.0'
 cuda_spec='cuda@11.4.0'
-
-# hdf5 and fftw depend on OpenMPI, which we install manually. To make sure this
-# dependency is picked by spack, we specify the compiler here explicitly. For
-# most other packages we dont really care about the compiler (i.e. system
-# compiler versus ${gcc_spec})
-hdf5_spec="hdf5@1.10.7+cxx+threadsafe%${gcc_spec}"
-fftw_spec="fftw%${gcc_spec}"
+hdf5_spec='hdf5@1.10.7'
 
 if (( $# != 1 )); then
     echo "usage: $(basename "$0") <env dir>" 1>&2
@@ -24,7 +18,7 @@ cd "${cwd}"
 
 # General configuration ########################################################
 # build with 128 tasks
-echo 'config:
+echo 'config:                                  
   build_jobs: 128
   build_stage:
     - $spack/var/spack/stage
@@ -44,23 +38,26 @@ rm external.yaml
 
 # Base compilers ###############################################################
 # configure system base
-
-spack env create base
-spack env activate base
 spack compiler find --scope site
 
-# install GCC, CUDA
-spack add ${gcc_spec} ${cuda_spec}
-spack concretize
-spack env depfile -o Makefile.tmp
-make -j128 -f Makefile.tmp
+# install GCC, CUDA & LLVM
+spack install ${gcc_spec} ${cuda_spec} llvm
+
+spack load llvm
 spack compiler find --scope site
+spack unload llvm
+
+spack load ${gcc_spec}
+spack compiler find --scope site
+spack unload ${gcc_spec}
 
 # Manual compilation of OpenMPI & UCX ##########################################
 # set build directories
 mkdir -p "${dir}"/build
 cd "${dir}"/build
 
+spack load ${gcc_spec} ${cuda_spec}
+
 cuda_path=$(spack find --format "{prefix}" cuda)
 gdrcopy_path=/mnt/lustre/tursafs1/apps/gdrcopy/2.3.1
 
@@ -127,8 +124,8 @@ mkdir build_gpu; cd build_gpu
              --with-cuda="${cuda_path}" --disable-getpwuid        \
              --with-verbs --with-slurm --enable-mpi-fortran=all   \
              --with-pmix=internal --with-libevent=internal
-make -j 128
-make install
+make -j 128 
+make install 
 cd ..
 
 # openmpi cpu build
@@ -144,62 +141,60 @@ make -j 128
 make install
 cd "${dir}"
 
-ucx_spec_gpu="ucx@1.12.0.GPU%${gcc_spec}"
-ucx_spec_cpu="ucx@1.12.0.CPU%${gcc_spec}"
-openmpi_spec_gpu="openmpi@4.1.1.GPU%${gcc_spec}"
-openmpi_spec_cpu="openmpi@4.1.1.CPU%${gcc_spec}"
-
 # Add externals to spack
 echo "packages:
   ucx:
     externals:
-    - spec: \"${ucx_spec_gpu}\"
+    - spec: \"ucx@1.12.0.GPU%gcc@9.4.0\"
       prefix: ${dir}/prefix/ucx_gpu
-    - spec: \"${ucx_spec_cpu}\"
+    - spec: \"ucx@1.12.0.CPU%gcc@9.4.0\"
       prefix: ${dir}/prefix/ucx_cpu
     buildable: False
   openmpi:
     externals:
-    - spec: \"${openmpi_spec_gpu}\"
+    - spec: \"openmpi@4.1.1.GPU%gcc@9.4.0\"
       prefix: ${dir}/prefix/ompi_gpu
-    - spec: \"${openmpi_spec_cpu}\"
+    - spec: \"openmpi@4.1.1.CPU%gcc@9.4.0\"
       prefix: ${dir}/prefix/ompi_cpu
     buildable: False" > spack.yaml
 
 spack config --scope site add -f spack.yaml
 rm spack.yaml
-spack env deactivate
+spack install ucx@1.12.0.GPU%gcc@9.4.0 openmpi@4.1.1.GPU%gcc@9.4.0
+spack install ucx@1.12.0.CPU%gcc@9.4.0 openmpi@4.1.1.CPU%gcc@9.4.0
 
 cd "${cwd}"
 
 # environments #################################################################
 dev_tools=("autoconf" "automake" "libtool" "jq" "git")
+ompi_gpu_hash=$(spack find --format "{hash}" openmpi@4.1.1.GPU)
+ompi_cpu_hash=$(spack find --format "{hash}" openmpi@4.1.1.CPU)
 
 spack env create grid-gpu
 spack env activate grid-gpu
-spack compiler find --scope site
-spack add ${gcc_spec} ${cuda_spec} ${ucx_spec_gpu} ${openmpi_spec_gpu}
-spack add ${hdf5_spec} ${fftw_spec}
-spack add openssl gmp mpfr c-lime "${dev_tools[@]}"
-spack concretize
-spack env depfile -o Makefile.tmp
-make -j128 -f Makefile.tmp
+spack add ${gcc_spec} ${cuda_spec} "${dev_tools[@]}" 
+spack add ucx@1.12.0.GPU%gcc@9.4.0 openmpi@4.1.1.GPU%gcc@9.4.0
+spack add ${hdf5_spec}+cxx+threadsafe ^/"${ompi_gpu_hash}"
+spack add fftw ^/"${ompi_gpu_hash}"
+spack add openssl gmp mpfr c-lime
+spack install
 spack env deactivate
 
 spack env create grid-cpu
 spack env activate grid-cpu
-spack compiler find --scope site
-spack add ${gcc_spec} ${ucx_spec_cpu} ${openmpi_spec_cpu}
-spack add ${hdf5_spec} ${fftw_spec}
-spack add openssl gmp mpfr c-lime "${dev_tools[@]}"
-spack concretize
-spack env depfile -o Makefile.tmp
-make -j128 -f Makefile.tmp
+spack add llvm "${dev_tools[@]}" 
+spack add ucx@1.12.0.CPU%gcc@9.4.0 openmpi@4.1.1.CPU%gcc@9.4.0
+spack add ${hdf5_spec}+cxx+threadsafe ^/"${ompi_cpu_hash}"
+spack add fftw ^/"${ompi_cpu_hash}"
+spack add openssl gmp mpfr c-lime
+spack install
 spack env deactivate
 
+spack install jq git
+
 # Final setup ##################################################################
 spack clean
-#spack gc -y  # "spack gc" tends to get hung up for unknown reasons
+spack gc -y
 
 # add more environment variables in module loading
 spack config --scope site add 'modules:prefix_inspections:lib:[LD_LIBRARY_PATH,LIBRARY_PATH]'

Quda/.clang-format

@@ -0,0 +1,14 @@
+{
+  BasedOnStyle: LLVM,
+  UseTab: Never,
+  IndentWidth: 2,
+  TabWidth: 2,
+  BreakBeforeBraces: Allman,
+  AllowShortIfStatementsOnASingleLine: false,
+  IndentCaseLabels: false,
+  ColumnLimit: 90,
+  AccessModifierOffset: -4,
+  NamespaceIndentation: All,
+  FixNamespaceComments: false,
+  SortIncludes: true,
+}

Quda/Benchmark_Quda.cpp

@@ -0,0 +1,381 @@
+#include <algorithm>
+#include <array>
+#include <blas_quda.h>
+#include <cassert>
+#include <color_spinor_field.h>
+#include <dirac_quda.h>
+#include <gauge_tools.h>
+#include <memory>
+#include <mpi.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+using namespace quda;
+
+// remove to use QUDA's own flop counting instead of Grid's convention
+#define FLOP_COUNTING_GRID
+
+// This is the MPI grid, i.e. the layout of ranks
+int nranks = -1;
+std::array<int, 4> mpi_grid = {1, 1, 1, 1};
+
+void initComms(int argc, char **argv)
+{
+  // init MPI communication
+  MPI_Init(&argc, &argv);
+
+  MPI_Comm_size(MPI_COMM_WORLD, &nranks);
+  assert(1 <= nranks && nranks <= 100000);
+
+  mpi_grid[3] = nranks;
+
+  // this maps coordinates to rank number
+  auto lex_rank_from_coords = [](int const *coords, void *)
+  {
+    int rank = coords[0];
+    for (int i = 1; i < 4; i++)
+      rank = mpi_grid[i] * rank + coords[i];
+    return rank;
+  };
+
+  initCommsGridQuda(4, mpi_grid.data(), lex_rank_from_coords, nullptr);
+
+  for (int d = 0; d < 4; d++)
+    if (mpi_grid[d] > 1)
+      commDimPartitionedSet(d);
+}
+
+// creates a random gauge field. L = local(!) size
+cudaGaugeField make_gauge_field(int L)
+{
+  GaugeFieldParam param;
+
+  // dimension and type of the lattice object
+  param.nDim = 4;
+  param.x[0] = L;
+  param.x[1] = L;
+  param.x[2] = L;
+  param.x[3] = L;
+
+  // number of colors. potentially confusingly, QUDA sometimes uses the word "color" to
+  // things unrelated with physical color. things like "nColor=32" do pop up in deflation
+  // solvers where it (to my understanding) refers to the number of (parallely processed)
+  // deflation vectors.
+  param.nColor = 3;
+
+  // boundary conditions (dont really care for benchmark)
+  param.t_boundary = QUDA_PERIODIC_T;
+
+  // for this benchmark we only need "SINGLE" and/or "DOUBLE" precision. But smaller
+  // precisions are available in QUDA too
+  param.setPrecision(QUDA_SINGLE_PRECISION);
+
+  // no even/odd subset, we want a full lattice
+  param.siteSubset = QUDA_FULL_SITE_SUBSET;
+
+  // what kind of 3x3 matrices the field contains. A proper gauge field has SU(3)
+  // matrices, but (for example) smeared/thick links could have non-unitary links.
+  param.link_type = QUDA_SU3_LINKS;
+
+  // "NULL" does not initialize the field upon creation, "ZERO" would set everything to 0
+  param.create = QUDA_NULL_FIELD_CREATE;
+
+  // field should be allocated directly on the accelerator/GPU
+  param.location = QUDA_CUDA_FIELD_LOCATION;
+
+  // "reconstruct" here means reconstructing a SU(3) matrix from fewer than 18 real
+  // numbers (=3x3 complex numbers). Great feature in production (saving
+  // memory/cache/network bandwidth), not used for this benchmark.
+  param.reconstruct = QUDA_RECONSTRUCT_NO;
+
+  // "ghostExchange" would often be called "halo exchange" outside of Quda. This has
+  // nothing to do with ghost fields from continuum/perturbative qcd.
+  param.ghostExchange = QUDA_GHOST_EXCHANGE_NO;
+
+  // This controls the physical order of elements. "float2" is the the default
+  param.order = QUDA_FLOAT2_GAUGE_ORDER;
+
+  // this means the field is a LORENTZ vector (which a gauge field must be). Has nothing
+  // to do with spin.
+  param.geometry = QUDA_VECTOR_GEOMETRY;
+
+  // create the field and fill with random SU(3) matrices
+  // std::cout << param << std::endl; // double-check parameters
+  auto U = cudaGaugeField(param);
+  gaugeGauss(U, /*seed=*/1234, 1.0);
+  return U;
+}
+
+// create a random source vector (L = local size)
+ColorSpinorField make_source(int L, int Ls = 1)
+{
+  // NOTE: `param.x` directly determines the size of the (local, per rank) memory
+  // allocation. Thus for checkerboarding, we have to specifly x=(L/2,L,L,L) to get a
+  // physical local volume of L^4, thus implicity choosing a dimension for the
+  // checkerboarding (shouldnt really matter of course which one).
+  ColorSpinorParam param;
+  param.nColor = 3;
+  param.nSpin = 4;
+  param.nVec = 1; // only a single vector
+  param.pad = 0;
+  param.siteSubset = QUDA_PARITY_SITE_SUBSET;
+  param.nDim = Ls == 1 ? 4 : 5;
+  param.x[0] = L / 2;
+  param.x[1] = L;
+  param.x[2] = L;
+  param.x[3] = L;
+  param.x[4] = Ls;
+  param.pc_type = QUDA_4D_PC;
+  param.siteOrder = QUDA_EVEN_ODD_SITE_ORDER;
+
+  // somewhat surprisingly, the DiracWilson::Dslash(...) function only works with the
+  // UKQCD_GAMMA_BASIS
+  param.gammaBasis = QUDA_UKQCD_GAMMA_BASIS;
+
+  param.create = QUDA_NULL_FIELD_CREATE; // do not (zero-) initilize the field
+  param.setPrecision(QUDA_SINGLE_PRECISION);
+  param.location = QUDA_CUDA_FIELD_LOCATION;
+
+  // create the field and fill it with random values
+  auto src = ColorSpinorField(param);
+  quda::RNG rng(src, 1234);
+  spinorNoise(src, rng, QUDA_NOISE_GAUSS);
+  /*printfQuda(
+      "created src with norm = %f (sanity check: should be close to %f) and %f bytes\n",
+      blas::norm2(src), 2.0 * 12 * geom[0] * geom[1] * geom[2] * geom[3],
+      src.Bytes() * 1.0);*/
+  // src.PrintDims();
+
+  return src;
+}
+
+void benchmark_wilson()
+{
+  int niter = 20;
+  int niter_warmup = 10;
+
+  printfQuda("==================== wilson dirac operator ====================\n");
+#ifdef FLOP_COUNTING_GRID
+  printfQuda("IMPORTANT: flop counting as in Benchmark_Grid\n");
+#else
+  printfQuda("IMPORTANT: flop counting by QUDA's own convention (different from "
+             "Benchmark_Grid)\n");
+#endif
+  printfQuda("%5s %15s %15s\n", "L", "time (usec)", "Gflop/s/rank");
+
+  for (int L : {8, 12, 16, 24, 32, 48})
+  {
+    auto U = make_gauge_field(L);
+    auto src = make_source(L);
+
+    // create (Wilson) dirac operator
+    DiracParam param;
+    param.kappa = 0.10;
+    param.dagger = QUDA_DAG_NO;
+    param.matpcType = QUDA_MATPC_EVEN_EVEN;
+    auto dirac = DiracWilson(param);
+
+    // insert gauge field into the dirac operator
+    // (the additional nullptr's are for smeared links and fancy preconditioners and such.
+    // Not used for simple Wilson fermions)
+    dirac.updateFields(&U, nullptr, nullptr, nullptr);
+
+    auto tmp = ColorSpinorField(ColorSpinorParam(src));
+
+    // couple iterations without timing to warm up
+    for (int iter = 0; iter < niter_warmup; ++iter)
+      dirac.Dslash(tmp, src, QUDA_EVEN_PARITY);
+
+    // actual benchmark with timings
+    dirac.Flops(); // reset flops counter
+    device_timer_t device_timer;
+    device_timer.start();
+    for (int iter = 0; iter < niter; ++iter)
+      dirac.Dslash(tmp, src, QUDA_EVEN_PARITY);
+    device_timer.stop();
+
+    double secs = device_timer.last() / niter;
+
+#ifdef FLOP_COUNTING_GRID
+    // this is the flop counting from Benchmark_Grid
+    double Nc = 3;
+    double Nd = 4;
+    double Ns = 4;
+    double flops =
+        (Nc * (6 + (Nc - 1) * 8) * Ns * Nd + 2 * Nd * Nc * Ns + 2 * Nd * Nc * Ns * 2);
+    flops *= L * L * L * L / 2.0;
+#else
+    double flops = 1.0 * dirac.Flops() / niter;
+#endif
+
+    printfQuda("%5d %15.2f %15.2f\n", L, secs * 1e6, flops / secs * 1e-9);
+  }
+}
+
+void benchmark_dwf()
+{
+  int niter = 20;
+  int niter_warmup = 10;
+
+  printfQuda("==================== domain wall dirac operator ====================\n");
+#ifdef FLOP_COUNTING_GRID
+  printfQuda("IMPORTANT: flop counting as in Benchmark_Grid\n");
+#else
+  printfQuda("IMPORTANT: flop counting by QUDA's own convention (different from "
+             "Benchmark_Grid)\n");
+#endif
+  printfQuda("%5s %15s %15s\n", "L", "time (usec)", "Gflop/s/rank");
+  int Ls = 12;
+  for (int L : {8, 12, 16, 24, 32, 48})
+  {
+    auto U = make_gauge_field(L);
+    auto src = make_source(L, Ls);
+
+    // create dirac operator
+    DiracParam param;
+    param.kappa = 0.10;
+    param.Ls = Ls;
+    param.m5 = 0.1;
+    param.dagger = QUDA_DAG_NO;
+    param.matpcType = QUDA_MATPC_EVEN_EVEN;
+    auto dirac = DiracDomainWall(param);
+
+    // insert gauge field into the dirac operator
+    // (the additional nullptr's are for smeared links and fancy preconditioners and such)
+    dirac.updateFields(&U, nullptr, nullptr, nullptr);
+
+    auto tmp = ColorSpinorField(ColorSpinorParam(src));
+
+    // couple iterations without timing to warm up
+    for (int iter = 0; iter < niter_warmup; ++iter)
+      dirac.Dslash(tmp, src, QUDA_EVEN_PARITY);
+
+    // actual benchmark with timings
+    dirac.Flops(); // reset flops counter
+    device_timer_t device_timer;
+    device_timer.start();
+    for (int iter = 0; iter < niter; ++iter)
+      dirac.Dslash(tmp, src, QUDA_EVEN_PARITY);
+    device_timer.stop();
+
+    double secs = device_timer.last() / niter;
+
+#ifdef FLOP_COUNTING_GRID
+    // this is the flop counting from Benchmark_Grid
+    double Nc = 3;
+    double Nd = 4;
+    double Ns = 4;
+    double flops =
+        (Nc * (6 + (Nc - 1) * 8) * Ns * Nd + 2 * Nd * Nc * Ns + 2 * Nd * Nc * Ns * 2);
+    flops *= L * L * L * L * Ls / 2.0;
+#else
+    double flops = 1.0 * dirac.Flops() / niter;
+#endif
+
+    printfQuda("%5d %15.2f %15.2f\n", L, secs * 1e6, flops / secs * 1e-9);
+  }
+}
+
+void benchmark_axpy()
+{
+  // number of iterations for warmup / measurement
+  // (feel free to change for noise/time tradeoff)
+  constexpr int niter_warmup = 10;
+  constexpr int niter = 20;
+
+  printfQuda("==================== axpy / memory ====================\n");
+
+  ColorSpinorParam param;
+  param.nDim = 4;   // 4-dimensional lattice
+  param.x[4] = 1;   // no fifth dimension
+  param.nColor = 3; // supported values for nSpin/nColor are configured when compiling
+                    // QUDA. "3*4" will probably always be enabled, so we stick with this
+  param.nSpin = 4;
+  param.nVec = 1;                            // just a single vector
+  param.siteSubset = QUDA_FULL_SITE_SUBSET;  // full lattice = no odd/even
+  param.pad = 0;                             // no padding
+  param.create = QUDA_NULL_FIELD_CREATE;     // do not (zero-) initilize the field
+  param.location = QUDA_CUDA_FIELD_LOCATION; // field should reside on GPU
+  param.setPrecision(QUDA_SINGLE_PRECISION);
+
+  // the following dont matter for an axpy benchmark, but need to choose something
+  param.pc_type = QUDA_4D_PC;
+  param.siteOrder = QUDA_EVEN_ODD_SITE_ORDER;
+  param.gammaBasis = QUDA_DEGRAND_ROSSI_GAMMA_BASIS;
+
+  printfQuda("%5s %15s %15s %15s %15s\n", "L", "size (MiB/rank)", "time (usec)",
+             "GiB/s/rank", "Gflop/s/rank");
+  std::vector L_list = {8, 12, 16, 24, 32};
+  for (int L : L_list)
+  {
+    // IMPORTANT: all of `param.x`, `field_elements`, `field.Bytes()`
+    //            are LOCAL, i.e. per rank / per GPU
+
+    param.x[0] = L;
+    param.x[1] = L;
+    param.x[2] = L;
+    param.x[3] = L;
+
+    // number of (real) elements in one (local) field
+    size_t field_elements = 2 * param.x[0] * param.x[1] * param.x[2] * param.x[3] *
+                            param.nColor * param.nSpin;
+
+    // create the field(s)
+    auto fieldA = ColorSpinorField(param);
+    auto fieldB = ColorSpinorField(param);
+    assert(fieldA.Bytes() == sizeof(float) * field_elements); // sanity check
+    assert(fieldB.Bytes() == sizeof(float) * field_elements); // sanity check
+
+    // fill fields with random values
+    quda::RNG rng(fieldA, 1234);
+    spinorNoise(fieldA, rng, QUDA_NOISE_GAUSS);
+    spinorNoise(fieldB, rng, QUDA_NOISE_GAUSS);
+
+    // number of operations / bytes per iteration
+    // axpy is one addition, one multiplication, two read, one write
+    double flops = 2 * field_elements;
+    double memory = 3 * sizeof(float) * field_elements;
+
+    // do some iterations to to let QUDA do its internal tuning and also stabilize cache
+    // behaviour and such
+    for (int iter = 0; iter < niter_warmup; ++iter)
+      blas::axpy(1.234, fieldA, fieldB);
+
+    // running the actual benchmark
+    device_timer_t device_timer;
+    device_timer.start();
+    for (int iter = 0; iter < niter; ++iter)
+      blas::axpy(1.234, fieldA, fieldB);
+    device_timer.stop();
+    double secs = device_timer.last() / niter; // seconds per iteration
+
+    printfQuda("%5d %15.2f %15.2f %15.2f %15.2f\n", L, memory / 1024. / 1024., secs * 1e6,
+               memory / secs / 1024. / 1024. / 1024., flops / secs * 1e-9);
+  }
+}
+
+int main(int argc, char **argv)
+{
+  initComms(argc, argv);
+
+  initQuda(-1); // -1 for multi-gpu. otherwise this selects the device to be used
+
+  //  verbosity options are:
+  //  SILENT, SUMMARIZE, VERBOSE, DEBUG_VERBOSE
+  setVerbosity(QUDA_SUMMARIZE);
+
+  printfQuda("MPI layout = %d %d %d %d\n", mpi_grid[0], mpi_grid[1], mpi_grid[2],
+             mpi_grid[3]);
+
+  benchmark_axpy();
+
+  setVerbosity(QUDA_SILENT);
+  benchmark_wilson();
+  benchmark_dwf();
+  setVerbosity(QUDA_SUMMARIZE);
+
+  printfQuda("==================== done with all benchmarks ====================\n");
+  endQuda();
+  quda::comm_finalize();
+  MPI_Finalize();
+}

Quda/build.sh

@@ -0,0 +1,10 @@
+#!/bin/bash
+#CXX=/home/dp207/dp207/shared/env/versions/220428/spack/opt/spack/linux-rhel8-zen/gcc-8.4.1/gcc-9.4.0-g3vyv3te4ah634euh7phyokb3fiurprp/bin/g++
+QUDA_BUILD=/home/dp207/dp207/dc-burg2/quda_build
+QUDA_SRC=/home/dp207/dp207/dc-burg2/quda
+#QUDA_BUILD=
+
+FLAGS="-DMPI_COMMS -DMULTI_GPU -DQUDA_PRECISION=14 -DQUDA_RECONSTRUCT=7 -g -O3 -Wall -Wextra -std=c++17 "
+$CXX $FLAGS -I$QUDA_BUILD/include/targets/cuda -I$QUDA_SRC/include -I$QUDA_BUILD/include -isystem $QUDA_SRC/include/externals -isystem $QUDA_BUILD/_deps/eigen-src -c -o Benchmark_Quda.o  Benchmark_Quda.cpp
+LINK_FLAGS="-Wl,-rpath,$QUDA_BUILD/tests:$QUDA_BUILD/lib:/home/dp207/dp207/shared/env/versions/220428/spack/opt/spack/linux-rhel8-zen2/gcc-9.4.0/cuda-11.4.0-etxow4jb23qdbs7j6txczy44cdatpj22/lib64/stubs: $QUDA_BUILD/lib/libquda.so /home/dp207/dp207/shared/env/versions/220428/spack/opt/spack/linux-rhel8-zen2/gcc-9.4.0/cuda-11.4.0-etxow4jb23qdbs7j6txczy44cdatpj22/lib64/stubs/libcuda.so /home/dp207/dp207/shared/env/versions/220428/spack/opt/spack/linux-rhel8-zen2/gcc-9.4.0/cuda-11.4.0-etxow4jb23qdbs7j6txczy44cdatpj22/lib64/stubs/libnvidia-ml.so /home/dp207/dp207/shared/env/versions/220428/spack/opt/spack/linux-rhel8-zen2/gcc-9.4.0/cuda-11.4.0-etxow4jb23qdbs7j6txczy44cdatpj22/lib64/libcudart_static.a -ldl /usr/lib64/librt.so /home/dp207/dp207/shared/env/versions/220428/spack/opt/spack/linux-rhel8-zen2/gcc-9.4.0/cuda-11.4.0-etxow4jb23qdbs7j6txczy44cdatpj22/lib64/libcublas.so /home/dp207/dp207/shared/env/versions/220428/spack/opt/spack/linux-rhel8-zen2/gcc-9.4.0/cuda-11.4.0-etxow4jb23qdbs7j6txczy44cdatpj22/lib64/libcufft.so -lpthread"
+$CXX -g -O3 Benchmark_Quda.o -o Benchmark_Quda $LINK_FLAGS -lmpi

Quda/env.sh

@@ -0,0 +1,21 @@
+module load gcc/9.3.0
+module load cuda/11.4.1
+module load openmpi/4.1.1-cuda11.4
+
+export QUDA_RESOURCE_PATH=$(pwd)/tuning
+export OMP_NUM_THREADS=4
+export OMPI_MCA_btl=^uct,openib
+export OMPI_MCA_pml=ucx # by fabian. no idea what this is
+#export UCX_TLS=rc,rc_x,sm,cuda_copy,cuda_ipc,gdr_copy
+export UCX_TLS=gdr_copy,rc,rc_x,sm,cuda_copy,cuda_ipc
+export UCX_RNDV_THRESH=16384
+export UCX_RNDV_SCHEME=put_zcopy
+export UCX_IB_GPU_DIRECT_RDMA=yes
+export UCX_MEMTYPE_CACHE=n
+
+export OMPI_MCA_io=romio321
+export OMPI_MCA_btl_openib_allow_ib=true
+export OMPI_MCA_btl_openib_device_type=infiniband
+export OMPI_MCA_btl_openib_if_exclude=mlx5_1,mlx5_2,mlx5_3
+
+export QUDA_REORDER_LOCATION=GPU # this is the default anyway