Merge pull request 'Updates GPU launch wrapper and install script' (#8 ) from aturner/lattice-benchmarks:main into main

Reviewed-on: portelli/lattice-benchmarks#8
Updates knem location in install script
2025-03-31 16:08:45 +01:00 · 2024-12-17 13:46:43 +00:00 · 2024-11-26 10:57:28 +00:00 · 2024-11-25 15:45:59 +00:00 · 2024-11-25 15:45:59 +00:00 · 2024-11-25 15:45:58 +00:00
23 changed files with 1542 additions and 1334 deletions
--- a/Grid/Benchmark_Grid.cpp
+++ b/Grid/Benchmark_Grid.cpp
--- a/Grid/Benchmark_IO.cpp
+++ b/Grid/Benchmark_IO.cpp
@ -32,7 +32,7 @@ along with this program. If not, see <http://www.gnu.org/licenses/>.
 #ifdef HAVE_LIME
 using namespace Grid;

-std::string filestem(const int l) { return "iobench_l" + std::to_string(l); }
+std::string filestem(const int l) { return "io/iobench_l" + std::to_string(l); }

 int vol(const int i) { return BENCH_IO_LMIN + 2 * i; }

@ -56,13 +56,6 @@ template <typename Mat> void stats(Mat &mean, Mat &stdDev, const std::vector<Mat
  mean /= n;
 }

-#define grid_printf(...)                                                                 \
-  {                                                                                      \
-    char _buf[1024];                                                                     \
-    sprintf(_buf, __VA_ARGS__);                                                          \
-    MSG << _buf;                                                                         \
-  }
-
 enum
 {
  sRead = 0,
@ -83,58 +76,58 @@ int main(int argc, char **argv)
  std::vector<Eigen::VectorXd> avPerf(BENCH_IO_NPASS, Eigen::VectorXd::Zero(4));
  std::vector<int> latt;

-  MSG << "Grid is setup to use " << threads << " threads" << std::endl;
-  MSG << "MPI partition " << mpi << std::endl;
+  GRID_MSG << "Grid is setup to use " << threads << " threads" << std::endl;
+  GRID_MSG << "MPI partition " << mpi << std::endl;
  for (unsigned int i = 0; i < BENCH_IO_NPASS; ++i)
  {
-    MSG << BIGSEP << std::endl;
-    MSG << "Pass " << i + 1 << "/" << BENCH_IO_NPASS << std::endl;
-    MSG << BIGSEP << std::endl;
-    MSG << SEP << std::endl;
-    MSG << "Benchmark std write" << std::endl;
-    MSG << SEP << std::endl;
+    grid_big_sep();
+    GRID_MSG << "Pass " << i + 1 << "/" << BENCH_IO_NPASS << std::endl;
+    grid_big_sep();
+    grid_small_sep();
+    GRID_MSG << "Benchmark std write" << std::endl;
+    grid_small_sep();
    for (int l = BENCH_IO_LMIN; l <= BENCH_IO_LMAX; l += 2)
    {
      latt = {l * mpi[0], l * mpi[1], l * mpi[2], l * mpi[3]};

-      MSG << "-- Local volume " << l << "^4" << std::endl;
+      GRID_MSG << "-- Local volume " << l << "^4" << std::endl;
      writeBenchmark<LatticeFermion>(latt, filestem(l), stdWrite<LatticeFermion>);
      perf[i](volInd(l), sWrite) = BinaryIO::lastPerf.mbytesPerSecond;
    }

-    MSG << SEP << std::endl;
-    MSG << "Benchmark std read" << std::endl;
-    MSG << SEP << std::endl;
+    grid_small_sep();
+    GRID_MSG << "Benchmark std read" << std::endl;
+    grid_small_sep();
    for (int l = BENCH_IO_LMIN; l <= BENCH_IO_LMAX; l += 2)
    {
      latt = {l * mpi[0], l * mpi[1], l * mpi[2], l * mpi[3]};

-      MSG << "-- Local volume " << l << "^4" << std::endl;
+      GRID_MSG << "-- Local volume " << l << "^4" << std::endl;
      readBenchmark<LatticeFermion>(latt, filestem(l), stdRead<LatticeFermion>);
      perf[i](volInd(l), sRead) = BinaryIO::lastPerf.mbytesPerSecond;
    }

 #ifdef HAVE_LIME
-    MSG << SEP << std::endl;
-    MSG << "Benchmark Grid C-Lime write" << std::endl;
-    MSG << SEP << std::endl;
+    grid_small_sep();
+    GRID_MSG << "Benchmark Grid C-Lime write" << std::endl;
+    grid_small_sep();
    for (int l = BENCH_IO_LMIN; l <= BENCH_IO_LMAX; l += 2)
    {
      latt = {l * mpi[0], l * mpi[1], l * mpi[2], l * mpi[3]};

-      MSG << "-- Local volume " << l << "^4" << std::endl;
+      GRID_MSG << "-- Local volume " << l << "^4" << std::endl;
      writeBenchmark<LatticeFermion>(latt, filestem(l), limeWrite<LatticeFermion>);
      perf[i](volInd(l), gWrite) = BinaryIO::lastPerf.mbytesPerSecond;
    }

-    MSG << SEP << std::endl;
-    MSG << "Benchmark Grid C-Lime read" << std::endl;
-    MSG << SEP << std::endl;
+    grid_small_sep();
+    GRID_MSG << "Benchmark Grid C-Lime read" << std::endl;
+    grid_small_sep();
    for (int l = BENCH_IO_LMIN; l <= BENCH_IO_LMAX; l += 2)
    {
      latt = {l * mpi[0], l * mpi[1], l * mpi[2], l * mpi[3]};

-      MSG << "-- Local volume " << l << "^4" << std::endl;
+      GRID_MSG << "-- Local volume " << l << "^4" << std::endl;
      readBenchmark<LatticeFermion>(latt, filestem(l), limeRead<LatticeFermion>);
      perf[i](volInd(l), gRead) = BinaryIO::lastPerf.mbytesPerSecond;
    }
@ -159,13 +152,13 @@ int main(int argc, char **argv)
  avRob.fill(100.);
  avRob -= 100. * avStdDev.cwiseQuotient(avMean.cwiseAbs());

-  MSG << BIGSEP << std::endl;
-  MSG << "SUMMARY" << std::endl;
-  MSG << BIGSEP << std::endl;
-  MSG << "Summary of individual results (all results in MB/s)." << std::endl;
-  MSG << "Every second colum gives the standard deviation of the previous column."
-      << std::endl;
-  MSG << std::endl;
+  grid_big_sep();
+  GRID_MSG << "SUMMARY" << std::endl;
+  grid_big_sep();
+  GRID_MSG << "Summary of individual results (all results in MB/s)." << std::endl;
+  GRID_MSG << "Every second colum gives the standard deviation of the previous column."
+           << std::endl;
+  GRID_MSG << std::endl;
  grid_printf("%4s %12s %12s %12s %12s %12s %12s %12s %12s\n", "L", "std read", "std dev",
              "std write", "std dev", "Grid read", "std dev", "Grid write", "std dev");
  for (int l = BENCH_IO_LMIN; l <= BENCH_IO_LMAX; l += 2)
@ -176,10 +169,10 @@ int main(int argc, char **argv)
                stdDev(volInd(l), gRead), mean(volInd(l), gWrite),
                stdDev(volInd(l), gWrite));
  }
-  MSG << std::endl;
-  MSG << "Robustness of individual results, in %. (rob = 100% - std dev / mean)"
-      << std::endl;
-  MSG << std::endl;
+  GRID_MSG << std::endl;
+  GRID_MSG << "Robustness of individual results, in %. (rob = 100% - std dev / mean)"
+           << std::endl;
+  GRID_MSG << std::endl;
  grid_printf("%4s %12s %12s %12s %12s\n", "L", "std read", "std write", "Grid read",
              "Grid write");
  for (int l = BENCH_IO_LMIN; l <= BENCH_IO_LMAX; l += 2)
@ -187,21 +180,21 @@ int main(int argc, char **argv)
    grid_printf("%4d %12.1f %12.1f %12.1f %12.1f\n", l, rob(volInd(l), sRead),
                rob(volInd(l), sWrite), rob(volInd(l), gRead), rob(volInd(l), gWrite));
  }
-  MSG << std::endl;
-  MSG << "Summary of results averaged over local volumes 24^4-" << BENCH_IO_LMAX
-      << "^4 (all results in MB/s)." << std::endl;
-  MSG << "Every second colum gives the standard deviation of the previous column."
-      << std::endl;
-  MSG << std::endl;
+  GRID_MSG << std::endl;
+  GRID_MSG << "Summary of results averaged over local volumes 24^4-" << BENCH_IO_LMAX
+           << "^4 (all results in MB/s)." << std::endl;
+  GRID_MSG << "Every second colum gives the standard deviation of the previous column."
+           << std::endl;
+  GRID_MSG << std::endl;
  grid_printf("%12s %12s %12s %12s %12s %12s %12s %12s\n", "std read", "std dev",
              "std write", "std dev", "Grid read", "std dev", "Grid write", "std dev");
  grid_printf("%12.1f %12.1f %12.1f %12.1f %12.1f %12.1f %12.1f %12.1f\n", avMean(sRead),
              avStdDev(sRead), avMean(sWrite), avStdDev(sWrite), avMean(gRead),
              avStdDev(gRead), avMean(gWrite), avStdDev(gWrite));
-  MSG << std::endl;
-  MSG << "Robustness of volume-averaged results, in %. (rob = 100% - std dev / mean)"
-      << std::endl;
-  MSG << std::endl;
+  GRID_MSG << std::endl;
+  GRID_MSG << "Robustness of volume-averaged results, in %. (rob = 100% - std dev / mean)"
+           << std::endl;
+  GRID_MSG << std::endl;
  grid_printf("%12s %12s %12s %12s\n", "std read", "std write", "Grid read",
              "Grid write");
  grid_printf("%12.1f %12.1f %12.1f %12.1f\n", avRob(sRead), avRob(sWrite), avRob(gRead),
--- a/Grid/Benchmark_IO.hpp
+++ b/Grid/Benchmark_IO.hpp
@ -18,12 +18,8 @@ along with this program. If not, see <http://www.gnu.org/licenses/>.
 #ifndef Benchmark_IO_hpp_
 #define Benchmark_IO_hpp_

+#include "Common.hpp"
 #include <Grid/Grid.h>
-#define MSG std::cout << GridLogMessage
-#define SEP                                                                              \
-  "-----------------------------------------------------------------------------"
-#define BIGSEP                                                                           \
-  "============================================================================="
 #ifdef HAVE_LIME

 namespace Grid
@ -50,9 +46,9 @@ namespace Grid
  //   crc = GridChecksum::crc32(vec_v.cpu_ptr, size);
  //   std::fwrite(&crc, sizeof(uint32_t), 1, file);
  //   crcWatch.Stop();
-  //   MSG << "Std I/O write: Data CRC32 " << std::hex << crc << std::dec << std::endl;
-  //   ioWatch.Start();
-  //   std::fwrite(vec_v.cpu_ptr, sizeof(typename Field::scalar_object),
+  //   GRID_MSG << "Std I/O write: Data CRC32 " << std::hex << crc << std::dec <<
+  //   std::endl; ioWatch.Start(); std::fwrite(vec_v.cpu_ptr, sizeof(typename
+  //   Field::scalar_object),
  //               vec.Grid()->lSites(), file);
  //   ioWatch.Stop();
  //   std::fclose(file);
@ -61,11 +57,11 @@ namespace Grid
  //   p.size = size;
  //   p.time = ioWatch.useconds();
  //   p.mbytesPerSecond = size / 1024. / 1024. / (ioWatch.useconds() / 1.e6);
-  //   MSG << "Std I/O write: Wrote " << p.size << " bytes in " << ioWatch.Elapsed()
+  //   GRID_MSG << "Std I/O write: Wrote " << p.size << " bytes in " << ioWatch.Elapsed()
  //       << ",
  //          "
  //       << p.mbytesPerSecond << " MB/s" << std::endl;
-  //   MSG << "Std I/O write: checksum overhead " << crcWatch.Elapsed() << std::endl;
+  //   GRID_MSG << "Std I/O write: checksum overhead " << crcWatch.Elapsed() << std::endl;
  // }

  // template <typename Field> void stdRead(Field &vec, const std::string filestem)
@ -94,16 +90,14 @@ namespace Grid
  //     crcData = GridChecksum::crc32(vec_v.cpu_ptr, size);
  //     crcWatch.Stop();
  //   }
-  //   MSG << "Std I/O read: Data CRC32 " << std::hex << crcData << std::dec << std::endl;
-  //   assert(crcData == crcRead);
-  //   size *= vec.Grid()->ProcessorCount();
-  //   auto &p = BinaryIO::lastPerf;
-  //   p.size = size;
-  //   p.time = ioWatch.useconds();
+  //   GRID_MSG << "Std I/O read: Data CRC32 " << std::hex << crcData << std::dec <<
+  //   std::endl; assert(crcData == crcRead); size *= vec.Grid()->ProcessorCount(); auto
+  //   &p = BinaryIO::lastPerf; p.size = size; p.time = ioWatch.useconds();
  //   p.mbytesPerSecond = size / 1024. / 1024. / (ioWatch.useconds() / 1.e6);
-  //   MSG << "Std I/O read: Read " << p.size << " bytes in " << ioWatch.Elapsed() << ", "
+  //   GRID_MSG << "Std I/O read: Read " << p.size << " bytes in " << ioWatch.Elapsed() <<
+  //   ", "
  //       << p.mbytesPerSecond << " MB/s" << std::endl;
-  //   MSG << "Std I/O read: checksum overhead " << crcWatch.Elapsed() << std::endl;
+  //   GRID_MSG << "Std I/O read: checksum overhead " << crcWatch.Elapsed() << std::endl;
  // }

  template <typename Field> void stdWrite(const std::string filestem, Field &vec)
@ -122,7 +116,7 @@ namespace Grid
    crc = GridChecksum::crc32(vec_v.cpu_ptr, size);
    file.write(reinterpret_cast<char *>(&crc), sizeof(uint32_t) / sizeof(char));
    crcWatch.Stop();
-    MSG << "Std I/O write: Data CRC32 " << std::hex << crc << std::dec << std::endl;
+    GRID_MSG << "Std I/O write: Data CRC32 " << std::hex << crc << std::dec << std::endl;
    ioWatch.Start();
    file.write(reinterpret_cast<char *>(vec_v.cpu_ptr), sizec);
    file.flush();
@ -132,9 +126,9 @@ namespace Grid
    p.size = size;
    p.time = ioWatch.useconds();
    p.mbytesPerSecond = size / 1024. / 1024. / (ioWatch.useconds() / 1.e6);
-    MSG << "Std I/O write: Wrote " << p.size << " bytes in " << ioWatch.Elapsed() << ", "
-        << p.mbytesPerSecond << " MB/s" << std::endl;
-    MSG << "Std I/O write: checksum overhead " << crcWatch.Elapsed() << std::endl;
+    GRID_MSG << "Std I/O write: Wrote " << p.size << " bytes in " << ioWatch.Elapsed()
+             << ", " << p.mbytesPerSecond << " MB/s" << std::endl;
+    GRID_MSG << "Std I/O write: checksum overhead " << crcWatch.Elapsed() << std::endl;
  }

  template <typename Field> void stdRead(Field &vec, const std::string filestem)
@ -163,16 +157,17 @@ namespace Grid
      crcData = GridChecksum::crc32(vec_v.cpu_ptr, size);
      crcWatch.Stop();
    }
-    MSG << "Std I/O read: Data CRC32 " << std::hex << crcData << std::dec << std::endl;
+    GRID_MSG << "Std I/O read: Data CRC32 " << std::hex << crcData << std::dec
+             << std::endl;
    assert(crcData == crcRead);
    size *= vec.Grid()->ProcessorCount();
    auto &p = BinaryIO::lastPerf;
    p.size = size;
    p.time = ioWatch.useconds();
    p.mbytesPerSecond = size / 1024. / 1024. / (ioWatch.useconds() / 1.e6);
-    MSG << "Std I/O read: Read " << p.size << " bytes in " << ioWatch.Elapsed() << ", "
-        << p.mbytesPerSecond << " MB/s" << std::endl;
-    MSG << "Std I/O read: checksum overhead " << crcWatch.Elapsed() << std::endl;
+    GRID_MSG << "Std I/O read: Read " << p.size << " bytes in " << ioWatch.Elapsed()
+             << ", " << p.mbytesPerSecond << " MB/s" << std::endl;
+    GRID_MSG << "Std I/O read: checksum overhead " << crcWatch.Elapsed() << std::endl;
  }

  template <typename Field> void limeWrite(const std::string filestem, Field &vec)
--- a/Grid/Benchmark_comms_host_device.cpp
+++ b/Grid/Benchmark_comms_host_device.cpp
@ -1,265 +0,0 @@
-/*
-Copyright © 2015 Peter Boyle <paboyle@ph.ed.ac.uk>
-Copyright © 2022 Antonin Portelli <antonin.portelli@me.com>
-
-This program is free software; you can redistribute it and/or
-modify it under the terms of the GNU General Public License
-as published by the Free Software Foundation; either version 2
-of the License, or (at your option) any later version.
-
-This program is distributed in the hope that it will be useful,
-but WITHOUT ANY WARRANTY; without even the implied warranty of
-MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-GNU General Public License for more details.
-
-You should have received a copy of the GNU General Public License
-along with this program. If not, see <http://www.gnu.org/licenses/>.
-*/
-
-#include <Grid/Grid.h>
-
-using namespace std;
-using namespace Grid;
-
-struct time_statistics
-{
-  double mean;
-  double err;
-  double min;
-  double max;
-
-  void statistics(std::vector<double> v)
-  {
-    double sum = std::accumulate(v.begin(), v.end(), 0.0);
-    mean = sum / v.size();
-
-    std::vector<double> diff(v.size());
-    std::transform(v.begin(), v.end(), diff.begin(), [=](double x) { return x - mean; });
-    double sq_sum = std::inner_product(diff.begin(), diff.end(), diff.begin(), 0.0);
-    err = std::sqrt(sq_sum / (v.size() * (v.size() - 1)));
-
-    auto result = std::minmax_element(v.begin(), v.end());
-    min = *result.first;
-    max = *result.second;
-  }
-};
-
-void header()
-{
-  std::cout << GridLogMessage << " L  "
-            << "\t"
-            << " Ls  "
-            << "\t" << std::setw(11) << "bytes\t\t"
-            << "MB/s uni"
-            << "\t"
-            << "MB/s bidi" << std::endl;
-};
-
-int main(int argc, char **argv)
-{
-  Grid_init(&argc, &argv);
-
-  Coordinate simd_layout = GridDefaultSimd(Nd, vComplexD::Nsimd());
-  Coordinate mpi_layout = GridDefaultMpi();
-  int threads = GridThread::GetThreads();
-  std::cout << GridLogMessage << "Grid is setup to use " << threads << " threads"
-            << std::endl;
-
-  int Nloop = 250;
-  int nmu = 0;
-  int maxlat = 32;
-  for (int mu = 0; mu < Nd; mu++)
-    if (mpi_layout[mu] > 1)
-      nmu++;
-
-  std::cout << GridLogMessage << "Number of iterations to average: " << Nloop
-            << std::endl;
-  std::vector<double> t_time(Nloop);
-  //  time_statistics timestat;
-
-  std::cout << GridLogMessage
-            << "========================================================================="
-               "==========================="
-            << std::endl;
-  std::cout << GridLogMessage
-            << "= Benchmarking sequential halo exchange from host memory " << std::endl;
-  std::cout << GridLogMessage
-            << "========================================================================="
-               "==========================="
-            << std::endl;
-  header();
-
-  for (int lat = 8; lat <= maxlat; lat += 4)
-  {
-    for (int Ls = 8; Ls <= 8; Ls *= 2)
-    {
-
-      Coordinate latt_size({lat * mpi_layout[0], lat * mpi_layout[1], lat * mpi_layout[2],
-                            lat * mpi_layout[3]});
-
-      GridCartesian Grid(latt_size, simd_layout, mpi_layout);
-      RealD Nrank = Grid._Nprocessors;
-      RealD Nnode = Grid.NodeCount();
-      RealD ppn = Nrank / Nnode;
-
-      std::vector<std::vector<HalfSpinColourVectorD>> xbuf(8);
-      std::vector<std::vector<HalfSpinColourVectorD>> rbuf(8);
-
-      for (int mu = 0; mu < 8; mu++)
-      {
-        xbuf[mu].resize(lat * lat * lat * Ls);
-        rbuf[mu].resize(lat * lat * lat * Ls);
-      }
-      uint64_t bytes = lat * lat * lat * Ls * sizeof(HalfSpinColourVectorD);
-
-      int ncomm;
-
-      for (int mu = 0; mu < 4; mu++)
-      {
-        if (mpi_layout[mu] > 1)
-        {
-          double start = usecond();
-          for (int i = 0; i < Nloop; i++)
-          {
-
-            ncomm = 0;
-
-            ncomm++;
-            int comm_proc = 1;
-            int xmit_to_rank;
-            int recv_from_rank;
-
-            {
-              std::vector<CommsRequest_t> requests;
-              Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
-              Grid.SendToRecvFrom((void *)&xbuf[mu][0], xmit_to_rank,
-                                  (void *)&rbuf[mu][0], recv_from_rank, bytes);
-            }
-
-            comm_proc = mpi_layout[mu] - 1;
-            {
-              std::vector<CommsRequest_t> requests;
-              Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
-              Grid.SendToRecvFrom((void *)&xbuf[mu + 4][0], xmit_to_rank,
-                                  (void *)&rbuf[mu + 4][0], recv_from_rank, bytes);
-            }
-          }
-          Grid.Barrier();
-          double stop = usecond();
-          double mean = (stop - start) / Nloop;
-          double dbytes = bytes * ppn;
-          double xbytes = dbytes * 2.0 * ncomm;
-          double rbytes = xbytes;
-          double bidibytes = xbytes + rbytes;
-
-          std::cout << GridLogMessage << std::setw(4) << lat << "\t" << Ls << "\t"
-                    << std::setw(11) << bytes << std::fixed << std::setprecision(1)
-                    << std::setw(7) << " " << std::right << xbytes / mean << "  "
-                    << "\t\t" << std::setw(7) << bidibytes / mean << std::endl;
-        }
-      }
-    }
-  }
-
-  std::cout << GridLogMessage
-            << "========================================================================="
-               "==========================="
-            << std::endl;
-  std::cout << GridLogMessage
-            << "= Benchmarking sequential halo exchange from GPU memory " << std::endl;
-  std::cout << GridLogMessage
-            << "========================================================================="
-               "==========================="
-            << std::endl;
-  header();
-
-  for (int lat = 8; lat <= maxlat; lat += 4)
-  {
-    for (int Ls = 8; Ls <= 8; Ls *= 2)
-    {
-
-      Coordinate latt_size({lat * mpi_layout[0], lat * mpi_layout[1], lat * mpi_layout[2],
-                            lat * mpi_layout[3]});
-
-      GridCartesian Grid(latt_size, simd_layout, mpi_layout);
-      RealD Nrank = Grid._Nprocessors;
-      RealD Nnode = Grid.NodeCount();
-      RealD ppn = Nrank / Nnode;
-
-      std::vector<HalfSpinColourVectorD *> xbuf(8);
-      std::vector<HalfSpinColourVectorD *> rbuf(8);
-
-      uint64_t bytes = lat * lat * lat * Ls * sizeof(HalfSpinColourVectorD);
-      for (int d = 0; d < 8; d++)
-      {
-        xbuf[d] = (HalfSpinColourVectorD *)acceleratorAllocDevice(bytes);
-        rbuf[d] = (HalfSpinColourVectorD *)acceleratorAllocDevice(bytes);
-      }
-
-      int ncomm;
-
-      for (int mu = 0; mu < 4; mu++)
-      {
-        if (mpi_layout[mu] > 1)
-        {
-          double start = usecond();
-          for (int i = 0; i < Nloop; i++)
-          {
-
-            ncomm = 0;
-
-            ncomm++;
-            int comm_proc = 1;
-            int xmit_to_rank;
-            int recv_from_rank;
-
-            {
-              std::vector<CommsRequest_t> requests;
-              Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
-              Grid.SendToRecvFrom((void *)&xbuf[mu][0], xmit_to_rank,
-                                  (void *)&rbuf[mu][0], recv_from_rank, bytes);
-            }
-
-            comm_proc = mpi_layout[mu] - 1;
-            {
-              std::vector<CommsRequest_t> requests;
-              Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
-              Grid.SendToRecvFrom((void *)&xbuf[mu + 4][0], xmit_to_rank,
-                                  (void *)&rbuf[mu + 4][0], recv_from_rank, bytes);
-            }
-          }
-          Grid.Barrier();
-          double stop = usecond();
-          double mean = (stop - start) / Nloop;
-          double dbytes = bytes * ppn;
-          double xbytes = dbytes * 2.0 * ncomm;
-          double rbytes = xbytes;
-          double bidibytes = xbytes + rbytes;
-
-          std::cout << GridLogMessage << std::setw(4) << lat << "\t" << Ls << "\t"
-                    << std::setw(11) << bytes << std::fixed << std::setprecision(1)
-                    << std::setw(7) << " " << std::right << xbytes / mean << "  "
-                    << "\t\t" << std::setw(7) << bidibytes / mean << std::endl;
-        }
-      }
-
-      for (int d = 0; d < 8; d++)
-      {
-        acceleratorFreeDevice(xbuf[d]);
-        acceleratorFreeDevice(rbuf[d]);
-      }
-    }
-  }
-
-  std::cout << GridLogMessage
-            << "========================================================================="
-               "==========================="
-            << std::endl;
-  std::cout << GridLogMessage << "= All done; Bye Bye" << std::endl;
-  std::cout << GridLogMessage
-            << "========================================================================="
-               "==========================="
-            << std::endl;
-
-  Grid_finalize();
-}
--- a/Grid/Benchmark_dwf_fp32.cpp
+++ b/Grid/Benchmark_dwf_fp32.cpp
@ -1,512 +0,0 @@
-/*
-Copyright © 2015 Peter Boyle <paboyle@ph.ed.ac.uk>
-Copyright © 2022 Antonin Portelli <antonin.portelli@me.com>
-Copyright © 2023 Simon Bürger <simon.buerger@rwth-aachen.de>
-
-This program is free software; you can redistribute it and/or
-modify it under the terms of the GNU General Public License
-as published by the Free Software Foundation; either version 2
-of the License, or (at your option) any later version.
-
-This program is distributed in the hope that it will be useful,
-but WITHOUT ANY WARRANTY; without even the implied warranty of
-MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-GNU General Public License for more details.
-
-You should have received a copy of the GNU General Public License
-along with this program. If not, see <http://www.gnu.org/licenses/>.
-*/
-
-#include "json.hpp"
-#include <Grid/Grid.h>
-#ifdef GRID_CUDA
-#define CUDA_PROFILE
-#endif
-
-#ifdef CUDA_PROFILE
-#include <cuda_profiler_api.h>
-#endif
-
-using namespace std;
-using namespace Grid;
-
-template <class d> struct scal
-{
-  d internal;
-};
-
-Gamma::Algebra Gmu[] = {Gamma::Algebra::GammaX, Gamma::Algebra::GammaY,
-                        Gamma::Algebra::GammaZ, Gamma::Algebra::GammaT};
-
-int main(int argc, char **argv)
-{
-  Grid_init(&argc, &argv);
-
-  int threads = GridThread::GetThreads();
-
-  Coordinate latt4 = GridDefaultLatt();
-  int Ls = 16;
-  std::string json_filename = ""; // empty indicates no json output
-  nlohmann::json json;
-
-  // benchmark specific command line arguments
-  for (int i = 0; i < argc; i++)
-  {
-    if (std::string(argv[i]) == "-Ls")
-    {
-      std::stringstream ss(argv[i + 1]);
-      ss >> Ls;
-    }
-    if (std::string(argv[i]) == "--json-out")
-      json_filename = argv[i + 1];
-  }
-
-  GridLogLayout();
-
-  long unsigned int single_site_flops = 8 * Nc * (7 + 16 * Nc);
-
-  json["single_site_flops"] = single_site_flops;
-
-  GridCartesian *UGrid = SpaceTimeGrid::makeFourDimGrid(
-      GridDefaultLatt(), GridDefaultSimd(Nd, vComplexF::Nsimd()), GridDefaultMpi());
-  GridRedBlackCartesian *UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
-
-  GridCartesian *FGrid = SpaceTimeGrid::makeFiveDimGrid(Ls, UGrid);
-  GridRedBlackCartesian *FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls, UGrid);
-
-  json["grid"] = FGrid->FullDimensions().toVector();
-  json["local_grid"] = FGrid->LocalDimensions().toVector();
-
-  std::cout << GridLogMessage << "Making s innermost grids" << std::endl;
-  GridCartesian *sUGrid =
-      SpaceTimeGrid::makeFourDimDWFGrid(GridDefaultLatt(), GridDefaultMpi());
-
-  GridRedBlackCartesian *sUrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(sUGrid);
-  GridCartesian *sFGrid = SpaceTimeGrid::makeFiveDimDWFGrid(Ls, UGrid);
-  GridRedBlackCartesian *sFrbGrid = SpaceTimeGrid::makeFiveDimDWFRedBlackGrid(Ls, UGrid);
-
-  std::vector<int> seeds4({1, 2, 3, 4});
-  std::vector<int> seeds5({5, 6, 7, 8});
-
-  std::cout << GridLogMessage << "Initialising 4d RNG" << std::endl;
-  GridParallelRNG RNG4(UGrid);
-  RNG4.SeedUniqueString(std::string("The 4D RNG"));
-  std::cout << GridLogMessage << "Initialising 5d RNG" << std::endl;
-  GridParallelRNG RNG5(FGrid);
-  RNG5.SeedUniqueString(std::string("The 5D RNG"));
-  std::cout << GridLogMessage << "Initialised RNGs" << std::endl;
-
-  LatticeFermionF src(FGrid);
-  random(RNG5, src);
-#if 0
-  src = Zero();
-  {
-    Coordinate origin({0,0,0,latt4[2]-1,0});
-    SpinColourVectorF tmp;
-    tmp=Zero();
-    tmp()(0)(0)=Complex(-2.0,0.0);
-    std::cout << " source site 0 " << tmp<<std::endl;
-    pokeSite(tmp,src,origin);
-  }
-#else
-  RealD N2 = 1.0 / ::sqrt(norm2(src));
-  src = src * N2;
-#endif
-
-  LatticeFermionF result(FGrid);
-  result = Zero();
-  LatticeFermionF ref(FGrid);
-  ref = Zero();
-  LatticeFermionF tmp(FGrid);
-  LatticeFermionF err(FGrid);
-
-  std::cout << GridLogMessage << "Drawing gauge field" << std::endl;
-  LatticeGaugeFieldF Umu(UGrid);
-  SU<Nc>::HotConfiguration(RNG4, Umu);
-  std::cout << GridLogMessage << "Random gauge initialised " << std::endl;
-#if 0
-  Umu=1.0;
-  for(int mu=0;mu<Nd;mu++){
-    LatticeColourMatrixF ttmp(UGrid);
-    ttmp = PeekIndex<LorentzIndex>(Umu,mu);
-    //    if (mu !=2 ) ttmp = 0;
-    //    ttmp = ttmp* pow(10.0,mu);
-    PokeIndex<LorentzIndex>(Umu,ttmp,mu);
-  }
-  std::cout << GridLogMessage << "Forced to diagonal " << std::endl;
-#endif
-
-  ////////////////////////////////////
-  // Naive wilson implementation
-  ////////////////////////////////////
-  // replicate across fifth dimension
-  //  LatticeGaugeFieldF Umu5d(FGrid);
-  std::vector<LatticeColourMatrixF> U(4, UGrid);
-  for (int mu = 0; mu < Nd; mu++)
-  {
-    U[mu] = PeekIndex<LorentzIndex>(Umu, mu);
-  }
-  std::cout << GridLogMessage << "Setting up Cshift based reference " << std::endl;
-
-  if (1)
-  {
-    ref = Zero();
-    for (int mu = 0; mu < Nd; mu++)
-    {
-
-      tmp = Cshift(src, mu + 1, 1);
-      {
-        autoView(tmp_v, tmp, CpuWrite);
-        autoView(U_v, U[mu], CpuRead);
-        for (int ss = 0; ss < U[mu].Grid()->oSites(); ss++)
-        {
-          for (int s = 0; s < Ls; s++)
-          {
-            tmp_v[Ls * ss + s] = U_v[ss] * tmp_v[Ls * ss + s];
-          }
-        }
-      }
-      ref = ref + tmp - Gamma(Gmu[mu]) * tmp;
-
-      {
-        autoView(tmp_v, tmp, CpuWrite);
-        autoView(U_v, U[mu], CpuRead);
-        autoView(src_v, src, CpuRead);
-        for (int ss = 0; ss < U[mu].Grid()->oSites(); ss++)
-        {
-          for (int s = 0; s < Ls; s++)
-          {
-            tmp_v[Ls * ss + s] = adj(U_v[ss]) * src_v[Ls * ss + s];
-          }
-        }
-      }
-      tmp = Cshift(tmp, mu + 1, -1);
-      ref = ref + tmp + Gamma(Gmu[mu]) * tmp;
-    }
-    ref = -0.5 * ref;
-  }
-
-  RealD mass = 0.1;
-  RealD M5 = 1.8;
-
-  RealD NP = UGrid->_Nprocessors;
-  RealD NN = UGrid->NodeCount();
-
-  json["ranks"] = NP;
-  json["nodes"] = NN;
-
-  std::cout << GridLogMessage
-            << "*****************************************************************"
-            << std::endl;
-  std::cout << GridLogMessage
-            << "* Kernel options --dslash-generic, --dslash-unroll, --dslash-asm"
-            << std::endl;
-  std::cout << GridLogMessage
-            << "*****************************************************************"
-            << std::endl;
-  std::cout << GridLogMessage
-            << "*****************************************************************"
-            << std::endl;
-  std::cout << GridLogMessage
-            << "* Benchmarking DomainWallFermionR::Dhop                  " << std::endl;
-  std::cout << GridLogMessage << "* Vectorising space-time by " << vComplexF::Nsimd()
-            << std::endl;
-  std::cout << GridLogMessage << "* VComplexF size is " << sizeof(vComplexF) << " B"
-            << std::endl;
-
-  if (sizeof(RealF) == 4)
-    std::cout << GridLogMessage << "* SINGLE precision " << std::endl;
-  if (sizeof(RealF) == 8)
-    std::cout << GridLogMessage << "* DOUBLE precision " << std::endl;
-#ifdef GRID_OMP
-  if (WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute)
-    std::cout << GridLogMessage << "* Using Overlapped Comms/Compute" << std::endl;
-  if (WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsThenCompute)
-    std::cout << GridLogMessage << "* Using sequential comms compute" << std::endl;
-#endif
-  if (WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGeneric)
-    std::cout << GridLogMessage << "* Using GENERIC Nc WilsonKernels" << std::endl;
-  if (WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptHandUnroll)
-    std::cout << GridLogMessage << "* Using Nc=3       WilsonKernels" << std::endl;
-  if (WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm)
-
-    std::cout << GridLogMessage << "* Using Asm Nc=3   WilsonKernels" << std::endl;
-  std::cout << GridLogMessage
-            << "*****************************************************************"
-            << std::endl;
-
-  DomainWallFermionF Dw(Umu, *FGrid, *FrbGrid, *UGrid, *UrbGrid, mass, M5);
-  int ncall = 300;
-
-  if (1)
-  {
-    FGrid->Barrier();
-    Dw.ZeroCounters();
-    Dw.Dhop(src, result, 0);
-    std::cout << GridLogMessage << "Called warmup" << std::endl;
-    double t0 = usecond();
-    for (int i = 0; i < ncall; i++)
-    {
-      __SSC_START;
-      Dw.Dhop(src, result, 0);
-      __SSC_STOP;
-    }
-    double t1 = usecond();
-    FGrid->Barrier();
-
-    double volume = Ls;
-    for (int mu = 0; mu < Nd; mu++)
-      volume = volume * latt4[mu];
-    double flops = single_site_flops * volume * ncall;
-
-    auto nsimd = vComplex::Nsimd();
-    auto simdwidth = sizeof(vComplex);
-
-    // RF: Nd Wilson * Ls, Nd gauge * Ls, Nc colors
-    double data_rf = volume * ((2 * Nd + 1) * Nd * Nc + 2 * Nd * Nc * Nc) * simdwidth /
-                     nsimd * ncall / (1024. * 1024. * 1024.);
-
-    // mem: Nd Wilson * Ls, Nd gauge, Nc colors
-    double data_mem =
-        (volume * (2 * Nd + 1) * Nd * Nc + (volume / Ls) * 2 * Nd * Nc * Nc) * simdwidth /
-        nsimd * ncall / (1024. * 1024. * 1024.);
-
-    json["Dw"]["calls"] = ncall;
-    json["Dw"]["time"] = t1 - t0;
-    json["Dw"]["mflops"] = flops / (t1 - t0);
-    json["Dw"]["mflops_per_rank"] = flops / (t1 - t0) / NP;
-    json["Dw"]["mflops_per_node"] = flops / (t1 - t0) / NN;
-    json["Dw"]["RF"] = 1000000. * data_rf / ((t1 - t0));
-    json["Dw"]["mem"] = 1000000. * data_mem / ((t1 - t0));
-
-    std::cout << GridLogMessage << "Called Dw " << ncall << " times in " << t1 - t0
-              << " us" << std::endl;
-    //    std::cout<<GridLogMessage << "norm result "<< norm2(result)<<std::endl;
-    //    std::cout<<GridLogMessage << "norm ref    "<< norm2(ref)<<std::endl;
-    std::cout << GridLogMessage << "mflop/s =   " << flops / (t1 - t0) << std::endl;
-    std::cout << GridLogMessage << "mflop/s per rank =  " << flops / (t1 - t0) / NP
-              << std::endl;
-    std::cout << GridLogMessage << "mflop/s per node =  " << flops / (t1 - t0) / NN
-              << std::endl;
-
-    std::cout << GridLogMessage
-              << "RF  GiB/s (base 2) =   " << 1000000. * data_rf / ((t1 - t0))
-              << std::endl;
-    std::cout << GridLogMessage
-              << "mem GiB/s (base 2) =   " << 1000000. * data_mem / ((t1 - t0))
-              << std::endl;
-    err = ref - result;
-    std::cout << GridLogMessage << "norm diff   " << norm2(err) << std::endl;
-    // exit(0);
-
-    if ((norm2(err) > 1.0e-4))
-    {
-      /*
-      std::cout << "RESULT\n " << result<<std::endl;
-      std::cout << "REF   \n " << ref   <<std::endl;
-      std::cout << "ERR   \n " << err   <<std::endl;
-      */
-      std::cout << GridLogMessage << "WRONG RESULT" << std::endl;
-      FGrid->Barrier();
-      exit(-1);
-    }
-    assert(norm2(err) < 1.0e-4);
-    Dw.Report();
-  }
-
-  if (1)
-  { // Naive wilson dag implementation
-    ref = Zero();
-    for (int mu = 0; mu < Nd; mu++)
-    {
-
-      //    ref =  src - Gamma(Gamma::Algebra::GammaX)* src ; // 1+gamma_x
-      tmp = Cshift(src, mu + 1, 1);
-      {
-        autoView(ref_v, ref, CpuWrite);
-        autoView(tmp_v, tmp, CpuRead);
-        autoView(U_v, U[mu], CpuRead);
-        for (int ss = 0; ss < U[mu].Grid()->oSites(); ss++)
-        {
-          for (int s = 0; s < Ls; s++)
-          {
-            int i = s + Ls * ss;
-            ref_v[i] += U_v[ss] * (tmp_v[i] + Gamma(Gmu[mu]) * tmp_v[i]);
-            ;
-          }
-        }
-      }
-
-      {
-        autoView(tmp_v, tmp, CpuWrite);
-        autoView(U_v, U[mu], CpuRead);
-        autoView(src_v, src, CpuRead);
-        for (int ss = 0; ss < U[mu].Grid()->oSites(); ss++)
-        {
-          for (int s = 0; s < Ls; s++)
-          {
-            tmp_v[Ls * ss + s] = adj(U_v[ss]) * src_v[Ls * ss + s];
-          }
-        }
-      }
-      //      tmp =adj(U[mu])*src;
-      tmp = Cshift(tmp, mu + 1, -1);
-      {
-        autoView(ref_v, ref, CpuWrite);
-        autoView(tmp_v, tmp, CpuRead);
-        for (int i = 0; i < ref_v.size(); i++)
-        {
-          ref_v[i] += tmp_v[i] - Gamma(Gmu[mu]) * tmp_v[i];
-          ;
-        }
-      }
-    }
-    ref = -0.5 * ref;
-  }
-  //  dump=1;
-  Dw.Dhop(src, result, 1);
-
-  std::cout << GridLogMessage
-            << "Compare to naive wilson implementation Dag to verify correctness"
-            << std::endl;
-  std::cout << GridLogMessage << "Called DwDag" << std::endl;
-  std::cout << GridLogMessage << "norm dag result " << norm2(result) << std::endl;
-  std::cout << GridLogMessage << "norm dag ref    " << norm2(ref) << std::endl;
-  err = ref - result;
-  std::cout << GridLogMessage << "norm dag diff   " << norm2(err) << std::endl;
-  if ((norm2(err) > 1.0e-4))
-  {
-    /*
-      std::cout<< "DAG RESULT\n "  <<ref     << std::endl;
-      std::cout<< "DAG sRESULT\n " <<result  << std::endl;
-      std::cout<< "DAG ERR   \n "  << err    <<std::endl;
-    */
-  }
-  LatticeFermionF src_e(FrbGrid);
-  LatticeFermionF src_o(FrbGrid);
-  LatticeFermionF r_e(FrbGrid);
-  LatticeFermionF r_o(FrbGrid);
-  LatticeFermionF r_eo(FGrid);
-
-  std::cout << GridLogMessage << "Calling Deo and Doe and //assert Deo+Doe == Dunprec"
-            << std::endl;
-  pickCheckerboard(Even, src_e, src);
-  pickCheckerboard(Odd, src_o, src);
-
-  std::cout << GridLogMessage << "src_e" << norm2(src_e) << std::endl;
-  std::cout << GridLogMessage << "src_o" << norm2(src_o) << std::endl;
-
-  // S-direction is INNERMOST and takes no part in the parity.
-  std::cout << GridLogMessage
-
-            << "*********************************************************" << std::endl;
-  std::cout << GridLogMessage
-            << "* Benchmarking DomainWallFermionF::DhopEO                " << std::endl;
-  std::cout << GridLogMessage << "* Vectorising space-time by " << vComplexF::Nsimd()
-            << std::endl;
-
-  if (sizeof(RealF) == 4)
-    std::cout << GridLogMessage << "* SINGLE precision " << std::endl;
-  if (sizeof(RealF) == 8)
-    std::cout << GridLogMessage << "* DOUBLE precision " << std::endl;
-#ifdef GRID_OMP
-  if (WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute)
-    std::cout << GridLogMessage << "* Using Overlapped Comms/Compute" << std::endl;
-  if (WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsThenCompute)
-    std::cout << GridLogMessage << "* Using sequential comms compute" << std::endl;
-#endif
-  if (WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGeneric)
-    std::cout << GridLogMessage << "* Using GENERIC Nc WilsonKernels" << std::endl;
-  if (WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptHandUnroll)
-    std::cout << GridLogMessage << "* Using Nc=3       WilsonKernels" << std::endl;
-  if (WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm)
-    std::cout << GridLogMessage << "* Using Asm Nc=3   WilsonKernels" << std::endl;
-  std::cout << GridLogMessage
-            << "*********************************************************" << std::endl;
-
-  {
-    Dw.ZeroCounters();
-    FGrid->Barrier();
-    Dw.DhopEO(src_o, r_e, DaggerNo);
-    double t0 = usecond();
-    for (int i = 0; i < ncall; i++)
-    {
-#ifdef CUDA_PROFILE
-      if (i == 10)
-        cudaProfilerStart();
-#endif
-      Dw.DhopEO(src_o, r_e, DaggerNo);
-#ifdef CUDA_PROFILE
-      if (i == 20)
-        cudaProfilerStop();
-#endif
-    }
-    double t1 = usecond();
-    FGrid->Barrier();
-
-    double volume = Ls;
-    for (int mu = 0; mu < Nd; mu++)
-      volume = volume * latt4[mu];
-    double flops = (single_site_flops * volume * ncall) / 2.0;
-
-    json["Deo"]["calls"] = ncall;
-    json["Deo"]["time"] = t1 - t0;
-    json["Deo"]["mflops"] = flops / (t1 - t0);
-    json["Deo"]["mflops_per_rank"] = flops / (t1 - t0) / NP;
-    json["Deo"]["mflops_per_node"] = flops / (t1 - t0) / NN;
-
-    std::cout << GridLogMessage << "Deo mflop/s =   " << flops / (t1 - t0) << std::endl;
-    std::cout << GridLogMessage << "Deo mflop/s per rank   " << flops / (t1 - t0) / NP
-              << std::endl;
-    std::cout << GridLogMessage << "Deo mflop/s per node   " << flops / (t1 - t0) / NN
-              << std::endl;
-
-    Dw.Report();
-  }
-  Dw.DhopEO(src_o, r_e, DaggerNo);
-  Dw.DhopOE(src_e, r_o, DaggerNo);
-  Dw.Dhop(src, result, DaggerNo);
-
-  std::cout << GridLogMessage << "r_e" << norm2(r_e) << std::endl;
-  std::cout << GridLogMessage << "r_o" << norm2(r_o) << std::endl;
-  std::cout << GridLogMessage << "res" << norm2(result) << std::endl;
-
-  setCheckerboard(r_eo, r_o);
-  setCheckerboard(r_eo, r_e);
-
-  err = r_eo - result;
-  std::cout << GridLogMessage << "norm diff   " << norm2(err) << std::endl;
-  if ((norm2(err) > 1.0e-4))
-  {
-    /*
-  std::cout<< "Deo RESULT\n " <<r_eo << std::endl;
-  std::cout<< "Deo REF\n " <<result  << std::endl;
-  std::cout<< "Deo ERR   \n " << err <<std::endl;
-    */
-  }
-
-  pickCheckerboard(Even, src_e, err);
-  pickCheckerboard(Odd, src_o, err);
-  std::cout << GridLogMessage << "norm diff even  " << norm2(src_e) << std::endl;
-  std::cout << GridLogMessage << "norm diff odd   " << norm2(src_o) << std::endl;
-
-  assert(norm2(src_e) < 1.0e-4);
-  assert(norm2(src_o) < 1.0e-4);
-
-  if (!json_filename.empty())
-  {
-    std::cout << GridLogMessage << "writing benchmark results to " << json_filename
-              << std::endl;
-
-    int me = 0;
-    MPI_Comm_rank(MPI_COMM_WORLD, &me);
-    if (me == 0)
-    {
-      std::ofstream json_file(json_filename);
-      json_file << std::setw(4) << json;
-    }
-  }
-
-  Grid_finalize();
-  exit(0);
-}
--- a/Grid/Common.hpp
+++ b/Grid/Common.hpp
@ -26,6 +26,20 @@ along with this program. If not, see <http://www.gnu.org/licenses/>.
 #define GRID_MSG_MAXSIZE 1024
 #endif

+#define GRID_BIG_SEP                                                                     \
+  "==============================================================================="
+#define GRID_SMALL_SEP "------------------------------------------"
+
+#define grid_big_sep()                                                                   \
+  {                                                                                      \
+    GRID_MSG << GRID_BIG_SEP << std::endl;                                               \
+  }
+
+#define grid_small_sep()                                                                 \
+  {                                                                                      \
+    GRID_MSG << GRID_SMALL_SEP << std::endl;                                             \
+  }
+
 #define grid_printf(...)                                                                 \
  {                                                                                      \
    char _buf[GRID_MSG_MAXSIZE];                                                         \
--- a/Grid/Makefile.am
+++ b/Grid/Makefile.am
@ -1,12 +1,8 @@
 ACLOCAL_AMFLAGS = -I .buildutils/m4
 
 bin_PROGRAMS =                \
-  Benchmark_comms_host_device \
-  Benchmark_dwf_fp32          \
  Benchmark_Grid              \
  Benchmark_IO
-  
-Benchmark_comms_host_device_SOURCES = Benchmark_comms_host_device.cpp
-Benchmark_dwf_fp32_SOURCES          = Benchmark_dwf_fp32.cpp
+
 Benchmark_Grid_SOURCES              = Benchmark_Grid.cpp
 Benchmark_IO_SOURCES                = Benchmark_IO.cpp
--- a/Grid/Readme.md
+++ b/Grid/Readme.md
@ -6,6 +6,7 @@ The benchmarks can be summarised as follows
 - `Benchmark_Grid`: This benchmark measure floating point performances for various fermion
 matrices, as well as bandwidth measurement for different operations. Measurements are
 performed for a fixed range of problem sizes.
+- `Benchmark_IO`: Parallel I/O benchmark.

 ## TL;DR
 Build and install Grid, all dependencies, and the benchmark with
@ -28,7 +29,7 @@ You should first deploy the environment for the specific system you are using, f
 systems/tursa/bootstrap-env.sh ./env
 ```
 will deploy the relevant environment for the [Tursa](https://www.epcc.ed.ac.uk/hpc-services/dirac-tursa-gpu) supercomputer in `./env`. This step might compile from source a large set
-of packages, and might take some time to complete.
+of packages, and take some time to complete.

 After that, the environment directory (`./env` in the example above) will contain a `env.sh` file that need to be sourced to activate the environment
 ```bash
@ -66,4 +67,84 @@ where `<env_dir>` is the environment directory and `<config>` is the build confi

 ## Running the benchmarks
 After building the benchmarks as above you can find the binaries in 
-`<env_dir>/prefix/gridbench_<config>`.
+`<env_dir>/prefix/gridbench_<config>`. Depending on the system selected, the environment
+directory might also contain batch script examples. More information about the benchmarks
+is provided below.
+
+### `Benchmark_Grid`
+This benchmark performs flop/s measurement for typical lattice QCD sparse matrices, as
+well as memory and inter-process bandwidth measurement using Grid routines. The benchmark
+command accept any Grid flag (see complete list with `--help`), as well as a 
+`--json-out <file>` flag to save the measurement results in JSON to `<file>`. The 
+benchmarks are performed on a fix set of problem sizes, and the Grid flag `--grid` will
+be ignored.
+
+The resulting metrics are as follows, all data size units are in base 2 
+(i.e. 1 kB = 1024 B).
+
+*Memory bandwidth*
+
+One sub-benchmark measure the memory bandwidth using a lattice version of the `axpy` BLAS
+routine, in a similar fashion to the STREAM benchmark. The JSON entries under `"axpy"` 
+have the form
+```json
+{
+  "GBps": 215.80653375861607,   // bandwidth in GB/s/node
+  "GFlops": 19.310041765757834, // FP performance (double precision)
+  "L": 8,                       // local lattice volume
+  "size_MB": 3.0                // memory size in MB/node
+}
+```
+
+A second benchmark performs site-wise SU(4) matrix multiplication, and has a higher
+arithmetic intensity than the `axpy` one (although it is still memory-bound). 
+The JSON entries under `"SU4"` have the form
+```json
+{
+  "GBps": 394.76639187026865,  // bandwidth in GB/s/node
+  "GFlops": 529.8464820758512, // FP performance (single precision)
+  "L": 8,                      // local lattice size
+  "size_MB": 6.0               // memory size in MB/node
+}
+```
+
+*Inter-process bandwidth*
+
+This sub-benchmark measures the achieved bidirectional bandwidth in threaded halo exchange
+using routines in Grid. The exchange is performed in each direction on the MPI Cartesian
+grid which is parallelised across at least 2 processes. The resulting bandwidth is related
+to node-local transfers (inter-CPU, NVLink, ...) or network transfers depending on the MPI
+decomposition. he JSON entries under `"comms"` have the form
+```json
+{
+  "L": 40,                       // local lattice size
+  "bytes": 73728000,             // payload size in B/rank
+  "dir": 2,                      // direction of the exchange, 8 possible directions
+                                 // (0: +x, 1: +y, ..., 5: -x, 6: -y, ...)
+  "rate_GBps": {
+    "error": 6.474271894240327,  // standard deviation across measurements (GB/s/node)
+    "max": 183.10546875,         // maximum measured bandwidth (GB/s/node)
+    "mean": 175.21747026766676   // average measured bandwidth (GB/s/node)
+  },
+  "time_usec": 3135.055          // average transfer time (microseconds)
+}
+```
+
+*Floating-point performances*
+
+This sub-benchmark measures the achieved floating-point performances using the 
+Wilson fermion, domain-wall fermion, and staggered fermion sparse matrices from Grid.
+In the `"flops"` and `"results"` section of the JSON output are recorded the best 
+performances, e.g.
+```json
+{
+  "Gflops_dwf4": 366.5251173474483,       // domain-wall in Gflop/s/node (single precision)
+  "Gflops_staggered": 7.5982861018529455, // staggered in Gflop/s/node (single precision)
+  "Gflops_wilson": 15.221839719288932,    // Wilson in Gflop/s/node (single precision)
+  "L": 8                                  // local lattice size
+}
+```
+Here "best" means across a number of different implementations of the routines. Please
+see the log of the benchmark for an additional breakdown. Finally, the JSON output
+contains a "comparison point", which is the average of the L=24 and L=32 best
+domain-wall performances.
--- a/Grid/build-benchmark.sh
+++ b/Grid/build-benchmark.sh
@ -20,8 +20,12 @@ mkdir -p "${build_dir}"
 source "${env_dir}/env.sh"
 entry=$(jq ".configs[]|select(.name==\"${cfg}\")" "${env_dir}"/grid-config.json)
 env_script=$(echo "${entry}" | jq -r ".\"env-script\"")
-cd "${build_dir}" || return
 source "${env_dir}/${env_script}"
+cd "${script_dir}"
+if [ ! -f configure ]; then
+    ./bootstrap.sh
+fi
+cd "${build_dir}"
 if [ ! -f Makefile ]; then
    "${script_dir}/configure" --with-grid="${env_dir}/prefix/grid_${cfg}" \
                            --prefix="${env_dir}/prefix/gridbench_${cfg}"
--- a/Grid/systems/tursa/env.sh
+++ b/Grid/systems/tursa/env.sh
@ -1,8 +1,8 @@
 #!/usr/bin/env bash
 # shellcheck disable=SC1091

-GRIDENVDIR="$(dirname "$(readlink -f "${BASH_SOURCE:-$0}")")"
-export GRIDENVDIR
-export PATH="${GRIDENVDIR}/prefix/base/bin:${PATH}"
-export ACLOCAL_PATH="${GRIDENVDIR}/prefix/base/share/aclocal:${ACLOCAL_PATH}"
-source "${GRIDENVDIR}"/spack/share/spack/setup-env.sh
+env_dir="$(dirname "$(readlink -f "${BASH_SOURCE:-$0}")")"
+mkdir -p ~/.config/lattice-benchmarks
+echo "${env_dir}" > ~/.config/lattice-benchmarks/grid-env
+source "${env_dir}/spack/share/spack/setup-env.sh"
+spack load jq git
--- a/Grid/systems/tursa/files/cpu-mpi-wrapper.sh
+++ b/Grid/systems/tursa/files/cpu-mpi-wrapper.sh
--- a/Grid/systems/tursa/files/gpu-mpi-wrapper.sh
+++ b/Grid/systems/tursa/files/gpu-mpi-wrapper.sh
@ -1,13 +1,12 @@
 #!/usr/bin/env bash

 lrank=$OMPI_COMM_WORLD_LOCAL_RANK
-numa1=$(( 2 * lrank))
-numa2=$(( 2 * lrank + 1 ))
+numa1=$((lrank))
 netdev=mlx5_${lrank}:1

 export CUDA_VISIBLE_DEVICES=$OMPI_COMM_WORLD_LOCAL_RANK
 export UCX_NET_DEVICES=${netdev}
-BINDING="--interleave=$numa1,$numa2"
+BINDING="--interleave=$numa1"

 echo "$(hostname) - $lrank device=$CUDA_VISIBLE_DEVICES binding=$BINDING"

--- a/Grid/systems/tursa/files/ompi-gpu.sh
+++ b/Grid/systems/tursa/files/ompi-gpu.sh
@ -0,0 +1,17 @@
+#!/usr/bin/env bash
+
+# OpenMP/OpenMPI/UCX environment ###############################################
+export OMP_NUM_THREADS=8
+export OMPI_MCA_btl=^uct,openib
+export OMPI_MCA_pml=ucx
+export UCX_TLS=gdr_copy,rc,rc_x,sm,cuda_copy,cuda_ipc
+export UCX_RNDV_SCHEME=put_zcopy
+export UCX_RNDV_THRESH=16384
+export UCX_IB_GPU_DIRECT_RDMA=yes
+export UCX_MEMTYPE_CACHE=n
+
+# IO environment ###############################################################
+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
--- a/Grid/systems/tursa/files/run.gpu.16nodes.sh
+++ b/Grid/systems/tursa/files/run.gpu.16nodes.sh
@ -0,0 +1,60 @@
+#!/usr/bin/env bash
+# shellcheck disable=SC1091,SC2050,SC2170
+
+#SBATCH -J benchmark-grid-16
+#SBATCH -t 1:00:00
+#SBATCH --nodes=16
+#SBATCH --ntasks=64
+#SBATCH --ntasks-per-node=4
+#SBATCH --cpus-per-task=8
+#SBATCH --partition=gpu
+#SBATCH --gres=gpu:4
+#SBATCH --output=%x.%j.out
+#SBATCH --error=%x.%j.err
+#SBATCH --qos=standard
+#SBATCH --no-requeue
+#SBATCH --gpu-freq=1410
+
+set -euo pipefail
+
+# load environment #############################################################
+env_cfg="${HOME}/.config/lattice-benchmarks/grid-env"
+if [ ! -f "${env_cfg}" ]; then
+	echo "error: ${env_cfg} does not exists, did you execute 'source env.sh' with your user account?"
+	exit 1
+fi
+env_dir="$(readlink -f "$(cat "${env_cfg}")")"
+source "${env_dir}/env.sh"      # load base Spack environment
+source "${env_dir}/env-gpu.sh"  # load GPU-sepcific packages
+source "${env_dir}/ompi-gpu.sh" # set GPU-specific OpenMPI variables
+
+# application and parameters ###################################################
+app="${env_dir}/prefix/gridbench_gpu/bin/Benchmark_Grid"
+
+# collect job information ######################################################
+job_info_dir=job/${SLURM_JOB_NAME}.${SLURM_JOB_ID}
+mkdir -p "${job_info_dir}"
+
+date                         > "${job_info_dir}/start-date"
+set                          > "${job_info_dir}/env"
+ldd "${app}"                 > "${job_info_dir}/ldd"
+md5sum "${app}"              > "${job_info_dir}/app-hash"
+readelf -a "${app}"          > "${job_info_dir}/elf"
+echo "${SLURM_JOB_NODELIST}" > "${job_info_dir}/nodes"
+cp "${BASH_SOURCE[0]}"       "${job_info_dir}/script"
+
+# run! #########################################################################
+mpirun -np "${SLURM_NTASKS}" -x LD_LIBRARY_PATH --bind-to none \
+	"${env_dir}/gpu-mpi-wrapper.sh" \
+  "${app}" \
+	--json-out "${job_info_dir}/result.json" \
+	--mpi 1.4.4.4 \
+  --accelerator-threads 8 \
+	--threads 8 \
+	--shm 2048 &> "${job_info_dir}/log"
+
+# if we reach that point the application exited successfully ###################
+touch "${job_info_dir}/success"
+date > "${job_info_dir}/end-date"
+
+################################################################################
--- a/Grid/systems/tursa/files/run.gpu.1nodes.sh
+++ b/Grid/systems/tursa/files/run.gpu.1nodes.sh
@ -0,0 +1,60 @@
+#!/usr/bin/env bash
+# shellcheck disable=SC1091,SC2050,SC2170
+
+#SBATCH -J benchmark-grid-1
+#SBATCH -t 1:00:00
+#SBATCH --nodes=1
+#SBATCH --ntasks=4
+#SBATCH --ntasks-per-node=4
+#SBATCH --cpus-per-task=8
+#SBATCH --partition=gpu
+#SBATCH --gres=gpu:4
+#SBATCH --output=%x.%j.out
+#SBATCH --error=%x.%j.err
+#SBATCH --qos=standard
+#SBATCH --no-requeue
+#SBATCH --gpu-freq=1410
+
+set -euo pipefail
+
+# load environment #############################################################
+env_cfg="${HOME}/.config/lattice-benchmarks/grid-env"
+if [ ! -f "${env_cfg}" ]; then
+	echo "error: ${env_cfg} does not exists, did you execute 'source env.sh' with your user account?"
+	exit 1
+fi
+env_dir="$(readlink -f "$(cat "${env_cfg}")")"
+source "${env_dir}/env.sh"      # load base Spack environment
+source "${env_dir}/env-gpu.sh"  # load GPU-sepcific packages
+source "${env_dir}/ompi-gpu.sh" # set GPU-specific OpenMPI variables
+
+# application and parameters ###################################################
+app="${env_dir}/prefix/gridbench_gpu/bin/Benchmark_Grid"
+
+# collect job information ######################################################
+job_info_dir=job/${SLURM_JOB_NAME}.${SLURM_JOB_ID}
+mkdir -p "${job_info_dir}"
+
+date                         > "${job_info_dir}/start-date"
+set                          > "${job_info_dir}/env"
+ldd "${app}"                 > "${job_info_dir}/ldd"
+md5sum "${app}"              > "${job_info_dir}/app-hash"
+readelf -a "${app}"          > "${job_info_dir}/elf"
+echo "${SLURM_JOB_NODELIST}" > "${job_info_dir}/nodes"
+cp "${BASH_SOURCE[0]}"       "${job_info_dir}/script"
+
+# run! #########################################################################
+mpirun -np "${SLURM_NTASKS}" -x LD_LIBRARY_PATH --bind-to none \
+	"${env_dir}/gpu-mpi-wrapper.sh" \
+  "${app}" \
+	--json-out "${job_info_dir}/result.json" \
+	--mpi 1.1.1.4 \
+  --accelerator-threads 8 \
+	--threads 8 \
+	--shm 2048 &> "${job_info_dir}/log"
+
+# if we reach that point the application exited successfully ###################
+touch "${job_info_dir}/success"
+date > "${job_info_dir}/end-date"
+
+################################################################################
--- a/Grid/systems/tursa/files/run.gpu.32nodes.sh
+++ b/Grid/systems/tursa/files/run.gpu.32nodes.sh
@ -0,0 +1,60 @@
+#!/usr/bin/env bash
+# shellcheck disable=SC1091,SC2050,SC2170
+
+#SBATCH -J benchmark-grid-32
+#SBATCH -t 1:00:00
+#SBATCH --nodes=32
+#SBATCH --ntasks=128
+#SBATCH --ntasks-per-node=4
+#SBATCH --cpus-per-task=8
+#SBATCH --partition=gpu
+#SBATCH --gres=gpu:4
+#SBATCH --output=%x.%j.out
+#SBATCH --error=%x.%j.err
+#SBATCH --qos=standard
+#SBATCH --no-requeue
+#SBATCH --gpu-freq=1410
+
+set -euo pipefail
+
+# load environment #############################################################
+env_cfg="${HOME}/.config/lattice-benchmarks/grid-env"
+if [ ! -f "${env_cfg}" ]; then
+	echo "error: ${env_cfg} does not exists, did you execute 'source env.sh' with your user account?"
+	exit 1
+fi
+env_dir="$(readlink -f "$(cat "${env_cfg}")")"
+source "${env_dir}/env.sh"      # load base Spack environment
+source "${env_dir}/env-gpu.sh"  # load GPU-sepcific packages
+source "${env_dir}/ompi-gpu.sh" # set GPU-specific OpenMPI variables
+
+# application and parameters ###################################################
+app="${env_dir}/prefix/gridbench_gpu/bin/Benchmark_Grid"
+
+# collect job information ######################################################
+job_info_dir=job/${SLURM_JOB_NAME}.${SLURM_JOB_ID}
+mkdir -p "${job_info_dir}"
+
+date                         > "${job_info_dir}/start-date"
+set                          > "${job_info_dir}/env"
+ldd "${app}"                 > "${job_info_dir}/ldd"
+md5sum "${app}"              > "${job_info_dir}/app-hash"
+readelf -a "${app}"          > "${job_info_dir}/elf"
+echo "${SLURM_JOB_NODELIST}" > "${job_info_dir}/nodes"
+cp "${BASH_SOURCE[0]}"       "${job_info_dir}/script"
+
+# run! #########################################################################
+mpirun -np "${SLURM_NTASKS}" -x LD_LIBRARY_PATH --bind-to none \
+	"${env_dir}/gpu-mpi-wrapper.sh" \
+  "${app}" \
+	--json-out "${job_info_dir}/result.json" \
+	--mpi 1.4.4.8 \
+  --accelerator-threads 8 \
+	--threads 8 \
+	--shm 2048 &> "${job_info_dir}/log"
+
+# if we reach that point the application exited successfully ###################
+touch "${job_info_dir}/success"
+date > "${job_info_dir}/end-date"
+
+################################################################################
--- a/Grid/systems/tursa/spack-bootstrap.sh
+++ b/Grid/systems/tursa/spack-bootstrap.sh
@ -4,7 +4,13 @@ set -euo pipefail

 gcc_spec='gcc@9.4.0'
 cuda_spec='cuda@11.4.0'
-hdf5_spec='hdf5@1.10.7'
+
+# 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}"

 if (( $# != 1 )); then
    echo "usage: $(basename "$0") <env dir>" 1>&2
@ -18,7 +24,7 @@ cd "${cwd}"

 # General configuration ########################################################
 # build with 128 tasks
-echo 'config:                                  
+echo 'config:
  build_jobs: 128
  build_stage:
    - $spack/var/spack/stage
@ -38,26 +44,23 @@ rm external.yaml

 # Base compilers ###############################################################
 # configure system base
+
+spack env create base
+spack env activate base
 spack compiler find --scope site

-# install GCC, CUDA & LLVM
-spack install ${gcc_spec} ${cuda_spec} llvm
-
-spack load llvm
+# install GCC, CUDA
+spack add ${gcc_spec} ${cuda_spec}
+spack concretize
+spack env depfile -o Makefile.tmp
+make -j128 -f Makefile.tmp
 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

@ -77,7 +80,7 @@ mkdir -p build_gpu; cd build_gpu
             --enable-devel-headers --enable-examples --enable-optimizations   \
             --with-gdrcopy=${gdrcopy_path} --with-verbs --disable-logging     \
             --disable-debug --disable-assertions --enable-cma                 \
-             --with-knem=/opt/knem-1.1.4.90mlnx1/ --with-rdmacm                \
+             --with-knem=/opt/knem-1.1.4.90mlnx2/ --with-rdmacm                \
             --without-rocm --without-ugni --without-java                      \
             --enable-compiler-opt=3 --with-cuda="${cuda_path}" --without-cm   \
             --with-rc --with-ud --with-dc --with-mlx5-dv --with-dm            \
@ -93,7 +96,7 @@ mkdir -p build_cpu; cd build_cpu
             --enable-devel-headers --enable-examples --enable-optimizations   \
             --with-verbs --disable-logging --disable-debug                    \
             --disable-assertions --enable-mt --enable-cma                     \
-             --with-knem=/opt/knem-1.1.4.90mlnx1/ --with-rdmacm                \
+              --with-knem=/opt/knem-1.1.4.90mlnx2/--with-rdmacm                \
             --without-rocm --without-ugni --without-java                      \
             --enable-compiler-opt=3 --without-cm --without-ugni --with-rc     \
             --with-ud --with-dc --with-mlx5-dv --with-dm --enable-mt --without-go
@ -119,13 +122,13 @@ mkdir build_gpu; cd build_gpu
 ../configure --prefix="${dir}"/prefix/ompi_gpu --without-xpmem    \
             --with-ucx="${dir}"/prefix/ucx_gpu                   \
             --with-ucx-libdir="${dir}"/prefix/ucx_gpu/lib        \
-             --with-knem=/opt/knem-1.1.4.90mlnx1/                 \
+             --with-knem=/opt/knem-1.1.4.90mlnx2/                 \
             --enable-mca-no-build=btl-uct                        \
             --with-cuda="${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
@ -133,7 +136,7 @@ mkdir build_cpu; cd build_cpu
 ../configure --prefix="${dir}"/prefix/ompi_cpu --without-xpmem    \
             --with-ucx="${dir}"/prefix/ucx_cpu                   \
             --with-ucx-libdir="${dir}"/prefix/ucx_cpu/lib        \
-             --with-knem=/opt/knem-1.1.4.90mlnx1/                 \
+             --with-knem=/opt/knem-1.1.4.90mlnx2/                 \
             --enable-mca-no-build=btl-uct --disable-getpwuid     \
             --with-verbs --with-slurm --enable-mpi-fortran=all   \
             --with-pmix=internal --with-libevent=internal
@ -141,65 +144,65 @@ make -j 128
 make install
 cd "${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@1.12.0.GPU%gcc@9.4.0\"
+    - spec: \"${ucx_spec_gpu}\"
      prefix: ${dir}/prefix/ucx_gpu
-    - spec: \"ucx@1.12.0.CPU%gcc@9.4.0\"
+    - spec: \"${ucx_spec_cpu}\"
      prefix: ${dir}/prefix/ucx_cpu
    buildable: False
  openmpi:
    externals:
-    - spec: \"openmpi@4.1.1.GPU%gcc@9.4.0\"
+    - spec: \"${openmpi_spec_gpu}\"
      prefix: ${dir}/prefix/ompi_gpu
-    - spec: \"openmpi@4.1.1.CPU%gcc@9.4.0\"
+    - spec: \"${openmpi_spec_cpu}\"
      prefix: ${dir}/prefix/ompi_cpu
    buildable: False" > spack.yaml

 spack config --scope site add -f spack.yaml
 rm spack.yaml
-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
+spack env deactivate

 cd "${cwd}"

 # environments #################################################################
-dev_tools=("autoconf" "automake" "libtool" "jq")
-ompi_gpu_hash=$(spack find --format "{hash}" openmpi@4.1.1.GPU)
-ompi_cpu_hash=$(spack find --format "{hash}" openmpi@4.1.1.CPU)
+dev_tools=("autoconf" "automake" "libtool" "jq" "git")

 spack env create grid-gpu
 spack env activate grid-gpu
-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 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 env deactivate

 spack env create grid-cpu
 spack env activate grid-cpu
-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 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 env deactivate

 # Final setup ##################################################################
 spack clean
-spack gc -y
+#spack gc -y  # "spack gc" tends to get hung up for unknown reasons

 # add more environment variables in module loading
 spack config --scope site add 'modules:prefix_inspections:lib:[LD_LIBRARY_PATH,LIBRARY_PATH]'
 spack config --scope site add 'modules:prefix_inspections:lib64:[LD_LIBRARY_PATH,LIBRARY_PATH]'
 spack config --scope site add 'modules:prefix_inspections:include:[C_INCLUDE_PATH,CPLUS_INCLUDE_PATH,INCLUDE]'
 spack module tcl refresh -y
-
-# permission change for group access
-chmod -R g+rw "${dir}/spack/var/spack/cache"
-setfacl -d -R -m g::rwX "${dir}/spack/var/spack/cache"
--- a/Quda/.clang-format
+++ b/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,
+}
--- a/Quda/Benchmark_Quda.cpp
+++ b/Quda/Benchmark_Quda.cpp
@ -0,0 +1,458 @@
+#include <algorithm>
+#include <array>
+#include <blas_quda.h>
+#include <cassert>
+#include <chrono>
+#include <color_spinor_field.h>
+#include <communicator_quda.h>
+#include <dirac_quda.h>
+#include <fstream>
+#include <gauge_tools.h>
+#include <memory>
+#include <mpi.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+// remove to use QUDA's own flop counting instead of Grid's convention
+#define FLOP_COUNTING_GRID
+
+#include "json.hpp"
+using nlohmann::json;
+json json_results;
+
+using namespace quda;
+
+// thanks chatGPT :)
+std::string get_timestamp()
+{
+  // Get the current time
+  auto now = std::chrono::system_clock::now();
+
+  // Convert the current time to a time_t object
+  std::time_t currentTime = std::chrono::system_clock::to_time_t(now);
+
+  // Format the time using std::put_time
+  std::stringstream ss;
+  ss << std::put_time(std::localtime(&currentTime), "%Y%m%d %H:%M:%S");
+
+  return ss.str();
+}
+
+// This is the MPI grid, i.e. the layout of ranks
+int nranks = -1;
+std::array<int, 4> mpi_grid = {1, 1, 1, 1};
+
+// run f() in a loop for roughly target_time seconds
+// returns seconds per iteration it took
+template <class F> double bench(F const &f, double target_time, int niter_warmup = 5)
+{
+  device_timer_t timer;
+  timer.start();
+  for (int iter = 0; iter < niter_warmup; ++iter)
+    f();
+  timer.stop();
+
+  double secs = timer.last() / niter_warmup;
+  int niter = std::max(1, int(target_time / secs));
+  // niter = std::min(1000, niter);
+  // printfQuda("during warmup took %f s/iter, deciding on %d iters\n", secs, niter);
+
+  // important: each rank has its own timer, so their measurements can slightly vary. But
+  // 'niter' needs to be consistent (bug took me a couple hours to track down)
+  comm_broadcast_global(&niter, sizeof(niter), 0);
+
+  timer.reset(__FUNCTION__, __FILE__, __LINE__);
+  timer.start();
+  for (int iter = 0; iter < niter; ++iter)
+    f();
+  timer.stop();
+  return timer.last() / niter;
+}
+
+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);
+
+  json_results["geometry"]["ranks"] = nranks;
+  json_results["geometry"]["mpi"] = mpi_grid;
+}
+
+// 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(std::vector<int> const &L_list, double target_time)
+{
+  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 : L_list)
+  {
+    // printfQuda("starting wilson L=%d\n", L);
+
+    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 res = ColorSpinorField(ColorSpinorParam(src));
+    auto f = [&]() { dirac.Dslash(res, src, QUDA_EVEN_PARITY); };
+
+    // first run to get the quda tuning out of the way
+    dirac.Flops(); // reset flops counter
+    f();
+    double flops = 1.0 * dirac.Flops();
+
+    // actual benchmarking
+    auto start_time = get_timestamp();
+    double secs = bench(f, target_time);
+    auto end_time = get_timestamp();
+
+#ifdef FLOP_COUNTING_GRID
+    // this is the flop counting from Benchmark_Grid
+    double Nc = 3;
+    double Nd = 4;
+    double Ns = 4;
+    flops = (Nc * (6 + (Nc - 1) * 8) * Ns * Nd + 2 * Nd * Nc * Ns + 2 * Nd * Nc * Ns * 2);
+    flops *= L * L * L * L / 2.0;
+#endif
+
+    printfQuda("%5d %15.2f %15.2f\n", L, secs * 1e6, flops / secs * 1e-9);
+
+    json tmp;
+    tmp["L"] = L;
+    tmp["Gflops_wilson"] = flops / secs * 1e-9;
+    tmp["start_time"] = start_time;
+    tmp["end_time"] = end_time;
+    json_results["flops"]["results"].push_back(tmp);
+  }
+}
+
+void benchmark_dwf(std::vector<int> const &L_list, double target_time)
+{
+  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 : L_list)
+  {
+    // printfQuda("starting dwf L=%d\n", L);
+    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 res = ColorSpinorField(ColorSpinorParam(src));
+    auto f = [&]() { dirac.Dslash(res, src, QUDA_EVEN_PARITY); };
+
+    // first run to get the quda tuning out of the way
+    dirac.Flops(); // reset flops counter
+    f();
+    double flops = 1.0 * dirac.Flops();
+
+    // actual benchmarking
+    auto start_time = get_timestamp();
+    double secs = bench(f, target_time);
+    auto end_time = get_timestamp();
+
+#ifdef FLOP_COUNTING_GRID
+    // this is the flop counting from Benchmark_Grid
+    double Nc = 3;
+    double Nd = 4;
+    double Ns = 4;
+    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;
+#endif
+
+    printfQuda("%5d %15.2f %15.2f\n", L, secs * 1e6, flops / secs * 1e-9);
+    json tmp;
+    tmp["L"] = L;
+    tmp["Gflops_dwf4"] = flops / secs * 1e-9;
+    tmp["start_time"] = start_time;
+    tmp["end_time"] = end_time;
+    json_results["flops"]["results"].push_back(tmp);
+  }
+}
+
+void benchmark_axpy(std::vector<int> const &L_list, double target_time)
+{
+  // number of iterations for warmup / measurement
+  // (feel free to change for noise/time tradeoff)
+  constexpr int niter_warmup = 5;
+
+  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");
+  for (int L : L_list)
+  {
+    // printfQuda("starting axpy L=%d\n", L);
+    //  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;
+
+    auto f = [&]() { blas::axpy(1.234, fieldA, fieldB); };
+
+    // first run to get the quda tuning out of the way
+    f();
+
+    // actual benchmarking
+    auto start_time = get_timestamp();
+    double secs = bench(f, target_time);
+    auto end_time = get_timestamp();
+
+    double mem_MiB = memory / 1024. / 1024.;
+    double GBps = mem_MiB / 1024 / secs;
+    printfQuda("%5d %15.2f %15.2f %15.2f %15.2f\n", L, mem_MiB, secs * 1e6, GBps,
+               flops / secs * 1e-9);
+
+    json tmp;
+    tmp["L"] = L;
+    tmp["size_MB"] = mem_MiB;
+    tmp["GBps"] = GBps;
+    tmp["GFlops"] = flops / secs * 1e-9;
+    tmp["start_time"] = start_time;
+    tmp["end_time"] = end_time;
+    json_results["axpy"].push_back(tmp);
+  }
+}
+
+int main(int argc, char **argv)
+{
+  std::string json_filename = ""; // empty indicates no json output
+  for (int i = 0; i < argc; i++)
+  {
+    if (std::string(argv[i]) == "--json-out")
+      json_filename = argv[i + 1];
+  }
+
+  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({8, 12, 16, 24, 32, 48}, 1.0);
+
+  setVerbosity(QUDA_SILENT);
+  benchmark_wilson({8, 12, 16, 24, 32, 48}, 1.0);
+  benchmark_dwf({8, 12, 16, 24, 32}, 1.0);
+  setVerbosity(QUDA_SUMMARIZE);
+
+  printfQuda("==================== done with all benchmarks ====================\n");
+
+  if (!json_filename.empty())
+  {
+    printfQuda("writing benchmark results to %s\n", json_filename.c_str());
+
+    int me = 0;
+    MPI_Comm_rank(MPI_COMM_WORLD, &me);
+    if (me == 0)
+    {
+      std::ofstream json_file(json_filename);
+      json_file << std::setw(2) << json_results;
+    }
+  }
+
+  endQuda();
+  quda::comm_finalize();
+  MPI_Finalize();
+}
--- a/Quda/Readme.md
+++ b/Quda/Readme.md
@ -0,0 +1,30 @@
+# QUDA benchmarks
+
+This folder contains benchmarks for the [QUDA](https://github.com/lattice/quda) library.
+
+- `Benchmark_Quda`: This benchmark measure floating point performances of fermion
+matrices (Wilson and DWF), as well as memory bandwidth (using a simple `axpy` operation). Measurements are
+performed for a fixed range of problem sizes.
+
+## Building
+After setting up your compilation environment (Tursa: `source /home/dp207/dp207/shared/env/production/env-{base,gpu}.sh`):
+```bash
+./build-quda.sh <env_dir>          # build Quda
+./build-benchmark.sh <env_dir>     # build benchmark
+```
+where `<env_dir>` is an arbitrary directory where every product will be stored.
+
+## Running the Benchmark
+
+The benchmark should be run as
+```bash
+mpirun -np <ranks> <env_dir>/prefix/qudabench/Benchmark_Quda
+```
+where `<ranks>` is the total number of GPU's to use. On Tursa this is 4 times the number of nodes.
+
+Note:
+- on Tursa, the `wrapper.sh` script that is typically used with Grid is not necessary.
+- due to Qudas automatic tuning, the benchmark might take significantly longer to run than `Benchmark_Grid` (even though it does fewer things).
+  - setting `QUDA_ENABLE_TUNING=0` disables all tuning (degrades performance severely). By default, it is turned on.
+  - setting `QUDA_RESOURCE_PATH=<some folder>` enables Quda to save and reuse optimal tuning parameters, making repeated runs much faster
+  
--- a/Quda/build-benchmark.sh
+++ b/Quda/build-benchmark.sh
@ -0,0 +1,32 @@
+#!/usr/bin/env bash
+# shellcheck disable=SC1090,SC1091
+
+set -euo pipefail
+
+if (( $# != 1 )); then
+    echo "usage: $(basename "$0") <environment directory>" 1>&2
+    exit 1
+fi
+env_dir=$1
+
+# TODO: this is Tursa specific. have not figured out the correct way to do this.
+EXTRA_LIBS="/home/dp207/dp207/shared/env/versions/220428/spack/opt/spack/linux-rhel8-zen2/gcc-9.4.0/cuda-11.4.0-etxow4jb23qdbs7j6txczy44cdatpj22/lib64/stubs/libcuda.so /home/dp207/dp207/shared/env/versions/220428/spack/opt/spack/linux-rhel8-zen2/gcc-9.4.0/cuda-11.4.0-etxow4jb23qdbs7j6txczy44cdatpj22/lib64/stubs/libnvidia-ml.so"
+
+# NOTE: these flags need to be in sync with Qudas compilation options (see build-quda.sh)
+BUILD_FLAGS="-O3 -std=c++17 -DMPI_COMMS -DMULTI_GPU -DQUDA_PRECISION=12 -DQUDA_RECONSTRUCT=4"
+
+call_dir=$(pwd -P)
+script_dir="$(dirname "$(readlink -f "${BASH_SOURCE:-$0}")")"
+cd "${env_dir}"
+env_dir=$(pwd -P)
+cd "${call_dir}"
+BUILD_DIR="${env_dir}/build/Quda-benchmarks"
+PREFIX_DIR="${env_dir}/prefix/qudabench"
+QUDA_DIR=${env_dir}/prefix/quda
+mkdir -p "${BUILD_DIR}"
+mkdir -p "${PREFIX_DIR}"
+
+LINK_FLAGS="-Wl,-rpath,$QUDA_DIR/lib: $QUDA_DIR/lib/libquda.so $EXTRA_LIBS -lpthread -lmpi"
+
+g++ $BUILD_FLAGS -I$QUDA_DIR/include/targets/cuda  -I$QUDA_DIR/include   -c -o $BUILD_DIR/Benchmark_Quda.o  $script_dir/Benchmark_Quda.cpp
+g++ -g -O3 $BUILD_DIR/Benchmark_Quda.o -o $PREFIX_DIR/Benchmark_Quda $LINK_FLAGS -lmpi
--- a/Quda/build-quda.sh
+++ b/Quda/build-quda.sh
@ -0,0 +1,36 @@
+#!/usr/bin/env bash
+# shellcheck disable=SC1090,SC1091
+
+BUILD_FLAGS="-O3 -std=c++17"
+QUDA_FLAGS="-DQUDA_MPI=ON -DQUDA_PRECISION=14 -DQUDA_RECONSTRUCT=4 -DQUDA_GPU_ARCH=sm_80"
+
+set -euo pipefail
+
+if (( $# != 1 )); then
+    echo "usage: $(basename "$0") <environment directory>" 1>&2
+    exit 1
+fi
+env_dir=$1
+
+call_dir=$(pwd -P)
+mkdir -p ${env_dir}
+cd "${env_dir}"
+env_dir=$(pwd -P)
+cd "${call_dir}"
+
+build_dir="${env_dir}/build/quda"
+if [ -d "${build_dir}" ]; then
+    echo "error: directory '${build_dir}' exists"
+    exit 1
+fi
+mkdir -p "${build_dir}"
+
+git clone https://github.com/lattice/quda.git "${build_dir}"
+cd "${build_dir}"
+
+mkdir build; cd build
+cmake .. $QUDA_FLAGS -DCMAKE_INSTALL_PREFIX=${env_dir}/prefix/quda
+make -j128
+make install
+
+cd "${call_dir}"
--- a/Quda/env.sh
+++ b/Quda/env.sh
@ -0,0 +1,21 @@
+module load gcc/9.3.0
+module load cuda/11.4.1
+module load openmpi/4.1.1-cuda11.4
+
+export QUDA_RESOURCE_PATH=$(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
Author	SHA1	Message	Date
Antonin Portelli	cb538bfbf1	Merge pull request 'Updates GPU launch wrapper and install script' (#8 ) from aturner/lattice-benchmarks:main into main Reviewed-on: portelli/lattice-benchmarks#8	2025-03-31 16:08:45 +01:00
Andy Turner	9056e9023c	Updates knem location in install script	2024-12-17 13:46:43 +00:00
Antonin Portelli	c049a2ad0b	Merge pull request 'benchmark-quda' (#3 ) from simon.buerger/lattice-benchmarks:benchmark-quda into main Reviewed-on: portelli/lattice-benchmarks#3	2024-11-26 10:57:28 +00:00
Simon Bürger	fb43d16830	log iso-timestamp instead of seconds-since-start	2024-11-25 15:45:59 +00:00
Simon Bürger	6fa2e6bcd0	fix bug that made benchmark_quda hang randomly	2024-11-25 15:45:59 +00:00
Simon Bürger	fb4c456776	choose iteration count automatically	2024-11-25 15:45:58 +00:00
Simon Bürger	3fbb8ea346	add timestamps to benchmarks	2024-11-25 15:45:58 +00:00
Simon Bürger	86b160cb5c	add json output to Benchmark_Quda	2024-11-25 15:45:58 +00:00
Simon Bürger	dc411017bb	Update 'Quda/Readme.md'	2024-11-25 15:45:58 +00:00
Simon Bürger	b2cc780690	add Readme.md to Quda benchmark	2024-11-25 15:45:58 +00:00
Simon Bürger	6d87396576	clean up build script a bit	2024-11-25 15:45:58 +00:00
Simon Bürger	e9d084ce09	better range of lattice sizes	2024-11-25 15:45:58 +00:00
Simon Bürger	32e301fc67	add DWF benchmark	2024-11-25 15:45:58 +00:00
Simon Bürger	eaa4feee43	benchmark Dslash(...) instead of full M(...)	2024-11-25 15:45:58 +00:00
Simon Bürger	025f9dab50	fix scaling conventions for multi-gpu	2024-11-25 15:45:57 +00:00
Simon Bürger	3a561091d9	tidy up the wilson benchmark and add environment script	2024-11-25 15:45:57 +00:00
Simon Bürger	191c0cfca5	add quda axpy/memory benchmark	2024-11-25 15:45:57 +00:00
Simon Bürger	6f9af8acad	first draft of Quda Benchmark	2024-11-25 15:45:57 +00:00
Antonin Portelli	371a329457	Merge pull request 'Point-to-Point latency' (#7 ) from simon.buerger/lattice-benchmarks:latency_benchmark into main Reviewed-on: portelli/lattice-benchmarks#7 Reviewed-by: Antonin Portelli <antonin.portelli@me.com>	2024-11-19 10:37:26 +00:00
Simon Bürger	f81cb198ab	add command line options to Benchmark_Grid	2024-11-18 23:50:45 +00:00
Simon Bürger	a7e1d9e67f	lower loop counts a bit for p2p/latency	2024-10-11 18:27:00 +01:00
Simon Bürger	19c9dcb6ae	fix order of ranks in latency/p2p	2024-10-10 11:40:44 +01:00
Simon Bürger	7d89380b80	point-to-point bandwith benchmark	2024-10-07 17:22:26 +01:00
Simon Bürger	4cd67805b9	make Latency benchmark proper one-way and increase statistics	2024-09-26 09:31:22 +01:00
Simon Bürger	f7e607eae4	proper warmup loop for latency	2024-05-09 23:33:04 +01:00
Simon Bürger	a267986800	naming consitency	2024-05-09 23:25:06 +01:00
Simon Bürger	a1ec08cdb3	point-to-point latency	2024-05-09 23:17:54 +01:00
Antonin Portelli	fb6c79d9ca	shm direction fix	2024-03-21 13:51:49 +09:00
Simon Bürger	d7647afa72	Merge remote-tracking branch 'upstream/main' into main	2024-01-23 10:10:52 +00:00
Antonin Portelli	ba00493c7d	Merge pull request 'fix incompatibility with latest Grid' (#5 ) from simon.buerger/lattice-benchmarks:main into main Reviewed-on: portelli/lattice-benchmarks#5	2023-12-20 15:06:41 +00:00
Simon Bürger	6055e0503c	simple latency benchmark	2023-12-20 13:43:51 +00:00
Simon Bürger	6ea093fc80	fix incompatibility with latest Grid	2023-12-18 16:48:19 +00:00
Antonin Portelli	fa47ec5bbe	Merge pull request 'refactor and repair the spack environment' (#4 ) from simon.buerger/lattice-benchmarks:fix_spack_environment into main Reviewed-on: portelli/lattice-benchmarks#4	2023-07-05 15:11:40 +01:00
Simon Bürger	7235bfde4c	refactor and repair the spack environment	2023-07-04 22:30:54 +01:00
Antonin Portelli	e5c61c2db1	Merge pull request 'add indication of shared-memory directions in comms benchmark' (#2 ) from simon.buerger/lattice-benchmarks:feature/grid-shared-mem into main Reviewed-on: portelli/lattice-benchmarks#2	2023-04-12 15:05:39 +01:00
Simon Bürger	80c80049d7	add indication of shared-memory directions in comms benchmark	2023-04-12 11:40:39 +01:00
Antonin Portelli	ce0d4d9457	Grid comms warmup sequence	2023-02-03 20:59:20 +00:00
Antonin Portelli	cc4c0255bc	Grid Tursa 32 node MPI layout change	2023-02-03 20:59:02 +00:00
Antonin Portelli	bdfb94bf11	Grid overflow fix	2023-02-03 20:58:41 +00:00
Antonin Portelli	af950e6e28	Grid Tursa 32 node batch script	2023-02-01 23:55:50 +00:00
Antonin Portelli	14fb2fddc2	Grid benchmark build script fix	2023-02-01 23:55:37 +00:00
Antonin Portelli	5198bbe1cd	Grid benchmark documentation update	2023-01-30 18:28:38 +00:00
Antonin Portelli	5f9abbb8d0	minor label fix	2023-01-30 18:28:17 +00:00
Antonin Portelli	9b6c6d4d40	Grid: geometry JSON output and cleanup	2023-01-30 17:43:35 +00:00
Antonin Portelli	43e264d29b	Proper runscript example for Tursa	2023-01-30 16:58:54 +00:00
Antonin Portelli	51eae5723e	Grid IO benchmark cleanup	2023-01-29 14:56:37 +00:00
Antonin Portelli	ce890a8fc2	Merge branch 'main' of git.dev.dirac.ed.ac.uk:portelli/lattice-benchmarks	2023-01-28 15:28:34 +00:00
Antonin Portelli	8f1a556afa	Finale cleanup!	2023-01-28 15:28:23 +00:00
Antonin Portelli	58080730ae	FInale cleanup!	2023-01-28 15:26:03 +00:00
Antonin Portelli	f180cbb8ec	JSON fix	2023-01-28 14:44:32 +00:00
Antonin Portelli	5098f57f08	Grid Tursa wrapper permission fix	2023-01-28 14:19:47 +00:00
Antonin Portelli	0e2b7225db	Gird Makefile.am fix	2023-01-28 14:19:25 +00:00
Antonin Portelli	7b689a8c94	removing old benchmarks	2023-01-28 14:18:56 +00:00
Antonin Portelli	77c75ea5b8	Memory benchmarks cleanup and fix	2023-01-28 14:18:48 +00:00
Antonin Portelli	78c464d1d8	Grid benchmark big formatting cleanup	2023-01-27 18:49:53 +00:00
Antonin Portelli	8f043343fb	Trusa Grid env fix	2023-01-27 17:56:10 +00:00