add quda axpy/memory benchmark

Quda/Benchmark_Quda.cpp

@@ -16,6 +16,7 @@ using namespace quda;
 
 QudaPrecision smoother_halo_prec = QUDA_INVALID_PRECISION;
 
+// This is the MPI grid, i.e. the layout of ranks, not the lattice volume
 std::array<int, 4> gridsize = {1, 1, 1, 4};
 
 void initComms(int argc, char **argv, std::array<int, 4> const &commDims)
@@ -31,6 +32,7 @@ void initComms(int argc, char **argv, std::array<int, 4> const &commDims)
       rank = gridsize[i] * rank + coords[i];
     return rank;
   };
+
   initCommsGridQuda(4, commDims.data(), lex_rank_from_coords, nullptr);
 
   for (int d = 0; d < 4; d++)
@@ -45,32 +47,57 @@ cudaGaugeField make_gauge_field(std::array<int, 4> const &geom)
 
   // dimension and type of the lattice object
   param.nDim = 4;
-  param.nColor = 3;
   param.x[0] = geom[0];
   param.x[1] = geom[1];
   param.x[2] = geom[2];
   param.x[3] = geom[3];
+
+  // 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;
-  param.siteSubset = QUDA_FULL_SITE_SUBSET; // no even/odd, just a full lattice
+
-  param.link_type = QUDA_SU3_LINKS;
+  // for this benchmark we only need "SINGLE" and/or "DOUBLE" precision. But smaller
+  // precisions are available in QUDA too
   param.setPrecision(QUDA_DOUBLE_PRECISION);
 
-  param.create = QUDA_NULL_FIELD_CREATE;     // do not (zero-) initilize the fields
+  // no even/odd subset, we want a full lattice
-  param.location = QUDA_CUDA_FIELD_LOCATION; // field should live on the accelerator
+  param.siteSubset = QUDA_FULL_SITE_SUBSET;
 
-  // turn off advanced features we dont care about for this benchmark
+  // 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;
 
-  // these control the physical data layout. Might be interesting to try out different
+  // This controls the physical order of elements. "float2" is the the default
-  // settings
   param.order = QUDA_FLOAT2_GAUGE_ORDER;
-  param.geometry = QUDA_SCALAR_GEOMETRY;
+
+  // 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);
-  quda::RNG rng(U, /*seed=*/1234);
+  gaugeGauss(U, /*seed=*/1234, 1.0);
-  gaugeGauss(U, rng, 1.0);
   return U;
 }
 
@@ -100,11 +127,11 @@ ColorSpinorField make_source(std::array<int, 4> const &geom)
   auto src = ColorSpinorField(param);
   quda::RNG rng(src, 1234);
   spinorNoise(src, rng, QUDA_NOISE_GAUSS);
-  printfQuda(
+  /*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.Bytes() * 1.0);*/
-  src.PrintDims();
+  // src.PrintDims();
 
   return src;
 }
@@ -113,8 +140,16 @@ void benchmark(int L, int niter)
 {
   std::array<int, 4> geom = {L, L, L, L};
 
+  printfQuda("=======================  benchmarking L=%d =======================\n", L);
+
   auto U = make_gauge_field(geom);
+  printfQuda("created random gauge field, %.3f GiB (sanity check: should be %.3f)\n",
+             U.Bytes() / 1024. / 1024. / 1024.,
+             1.0 * L * L * L * L * 4 * 18 * 8 / 1024. / 1024. / 1024.);
   auto src = make_source(geom);
+  printfQuda("created random source, %.3f GiB (sanity check: should be %.3f)\n",
+             src.Bytes() / 1024. / 1024. / 1024.,
+             1.0 * L * L * L * L * 12 * 2 * 8 / 1024. / 1024. / 1024.);
 
   // create (Wilson) dirac operator
   DiracParam param;
@@ -134,9 +169,11 @@ void benchmark(int L, int niter)
              geom[1], geom[2], geom[3], niter);
 
   // couple iterations without timing to warm up
+  printfQuda("warmup...\n");
   for (int iter = 0; iter < 20; ++iter)
     dirac.M(tmp, src);
 
+  printfQuda("running...\n");
   dirac.Flops(); // reset flops counter
   device_timer_t device_timer;
   device_timer.start();
@@ -146,22 +183,82 @@ void benchmark(int L, int niter)
 
   double secs = device_timer.last();
   double gflops = (dirac.Flops() * 1e-9) / secs;
-  printfQuda("Gflops = %6.1f\n", gflops);
+  printfQuda("Gflops = %6.2f\n", gflops);
+}
+
+void benchmark_axpy(int L)
+{
+  printfQuda("================ axpy L=%d ==============\n", L);
+
+  ColorSpinorParam param;
+  param.nColor = 3;
+  param.nSpin = 4;
+  param.nVec = 1;
+  param.pad = 0;
+  param.siteSubset = QUDA_FULL_SITE_SUBSET;
+  param.nDim = 4;
+  param.x[0] = L;
+  param.x[1] = L;
+  param.x[2] = L;
+  param.x[3] = L;
+  param.x[4] = 1; // no fifth dimension
+  param.pc_type = QUDA_4D_PC;
+  param.siteOrder = QUDA_EVEN_ODD_SITE_ORDER;
+  param.gammaBasis = QUDA_DEGRAND_ROSSI_GAMMA_BASIS;
+  param.create = QUDA_NULL_FIELD_CREATE; // do not (zero-) initilize the field
+  param.setPrecision(QUDA_DOUBLE_PRECISION);
+  param.location = QUDA_CUDA_FIELD_LOCATION;
+
+  // create the field and fill it with random values
+  auto fieldA = ColorSpinorField(param);
+  auto fieldB = ColorSpinorField(param);
+  quda::RNG rng(fieldA, 1234);
+  auto size_bytes = size_t(8) * 2 * param.x[0] * param.x[1] * param.x[2] * param.x[3] *
+                    param.nColor * param.nSpin;
+  assert(fieldA.Bytes() == size_bytes); // sanity check
+  assert(fieldB.Bytes() == size_bytes); // sanity check
+  spinorNoise(fieldA, rng, QUDA_NOISE_GAUSS);
+  spinorNoise(fieldB, rng, QUDA_NOISE_GAUSS);
+
+  // number of (real) elements in the field = number of fma instructions to do
+  double flops_per_iter =
+      2 * param.x[0] * param.x[1] * param.x[2] * param.x[3] * param.nColor * param.nSpin;
+
+  int niter = 20;
+
+  printfQuda("warmup...\n");
+  for (int iter = 0; iter < 10; ++iter)
+    blas::axpy(1.234, fieldA, fieldB);
+
+  printfQuda("running...\n");
+  device_timer_t device_timer;
+  device_timer.start();
+  for (int iter = 0; iter < niter; ++iter)
+    blas::axpy(1.234, fieldA, fieldB); // fieldB += 1.234*fieldA
+  device_timer.stop();
+
+  double secs = device_timer.last();
+  double gflops = (flops_per_iter * niter) * 1e-9 / secs;
+  printfQuda("Gflops = %6.2f\n", gflops);
+  printfQuda("bytes = %6.2f GiB\n", 3. * fieldA.Bytes() / 1024. / 1024. / 1024.);
+  printfQuda("bandwidth = %6.2f GiB/s\n",
+             fieldA.Bytes() * 3 / 1024. / 1024. / 1024. * niter / secs);
 }
 
 int main(int argc, char **argv)
 {
   initComms(argc, argv, gridsize);
 
-  // -1 for multi-gpu. otherwise this selects the device to be used
+  initQuda(-1); // -1 for multi-gpu. otherwise this selects the device to be used
-  initQuda(-1);
 
-  // verbosity options are:
+  //  verbosity options are:
-  // SILENT, SUMMARIZE, VERBOSE, DEBUG_VERBOSE
+  //  SILENT, SUMMARIZE, VERBOSE, DEBUG_VERBOSE
-  setVerbosity(QUDA_SUMMARIZE);
+  setVerbosity(QUDA_VERBOSE);
 
-  for (int L : {8, 16, 24, 32})
+  for (int L : {8, 12, 16, 24, 32})
-    benchmark(L, 1000);
+    benchmark_axpy(L);
+  for (int L : {16, 24, 32, 48, 64})
+    benchmark(L, 100);
 
   endQuda();
   quda::comm_finalize();

Quda/build.sh

@@ -1,11 +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=
 
-set -e
+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: $QUDA_BUILD/lib/libquda.so /home/dp207/dp207/shared/env/versions/220428/spack/opt/spack/linux-rhel8-zen2/gcc-9.4.0/cuda-11.4.0-etxow4jb23qdbs7j6txczy44cdatpj22/lib64/stubs/libcuda.so /home/dp207/dp207/shared/env/versions/220428/spack/opt/spack/linux-rhel8-zen2/gcc-9.4.0/cuda-11.4.0-etxow4jb23qdbs7j6txczy44cdatpj22/lib64/stubs/libnvidia-ml.so /home/dp207/dp207/shared/env/versions/220428/spack/opt/spack/linux-rhel8-zen2/gcc-9.4.0/cuda-11.4.0-etxow4jb23qdbs7j6txczy44cdatpj22/lib64/libcudart_static.a -ldl /usr/lib64/librt.so /home/dp207/dp207/shared/env/versions/220428/spack/opt/spack/linux-rhel8-zen2/gcc-9.4.0/cuda-11.4.0-etxow4jb23qdbs7j6txczy44cdatpj22/lib64/libcublas.so /home/dp207/dp207/shared/env/versions/220428/spack/opt/spack/linux-rhel8-zen2/gcc-9.4.0/cuda-11.4.0-etxow4jb23qdbs7j6txczy44cdatpj22/lib64/libcufft.so -lpthread"
-COMPILE_FLAGS="-DMPI_COMMS -DMULTI_GPU -DQUDA_PRECISION=14 -DQUDA_RECONSTRUCT=7 -g -O3 -Wall -Wextra -pthread -std=c++17"
+$CXX -g -O3 Benchmark_Quda.o -o Benchmark_Quda $LINK_FLAGS -lmpi
-LINK_FLAGS="-g -O3 -Wl,-rpath -Wl,/mnt/lustre/tursafs1/home/dp207/dp207/shared/env/versions/220428/prefix/ompi_gpu/lib -Wl,--enable-new-dtags -L/mnt/lustre/tursafs1/home/dp207/dp207/shared/env/versions/220428/prefix/ompi_gpu/lib -pthread -Wl,-rpath,/home/dp207/dp207/dc-burg2/quda_build/tests:/home/dp207/dp207/dc-burg2/quda_build/lib:/home/dp207/dp207/shared/env/versions/220428/spack/opt/spack/linux-rhel8-zen2/gcc-9.4.0/cuda-11.4.0-etxow4jb23qdbs7j6txczy44cdatpj22/lib64/stubs:/home/dp207/dp207/shared/env/versions/220428/spack/opt/spack/linux-rhel8-zen2/gcc-9.4.0/cuda-11.4.0-etxow4jb23qdbs7j6txczy44cdatpj22/lib64: ../../quda_install/lib/libquda.so /home/dp207/dp207/shared/env/versions/220428/spack/opt/spack/linux-rhel8-zen2/gcc-9.4.0/cuda-11.4.0-etxow4jb23qdbs7j6txczy44cdatpj22/lib64/stubs/libcuda.so /home/dp207/dp207/shared/env/versions/220428/spack/opt/spack/linux-rhel8-zen2/gcc-9.4.0/cuda-11.4.0-etxow4jb23qdbs7j6txczy44cdatpj22/lib64/stubs/libnvidia-ml.so /home/dp207/dp207/shared/env/versions/220428/spack/opt/spack/linux-rhel8-zen2/gcc-9.4.0/cuda-11.4.0-etxow4jb23qdbs7j6txczy44cdatpj22/lib64/libcudart_static.a -ldl /usr/lib64/librt.so /home/dp207/dp207/shared/env/versions/220428/spack/opt/spack/linux-rhel8-zen2/gcc-9.4.0/cuda-11.4.0-etxow4jb23qdbs7j6txczy44cdatpj22/lib64/libcublas.so /home/dp207/dp207/shared/env/versions/220428/spack/opt/spack/linux-rhel8-zen2/gcc-9.4.0/cuda-11.4.0-etxow4jb23qdbs7j6txczy44cdatpj22/lib64/libcufft.so -lpthread /home/dp207/dp207/shared/env/versions/220428/prefix/ompi_gpu/lib/libmpi.so"
-
-$CXX $COMPILE_FLAGS -I/home/dp207/dp207/dc-burg2/quda_install/include/ -o Benchmark_Quda.o -c Benchmark_Quda.cpp
-$CXX  $LINK_FLAGS Benchmark_Quda.o -o Benchmark_Quda
-