Quda/Benchmark_Quda.cpp

@@ -12,6 +12,9 @@
 
 using namespace quda;
 
+// remove to use QUDA's own flop counting instead of Grid's convention
+#define FLOP_COUNTING_GRID
+
 // This is the MPI grid, i.e. the layout of ranks
 int nranks = -1;
 std::array<int, 4> mpi_grid = {1, 1, 1, 1};
@@ -65,7 +68,7 @@ cudaGaugeField make_gauge_field(int L)
 
   // 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.setPrecision(QUDA_SINGLE_PRECISION);
 
   // no even/odd subset, we want a full lattice
   param.siteSubset = QUDA_FULL_SITE_SUBSET;
@@ -104,7 +107,7 @@ cudaGaugeField make_gauge_field(int L)
 }
 
 // create a random source vector (L = local size)
-ColorSpinorField make_source(int L)
+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
@@ -116,12 +119,12 @@ ColorSpinorField make_source(int L)
   param.nVec = 1; // only a single vector
   param.pad = 0;
   param.siteSubset = QUDA_PARITY_SITE_SUBSET;
-  param.nDim = 4;
+  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] = 1; // no fifth dimension
+  param.x[4] = Ls;
   param.pc_type = QUDA_4D_PC;
   param.siteOrder = QUDA_EVEN_ODD_SITE_ORDER;
 
@@ -130,7 +133,7 @@ ColorSpinorField make_source(int L)
   param.gammaBasis = QUDA_UKQCD_GAMMA_BASIS;
 
   param.create = QUDA_NULL_FIELD_CREATE; // do not (zero-) initilize the field
-  param.setPrecision(QUDA_DOUBLE_PRECISION);
+  param.setPrecision(QUDA_SINGLE_PRECISION);
   param.location = QUDA_CUDA_FIELD_LOCATION;
 
   // create the field and fill it with random values
@@ -152,10 +155,15 @@ void benchmark_wilson()
   int niter_warmup = 10;
 
   printfQuda("==================== wilson dirac operator ====================\n");
-  printfQuda("IMPORTANT: QUDAs own flop counting. Probably not the same as in Grid.\n");
+#ifdef FLOP_COUNTING_GRID
+  printfQuda("IMPORTANT: flop counting as in Benchmark_Grid\n");
+#else
+  printfQuda("IMPORTANT: flop counting by QUDA's own convention (different from "
+             "Benchmark_Grid)\n");
+#endif
   printfQuda("%5s %15s %15s\n", "L", "time (usec)", "Gflop/s/rank");
 
-  for (int L : {8, 12, 16, 24, 32})
+  for (int L : {8, 12, 16, 24, 32, 48})
   {
     auto U = make_gauge_field(L);
     auto src = make_source(L);
@@ -187,7 +195,82 @@ void benchmark_wilson()
     device_timer.stop();
 
     double secs = device_timer.last() / niter;
+
+#ifdef FLOP_COUNTING_GRID
+    // this is the flop counting from Benchmark_Grid
+    double Nc = 3;
+    double Nd = 4;
+    double Ns = 4;
+    double flops =
+        (Nc * (6 + (Nc - 1) * 8) * Ns * Nd + 2 * Nd * Nc * Ns + 2 * Nd * Nc * Ns * 2);
+    flops *= L * L * L * L / 2.0;
+#else
     double flops = 1.0 * dirac.Flops() / niter;
+#endif
+
+    printfQuda("%5d %15.2f %15.2f\n", L, secs * 1e6, flops / secs * 1e-9);
+  }
+}
+
+void benchmark_dwf()
+{
+  int niter = 20;
+  int niter_warmup = 10;
+
+  printfQuda("==================== domain wall dirac operator ====================\n");
+#ifdef FLOP_COUNTING_GRID
+  printfQuda("IMPORTANT: flop counting as in Benchmark_Grid\n");
+#else
+  printfQuda("IMPORTANT: flop counting by QUDA's own convention (different from "
+             "Benchmark_Grid)\n");
+#endif
+  printfQuda("%5s %15s %15s\n", "L", "time (usec)", "Gflop/s/rank");
+  int Ls = 12;
+  for (int L : {8, 12, 16, 24, 32, 48})
+  {
+    auto U = make_gauge_field(L);
+    auto src = make_source(L, Ls);
+
+    // create dirac operator
+    DiracParam param;
+    param.kappa = 0.10;
+    param.Ls = Ls;
+    param.m5 = 0.1;
+    param.dagger = QUDA_DAG_NO;
+    param.matpcType = QUDA_MATPC_EVEN_EVEN;
+    auto dirac = DiracDomainWall(param);
+
+    // insert gauge field into the dirac operator
+    // (the additional nullptr's are for smeared links and fancy preconditioners and such)
+    dirac.updateFields(&U, nullptr, nullptr, nullptr);
+
+    auto tmp = ColorSpinorField(ColorSpinorParam(src));
+
+    // couple iterations without timing to warm up
+    for (int iter = 0; iter < niter_warmup; ++iter)
+      dirac.Dslash(tmp, src, QUDA_EVEN_PARITY);
+
+    // actual benchmark with timings
+    dirac.Flops(); // reset flops counter
+    device_timer_t device_timer;
+    device_timer.start();
+    for (int iter = 0; iter < niter; ++iter)
+      dirac.Dslash(tmp, src, QUDA_EVEN_PARITY);
+    device_timer.stop();
+
+    double secs = device_timer.last() / niter;
+
+#ifdef FLOP_COUNTING_GRID
+    // this is the flop counting from Benchmark_Grid
+    double Nc = 3;
+    double Nd = 4;
+    double Ns = 4;
+    double flops =
+        (Nc * (6 + (Nc - 1) * 8) * Ns * Nd + 2 * Nd * Nc * Ns + 2 * Nd * Nc * Ns * 2);
+    flops *= L * L * L * L * Ls / 2.0;
+#else
+    double flops = 1.0 * dirac.Flops() / niter;
+#endif
 
     printfQuda("%5d %15.2f %15.2f\n", L, secs * 1e6, flops / secs * 1e-9);
   }
@@ -213,7 +296,7 @@ void benchmark_axpy()
   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_DOUBLE_PRECISION);
+  param.setPrecision(QUDA_SINGLE_PRECISION);
 
   // the following dont matter for an axpy benchmark, but need to choose something
   param.pc_type = QUDA_4D_PC;
@@ -240,8 +323,8 @@ void benchmark_axpy()
     // create the field(s)
     auto fieldA = ColorSpinorField(param);
     auto fieldB = ColorSpinorField(param);
-    assert(fieldA.Bytes() == sizeof(double) * field_elements); // sanity check
+    assert(fieldA.Bytes() == sizeof(float) * field_elements); // sanity check
-    assert(fieldB.Bytes() == sizeof(double) * field_elements); // sanity check
+    assert(fieldB.Bytes() == sizeof(float) * field_elements); // sanity check
 
     // fill fields with random values
     quda::RNG rng(fieldA, 1234);
@@ -251,7 +334,7 @@ void benchmark_axpy()
     // 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(double) * field_elements;
+    double memory = 3 * sizeof(float) * field_elements;
 
     // do some iterations to to let QUDA do its internal tuning and also stabilize cache
     // behaviour and such
@@ -288,6 +371,7 @@ int main(int argc, char **argv)
 
   setVerbosity(QUDA_SILENT);
   benchmark_wilson();
+  benchmark_dwf();
   setVerbosity(QUDA_SUMMARIZE);
 
   printfQuda("==================== done with all benchmarks ====================\n");

Quda/env.sh

@@ -2,6 +2,7 @@ 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