From b4b70702fd5c3897367afd67700e71730c8b2b45 Mon Sep 17 00:00:00 2001
From: Peter Boyle <paboyle@ph.ed.ac.uk>
Date: Fri, 15 May 2015 14:41:19 +0100
Subject: [PATCH] Added su3 matrix benchmark.

---
 benchmarks/Grid_su3.cc | 139 +++++++++++++++++++++++++++++++++++++++++
 benchmarks/Makefile.am |   5 +-
 2 files changed, 143 insertions(+), 1 deletion(-)
 create mode 100644 benchmarks/Grid_su3.cc

diff --git a/benchmarks/Grid_su3.cc b/benchmarks/Grid_su3.cc
new file mode 100644
index 00000000..7f7bff90
--- /dev/null
+++ b/benchmarks/Grid_su3.cc
@@ -0,0 +1,139 @@
+#include <Grid.h>
+
+using namespace std;
+using namespace Grid;
+using namespace Grid::QCD;
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  int Nloop=1000;
+
+  std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
+  std::vector<int> mpi_layout  = GridDefaultMpi();
+
+  std::cout << "===================================================================================================="<<std::endl;
+  std::cout << "= Benchmarking SU3xSU3  x= x*y"<<std::endl;
+  std::cout << "===================================================================================================="<<std::endl;
+  std::cout << "  L  "<<"\t\t"<<"bytes"<<"\t\t"<<"GB/s\t\t GFlop/s"<<std::endl;
+  std::cout << "----------------------------------------------------------"<<std::endl;
+
+  for(int lat=2;lat<=24;lat+=2){
+
+      std::vector<int> latt_size  ({lat,lat,lat,lat});
+
+      GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
+      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedRandomDevice();
+
+      LatticeColourMatrix z(&Grid);// random(pRNG,z);
+      LatticeColourMatrix x(&Grid);// random(pRNG,x);
+      LatticeColourMatrix y(&Grid);// random(pRNG,y);
+
+      double start=usecond();
+      for(int i=0;i<Nloop;i++){
+	x=x*y;
+      }
+      double stop=usecond();
+      double time = (stop-start)/Nloop*1000.0;
+      
+      double bytes=3.0*lat*lat*lat*lat*Nc*Nc*sizeof(Complex);
+      double footprint=2.0*lat*lat*lat*lat*Nc*Nc*sizeof(Complex);
+      double flops=Nc*Nc*(6.0+8.0+8.0)*lat*lat*lat*lat;
+      std::cout<<std::setprecision(2) << lat<<"\t\t"<<footprint<<"\t\t"<<bytes/time<<"\t\t" << flops/time<<std::endl;
+
+    }
+
+
+  std::cout << "===================================================================================================="<<std::endl;
+  std::cout << "= Benchmarking SU3xSU3  z= x*y"<<std::endl;
+  std::cout << "===================================================================================================="<<std::endl;
+  std::cout << "  L  "<<"\t\t"<<"bytes"<<"\t\t"<<"GB/s\t\t GFlop/s"<<std::endl;
+  std::cout << "----------------------------------------------------------"<<std::endl;
+
+  for(int lat=2;lat<=24;lat+=2){
+
+      std::vector<int> latt_size  ({lat,lat,lat,lat});
+
+      GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
+      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedRandomDevice();
+
+      LatticeColourMatrix z(&Grid); //random(pRNG,z);
+      LatticeColourMatrix x(&Grid); //random(pRNG,x);
+      LatticeColourMatrix y(&Grid); //random(pRNG,y);
+
+      double start=usecond();
+      for(int i=0;i<Nloop;i++){
+	z=x*y;
+      }
+      double stop=usecond();
+      double time = (stop-start)/Nloop*1000.0;
+      
+      double bytes=3*lat*lat*lat*lat*Nc*Nc*sizeof(Complex);
+      double flops=Nc*Nc*(6+8+8)*lat*lat*lat*lat;
+      std::cout<<std::setprecision(2) << lat<<"\t\t"<<bytes<<"\t\t"<<bytes/time<<"\t\t" << flops/time<<std::endl;
+
+    }
+
+  std::cout << "===================================================================================================="<<std::endl;
+  std::cout << "= Benchmarking SU3xSU3  mult(z,x,y)"<<std::endl;
+  std::cout << "===================================================================================================="<<std::endl;
+  std::cout << "  L  "<<"\t\t"<<"bytes"<<"\t\t"<<"GB/s\t\t GFlop/s"<<std::endl;
+  std::cout << "----------------------------------------------------------"<<std::endl;
+
+  for(int lat=2;lat<=24;lat+=2){
+
+      std::vector<int> latt_size  ({lat,lat,lat,lat});
+
+      GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
+      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedRandomDevice();
+
+      LatticeColourMatrix z(&Grid); //random(pRNG,z);
+      LatticeColourMatrix x(&Grid); //random(pRNG,x);
+      LatticeColourMatrix y(&Grid); //random(pRNG,y);
+
+      double start=usecond();
+      for(int i=0;i<Nloop;i++){
+	mult(z,x,y);
+      }
+      double stop=usecond();
+      double time = (stop-start)/Nloop*1000.0;
+      
+      double bytes=3*lat*lat*lat*lat*Nc*Nc*sizeof(Complex);
+      double flops=Nc*Nc*(6+8+8)*lat*lat*lat*lat;
+      std::cout<<std::setprecision(2) << lat<<"\t\t"<<bytes<<"\t\t"<<bytes/time<<"\t\t" << flops/time<<std::endl;
+
+    }
+
+  std::cout << "===================================================================================================="<<std::endl;
+  std::cout << "= Benchmarking SU3xSU3  mac(z,x,y)"<<std::endl;
+  std::cout << "===================================================================================================="<<std::endl;
+  std::cout << "  L  "<<"\t\t"<<"bytes"<<"\t\t"<<"GB/s\t\t GFlop/s"<<std::endl;
+  std::cout << "----------------------------------------------------------"<<std::endl;
+
+  for(int lat=2;lat<=24;lat+=2){
+
+      std::vector<int> latt_size  ({lat,lat,lat,lat});
+
+      GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
+      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedRandomDevice();
+
+      LatticeColourMatrix z(&Grid); //random(pRNG,z);
+      LatticeColourMatrix x(&Grid); //random(pRNG,x);
+      LatticeColourMatrix y(&Grid); //random(pRNG,y);
+
+      double start=usecond();
+      for(int i=0;i<Nloop;i++){
+	mac(z,x,y);
+      }
+      double stop=usecond();
+      double time = (stop-start)/Nloop*1000.0;
+      
+      double bytes=3*lat*lat*lat*lat*Nc*Nc*sizeof(Complex);
+      double flops=Nc*Nc*(8+8+8)*lat*lat*lat*lat;
+      std::cout<<std::setprecision(2) << lat<<"\t\t"<<bytes<<"\t\t"<<bytes/time<<"\t\t" << flops/time<<std::endl;
+
+    }
+
+  Grid_finalize();
+}
diff --git a/benchmarks/Makefile.am b/benchmarks/Makefile.am
index 76c72279..2fe40feb 100644
--- a/benchmarks/Makefile.am
+++ b/benchmarks/Makefile.am
@@ -5,7 +5,7 @@ AM_LDFLAGS = -L$(top_builddir)/lib
 #
 # Test code
 #
-bin_PROGRAMS = Grid_wilson Grid_comms Grid_memory_bandwidth
+bin_PROGRAMS = Grid_wilson Grid_comms Grid_memory_bandwidth Grid_su3
 
 
 Grid_wilson_SOURCES = Grid_wilson.cc
@@ -14,6 +14,9 @@ Grid_wilson_LDADD = -lGrid
 Grid_comms_SOURCES = Grid_comms.cc
 Grid_comms_LDADD = -lGrid
 
+Grid_su3_SOURCES = Grid_su3.cc
+Grid_su3_LDADD = -lGrid
+
 Grid_memory_bandwidth_SOURCES = Grid_memory_bandwidth.cc
 Grid_memory_bandwidth_LDADD = -lGrid