From 81441e98f4feded0c3b09735713b2cb49ae18841 Mon Sep 17 00:00:00 2001
From: Peter Boyle <Peter Boyle>
Date: Wed, 16 Sep 2020 03:35:03 +0100
Subject: [PATCH] HIP runs sensible

---
 Grid/lattice/Lattice_arith.h    |  6 ++---
 Grid/threads/Accelerator.cc     |  2 +-
 benchmarks/Benchmark_su3.cc     | 42 +++++++++++++++++++++++++++++----
 benchmarks/Benchmark_su3_gpu.cc |  3 ++-
 4 files changed, 44 insertions(+), 9 deletions(-)

diff --git a/Grid/lattice/Lattice_arith.h b/Grid/lattice/Lattice_arith.h
index a3ae1f28..3c269c58 100644
--- a/Grid/lattice/Lattice_arith.h
+++ b/Grid/lattice/Lattice_arith.h
@@ -60,9 +60,9 @@ void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
   autoView( lhs_v , lhs, AcceleratorRead);
   autoView( rhs_v , rhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
-    decltype(coalescedRead(obj1())) tmp;
     auto lhs_t=lhs_v(ss);
     auto rhs_t=rhs_v(ss);
+    auto tmp  =ret_v(ss);
     mac(&tmp,&lhs_t,&rhs_t);
     coalescedWrite(ret_v[ss],tmp);
   });
@@ -124,7 +124,7 @@ void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
   autoView( ret_v , ret, AcceleratorWrite);
   autoView( lhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
-    decltype(coalescedRead(obj1())) tmp;
+    auto tmp  =ret_v(ss);
     auto lhs_t=lhs_v(ss);
     mac(&tmp,&lhs_t,&rhs);
     coalescedWrite(ret_v[ss],tmp);
@@ -182,7 +182,7 @@ void mac(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
   autoView( ret_v , ret, AcceleratorWrite);
   autoView( rhs_v , lhs, AcceleratorRead);
   accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
-    decltype(coalescedRead(obj1())) tmp;
+    auto tmp  =ret_v(ss);
     auto rhs_t=rhs_v(ss);
     mac(&tmp,&lhs,&rhs_t);
     coalescedWrite(ret_v[ss],tmp);
diff --git a/Grid/threads/Accelerator.cc b/Grid/threads/Accelerator.cc
index 35e10d31..2134d158 100644
--- a/Grid/threads/Accelerator.cc
+++ b/Grid/threads/Accelerator.cc
@@ -127,7 +127,7 @@ void acceleratorInit(void)
       printf("AcceleratorHipInit: ========================\n");
       printf("AcceleratorHipInit: Device identifier: %s\n", prop.name);
 
-      GPU_PROP_FMT(totalGlobalMem,"%lld");
+      GPU_PROP_FMT(totalGlobalMem,"%lu");
       //      GPU_PROP(managedMemory);
       GPU_PROP(isMultiGpuBoard);
       GPU_PROP(warpSize);
diff --git a/benchmarks/Benchmark_su3.cc b/benchmarks/Benchmark_su3.cc
index d24a3e25..d094da0c 100644
--- a/benchmarks/Benchmark_su3.cc
+++ b/benchmarks/Benchmark_su3.cc
@@ -36,12 +36,12 @@ int main (int argc, char ** argv)
 {
   Grid_init(&argc,&argv);
 
-#define LMAX (48)
+#define LMAX (40)
 #define LMIN (8)
 #define LADD (8)
 
-  int64_t Nwarm=50;
-  int64_t Nloop=500;
+  int64_t Nwarm=10;
+  int64_t Nloop=100;
 
   Coordinate simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
   Coordinate mpi_layout  = GridDefaultMpi();
@@ -118,6 +118,41 @@ int main (int argc, char ** argv)
 
     }
 
+
+  std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
+  std::cout<<GridLogMessage << "= Benchmarking SU3xSU3  z=z+ x*y"<<std::endl;
+  std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
+  std::cout<<GridLogMessage << "  L  "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s\t\t GFlop/s"<<std::endl;
+  std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
+
+  for(int lat=LMIN;lat<=LMAX;lat+=LADD){
+
+      Coordinate latt_size  ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
+      int64_t vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
+
+      GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
+      GridParallelRNG          pRNG(&Grid);      pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9}));
+
+      LatticeColourMatrix z(&Grid); random(pRNG,z);
+      LatticeColourMatrix x(&Grid); random(pRNG,x);
+      LatticeColourMatrix y(&Grid); random(pRNG,y);
+
+      for(int64_t i=0;i<Nwarm;i++){
+	z=z+x*y;
+      }
+      double start=usecond();
+      for(int64_t i=0;i<Nloop;i++){
+	z=z+x*y;
+      }
+      double stop=usecond();
+      double time = (stop-start)/Nloop*1000.0;
+      
+      double bytes=4*vol*Nc*Nc*sizeof(Complex);
+      double flops=Nc*Nc*(6+8+8)*vol;
+      std::cout<<GridLogMessage<<std::setprecision(3) << lat<<"\t\t"<<bytes<<"    \t\t"<<bytes/time<<"\t\t" << flops/time<<std::endl;
+
+    }
+
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
   std::cout<<GridLogMessage << "= Benchmarking SU3xSU3  mult(z,x,y)"<<std::endl;
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
@@ -143,7 +178,6 @@ int main (int argc, char ** argv)
       double start=usecond();
       for(int64_t i=0;i<Nloop;i++){
 	mult(z,x,y);
-	//	mac(z,x,y);
       }
       double stop=usecond();
       double time = (stop-start)/Nloop*1000.0;
diff --git a/benchmarks/Benchmark_su3_gpu.cc b/benchmarks/Benchmark_su3_gpu.cc
index 93592ca2..f7412054 100644
--- a/benchmarks/Benchmark_su3_gpu.cc
+++ b/benchmarks/Benchmark_su3_gpu.cc
@@ -187,7 +187,8 @@ int main (int argc, char ** argv)
 	  auto xx = coalescedRead(x_v[ss]);
 	  auto yy = coalescedRead(y_v[ss]);
 	  auto zz = coalescedRead(z_v[ss]);
-	  zz = zz+xx*yy;
+	  //zz = zz+xx*yy;
+	  mac(&zz,&xx,&yy);
 	  coalescedWrite(z_v[ss],zz);
         });
       }