Perf on Aurora

Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h

@@ -74,7 +74,7 @@ public:
 
   void operator() (const Field &src, Field &psi){
 
-    psi=Zero();
+    //    psi=Zero();
     RealD cp, ssq,rsq;
     ssq=norm2(src);
     rsq=Tolerance*Tolerance*ssq;

Grid/algorithms/multigrid/Aggregates.h

@@ -30,6 +30,8 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 /*  END LEGAL */
 #pragma once
 
+#include <Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h>
+
 NAMESPACE_BEGIN(Grid);
 
 inline RealD AggregatePowerLaw(RealD x)
@@ -124,6 +126,53 @@ public:
     }
   }
 
+  virtual void CreateSubspaceGCR(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &DiracOp,int nn=nbasis)
+  {
+    RealD scale;
+
+    TrivialPrecon<FineField> simple_fine;
+    PrecGeneralisedConjugateResidualNonHermitian<FineField> GCR(0.001,30,DiracOp,simple_fine,12,12);
+    FineField noise(FineGrid);
+    FineField src(FineGrid);
+    FineField guess(FineGrid);
+    FineField Mn(FineGrid);
+
+    for(int b=0;b<nn;b++){
+      
+      subspace[b] = Zero();
+      gaussian(RNG,noise);
+      scale = std::pow(norm2(noise),-0.5); 
+      noise=noise*scale;
+      
+      DiracOp.Op(noise,Mn); std::cout<<GridLogMessage << "noise   ["<<b<<"] <n|Op|n> "<<innerProduct(noise,Mn)<<std::endl;
+
+      for(int i=0;i<3;i++){
+	//  void operator() (const Field &src, Field &psi){
+#if 1
+	std::cout << GridLogMessage << " inverting on noise "<<std::endl;
+	src = noise;
+	guess=Zero();
+	GCR(src,guess);
+	subspace[b] = guess;
+#else
+	std::cout << GridLogMessage << " inverting on zero "<<std::endl;
+	src=Zero();
+	guess = noise;
+	GCR(src,guess);
+	subspace[b] = guess;
+#endif
+	noise = subspace[b];
+	scale = std::pow(norm2(noise),-0.5); 
+	noise=noise*scale;
+
+      }
+
+      DiracOp.Op(noise,Mn); std::cout<<GridLogMessage << "filtered["<<b<<"] <f|Op|f> "<<innerProduct(noise,Mn)<<std::endl;
+      subspace[b]   = noise;
+
+    }
+  }
+
   ////////////////////////////////////////////////////////////////////////////////////////////////
   // World of possibilities here. But have tried quite a lot of experiments (250+ jobs run on Summit)
   // and this is the best I found
@@ -160,14 +209,21 @@ public:
 
     int b =0;
     {
+      ComplexD ip;
       // Filter
       Chebyshev<FineField> Cheb(lo,hi,orderfilter);
       Cheb(hermop,noise,Mn);
       // normalise
       scale = std::pow(norm2(Mn),-0.5); 	Mn=Mn*scale;
       subspace[b]   = Mn;
-      hermop.Op(Mn,tmp); 
-      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+
+      hermop.Op(Mn,tmp);
+      ip= innerProduct(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|Op|n> "<<norm2(tmp)<<" "<<ip<<std::endl;
+
+      hermop.AdjOp(Mn,tmp); 
+      ip = innerProduct(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|AdjOp|n> "<<norm2(tmp)<<" "<<ip<<std::endl;
       b++;
     }
 
@@ -213,8 +269,18 @@ public:
 	  Mn=*Tnp;
 	  scale = std::pow(norm2(Mn),-0.5);         Mn=Mn*scale;
 	  subspace[b] = Mn;
-	  hermop.Op(Mn,tmp); 
-	  std::cout<<GridLogMessage << n<<" filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+
+
+	  ComplexD ip;
+
+	  hermop.Op(Mn,tmp);
+	  ip= innerProduct(Mn,tmp); 
+	  std::cout<<GridLogMessage << "filt ["<<b<<"] <n|Op|n> "<<norm2(tmp)<<" "<<ip<<std::endl;
+
+	  hermop.AdjOp(Mn,tmp); 
+	  ip = innerProduct(Mn,tmp); 
+	  std::cout<<GridLogMessage << "filt ["<<b<<"] <n|AdjOp|n> "<<norm2(tmp)<<" "<<ip<<std::endl;
+	  
 	  b++;
 	}
 

systems/Aurora/benchmarks/bench16.pbs

@@ -0,0 +1,80 @@
+#!/bin/bash
+
+##PBS -q LatticeQCD_aesp_CNDA
+#PBS -q debug-scaling
+##PBS -q prod
+#PBS -l select=16
+#PBS -l walltime=00:20:00
+#PBS -A LatticeQCD_aesp_CNDA
+
+cd $PBS_O_WORKDIR
+
+source ../sourceme.sh
+
+cp $PBS_NODEFILE nodefile
+
+export OMP_NUM_THREADS=4
+export MPICH_OFI_NIC_POLICY=GPU
+
+#export MPIR_CVAR_CH4_OFI_ENABLE_GPU_PIPELINE=1
+#unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE
+#unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE
+#unset MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE=0
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE=0
+#export MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST=1
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_BUFFER_SZ=1048576
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_THRESHOLD=131072
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_NUM_BUFFERS_PER_CHUNK=16
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_MAX_NUM_BUFFERS=16
+
+#
+# Local vol 16.16.16.32
+#
+
+VOL 128.64.128.96
+MPI 4.4.4.3
+NPROC 192
+mpiexec -np 192 -ppn 12 -envall ./gpu_tile.sh ./Benchmark_dwf_fp32 --mpi 4.4.4.3 --grid 128.64.128.96 --shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap
+
+
+LX=32
+LY=16
+LZ=32
+LT=32
+
+NX=2
+NY=2
+NZ=4
+NT=1
+
+GX=2
+GY=2
+GZ=1
+GT=3
+
+PX=$((NX * GX ))
+PY=$((NY * GY ))
+PZ=$((NZ * GZ ))
+PT=$((NT * GT ))
+
+VX=$((PX * LX ))
+VY=$((PY * LY ))
+VZ=$((PZ * LZ ))
+VT=$((PT * LT ))
+
+NP=$((PX*PY*PZ*PT))
+VOL=${VX}.${VY}.${VZ}.${VT}
+AT=8
+MPI=${PX}.${PY}.${PZ}.${PT}
+
+CMD="mpiexec -np $NP -ppn 12  -envall \
+	     ./gpu_tile.sh ./Benchmark_dwf_fp32 --mpi $MPI --grid $VOL \
+		--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads $AT --comms-overlap "
+
+echo VOL $VOL
+echo MPI $MPI
+echo NPROC $NP
+echo $CMD
+$CMD
+

systems/Aurora/benchmarks/gpu_tile.sh

@@ -30,8 +30,8 @@ echo "rank $PALS_RANKID ; local rank $PALS_LOCAL_RANKID ; ZE_AFFINITY_MASK=$ZE_A
 
 if [ $PALS_RANKID = "0" ]
 then
-    numactl -p $NUMAP -N $NUMAP unitrace --chrome-kernel-logging --chrome-mpi-logging --chrome-sycl-logging --demangle "$@"
-#    numactl -p $NUMAP -N $NUMAP  "$@"
+#    numactl -p $NUMAP -N $NUMAP unitrace --chrome-kernel-logging --chrome-mpi-logging --chrome-sycl-logging --demangle "$@"
+    numactl -p $NUMAP -N $NUMAP  "$@"
 else 
     numactl -p $NUMAP -N $NUMAP  "$@"
 fi

tests/debug/Test_general_coarse_pvdagm.cc

@@ -154,6 +154,8 @@ public:
     //    std::cout<<GridLogMessage << "Calling PreSmoother input residual "<<norm2(in) <<std::endl;
     double t;
     // Fine Smoother
+    //    out = in;
+    out = Zero();
     t=-usecond();
     _PreSmoother(in,out);
     t+=usecond();
@@ -172,6 +174,7 @@ public:
 
     // Coarse correction
     t=-usecond();
+    Csol = Zero();
     _CoarseSolve(Csrc,Csol);
     //Csol=Zero();
     t+=usecond();
@@ -191,6 +194,8 @@ public:
 
     // Fine Smoother
     t=-usecond();
+    //    vec2=vec1;
+    vec2=Zero();
     _PostSmoother(vec1,vec2);
     t+=usecond();
     std::cout<<GridLogMessage << "PostSmoother took "<< t/1000.0<< "ms" <<std::endl;
@@ -215,7 +220,8 @@ int main (int argc, char ** argv)
   // Construct a coarsened grid
   Coordinate clatt = GridDefaultLatt();
   for(int d=0;d<clatt.size();d++){
-    clatt[d] = clatt[d]/4;
+    clatt[d] = clatt[d]/2;
+    //    clatt[d] = clatt[d]/4;
   }
   GridCartesian *Coarse4d =  SpaceTimeGrid::makeFourDimGrid(clatt, GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());;
   GridCartesian *Coarse5d =  SpaceTimeGrid::makeFiveDimGrid(1,Coarse4d);
@@ -244,7 +250,7 @@ int main (int argc, char ** argv)
   DomainWallFermionD Ddwf(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
   DomainWallFermionD Dpv(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,1.0,M5);
 
-  const int nbasis = 8;
+  const int nbasis = 20;
   const int cb = 0 ;
   LatticeFermion prom(FGrid);
 
@@ -260,7 +266,25 @@ int main (int argc, char ** argv)
   typedef PVdagMLinearOperator<DomainWallFermionD,LatticeFermionD> PVdagM_t;
   typedef ShiftedPVdagMLinearOperator<DomainWallFermionD,LatticeFermionD> ShiftedPVdagM_t;
   PVdagM_t PVdagM(Ddwf,Dpv);
-  ShiftedPVdagM_t ShiftedPVdagM(2.0,Ddwf,Dpv);
+  //  ShiftedPVdagM_t ShiftedPVdagM(2.0,Ddwf,Dpv); // 355
+  //  ShiftedPVdagM_t ShiftedPVdagM(1.0,Ddwf,Dpv); // 246
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.5,Ddwf,Dpv); // 183
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.25,Ddwf,Dpv); // 145
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.1,Ddwf,Dpv); // 134
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.1,Ddwf,Dpv); // 127 -- NULL space via inverse iteration
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.1,Ddwf,Dpv); // 57 -- NULL space via inverse iteration; 3 iterations
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.25,Ddwf,Dpv); // 57 , tighter inversion
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.25,Ddwf,Dpv); // nbasis 20 -- 49 iters
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.25,Ddwf,Dpv); // nbasis 20 -- 70 iters; asymmetric 
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.25,Ddwf,Dpv); // 58; Loosen coarse, tighten fine
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.1,Ddwf,Dpv); // 56 ... 
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.1,Ddwf,Dpv); // 51 ...  with 24 vecs
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.1,Ddwf,Dpv); // 31 ...  with 24 vecs and 2^4 blocking
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.1,Ddwf,Dpv); // 43 ...  with 16 vecs and 2^4 blocking, sloppier
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.1,Ddwf,Dpv); // 35  ...  with 20 vecs and 2^4 blocking
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.1,Ddwf,Dpv); // 35  ...  with 20 vecs and 2^4 blocking, looser coarse
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.1,Ddwf,Dpv); // 64  ...  with 20 vecs, Christoph setup, and 2^4 blocking, looser coarse
+  ShiftedPVdagM_t ShiftedPVdagM(0.01,Ddwf,Dpv); // 
 
 
   // Run power method on HOA??
@@ -269,6 +293,7 @@ int main (int argc, char ** argv)
   // Warning: This routine calls PVdagM.Op, not PVdagM.HermOp
   typedef Aggregation<vSpinColourVector,vTComplex,nbasis> Subspace;
   Subspace AggregatesPD(Coarse5d,FGrid,cb);
+  /*
   AggregatesPD.CreateSubspaceChebyshev(RNG5,
 				       PVdagM,
 				       nbasis,
@@ -278,6 +303,10 @@ int main (int argc, char ** argv)
 				       200,
 				       200,
 				       0.0);
+  */
+  AggregatesPD.CreateSubspaceGCR(RNG5,
+				 PVdagM,
+				 nbasis);
   
   LittleDiracOperator LittleDiracOpPV(geom,FGrid,Coarse5d);
   LittleDiracOpPV.CoarsenOperator(PVdagM,AggregatesPD);
@@ -334,12 +363,13 @@ int main (int argc, char ** argv)
   ///////////////////////////////////////
 
   std::cout<<GridLogMessage<<"******************* "<<std::endl;
-  std::cout<<GridLogMessage<<" Coarse Grid Solve "<<std::endl;
+  std::cout<<GridLogMessage<<" Coarse Grid Solve -- Level 3 "<<std::endl;
   std::cout<<GridLogMessage<<"******************* "<<std::endl;
   TrivialPrecon<CoarseVector> simple;
   NonHermitianLinearOperator<LittleDiracOperator,CoarseVector> LinOpCoarse(LittleDiracOpPV);
-  PrecGeneralisedConjugateResidualNonHermitian<CoarseVector>  L2PGCR(1.0e-8, 100, LinOpCoarse,simple,10,10); 
-  L2PGCR.Level(2);
+  //  PrecGeneralisedConjugateResidualNonHermitian<CoarseVector>  L2PGCR(1.0e-4, 100, LinOpCoarse,simple,10,10); 
+  PrecGeneralisedConjugateResidualNonHermitian<CoarseVector>  L2PGCR(3.0e-2, 100, LinOpCoarse,simple,10,10); 
+  L2PGCR.Level(3);
   c_res=Zero();
   L2PGCR(c_src,c_res);
 
@@ -347,11 +377,12 @@ int main (int argc, char ** argv)
   // Fine grid smoother
   ////////////////////////////////////////
   std::cout<<GridLogMessage<<"******************* "<<std::endl;
-  std::cout<<GridLogMessage<<" Fine Grid Smoother "<<std::endl;
+  std::cout<<GridLogMessage<<" Fine Grid Smoother -- Level 2 "<<std::endl;
   std::cout<<GridLogMessage<<"******************* "<<std::endl;
   TrivialPrecon<LatticeFermionD> simple_fine;
   //  NonHermitianLinearOperator<PVdagM_t,LatticeFermionD> LinOpSmooth(PVdagM);
-  PrecGeneralisedConjugateResidualNonHermitian<LatticeFermionD> SmootherGCR(0.01,10,ShiftedPVdagM,simple_fine,4,4);
+  PrecGeneralisedConjugateResidualNonHermitian<LatticeFermionD> SmootherGCR(0.01,1,ShiftedPVdagM,simple_fine,16,16);
+  SmootherGCR.Level(2);
   
   LatticeFermionD f_src(FGrid);
   LatticeFermionD f_res(FGrid);
@@ -364,12 +395,12 @@ int main (int argc, char ** argv)
 
   TwoLevelMG TwoLevelPrecon(AggregatesPD,
 			    PVdagM,
-			    SmootherGCR,
+			    simple_fine,
 			    SmootherGCR,
 			    LinOpCoarse,
 			    L2PGCR);
   
-  PrecGeneralisedConjugateResidualNonHermitian<LatticeFermion> L1PGCR(1.0e-8,1000,PVdagM,TwoLevelPrecon,8,8);
+  PrecGeneralisedConjugateResidualNonHermitian<LatticeFermion> L1PGCR(1.0e-8,1000,PVdagM,TwoLevelPrecon,16,16);
   L1PGCR.Level(1);
 
   f_res=Zero();