Merge branch 'develop' of https://github.com/paboyle/Grid into develop

THe bye counts are out and should be doubled for SIMD directions

Grid/qcd/action/fermion/WilsonCompressor.h

@@ -117,19 +117,19 @@ public:
     typedef decltype(coalescedRead(*in))    sobj;
     typedef decltype(coalescedRead(*out0)) hsobj;
 
-    unsigned int Nsimd = vobj::Nsimd();
+    constexpr unsigned int Nsimd = vobj::Nsimd();
     unsigned int mask = Nsimd >> (type + 1);
     int lane = acceleratorSIMTlane(Nsimd);
     int j0 = lane &(~mask); // inner coor zero
     int j1 = lane |(mask) ; // inner coor one
-    const vobj *vp0 = &in[k];
+    const vobj *vp0 = &in[k];  // out0[j] = merge low bit of type from in[k] and in[m] 
-    const vobj *vp1 = &in[m];
+    const vobj *vp1 = &in[m];  // out1[j] = merge hi  bit of type from in[k] and in[m]
-    const vobj *vp = (lane&mask) ? vp1:vp0;
+    const vobj *vp = (lane&mask) ? vp1:vp0;// if my lane has high bit take vp1, low bit take vp0
-    auto sa = coalescedRead(*vp,j0);
+    auto sa = coalescedRead(*vp,j0); // lane to read for out 0, NB 50% read coalescing
-    auto sb = coalescedRead(*vp,j1);
+    auto sb = coalescedRead(*vp,j1); // lane to read for out 1
     hsobj psa, psb;
-    projector::Proj(psa,sa,mu,dag);
+    projector::Proj(psa,sa,mu,dag);  // spin project the result0
-    projector::Proj(psb,sb,mu,dag);
+    projector::Proj(psb,sb,mu,dag);  // spin project the result1
     coalescedWrite(out0[j],psa);
     coalescedWrite(out1[j],psb);
 #else

Grid/stencil/Stencil.h

@@ -80,11 +80,14 @@ void Gather_plane_simple_table (commVector<std::pair<int,int> >& table,const Lat
 ///////////////////////////////////////////////////////////////////
 template<class cobj,class vobj,class compressor>
 void Gather_plane_exchange_table(const Lattice<vobj> &rhs,
-				 commVector<cobj *> pointers,int dimension,int plane,int cbmask,compressor &compress,int type) __attribute__((noinline));
+				 commVector<cobj *> pointers,
+				 int dimension,int plane,
+				 int cbmask,compressor &compress,int type) __attribute__((noinline));
 
 template<class cobj,class vobj,class compressor>
-void Gather_plane_exchange_table(commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
+void Gather_plane_exchange_table(commVector<std::pair<int,int> >& table,
-				 Vector<cobj *> pointers,int dimension,int plane,int cbmask,
+				 const Lattice<vobj> &rhs,
+				 std::vector<cobj *> &pointers,int dimension,int plane,int cbmask,
 				 compressor &compress,int type)
 {
   assert( (table.size()&0x1)==0);
@@ -92,14 +95,15 @@ void Gather_plane_exchange_table(commVector<std::pair<int,int> >& table,const La
   int so  = plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane
 
   auto rhs_v = rhs.View(AcceleratorRead);
+  auto rhs_p = &rhs_v[0];
   auto p0=&pointers[0][0];
   auto p1=&pointers[1][0];
   auto tp=&table[0];
   accelerator_forNB(j, num, vobj::Nsimd(), {
-      compress.CompressExchange(p0,p1, &rhs_v[0], j,
+      compress.CompressExchange(p0,p1, rhs_p, j,
-			      so+tp[2*j  ].second,
+				so+tp[2*j  ].second,
-			      so+tp[2*j+1].second,
+				so+tp[2*j+1].second,
-			      type);
+				type);
   });
   rhs_v.ViewClose();
 }
@@ -230,8 +234,8 @@ public:
   };
   struct Merge {
     cobj * mpointer;
-    Vector<scalar_object *> rpointers;
+    //    std::vector<scalar_object *> rpointers;
-    Vector<cobj *> vpointers;
+    std::vector<cobj *> vpointers;
     Integer buffer_size;
     Integer type;
   };
@@ -406,6 +410,7 @@ public:
       comms_bytes+=bytes;
       shm_bytes  +=2*Packets[i].bytes-bytes;
     }
+    _grid->StencilBarrier();// Synch shared memory on a single nodes
   }
 
   void CommunicateComplete(std::vector<std::vector<CommsRequest_t> > &reqs)
@@ -420,7 +425,7 @@ public:
   ////////////////////////////////////////////////////////////////////////
   void Communicate(void)
   {
-    if ( CartesianCommunicator::CommunicatorPolicy == CartesianCommunicator::CommunicatorPolicySequential ){
+    if ( 0 ){
       thread_region {
 	// must be called in parallel region
 	int mythread  = thread_num();
@@ -569,7 +574,7 @@ public:
     d.buffer_size = buffer_size;
     dv.push_back(d);
   }
-  void AddMerge(cobj *merge_p,Vector<cobj *> &rpointers,Integer buffer_size,Integer type,std::vector<Merge> &mv) {
+  void AddMerge(cobj *merge_p,std::vector<cobj *> &rpointers,Integer buffer_size,Integer type,std::vector<Merge> &mv) {
     Merge m;
     m.type     = type;
     m.mpointer = merge_p;
@@ -582,6 +587,7 @@ public:
   }
   template<class decompressor>  void CommsMergeSHM(decompressor decompress) {
     mpi3synctime-=usecond();
+    accelerator_barrier();
     _grid->StencilBarrier();// Synch shared memory on a single nodes
     mpi3synctime+=usecond();
     shmmergetime-=usecond();
@@ -1114,8 +1120,8 @@ public:
     int bytes = (reduced_buffer_size*datum_bytes)/simd_layout;
     assert(bytes*simd_layout == reduced_buffer_size*datum_bytes);
 
-    Vector<cobj *> rpointers(maxl);
+    std::vector<cobj *> rpointers(maxl);
-    Vector<cobj *> spointers(maxl);
+    std::vector<cobj *> spointers(maxl);
 
     ///////////////////////////////////////////
     // Work out what to send where

Grid/threads/Accelerator.h

@@ -298,14 +298,14 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
      });	   			              \
     });
 
-#define accelerator_barrier(dummy) { printf(" theGridAccelerator::wait()\n");  theGridAccelerator->wait(); }
+#define accelerator_barrier(dummy) { theGridAccelerator->wait(); }
 
 inline void *acceleratorAllocShared(size_t bytes){ return malloc_shared(bytes,*theGridAccelerator);};
 inline void *acceleratorAllocDevice(size_t bytes){ return malloc_device(bytes,*theGridAccelerator);};
 inline void acceleratorFreeShared(void *ptr){free(ptr,*theGridAccelerator);};
 inline void acceleratorFreeDevice(void *ptr){free(ptr,*theGridAccelerator);};
 
-inline void acceleratorCopySynchronise(void) {  printf(" theCopyAccelerator::wait()\n"); theCopyAccelerator->wait(); }
+inline void acceleratorCopySynchronise(void) {  theCopyAccelerator->wait(); }
 inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  {  theCopyAccelerator->memcpy(to,from,bytes);}
 inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { theCopyAccelerator->memcpy(to,from,bytes); theCopyAccelerator->wait();}
 inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ theCopyAccelerator->memcpy(to,from,bytes); theCopyAccelerator->wait();}

benchmarks/Benchmark_halo.cc

@@ -0,0 +1,131 @@
+ /*************************************************************************************
+    Grid physics library, www.github.com/paboyle/Grid
+    Source file: ./benchmarks/Benchmark_dwf.cc
+    Copyright (C) 2015
+
+    Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+    Author: paboyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#include <Grid/Grid.h>
+#ifdef GRID_CUDA
+#define CUDA_PROFILE
+#endif
+
+#ifdef CUDA_PROFILE
+#include <cuda_profiler_api.h>
+#endif
+
+using namespace std;
+using namespace Grid;
+
+template<class d>
+struct scal {
+  d internal;
+};
+
+  Gamma::Algebra Gmu [] = {
+    Gamma::Algebra::GammaX,
+    Gamma::Algebra::GammaY,
+    Gamma::Algebra::GammaZ,
+    Gamma::Algebra::GammaT
+  };
+
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  Coordinate latt4= GridDefaultLatt();
+  Coordinate mpi  = GridDefaultMpi();
+  Coordinate simd = GridDefaultSimd(Nd,vComplexF::Nsimd());
+
+  GridLogLayout();
+
+  int Ls=16;
+  for(int i=0;i<argc;i++)
+    if(std::string(argv[i]) == "-Ls"){
+      std::stringstream ss(argv[i+1]); ss >> Ls;
+    }
+
+  
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(latt4,simd ,mpi);
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
+  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
+
+  std::cout << GridLogMessage << "Making s innermost grids"<<std::endl;
+  GridCartesian         * sUGrid   = SpaceTimeGrid::makeFourDimDWFGrid(GridDefaultLatt(),GridDefaultMpi());
+  GridRedBlackCartesian * sUrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(sUGrid);
+  GridCartesian         * sFGrid   = SpaceTimeGrid::makeFiveDimDWFGrid(Ls,UGrid);
+  GridRedBlackCartesian * sFrbGrid = SpaceTimeGrid::makeFiveDimDWFRedBlackGrid(Ls,UGrid);
+
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+
+  std::cout << GridLogMessage << "Initialising 4d RNG" << std::endl;
+  GridParallelRNG          RNG4(UGrid);  RNG4.SeedUniqueString(std::string("The 4D RNG"));
+  std::cout << GridLogMessage << "Initialising 5d RNG" << std::endl;
+  GridParallelRNG          RNG5(FGrid);  RNG5.SeedUniqueString(std::string("The 5D RNG"));
+  std::cout << GridLogMessage << "Initialised RNGs" << std::endl;
+
+  LatticeFermionF src   (FGrid); random(RNG5,src);
+  RealD N2 = 1.0/::sqrt(norm2(src));
+  src = src*N2;
+
+  std::cout << GridLogMessage << "Drawing gauge field" << std::endl;
+  LatticeGaugeFieldF Umu(UGrid);
+  SU<Nc>::HotConfiguration(RNG4,Umu);
+  std::cout << GridLogMessage << "Random gauge initialised " << std::endl;
+
+  RealD mass=0.1;
+  RealD M5  =1.8;
+
+  RealD NP = UGrid->_Nprocessors;
+  RealD NN = UGrid->NodeCount();
+
+  DomainWallFermionF Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+
+  const int ncall = 500;
+  std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
+  std::cout << GridLogMessage<< "* Benchmarking DomainWallFermionF::HaloGatherOpt         "<<std::endl;
+  std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
+  {
+    typename DomainWallFermionF::Compressor compressor(0);
+    FGrid->Barrier();
+    Dw.Stencil.HaloExchangeOptGather(src,compressor);
+    double t0=usecond();
+    for(int i=0;i<ncall;i++){
+      Dw.Stencil.HaloExchangeOptGather(src,compressor);
+    }
+    double t1=usecond();
+    FGrid->Barrier();
+
+    double bytes=0.0;
+    if(mpi[0]) bytes+=latt4[1]*latt4[2]*latt4[3];
+    if(mpi[1]) bytes+=latt4[0]*latt4[2]*latt4[3];
+    if(mpi[2]) bytes+=latt4[0]*latt4[1]*latt4[3];
+    if(mpi[3]) bytes+=latt4[0]*latt4[1]*latt4[2];
+    bytes = bytes * Ls * 8.* (24.+12.)* 2.0;
+
+    std::cout<<GridLogMessage << "Gather us /call =   "<< (t1-t0)/ncall<<std::endl;
+    std::cout<<GridLogMessage << "Gather MBs /call =   "<< bytes*ncall/(t1-t0)<<std::endl;
+
+  }
+
+  Grid_finalize();
+  exit(0);
+}

systems/PVC/benchmarks/run-2tile-mpi.sh

@@ -1,6 +1,7 @@
 #!/bin/bash
 ##SBATCH -p PVC-SPR-QZEH 
 ##SBATCH -p PVC-ICX-QZNW
+
 #SBATCH -p QZ1J-ICX-PVC
 
 source /nfs/site/home/paboylex/ATS/GridNew/Grid/systems/PVC-nightly/setup.sh
@@ -19,8 +20,14 @@ export SYCL_DEVICE_FILTER=gpu,level_zero
 export I_MPI_OFFLOAD_CELL=tile
 export EnableImplicitScaling=0
 export EnableWalkerPartition=0
+#export SYCL_PI_LEVEL_ZERO_DEVICE_SCOPE_EVENTS=1
+export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
+#export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE=0
 
-mpiexec -launcher ssh -n 1 -host localhost  ./wrap.sh ./Benchmark_dwf_fp32 --mpi 1.1.1.1 --grid 32.32.32.32 --accelerator-threads $NT --comms-sequential --shm-mpi 1 > 1tile.log
+mpiexec -launcher ssh -n 2 -host localhost  ./wrap4gpu.sh ./Benchmark_dwf_fp32 --mpi 1.1.1.2 --grid 32.32.32.64 --accelerator-threads $NT  --shm-mpi 1 > dw.2tile.1x2.log
+mpiexec -launcher ssh -n 2 -host localhost  ./wrap4gpu.sh ./Benchmark_dwf_fp32 --mpi 2.1.1.1 --grid 64.32.32.32 --accelerator-threads $NT  --shm-mpi 1 > dw.2tile.2x1.log
+
+mpiexec -launcher ssh -n 2 -host localhost  ./wrap4gpu.sh ./Benchmark_halo --mpi 1.1.1.2 --grid 32.32.32.64 --accelerator-threads $NT  --shm-mpi 1 > halo.2tile.1x2.log
+mpiexec -launcher ssh -n 2 -host localhost  ./wrap4gpu.sh ./Benchmark_halo --mpi 2.1.1.1 --grid 64.32.32.32 --accelerator-threads $NT  --shm-mpi 1 > halo.2tile.2x1.log
 
-mpiexec -launcher ssh -n 2 -host localhost  ./wrap.sh ./Benchmark_dwf_fp32 --mpi 2.1.1.1 --grid 64.32.32.32 --accelerator-threads $NT --comms-sequential --shm-mpi 1 > 2tile.log
 

systems/PVC/benchmarks/wrap.sh

@@ -7,8 +7,8 @@ echo Ranke $MPI_LOCALRANKID ZE_AFFINITY_MASK is $ZE_AFFINITY_MASK
 
 if [ $MPI_LOCALRANKID = "0" ] 
 then
-#  ~psteinbr/build_pti/ze_tracer -c $@
+#  ~psteinbr/build_pti/ze_tracer -h $@
-  onetrace --chrome-kernel-timeline $@
+  onetrace --chrome-device-timeline $@
 else
   $@
 fi