10TF/s on 32^3 x 64 on single node

Grid/communicator/Communicator_mpi3.cc

@@ -406,6 +406,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
 void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
 {
   acceleratorCopySynchronise();
+  StencilBarrier();// Synch shared memory on a single nodes
 
   int nreq=list.size();
 

Grid/qcd/action/fermion/WilsonCompressor.h

@@ -400,7 +400,6 @@ public:
     }
     this->face_table_computed=1;
     assert(this->u_comm_offset==this->_unified_buffer_size);
-    accelerator_barrier();
   }
 
 };

Grid/qcd/action/fermion/implementation/WilsonFermion5DImplementation.h

@@ -233,10 +233,10 @@ void WilsonFermion5D<Impl>::ImportGauge(const GaugeField &_Umu)
   GaugeField HUmu(_Umu.Grid());
   HUmu = _Umu*(-0.5);
   if ( Dirichlet ) {
-    std::cout << GridLogMessage << " Dirichlet BCs 5d " <<Block<<std::endl;
+    std::cout << GridLogDslash << " Dirichlet BCs 5d " <<Block<<std::endl;
     Coordinate GaugeBlock(Nd);
     for(int d=0;d<Nd;d++) GaugeBlock[d] = Block[d+1];
-    std::cout << GridLogMessage << " Dirichlet BCs 4d " <<GaugeBlock<<std::endl;
+    std::cout << GridLogDslash << " Dirichlet BCs 4d " <<GaugeBlock<<std::endl;
     DirichletFilter<GaugeField> Filter(GaugeBlock);
     Filter.applyFilter(HUmu);
   }
@@ -382,12 +382,14 @@ void WilsonFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOrder &lo,
                                          DoubledGaugeField & U,
                                          const FermionField &in, FermionField &out,int dag)
 {
-  DhopTotalTime-=usecond();
+  //  std::cout << GridLogDslash<<"Dhop internal"<<std::endl;
+  DhopTotalTime=-usecond();
   if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute )
     DhopInternalOverlappedComms(st,lo,U,in,out,dag);
   else 
     DhopInternalSerialComms(st,lo,U,in,out,dag);
   DhopTotalTime+=usecond();
+  //  std::cout << GridLogDslash<<"Dhop took"<<DhopTotalTime<<std::endl;
 }
 
 
@@ -404,53 +406,59 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
   /////////////////////////////
   // Start comms  // Gather intranode and extra node differentiated??
   /////////////////////////////
-  DhopFaceTime-=usecond();
+  DhopFaceTime=-usecond();
   st.HaloExchangeOptGather(in,compressor);
   DhopFaceTime+=usecond();
+  //  std::cout << GridLogDslash<< " Dhop Gather end "<< DhopFaceTime<<" us " <<std::endl;
 
-  DhopCommTime -=usecond();
+  DhopCommTime =-usecond();
   std::vector<std::vector<CommsRequest_t> > requests;
   st.CommunicateBegin(requests);
 
   /////////////////////////////
   // Overlap with comms
   /////////////////////////////
-  DhopFaceTime-=usecond();
+  DhopFaceTime=-usecond();
   st.CommsMergeSHM(compressor);// Could do this inside parallel region overlapped with comms
   DhopFaceTime+=usecond();
+  //  std::cout << GridLogDslash<< " Dhop Commsmerge end "<<DhopFaceTime<< " us "<<std::endl;
       
   /////////////////////////////
   // do the compute interior
   /////////////////////////////
   int Opt = WilsonKernelsStatic::Opt; // Why pass this. Kernels should know
-  DhopComputeTime-=usecond();
+  DhopComputeTime=-usecond();
   if (dag == DaggerYes) {
     Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out,1,0);
   } else {
     Kernels::DhopKernel   (Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out,1,0);
   }
   DhopComputeTime+=usecond();
+  //  std::cout << GridLogDslash<< " Dhop Compute 1 end "<< DhopComputeTime<<" us" <<std::endl;
 
   /////////////////////////////
   // Complete comms
   /////////////////////////////
   st.CommunicateComplete(requests);
   DhopCommTime   +=usecond();
+  //  std::cout << GridLogDslash<< " Dhop Comunicate end "<< DhopCommTime << " us" <<std::endl;
 
   /////////////////////////////
   // do the compute exterior
   /////////////////////////////
-  DhopFaceTime-=usecond();
+  DhopFaceTime=-usecond();
   st.CommsMerge(compressor);
   DhopFaceTime+=usecond();
+  //  std::cout << GridLogDslash<< " Dhop CommsMerge2 end "<<DhopFaceTime << " us "<<std::endl;
 
-  DhopComputeTime2-=usecond();
+  DhopComputeTime2=-usecond();
   if (dag == DaggerYes) {
     Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out,0,1);
   } else {
     Kernels::DhopKernel   (Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out,0,1);
   }
   DhopComputeTime2+=usecond();
+  //  std::cout << GridLogDslash<< " Dhop Ext end "<<DhopComputeTime2 <<"us  "<<std::endl;
 }
 
 
@@ -464,11 +472,13 @@ void WilsonFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st, LebesgueOr
 
   int LLs = in.Grid()->_rdimensions[0];
 
-  DhopCommTime-=usecond();
+  //  std::cout << GridLogDslash<< " Dhop Halo exchange begine " <<std::endl;
+  DhopCommTime=-usecond();
   st.HaloExchangeOpt(in,compressor);
   DhopCommTime+=usecond();
+  //  std::cout << GridLogDslash<< " Dhop Comms end "<<DhopCommTime<<" us"<<std::endl;
   
-  DhopComputeTime-=usecond();
+  DhopComputeTime=-usecond();
   int Opt = WilsonKernelsStatic::Opt;
   if (dag == DaggerYes) {
     Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out);
@@ -476,6 +486,7 @@ void WilsonFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st, LebesgueOr
     Kernels::DhopKernel(Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out);
   }
   DhopComputeTime+=usecond();
+  //  std::cout << GridLogDslash<< " Dhop Compute end "<<DhopComputeTime<<" us" <<std::endl;
 }
 
 

Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h

@@ -416,19 +416,6 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
 #undef LoopBody
 }
 
-#define KERNEL_CALL_TMP(A) \
-  const uint64_t    NN = Nsite*Ls;					\
-  auto U_p = & U_v[0];							\
-  auto in_p = & in_v[0];						\
-  auto out_p = & out_v[0];						\
-  auto st_p = st_v._entries_p;						\
-  auto st_perm = st_v._permute_type;					\
-  accelerator_forNB( ss, NN, Simd::Nsimd(), {				\
-      int sF = ss;							\
-      int sU = ss/Ls;							\
-      WilsonKernels<Impl>::A(st_perm,st_p,U_p,buf,sF,sU,in_p,out_p);	\
-    });									\
-  accelerator_barrier();
 
 #define KERNEL_CALLNB(A)						\
   const uint64_t    NN = Nsite*Ls;					\
@@ -448,8 +435,7 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
       int sF = ptr[ss];							\
       int sU = ss/Ls;							\
       WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,in_v,out_v);		\
-    });									\
-  accelerator_barrier();
+    });									
 
 #define ASM_CALL(A)							\
   thread_for( ss, Nsite, {						\
@@ -471,7 +457,7 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
    if( interior && exterior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSite); return;}
 #ifdef SYCL_HACK     
-     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL_TMP(HandDhopSiteSycl);    return; }
+     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteSycl);    return; }
 #else
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSite);    return;}
 #endif     

Grid/stencil/Stencil.h

@@ -359,6 +359,7 @@ public:
   ////////////////////////////////////////////////////////////////////////
   void CommunicateBegin(std::vector<std::vector<CommsRequest_t> > &reqs)
   {
+    accelerator_barrier();
     for(int i=0;i<Packets.size();i++){
       _grid->StencilSendToRecvFromBegin(MpiReqs,
 					Packets[i].send_buf,
@@ -371,32 +372,13 @@ public:
 
   void CommunicateComplete(std::vector<std::vector<CommsRequest_t> > &reqs)
   {
-    _grid->StencilSendToRecvFromComplete(MpiReqs,i);
+    _grid->StencilSendToRecvFromComplete(MpiReqs,0);
   }
   ////////////////////////////////////////////////////////////////////////
   // Blocking send and receive. Either sequential or parallel.
   ////////////////////////////////////////////////////////////////////////
   void Communicate(void)
   {
-    if ( CartesianCommunicator::CommunicatorPolicy == CartesianCommunicator::CommunicatorPolicySequential ){
-      /////////////////////////////////////////////////////////
-      // several way threaded on different communicators.
-      // Cannot combine with Dirichlet operators
-      // This scheme is needed on Intel Omnipath for best performance
-      // Deprecate once there are very few omnipath clusters
-      /////////////////////////////////////////////////////////
-      int nthreads = CartesianCommunicator::nCommThreads;
-      int old = GridThread::GetThreads();
-      GridThread::SetThreads(nthreads);
-      thread_for(i,Packets.size(),{
-	  _grid->StencilSendToRecvFrom(Packets[i].send_buf,
-				       Packets[i].to_rank,Packets[i].do_send,
-				       Packets[i].recv_buf,
-				       Packets[i].from_rank,Packets[i].do_recv,
-				       Packets[i].bytes,i);
-      });
-      GridThread::SetThreads(old);
-    } else { 
     /////////////////////////////////////////////////////////
     // Concurrent and non-threaded asynch calls to MPI
     /////////////////////////////////////////////////////////
@@ -404,7 +386,6 @@ public:
     this->CommunicateBegin(reqs);
     this->CommunicateComplete(reqs);
   }
-  }
 
   template<class compressor> void HaloExchange(const Lattice<vobj> &source,compressor &compress)
   {
@@ -483,7 +464,6 @@ public:
     face_table_computed=1;
     assert(u_comm_offset==_unified_buffer_size);
 
-    accelerator_barrier();
   }
 
   /////////////////////////

Grid/threads/Accelerator.cc

@@ -1,6 +1,7 @@
 #include <Grid/GridCore.h>
 
 NAMESPACE_BEGIN(Grid);
+int      world_rank; // Use to control world rank for print guarding
 int      acceleratorAbortOnGpuError=1;
 uint32_t accelerator_threads=2;
 uint32_t acceleratorThreads(void)       {return accelerator_threads;};
@@ -16,7 +17,7 @@ void     acceleratorThreads(uint32_t t) {accelerator_threads = t;};
 #ifdef GRID_CUDA
 cudaDeviceProp *gpu_props;
 cudaStream_t copyStream;
-cudaStream_t cpuStream;
+cudaStream_t computeStream;
 void acceleratorInit(void)
 {
   int nDevices = 1;
@@ -24,7 +25,8 @@ void acceleratorInit(void)
   gpu_props = new cudaDeviceProp[nDevices];
 
   char * localRankStr = NULL;
-  int rank = 0, world_rank=0; 
+  int rank = 0;
+  world_rank=0; 
   if ((localRankStr = getenv(ENV_RANK_OMPI   )) != NULL) { world_rank = atoi(localRankStr);}
   if ((localRankStr = getenv(ENV_RANK_MVAPICH)) != NULL) { world_rank = atoi(localRankStr);}
   if ((localRankStr = getenv(ENV_RANK_SLURM  )) != NULL) { world_rank = atoi(localRankStr);}
@@ -99,7 +101,7 @@ void acceleratorInit(void)
 
   cudaSetDevice(device);
   cudaStreamCreate(&copyStream);
-  cudaStreamCreate(&cpuStream);
+  cudaStreamCreate(&computeStream);
   const int len=64;
   char busid[len];
   if( rank == world_rank ) { 
@@ -114,7 +116,7 @@ void acceleratorInit(void)
 #ifdef GRID_HIP
 hipDeviceProp_t *gpu_props;
 hipStream_t copyStream;
-hipStream_t cpuStream;
+hipStream_t computeStream;
 void acceleratorInit(void)
 {
   int nDevices = 1;
@@ -122,7 +124,8 @@ void acceleratorInit(void)
   gpu_props = new hipDeviceProp_t[nDevices];
 
   char * localRankStr = NULL;
-  int rank = 0, world_rank=0; 
+  int rank = 0;
+  world_rank=0; 
   // We extract the local rank initialization using an environment variable
   if ((localRankStr = getenv(ENV_LOCAL_RANK_OMPI)) != NULL)
   {
@@ -183,7 +186,7 @@ void acceleratorInit(void)
 #endif
   hipSetDevice(device);
   hipStreamCreate(&copyStream);
-  hipStreamCreate(&cpuStream);
+  hipStreamCreate(&computeStream);
   const int len=64;
   char busid[len];
   if( rank == world_rank ) { 
@@ -210,7 +213,8 @@ void acceleratorInit(void)
 #endif
   
   char * localRankStr = NULL;
-  int rank = 0, world_rank=0; 
+  int rank = 0;
+  world_rank=0; 
 
   // We extract the local rank initialization using an environment variable
   if ((localRankStr = getenv(ENV_LOCAL_RANK_OMPI)) != NULL)

Grid/threads/Accelerator.h

@@ -107,7 +107,7 @@ void     acceleratorInit(void);
 
 extern int acceleratorAbortOnGpuError;
 extern cudaStream_t copyStream;
-extern cudaStream_t cpuStream;
+extern cudaStream_t computeStream;
 
 accelerator_inline int acceleratorSIMTlane(int Nsimd) {
 #ifdef GRID_SIMT
@@ -135,7 +135,7 @@ inline void cuda_mem(void)
     };									\
     dim3 cu_threads(nsimd,acceleratorThreads(),1);			\
     dim3 cu_blocks ((num1+nt-1)/nt,num2,1);				\
-    LambdaApply<<<cu_blocks,cu_threads,0,cpuStream>>>(num1,num2,nsimd,lambda);	\
+    LambdaApply<<<cu_blocks,cu_threads,0,computeStream>>>(num1,num2,nsimd,lambda);	\
   }
 
 #define accelerator_for6dNB(iter1, num1,				\
@@ -154,7 +154,7 @@ inline void cuda_mem(void)
     };									\
     dim3 cu_blocks (num1,num2,num3);					\
     dim3 cu_threads(num4,num5,num6);					\
-    Lambda6Apply<<<cu_blocks,cu_threads,0,cpuStream>>>(num1,num2,num3,num4,num5,num6,lambda); \
+    Lambda6Apply<<<cu_blocks,cu_threads,0,computeStream>>>(num1,num2,num3,num4,num5,num6,lambda); \
   }
 
 template<typename lambda>  __global__
@@ -190,7 +190,7 @@ void Lambda6Apply(uint64_t num1, uint64_t num2, uint64_t num3,
 
 #define accelerator_barrier(dummy)					\
   {									\
-    cudaDeviceSynchronize();						\
+    cudaStreamSynchronize(computeStream);				\
     cudaError err = cudaGetLastError();					\
     if ( cudaSuccess != err ) {						\
       printf("accelerator_barrier(): Cuda error %s \n",			\
@@ -340,7 +340,7 @@ NAMESPACE_BEGIN(Grid);
 #define accelerator_inline __host__ __device__ inline
 
 extern hipStream_t copyStream;
-extern hipStream_t cpuStream;
+extern hipStream_t computeStream;
 /*These routines define mapping from thread grid to loop & vector lane indexing */
 accelerator_inline int acceleratorSIMTlane(int Nsimd) {
 #ifdef GRID_SIMT
@@ -362,16 +362,15 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
     dim3 hip_blocks ((num1+nt-1)/nt,num2,1); \
     if(hip_threads.x * hip_threads.y * hip_threads.z <= 64){ \
       hipLaunchKernelGGL(LambdaApply64,hip_blocks,hip_threads,		\
-			 0,0/*cpuStream*/,				\
+   	                 0,computeStream,						\
 			 num1,num2,nsimd, lambda);			\
     } else { \
       hipLaunchKernelGGL(LambdaApply,hip_blocks,hip_threads,		\
-			 0,0/*cpuStream*/,				\
+			 0,computeStream,				\
 			 num1,num2,nsimd, lambda);			\
     } \
   }
 
-
 template<typename lambda>  __global__
 __launch_bounds__(64,1)
 void LambdaApply64(uint64_t numx, uint64_t numy, uint64_t numz, lambda Lambda)
@@ -400,7 +399,7 @@ void LambdaApply(uint64_t numx, uint64_t numy, uint64_t numz, lambda Lambda)
 
 #define accelerator_barrier(dummy)				\
   {								\
-    hipStreamSynchronize(cpuStream);					\
+    hipStreamSynchronize(computeStream);			\
     auto err = hipGetLastError();				\
     if ( err != hipSuccess ) {					\
       printf("After hipDeviceSynchronize() : HIP error %s \n", hipGetErrorString( err )); \
@@ -443,7 +442,7 @@ inline void acceleratorMemSet(void *base,int value,size_t bytes) { hipMemset(bas
 
 inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) // Asynch
 {
-  hipMemcpy(to,from,bytes, hipMemcpyDeviceToDevice);
+  hipMemcpyDtoDAsync(to,from,bytes, copyStream);
 }
 inline void acceleratorCopySynchronise(void) { hipStreamSynchronize(copyStream); };
 

HMC/Mobius2p1f_DD_RHMC_96I_mixedmshift.cc

@@ -332,9 +332,9 @@ int main(int argc, char **argv) {
   OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy> StrangePseudoFermionBdy(StrangeOpDir,StrangeOp,SFRp);
   OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy> StrangePseudoFermionLocal(StrangePauliVillarsOpDir,StrangeOpDir,SFRp);
   OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy> StrangePseudoFermionPVBdy(StrangePauliVillarsOp,StrangePauliVillarsOpDir,SFRp);
-  Level1.push_back(&StrangePseudoFermionBdy);
+  Level1.push_back(&StrangePseudoFermionBdy); // ok
   Level2.push_back(&StrangePseudoFermionLocal);
-  Level1.push_back(&StrangePseudoFermionPVBdy);
+  Level1.push_back(&StrangePseudoFermionPVBdy); //ok
 
   ////////////////////////////////////
   // up down action
@@ -436,6 +436,10 @@ int main(int argc, char **argv) {
 			   *Numerators[h],*Denominators[h],
 			   *NumeratorsF[h],*DenominatorsF[h],
 			   OFRp, 500) );
+      Bdys.push_back( new OneFlavourEvenOddRatioRationalMixedPrecPseudoFermionAction<FermionImplPolicy,FermionImplPolicyF>(
+			   *Numerators[h],*Denominators[h],
+			   *NumeratorsF[h],*DenominatorsF[h],
+			   OFRp, 500) );
 #else
       Bdys.push_back( new OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy>(*Numerators[h],*Denominators[h],OFRp));
       Bdys.push_back( new OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy>(*Numerators[h],*Denominators[h],OFRp));