Adding a better controlled threading class, preparing to

force in deterministic reduction.

benchmarks/Grid_comms.cc

@@ -8,11 +8,8 @@ int main (int argc, char ** argv)
 {
   Grid_init(&argc,&argv);
 
-  std::vector<int> latt_size;
-  std::vector<int> simd_layout;
-  std::vector<int> mpi_layout;
-
-  GridParseLayout(argv,argc,latt_size,simd_layout,mpi_layout);
+  std::vector<int> simd_layout = GridDefaultSimd();
+  std::vector<int> mpi_layout  = GridDefaultMpi();
 
   int Nloop=10;
   int nmu=0;

benchmarks/Grid_memory_bandwidth.cc

@@ -8,17 +8,14 @@ int main (int argc, char ** argv)
 {
   Grid_init(&argc,&argv);
 
-  std::vector<int> tmp_latt_size;
-  std::vector<int> simd_layout;
-  std::vector<int> mpi_layout;
-
-  GridParseLayout(argv,argc,tmp_latt_size,simd_layout,mpi_layout);
-
   const int Nvec=8;
   typedef Lattice< iVector< vReal,Nvec> > LatticeVec;
 
   int Nloop=1000;
 
+  std::vector<int> simd_layout = GridDefaultSimd();
+  std::vector<int> mpi_layout  = GridDefaultMpi();
+
   std::cout << "===================================================================================================="<<std::endl;
   std::cout << "= Benchmarking fused AXPY bandwidth"<<std::endl;
   std::cout << "===================================================================================================="<<std::endl;

benchmarks/Grid_wilson.cc

@@ -20,13 +20,12 @@ int main (int argc, char ** argv)
 {
   Grid_init(&argc,&argv);
 
-  std::vector<int> latt_size;
-  std::vector<int> simd_layout;
-  std::vector<int> mpi_layout;
-
-  GridParseLayout(argv,argc,latt_size,simd_layout,mpi_layout);
 
+  std::vector<int> latt_size   = GridDefaultLatt();
+  std::vector<int> simd_layout = GridDefaultSimd();
+  std::vector<int> mpi_layout  = GridDefaultMpi();
   GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
+
   std::vector<int> seeds({1,2,3,4});
 
   GridParallelRNG          pRNG(&Grid);

lib/Grid.h

@@ -45,8 +45,11 @@
 
 #include <Grid_aligned_allocator.h>
 #include <Grid_simd.h>
-#include <Grid_math.h>
+#include <Grid_threads.h>
+
 #include <Grid_cartesian.h>
+
+#include <Grid_math.h>
 #include <Grid_lattice.h>
 #include <Grid_comparison.h>
 #include <Grid_cshift.h>
@@ -60,6 +63,7 @@ namespace Grid {
 
   void Grid_init(int *argc,char ***argv);
   void Grid_finalize(void);
+  // internal, controled with --handle
   void Grid_sa_signal_handler(int sig,siginfo_t *si,void * ptr);
   void Grid_debug_handler_init(void);
   void Grid_quiesce_nodes(void);
@@ -68,6 +72,11 @@ namespace Grid {
   // C++11 time facilities better?
   double usecond(void);
 
+  const std::vector<int> &GridDefaultSimd(void);
+  const std::vector<int> &GridDefaultLatt(void);
+  const std::vector<int> &GridDefaultMpi(void);
+  const int              &GridThreads(void)  ;
+  void                 GridSetThreads(int t) ;
 
   // Common parsing chores
   std::string GridCmdOptionPayload(char ** begin, char ** end, const std::string & option);
@@ -75,8 +84,8 @@ namespace Grid {
   void        GridParseIntVector(std::string &str,std::vector<int> & vec);
 
   void GridParseLayout(char **argv,int argc,
-		       std::vector<int> &simd,
 		       std::vector<int> &latt,
+		       std::vector<int> &simd,
 		       std::vector<int> &mpi);
 
 

lib/Grid_init.cc

@@ -1,5 +1,5 @@
 /****************************************************************************/
-/* PAB: Signal magic. Processor state dump is x86-64 specific               */
+/* pab: Signal magic. Processor state dump is x86-64 specific               */
 /****************************************************************************/
 
 #include <stdlib.h>
@@ -23,23 +23,25 @@
 
 namespace Grid {
 
-  void Grid_quiesce_nodes(void)
-  {
-#ifdef GRID_COMMS_MPI
-    int me;
-    MPI_Comm_rank(MPI_COMM_WORLD,&me);
-    if ( me ) { 
-      std::cout.setstate(std::ios::badbit);
-    }
-#endif
-  }
-  void Grid_unquiesce_nodes(void)
-  {
-#ifdef GRID_COMMS_MPI
-    std::cout.clear();
-#endif
-  }
+  //////////////////////////////////////////////////////
+  // Convenience functions to access stadard command line arg
+  // driven parallelism controls
+  //////////////////////////////////////////////////////
+  static std::vector<int> Grid_default_simd;
+  static std::vector<int> Grid_default_latt;
+  static std::vector<int> Grid_default_mpi;
 
+  int GridThread::_threads;
+
+
+  const std::vector<int> &GridDefaultSimd(void)     {return Grid_default_simd;};
+  const std::vector<int> &GridDefaultLatt(void)     {return Grid_default_latt;};
+  const std::vector<int> &GridDefaultMpi(void)      {return Grid_default_mpi;};
+
+
+  ////////////////////////////////////////////////////////////
+  // Command line parsing assist for stock controls
+  ////////////////////////////////////////////////////////////
 std::string GridCmdOptionPayload(char ** begin, char ** end, const std::string & option)
 {
   char ** itr = std::find(begin, end, option);
@@ -53,15 +55,6 @@ bool GridCmdOptionExists(char** begin, char** end, const std::string& option)
 {
   return std::find(begin, end, option) != end;
 }
-void Grid_init(int *argc,char ***argv)
-{
-#ifdef GRID_COMMS_MPI
-  MPI_Init(argc,argv);
-#endif
-  // Parse command line args.
-  Grid_quiesce_nodes();
-
-}
 
 void GridCmdOptionIntVector(std::string &str,std::vector<int> & vec)
 {
@@ -70,7 +63,7 @@ void GridCmdOptionIntVector(std::string &str,std::vector<int> & vec)
   int i;
   while (ss >> i){
     vec.push_back(i);
-    if (ss.peek() == ',')
+    if(std::ispunct(ss.peek()))
       ss.ignore();
   }    
   return;
@@ -94,38 +87,74 @@ void GridParseLayout(char **argv,int argc,
   simd=std::vector<int>({1,2,2,2});
 #endif
 
+  GridThread::SetMaxThreads();
 
   std::string arg;
   if( GridCmdOptionExists(argv,argv+argc,"--mpi") ){
     arg = GridCmdOptionPayload(argv,argv+argc,"--mpi");
     GridCmdOptionIntVector(arg,mpi);
-    std::cout<<"MPI  ";
-    for(int i=0;i<mpi.size();i++){
-      std::cout<<mpi[i]<<" ";
-    }
-    std::cout<<std::endl;
   }
   if( GridCmdOptionExists(argv,argv+argc,"--simd") ){
     arg= GridCmdOptionPayload(argv,argv+argc,"--simd");
     GridCmdOptionIntVector(arg,simd);
-    std::cout<<"SIMD ";
-    for(int i=0;i<simd.size();i++){
-      std::cout<<simd[i]<<" ";
-    }
-    std::cout<<std::endl;
   }
   if( GridCmdOptionExists(argv,argv+argc,"--grid") ){
     arg= GridCmdOptionPayload(argv,argv+argc,"--grid");
     GridCmdOptionIntVector(arg,latt);
-    std::cout<<"Grid ";
-    for(int i=0;i<latt.size();i++){
-      std::cout<<latt[i]<<" ";
   }
-    std::cout<<std::endl;
+  if( GridCmdOptionExists(argv,argv+argc,"--omp") ){
+    std::vector<int> ompthreads(0);
+    arg= GridCmdOptionPayload(argv,argv+argc,"--omp");
+    GridCmdOptionIntVector(arg,ompthreads);
+    assert(ompthreads.size()==1);
+    GridThread::SetThreads(ompthreads[0]);
+  }
+
+}
+
+  /////////////////////////////////////////////////////////
+  /////////////////////////////////////////////////////////
+void Grid_init(int *argc,char ***argv)
+{
+#ifdef GRID_COMMS_MPI
+  MPI_Init(argc,argv);
+#endif
+  // Parse command line args.
+
+  if( GridCmdOptionExists(*argv,*argv+*argc,"--debug-signals") ){
+    Grid_debug_handler_init();
+  }
+  if( !GridCmdOptionExists(*argv,*argv+*argc,"--debug-stdout") ){
+    Grid_quiesce_nodes();
+  }
+  GridParseLayout(*argv,*argc,
+		  Grid_default_latt,
+		  Grid_default_simd,
+		  Grid_default_mpi);
+
 }
 
 
+  ////////////////////////////////////////////////////////////
+  // Verbose limiter on MPI tasks
+  ////////////////////////////////////////////////////////////
+  void Grid_quiesce_nodes(void)
+  {
+#ifdef GRID_COMMS_MPI
+    int me;
+    MPI_Comm_rank(MPI_COMM_WORLD,&me);
+    if ( me ) { 
+      std::cout.setstate(std::ios::badbit);
     }
+#endif
+  }
+  void Grid_unquiesce_nodes(void)
+  {
+#ifdef GRID_COMMS_MPI
+    std::cout.clear();
+#endif
+  }
+
   
 void Grid_finalize(void)
 {
@@ -146,13 +175,13 @@ void * Grid_backtrace_buffer[_NBACKTRACE];
 void Grid_sa_signal_handler(int sig,siginfo_t *si,void * ptr)
 {
   printf("Caught signal %d\n",si->si_signo);
-  printf("  mem address %lx\n",(uint64_t)si->si_addr);
+  printf("  mem address %llx\n",(unsigned long long)si->si_addr);
   printf("         code %d\n",si->si_code);
 
 #ifdef __X86_64
     ucontext_t * uc= (ucontext_t *)ptr;
   struct sigcontext *sc = (struct sigcontext *)&uc->uc_mcontext;
-  printf("  instruction %llx\n",(uint64_t)sc->rip);
+  printf("  instruction %llx\n",(unsigned long long)sc->rip);
 #define REG(A)  printf("  %s %lx\n",#A,sc-> A);
   REG(rdi);
   REG(rsi);

lib/Grid_threads.h

@@ -0,0 +1,80 @@
+#ifndef GRID_THREADS_H
+#define GRID_THREADS_H
+
+#ifdef HAVE_OPENMP
+#include <omp.h>
+#define PARALLEL_FOR_LOOP _Pragma("omp parallel for")
+#define PARALLEL_NESTED_LOOP(n) _Pragma("omp parallel for collapse(" #n ")")
+#else
+#define PARALLEL_FOR_LOOP 
+#define PARALLEL_NESTED_LOOP(n) 
+#endif
+
+namespace Grid {
+
+  // Introduce a class to gain deterministic bit reproducible reduction.
+  // make static; perhaps just a namespace is required.
+
+class GridThread {
+ public:
+  static int _threads;
+
+  static void SetThreads(int thr) { 
+#ifdef HAVE_OPENMP
+    _threads = MIN(thr,omp_get_max_threads()) ;
+    omp_set_num_threads(_threads);
+#else 
+    _threads = 1;
+#endif
+  };
+  static void SetMaxThreads(void) { 
+#ifdef HAVE_OPENMP
+    _threads = omp_get_max_threads();
+    omp_set_num_threads(_threads);
+#else 
+    _threads = 1;
+#endif
+  };
+  static int GetThreads(void) { return _threads; };
+  static int SumArraySize(void) {return _threads;};
+
+  static void GetWork(int nwork, int me, int & mywork, int & myoff){
+    int basework = nwork/_threads;
+    int backfill = _threads-(nwork%_threads);
+    if ( me >= _threads ) { 
+      mywork = myoff = 0;
+    } else { 
+      mywork = (nwork+me)/_threads;
+      myoff  = basework * me;
+      if ( me > backfill ) 
+	myoff+= (me-backfill);
+    }
+    return;
+  };
+
+  static void GetWorkBarrier(int nwork, int &me, int & mywork, int & myoff){
+    me     = ThreadBarrier();
+    GetWork(nwork,me,mywork,myoff);
+  };
+
+  static int  ThreadBarrier(void) {
+#ifdef HAVE_OPENMP
+#pragma omp barrier
+    return omp_get_thread_num();
+#else
+    return 0;
+#endif
+  };
+  
+  template<class obj> static void ThreadSum( std::vector<obj> &sum_array,obj &val,int me){
+    sum_array[me] = val;
+    val=zero;
+    ThreadBarrier();
+    for(int i=0;i<_threads;i++) val+= sum_array[i];
+    ThreadBarrier();
+  };
+
+};
+
+}
+#endif

lib/cartesian/Grid_cartesian_base.h

@@ -14,7 +14,7 @@ namespace Grid{
   //    int              _processor;  // linear processor rank
   //    std::vector<int> _processor_coor;  // linear processor rank
   //////////////////////////////////////////////////////////////////////
-  class GridBase : public CartesianCommunicator {
+  class GridBase : public CartesianCommunicator , public GridThread {
 
 public:
 
@@ -23,6 +23,7 @@ public:
 
     GridBase(std::vector<int> & processor_grid) : CartesianCommunicator(processor_grid) {};
 
+
     // Physics Grid information.
     std::vector<int> _simd_layout;// Which dimensions get relayed out over simd lanes.
     std::vector<int> _fdimensions;// Global dimensions of array prior to cb removal

tests/Grid_cshift.cc

@@ -8,11 +8,10 @@ int main (int argc, char ** argv)
 {
   Grid_init(&argc,&argv);
 
-  std::vector<int> simd_layout;
-  std::vector<int> mpi_layout;
-  std::vector<int> latt_size;
+  std::vector<int> latt_size   = GridDefaultLatt();
+  std::vector<int> simd_layout = GridDefaultSimd();
+  std::vector<int> mpi_layout  = GridDefaultMpi();
 
-  GridParseLayout(argv,argc,latt_size,simd_layout,mpi_layout);
   GridCartesian        Fine(latt_size,simd_layout,mpi_layout);
 
   GridParallelRNG      FineRNG(&Fine);  FineRNG.SeedRandomDevice();

tests/Grid_gamma.cc

@@ -14,11 +14,9 @@ int main (int argc, char ** argv)
 {
   Grid_init(&argc,&argv);
 
-  std::vector<int> latt_size;
-  std::vector<int> simd_layout;
-  std::vector<int> mpi_layout;
-
-  GridParseLayout(argv,argc,latt_size,simd_layout,mpi_layout);
+  std::vector<int> latt_size   = GridDefaultLatt();
+  std::vector<int> simd_layout = GridDefaultSimd();
+  std::vector<int> mpi_layout  = GridDefaultMpi();
      
   GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
 

tests/Grid_main.cc

@@ -25,11 +25,10 @@ int main (int argc, char ** argv)
 {
   Grid_init(&argc,&argv);
 
-  std::vector<int> latt_size;
-  std::vector<int> simd_layout;
-  std::vector<int> mpi_layout;
+  std::vector<int> latt_size   = GridDefaultLatt();
+  std::vector<int> simd_layout = GridDefaultSimd();
+  std::vector<int> mpi_layout  = GridDefaultMpi();
 
-  GridParseLayout(argv,argc,latt_size,simd_layout,mpi_layout);
   latt_size.resize(4);
 
 #ifdef AVX512

tests/Grid_nersc_io.cc

@@ -10,12 +10,9 @@ int main (int argc, char ** argv)
 {
   Grid_init(&argc,&argv);
 
-  std::vector<int> tmp_latt_size;
-  std::vector<int> simd_layout;
-  std::vector<int> mpi_layout;
-
-  GridParseLayout(argv,argc,tmp_latt_size,simd_layout,mpi_layout);
 
+  std::vector<int> simd_layout = GridDefaultSimd();
+  std::vector<int> mpi_layout  = GridDefaultMpi();
   std::vector<int> latt_size  ({16,16,16,32});
   std::vector<int> clatt_size  ({4,4,4,8});
   int orthodir=3;

tests/Grid_simd.cc

@@ -106,11 +106,9 @@ int main (int argc, char ** argv)
 {
   Grid_init(&argc,&argv);
 
-  std::vector<int> latt_size;
-  std::vector<int> simd_layout;
-  std::vector<int> mpi_layout;
-
-  GridParseLayout(argv,argc,latt_size,simd_layout,mpi_layout);
+  std::vector<int> latt_size   = GridDefaultLatt();
+  std::vector<int> simd_layout = GridDefaultSimd();
+  std::vector<int> mpi_layout  = GridDefaultMpi();
     
   GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
   std::vector<int> seeds({1,2,3,4});

tests/Grid_stencil.cc

@@ -8,11 +8,10 @@ int main (int argc, char ** argv)
 {
   Grid_init(&argc,&argv);
 
-  std::vector<int> latt_size;
-  std::vector<int> simd_layout;
-  std::vector<int> mpi_layout;
 
-  GridParseLayout(argv,argc,latt_size,simd_layout,mpi_layout);
+  std::vector<int> latt_size   = GridDefaultLatt();
+  std::vector<int> simd_layout = GridDefaultSimd();
+  std::vector<int> mpi_layout  = GridDefaultMpi();
 
   double volume = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];