Huayno accelerate zero mass particles and assorted fixes

src/amuse/community/huayno/Makefile

@@ -20,13 +20,6 @@ CXXFLAGS +=
 LIBS += -lm
 INCLUDE =
 
-
-CODE_GENERATOR ?= $(AMUSE_DIR)/build.py
-
-# set OpenCL and STDCL libraries
-OPENCL = /home/inti/libraries/cuda
-STDCL = /home/inti/libraries/coprthr
-
 #---------------------------------------------
 
 OBJS = interface.o
@@ -40,17 +33,14 @@ OPENMP_BUILDDIR = build_mp
 AM_LIBS = -L$(AMUSE_DIR)/lib/amuse_mpi -lamuse_mpi
 AM_CFLAGS = -I$(AMUSE_DIR)/lib/amuse_mpi
 
-CL_LIBS?=-L$(OPENCL)/lib/x86_64 -lOpenCL -lpthread -ldl -L$(STDCL)/lib -lstdcl
-CL_FLAGS?=-I$(OPENCL)/include -I$(STDCL)/include
-
 all:	huayno_worker
 
 clean:
 	rm -f *.o *.bck *.pyc *.clh worker_code.cc worker_code.h
 	rm -Rf $(BUILDDIR)
 	rm -Rf $(OPENCL_BUILDDIR)
 	rm -Rf $(OPENMP_BUILDDIR)
-	rm -f huayno_worker huayno_worker_cl huayno_worker_mp huayno_worker_sockets
+	rm -f huayno_worker huayno_worker_cl huayno_worker_mp
 	-make -C $(SRCDIR) clean
 
 distclean: clean
@@ -60,8 +50,7 @@ $(BUILDDIR)/Makefile: $(SRCDIR)/Makefile
 	cp $(SRCDIR)/Makefile $(BUILDDIR)/Makefile
 
 $(BUILDDIR)/libhuayno.a: $(BUILDDIR)/Makefile $(LIBFILES) src/Makefile
-	make -C $(BUILDDIR) all VPATH=../src 
-
+	make -C $(BUILDDIR) all VPATH=../src CFLAGS="$(CFLAGS)" CXXFLAGS="$(CXXFLAGS)"
 
 $(OPENMP_BUILDDIR)/Makefile: $(SRCDIR)/Makefile 
 	-mkdir $(OPENMP_BUILDDIR) 
@@ -76,8 +65,7 @@ $(OPENCL_BUILDDIR)/Makefile: $(SRCDIR)/Makefile
 
 $(OPENCL_BUILDDIR)/libhuayno_cl.a: $(OPENCL_BUILDDIR)/Makefile $(LIBFILES) src/Makefile
 	make -C $(OPENCL_BUILDDIR) libhuayno_cl.a VPATH=../src CFLAGS="$(CFLAGS)" CXXFLAGS="$(CXXFLAGS)" LIBS="$(LIBS)" INCLUDE="$(INCLUDE)"
-
-
+
 worker_code.cc: interface.py
 	$(CODE_GENERATOR) --type=c interface.py HuaynoInterface -o $@
 
@@ -92,32 +80,27 @@ huayno_worker_mp: CXXFLAGS   += $(OPENMP_CFLAGS)
 huayno_worker_mp: __init__.py worker_code.cc worker_code.h $(OPENMP_BUILDDIR)/libhuayno.a $(OPENMP_BUILDDIR)/interface.o 
 	$(MPICXX) $(SC_FLAGS) $(CXXFLAGS) $(INCLUDE)  $(LDFLAGS) worker_code.cc $(OPENMP_BUILDDIR)/interface.o $(OPENMP_BUILDDIR)/libhuayno.a -o $@ $(LIBS) $(SC_CLIBS)
 
-huayno_worker_mp.so: CFLAGS   += $(OPENMP_CFLAGS) 
-huayno_worker_mp.so: CXXFLAGS   += $(OPENMP_CFLAGS)	
+#~ huayno_worker_mp.so: CFLAGS   += $(OPENMP_CFLAGS) 
+#~ huayno_worker_mp.so: CXXFLAGS   += $(OPENMP_CFLAGS)	
 
-huayno_worker_cl: CFLAGS   += -fopenmp -DEVOLVE_OPENCL
-huayno_worker_cl: CXXFLAGS   += -fopenmp -DEVOLVE_OPENCL
-huayno_worker_cl: LIBS += $(CL_LIBS)  -lstdcl
-huayno_worker_cl: INCLUDE += $(CL_FLAGS)
+huayno_worker_cl: CFLAGS   += -DEVOLVE_OPENCL
+huayno_worker_cl: CXXFLAGS   += -DEVOLVE_OPENCL
+huayno_worker_cl: LIBS += $(CL_LIBS)
+huayno_worker_cl: INCLUDE += $(CL_FLAGS) -I.
 huayno_worker_cl: __init__.py worker_code.cc worker_code.h $(OPENCL_BUILDDIR)/libhuayno_cl.a $(OPENCL_BUILDDIR)/interface.o 
 	$(MPICXX) $(CXXFLAGS) $(INCLUDE) $(LDFLAGS) worker_code.cc $(OPENCL_BUILDDIR)/interface.o $(OPENCL_BUILDDIR)/libhuayno_cl.a -o $@ $(LIBS) $(SC_CLIBS)
 
-
 .cc.o: $<
 	$(MPICXX) -I$(SRCDIR)  $(SC_FLAGS) $(CXXFLAGS) $(INCLUDE) -c -o $@ $< 
 
 $(BUILDDIR)/%.o: %.c
-	$(MPICC) -I$(SRCDIR)  $(SC_FLAGS) $(CFLAGS) $(INCLUDE) -c -o $@ $< 
+	$(CC) -I$(SRCDIR)  $(SC_FLAGS) $(CFLAGS) $(INCLUDE) -c -o $@ $< 
 
 $(OPENMP_BUILDDIR)/%.o: %.c
-	$(MPICC) -I$(SRCDIR)  $(SC_FLAGS) $(CFLAGS) $(INCLUDE) -c -o $@ $< 
+	$(CC) -I$(SRCDIR)  $(SC_FLAGS) $(CFLAGS) $(INCLUDE) -c -o $@ $< 
 
-
 $(OPENCL_BUILDDIR)/%.o: %.c
-	$(MPICC) -I$(SRCDIR)  $(SC_FLAGS) $(CFLAGS) $(INCLUDE) -c -o $@ $< 
+	$(CC) -I$(SRCDIR)  $(SC_FLAGS) $(CFLAGS) $(INCLUDE) -c -o $@ $< 
 
 .c.o: $<
 	$(MPICC) -I$(SRCDIR)  $(SC_FLAGS) $(CFLAGS) $(INCLUDE) -c -o $@ $< 
-
-%.clh: %.cl
-	awk 'BEGIN{print "const char srcstr[]=" } {if(substr($$0,length($0))=="\\"){$$0=$$0"\\"};print "\""$$0"\\n\""} END{print ";"}' $< > $@

src/amuse/community/huayno/interface.c

@@ -31,16 +31,16 @@ int initialize_code()
   nmax=NMAX;
   err=LOOKUPSYMBOL(init_,)(&lookup, nmax*sizeof(struct particle)/sizeof(int));
   if(err != 0) return err;
-  mainsys.n=0;
+  mainsys=zerosys;
   mainsys.part=(struct particle*) malloc(nmax*sizeof(struct particle));
   if(mainsys.part == NULL) return -1;
-  mainsys.last=NULL;
-  dt_param=.03; 
+  dt_param=.03;
+  accel_zero_mass=1; 
   eps2=0.;
   inttype=8;
+  opencl_device_type=0;
   t_now=0.;
   dtime=0.;
-  init_code();
   // AMUSE STOPPING CONDITIONS SUPPORT
   set_support_for_condition(COLLISION_DETECTION);
   set_support_for_condition(OUT_OF_BOX_DETECTION);
@@ -56,8 +56,8 @@ int cleanup_code()
     	free(mainsys.part);
     	mainsys.part=NULL;
     }
-    mainsys.last=NULL;
     dt_param=.03; 
+    accel_zero_mass=1;
     t_now=0.;
     dtime=0.;
     eps2=0.;
@@ -67,6 +67,11 @@ int cleanup_code()
     return 0;
 }
 
+void refresh_lookup()
+{
+    for(int p=0;p<mainsys.n;p++) LOOKUPSYMBOL(,_update)(&lookup, mainsys.part[p].id,p);
+}
+
 int new_particle(int *id, double mass,
                  double x, double y, double z,
                  double vx, double vy, double vz,
@@ -99,7 +104,6 @@ int new_particle(int *id, double mass,
  mainsys.part[p].postime=0;
  mainsys.n++;
  pcounter++;
- mainsys.last=&mainsys.part[p];
  return 0;
 }
 
@@ -111,12 +115,7 @@ int delete_particle(int id)
   if(err!=0) return err;
   LOOKUPSYMBOL(,_delete)(&lookup,id);  
   mainsys.n--;
-  if(mainsys.n==0)
-  {
-    mainsys.last=NULL;
-    return 0;
-  }
-  mainsys.last--;
+  if(mainsys.n==0) return 0;
   mainsys.part[p]=mainsys.part[mainsys.n];
   LOOKUPSYMBOL(,_update)(&lookup,mainsys.part[p].id,p);
   return 0; 
@@ -277,13 +276,13 @@ int get_index_of_next_particle(int  id, int *nout)
 
 int get_kinetic_energy(double *kinetic_energy)
 {
-  *kinetic_energy=system_kinetic_energy(mainsys);
+  *kinetic_energy=(double) system_kinetic_energy(mainsys);
   return 0;
 }
 
 int get_potential_energy(double *potential_energy)
 {
-  *potential_energy=system_potential_energy(mainsys);
+  *potential_energy=(double) system_potential_energy(mainsys);
   return 0;
 }
 
@@ -299,6 +298,19 @@ int set_inttype_parameter(int i)
   return 0;
 }
 
+int get_opencl_device_type(int *i)
+{
+  *i=opencl_device_type;
+  return 0;
+}
+
+int set_opencl_device_type(int i)
+{
+  opencl_device_type=i;
+  return 0;
+}
+
+
 int set_timestep_parameter(double t)
 {
   dt_param=t;
@@ -335,6 +347,18 @@ int get_verbosity_parameter(int *t)
   return 0;
 }
 
+int set_accel_zero_mass_parameter(int t)
+{
+  accel_zero_mass=t;
+  return 0;
+}
+
+int get_accel_zero_mass_parameter(int *t)
+{
+  *t=accel_zero_mass;
+  return 0;
+}
+
 
 int evolve_model(double t_end)
 {
@@ -396,7 +420,7 @@ int evolve_model(double t_end)
       t_now+=dt;
     }
   }
-  for(int p=0;p<mainsys.n;p++) LOOKUPSYMBOL(,_update)(&lookup, mainsys.part[p].id,p);
+  refresh_lookup();
 
   if (set_conditions & enabled_conditions) {
     int err,type, number_of_particles, id;
@@ -433,6 +457,7 @@ int get_begin_time(double * output) {
 int commit_particles()
 {
   init_evolve(mainsys,inttype);
+  refresh_lookup();
   return 0;
 }
 
@@ -467,6 +492,7 @@ int synchronize_model()
 int recommit_particles()
 {
   init_evolve(mainsys,inttype);
+  refresh_lookup();
   return 0;
 }
 
@@ -528,10 +554,12 @@ int get_center_of_mass_velocity(double *vx,double *vy,double *vz)
 int recommit_parameters()
 {
   init_evolve(mainsys,inttype);
+  refresh_lookup();
   return 0;
 }
 int commit_parameters()
 {
+    init_code();
     t_now = begin_time;
     return 0;
 }
@@ -564,10 +592,9 @@ int get_indices_of_colliding_particles(int *p1,int*p2)
 int get_gravity_at_point(double * eps, double * x, double * y, double * z, 
 			 double * ax, double * ay, double * az, int n)
 {
-  struct sys tmpsys;
+  struct sys tmpsys=zerosys;
   tmpsys.n=n;
   tmpsys.part=(struct particle*) malloc(n*sizeof(struct particle));
-  tmpsys.last=&tmpsys.part[n];
   for(int p=0;p<n;p++)
   {
     tmpsys.part[p].pos[0]=x[p];
@@ -598,10 +625,9 @@ int get_potential_at_point(double * eps,
 			   double * x, double * y, double * z, 
 			   double * phi, int n)
 {
-  struct sys tmpsys;
+  struct sys tmpsys=zerosys;
   tmpsys.n=n;
   tmpsys.part=(struct particle*) malloc(n*sizeof(struct particle));
-  tmpsys.last=&tmpsys.part[n];
   for(int p=0;p<n;p++)
   {
     tmpsys.part[p].pos[0]=x[p];

src/amuse/community/huayno/interface.py

@@ -152,6 +152,35 @@ def set_eps2_parameter():
         function.result_type = 'i'
         return function
 
+    @legacy_function
+    def get_accel_zero_mass_parameter():
+        function = LegacyFunctionSpecification()
+        function.addParameter('accelerate_zero_mass', dtype='b', direction=function.OUT)
+        function.result_type = 'i'
+        return function
+
+    @legacy_function
+    def set_accel_zero_mass_parameter():
+        function = LegacyFunctionSpecification()
+        function.addParameter('accelerate_zero_mass', dtype='b', direction=function.IN)
+        function.result_type = 'i'
+        return function
+
+    @legacy_function
+    def get_opencl_device_type():
+        function = LegacyFunctionSpecification()
+        function.addParameter('opencl_device_type', dtype='i', direction=function.OUT)
+        function.result_type = 'i'
+        return function
+
+    @legacy_function
+    def set_opencl_device_type():
+        function = LegacyFunctionSpecification()
+        function.addParameter('opencl_device_type', dtype='i', direction=function.IN)
+        function.result_type = 'i'
+        return function
+
+
     def set_eps2(self, e):
         return self.set_eps2_parameter(e)
 
@@ -197,7 +226,7 @@ def _list(cls):
       CC_BS = 24, CCC_BS = 25, BS_CC_KEPLER = 26, CC_BSA = 27, CCC_BSA = 28, SHARED2_COLLISIONS = 29,
       SHARED4_COLLISIONS = 30, SHARED6_COLLISIONS = 31, SHARED8_COLLISIONS = 32, 
       SHARED10_COLLISIONS = 33, CONSTANT2 = 34, CONSTANT4 = 35, CONSTANT6 = 36, 
-      CONSTANT8 = 37, CONSTANT10 = 38, )
+      CONSTANT8 = 37, CONSTANT10 = 38, ERROR_CONTROL=39, CC_SHARED10=40, CCC_SHARED10=41)
 
     for key, val in all_inttypes.items():
       setattr(inttypes, key, val)
@@ -260,6 +289,14 @@ def define_parameters(self, handler):
             default_value = 0
         )
 
+        handler.add_boolean_parameter(
+            "get_accel_zero_mass_parameter",
+            "set_accel_zero_mass_parameter",
+            "accelerate_zero_mass",
+            "accelerate zero mass particle interactions (should always be true, except for testing)",
+            default_value = True
+        )
+
         inttypes=sorted([(getattr(self.inttypes,t),t ) 
                    for i,t in enumerate(sorted(self.inttypes._list()))])
 
@@ -289,6 +326,13 @@ def define_parameters(self, handler):
             default_value = 0.0 | nbody_system.time
         )
 
+        handler.add_method_parameter(
+            "get_opencl_device_type",
+            "set_opencl_device_type",
+            "opencl_device_type",
+            "set preferred OpenCL device type (0=default, 1=cpu, 2=gpu)",
+            default_value = 0
+        )
 
     def define_methods(self, handler):
         GravitationalDynamics.define_methods(self, handler)