diff --git a/src/KOKKOS/Install.sh b/src/KOKKOS/Install.sh
index 0eea611d20..811164ff37 100644
--- a/src/KOKKOS/Install.sh
+++ b/src/KOKKOS/Install.sh
@@ -155,19 +155,20 @@ if (test $1 = 1) then
     sed -i -e 's/[^ \t]*KOKKOS[^ \t]* //g' ../Makefile.package
     sed -i -e 's|^PKG_INC =[ \t]*|&-DLMP_KOKKOS |' ../Makefile.package
 #    sed -i -e 's|^PKG_PATH =[ \t]*|&-L..\/..\/lib\/kokkos\/core\/src |' ../Makefile.package
-    sed -i -e 's|^PKG_LIB =[ \t]*|&-lkokkoscore |' ../Makefile.package
-    sed -i -e 's|^PKG_SYSINC =[ \t]*|&$(KOKKOS_INC) |' ../Makefile.package
-    sed -i -e 's|^PKG_SYSLIB =[ \t]*|&$(KOKKOS_LINK) |' ../Makefile.package
+    sed -i -e 's|^PKG_CPP_DEPENDS =[ \t]*|&$(KOKKOS_CPP_DEPENDS) |' ../Makefile.package
+    sed -i -e 's|^PKG_LIB =[ \t]*|&$(KOKKOS_LIBS) |' ../Makefile.package
+    sed -i -e 's|^PKG_LINK_DEPENDS =[ \t]*|&$(KOKKOS_LINK_DEPENDS) |' ../Makefile.package
+    sed -i -e 's|^PKG_SYSINC =[ \t]*|&$(KOKKOS_CPPFLAGS) $(KOKKOS_CXXFLAGS) |' ../Makefile.package
+    sed -i -e 's|^PKG_SYSLIB =[ \t]*|&$(KOKKOS_LDFLAGS) |' ../Makefile.package
 #    sed -i -e 's|^PKG_SYSPATH =[ \t]*|&$(kokkos_SYSPATH) |' ../Makefile.package
   fi
 
   if (test -e ../Makefile.package.settings) then
+    sed -i -e '/CXX\ =\ \$(CC)/d' ../Makefile.package.settings
     sed -i -e '/^include.*kokkos.*$/d' ../Makefile.package.settings
     # multiline form needed for BSD sed on Macs
-    sed -i -e '4 i \
-include ..\/..\/lib\/kokkos\/Makefile.lammps
-' ../Makefile.package.settings
-
+    sed -i -e '4 i \CXX = $(CC)' ../Makefile.package.settings
+    sed -i -e '5 i \include ..\/..\/lib\/kokkos\/Makefile.kokkos' ../Makefile.package.settings
   fi
 
 elif (test $1 = 0) then
@@ -178,6 +179,7 @@ elif (test $1 = 0) then
   fi
 
   if (test -e ../Makefile.package.settings) then
+    sed -i -e '/CXX\ =\ \$(CC)/d' ../Makefile.package.settings
     sed -i -e '/^include.*kokkos.*$/d' ../Makefile.package.settings
   fi
 
diff --git a/src/KOKKOS/kokkos_type.h b/src/KOKKOS/kokkos_type.h
index 123bbd1a8c..1f9087c3cc 100644
--- a/src/KOKKOS/kokkos_type.h
+++ b/src/KOKKOS/kokkos_type.h
@@ -22,6 +22,68 @@
 #define MAX_TYPES_STACKPARAMS 12
 #define NeighClusterSize 8
 
+  struct lmp_float3 {
+    float x,y,z;
+    KOKKOS_INLINE_FUNCTION
+    lmp_float3():x(0.0f),z(0.0f),y(0.0f) {}
+
+    KOKKOS_INLINE_FUNCTION
+    void operator += (const lmp_float3& tmp) {
+      x+=tmp.x;
+      y+=tmp.y;
+      z+=tmp.z;
+    }
+    KOKKOS_INLINE_FUNCTION
+    void operator += (const lmp_float3& tmp) volatile {
+      x+=tmp.x;
+      y+=tmp.y;
+      z+=tmp.z;
+    }
+    KOKKOS_INLINE_FUNCTION
+    void operator = (const lmp_float3& tmp) {
+      x=tmp.x;
+      y=tmp.y;
+      z=tmp.z;
+    }
+    KOKKOS_INLINE_FUNCTION
+    void operator = (const lmp_float3& tmp) volatile {
+      x=tmp.x;
+      y=tmp.y;
+      z=tmp.z;
+    }
+  };
+
+  struct lmp_double3 {
+    double x,y,z;
+    KOKKOS_INLINE_FUNCTION
+    lmp_double3():x(0.0),z(0.0),y(0.0) {}
+
+    KOKKOS_INLINE_FUNCTION
+    void operator += (const lmp_double3& tmp) {
+      x+=tmp.x;
+      y+=tmp.y;
+      z+=tmp.z;
+    }
+    KOKKOS_INLINE_FUNCTION
+    void operator += (const lmp_double3& tmp) volatile {
+      x+=tmp.x;
+      y+=tmp.y;
+      z+=tmp.z;
+    }
+    KOKKOS_INLINE_FUNCTION
+    void operator = (const lmp_double3& tmp) {
+      x=tmp.x;
+      y=tmp.y;
+      z=tmp.z;
+    }
+    KOKKOS_INLINE_FUNCTION
+    void operator = (const lmp_double3& tmp) volatile {
+      x=tmp.x;
+      y=tmp.y;
+      z=tmp.z;
+    }
+  };
+
 #if !defined(__CUDACC__) && !defined(__VECTOR_TYPES_H__)
   struct double2 {
     double x, y;
@@ -29,14 +91,13 @@
   struct float2 {
     float x, y;
   };
-  struct double4 {
-    double x, y, z, w;
-  };
   struct float4 {
     float x, y, z, w;
   };
+  struct double4 {
+    double x, y, z, w;
+  };
 #endif
-
 // set LMPHostype and LMPDeviceType from Kokkos Default Types
 typedef Kokkos::DefaultExecutionSpace LMPDeviceType;
 typedef Kokkos::HostSpace::execution_space LMPHostType;
@@ -66,10 +127,12 @@ struct ExecutionSpaceFromDevice<Kokkos::Cuda> {
 #if PRECISION==1
 typedef float LMP_FLOAT;
 typedef float2 LMP_FLOAT2;
+typedef lmp_float3 LMP_FLOAT3;
 typedef float4 LMP_FLOAT4;
 #else
 typedef double LMP_FLOAT;
 typedef double2 LMP_FLOAT2;
+typedef lmp_double3 LMP_FLOAT3;
 typedef double4 LMP_FLOAT4;
 #endif
 
@@ -80,10 +143,12 @@ typedef double4 LMP_FLOAT4;
 #if PREC_FORCE==1
 typedef float F_FLOAT;
 typedef float2 F_FLOAT2;
+typedef lmp_float3 F_FLOAT3;
 typedef float4 F_FLOAT4;
 #else
 typedef double F_FLOAT;
 typedef double2 F_FLOAT2;
+typedef lmp_double3 F_FLOAT3;
 typedef double4 F_FLOAT4;
 #endif
 
@@ -664,7 +729,7 @@ void buffer_view(BufferView &buf, DualView &view,
 
 template<class DeviceType>
 struct MemsetZeroFunctor {
-  typedef DeviceType  device_type ;
+  typedef DeviceType  execution_space ;
   void* ptr;
   KOKKOS_INLINE_FUNCTION void operator()(const int i) const {
     ((int*)ptr)[i] = 0;
@@ -673,10 +738,10 @@ struct MemsetZeroFunctor {
 
 template<class ViewType>
 void memset_kokkos (ViewType &view) {
-  static MemsetZeroFunctor<typename ViewType::device_type> f;
+  static MemsetZeroFunctor<typename ViewType::execution_space> f;
   f.ptr = view.ptr_on_device();
   Kokkos::parallel_for(view.capacity()*sizeof(typename ViewType::value_type)/4, f);
-  ViewType::device_type::fence();
+  ViewType::execution_space::fence();
 }
 
 
diff --git a/src/KOKKOS/pair_kokkos.h b/src/KOKKOS/pair_kokkos.h
index f3bef77b86..5f9b347ddf 100644
--- a/src/KOKKOS/pair_kokkos.h
+++ b/src/KOKKOS/pair_kokkos.h
@@ -336,7 +336,6 @@ struct PairComputeFunctor  {
 template <class PairStyle, bool STACKPARAMS, class Specialisation>
 struct PairComputeFunctor<PairStyle,FULLCLUSTER,STACKPARAMS,Specialisation>  {
   typedef typename PairStyle::device_type device_type ;
-  typedef Kokkos::Vectorization<device_type,NeighClusterSize> vectorization;
   typedef EV_FLOAT value_type;
 
   PairStyle c;
@@ -356,7 +355,7 @@ struct PairComputeFunctor<PairStyle,FULLCLUSTER,STACKPARAMS,Specialisation>  {
   EV_FLOAT compute_item(const typename Kokkos::TeamPolicy<device_type>::member_type& dev,
                         const NeighListKokkos<device_type> &list, const NoCoulTag& ) const {
     EV_FLOAT ev;
-    const int i = vectorization::global_thread_rank(dev);
+    const int i = dev.league_rank()*dev.team_size() + dev.team_rank();
 
     const X_FLOAT xtmp = c.c_x(i,0);
     const X_FLOAT ytmp = c.c_x(i,1);
@@ -366,17 +365,15 @@ struct PairComputeFunctor<PairStyle,FULLCLUSTER,STACKPARAMS,Specialisation>  {
     const AtomNeighborsConst neighbors_i = list.get_neighbors_const(i);
     const int jnum = list.d_numneigh[i];
 
-    F_FLOAT fxtmp = 0.0;
-    F_FLOAT fytmp = 0.0;
-    F_FLOAT fztmp = 0.0;
+    F_FLOAT3 ftmp;
 
     for (int jj = 0; jj < jnum; jj++) {
       const int jjj = neighbors_i(jj);
 
-      for (int k = vectorization::begin(); k<NeighClusterSize; k+=vectorization::increment) {
+      Kokkos::parallel_reduce(Kokkos::ThreadVectorRange(dev,NeighClusterSize),[&] (const int& k, F_FLOAT3& fftmp) {
         const F_FLOAT factor_lj = c.special_lj[sbmask(jjj+k)];
         const int j = (jjj + k)&NEIGHMASK;
-        if((j==i)||(j>=c.nall)) continue;
+        if((j==i)||(j>=c.nall)) return;
         const X_FLOAT delx = xtmp - c.c_x(j,0);
         const X_FLOAT dely = ytmp - c.c_x(j,1);
         const X_FLOAT delz = ztmp - c.c_x(j,2);
@@ -386,9 +383,9 @@ struct PairComputeFunctor<PairStyle,FULLCLUSTER,STACKPARAMS,Specialisation>  {
         if(rsq < (STACKPARAMS?c.m_cutsq[itype][jtype]:c.d_cutsq(itype,jtype))) {
 
           const F_FLOAT fpair = factor_lj*c.template compute_fpair<STACKPARAMS,Specialisation>(rsq,i,j,itype,jtype);
-          fxtmp += delx*fpair;
-          fytmp += dely*fpair;
-          fztmp += delz*fpair;
+          fftmp.x += delx*fpair;
+          fftmp.y += dely*fpair;
+          fftmp.z += delz*fpair;
 
           if (EVFLAG) {
             F_FLOAT evdwl = 0.0;
@@ -401,17 +398,14 @@ struct PairComputeFunctor<PairStyle,FULLCLUSTER,STACKPARAMS,Specialisation>  {
             if (c.vflag_either || c.eflag_atom) ev_tally(ev,i,j,evdwl,fpair,delx,dely,delz);
           }
         }
-      }
+      },ftmp);
     }
 
-    const F_FLOAT fx = vectorization::reduce(fxtmp);
-    const F_FLOAT fy = vectorization::reduce(fytmp);
-    const F_FLOAT fz = vectorization::reduce(fztmp);
-    if(vectorization::is_lane_0(dev)) {
-      c.f(i,0) += fx;
-      c.f(i,1) += fy;
-      c.f(i,2) += fz;
-    }
+    Kokkos::single(Kokkos::PerThread(dev), [&]() {
+      c.f(i,0) += ftmp.x;
+      c.f(i,1) += ftmp.y;
+      c.f(i,2) += ftmp.z;
+    });
 
     return ev;
   }
@@ -659,12 +653,12 @@ EV_FLOAT pair_compute_fullcluster (PairStyle* fpair, typename Kokkos::Impl::enab
       f_type;
     f_type ff(fpair, list);
     #ifdef KOKKOS_HAVE_CUDA
-      const int teamsize = Kokkos::Impl::is_same<typename f_type::device_type, Kokkos::Cuda>::value ? 256 : 1;
+      const int teamsize = Kokkos::Impl::is_same<typename f_type::device_type, Kokkos::Cuda>::value ? 32 : 1;
     #else
       const int teamsize = 1;
     #endif
-    const int nteams = (list->inum*f_type::vectorization::increment+teamsize-1)/teamsize;
-    Kokkos::TeamPolicy<typename f_type::device_type> config(nteams,teamsize);
+    const int nteams = (list->inum*+teamsize-1)/teamsize;
+    Kokkos::TeamPolicy<typename f_type::device_type> config(nteams,teamsize,NeighClusterSize);
     if (fpair->eflag || fpair->vflag) Kokkos::parallel_reduce(config,ff,ev);
     else Kokkos::parallel_for(config,ff);
   } else {
@@ -672,12 +666,12 @@ EV_FLOAT pair_compute_fullcluster (PairStyle* fpair, typename Kokkos::Impl::enab
       f_type;
     f_type ff(fpair, list);
     #ifdef KOKKOS_HAVE_CUDA
-      const int teamsize = Kokkos::Impl::is_same<typename f_type::device_type, Kokkos::Cuda>::value ? 256 : 1;
+      const int teamsize = Kokkos::Impl::is_same<typename f_type::device_type, Kokkos::Cuda>::value ? 32 : 1;
     #else
       const int teamsize = 1;
     #endif
-    const int nteams = (list->inum*f_type::vectorization::increment+teamsize-1)/teamsize;
-    Kokkos::TeamPolicy<typename f_type::device_type> config(nteams,teamsize);
+      const int nteams = (list->inum*+teamsize-1)/teamsize;
+      Kokkos::TeamPolicy<typename f_type::device_type> config(nteams,teamsize,NeighClusterSize);
     if (fpair->eflag || fpair->vflag) Kokkos::parallel_reduce(config,ff,ev);
     else Kokkos::parallel_for(config,ff);
   }
@@ -706,12 +700,12 @@ template<class DeviceType>
 struct PairVirialFDotRCompute {
   typedef ArrayTypes<DeviceType> AT;
   typedef EV_FLOAT value_type;
-  typename AT::t_x_array_const x;
-  typename AT::t_f_array_const f;
+  typename AT::t_x_array_const_um x;
+  typename AT::t_f_array_const_um f;
   const int offset;
 
-  PairVirialFDotRCompute(  typename AT::t_x_array_const x_,
-  typename AT::t_f_array_const f_,
+  PairVirialFDotRCompute(  typename AT::t_x_array_const_um x_,
+  typename AT::t_f_array_const_um f_,
   const int offset_):x(x_),f(f_),offset(offset_) {}
 
   KOKKOS_INLINE_FUNCTION
diff --git a/src/KOKKOS/pair_table_kokkos.cpp b/src/KOKKOS/pair_table_kokkos.cpp
index dfd6787c95..f0c6068bbe 100644
--- a/src/KOKKOS/pair_table_kokkos.cpp
+++ b/src/KOKKOS/pair_table_kokkos.cpp
@@ -147,12 +147,12 @@ void PairTableKokkos<DeviceType>::compute_style(int eflag_in, int vflag_in)
         f_type;
       f_type f(this,(NeighListKokkos<DeviceType>*) list);
       #ifdef KOKKOS_HAVE_CUDA
-        const int teamsize = Kokkos::Impl::is_same<typename f_type::device_type, Kokkos::Cuda>::value ? 256 : 1;
+        const int teamsize = Kokkos::Impl::is_same<DeviceType, Kokkos::Cuda>::value ? 32 : 1;
       #else
         const int teamsize = 1;
       #endif
-      const int nteams = (list->inum*f_type::vectorization::increment+teamsize-1)/teamsize;
-      Kokkos::TeamPolicy<DeviceType> config(nteams,teamsize);
+      const int nteams = (list->inum*+teamsize-1)/teamsize;
+      Kokkos::TeamPolicy<DeviceType> config(nteams,teamsize,NeighClusterSize);
       if (eflag || vflag) Kokkos::parallel_reduce(config,f,ev);
       else Kokkos::parallel_for(config,f);
     }
@@ -182,12 +182,12 @@ void PairTableKokkos<DeviceType>::compute_style(int eflag_in, int vflag_in)
         f_type;
       f_type f(this,(NeighListKokkos<DeviceType>*) list);
       #ifdef KOKKOS_HAVE_CUDA
-        const int teamsize = Kokkos::Impl::is_same<typename f_type::device_type, Kokkos::Cuda>::value ? 256 : 1;
+        const int teamsize = Kokkos::Impl::is_same<DeviceType, Kokkos::Cuda>::value ? 32 : 1;
       #else
         const int teamsize = 1;
       #endif
-      const int nteams = (list->inum*f_type::vectorization::increment+teamsize-1)/teamsize;
-      Kokkos::TeamPolicy<DeviceType> config(nteams,teamsize);
+      const int nteams = (list->inum*+teamsize-1)/teamsize;
+      Kokkos::TeamPolicy<DeviceType> config(nteams,teamsize,NeighClusterSize);
       if (eflag || vflag) Kokkos::parallel_reduce(config,f,ev);
       else Kokkos::parallel_for(config,f);
     }
diff --git a/src/MAKE/Makefile.mpi b/src/MAKE/Makefile.mpi
index 3d1766bbd0..9497fc7c92 100755
--- a/src/MAKE/Makefile.mpi
+++ b/src/MAKE/Makefile.mpi
@@ -74,7 +74,9 @@ include	Makefile.package
 
 EXTRA_INC = $(LMP_INC) $(PKG_INC) $(MPI_INC) $(FFT_INC) $(JPG_INC) $(PKG_SYSINC)
 EXTRA_PATH = $(PKG_PATH) $(MPI_PATH) $(FFT_PATH) $(JPG_PATH) $(PKG_SYSPATH)
+EXTRA_CPP_DEPENDS = $(PKG_CPP_DEPENDS)
 EXTRA_LIB = $(PKG_LIB) $(MPI_LIB) $(FFT_LIB) $(JPG_LIB) $(PKG_SYSLIB)
+EXTRA_LINK_DEPENDS = $(PKG_LINK_DEPENDS)
 
 # Path to src files
 
@@ -83,28 +85,28 @@ vpath %.h ..
 
 # Link target
 
-$(EXE):	$(OBJ)
+$(EXE):	$(OBJ) $(EXTRA_LINK_DEPENDS)
 	$(LINK) $(LINKFLAGS) $(EXTRA_PATH) $(OBJ) $(EXTRA_LIB) $(LIB) -o $(EXE)
 	$(SIZE) $(EXE)
 
 # Library targets
 
-lib:	$(OBJ)
+lib:	$(OBJ) $(EXTRA_LINK_DEPENDS)
 	$(ARCHIVE) $(ARFLAGS) $(EXE) $(OBJ)
 
-shlib:	$(OBJ)
+shlib:	$(OBJ) $(EXTRA_LINK_DEPENDS)
 	$(CC) $(CCFLAGS) $(SHFLAGS) $(SHLIBFLAGS) $(EXTRA_PATH) -o $(EXE) \
         $(OBJ) $(EXTRA_LIB) $(LIB)
 
 # Compilation rules
 
-%.o:%.cpp
+%.o:%.cpp $(EXTRA_CPP_DEPENDS)
 	$(CC) $(CCFLAGS) $(SHFLAGS) $(EXTRA_INC) -c $<
 
-%.d:%.cpp
+%.d:%.cpp $(EXTRA_CPP_DEPENDS)
 	$(CC) $(CCFLAGS) $(EXTRA_INC) $(DEPFLAGS) $< > $@
 
-%.o:%.cu
+%.o:%.cu $(EXTRA_CPP_DEPENDS)
 	$(CC) $(CCFLAGS) $(SHFLAGS) $(EXTRA_INC) -c $<
 
 # Individual dependencies
diff --git a/src/MAKE/Makefile.serial b/src/MAKE/Makefile.serial
index 767624c6de..efc5fed755 100755
--- a/src/MAKE/Makefile.serial
+++ b/src/MAKE/Makefile.serial
@@ -74,7 +74,9 @@ include	Makefile.package
 
 EXTRA_INC = $(LMP_INC) $(PKG_INC) $(MPI_INC) $(FFT_INC) $(JPG_INC) $(PKG_SYSINC)
 EXTRA_PATH = $(PKG_PATH) $(MPI_PATH) $(FFT_PATH) $(JPG_PATH) $(PKG_SYSPATH)
+EXTRA_CCP_DEPENDS = $(PKG_CPP_DEPENDS)
 EXTRA_LIB = $(PKG_LIB) $(MPI_LIB) $(FFT_LIB) $(JPG_LIB) $(PKG_SYSLIB)
+EXTRA_LINK_DEPENDS = $(PKG_LINK_DEPENDS)
 
 # Path to src files
 
@@ -83,28 +85,28 @@ vpath %.h ..
 
 # Link target
 
-$(EXE):	$(OBJ)
+$(EXE):	$(OBJ) $(EXTRA_LINK_DEPENDS)
 	$(LINK) $(LINKFLAGS) $(EXTRA_PATH) $(OBJ) $(EXTRA_LIB) $(LIB) -o $(EXE)
 	$(SIZE) $(EXE)
 
 # Library targets
 
-lib:	$(OBJ)
+lib:	$(OBJ) $(EXTRA_LINK_DEPENDS)
 	$(ARCHIVE) $(ARFLAGS) $(EXE) $(OBJ)
 
-shlib:	$(OBJ)
+shlib:	$(OBJ) $(EXTRA_LINK_DEPENDS)
 	$(CC) $(CCFLAGS) $(SHFLAGS) $(SHLIBFLAGS) $(EXTRA_PATH) -o $(EXE) \
         $(OBJ) $(EXTRA_LIB) $(LIB)
 
 # Compilation rules
 
-%.o:%.cpp
+%.o:%.cpp $(EXTRA_CPP_DEPENDS)
 	$(CC) $(CCFLAGS) $(SHFLAGS) $(EXTRA_INC) -c $<
 
-%.d:%.cpp
+%.d:%.cpp $(EXTRA_CPP_DEPENDS)
 	$(CC) $(CCFLAGS) $(EXTRA_INC) $(DEPFLAGS) $< > $@
 
-%.o:%.cu
+%.o:%.cu $(EXTRA_CPP_DEPENDS)
 	$(CC) $(CCFLAGS) $(SHFLAGS) $(EXTRA_INC) -c $<
 
 # Individual dependencies
diff --git a/src/Makefile.package.empty b/src/Makefile.package.empty
index 0d8e6c175f..d421877e2d 100644
--- a/src/Makefile.package.empty
+++ b/src/Makefile.package.empty
@@ -4,6 +4,8 @@
 PKG_INC =   
 PKG_PATH =  
 PKG_LIB =   
+PKG_CPP_DEPENDS = 
+PKG_LINK_DEPENDS = 
 
 PKG_SYSINC =  
 PKG_SYSLIB =