diff --git a/src/USER-CUDA/cuda.h b/src/USER-CUDA/cuda.h
index 2494796da2..9d77dcdb63 100644
--- a/src/USER-CUDA/cuda.h
+++ b/src/USER-CUDA/cuda.h
@@ -38,116 +38,118 @@
 
 namespace LAMMPS_NS
 {
-        class Cuda : protected Pointers
-        {
-                public:
-                Cuda(class LAMMPS *);
-                ~Cuda();
-                //static void setDevice(class LAMMPS*);
-                void allocate();
+class Cuda : protected Pointers
+{
+  public:
+    Cuda(class LAMMPS*);
+    ~Cuda();
+    //static void setDevice(class LAMMPS*);
+    void allocate();
 
-                void accelerator(int, char **);
+    void accelerator(int, char**);
 
-                void setSharedDataZero();
-                void setSystemParams();
+    void setSharedDataZero();
+    void setSystemParams();
 
-                void setDomainParams();
+    void setDomainParams();
 
-                void checkResize();
-                void evsetup_eatom_vatom(int eflag_atom,int vflag_atom);
-                void uploadAll();
-                void downloadAll();
-                void downloadX();
+    void checkResize();
+    void evsetup_eatom_vatom(int eflag_atom, int vflag_atom);
+    void uploadAll();
+    void downloadAll();
+    void upload(int datamask);
+    void download(int datamask);
+    void downloadX();
 
-                class CudaNeighList* registerNeighborList(class NeighList* neigh_list);
-                void uploadAllNeighborLists();
-                void downloadAllNeighborLists();
-                void set_neighinit(int dist_check, double triggerneighsq)
-                {
-              shared_data.atom.dist_check=dist_check;
-              shared_data.atom.triggerneighsq = triggerneighsq;
-                }
-                bool decide_by_integrator()
-                {
-                 return neighbor_decide_by_integrator  && cu_xhold && finished_setup;
-            }
-            void update_xhold(int &maxhold,double* xhold);
+    class CudaNeighList* registerNeighborList(class NeighList* neigh_list);
+    void uploadAllNeighborLists();
+    void downloadAllNeighborLists();
+    void set_neighinit(int dist_check, double triggerneighsq) {
+      shared_data.atom.dist_check = dist_check;
+      shared_data.atom.triggerneighsq = triggerneighsq;
+    }
+    bool decide_by_integrator() {
+      return neighbor_decide_by_integrator  && cu_xhold && finished_setup;
+    }
+    void update_xhold(int &maxhold, double* xhold);
 
-                void setTimingsZero();
-                void print_timings();
+    void setTimingsZero();
+    void print_timings();
 
-                void cu_x_download() {cu_x->download();}
-                bool device_set;
-                bool dotiming;
-                bool dotestatom;
-                int testatom;
+    void cu_x_download() {
+      cu_x->download();
+    }
+    bool device_set;
+    bool dotiming;
+    bool dotestatom;
+    int testatom;
 
-                double uploadtime,downloadtime;
-                bool finished_setup,begin_setup;
-                bool oncpu;
-                bool finished_run;
+    double uploadtime, downloadtime;
+    bool finished_setup, begin_setup;
+    bool oncpu;
+    bool finished_run;
 
-                int self_comm;
+    int self_comm;
 
-                int cuda_exists;
+    int cuda_exists;
 
-                double extent[6];
-                int* debugdata;
-                // data shared between host code and device code
-                // (number of atoms, device pointers for up- & download)
-                cuda_shared_data shared_data;
+    double extent[6];
+    int* debugdata;
+    // data shared between host code and device code
+    // (number of atoms, device pointers for up- & download)
+    cuda_shared_data shared_data;
 
-                cCudaData<double  , F_FLOAT , x >* cu_q;
-                cCudaData<double  , F_FLOAT , yx>* cu_f;
-                cCudaData<double  , V_FLOAT , x >* cu_mass;
-                cCudaData<double  , V_FLOAT , x >* cu_rmass;
-                cCudaData<double  , V_FLOAT , yx>* cu_v;
-                cCudaData<double  , X_FLOAT , yx>* cu_x;
-                cCudaData<double  , X_FLOAT , yx>* cu_xhold;
-                cCudaData<int     , int     , x >* cu_mask;
-                cCudaData<int     , int     , x >* cu_tag;
-                cCudaData<int     , int     , x >* cu_type;
-                cCudaData<int     , int     , x >* cu_image;
-                cCudaData<double  , ENERGY_FLOAT, x >* cu_eatom;
-                cCudaData<double  , ENERGY_FLOAT, yx>* cu_vatom;
-                cCudaData<double  , ENERGY_FLOAT, x >* cu_virial;
-                cCudaData<double  , ENERGY_FLOAT, x >* cu_eng_vdwl;
-                cCudaData<double  , ENERGY_FLOAT, x >* cu_eng_coul;
-                cCudaData<double  , double  , x >* cu_extent;
-                int* binned_id;
-                cCudaData<int           , int            , xx >* cu_binned_id;
-                int* binned_idnew;
-                cCudaData<int           , int            , xx >* cu_binned_idnew;
-                cCudaData<int           , int            , x >* cu_debugdata;
-                cCudaData<double  , X_FLOAT , x>* cu_radius;
-                cCudaData<double  , F_FLOAT , x>* cu_density;
-                cCudaData<double  , V_FLOAT , yx>* cu_omega;
-                cCudaData<double  , F_FLOAT , yx>* cu_torque;
-                cCudaData<int           , int            , yx >* cu_special;
-                cCudaData<int           , int            , yx >* cu_nspecial;
-                cCudaData<int     , int     , x >* cu_molecule;
+    cCudaData<double  , F_FLOAT , x >* cu_q;
+    cCudaData<double  , F_FLOAT , yx>* cu_f;
+    cCudaData<double  , V_FLOAT , x >* cu_mass;
+    cCudaData<double  , V_FLOAT , x >* cu_rmass;
+    cCudaData<double  , V_FLOAT , yx>* cu_v;
+    cCudaData<double  , X_FLOAT , yx>* cu_x;
+    cCudaData<double  , X_FLOAT , yx>* cu_xhold;
+    cCudaData<int     , int     , x >* cu_mask;
+    cCudaData<int     , int     , x >* cu_tag;
+    cCudaData<int     , int     , x >* cu_type;
+    cCudaData<int     , int     , x >* cu_image;
+    cCudaData<double  , ENERGY_FLOAT, x >* cu_eatom;
+    cCudaData<double  , ENERGY_FLOAT, yx>* cu_vatom;
+    cCudaData<double  , ENERGY_FLOAT, x >* cu_virial;
+    cCudaData<double  , ENERGY_FLOAT, x >* cu_eng_vdwl;
+    cCudaData<double  , ENERGY_FLOAT, x >* cu_eng_coul;
+    cCudaData<double  , double  , x >* cu_extent;
+    int* binned_id;
+    cCudaData<int           , int            , xx >* cu_binned_id;
+    int* binned_idnew;
+    cCudaData<int           , int            , xx >* cu_binned_idnew;
+    cCudaData<int           , int            , x >* cu_debugdata;
+    cCudaData<double  , X_FLOAT , x>* cu_radius;
+    cCudaData<double  , F_FLOAT , x>* cu_density;
+    cCudaData<double  , V_FLOAT , yx>* cu_omega;
+    cCudaData<double  , F_FLOAT , yx>* cu_torque;
+    cCudaData<int           , int            , yx >* cu_special;
+    cCudaData<int           , int            , yx >* cu_nspecial;
+    cCudaData<int     , int     , x >* cu_molecule;
 
 
-                cCudaData<X_FLOAT  , X_FLOAT , x>* cu_x_type;
-                X_FLOAT* x_type;
+    cCudaData<X_FLOAT  , X_FLOAT , x>* cu_x_type;
+    X_FLOAT* x_type;
 
-                cCudaData<V_FLOAT  , V_FLOAT , x>* cu_v_radius;
-                V_FLOAT* v_radius;
+    cCudaData<V_FLOAT  , V_FLOAT , x>* cu_v_radius;
+    V_FLOAT* v_radius;
 
-                cCudaData<V_FLOAT  , V_FLOAT , x>* cu_omega_rmass;
-                V_FLOAT* omega_rmass;
+    cCudaData<V_FLOAT  , V_FLOAT , x>* cu_omega_rmass;
+    V_FLOAT* omega_rmass;
 
-                cCudaData<int     , int     , x >* cu_map_array;
-                int neighbor_decide_by_integrator;
+    cCudaData<int     , int     , x >* cu_map_array;
+    int neighbor_decide_by_integrator;
 
-        bool pinned;
+    bool pinned;
 
-                void* copy_buffer;
-                int copy_buffersize;
+    void* copy_buffer;
+    int copy_buffersize;
 
-                private:
-                std::map<class NeighList*, class CudaNeighList*> neigh_lists;
-        };
+  private:
+    std::map<class NeighList*, class CudaNeighList*> neigh_lists;
+};
 }
 
 #endif // CUDA_H
diff --git a/src/group.cpp b/src/group.cpp
index e15891d3a2..21bea8f750 100644
--- a/src/group.cpp
+++ b/src/group.cpp
@@ -25,7 +25,10 @@
 #include "fix.h"
 #include "compute.h"
 #include "output.h"
+#include "input.h"
+#include "variable.h"
 #include "dump.h"
+#include "memory.h"
 #include "error.h"
 
 using namespace LAMMPS_NS;
@@ -241,6 +244,29 @@ void Group::assign(int narg, char **arg)
 
     delete [] list;
 
+  // style = variable
+
+  } else if (strcmp(arg[1],"variable") == 0) {
+
+    int ivar = input->variable->find(arg[2]);
+    if (ivar < 0) error->all(FLERR,"Variable name for group does not exist");
+    if (!input->variable->atomstyle(ivar))
+      error->all(FLERR,"Variable for group is invalid style");
+
+    double *aflag;
+    
+    // aflag = evaluation of per-atom variable
+
+    memory->create(aflag,nlocal,"group:aflag");
+    input->variable->compute_atom(ivar,0,aflag,1,0);
+
+    // add to group if per-atom variable evaluated to non-zero
+
+    for (i = 0; i < nlocal; i++)
+      if (aflag[i] != 0.0) mask[i] |= bit;
+
+    memory->destroy(aflag);
+
   // style = subtract
 
   } else if (strcmp(arg[1],"subtract") == 0) {