diff --git a/lib/gpu/Nvidia.makefile b/lib/gpu/Nvidia.makefile
index 1776cb44a9..2f65e49acb 100644
--- a/lib/gpu/Nvidia.makefile
+++ b/lib/gpu/Nvidia.makefile
@@ -33,7 +33,7 @@ NVD_H  = $(wildcard ./geryon/nvd*.h) $(UCL_H) nv_kernel_def.h
 # Headers for Pair Stuff
 PAIR_H  = pair_gpu_atom.h pair_gpu_ans.h pair_gpu_nbor_shared.h \
           pair_gpu_nbor.h pair_gpu_precision.h pair_gpu_device.h \
-          pair_gpu_balance.h pppm_gpu_memory.h
+          pair_gpu_balance.h pppm.h
 
 ALL_H = $(NVD_H) $(PAIR_H)
 
@@ -45,20 +45,20 @@ OBJS = $(OBJ_DIR)/pair_gpu_atom.o $(OBJ_DIR)/pair_gpu_ans.o \
        $(OBJ_DIR)/pair_gpu_nbor.o $(OBJ_DIR)/pair_gpu_nbor_shared.o \
        $(OBJ_DIR)/pair_gpu_device.o $(OBJ_DIR)/atomic_gpu_memory.o \
        $(OBJ_DIR)/charge_gpu_memory.o $(OBJ_DIR)/base_ellipsoid.o \
-       $(OBJ_DIR)/pppm_gpu_memory.o $(OBJ_DIR)/pppm_l_gpu.o \
+       $(OBJ_DIR)/pppm.o $(OBJ_DIR)/pppm_ext.o \
        $(OBJ_DIR)/gayberne.o $(OBJ_DIR)/gayberne_ext.o \
        $(OBJ_DIR)/re_squared.o $(OBJ_DIR)/re_squared_ext.o \
-       $(OBJ_DIR)/lj_cut_gpu_memory.o $(OBJ_DIR)/lj_cut_gpu.o \
-       $(OBJ_DIR)/lj96_cut_gpu_memory.o $(OBJ_DIR)/lj96_cut_gpu.o \
-       $(OBJ_DIR)/lj_expand_gpu_memory.o $(OBJ_DIR)/lj_expand_gpu.o \
-       $(OBJ_DIR)/ljc_cut_gpu_memory.o $(OBJ_DIR)/ljc_cut_gpu.o \
-       $(OBJ_DIR)/ljcl_cut_gpu_memory.o $(OBJ_DIR)/ljcl_cut_gpu.o \
+       $(OBJ_DIR)/lj.o $(OBJ_DIR)/lj_ext.o \
+       $(OBJ_DIR)/lj96.o $(OBJ_DIR)/lj96_ext.o \
+       $(OBJ_DIR)/lj_expand.o $(OBJ_DIR)/lj_expand_ext.o \
+       $(OBJ_DIR)/lj_coul.o $(OBJ_DIR)/lj_coul_ext.o \
+       $(OBJ_DIR)/lj_coul_long.o $(OBJ_DIR)/lj_coul_long_ext.o \
        $(OBJ_DIR)/lj_class2_long.o $(OBJ_DIR)/lj_class2_long_ext.o \
-       $(OBJ_DIR)/morse_gpu_memory.o $(OBJ_DIR)/morse_gpu.o \
-       $(OBJ_DIR)/crml_gpu_memory.o $(OBJ_DIR)/crml_gpu.o \
-       $(OBJ_DIR)/cmm_cut_gpu_memory.o $(OBJ_DIR)/cmm_cut_gpu.o \
-       $(OBJ_DIR)/cmmc_long_gpu_memory.o $(OBJ_DIR)/cmmc_long_gpu.o \
-       $(OBJ_DIR)/cmmc_msm_gpu_memory.o $(OBJ_DIR)/cmmc_msm_gpu.o \
+       $(OBJ_DIR)/morse.o $(OBJ_DIR)/morse_ext.o \
+       $(OBJ_DIR)/charmm_long.o $(OBJ_DIR)/charmm_long_ext.o \
+       $(OBJ_DIR)/cg_cmm.o $(OBJ_DIR)/cg_cmm_ext.o \
+       $(OBJ_DIR)/cg_cmm_long.o $(OBJ_DIR)/cg_cmm_long_ext.o \
+       $(OBJ_DIR)/cg_cmm_msm.o $(OBJ_DIR)/cg_cmm_msm_ext.o \
        $(CUDPP)
 PTXS = $(OBJ_DIR)/pair_gpu_dev_kernel.ptx \
        $(OBJ_DIR)/pair_gpu_atom_kernel.ptx $(OBJ_DIR)/pair_gpu_atom_ptx.h \
@@ -70,17 +70,17 @@ PTXS = $(OBJ_DIR)/pair_gpu_dev_kernel.ptx \
        $(OBJ_DIR)/gayberne.ptx $(OBJ_DIR)/gayberne_lj.ptx \
        $(OBJ_DIR)/gayberne_ptx.h $(OBJ_DIR)/re_squared.ptx \
        $(OBJ_DIR)/re_squared_lj.ptx $(OBJ_DIR)/re_squared_ptx.h \
-       $(OBJ_DIR)/lj_cut_gpu_kernel.ptx $(OBJ_DIR)/lj_cut_gpu_ptx.h \
-       $(OBJ_DIR)/lj96_cut_gpu_kernel.ptx $(OBJ_DIR)/lj96_cut_gpu_ptx.h \
-       $(OBJ_DIR)/lj_expand_gpu_kernel.ptx $(OBJ_DIR)/lj_expand_gpu_ptx.h \
-       $(OBJ_DIR)/ljc_cut_gpu_kernel.ptx $(OBJ_DIR)/ljc_cut_gpu_ptx.h \
-       $(OBJ_DIR)/ljcl_cut_gpu_kernel.ptx $(OBJ_DIR)/ljcl_cut_gpu_ptx.h \
+       $(OBJ_DIR)/lj.ptx $(OBJ_DIR)/lj_ext_ptx.h \
+       $(OBJ_DIR)/lj96.ptx $(OBJ_DIR)/lj96_ext_ptx.h \
+       $(OBJ_DIR)/lj_expand.ptx $(OBJ_DIR)/lj_expand_ext_ptx.h \
+       $(OBJ_DIR)/lj_coul.ptx $(OBJ_DIR)/lj_coul_ext_ptx.h \
+       $(OBJ_DIR)/lj_coul_long.ptx $(OBJ_DIR)/lj_coul_long_ext_ptx.h \
        $(OBJ_DIR)/lj_class2_long.ptx $(OBJ_DIR)/lj_class2_long_ptx.h \
-       $(OBJ_DIR)/morse_gpu_kernel.ptx $(OBJ_DIR)/morse_gpu_ptx.h \
-       $(OBJ_DIR)/crml_gpu_kernel.ptx $(OBJ_DIR)/crml_gpu_ptx.h \
-       $(OBJ_DIR)/cmm_cut_gpu_kernel.ptx $(OBJ_DIR)/cmm_cut_gpu_ptx.h \
-       $(OBJ_DIR)/cmmc_long_gpu_kernel.ptx $(OBJ_DIR)/cmmc_long_gpu_ptx.h \
-       $(OBJ_DIR)/cmmc_msm_gpu_kernel.ptx $(OBJ_DIR)/cmmc_msm_gpu_ptx.h
+       $(OBJ_DIR)/morse.ptx $(OBJ_DIR)/morse_ext_ptx.h \
+       $(OBJ_DIR)/charmm_long.ptx $(OBJ_DIR)/charmm_long_ext_ptx.h \
+       $(OBJ_DIR)/cg_cmm.ptx $(OBJ_DIR)/cg_cmm_ext_ptx.h \
+       $(OBJ_DIR)/cg_cmm_long.ptx $(OBJ_DIR)/cg_cmm_long_ext_ptx.h \
+       $(OBJ_DIR)/cg_cmm_msm.ptx $(OBJ_DIR)/cg_cmm_msm_ext_ptx.h
 
 all: $(GPU_LIB) $(EXECS)
 
@@ -150,23 +150,23 @@ $(OBJ_DIR)/charge_gpu_memory.o: $(ALL_H) charge_gpu_memory.h charge_gpu_memory.c
 $(OBJ_DIR)/base_ellipsoid.o: $(ALL_H) base_ellipsoid.h base_ellipsoid.cpp $(OBJ_DIR)/ellipsoid_nbor_ptx.h
 	$(CUDR) -o $@ -c base_ellipsoid.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/pppm_f_gpu_kernel.ptx: pppm_gpu_kernel.cu pair_gpu_precision.h
-	$(CUDA) --ptx -DNV_KERNEL -Dgrdtyp=float -Dgrdtyp4=float4 -o $@ pppm_gpu_kernel.cu
+$(OBJ_DIR)/pppm_f_gpu_kernel.ptx: pppm.cu pair_gpu_precision.h
+	$(CUDA) --ptx -DNV_KERNEL -Dgrdtyp=float -Dgrdtyp4=float4 -o $@ pppm.cu
 
 $(OBJ_DIR)/pppm_f_gpu_ptx.h: $(OBJ_DIR)/pppm_f_gpu_kernel.ptx
 	$(BSH) ./geryon/file_to_cstr.sh $(OBJ_DIR)/pppm_f_gpu_kernel.ptx $(OBJ_DIR)/pppm_f_gpu_ptx.h
 
-$(OBJ_DIR)/pppm_d_gpu_kernel.ptx: pppm_gpu_kernel.cu pair_gpu_precision.h
-	$(CUDA) --ptx -DNV_KERNEL -Dgrdtyp=double -Dgrdtyp4=double4 -o $@ pppm_gpu_kernel.cu
+$(OBJ_DIR)/pppm_d_gpu_kernel.ptx: pppm.cu pair_gpu_precision.h
+	$(CUDA) --ptx -DNV_KERNEL -Dgrdtyp=double -Dgrdtyp4=double4 -o $@ pppm.cu
 
 $(OBJ_DIR)/pppm_d_gpu_ptx.h: $(OBJ_DIR)/pppm_d_gpu_kernel.ptx
 	$(BSH) ./geryon/file_to_cstr.sh $(OBJ_DIR)/pppm_d_gpu_kernel.ptx $(OBJ_DIR)/pppm_d_gpu_ptx.h
 
-$(OBJ_DIR)/pppm_gpu_memory.o: $(ALL_H) pppm_gpu_memory.h pppm_gpu_memory.cpp $(OBJ_DIR)/pppm_f_gpu_ptx.h $(OBJ_DIR)/pppm_d_gpu_ptx.h
-	$(CUDR) -o $@ -c pppm_gpu_memory.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/pppm.o: $(ALL_H) pppm.h pppm.cpp $(OBJ_DIR)/pppm_f_gpu_ptx.h $(OBJ_DIR)/pppm_d_gpu_ptx.h
+	$(CUDR) -o $@ -c pppm.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/pppm_l_gpu.o: $(ALL_H) pppm_gpu_memory.h pppm_l_gpu.cpp
-	$(CUDR) -o $@ -c pppm_l_gpu.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/pppm_ext.o: $(ALL_H) pppm.h pppm_ext.cpp
+	$(CUDR) -o $@ -c pppm_ext.cpp -I$(OBJ_DIR)
 
 $(OBJ_DIR)/ellipsoid_nbor.ptx: ellipsoid_nbor.cu pair_gpu_precision.h
 	$(CUDA) --ptx -DNV_KERNEL -o $@ ellipsoid_nbor.cu
@@ -204,29 +204,29 @@ $(OBJ_DIR)/re_squared.o: $(ALL_H) re_squared.h re_squared.cpp $(OBJ_DIR)/re_squa
 $(OBJ_DIR)/re_squared_ext.o: $(ALL_H) $(OBJ_DIR)/re_squared.o re_squared_ext.cpp
 	$(CUDR) -o $@ -c re_squared_ext.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/lj_cut_gpu_kernel.ptx: lj_cut_gpu_kernel.cu pair_gpu_precision.h
-	$(CUDA) --ptx -DNV_KERNEL -o $@ lj_cut_gpu_kernel.cu
+$(OBJ_DIR)/lj.ptx: lj.cu pair_gpu_precision.h
+	$(CUDA) --ptx -DNV_KERNEL -o $@ lj.cu
 
-$(OBJ_DIR)/lj_cut_gpu_ptx.h: $(OBJ_DIR)/lj_cut_gpu_kernel.ptx $(OBJ_DIR)/lj_cut_gpu_kernel.ptx
-	$(BSH) ./geryon/file_to_cstr.sh $(OBJ_DIR)/lj_cut_gpu_kernel.ptx $(OBJ_DIR)/lj_cut_gpu_ptx.h
+$(OBJ_DIR)/lj_ext_ptx.h: $(OBJ_DIR)/lj.ptx $(OBJ_DIR)/lj.ptx
+	$(BSH) ./geryon/file_to_cstr.sh $(OBJ_DIR)/lj.ptx $(OBJ_DIR)/lj_ext_ptx.h
 
-$(OBJ_DIR)/lj_cut_gpu_memory.o: $(ALL_H) lj_cut_gpu_memory.h lj_cut_gpu_memory.cpp $(OBJ_DIR)/lj_cut_gpu_ptx.h $(OBJ_DIR)/atomic_gpu_memory.o
-	$(CUDR) -o $@ -c lj_cut_gpu_memory.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/lj.o: $(ALL_H) lj.h lj.cpp $(OBJ_DIR)/lj_ext_ptx.h $(OBJ_DIR)/atomic_gpu_memory.o
+	$(CUDR) -o $@ -c lj.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/lj_cut_gpu.o: $(ALL_H) lj_cut_gpu_memory.h lj_cut_gpu.cpp atomic_gpu_memory.h
-	$(CUDR) -o $@ -c lj_cut_gpu.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/lj_ext.o: $(ALL_H) lj.h lj_ext.cpp atomic_gpu_memory.h
+	$(CUDR) -o $@ -c lj_ext.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/ljc_cut_gpu_kernel.ptx: ljc_cut_gpu_kernel.cu pair_gpu_precision.h
-	$(CUDA) --ptx -DNV_KERNEL -o $@ ljc_cut_gpu_kernel.cu
+$(OBJ_DIR)/lj_coul.ptx: lj_coul.cu pair_gpu_precision.h
+	$(CUDA) --ptx -DNV_KERNEL -o $@ lj_coul.cu
 
-$(OBJ_DIR)/ljc_cut_gpu_ptx.h: $(OBJ_DIR)/ljc_cut_gpu_kernel.ptx $(OBJ_DIR)/ljc_cut_gpu_kernel.ptx
-	$(BSH) ./geryon/file_to_cstr.sh $(OBJ_DIR)/ljc_cut_gpu_kernel.ptx $(OBJ_DIR)/ljc_cut_gpu_ptx.h
+$(OBJ_DIR)/lj_coul_ext_ptx.h: $(OBJ_DIR)/lj_coul.ptx $(OBJ_DIR)/lj_coul.ptx
+	$(BSH) ./geryon/file_to_cstr.sh $(OBJ_DIR)/lj_coul.ptx $(OBJ_DIR)/lj_coul_ext_ptx.h
 
-$(OBJ_DIR)/ljc_cut_gpu_memory.o: $(ALL_H) ljc_cut_gpu_memory.h ljc_cut_gpu_memory.cpp $(OBJ_DIR)/ljc_cut_gpu_ptx.h $(OBJ_DIR)/charge_gpu_memory.o
-	$(CUDR) -o $@ -c ljc_cut_gpu_memory.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/lj_coul.o: $(ALL_H) lj_coul.h lj_coul.cpp $(OBJ_DIR)/lj_coul_ext_ptx.h $(OBJ_DIR)/charge_gpu_memory.o
+	$(CUDR) -o $@ -c lj_coul.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/ljc_cut_gpu.o: $(ALL_H) ljc_cut_gpu_memory.h ljc_cut_gpu.cpp charge_gpu_memory.h
-	$(CUDR) -o $@ -c ljc_cut_gpu.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/lj_coul_ext.o: $(ALL_H) lj_coul.h lj_coul_ext.cpp charge_gpu_memory.h
+	$(CUDR) -o $@ -c lj_coul_ext.cpp -I$(OBJ_DIR)
 
 $(OBJ_DIR)/lj_class2_long.ptx: lj_class2_long.cu pair_gpu_precision.h
 	$(CUDA) --ptx -DNV_KERNEL -o $@ lj_class2_long.cu
@@ -240,101 +240,101 @@ $(OBJ_DIR)/lj_class2_long.o: $(ALL_H) lj_class2_long.h lj_class2_long.cpp $(OBJ_
 $(OBJ_DIR)/lj_class2_long_ext.o: $(ALL_H) lj_class2_long.h lj_class2_long_ext.cpp charge_gpu_memory.h
 	$(CUDR) -o $@ -c lj_class2_long_ext.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/ljcl_cut_gpu_kernel.ptx: ljcl_cut_gpu_kernel.cu pair_gpu_precision.h
-	$(CUDA) --ptx -DNV_KERNEL -o $@ ljcl_cut_gpu_kernel.cu
+$(OBJ_DIR)/lj_coul_long.ptx: lj_coul_long.cu pair_gpu_precision.h
+	$(CUDA) --ptx -DNV_KERNEL -o $@ lj_coul_long.cu
 
-$(OBJ_DIR)/ljcl_cut_gpu_ptx.h: $(OBJ_DIR)/ljcl_cut_gpu_kernel.ptx $(OBJ_DIR)/ljcl_cut_gpu_kernel.ptx
-	$(BSH) ./geryon/file_to_cstr.sh $(OBJ_DIR)/ljcl_cut_gpu_kernel.ptx $(OBJ_DIR)/ljcl_cut_gpu_ptx.h
+$(OBJ_DIR)/lj_coul_long_ext_ptx.h: $(OBJ_DIR)/lj_coul_long.ptx $(OBJ_DIR)/lj_coul_long.ptx
+	$(BSH) ./geryon/file_to_cstr.sh $(OBJ_DIR)/lj_coul_long.ptx $(OBJ_DIR)/lj_coul_long_ext_ptx.h
 
-$(OBJ_DIR)/ljcl_cut_gpu_memory.o: $(ALL_H) ljcl_cut_gpu_memory.h ljcl_cut_gpu_memory.cpp $(OBJ_DIR)/ljcl_cut_gpu_ptx.h $(OBJ_DIR)/charge_gpu_memory.o
-	$(CUDR) -o $@ -c ljcl_cut_gpu_memory.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/lj_coul_long.o: $(ALL_H) lj_coul_long.h lj_coul_long.cpp $(OBJ_DIR)/lj_coul_long_ext_ptx.h $(OBJ_DIR)/charge_gpu_memory.o
+	$(CUDR) -o $@ -c lj_coul_long.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/ljcl_cut_gpu.o: $(ALL_H) ljcl_cut_gpu_memory.h ljcl_cut_gpu.cpp charge_gpu_memory.h
-	$(CUDR) -o $@ -c ljcl_cut_gpu.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/lj_coul_long_ext.o: $(ALL_H) lj_coul_long.h lj_coul_long_ext.cpp charge_gpu_memory.h
+	$(CUDR) -o $@ -c lj_coul_long_ext.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/morse_gpu_kernel.ptx: morse_gpu_kernel.cu pair_gpu_precision.h
-	$(CUDA) --ptx -DNV_KERNEL -o $@ morse_gpu_kernel.cu
+$(OBJ_DIR)/morse.ptx: morse.cu pair_gpu_precision.h
+	$(CUDA) --ptx -DNV_KERNEL -o $@ morse.cu
 
-$(OBJ_DIR)/morse_gpu_ptx.h: $(OBJ_DIR)/morse_gpu_kernel.ptx $(OBJ_DIR)/morse_gpu_kernel.ptx
-	$(BSH) ./geryon/file_to_cstr.sh $(OBJ_DIR)/morse_gpu_kernel.ptx $(OBJ_DIR)/morse_gpu_ptx.h
+$(OBJ_DIR)/morse_ext_ptx.h: $(OBJ_DIR)/morse.ptx $(OBJ_DIR)/morse.ptx
+	$(BSH) ./geryon/file_to_cstr.sh $(OBJ_DIR)/morse.ptx $(OBJ_DIR)/morse_ext_ptx.h
 
-$(OBJ_DIR)/morse_gpu_memory.o: $(ALL_H) morse_gpu_memory.h morse_gpu_memory.cpp $(OBJ_DIR)/morse_gpu_ptx.h $(OBJ_DIR)/atomic_gpu_memory.o
-	$(CUDR) -o $@ -c morse_gpu_memory.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/morse.o: $(ALL_H) morse.h morse.cpp $(OBJ_DIR)/morse_ext_ptx.h $(OBJ_DIR)/atomic_gpu_memory.o
+	$(CUDR) -o $@ -c morse.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/morse_gpu.o: $(ALL_H) morse_gpu_memory.h morse_gpu.cpp atomic_gpu_memory.h
-	$(CUDR) -o $@ -c morse_gpu.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/morse_ext.o: $(ALL_H) morse.h morse_ext.cpp atomic_gpu_memory.h
+	$(CUDR) -o $@ -c morse_ext.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/crml_gpu_kernel.ptx: crml_gpu_kernel.cu pair_gpu_precision.h
-	$(CUDA) --ptx -DNV_KERNEL -o $@ crml_gpu_kernel.cu
+$(OBJ_DIR)/charmm_long.ptx: charmm_long.cu pair_gpu_precision.h
+	$(CUDA) --ptx -DNV_KERNEL -o $@ charmm_long.cu
 
-$(OBJ_DIR)/crml_gpu_ptx.h: $(OBJ_DIR)/crml_gpu_kernel.ptx $(OBJ_DIR)/crml_gpu_kernel.ptx
-	$(BSH) ./geryon/file_to_cstr.sh $(OBJ_DIR)/crml_gpu_kernel.ptx $(OBJ_DIR)/crml_gpu_ptx.h
+$(OBJ_DIR)/charmm_long_ext_ptx.h: $(OBJ_DIR)/charmm_long.ptx $(OBJ_DIR)/charmm_long.ptx
+	$(BSH) ./geryon/file_to_cstr.sh $(OBJ_DIR)/charmm_long.ptx $(OBJ_DIR)/charmm_long_ext_ptx.h
 
-$(OBJ_DIR)/crml_gpu_memory.o: $(ALL_H) crml_gpu_memory.h crml_gpu_memory.cpp $(OBJ_DIR)/crml_gpu_ptx.h $(OBJ_DIR)/charge_gpu_memory.o
-	$(CUDR) -o $@ -c crml_gpu_memory.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/charmm_long.o: $(ALL_H) charmm_long.h charmm_long.cpp $(OBJ_DIR)/charmm_long_ext_ptx.h $(OBJ_DIR)/charge_gpu_memory.o
+	$(CUDR) -o $@ -c charmm_long.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/crml_gpu.o: $(ALL_H) crml_gpu_memory.h crml_gpu.cpp charge_gpu_memory.h
-	$(CUDR) -o $@ -c crml_gpu.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/charmm_long_ext.o: $(ALL_H) charmm_long.h charmm_long_ext.cpp charge_gpu_memory.h
+	$(CUDR) -o $@ -c charmm_long_ext.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/lj96_cut_gpu_kernel.ptx: lj96_cut_gpu_kernel.cu pair_gpu_precision.h
-	$(CUDA) --ptx -DNV_KERNEL -o $@ lj96_cut_gpu_kernel.cu
+$(OBJ_DIR)/lj96.ptx: lj96.cu pair_gpu_precision.h
+	$(CUDA) --ptx -DNV_KERNEL -o $@ lj96.cu
 
-$(OBJ_DIR)/lj96_cut_gpu_ptx.h: $(OBJ_DIR)/lj96_cut_gpu_kernel.ptx $(OBJ_DIR)/lj96_cut_gpu_kernel.ptx
-	$(BSH) ./geryon/file_to_cstr.sh $(OBJ_DIR)/lj96_cut_gpu_kernel.ptx $(OBJ_DIR)/lj96_cut_gpu_ptx.h
+$(OBJ_DIR)/lj96_ext_ptx.h: $(OBJ_DIR)/lj96.ptx $(OBJ_DIR)/lj96.ptx
+	$(BSH) ./geryon/file_to_cstr.sh $(OBJ_DIR)/lj96.ptx $(OBJ_DIR)/lj96_ext_ptx.h
 
-$(OBJ_DIR)/lj96_cut_gpu_memory.o: $(ALL_H) lj96_cut_gpu_memory.h lj96_cut_gpu_memory.cpp $(OBJ_DIR)/lj96_cut_gpu_ptx.h $(OBJ_DIR)/atomic_gpu_memory.o
-	$(CUDR) -o $@ -c lj96_cut_gpu_memory.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/lj96.o: $(ALL_H) lj96.h lj96.cpp $(OBJ_DIR)/lj96_ext_ptx.h $(OBJ_DIR)/atomic_gpu_memory.o
+	$(CUDR) -o $@ -c lj96.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/lj96_cut_gpu.o: $(ALL_H) lj96_cut_gpu_memory.h lj96_cut_gpu.cpp atomic_gpu_memory.h
-	$(CUDR) -o $@ -c lj96_cut_gpu.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/lj96_ext.o: $(ALL_H) lj96.h lj96_ext.cpp atomic_gpu_memory.h
+	$(CUDR) -o $@ -c lj96_ext.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/lj_expand_gpu_kernel.ptx: lj_expand_gpu_kernel.cu pair_gpu_precision.h
-	$(CUDA) --ptx -DNV_KERNEL -o $@ lj_expand_gpu_kernel.cu
+$(OBJ_DIR)/lj_expand.ptx: lj_expand.cu pair_gpu_precision.h
+	$(CUDA) --ptx -DNV_KERNEL -o $@ lj_expand.cu
 
-$(OBJ_DIR)/lj_expand_gpu_ptx.h: $(OBJ_DIR)/lj_expand_gpu_kernel.ptx $(OBJ_DIR)/lj_expand_gpu_kernel.ptx
-	$(BSH) ./geryon/file_to_cstr.sh $(OBJ_DIR)/lj_expand_gpu_kernel.ptx $(OBJ_DIR)/lj_expand_gpu_ptx.h
+$(OBJ_DIR)/lj_expand_ext_ptx.h: $(OBJ_DIR)/lj_expand.ptx $(OBJ_DIR)/lj_expand.ptx
+	$(BSH) ./geryon/file_to_cstr.sh $(OBJ_DIR)/lj_expand.ptx $(OBJ_DIR)/lj_expand_ext_ptx.h
 
-$(OBJ_DIR)/lj_expand_gpu_memory.o: $(ALL_H) lj_expand_gpu_memory.h lj_expand_gpu_memory.cpp $(OBJ_DIR)/lj_expand_gpu_ptx.h $(OBJ_DIR)/atomic_gpu_memory.o
-	$(CUDR) -o $@ -c lj_expand_gpu_memory.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/lj_expand.o: $(ALL_H) lj_expand.h lj_expand.cpp $(OBJ_DIR)/lj_expand_ext_ptx.h $(OBJ_DIR)/atomic_gpu_memory.o
+	$(CUDR) -o $@ -c lj_expand.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/lj_expand_gpu.o: $(ALL_H) lj_expand_gpu_memory.h lj_expand_gpu.cpp atomic_gpu_memory.h
-	$(CUDR) -o $@ -c lj_expand_gpu.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/lj_expand_ext.o: $(ALL_H) lj_expand.h lj_expand_ext.cpp atomic_gpu_memory.h
+	$(CUDR) -o $@ -c lj_expand_ext.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/cmm_cut_gpu_kernel.ptx: cmm_cut_gpu_kernel.cu pair_gpu_precision.h
-	$(CUDA) --ptx -DNV_KERNEL -o $@ cmm_cut_gpu_kernel.cu
+$(OBJ_DIR)/cg_cmm.ptx: cg_cmm.cu pair_gpu_precision.h
+	$(CUDA) --ptx -DNV_KERNEL -o $@ cg_cmm.cu
 
-$(OBJ_DIR)/cmm_cut_gpu_ptx.h: $(OBJ_DIR)/cmm_cut_gpu_kernel.ptx $(OBJ_DIR)/cmm_cut_gpu_kernel.ptx
-	$(BSH) ./geryon/file_to_cstr.sh $(OBJ_DIR)/cmm_cut_gpu_kernel.ptx $(OBJ_DIR)/cmm_cut_gpu_ptx.h
+$(OBJ_DIR)/cg_cmm_ext_ptx.h: $(OBJ_DIR)/cg_cmm.ptx $(OBJ_DIR)/cg_cmm.ptx
+	$(BSH) ./geryon/file_to_cstr.sh $(OBJ_DIR)/cg_cmm.ptx $(OBJ_DIR)/cg_cmm_ext_ptx.h
 
-$(OBJ_DIR)/cmm_cut_gpu_memory.o: $(ALL_H) cmm_cut_gpu_memory.h cmm_cut_gpu_memory.cpp $(OBJ_DIR)/cmm_cut_gpu_ptx.h $(OBJ_DIR)/atomic_gpu_memory.o
-	$(CUDR) -o $@ -c cmm_cut_gpu_memory.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/cg_cmm.o: $(ALL_H) cg_cmm.h cg_cmm.cpp $(OBJ_DIR)/cg_cmm_ext_ptx.h $(OBJ_DIR)/atomic_gpu_memory.o
+	$(CUDR) -o $@ -c cg_cmm.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/cmm_cut_gpu.o: $(ALL_H) cmm_cut_gpu_memory.h cmm_cut_gpu.cpp atomic_gpu_memory.h
-	$(CUDR) -o $@ -c cmm_cut_gpu.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/cg_cmm_ext.o: $(ALL_H) cg_cmm.h cg_cmm_ext.cpp atomic_gpu_memory.h
+	$(CUDR) -o $@ -c cg_cmm_ext.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/cmmc_long_gpu_kernel.ptx: cmmc_long_gpu_kernel.cu pair_gpu_precision.h
-	$(CUDA) --ptx -DNV_KERNEL -o $@ cmmc_long_gpu_kernel.cu
+$(OBJ_DIR)/cg_cmm_long.ptx: cg_cmm_long.cu pair_gpu_precision.h
+	$(CUDA) --ptx -DNV_KERNEL -o $@ cg_cmm_long.cu
 
-$(OBJ_DIR)/cmmc_long_gpu_ptx.h: $(OBJ_DIR)/cmmc_long_gpu_kernel.ptx $(OBJ_DIR)/cmmc_long_gpu_kernel.ptx
-	$(BSH) ./geryon/file_to_cstr.sh $(OBJ_DIR)/cmmc_long_gpu_kernel.ptx $(OBJ_DIR)/cmmc_long_gpu_ptx.h
+$(OBJ_DIR)/cg_cmm_long_ext_ptx.h: $(OBJ_DIR)/cg_cmm_long.ptx $(OBJ_DIR)/cg_cmm_long.ptx
+	$(BSH) ./geryon/file_to_cstr.sh $(OBJ_DIR)/cg_cmm_long.ptx $(OBJ_DIR)/cg_cmm_long_ext_ptx.h
 
-$(OBJ_DIR)/cmmc_long_gpu_memory.o: $(ALL_H) cmmc_long_gpu_memory.h cmmc_long_gpu_memory.cpp $(OBJ_DIR)/cmmc_long_gpu_ptx.h $(OBJ_DIR)/atomic_gpu_memory.o
-	$(CUDR) -o $@ -c cmmc_long_gpu_memory.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/cg_cmm_long.o: $(ALL_H) cg_cmm_long.h cg_cmm_long.cpp $(OBJ_DIR)/cg_cmm_long_ext_ptx.h $(OBJ_DIR)/atomic_gpu_memory.o
+	$(CUDR) -o $@ -c cg_cmm_long.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/cmmc_long_gpu.o: $(ALL_H) cmmc_long_gpu_memory.h cmmc_long_gpu.cpp charge_gpu_memory.h
-	$(CUDR) -o $@ -c cmmc_long_gpu.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/cg_cmm_long_ext.o: $(ALL_H) cg_cmm_long.h cg_cmm_long_ext.cpp charge_gpu_memory.h
+	$(CUDR) -o $@ -c cg_cmm_long_ext.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/cmmc_msm_gpu_kernel.ptx: cmmc_msm_gpu_kernel.cu pair_gpu_precision.h
-	$(CUDA) --ptx -DNV_KERNEL -o $@ cmmc_msm_gpu_kernel.cu
+$(OBJ_DIR)/cg_cmm_msm.ptx: cg_cmm_msm.cu pair_gpu_precision.h
+	$(CUDA) --ptx -DNV_KERNEL -o $@ cg_cmm_msm.cu
 
-$(OBJ_DIR)/cmmc_msm_gpu_ptx.h: $(OBJ_DIR)/cmmc_msm_gpu_kernel.ptx $(OBJ_DIR)/cmmc_msm_gpu_kernel.ptx
-	$(BSH) ./geryon/file_to_cstr.sh $(OBJ_DIR)/cmmc_msm_gpu_kernel.ptx $(OBJ_DIR)/cmmc_msm_gpu_ptx.h
+$(OBJ_DIR)/cg_cmm_msm_ext_ptx.h: $(OBJ_DIR)/cg_cmm_msm.ptx $(OBJ_DIR)/cg_cmm_msm.ptx
+	$(BSH) ./geryon/file_to_cstr.sh $(OBJ_DIR)/cg_cmm_msm.ptx $(OBJ_DIR)/cg_cmm_msm_ext_ptx.h
 
-$(OBJ_DIR)/cmmc_msm_gpu_memory.o: $(ALL_H) cmmc_msm_gpu_memory.h cmmc_msm_gpu_memory.cpp $(OBJ_DIR)/cmmc_msm_gpu_ptx.h $(OBJ_DIR)/atomic_gpu_memory.o
-	$(CUDR) -o $@ -c cmmc_msm_gpu_memory.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/cg_cmm_msm.o: $(ALL_H) cg_cmm_msm.h cg_cmm_msm.cpp $(OBJ_DIR)/cg_cmm_msm_ext_ptx.h $(OBJ_DIR)/atomic_gpu_memory.o
+	$(CUDR) -o $@ -c cg_cmm_msm.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/cmmc_msm_gpu.o: $(ALL_H) cmmc_msm_gpu_memory.h cmmc_msm_gpu.cpp charge_gpu_memory.h
-	$(CUDR) -o $@ -c cmmc_msm_gpu.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/cg_cmm_msm_ext.o: $(ALL_H) cg_cmm_msm.h cg_cmm_msm_ext.cpp charge_gpu_memory.h
+	$(CUDR) -o $@ -c cg_cmm_msm_ext.cpp -I$(OBJ_DIR)
 
 $(BIN_DIR)/nvc_get_devices: ./geryon/ucl_get_devices.cpp $(NVC_H)
 	$(CUDR) -o $@ ./geryon/ucl_get_devices.cpp -DUCL_CUDART $(CUDA_LINK) 
diff --git a/lib/gpu/Opencl.makefile b/lib/gpu/Opencl.makefile
index 26f5fdc19c..3abe63bd91 100644
--- a/lib/gpu/Opencl.makefile
+++ b/lib/gpu/Opencl.makefile
@@ -26,7 +26,7 @@ OCL_H  = $(wildcard ./geryon/ocl*.h) $(UCL_H)
 # Headers for Pair Stuff
 PAIR_H  = pair_gpu_atom.h pair_gpu_ans.h pair_gpu_nbor_shared.h \
           pair_gpu_nbor.h pair_gpu_precision.h pair_gpu_device.h \
-          pair_gpu_balance.h pppm_gpu_memory.h
+          pair_gpu_balance.h pppm.h
 
 ALL_H = $(OCL_H) $(PAIR_H)
 
@@ -35,29 +35,29 @@ OBJS = $(OBJ_DIR)/pair_gpu_atom.o $(OBJ_DIR)/pair_gpu_ans.o \
        $(OBJ_DIR)/pair_gpu_nbor_shared.o $(OBJ_DIR)/pair_gpu_nbor.o \
        $(OBJ_DIR)/pair_gpu_device.o $(OBJ_DIR)/atomic_gpu_memory.o \
        $(OBJ_DIR)/charge_gpu_memory.o $(OBJ_DIR)/base_ellipsoid.o \
-       $(OBJ_DIR)/pppm_gpu_memory.o $(OBJ_DIR)/pppm_l_gpu.o \
+       $(OBJ_DIR)/pppm.o $(OBJ_DIR)/pppm_ext.o \
        $(OBJ_DIR)/gayberne.o $(OBJ_DIR)/gayberne_ext.o \
        $(OBJ_DIR)/re_squared.o $(OBJ_DIR)/re_squared_ext.o \
-       $(OBJ_DIR)/lj_cut_gpu_memory.o $(OBJ_DIR)/lj_cut_gpu.o \
-       $(OBJ_DIR)/lj96_cut_gpu_memory.o $(OBJ_DIR)/lj96_cut_gpu.o \
-       $(OBJ_DIR)/lj_expand_gpu_memory.o $(OBJ_DIR)/lj_expand_gpu.o \
-       $(OBJ_DIR)/ljc_cut_gpu_memory.o $(OBJ_DIR)/ljc_cut_gpu.o \
-       $(OBJ_DIR)/ljcl_cut_gpu_memory.o $(OBJ_DIR)/ljcl_cut_gpu.o \
+       $(OBJ_DIR)/lj.o $(OBJ_DIR)/lj_ext.o \
+       $(OBJ_DIR)/lj96.o $(OBJ_DIR)/lj96_ext.o \
+       $(OBJ_DIR)/lj_expand.o $(OBJ_DIR)/lj_expand_ext.o \
+       $(OBJ_DIR)/lj_coul.o $(OBJ_DIR)/lj_coul_ext.o \
+       $(OBJ_DIR)/lj_coul_long.o $(OBJ_DIR)/lj_coul_long_ext.o \
        $(OBJ_DIR)/lj_class2_long.o $(OBJ_DIR)/lj_class2_long_ext.o \
-       $(OBJ_DIR)/morse_gpu_memory.o $(OBJ_DIR)/morse_gpu.o \
-       $(OBJ_DIR)/crml_gpu_memory.o $(OBJ_DIR)/crml_gpu.o \
-       $(OBJ_DIR)/cmm_cut_gpu_memory.o $(OBJ_DIR)/cmm_cut_gpu.o \
-       $(OBJ_DIR)/cmmc_long_gpu_memory.o $(OBJ_DIR)/cmmc_long_gpu.o 
+       $(OBJ_DIR)/morse.o $(OBJ_DIR)/morse_ext.o \
+       $(OBJ_DIR)/charmm_long.o $(OBJ_DIR)/charmm_long_ext.o \
+       $(OBJ_DIR)/cg_cmm.o $(OBJ_DIR)/cg_cmm_ext.o \
+       $(OBJ_DIR)/cg_cmm_long.o $(OBJ_DIR)/cg_cmm_long_ext.o 
 KERS = $(OBJ_DIR)/pair_gpu_dev_cl.h $(OBJ_DIR)/pair_gpu_atom_cl.h \
        $(OBJ_DIR)/pair_gpu_nbor_cl.h $(OBJ_DIR)/pppm_gpu_cl.h \
        $(OBJ_DIR)/ellipsoid_nbor_cl.h $(OBJ_DIR)/gayberne_cl.h \
        $(OBJ_DIR)/re_squared_cl.h \
-       $(OBJ_DIR)/lj_cut_gpu_cl.h $(OBJ_DIR)/lj96_cut_gpu_cl.h \
-       $(OBJ_DIR)/lj_expand_gpu_cl.h $(OBJ_DIR)/ljc_cut_gpu_cl.h \
-       $(OBJ_DIR)/ljcl_cut_gpu_cl.h $(OBJ_DIR)/lj_class2_long_cl.h \
-       $(OBJ_DIR)/morse_gpu_cl.h \
-       $(OBJ_DIR)/crml_gpu_cl.h $(OBJ_DIR)/cmm_cut_gpu_cl.h \
-       $(OBJ_DIR)/cmmc_long_gpu_cl.h 
+       $(OBJ_DIR)/lj_ext_cl.h $(OBJ_DIR)/lj96_ext_cl.h \
+       $(OBJ_DIR)/lj_expand_ext_cl.h $(OBJ_DIR)/lj_coul_ext_cl.h \
+       $(OBJ_DIR)/lj_coul_long_ext_cl.h $(OBJ_DIR)/lj_class2_long_cl.h \
+       $(OBJ_DIR)/morse_ext_cl.h \
+       $(OBJ_DIR)/charmm_long_ext_cl.h $(OBJ_DIR)/cg_cmm_ext_cl.h \
+       $(OBJ_DIR)/cg_cmm_long_ext_cl.h 
 
 OCL_EXECS = $(BIN_DIR)/ocl_get_devices
 
@@ -96,14 +96,14 @@ $(OBJ_DIR)/charge_gpu_memory.o: $(OCL_H) charge_gpu_memory.h charge_gpu_memory.c
 $(OBJ_DIR)/base_ellipsoid.o: $(OCL_H) base_ellipsoid.h base_ellipsoid.cpp $(OBJ_DIR)/ellipsoid_nbor_cl.h
 	$(OCL) -o $@ -c base_ellipsoid.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/pppm_gpu_cl.h: pppm_gpu_kernel.cu
-	$(BSH) ./geryon/file_to_cstr.sh pppm_gpu_kernel.cu $(OBJ_DIR)/pppm_gpu_cl.h;
+$(OBJ_DIR)/pppm_gpu_cl.h: pppm.cu
+	$(BSH) ./geryon/file_to_cstr.sh pppm.cu $(OBJ_DIR)/pppm_gpu_cl.h;
 
-$(OBJ_DIR)/pppm_gpu_memory.o: $(ALL_H) pppm_gpu_memory.h pppm_gpu_memory.cpp  $(OBJ_DIR)/pppm_gpu_cl.h $(OBJ_DIR)/pppm_gpu_cl.h
-	$(OCL) -o $@ -c pppm_gpu_memory.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/pppm.o: $(ALL_H) pppm.h pppm.cpp  $(OBJ_DIR)/pppm_gpu_cl.h $(OBJ_DIR)/pppm_gpu_cl.h
+	$(OCL) -o $@ -c pppm.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/pppm_l_gpu.o: $(ALL_H) pppm_gpu_memory.h pppm_l_gpu.cpp
-	$(OCL) -o $@ -c pppm_l_gpu.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/pppm_ext.o: $(ALL_H) pppm.h pppm_ext.cpp
+	$(OCL) -o $@ -c pppm_ext.cpp -I$(OBJ_DIR)
 
 $(OBJ_DIR)/ellipsoid_nbor_cl.h: ellipsoid_nbor.cu
 	$(BSH) ./geryon/file_to_cstr.sh ellipsoid_nbor.cu $(OBJ_DIR)/ellipsoid_nbor_cl.h
@@ -132,32 +132,32 @@ $(OBJ_DIR)/re_squared.o: $(ALL_H) re_squared.h re_squared.cpp $(OBJ_DIR)/re_squa
 $(OBJ_DIR)/re_squared_ext.o: $(ALL_H) $(OBJ_DIR)/re_squared.o re_squared_ext.cpp
 	$(OCL) -o $@ -c re_squared_ext.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/lj_cut_gpu_cl.h: lj_cut_gpu_kernel.cu
-	$(BSH) ./geryon/file_to_cstr.sh lj_cut_gpu_kernel.cu $(OBJ_DIR)/lj_cut_gpu_cl.h;
+$(OBJ_DIR)/lj_ext_cl.h: lj.cu
+	$(BSH) ./geryon/file_to_cstr.sh lj.cu $(OBJ_DIR)/lj_ext_cl.h;
 
-$(OBJ_DIR)/lj_cut_gpu_memory.o: $(ALL_H) lj_cut_gpu_memory.h lj_cut_gpu_memory.cpp  $(OBJ_DIR)/lj_cut_gpu_cl.h $(OBJ_DIR)/pair_gpu_nbor_cl.h $(OBJ_DIR)/lj_cut_gpu_cl.h $(OBJ_DIR)/atomic_gpu_memory.o
-	$(OCL) -o $@ -c lj_cut_gpu_memory.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/lj.o: $(ALL_H) lj.h lj.cpp  $(OBJ_DIR)/lj_ext_cl.h $(OBJ_DIR)/pair_gpu_nbor_cl.h $(OBJ_DIR)/lj_ext_cl.h $(OBJ_DIR)/atomic_gpu_memory.o
+	$(OCL) -o $@ -c lj.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/lj_cut_gpu.o: $(ALL_H) lj_cut_gpu_memory.h lj_cut_gpu.cpp atomic_gpu_memory.h
-	$(OCL) -o $@ -c lj_cut_gpu.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/lj_ext.o: $(ALL_H) lj.h lj_ext.cpp atomic_gpu_memory.h
+	$(OCL) -o $@ -c lj_ext.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/ljc_cut_gpu_cl.h: ljc_cut_gpu_kernel.cu
-	$(BSH) ./geryon/file_to_cstr.sh ljc_cut_gpu_kernel.cu $(OBJ_DIR)/ljc_cut_gpu_cl.h;
+$(OBJ_DIR)/lj_coul_ext_cl.h: lj_coul.cu
+	$(BSH) ./geryon/file_to_cstr.sh lj_coul.cu $(OBJ_DIR)/lj_coul_ext_cl.h;
 
-$(OBJ_DIR)/ljc_cut_gpu_memory.o: $(ALL_H) ljc_cut_gpu_memory.h ljc_cut_gpu_memory.cpp  $(OBJ_DIR)/ljc_cut_gpu_cl.h $(OBJ_DIR)/pair_gpu_nbor_cl.h $(OBJ_DIR)/ljc_cut_gpu_cl.h $(OBJ_DIR)/charge_gpu_memory.o
-	$(OCL) -o $@ -c ljc_cut_gpu_memory.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/lj_coul.o: $(ALL_H) lj_coul.h lj_coul.cpp  $(OBJ_DIR)/lj_coul_ext_cl.h $(OBJ_DIR)/pair_gpu_nbor_cl.h $(OBJ_DIR)/lj_coul_ext_cl.h $(OBJ_DIR)/charge_gpu_memory.o
+	$(OCL) -o $@ -c lj_coul.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/ljc_cut_gpu.o: $(ALL_H) ljc_cut_gpu_memory.h ljc_cut_gpu.cpp charge_gpu_memory.h
-	$(OCL) -o $@ -c ljc_cut_gpu.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/lj_coul_ext.o: $(ALL_H) lj_coul.h lj_coul_ext.cpp charge_gpu_memory.h
+	$(OCL) -o $@ -c lj_coul_ext.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/ljcl_cut_gpu_cl.h: ljcl_cut_gpu_kernel.cu
-	$(BSH) ./geryon/file_to_cstr.sh ljcl_cut_gpu_kernel.cu $(OBJ_DIR)/ljcl_cut_gpu_cl.h;
+$(OBJ_DIR)/lj_coul_long_ext_cl.h: lj_coul_long.cu
+	$(BSH) ./geryon/file_to_cstr.sh lj_coul_long.cu $(OBJ_DIR)/lj_coul_long_ext_cl.h;
 
-$(OBJ_DIR)/ljcl_cut_gpu_memory.o: $(ALL_H) ljcl_cut_gpu_memory.h ljcl_cut_gpu_memory.cpp  $(OBJ_DIR)/ljcl_cut_gpu_cl.h $(OBJ_DIR)/pair_gpu_nbor_cl.h $(OBJ_DIR)/charge_gpu_memory.o
-	$(OCL) -o $@ -c ljcl_cut_gpu_memory.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/lj_coul_long.o: $(ALL_H) lj_coul_long.h lj_coul_long.cpp  $(OBJ_DIR)/lj_coul_long_ext_cl.h $(OBJ_DIR)/pair_gpu_nbor_cl.h $(OBJ_DIR)/charge_gpu_memory.o
+	$(OCL) -o $@ -c lj_coul_long.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/ljcl_cut_gpu.o: $(ALL_H) ljcl_cut_gpu_memory.h ljcl_cut_gpu.cpp charge_gpu_memory.h
-	$(OCL) -o $@ -c ljcl_cut_gpu.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/lj_coul_long_ext.o: $(ALL_H) lj_coul_long.h lj_coul_long_ext.cpp charge_gpu_memory.h
+	$(OCL) -o $@ -c lj_coul_long_ext.cpp -I$(OBJ_DIR)
 
 $(OBJ_DIR)/lj_class2_long_cl.h: lj_class2_long.cu
 	$(BSH) ./geryon/file_to_cstr.sh lj_class2_long.cu $(OBJ_DIR)/lj_class2_long_cl.h;
@@ -168,59 +168,59 @@ $(OBJ_DIR)/lj_class2_long.o: $(ALL_H) lj_class2_long.h lj_class2_long.cpp  $(OBJ
 $(OBJ_DIR)/lj_class2_long_ext.o: $(ALL_H) lj_class2_long.h lj_class2_long_ext.cpp charge_gpu_memory.h
 	$(OCL) -o $@ -c lj_class2_long_ext.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/morse_gpu_cl.h: morse_gpu_kernel.cu
-	$(BSH) ./geryon/file_to_cstr.sh morse_gpu_kernel.cu $(OBJ_DIR)/morse_gpu_cl.h;
+$(OBJ_DIR)/morse_ext_cl.h: morse.cu
+	$(BSH) ./geryon/file_to_cstr.sh morse.cu $(OBJ_DIR)/morse_ext_cl.h;
 
-$(OBJ_DIR)/morse_gpu_memory.o: $(ALL_H) morse_gpu_memory.h morse_gpu_memory.cpp  $(OBJ_DIR)/morse_gpu_cl.h $(OBJ_DIR)/pair_gpu_nbor_cl.h $(OBJ_DIR)/morse_gpu_cl.h $(OBJ_DIR)/atomic_gpu_memory.o
-	$(OCL) -o $@ -c morse_gpu_memory.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/morse.o: $(ALL_H) morse.h morse.cpp  $(OBJ_DIR)/morse_ext_cl.h $(OBJ_DIR)/pair_gpu_nbor_cl.h $(OBJ_DIR)/morse_ext_cl.h $(OBJ_DIR)/atomic_gpu_memory.o
+	$(OCL) -o $@ -c morse.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/morse_gpu.o: $(ALL_H) morse_gpu_memory.h morse_gpu.cpp atomic_gpu_memory.h
-	$(OCL) -o $@ -c morse_gpu.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/morse_ext.o: $(ALL_H) morse.h morse_ext.cpp atomic_gpu_memory.h
+	$(OCL) -o $@ -c morse_ext.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/crml_gpu_cl.h: crml_gpu_kernel.cu
-	$(BSH) ./geryon/file_to_cstr.sh crml_gpu_kernel.cu $(OBJ_DIR)/crml_gpu_cl.h;
+$(OBJ_DIR)/charmm_long_ext_cl.h: charmm_long.cu
+	$(BSH) ./geryon/file_to_cstr.sh charmm_long.cu $(OBJ_DIR)/charmm_long_ext_cl.h;
 
-$(OBJ_DIR)/crml_gpu_memory.o: $(ALL_H) crml_gpu_memory.h crml_gpu_memory.cpp  $(OBJ_DIR)/crml_gpu_cl.h $(OBJ_DIR)/pair_gpu_nbor_cl.h $(OBJ_DIR)/crml_gpu_cl.h $(OBJ_DIR)/charge_gpu_memory.o
-	$(OCL) -o $@ -c crml_gpu_memory.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/charmm_long.o: $(ALL_H) charmm_long.h charmm_long.cpp  $(OBJ_DIR)/charmm_long_ext_cl.h $(OBJ_DIR)/pair_gpu_nbor_cl.h $(OBJ_DIR)/charmm_long_ext_cl.h $(OBJ_DIR)/charge_gpu_memory.o
+	$(OCL) -o $@ -c charmm_long.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/crml_gpu.o: $(ALL_H) crml_gpu_memory.h crml_gpu.cpp charge_gpu_memory.h
-	$(OCL) -o $@ -c crml_gpu.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/charmm_long_ext.o: $(ALL_H) charmm_long.h charmm_long_ext.cpp charge_gpu_memory.h
+	$(OCL) -o $@ -c charmm_long_ext.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/lj96_cut_gpu_cl.h: lj96_cut_gpu_kernel.cu
-	$(BSH) ./geryon/file_to_cstr.sh lj96_cut_gpu_kernel.cu $(OBJ_DIR)/lj96_cut_gpu_cl.h;
+$(OBJ_DIR)/lj96_ext_cl.h: lj96.cu
+	$(BSH) ./geryon/file_to_cstr.sh lj96.cu $(OBJ_DIR)/lj96_ext_cl.h;
 
-$(OBJ_DIR)/lj96_cut_gpu_memory.o: $(ALL_H) lj96_cut_gpu_memory.h lj96_cut_gpu_memory.cpp  $(OBJ_DIR)/lj96_cut_gpu_cl.h $(OBJ_DIR)/pair_gpu_nbor_cl.h $(OBJ_DIR)/lj96_cut_gpu_cl.h $(OBJ_DIR)/atomic_gpu_memory.o
-	$(OCL) -o $@ -c lj96_cut_gpu_memory.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/lj96.o: $(ALL_H) lj96.h lj96.cpp  $(OBJ_DIR)/lj96_ext_cl.h $(OBJ_DIR)/pair_gpu_nbor_cl.h $(OBJ_DIR)/lj96_ext_cl.h $(OBJ_DIR)/atomic_gpu_memory.o
+	$(OCL) -o $@ -c lj96.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/lj96_cut_gpu.o: $(ALL_H) lj96_cut_gpu_memory.h lj96_cut_gpu.cpp atomic_gpu_memory.h
-	$(OCL) -o $@ -c lj96_cut_gpu.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/lj96_ext.o: $(ALL_H) lj96.h lj96_ext.cpp atomic_gpu_memory.h
+	$(OCL) -o $@ -c lj96_ext.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/lj_expand_gpu_cl.h: lj_expand_gpu_kernel.cu
-	$(BSH) ./geryon/file_to_cstr.sh lj_expand_gpu_kernel.cu $(OBJ_DIR)/lj_expand_gpu_cl.h;
+$(OBJ_DIR)/lj_expand_ext_cl.h: lj_expand.cu
+	$(BSH) ./geryon/file_to_cstr.sh lj_expand.cu $(OBJ_DIR)/lj_expand_ext_cl.h;
 
-$(OBJ_DIR)/lj_expand_gpu_memory.o: $(ALL_H) lj_expand_gpu_memory.h lj_expand_gpu_memory.cpp  $(OBJ_DIR)/lj_expand_gpu_cl.h $(OBJ_DIR)/pair_gpu_nbor_cl.h $(OBJ_DIR)/lj_expand_gpu_cl.h $(OBJ_DIR)/atomic_gpu_memory.o
-	$(OCL) -o $@ -c lj_expand_gpu_memory.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/lj_expand.o: $(ALL_H) lj_expand.h lj_expand.cpp  $(OBJ_DIR)/lj_expand_ext_cl.h $(OBJ_DIR)/pair_gpu_nbor_cl.h $(OBJ_DIR)/lj_expand_ext_cl.h $(OBJ_DIR)/atomic_gpu_memory.o
+	$(OCL) -o $@ -c lj_expand.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/lj_expand_gpu.o: $(ALL_H) lj_expand_gpu_memory.h lj_expand_gpu.cpp atomic_gpu_memory.h
-	$(OCL) -o $@ -c lj_expand_gpu.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/lj_expand_ext.o: $(ALL_H) lj_expand.h lj_expand_ext.cpp atomic_gpu_memory.h
+	$(OCL) -o $@ -c lj_expand_ext.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/cmm_cut_gpu_cl.h: cmm_cut_gpu_kernel.cu
-	$(BSH) ./geryon/file_to_cstr.sh cmm_cut_gpu_kernel.cu $(OBJ_DIR)/cmm_cut_gpu_cl.h;
+$(OBJ_DIR)/cg_cmm_ext_cl.h: cg_cmm.cu
+	$(BSH) ./geryon/file_to_cstr.sh cg_cmm.cu $(OBJ_DIR)/cg_cmm_ext_cl.h;
 
-$(OBJ_DIR)/cmm_cut_gpu_memory.o: $(ALL_H) cmm_cut_gpu_memory.h cmm_cut_gpu_memory.cpp  $(OBJ_DIR)/cmm_cut_gpu_cl.h $(OBJ_DIR)/pair_gpu_nbor_cl.h $(OBJ_DIR)/cmm_cut_gpu_cl.h $(OBJ_DIR)/atomic_gpu_memory.o
-	$(OCL) -o $@ -c cmm_cut_gpu_memory.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/cg_cmm.o: $(ALL_H) cg_cmm.h cg_cmm.cpp  $(OBJ_DIR)/cg_cmm_ext_cl.h $(OBJ_DIR)/pair_gpu_nbor_cl.h $(OBJ_DIR)/cg_cmm_ext_cl.h $(OBJ_DIR)/atomic_gpu_memory.o
+	$(OCL) -o $@ -c cg_cmm.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/cmm_cut_gpu.o: $(ALL_H) cmm_cut_gpu_memory.h cmm_cut_gpu.cpp atomic_gpu_memory.h
-	$(OCL) -o $@ -c cmm_cut_gpu.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/cg_cmm_ext.o: $(ALL_H) cg_cmm.h cg_cmm_ext.cpp atomic_gpu_memory.h
+	$(OCL) -o $@ -c cg_cmm_ext.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/cmmc_long_gpu_cl.h: cmmc_long_gpu_kernel.cu
-	$(BSH) ./geryon/file_to_cstr.sh cmmc_long_gpu_kernel.cu $(OBJ_DIR)/cmmc_long_gpu_cl.h;
+$(OBJ_DIR)/cg_cmm_long_ext_cl.h: cg_cmm_long.cu
+	$(BSH) ./geryon/file_to_cstr.sh cg_cmm_long.cu $(OBJ_DIR)/cg_cmm_long_ext_cl.h;
 
-$(OBJ_DIR)/cmmc_long_gpu_memory.o: $(ALL_H) cmmc_long_gpu_memory.h cmmc_long_gpu_memory.cpp  $(OBJ_DIR)/cmmc_long_gpu_cl.h $(OBJ_DIR)/pair_gpu_nbor_cl.h $(OBJ_DIR)/cmmc_long_gpu_cl.h $(OBJ_DIR)/atomic_gpu_memory.o
-	$(OCL) -o $@ -c cmmc_long_gpu_memory.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/cg_cmm_long.o: $(ALL_H) cg_cmm_long.h cg_cmm_long.cpp  $(OBJ_DIR)/cg_cmm_long_ext_cl.h $(OBJ_DIR)/pair_gpu_nbor_cl.h $(OBJ_DIR)/cg_cmm_long_ext_cl.h $(OBJ_DIR)/atomic_gpu_memory.o
+	$(OCL) -o $@ -c cg_cmm_long.cpp -I$(OBJ_DIR)
 
-$(OBJ_DIR)/cmmc_long_gpu.o: $(ALL_H) cmmc_long_gpu_memory.h cmmc_long_gpu.cpp charge_gpu_memory.h
-	$(OCL) -o $@ -c cmmc_long_gpu.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/cg_cmm_long_ext.o: $(ALL_H) cg_cmm_long.h cg_cmm_long_ext.cpp charge_gpu_memory.h
+	$(OCL) -o $@ -c cg_cmm_long_ext.cpp -I$(OBJ_DIR)
 
 $(BIN_DIR)/ocl_get_devices: ./geryon/ucl_get_devices.cpp
 	$(OCL) -o $@ ./geryon/ucl_get_devices.cpp -DUCL_OPENCL $(OCL_LINK) 
diff --git a/lib/gpu/cmm_cut_gpu_memory.cpp b/lib/gpu/cg_cmm.cpp
similarity index 97%
rename from lib/gpu/cmm_cut_gpu_memory.cpp
rename to lib/gpu/cg_cmm.cpp
index 8d18c99d7c..370acbc5e1 100644
--- a/lib/gpu/cmm_cut_gpu_memory.cpp
+++ b/lib/gpu/cg_cmm.cpp
@@ -16,12 +16,12 @@
 ------------------------------------------------------------------------- */
 
 #ifdef USE_OPENCL
-#include "cmm_cut_gpu_cl.h"
+#include "cg_cmm_ext_cl.h"
 #else
-#include "cmm_cut_gpu_ptx.h"
+#include "cg_cmm_ext_ptx.h"
 #endif
 
-#include "cmm_cut_gpu_memory.h"
+#include "cg_cmm.h"
 #include <cassert>
 #define CMM_GPU_MemoryT CMM_GPU_Memory<numtyp, acctyp>
 
@@ -52,7 +52,7 @@ int CMM_GPU_MemoryT::init(const int ntypes, double **host_cutsq,
                           const double gpu_split, FILE *_screen) {
   int success;
   success=this->init_atomic(nlocal,nall,max_nbors,maxspecial,cell_size,gpu_split,
-                            _screen,cmm_cut_gpu_kernel);
+                            _screen,cg_cmm);
   if (success!=0)
     return success;
 
diff --git a/lib/gpu/cmm_cut_gpu_kernel.cu b/lib/gpu/cg_cmm.cu
similarity index 100%
rename from lib/gpu/cmm_cut_gpu_kernel.cu
rename to lib/gpu/cg_cmm.cu
diff --git a/lib/gpu/cmm_cut_gpu_memory.h b/lib/gpu/cg_cmm.h
similarity index 100%
rename from lib/gpu/cmm_cut_gpu_memory.h
rename to lib/gpu/cg_cmm.h
diff --git a/lib/gpu/cmm_cut_gpu.cpp b/lib/gpu/cg_cmm_ext.cpp
similarity index 99%
rename from lib/gpu/cmm_cut_gpu.cpp
rename to lib/gpu/cg_cmm_ext.cpp
index 7be958615a..519fad118c 100644
--- a/lib/gpu/cmm_cut_gpu.cpp
+++ b/lib/gpu/cg_cmm_ext.cpp
@@ -19,7 +19,7 @@
 #include <cassert>
 #include <math.h>
 
-#include "cmm_cut_gpu_memory.h"
+#include "cg_cmm.h"
 
 using namespace std;
 
diff --git a/lib/gpu/cmmc_long_gpu_memory.cpp b/lib/gpu/cg_cmm_long.cpp
similarity index 97%
rename from lib/gpu/cmmc_long_gpu_memory.cpp
rename to lib/gpu/cg_cmm_long.cpp
index c6291982d6..1bb4fab4c2 100644
--- a/lib/gpu/cmmc_long_gpu_memory.cpp
+++ b/lib/gpu/cg_cmm_long.cpp
@@ -16,12 +16,12 @@
 ------------------------------------------------------------------------- */
 
 #ifdef USE_OPENCL
-#include "cmmc_long_gpu_cl.h"
+#include "cg_cmm_long_ext_cl.h"
 #else
-#include "cmmc_long_gpu_ptx.h"
+#include "cg_cmm_long_ext_ptx.h"
 #endif
 
-#include "cmmc_long_gpu_memory.h"
+#include "cg_cmm_long.h"
 #include <cassert>
 #define CMML_GPU_MemoryT CMML_GPU_Memory<numtyp, acctyp>
 
@@ -57,7 +57,7 @@ int CMML_GPU_MemoryT::init(const int ntypes, double **host_cutsq,
                            const double g_ewald) {
   int success;
   success=this->init_atomic(nlocal,nall,max_nbors,maxspecial,cell_size,gpu_split,
-                            _screen,cmmc_long_gpu_kernel);
+                            _screen,cg_cmm_long);
   if (success!=0)
     return success;
 
diff --git a/lib/gpu/cmmc_long_gpu_kernel.cu b/lib/gpu/cg_cmm_long.cu
similarity index 100%
rename from lib/gpu/cmmc_long_gpu_kernel.cu
rename to lib/gpu/cg_cmm_long.cu
diff --git a/lib/gpu/cmmc_long_gpu_memory.h b/lib/gpu/cg_cmm_long.h
similarity index 100%
rename from lib/gpu/cmmc_long_gpu_memory.h
rename to lib/gpu/cg_cmm_long.h
diff --git a/lib/gpu/cmmc_long_gpu.cpp b/lib/gpu/cg_cmm_long_ext.cpp
similarity index 99%
rename from lib/gpu/cmmc_long_gpu.cpp
rename to lib/gpu/cg_cmm_long_ext.cpp
index a6f3d090af..5e3d64f3f4 100644
--- a/lib/gpu/cmmc_long_gpu.cpp
+++ b/lib/gpu/cg_cmm_long_ext.cpp
@@ -19,7 +19,7 @@
 #include <cassert>
 #include <math.h>
 
-#include "cmmc_long_gpu_memory.h"
+#include "cg_cmm_long.h"
 
 using namespace std;
 
diff --git a/lib/gpu/cmmc_msm_gpu_memory.cpp b/lib/gpu/cg_cmm_msm.cpp
similarity index 97%
rename from lib/gpu/cmmc_msm_gpu_memory.cpp
rename to lib/gpu/cg_cmm_msm.cpp
index 22d69a33e2..25ebcf8657 100644
--- a/lib/gpu/cmmc_msm_gpu_memory.cpp
+++ b/lib/gpu/cg_cmm_msm.cpp
@@ -16,12 +16,12 @@
 ------------------------------------------------------------------------- */
 
 #ifdef USE_OPENCL
-#include "cmmc_msm_gpu_cl.h"
+#include "cg_cmm_msm_ext_cl.h"
 #else
-#include "cmmc_msm_gpu_ptx.h"
+#include "cg_cmm_msm_ext_ptx.h"
 #endif
 
-#include "cmmc_msm_gpu_memory.h"
+#include "cg_cmm_msm.h"
 #include <cassert>
 #define CMMM_GPU_MemoryT CMMM_GPU_Memory<numtyp, acctyp>
 
@@ -57,7 +57,7 @@ int CMMM_GPU_MemoryT::init(const int ntypes, double **host_cutsq,
                            const int smooth) {
   int success;
   success=this->init_atomic(nlocal,nall,max_nbors,maxspecial,cell_size,gpu_split,
-                            _screen,cmmc_msm_gpu_kernel);
+                            _screen,cg_cmm_msm);
   if (success!=0)
     return success;
 
diff --git a/lib/gpu/cmmc_msm_gpu_kernel.cu b/lib/gpu/cg_cmm_msm.cu
similarity index 100%
rename from lib/gpu/cmmc_msm_gpu_kernel.cu
rename to lib/gpu/cg_cmm_msm.cu
diff --git a/lib/gpu/cmmc_msm_gpu_memory.h b/lib/gpu/cg_cmm_msm.h
similarity index 100%
rename from lib/gpu/cmmc_msm_gpu_memory.h
rename to lib/gpu/cg_cmm_msm.h
diff --git a/lib/gpu/cmmc_msm_gpu.cpp b/lib/gpu/cg_cmm_msm_ext.cpp
similarity index 99%
rename from lib/gpu/cmmc_msm_gpu.cpp
rename to lib/gpu/cg_cmm_msm_ext.cpp
index cfa6c50453..681f4cf94a 100644
--- a/lib/gpu/cmmc_msm_gpu.cpp
+++ b/lib/gpu/cg_cmm_msm_ext.cpp
@@ -19,7 +19,7 @@
 #include <cassert>
 #include <math.h>
 
-#include "cmmc_msm_gpu_memory.h"
+#include "cg_cmm_msm.h"
 
 using namespace std;
 
diff --git a/lib/gpu/crml_gpu_memory.cpp b/lib/gpu/charmm_long.cpp
similarity index 97%
rename from lib/gpu/crml_gpu_memory.cpp
rename to lib/gpu/charmm_long.cpp
index 9c41cced1a..6fdf620f7c 100644
--- a/lib/gpu/crml_gpu_memory.cpp
+++ b/lib/gpu/charmm_long.cpp
@@ -16,12 +16,12 @@
 ------------------------------------------------------------------------- */
 
 #ifdef USE_OPENCL
-#include "crml_gpu_cl.h"
+#include "charmm_long_ext_cl.h"
 #else
-#include "crml_gpu_ptx.h"
+#include "charmm_long_ext_ptx.h"
 #endif
 
-#include "crml_gpu_memory.h"
+#include "charmm_long.h"
 #include <cassert>
 #define CRML_GPU_MemoryT CRML_GPU_Memory<numtyp, acctyp>
 
@@ -58,7 +58,7 @@ int CRML_GPU_MemoryT::init(const int ntypes,
                            double **sigma, const bool mix_arithmetic) {
   int success;
   success=this->init_atomic(nlocal,nall,max_nbors,maxspecial,cell_size,gpu_split,
-                            _screen,crml_gpu_kernel);
+                            _screen,charmm_long);
   if (success!=0)
     return success;
 
diff --git a/lib/gpu/crml_gpu_kernel.cu b/lib/gpu/charmm_long.cu
similarity index 100%
rename from lib/gpu/crml_gpu_kernel.cu
rename to lib/gpu/charmm_long.cu
diff --git a/lib/gpu/crml_gpu_memory.h b/lib/gpu/charmm_long.h
similarity index 100%
rename from lib/gpu/crml_gpu_memory.h
rename to lib/gpu/charmm_long.h
diff --git a/lib/gpu/crml_gpu.cpp b/lib/gpu/charmm_long_ext.cpp
similarity index 94%
rename from lib/gpu/crml_gpu.cpp
rename to lib/gpu/charmm_long_ext.cpp
index 1e59562ed5..24064b3f48 100644
--- a/lib/gpu/crml_gpu.cpp
+++ b/lib/gpu/charmm_long_ext.cpp
@@ -19,7 +19,7 @@
 #include <cassert>
 #include <math.h>
 
-#include "crml_gpu_memory.h"
+#include "charmm_long.h"
 
 using namespace std;
 
@@ -28,7 +28,7 @@ static CRML_GPU_Memory<PRECISION,ACC_PRECISION> CRMLMF;
 // ---------------------------------------------------------------------------
 // Allocate memory on host and device and copy constants to device
 // ---------------------------------------------------------------------------
-int crml_gpu_init(const int ntypes, double cut_bothsq, double **host_lj1,
+int charmm_long_ext_init(const int ntypes, double cut_bothsq, double **host_lj1,
                   double **host_lj2, double **host_lj3, double **host_lj4,
                   double **offset, double *special_lj, const int inum,
                   const int nall, const int max_nbors, const int maxspecial,
@@ -99,11 +99,11 @@ int crml_gpu_init(const int ntypes, double cut_bothsq, double **host_lj1,
   return init_ok;
 }
 
-void crml_gpu_clear() {
+void charmm_long_ext_clear() {
   CRMLMF.clear();
 }
 
-int** crml_gpu_compute_n(const int ago, const int inum_full,
+int** charmm_long_ext_compute_n(const int ago, const int inum_full,
                          const int nall, double **host_x, int *host_type,
                          double *sublo, double *subhi, int *tag, int **nspecial, 
                          int **special, const bool eflag, const bool vflag,
@@ -117,7 +117,7 @@ int** crml_gpu_compute_n(const int ago, const int inum_full,
                         host_q, boxlo, prd);
 }  
 			
-void crml_gpu_compute(const int ago, const int inum_full,
+void charmm_long_ext_compute(const int ago, const int inum_full,
 	 	                  const int nall, double **host_x, int *host_type,
                       int *ilist, int *numj, int **firstneigh,
 		                  const bool eflag, const bool vflag, const bool eatom,
@@ -129,7 +129,7 @@ void crml_gpu_compute(const int ago, const int inum_full,
                  nlocal,boxlo,prd);
 }
 
-double crml_gpu_bytes() {
+double charmm_long_ext_bytes() {
   return CRMLMF.host_memory_usage();
 }
 
diff --git a/lib/gpu/lj_cut_gpu_memory.cpp b/lib/gpu/lj.cpp
similarity index 97%
rename from lib/gpu/lj_cut_gpu_memory.cpp
rename to lib/gpu/lj.cpp
index 40622c576c..e3d1543dbf 100644
--- a/lib/gpu/lj_cut_gpu_memory.cpp
+++ b/lib/gpu/lj.cpp
@@ -16,12 +16,12 @@
 ------------------------------------------------------------------------- */
 
 #ifdef USE_OPENCL
-#include "lj_cut_gpu_cl.h"
+#include "lj_ext_cl.h"
 #else
-#include "lj_cut_gpu_ptx.h"
+#include "lj_ext_ptx.h"
 #endif
 
-#include "lj_cut_gpu_memory.h"
+#include "lj.h"
 #include <cassert>
 #define LJL_GPU_MemoryT LJL_GPU_Memory<numtyp, acctyp>
 
@@ -52,7 +52,7 @@ int LJL_GPU_MemoryT::init(const int ntypes,
                           const double gpu_split, FILE *_screen) {
   int success;
   success=this->init_atomic(nlocal,nall,max_nbors,maxspecial,cell_size,gpu_split,
-                            _screen,lj_cut_gpu_kernel);
+                            _screen,lj);
   if (success!=0)
     return success;
 
diff --git a/lib/gpu/lj_cut_gpu_kernel.cu b/lib/gpu/lj.cu
similarity index 100%
rename from lib/gpu/lj_cut_gpu_kernel.cu
rename to lib/gpu/lj.cu
diff --git a/lib/gpu/lj_cut_gpu_memory.h b/lib/gpu/lj.h
similarity index 100%
rename from lib/gpu/lj_cut_gpu_memory.h
rename to lib/gpu/lj.h
diff --git a/lib/gpu/lj96_cut_gpu_memory.cpp b/lib/gpu/lj96.cpp
similarity index 97%
rename from lib/gpu/lj96_cut_gpu_memory.cpp
rename to lib/gpu/lj96.cpp
index 4cc61a5d03..8850c8f655 100644
--- a/lib/gpu/lj96_cut_gpu_memory.cpp
+++ b/lib/gpu/lj96.cpp
@@ -16,12 +16,12 @@
 ------------------------------------------------------------------------- */
 
 #ifdef USE_OPENCL
-#include "lj96_cut_gpu_cl.h"
+#include "lj96_ext_cl.h"
 #else
-#include "lj96_cut_gpu_ptx.h"
+#include "lj96_ext_ptx.h"
 #endif
 
-#include "lj96_cut_gpu_memory.h"
+#include "lj96.h"
 #include <cassert>
 #define LJ96_GPU_MemoryT LJ96_GPU_Memory<numtyp, acctyp>
 
@@ -52,7 +52,7 @@ int LJ96_GPU_MemoryT::init(const int ntypes,
                            const double gpu_split, FILE *_screen) {
   int success;
   success=this->init_atomic(nlocal,nall,max_nbors,maxspecial,cell_size,gpu_split,
-                            _screen,lj96_cut_gpu_kernel);
+                            _screen,lj96);
   if (success!=0)
     return success;
 
diff --git a/lib/gpu/lj96_cut_gpu_kernel.cu b/lib/gpu/lj96.cu
similarity index 100%
rename from lib/gpu/lj96_cut_gpu_kernel.cu
rename to lib/gpu/lj96.cu
diff --git a/lib/gpu/lj96_cut_gpu_memory.h b/lib/gpu/lj96.h
similarity index 100%
rename from lib/gpu/lj96_cut_gpu_memory.h
rename to lib/gpu/lj96.h
diff --git a/lib/gpu/lj96_cut_gpu.cpp b/lib/gpu/lj96_ext.cpp
similarity index 99%
rename from lib/gpu/lj96_cut_gpu.cpp
rename to lib/gpu/lj96_ext.cpp
index df83afd521..0ec5e7d48a 100644
--- a/lib/gpu/lj96_cut_gpu.cpp
+++ b/lib/gpu/lj96_ext.cpp
@@ -19,7 +19,7 @@
 #include <cassert>
 #include <math.h>
 
-#include "lj96_cut_gpu_memory.h"
+#include "lj96.h"
 
 using namespace std;
 
diff --git a/lib/gpu/ljc_cut_gpu_memory.cpp b/lib/gpu/lj_coul.cpp
similarity index 97%
rename from lib/gpu/ljc_cut_gpu_memory.cpp
rename to lib/gpu/lj_coul.cpp
index df089a626a..9da19b7123 100644
--- a/lib/gpu/ljc_cut_gpu_memory.cpp
+++ b/lib/gpu/lj_coul.cpp
@@ -16,12 +16,12 @@
 ------------------------------------------------------------------------- */
 
 #ifdef USE_OPENCL
-#include "ljc_cut_gpu_cl.h"
+#include "lj_coul_ext_cl.h"
 #else
-#include "ljc_cut_gpu_ptx.h"
+#include "lj_coul_ext_ptx.h"
 #endif
 
-#include "ljc_cut_gpu_memory.h"
+#include "lj_coul.h"
 #include <cassert>
 #define LJC_GPU_MemoryT LJC_GPU_Memory<numtyp, acctyp>
 
@@ -55,7 +55,7 @@ int LJC_GPU_MemoryT::init(const int ntypes,
                           double *host_special_coul, const double qqrd2e) {
   int success;
   success=this->init_atomic(nlocal,nall,max_nbors,maxspecial,cell_size,gpu_split,
-                            _screen,ljc_cut_gpu_kernel);
+                            _screen,lj_coul);
   if (success!=0)
     return success;
 
diff --git a/lib/gpu/ljc_cut_gpu_kernel.cu b/lib/gpu/lj_coul.cu
similarity index 100%
rename from lib/gpu/ljc_cut_gpu_kernel.cu
rename to lib/gpu/lj_coul.cu
diff --git a/lib/gpu/ljc_cut_gpu_memory.h b/lib/gpu/lj_coul.h
similarity index 100%
rename from lib/gpu/ljc_cut_gpu_memory.h
rename to lib/gpu/lj_coul.h
diff --git a/lib/gpu/ljc_cut_gpu.cpp b/lib/gpu/lj_coul_ext.cpp
similarity index 99%
rename from lib/gpu/ljc_cut_gpu.cpp
rename to lib/gpu/lj_coul_ext.cpp
index de6f4f3e62..1909384395 100644
--- a/lib/gpu/ljc_cut_gpu.cpp
+++ b/lib/gpu/lj_coul_ext.cpp
@@ -19,7 +19,7 @@
 #include <cassert>
 #include <math.h>
 
-#include "ljc_cut_gpu_memory.h"
+#include "lj_coul.h"
 
 using namespace std;
 
diff --git a/lib/gpu/ljcl_cut_gpu_memory.cpp b/lib/gpu/lj_coul_long.cpp
similarity index 97%
rename from lib/gpu/ljcl_cut_gpu_memory.cpp
rename to lib/gpu/lj_coul_long.cpp
index f36d2a68c2..41c1a1483c 100644
--- a/lib/gpu/ljcl_cut_gpu_memory.cpp
+++ b/lib/gpu/lj_coul_long.cpp
@@ -16,12 +16,12 @@
 ------------------------------------------------------------------------- */
 
 #ifdef USE_OPENCL
-#include "ljcl_cut_gpu_cl.h"
+#include "lj_coul_long_ext_cl.h"
 #else
-#include "ljcl_cut_gpu_ptx.h"
+#include "lj_coul_long_ext_ptx.h"
 #endif
 
-#include "ljcl_cut_gpu_memory.h"
+#include "lj_coul_long.h"
 #include <cassert>
 #define LJCL_GPU_MemoryT LJCL_GPU_Memory<numtyp, acctyp>
 
@@ -56,7 +56,7 @@ int LJCL_GPU_MemoryT::init(const int ntypes,
                            const double g_ewald) {
   int success;
   success=this->init_atomic(nlocal,nall,max_nbors,maxspecial,cell_size,gpu_split,
-                            _screen,ljcl_cut_gpu_kernel);
+                            _screen,lj_coul_long);
   if (success!=0)
     return success;
 
diff --git a/lib/gpu/ljcl_cut_gpu_kernel.cu b/lib/gpu/lj_coul_long.cu
similarity index 100%
rename from lib/gpu/ljcl_cut_gpu_kernel.cu
rename to lib/gpu/lj_coul_long.cu
diff --git a/lib/gpu/ljcl_cut_gpu_memory.h b/lib/gpu/lj_coul_long.h
similarity index 100%
rename from lib/gpu/ljcl_cut_gpu_memory.h
rename to lib/gpu/lj_coul_long.h
diff --git a/lib/gpu/ljcl_cut_gpu.cpp b/lib/gpu/lj_coul_long_ext.cpp
similarity index 99%
rename from lib/gpu/ljcl_cut_gpu.cpp
rename to lib/gpu/lj_coul_long_ext.cpp
index 167f41b374..a90e6c714a 100644
--- a/lib/gpu/ljcl_cut_gpu.cpp
+++ b/lib/gpu/lj_coul_long_ext.cpp
@@ -19,7 +19,7 @@
 #include <cassert>
 #include <math.h>
 
-#include "ljcl_cut_gpu_memory.h"
+#include "lj_coul_long.h"
 
 using namespace std;
 
diff --git a/lib/gpu/lj_expand_gpu_memory.cpp b/lib/gpu/lj_expand.cpp
similarity index 97%
rename from lib/gpu/lj_expand_gpu_memory.cpp
rename to lib/gpu/lj_expand.cpp
index 714aa7c77f..b12ff98a78 100644
--- a/lib/gpu/lj_expand_gpu_memory.cpp
+++ b/lib/gpu/lj_expand.cpp
@@ -16,12 +16,12 @@
 ------------------------------------------------------------------------- */
 
 #ifdef USE_OPENCL
-#include "lj_expand_gpu_cl.h"
+#include "lj_expand_ext_cl.h"
 #else
-#include "lj_expand_gpu_ptx.h"
+#include "lj_expand_ext_ptx.h"
 #endif
 
-#include "lj_expand_gpu_memory.h"
+#include "lj_expand.h"
 #include <cassert>
 #define LJE_GPU_MemoryT LJE_GPU_Memory<numtyp, acctyp>
 
@@ -52,7 +52,7 @@ int LJE_GPU_MemoryT::init(const int ntypes, double **host_cutsq,
                           const double gpu_split, FILE *_screen) {
   int success;
   success=this->init_atomic(nlocal,nall,max_nbors,maxspecial,cell_size,gpu_split,
-                            _screen,lj_expand_gpu_kernel);
+                            _screen,lj_expand);
   if (success!=0)
     return success;
 
diff --git a/lib/gpu/lj_expand_gpu_kernel.cu b/lib/gpu/lj_expand.cu
similarity index 100%
rename from lib/gpu/lj_expand_gpu_kernel.cu
rename to lib/gpu/lj_expand.cu
diff --git a/lib/gpu/lj_expand_gpu_memory.h b/lib/gpu/lj_expand.h
similarity index 100%
rename from lib/gpu/lj_expand_gpu_memory.h
rename to lib/gpu/lj_expand.h
diff --git a/lib/gpu/lj_expand_gpu.cpp b/lib/gpu/lj_expand_ext.cpp
similarity index 99%
rename from lib/gpu/lj_expand_gpu.cpp
rename to lib/gpu/lj_expand_ext.cpp
index fd0392f3fc..60fc19f77c 100644
--- a/lib/gpu/lj_expand_gpu.cpp
+++ b/lib/gpu/lj_expand_ext.cpp
@@ -19,7 +19,7 @@
 #include <cassert>
 #include <math.h>
 
-#include "lj_expand_gpu_memory.h"
+#include "lj_expand.h"
 
 using namespace std;
 
diff --git a/lib/gpu/lj_cut_gpu.cpp b/lib/gpu/lj_ext.cpp
similarity index 99%
rename from lib/gpu/lj_cut_gpu.cpp
rename to lib/gpu/lj_ext.cpp
index aef085f7c9..ed0b4ec179 100644
--- a/lib/gpu/lj_cut_gpu.cpp
+++ b/lib/gpu/lj_ext.cpp
@@ -19,7 +19,7 @@
 #include <cassert>
 #include <math.h>
 
-#include "lj_cut_gpu_memory.h"
+#include "lj.h"
 
 using namespace std;
 
diff --git a/lib/gpu/morse_gpu_memory.cpp b/lib/gpu/morse.cpp
similarity index 97%
rename from lib/gpu/morse_gpu_memory.cpp
rename to lib/gpu/morse.cpp
index d188683915..feb6e4258f 100644
--- a/lib/gpu/morse_gpu_memory.cpp
+++ b/lib/gpu/morse.cpp
@@ -16,12 +16,12 @@
 ------------------------------------------------------------------------- */
 
 #ifdef USE_OPENCL
-#include "morse_gpu_cl.h"
+#include "morse_ext_cl.h"
 #else
-#include "morse_gpu_ptx.h"
+#include "morse_ext_ptx.h"
 #endif
 
-#include "morse_gpu_memory.h"
+#include "morse.h"
 #include <cassert>
 #define MOR_GPU_MemoryT MOR_GPU_Memory<numtyp, acctyp>
 
@@ -52,7 +52,7 @@ int MOR_GPU_MemoryT::init(const int ntypes,
                           const double gpu_split, FILE *_screen) {
   int success;
   success=this->init_atomic(nlocal,nall,max_nbors,maxspecial,cell_size,gpu_split,
-                            _screen,morse_gpu_kernel);
+                            _screen,morse);
   if (success!=0)
     return success;
 
diff --git a/lib/gpu/morse_gpu_kernel.cu b/lib/gpu/morse.cu
similarity index 100%
rename from lib/gpu/morse_gpu_kernel.cu
rename to lib/gpu/morse.cu
diff --git a/lib/gpu/morse_gpu_memory.h b/lib/gpu/morse.h
similarity index 100%
rename from lib/gpu/morse_gpu_memory.h
rename to lib/gpu/morse.h
diff --git a/lib/gpu/morse_gpu.cpp b/lib/gpu/morse_ext.cpp
similarity index 99%
rename from lib/gpu/morse_gpu.cpp
rename to lib/gpu/morse_ext.cpp
index 96636128db..00eab1c2c1 100644
--- a/lib/gpu/morse_gpu.cpp
+++ b/lib/gpu/morse_ext.cpp
@@ -19,7 +19,7 @@
 #include <cassert>
 #include <math.h>
 
-#include "morse_gpu_memory.h"
+#include "morse.h"
 
 using namespace std;
 
diff --git a/lib/gpu/pair_gpu_atom.cpp b/lib/gpu/pair_gpu_atom.cpp
index e34a15c0b9..28550b8419 100644
--- a/lib/gpu/pair_gpu_atom.cpp
+++ b/lib/gpu/pair_gpu_atom.cpp
@@ -19,15 +19,6 @@
 
 #define PairGPUAtomT PairGPUAtom<numtyp,acctyp>
 
-#ifdef WINDLL
-#include <windows.h>
-typedef bool (*__win_sort_alloc)(const int max_atoms);
-typedef void (*__win_sort)(const int max_atoms, unsigned *cell_begin,
-                           int *particle_begin);
-__win_sort_alloc _win_sort_alloc;
-__win_sort _win_sort;
-#endif
-
 template <class numtyp, class acctyp>
 PairGPUAtomT::PairGPUAtom() : _compiled(false),_allocated(false),
                               _max_gpu_bytes(0) {
@@ -36,17 +27,6 @@ PairGPUAtomT::PairGPUAtom() : _compiled(false),_allocated(false),
   sort_config.datatype = CUDPP_UINT;
   sort_config.algorithm = CUDPP_SORT_RADIX;
   sort_config.options = CUDPP_OPTION_KEY_VALUE_PAIRS;
-
-  #ifdef WINDLL
-  HINSTANCE hinstLib = LoadLibrary(TEXT("gpu.dll"));
-  if (hinstLib == NULL) {
-    printf("\nUnable to load gpu.dll\n");
-    exit(1);
-  }
-  _win_sort_alloc=(__win_sort_alloc)GetProcAddress(hinstLib,"_win_sort_alloc");
-  _win_sort=(__win_sort)GetProcAddress(hinstLib,"_win_sort");
-  #endif
-
   #endif
 }
 
@@ -76,16 +56,12 @@ bool PairGPUAtomT::alloc(const int nall) {
     
   // Allocate storage for CUDPP sort
   #ifndef USE_OPENCL
-  #ifdef WINDLL
-  _win_sort_alloc(_max_atoms);
-  #else
   if (_gpu_nbor) {
     CUDPPResult result = cudppPlan(&sort_plan, sort_config, _max_atoms, 1, 0);  
     if (CUDPP_SUCCESS != result)
       return false;
   }
   #endif
-  #endif
 
   // --------------------------   Host allocations
   // Get a host write only buffer
@@ -259,10 +235,8 @@ void PairGPUAtomT::clear_resize() {
   #endif
 
   #ifndef USE_OPENCL
-  #ifndef WINDLL
   if (_gpu_nbor) cudppDestroyPlan(sort_plan);
   #endif
-  #endif
 }
 
 template <class numtyp, class acctyp>
@@ -300,10 +274,6 @@ double PairGPUAtomT::host_memory_usage() const {
 template <class numtyp, class acctyp>
 void PairGPUAtomT::sort_neighbor(const int num_atoms) {
   #ifndef USE_OPENCL
-  #ifdef WINDLL
-  _win_sort(num_atoms,(unsigned *)dev_cell_id.begin(),
-            (int *)dev_particle_id.begin());
-  #else
   CUDPPResult result = cudppSort(sort_plan, (unsigned *)dev_cell_id.begin(), 
                                  (int *)dev_particle_id.begin(), 
                                  8*sizeof(unsigned), num_atoms);
@@ -312,7 +282,6 @@ void PairGPUAtomT::sort_neighbor(const int num_atoms) {
     NVD_GERYON_EXIT;
   }
   #endif
-  #endif
 }
 
 #ifdef GPU_CAST
diff --git a/lib/gpu/pair_gpu_device.h b/lib/gpu/pair_gpu_device.h
index 52b35cfcf2..dbc93f70cd 100644
--- a/lib/gpu/pair_gpu_device.h
+++ b/lib/gpu/pair_gpu_device.h
@@ -21,7 +21,7 @@
 #include "pair_gpu_atom.h"
 #include "pair_gpu_ans.h"
 #include "pair_gpu_nbor.h"
-#include "pppm_gpu_memory.h"
+#include "pppm.h"
 #include "mpi.h"
 #include <sstream>
 #include "stdio.h"
diff --git a/lib/gpu/pair_win_sort.cpp b/lib/gpu/pair_win_sort.cpp
deleted file mode 100644
index 4843a7ca17..0000000000
--- a/lib/gpu/pair_win_sort.cpp
+++ /dev/null
@@ -1,82 +0,0 @@
-/* ----------------------------------------------------------------------
-   LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator
-   http://lammps.sandia.gov, Sandia National Laboratories
-   Steve Plimpton, sjplimp@sandia.gov
-
-   Copyright (2003) Sandia Corporation.  Under the terms of Contract
-   DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains
-   certain rights in this software.  This software is distributed under 
-   the GNU General Public License.
-
-   See the README file in the top-level LAMMPS directory.
-------------------------------------------------------------------------- */
- 
-/* ----------------------------------------------------------------------
-   Contributing authors: Mike Brown (ORNL), brownw@ornl.gov
-------------------------------------------------------------------------- */
-
-#ifndef USE_OPENCL
-#include "cudpp.h"
-#endif
-
-class PairWinSort {
- public:
-  inline PairWinSort() : _allocated(false) {
-    #ifndef USE_OPENCL
-    sort_config.op = CUDPP_ADD;
-    sort_config.datatype = CUDPP_UINT;
-    sort_config.algorithm = CUDPP_SORT_RADIX;
-    sort_config.options = CUDPP_OPTION_KEY_VALUE_PAIRS;
-    #endif
-  }
-  inline ~PairWinSort() { clear(); }
-
-  /// Free all memory on host and device
-  inline void clear() {
-    #ifndef USE_OPENCL
-    if (_allocated) { cudppDestroyPlan(sort_plan); _allocated=false; }
-    #endif
-  }
- 
-  inline bool alloc(const int max_atoms) {
-    #ifndef USE_OPENCL
-    clear();
-    CUDPPResult result = cudppPlan(&sort_plan, sort_config, max_atoms, 1, 0);  
-    if (CUDPP_SUCCESS != result)
-      return false;
-    #endif
-    return true;
-  }
-
-  /// Sort arrays for neighbor list calculation
-  void sort_neighbor(const int num_atoms, unsigned *cell_begin, int *particle_begin) {
-    #ifndef USE_OPENCL
-    CUDPPResult result = cudppSort(sort_plan, cell_begin, particle_begin, 
-                                   8*sizeof(unsigned), num_atoms);
-    if (CUDPP_SUCCESS != result) {
-      printf("Error in cudppSort\n");
-      assert(1==0);
-    }
-    #endif
-  }
-  
- private:
-  
-  bool allocated;
-
-  #ifndef USE_OPENCL
-  CUDPPConfiguration sort_config;
-  CUDPPHandle sort_plan;
-  #endif
-};
-
-static PairWinSort win_sort;
-
-extern "C" __declspec(dllexport) bool _win_sort_alloc(const int max_atoms) {
-  win_sort.alloc(max_atoms);
-}
-
-extern "C" __declspec(dllexport) bool _win_sort(const int max_atoms, unsigned *cell_begin,
-                                                int *particle_begin) {
-  win_sort.sort(num_atoms,cell_begin,particle_begin);
-}
diff --git a/lib/gpu/pppm_gpu_memory.cpp b/lib/gpu/pppm.cpp
similarity index 99%
rename from lib/gpu/pppm_gpu_memory.cpp
rename to lib/gpu/pppm.cpp
index 2f7b35d051..2bd394e9f6 100644
--- a/lib/gpu/pppm_gpu_memory.cpp
+++ b/lib/gpu/pppm.cpp
@@ -21,7 +21,7 @@
 #include "pppm_f_gpu_ptx.h"
 #include "pppm_d_gpu_ptx.h"
 #endif
-#include "pppm_gpu_memory.h"
+#include "pppm.h"
 #include <cassert>
 
 #define PPPMGPUMemoryT PPPMGPUMemory<numtyp, acctyp, grdtyp, grdtyp4>
@@ -387,7 +387,7 @@ void PPPMGPUMemoryT::compile_kernels(UCL_Device &dev) {
   pppm_program=new UCL_Program(dev);
   
   #ifdef USE_OPENCL
-  pppm_program->load_string(pppm_gpu_kernel,flags.c_str());
+  pppm_program->load_string(pppm,flags.c_str());
   #else
   if (sizeof(grdtyp)==sizeof(float))
     pppm_program->load_string(pppm_f_gpu_kernel,flags.c_str());
diff --git a/lib/gpu/pppm_gpu_kernel.cu b/lib/gpu/pppm.cu
similarity index 100%
rename from lib/gpu/pppm_gpu_kernel.cu
rename to lib/gpu/pppm.cu
diff --git a/lib/gpu/pppm_gpu_memory.h b/lib/gpu/pppm.h
similarity index 100%
rename from lib/gpu/pppm_gpu_memory.h
rename to lib/gpu/pppm.h
diff --git a/lib/gpu/pppm_l_gpu.cpp b/lib/gpu/pppm_ext.cpp
similarity index 99%
rename from lib/gpu/pppm_l_gpu.cpp
rename to lib/gpu/pppm_ext.cpp
index 200d2f1685..8b0206c89e 100644
--- a/lib/gpu/pppm_l_gpu.cpp
+++ b/lib/gpu/pppm_ext.cpp
@@ -19,7 +19,7 @@
 #include <cassert>
 #include <math.h>
 
-#include "pppm_gpu_memory.h"
+#include "pppm.h"
 
 using namespace std;
 
diff --git a/lib/gpu/re_squared.h b/lib/gpu/re_squared.h
index d0c8a4dd4e..6ba94ff4ca 100644
--- a/lib/gpu/re_squared.h
+++ b/lib/gpu/re_squared.h
@@ -1,7 +1,7 @@
 /***************************************************************************
                                  re_squared.h
                              -------------------
-                               W. Michael Brown
+                            W. Michael Brown (ORNL)
 
   Host code for RE-Squared potential acceleration