diff --git a/cuda/cintrf.h b/cuda/cintrf.h
index 58778e6c..4f1cb015 100644
--- a/cuda/cintrf.h
+++ b/cuda/cintrf.h
@@ -35,7 +35,7 @@
 
 #include <stdlib.h>
 #include <stdio.h>
-
+#include "psb_config.h"
 #include "core.h"
 
 #define ELL_PITCH_ALIGN_S 32
diff --git a/cuda/spgpu/Makefile b/cuda/spgpu/Makefile
index 1a58fdc2..ae0b1be6 100644
--- a/cuda/spgpu/Makefile
+++ b/cuda/spgpu/Makefile
@@ -9,6 +9,7 @@ MODDIR=$(TOP)/modules
 LIBNAME=libspgpu.a
 
 OBJS=coo.o core.o dia.o ell.o hdia.o hell.o 
+CU_INCLUDES=-I$(INCDIR) $(CUDA_INCLUDES)
 
 all: includes objs
 
@@ -31,9 +32,9 @@ clean: iclean
 iclean:
 	$(MAKE) -C kernels clean 
 .c.o:
-	$(CC) $(CCOPT) $(CINCLUDES) $(CDEFINES) $(CUDA_INCLUDES) -c $<
+	$(CC) $(CCOPT) $(CINCLUDES) $(CDEFINES) $(CU_INCLUDES) -c $<
 .cpp.o:
-	$(CXX) $(CXXOPT) $(CXXINCLUDES) $(CXXDEFINES) $(CUDA_INCLUDES) -c $< -o $@
+	$(CXX) $(CXXOPT) $(CXXINCLUDES) $(CXXDEFINES) $(CU_INCLUDES) -c $< -o $@
 
 .cu.o:
-	$(NVCC)   $(CINCLUDES) $(CDEFINES) $(CUDEFINES) $(CUDA_INCLUDES) -c $<
+	$(NVCC)   $(CINCLUDES) $(CDEFINES) $(CUDEFINES) $(CU_INCLUDES) -c $<
diff --git a/cuda/spgpu/core.h b/cuda/spgpu/core.h
index 46d98f64..4ba85fd9 100644
--- a/cuda/spgpu/core.h
+++ b/cuda/spgpu/core.h
@@ -46,7 +46,7 @@
  * GNU General Public License for more details.
  */
 
- 
+#include "psb_config.h"
 #include "driver_types.h"
 #include "cuComplex.h"
  
diff --git a/cuda/spgpu/kernels/Makefile b/cuda/spgpu/kernels/Makefile
index 43cf2724..37113c75 100644
--- a/cuda/spgpu/kernels/Makefile
+++ b/cuda/spgpu/kernels/Makefile
@@ -9,7 +9,7 @@ MODDIR=$(TOP)/modules
 UP=..
 LIBNAME=$(UP)/libspgpu.a
 CINCLUDES=-I$(INCDIR)
-
+CU_INCLUDES=-I$(INCDIR) $(CUDA_INCLUDES)
 OBJS=cabs.o  camax.o  casum.o  caxpby.o  caxy.o  cdot.o  cgath.o \
   cnrm2.o cscal.o cscat.o csetscal.o cupd_xyz.o\
   dabs.o damax.o dasum.o daxpby.o daxy.o ddot.o dgath.o dupd_xyz.o\
@@ -30,5 +30,5 @@ lib: objs
 clean:
 	/bin/rm -fr $(OBJS)
 .cu.o:
-	$(NVCC)   $(CINCLUDES) $(CDEFINES) $(CUDEFINES) $(CUDA_INCLUDES) -c $<
+	$(NVCC)   $(CINCLUDES) $(CDEFINES) $(CUDEFINES) $(CU_INCLUDES) -c $<
 
diff --git a/cuda/spgpu/kernels/cabs.cu b/cuda/spgpu/kernels/cabs.cu
index f5e9d3ea..4c6df219 100644
--- a/cuda/spgpu/kernels/cabs.cu
+++ b/cuda/spgpu/kernels/cabs.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/camax.cu b/cuda/spgpu/kernels/camax.cu
index 049bbbe3..05d701e4 100644
--- a/cuda/spgpu/kernels/camax.cu
+++ b/cuda/spgpu/kernels/camax.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/casum.cu b/cuda/spgpu/kernels/casum.cu
index 7f4c92c5..7362cff8 100644
--- a/cuda/spgpu/kernels/casum.cu
+++ b/cuda/spgpu/kernels/casum.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/caxpby.cu b/cuda/spgpu/kernels/caxpby.cu
index 817fdf53..bea30d94 100644
--- a/cuda/spgpu/kernels/caxpby.cu
+++ b/cuda/spgpu/kernels/caxpby.cu
@@ -16,11 +16,10 @@
 #include "cudadebug.h"
 #include "cudalang.h"
 #include "cuComplex.h"
-
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
   int getGPUMultiProcessors();
   int getGPUMaxThreadsPerMP();
diff --git a/cuda/spgpu/kernels/caxy.cu b/cuda/spgpu/kernels/caxy.cu
index 6d7713b7..7b2a06cc 100644
--- a/cuda/spgpu/kernels/caxy.cu
+++ b/cuda/spgpu/kernels/caxy.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/cdot.cu b/cuda/spgpu/kernels/cdot.cu
index b6b19ec8..6ad40bae 100644
--- a/cuda/spgpu/kernels/cdot.cu
+++ b/cuda/spgpu/kernels/cdot.cu
@@ -19,10 +19,10 @@
 #include "cudadebug.h"
 #include "cuComplex.h"
 
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/cgath.cu b/cuda/spgpu/kernels/cgath.cu
index 103b6171..256254c3 100644
--- a/cuda/spgpu/kernels/cgath.cu
+++ b/cuda/spgpu/kernels/cgath.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/cnrm2.cu b/cuda/spgpu/kernels/cnrm2.cu
index 69856fb1..23e133e8 100644
--- a/cuda/spgpu/kernels/cnrm2.cu
+++ b/cuda/spgpu/kernels/cnrm2.cu
@@ -19,10 +19,10 @@
 #include "cudadebug.h"
 #include "cuComplex.h"
 
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/cscal.cu b/cuda/spgpu/kernels/cscal.cu
index c6239c05..26ffaf3d 100644
--- a/cuda/spgpu/kernels/cscal.cu
+++ b/cuda/spgpu/kernels/cscal.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/cscat.cu b/cuda/spgpu/kernels/cscat.cu
index 60942d1f..683f0ddd 100644
--- a/cuda/spgpu/kernels/cscat.cu
+++ b/cuda/spgpu/kernels/cscat.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/csetscal.cu b/cuda/spgpu/kernels/csetscal.cu
index 5934bdd9..f8774261 100644
--- a/cuda/spgpu/kernels/csetscal.cu
+++ b/cuda/spgpu/kernels/csetscal.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/cupd_xyz.cu b/cuda/spgpu/kernels/cupd_xyz.cu
index 34d637c6..04c137ff 100644
--- a/cuda/spgpu/kernels/cupd_xyz.cu
+++ b/cuda/spgpu/kernels/cupd_xyz.cu
@@ -17,10 +17,10 @@
 #include "cudadebug.h"
 #include "cudalang.h"
 #include <cuda_runtime.h>
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
   int getGPUMultiProcessors();
   int getGPUMaxThreadsPerMP();
diff --git a/cuda/spgpu/kernels/cxyzw.cu b/cuda/spgpu/kernels/cxyzw.cu
index d2b332b1..3047d65c 100644
--- a/cuda/spgpu/kernels/cxyzw.cu
+++ b/cuda/spgpu/kernels/cxyzw.cu
@@ -17,10 +17,10 @@
 #include "cudadebug.h"
 #include "cudalang.h"
 #include <cuda_runtime.h>
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
   int getGPUMultiProcessors();
   int getGPUMaxThreadsPerMP();
diff --git a/cuda/spgpu/kernels/dabs.cu b/cuda/spgpu/kernels/dabs.cu
index 7fb9c074..a27f21e4 100644
--- a/cuda/spgpu/kernels/dabs.cu
+++ b/cuda/spgpu/kernels/dabs.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/damax.cu b/cuda/spgpu/kernels/damax.cu
index 1e78cc1f..eee0b065 100644
--- a/cuda/spgpu/kernels/damax.cu
+++ b/cuda/spgpu/kernels/damax.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/dasum.cu b/cuda/spgpu/kernels/dasum.cu
index 71a37e80..5d7d7f9b 100644
--- a/cuda/spgpu/kernels/dasum.cu
+++ b/cuda/spgpu/kernels/dasum.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/daxpby.cu b/cuda/spgpu/kernels/daxpby.cu
index e4823b34..3a388988 100644
--- a/cuda/spgpu/kernels/daxpby.cu
+++ b/cuda/spgpu/kernels/daxpby.cu
@@ -17,10 +17,10 @@
 #include "cudadebug.h"
 #include "cudalang.h"
 #include <cuda_runtime.h>
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
   int getGPUMultiProcessors();
   int getGPUMaxThreadsPerMP();
diff --git a/cuda/spgpu/kernels/daxy.cu b/cuda/spgpu/kernels/daxy.cu
index 4a78c13c..3131a7af 100644
--- a/cuda/spgpu/kernels/daxy.cu
+++ b/cuda/spgpu/kernels/daxy.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/ddot.cu b/cuda/spgpu/kernels/ddot.cu
index f364ea2c..49f18943 100644
--- a/cuda/spgpu/kernels/ddot.cu
+++ b/cuda/spgpu/kernels/ddot.cu
@@ -18,10 +18,10 @@
 #include "cudalang.h"
 #include "cudadebug.h"
 
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/dgath.cu b/cuda/spgpu/kernels/dgath.cu
index 02b907f0..26f12647 100644
--- a/cuda/spgpu/kernels/dgath.cu
+++ b/cuda/spgpu/kernels/dgath.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/dia_cspmv.cu b/cuda/spgpu/kernels/dia_cspmv.cu
index c33d51bc..d4d56dc9 100644
--- a/cuda/spgpu/kernels/dia_cspmv.cu
+++ b/cuda/spgpu/kernels/dia_cspmv.cu
@@ -17,10 +17,10 @@
 #include "cudadebug.h"
 #include "cudalang.h"
 #include "cuComplex.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "dia.h"
 }
 
diff --git a/cuda/spgpu/kernels/dia_dspmv.cu b/cuda/spgpu/kernels/dia_dspmv.cu
index 0c720d42..21d16570 100644
--- a/cuda/spgpu/kernels/dia_dspmv.cu
+++ b/cuda/spgpu/kernels/dia_dspmv.cu
@@ -16,10 +16,10 @@
 
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "dia.h"
 }
 
diff --git a/cuda/spgpu/kernels/dia_sspmv.cu b/cuda/spgpu/kernels/dia_sspmv.cu
index 2870d780..17464b43 100644
--- a/cuda/spgpu/kernels/dia_sspmv.cu
+++ b/cuda/spgpu/kernels/dia_sspmv.cu
@@ -16,10 +16,10 @@
 
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "dia.h"
 }
 
diff --git a/cuda/spgpu/kernels/dia_zspmv.cu b/cuda/spgpu/kernels/dia_zspmv.cu
index 7e2fe68f..4f53bf5e 100644
--- a/cuda/spgpu/kernels/dia_zspmv.cu
+++ b/cuda/spgpu/kernels/dia_zspmv.cu
@@ -17,10 +17,10 @@
 #include "cudadebug.h"
 #include "cudalang.h"
 #include "cuComplex.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "dia.h"
 }
 
diff --git a/cuda/spgpu/kernels/dnrm2.cu b/cuda/spgpu/kernels/dnrm2.cu
index 450882c4..aae3edfc 100644
--- a/cuda/spgpu/kernels/dnrm2.cu
+++ b/cuda/spgpu/kernels/dnrm2.cu
@@ -18,10 +18,10 @@
 #include "cudalang.h"
 #include "cudadebug.h"
 
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/dscal.cu b/cuda/spgpu/kernels/dscal.cu
index 59502231..9e3b240c 100644
--- a/cuda/spgpu/kernels/dscal.cu
+++ b/cuda/spgpu/kernels/dscal.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/dscat.cu b/cuda/spgpu/kernels/dscat.cu
index 9582e1ea..c6bcf907 100644
--- a/cuda/spgpu/kernels/dscat.cu
+++ b/cuda/spgpu/kernels/dscat.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/dsetscal.cu b/cuda/spgpu/kernels/dsetscal.cu
index f452fa82..46af07b3 100644
--- a/cuda/spgpu/kernels/dsetscal.cu
+++ b/cuda/spgpu/kernels/dsetscal.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/dupd_xyz.cu b/cuda/spgpu/kernels/dupd_xyz.cu
index 38957fe1..d0060694 100644
--- a/cuda/spgpu/kernels/dupd_xyz.cu
+++ b/cuda/spgpu/kernels/dupd_xyz.cu
@@ -17,10 +17,10 @@
 #include "cudadebug.h"
 #include "cudalang.h"
 #include <cuda_runtime.h>
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
   int getGPUMultiProcessors();
   int getGPUMaxThreadsPerMP();
diff --git a/cuda/spgpu/kernels/dxyzw.cu b/cuda/spgpu/kernels/dxyzw.cu
index afd36651..ab93a379 100644
--- a/cuda/spgpu/kernels/dxyzw.cu
+++ b/cuda/spgpu/kernels/dxyzw.cu
@@ -17,10 +17,10 @@
 #include "cudadebug.h"
 #include "cudalang.h"
 #include <cuda_runtime.h>
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
   int getGPUMultiProcessors();
   int getGPUMaxThreadsPerMP();
diff --git a/cuda/spgpu/kernels/ell_ccsput.cu b/cuda/spgpu/kernels/ell_ccsput.cu
index 29e4d54c..66126bd0 100644
--- a/cuda/spgpu/kernels/ell_ccsput.cu
+++ b/cuda/spgpu/kernels/ell_ccsput.cu
@@ -17,10 +17,10 @@
 #include "cudadebug.h"
 #include "cudalang.h"
 #include "cuComplex.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "ell.h"
 }
 
diff --git a/cuda/spgpu/kernels/ell_cspmv.cu b/cuda/spgpu/kernels/ell_cspmv.cu
index 88bf47f7..b2a5ca3a 100644
--- a/cuda/spgpu/kernels/ell_cspmv.cu
+++ b/cuda/spgpu/kernels/ell_cspmv.cu
@@ -17,10 +17,10 @@
 #include "cudadebug.h"
 #include "cudalang.h"
 #include "cuComplex.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "ell.h"
 }
 
diff --git a/cuda/spgpu/kernels/ell_dcsput.cu b/cuda/spgpu/kernels/ell_dcsput.cu
index 00e60565..b1a25d52 100644
--- a/cuda/spgpu/kernels/ell_dcsput.cu
+++ b/cuda/spgpu/kernels/ell_dcsput.cu
@@ -16,10 +16,10 @@
 
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "ell.h"
 }
 
diff --git a/cuda/spgpu/kernels/ell_dspmv.cu b/cuda/spgpu/kernels/ell_dspmv.cu
index 4e682772..e4e2e815 100644
--- a/cuda/spgpu/kernels/ell_dspmv.cu
+++ b/cuda/spgpu/kernels/ell_dspmv.cu
@@ -16,10 +16,10 @@
 
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "ell.h"
 }
 
diff --git a/cuda/spgpu/kernels/ell_scsput.cu b/cuda/spgpu/kernels/ell_scsput.cu
index eeb5c6bb..0e3ed7b3 100644
--- a/cuda/spgpu/kernels/ell_scsput.cu
+++ b/cuda/spgpu/kernels/ell_scsput.cu
@@ -16,10 +16,10 @@
  
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "ell.h"
 }
 
diff --git a/cuda/spgpu/kernels/ell_sspmv.cu b/cuda/spgpu/kernels/ell_sspmv.cu
index e64e03e8..298ff1cb 100644
--- a/cuda/spgpu/kernels/ell_sspmv.cu
+++ b/cuda/spgpu/kernels/ell_sspmv.cu
@@ -16,10 +16,10 @@
  
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "ell.h"
 }
 
diff --git a/cuda/spgpu/kernels/ell_zcsput.cu b/cuda/spgpu/kernels/ell_zcsput.cu
index 3d0f3293..9fbd4311 100644
--- a/cuda/spgpu/kernels/ell_zcsput.cu
+++ b/cuda/spgpu/kernels/ell_zcsput.cu
@@ -17,10 +17,10 @@
 #include "cudadebug.h"
 #include "cudalang.h"
 #include "cuComplex.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "ell.h"
 }
 
diff --git a/cuda/spgpu/kernels/ell_zspmv.cu b/cuda/spgpu/kernels/ell_zspmv.cu
index a9c0fe64..dd1bc72d 100644
--- a/cuda/spgpu/kernels/ell_zspmv.cu
+++ b/cuda/spgpu/kernels/ell_zspmv.cu
@@ -17,10 +17,10 @@
 #include "cudadebug.h"
 #include "cudalang.h"
 #include "cuComplex.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "ell.h"
 }
 
diff --git a/cuda/spgpu/kernels/hdia_cspmv.cu b/cuda/spgpu/kernels/hdia_cspmv.cu
index 52360b40..27bb30e1 100644
--- a/cuda/spgpu/kernels/hdia_cspmv.cu
+++ b/cuda/spgpu/kernels/hdia_cspmv.cu
@@ -17,10 +17,10 @@
 #include "cudadebug.h"
 #include "cudalang.h"
 #include "cuComplex.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "hdia.h"
 }
 
diff --git a/cuda/spgpu/kernels/hdia_dspmv.cu b/cuda/spgpu/kernels/hdia_dspmv.cu
index 5fdb08e9..c1ac9af1 100644
--- a/cuda/spgpu/kernels/hdia_dspmv.cu
+++ b/cuda/spgpu/kernels/hdia_dspmv.cu
@@ -16,10 +16,10 @@
 
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "hdia.h"
 }
 
diff --git a/cuda/spgpu/kernels/hdia_sspmv.cu b/cuda/spgpu/kernels/hdia_sspmv.cu
index 7370f368..ad1d77e7 100644
--- a/cuda/spgpu/kernels/hdia_sspmv.cu
+++ b/cuda/spgpu/kernels/hdia_sspmv.cu
@@ -16,10 +16,10 @@
  
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "hdia.h"
 }
 
diff --git a/cuda/spgpu/kernels/hdia_zspmv.cu b/cuda/spgpu/kernels/hdia_zspmv.cu
index 4f955ed9..f0b592f6 100644
--- a/cuda/spgpu/kernels/hdia_zspmv.cu
+++ b/cuda/spgpu/kernels/hdia_zspmv.cu
@@ -17,10 +17,10 @@
 #include "cudadebug.h"
 #include "cudalang.h"
 #include "cuComplex.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "hdia.h"
 }
 
diff --git a/cuda/spgpu/kernels/hell_cspmv.cu b/cuda/spgpu/kernels/hell_cspmv.cu
index 86fbfc2b..c2b17419 100644
--- a/cuda/spgpu/kernels/hell_cspmv.cu
+++ b/cuda/spgpu/kernels/hell_cspmv.cu
@@ -17,10 +17,10 @@
 #include "cudadebug.h"
 #include "cudalang.h"
 #include "cuComplex.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "hell.h"
 }
 
diff --git a/cuda/spgpu/kernels/hell_dspmv.cu b/cuda/spgpu/kernels/hell_dspmv.cu
index b06f8cfb..fe379841 100644
--- a/cuda/spgpu/kernels/hell_dspmv.cu
+++ b/cuda/spgpu/kernels/hell_dspmv.cu
@@ -16,10 +16,10 @@
 
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "hell.h"
 }
 
diff --git a/cuda/spgpu/kernels/hell_sspmv.cu b/cuda/spgpu/kernels/hell_sspmv.cu
index 8d296b81..86b4c7b3 100644
--- a/cuda/spgpu/kernels/hell_sspmv.cu
+++ b/cuda/spgpu/kernels/hell_sspmv.cu
@@ -16,10 +16,10 @@
  
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "hell.h"
 }
 
diff --git a/cuda/spgpu/kernels/hell_zspmv.cu b/cuda/spgpu/kernels/hell_zspmv.cu
index 784ab719..aa8f6418 100644
--- a/cuda/spgpu/kernels/hell_zspmv.cu
+++ b/cuda/spgpu/kernels/hell_zspmv.cu
@@ -17,10 +17,10 @@
 #include "cudadebug.h"
 #include "cudalang.h"
 #include "cuComplex.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "hell.h"
 }
 
diff --git a/cuda/spgpu/kernels/igath.cu b/cuda/spgpu/kernels/igath.cu
index 8c0fe1b6..a97a1c98 100644
--- a/cuda/spgpu/kernels/igath.cu
+++ b/cuda/spgpu/kernels/igath.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/iscat.cu b/cuda/spgpu/kernels/iscat.cu
index 2001a740..dfb7e3ca 100644
--- a/cuda/spgpu/kernels/iscat.cu
+++ b/cuda/spgpu/kernels/iscat.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/isetscal.cu b/cuda/spgpu/kernels/isetscal.cu
index d54eae95..fe028b0e 100644
--- a/cuda/spgpu/kernels/isetscal.cu
+++ b/cuda/spgpu/kernels/isetscal.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/sabs.cu b/cuda/spgpu/kernels/sabs.cu
index e335e773..ee0efd8b 100644
--- a/cuda/spgpu/kernels/sabs.cu
+++ b/cuda/spgpu/kernels/sabs.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/samax.cu b/cuda/spgpu/kernels/samax.cu
index aaef17e2..8366e3e1 100644
--- a/cuda/spgpu/kernels/samax.cu
+++ b/cuda/spgpu/kernels/samax.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/sasum.cu b/cuda/spgpu/kernels/sasum.cu
index 33392203..3620775b 100644
--- a/cuda/spgpu/kernels/sasum.cu
+++ b/cuda/spgpu/kernels/sasum.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/saxpby.cu b/cuda/spgpu/kernels/saxpby.cu
index 2f06e39c..ebb41350 100644
--- a/cuda/spgpu/kernels/saxpby.cu
+++ b/cuda/spgpu/kernels/saxpby.cu
@@ -15,10 +15,10 @@
  */
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
   int getGPUMultiProcessors();
   int getGPUMaxThreadsPerMP();
diff --git a/cuda/spgpu/kernels/saxy.cu b/cuda/spgpu/kernels/saxy.cu
index be1591f5..9277a2c8 100644
--- a/cuda/spgpu/kernels/saxy.cu
+++ b/cuda/spgpu/kernels/saxy.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/sdot.cu b/cuda/spgpu/kernels/sdot.cu
index f91fab43..5952c34e 100644
--- a/cuda/spgpu/kernels/sdot.cu
+++ b/cuda/spgpu/kernels/sdot.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudalang.h"
 #include "cudadebug.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/sgath.cu b/cuda/spgpu/kernels/sgath.cu
index d1980bc6..6ccfa796 100644
--- a/cuda/spgpu/kernels/sgath.cu
+++ b/cuda/spgpu/kernels/sgath.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/snrm2.cu b/cuda/spgpu/kernels/snrm2.cu
index 27b8a6e9..1186cbe9 100644
--- a/cuda/spgpu/kernels/snrm2.cu
+++ b/cuda/spgpu/kernels/snrm2.cu
@@ -18,10 +18,10 @@
 #include "cudalang.h"
 #include "cudadebug.h"
 
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/sscal.cu b/cuda/spgpu/kernels/sscal.cu
index 4409390b..496d37a0 100644
--- a/cuda/spgpu/kernels/sscal.cu
+++ b/cuda/spgpu/kernels/sscal.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/sscat.cu b/cuda/spgpu/kernels/sscat.cu
index 6141ddb6..44e5dad9 100644
--- a/cuda/spgpu/kernels/sscat.cu
+++ b/cuda/spgpu/kernels/sscat.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/ssetscal.cu b/cuda/spgpu/kernels/ssetscal.cu
index 91bd3dff..dee8e9ab 100644
--- a/cuda/spgpu/kernels/ssetscal.cu
+++ b/cuda/spgpu/kernels/ssetscal.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/supd_xyz.cu b/cuda/spgpu/kernels/supd_xyz.cu
index d4ad1d0e..e8d7b028 100644
--- a/cuda/spgpu/kernels/supd_xyz.cu
+++ b/cuda/spgpu/kernels/supd_xyz.cu
@@ -17,10 +17,10 @@
 #include "cudadebug.h"
 #include "cudalang.h"
 #include <cuda_runtime.h>
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
   int getGPUMultiProcessors();
   int getGPUMaxThreadsPerMP();
diff --git a/cuda/spgpu/kernels/sxyzw.cu b/cuda/spgpu/kernels/sxyzw.cu
index 9cedd02f..f69d087e 100644
--- a/cuda/spgpu/kernels/sxyzw.cu
+++ b/cuda/spgpu/kernels/sxyzw.cu
@@ -17,10 +17,10 @@
 #include "cudadebug.h"
 #include "cudalang.h"
 #include <cuda_runtime.h>
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
   int getGPUMultiProcessors();
   int getGPUMaxThreadsPerMP();
diff --git a/cuda/spgpu/kernels/zabs.cu b/cuda/spgpu/kernels/zabs.cu
index ba0deba6..4bef9294 100644
--- a/cuda/spgpu/kernels/zabs.cu
+++ b/cuda/spgpu/kernels/zabs.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/zamax.cu b/cuda/spgpu/kernels/zamax.cu
index b4267180..4e65cf22 100644
--- a/cuda/spgpu/kernels/zamax.cu
+++ b/cuda/spgpu/kernels/zamax.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/zasum.cu b/cuda/spgpu/kernels/zasum.cu
index ffa28a8d..fa40b58c 100644
--- a/cuda/spgpu/kernels/zasum.cu
+++ b/cuda/spgpu/kernels/zasum.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/zaxpby.cu b/cuda/spgpu/kernels/zaxpby.cu
index 8efc40d2..c576acbd 100644
--- a/cuda/spgpu/kernels/zaxpby.cu
+++ b/cuda/spgpu/kernels/zaxpby.cu
@@ -18,10 +18,10 @@
 #include "cudalang.h"
 #include "cuComplex.h"
 
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
   int getGPUMultiProcessors();
   int getGPUMaxThreadsPerMP();
diff --git a/cuda/spgpu/kernels/zaxy.cu b/cuda/spgpu/kernels/zaxy.cu
index 1defbde2..cdc21adb 100644
--- a/cuda/spgpu/kernels/zaxy.cu
+++ b/cuda/spgpu/kernels/zaxy.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/zdot.cu b/cuda/spgpu/kernels/zdot.cu
index 6e1d1846..1b4b12c9 100644
--- a/cuda/spgpu/kernels/zdot.cu
+++ b/cuda/spgpu/kernels/zdot.cu
@@ -20,10 +20,10 @@
 #include "cuComplex.h"
 
 
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/zgath.cu b/cuda/spgpu/kernels/zgath.cu
index df381aca..59cebf6e 100644
--- a/cuda/spgpu/kernels/zgath.cu
+++ b/cuda/spgpu/kernels/zgath.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/znrm2.cu b/cuda/spgpu/kernels/znrm2.cu
index 3905fafe..3d1ed304 100644
--- a/cuda/spgpu/kernels/znrm2.cu
+++ b/cuda/spgpu/kernels/znrm2.cu
@@ -19,10 +19,10 @@
 #include "cudadebug.h"
 #include "cuComplex.h"
 
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/zscal.cu b/cuda/spgpu/kernels/zscal.cu
index f0520f90..13f3ad15 100644
--- a/cuda/spgpu/kernels/zscal.cu
+++ b/cuda/spgpu/kernels/zscal.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/zscat.cu b/cuda/spgpu/kernels/zscat.cu
index e46e99ff..b43ab761 100644
--- a/cuda/spgpu/kernels/zscat.cu
+++ b/cuda/spgpu/kernels/zscat.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/zsetscal.cu b/cuda/spgpu/kernels/zsetscal.cu
index e52e3ff9..70afa01c 100644
--- a/cuda/spgpu/kernels/zsetscal.cu
+++ b/cuda/spgpu/kernels/zsetscal.cu
@@ -17,10 +17,10 @@
 #include "stdio.h"
 #include "cudadebug.h"
 #include "cudalang.h"
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 
diff --git a/cuda/spgpu/kernels/zupd_xyz.cu b/cuda/spgpu/kernels/zupd_xyz.cu
index a3936ee9..360b87bd 100644
--- a/cuda/spgpu/kernels/zupd_xyz.cu
+++ b/cuda/spgpu/kernels/zupd_xyz.cu
@@ -17,10 +17,10 @@
 #include "cudadebug.h"
 #include "cudalang.h"
 #include <cuda_runtime.h>
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
   int getGPUMultiProcessors();
   int getGPUMaxThreadsPerMP();
diff --git a/cuda/spgpu/kernels/zxyzw.cu b/cuda/spgpu/kernels/zxyzw.cu
index 7a61edee..b07b6aa9 100644
--- a/cuda/spgpu/kernels/zxyzw.cu
+++ b/cuda/spgpu/kernels/zxyzw.cu
@@ -17,10 +17,10 @@
 #include "cudadebug.h"
 #include "cudalang.h"
 #include <cuda_runtime.h>
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
   int getGPUMultiProcessors();
   int getGPUMaxThreadsPerMP();