Fix cuda sources to use psb_config.h

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

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 $<

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"
  

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 $<
 

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"
 }
 

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"
 }
 

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"
 }
 

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();

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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();

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();

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"
 }
 

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"
 }
 

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"
 }
 

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();

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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();

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();

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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();

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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();

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();

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"
 }
 

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"
 }
 

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"
 }
 

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();

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"
 }
 

cuda/spgpu/kernels/zdot.cu

@@ -20,10 +20,10 @@
 #include "cuComplex.h"
 
 
+#include "core.h"
 
 extern "C"
 {
-#include "core.h"
 #include "vector.h"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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"
 }
 

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();

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();