compilation with CUDA5.0 (#1982, thanks to Rodrigo)

modules/gpu/src/element_operations.cpp

@@ -998,7 +998,11 @@ namespace
         typedef NppTypeTraits<CV_16U>::npp_t npp_t;
         typedef Npp32u scalar_t;
 
+#if (CUDA_VERSION <= 4020)
         typedef NppStatus (*func_t)(const Npp16u* pSrc1, int nSrc1Step, Npp16u* pDst, int nDstStep, NppiSize oSizeROI, Npp32u nConstant);
+#else
+        typedef NppStatus (*func_t)(const Npp16u * pSrc1, int nSrc1Step, Npp16u * pDst,  int nDstStep,  NppiSize oSizeROI, Npp16u nConstant);
+#endif
     };
 
     template <int DEPTH, typename NppAbsDiffCFunc<DEPTH>::func_t func> struct NppAbsDiffC

modules/gpu/src/matrix_reductions.cpp

@@ -128,7 +128,11 @@ void cv::gpu::meanStdDev(const GpuMat& src, Scalar& mean, Scalar& stddev, GpuMat
     DeviceBuffer dbuf(2);
 
     int bufSize;
+#if (CUDA_VERSION <= 4020)
     nppSafeCall( nppiMeanStdDev8uC1RGetBufferHostSize(sz, &bufSize) );
+#else
+    nppSafeCall( nppiMeanStdDevGetBufferHostSize_8u_C1R(sz, &bufSize) );
+#endif
 
     ensureSizeIsEnough(1, bufSize, CV_8UC1, buf);