optimizations:

- new reduce implementation (with kepler optimizations)
- saturate_cast via asm command
- video SIMD instructions in element operations
- float arithmetics instead of double
- new deviceSupports function

modules/core/include/opencv2/core/gpumat.hpp

@@ -79,6 +79,8 @@ namespace cv { namespace gpu
         WARP_SHUFFLE_FUNCTIONS = FEATURE_SET_COMPUTE_30
     };
 
+    CV_EXPORTS bool deviceSupports(FeatureSet feature_set);
+
     // Gives information about what GPU archs this OpenCV GPU module was
     // compiled for
     class CV_EXPORTS TargetArchs

modules/core/src/cuda/matrix_operations.cu

@@ -44,6 +44,7 @@
 #include "opencv2/gpu/device/saturate_cast.hpp"
 #include "opencv2/gpu/device/transform.hpp"
 #include "opencv2/gpu/device/functional.hpp"
+#include "opencv2/gpu/device/type_traits.hpp"
 
 namespace cv { namespace gpu { namespace device
 {
@@ -54,6 +55,7 @@ namespace cv { namespace gpu { namespace device
     void writeScalar(const int*);
     void writeScalar(const float*);
     void writeScalar(const double*);
+    void copyToWithMask_gpu(PtrStepSzb src, PtrStepSzb dst, size_t elemSize1, int cn, PtrStepSzb mask, bool colorMask, cudaStream_t stream);
     void convert_gpu(PtrStepSzb, int, PtrStepSzb, int, double, double, cudaStream_t);
 }}}
 
@@ -226,16 +228,16 @@ namespace cv { namespace gpu { namespace device
     //////////////////////////////// ConvertTo ////////////////////////////////
     ///////////////////////////////////////////////////////////////////////////
 
-    template <typename T, typename D> struct Convertor : unary_function<T, D>
+    template <typename T, typename D, typename S> struct Convertor : unary_function<T, D>
     {
-        Convertor(double alpha_, double beta_) : alpha(alpha_), beta(beta_) {}
+        Convertor(S alpha_, S beta_) : alpha(alpha_), beta(beta_) {}
 
-        __device__ __forceinline__ D operator()(const T& src) const
+        __device__ __forceinline__ D operator()(typename TypeTraits<T>::ParameterType src) const
         {
             return saturate_cast<D>(alpha * src + beta);
         }
 
-        double alpha, beta;
+        S alpha, beta;
     };
 
     namespace detail
@@ -282,16 +284,16 @@ namespace cv { namespace gpu { namespace device
         };
     }
 
-    template <typename T, typename D> struct TransformFunctorTraits< Convertor<T, D> > : detail::ConvertTraits< Convertor<T, D> >
+    template <typename T, typename D, typename S> struct TransformFunctorTraits< Convertor<T, D, S> > : detail::ConvertTraits< Convertor<T, D, S> >
     {
     };
 
-    template<typename T, typename D>
+    template<typename T, typename D, typename S>
     void cvt_(PtrStepSzb src, PtrStepSzb dst, double alpha, double beta, cudaStream_t stream)
     {
         cudaSafeCall( cudaSetDoubleForDevice(&alpha) );
         cudaSafeCall( cudaSetDoubleForDevice(&beta) );
-        Convertor<T, D> op(alpha, beta);
+        Convertor<T, D, S> op(static_cast<S>(alpha), static_cast<S>(beta));
         cv::gpu::device::transform((PtrStepSz<T>)src, (PtrStepSz<D>)dst, op, WithOutMask(), stream);
     }
 
@@ -304,36 +306,74 @@ namespace cv { namespace gpu { namespace device
     {
         typedef void (*caller_t)(PtrStepSzb src, PtrStepSzb dst, double alpha, double beta, cudaStream_t stream);
 
-        static const caller_t tab[8][8] =
+        static const caller_t tab[7][7] =
         {
-            {cvt_<uchar, uchar>, cvt_<uchar, schar>, cvt_<uchar, ushort>, cvt_<uchar, short>,
-            cvt_<uchar, int>, cvt_<uchar, float>, cvt_<uchar, double>, 0},
-
-            {cvt_<schar, uchar>, cvt_<schar, schar>, cvt_<schar, ushort>, cvt_<schar, short>,
-            cvt_<schar, int>, cvt_<schar, float>, cvt_<schar, double>, 0},
-
-            {cvt_<ushort, uchar>, cvt_<ushort, schar>, cvt_<ushort, ushort>, cvt_<ushort, short>,
-            cvt_<ushort, int>, cvt_<ushort, float>, cvt_<ushort, double>, 0},
-
-            {cvt_<short, uchar>, cvt_<short, schar>, cvt_<short, ushort>, cvt_<short, short>,
-            cvt_<short, int>, cvt_<short, float>, cvt_<short, double>, 0},
-
-            {cvt_<int, uchar>, cvt_<int, schar>, cvt_<int, ushort>,
-            cvt_<int, short>, cvt_<int, int>, cvt_<int, float>, cvt_<int, double>, 0},
-
-            {cvt_<float, uchar>, cvt_<float, schar>, cvt_<float, ushort>,
-            cvt_<float, short>, cvt_<float, int>, cvt_<float, float>, cvt_<float, double>, 0},
-
-            {cvt_<double, uchar>, cvt_<double, schar>, cvt_<double, ushort>,
-            cvt_<double, short>, cvt_<double, int>, cvt_<double, float>, cvt_<double, double>, 0},
-
-            {0,0,0,0,0,0,0,0}
+            {
+                cvt_<uchar, uchar, float>,
+                cvt_<uchar, schar, float>,
+                cvt_<uchar, ushort, float>,
+                cvt_<uchar, short, float>,
+                cvt_<uchar, int, float>,
+                cvt_<uchar, float, float>,
+                cvt_<uchar, double, double>
+            },
+            {
+                cvt_<schar, uchar, float>,
+                cvt_<schar, schar, float>,
+                cvt_<schar, ushort, float>,
+                cvt_<schar, short, float>,
+                cvt_<schar, int, float>,
+                cvt_<schar, float, float>,
+                cvt_<schar, double, double>
+            },
+            {
+                cvt_<ushort, uchar, float>,
+                cvt_<ushort, schar, float>,
+                cvt_<ushort, ushort, float>,
+                cvt_<ushort, short, float>,
+                cvt_<ushort, int, float>,
+                cvt_<ushort, float, float>,
+                cvt_<ushort, double, double>
+            },
+            {
+                cvt_<short, uchar, float>,
+                cvt_<short, schar, float>,
+                cvt_<short, ushort, float>,
+                cvt_<short, short, float>,
+                cvt_<short, int, float>,
+                cvt_<short, float, float>,
+                cvt_<short, double, double>
+            },
+            {
+                cvt_<int, uchar, float>,
+                cvt_<int, schar, float>,
+                cvt_<int, ushort, float>,
+                cvt_<int, short, float>,
+                cvt_<int, int, double>,
+                cvt_<int, float, double>,
+                cvt_<int, double, double>
+            },
+            {
+                cvt_<float, uchar, float>,
+                cvt_<float, schar, float>,
+                cvt_<float, ushort, float>,
+                cvt_<float, short, float>,
+                cvt_<float, int, float>,
+                cvt_<float, float, float>,
+                cvt_<float, double, double>
+            },
+            {
+                cvt_<double, uchar, double>,
+                cvt_<double, schar, double>,
+                cvt_<double, ushort, double>,
+                cvt_<double, short, double>,
+                cvt_<double, int, double>,
+                cvt_<double, float, double>,
+                cvt_<double, double, double>
+            }
         };
 
         caller_t func = tab[sdepth][ddepth];
-        if (!func)
-            cv::gpu::error("Unsupported convert operation", __FILE__, __LINE__, "convert_gpu");
-
         func(src, dst, alpha, beta, stream);
     }
 

modules/core/src/gpumat.cpp

@@ -45,8 +45,7 @@
 #include <iostream>
 
 #ifdef HAVE_CUDA
-    #include <cuda.h>
-    #include <cuda_runtime_api.h>
+    #include <cuda_runtime.h>
     #include <npp.h>
 
     #define CUDART_MINIMUM_REQUIRED_VERSION 4010
@@ -69,33 +68,89 @@ using namespace cv::gpu;
 
 namespace
 {
-    // Compares value to set using the given comparator. Returns true if
-    // there is at least one element x in the set satisfying to: x cmp value
-    // predicate.
-    template <typename Comparer>
-    bool compareToSet(const std::string& set_as_str, int value, Comparer cmp)
+    class CudaArch
+    {
+    public:
+        CudaArch();
+
+        bool builtWith(FeatureSet feature_set) const;
+        bool hasPtx(int major, int minor) const;
+        bool hasBin(int major, int minor) const;
+        bool hasEqualOrLessPtx(int major, int minor) const;
+        bool hasEqualOrGreaterPtx(int major, int minor) const;
+        bool hasEqualOrGreaterBin(int major, int minor) const;
+
+    private:
+        static void fromStr(const string& set_as_str, vector<int>& arr);
+
+        vector<int> bin;
+        vector<int> ptx;
+        vector<int> features;
+    };
+
+    const CudaArch cudaArch;
+
+    CudaArch::CudaArch()
+    {
+    #ifdef HAVE_CUDA
+        fromStr(CUDA_ARCH_BIN, bin);
+        fromStr(CUDA_ARCH_PTX, ptx);
+        fromStr(CUDA_ARCH_FEATURES, features);
+    #endif
+    }
+
+    bool CudaArch::builtWith(FeatureSet feature_set) const
+    {
+        return !features.empty() && (features.back() >= feature_set);
+    }
+
+    bool CudaArch::hasPtx(int major, int minor) const
+    {
+        return find(ptx.begin(), ptx.end(), major * 10 + minor) != ptx.end();
+    }
+
+    bool CudaArch::hasBin(int major, int minor) const
+    {
+        return find(bin.begin(), bin.end(), major * 10 + minor) != bin.end();
+    }
+
+    bool CudaArch::hasEqualOrLessPtx(int major, int minor) const
+    {
+        return !ptx.empty() && (ptx.front() <= major * 10 + minor);
+    }
+
+    bool CudaArch::hasEqualOrGreaterPtx(int major, int minor) const
+    {
+        return !ptx.empty() && (ptx.back() >= major * 10 + minor);
+    }
+
+    bool CudaArch::hasEqualOrGreaterBin(int major, int minor) const
+    {
+        return !bin.empty() && (bin.back() >= major * 10 + minor);
+    }
+
+    void CudaArch::fromStr(const string& set_as_str, vector<int>& arr)
     {
         if (set_as_str.find_first_not_of(" ") == string::npos)
-            return false;
+            return;
 
-        std::stringstream stream(set_as_str);
+        istringstream stream(set_as_str);
         int cur_value;
 
         while (!stream.eof())
         {
             stream >> cur_value;
-            if (cmp(cur_value, value))
-                return true;
+            arr.push_back(cur_value);
         }
 
-        return false;
+        sort(arr.begin(), arr.end());
     }
 }
 
 bool cv::gpu::TargetArchs::builtWith(cv::gpu::FeatureSet feature_set)
 {
 #if defined (HAVE_CUDA)
-    return ::compareToSet(CUDA_ARCH_FEATURES, feature_set, std::greater_equal<int>());
+    return cudaArch.builtWith(feature_set);
 #else
     (void)feature_set;
     return false;
@@ -110,7 +165,7 @@ bool cv::gpu::TargetArchs::has(int major, int minor)
 bool cv::gpu::TargetArchs::hasPtx(int major, int minor)
 {
 #if defined (HAVE_CUDA)
-    return ::compareToSet(CUDA_ARCH_PTX, major * 10 + minor, std::equal_to<int>());
+    return cudaArch.hasPtx(major, minor);
 #else
     (void)major;
     (void)minor;
@@ -121,7 +176,7 @@ bool cv::gpu::TargetArchs::hasPtx(int major, int minor)
 bool cv::gpu::TargetArchs::hasBin(int major, int minor)
 {
 #if defined (HAVE_CUDA)
-    return ::compareToSet(CUDA_ARCH_BIN, major * 10 + minor, std::equal_to<int>());
+    return cudaArch.hasBin(major, minor);
 #else
     (void)major;
     (void)minor;
@@ -132,8 +187,7 @@ bool cv::gpu::TargetArchs::hasBin(int major, int minor)
 bool cv::gpu::TargetArchs::hasEqualOrLessPtx(int major, int minor)
 {
 #if defined (HAVE_CUDA)
-    return ::compareToSet(CUDA_ARCH_PTX, major * 10 + minor,
-                     std::less_equal<int>());
+    return cudaArch.hasEqualOrLessPtx(major, minor);
 #else
     (void)major;
     (void)minor;
@@ -143,14 +197,13 @@ bool cv::gpu::TargetArchs::hasEqualOrLessPtx(int major, int minor)
 
 bool cv::gpu::TargetArchs::hasEqualOrGreater(int major, int minor)
 {
-    return hasEqualOrGreaterPtx(major, minor) ||
-           hasEqualOrGreaterBin(major, minor);
+    return hasEqualOrGreaterPtx(major, minor) || hasEqualOrGreaterBin(major, minor);
 }
 
 bool cv::gpu::TargetArchs::hasEqualOrGreaterPtx(int major, int minor)
 {
 #if defined (HAVE_CUDA)
-    return ::compareToSet(CUDA_ARCH_PTX, major * 10 + minor, std::greater_equal<int>());
+    return cudaArch.hasEqualOrGreaterPtx(major, minor);
 #else
     (void)major;
     (void)minor;
@@ -161,8 +214,7 @@ bool cv::gpu::TargetArchs::hasEqualOrGreaterPtx(int major, int minor)
 bool cv::gpu::TargetArchs::hasEqualOrGreaterBin(int major, int minor)
 {
 #if defined (HAVE_CUDA)
-    return ::compareToSet(CUDA_ARCH_BIN, major * 10 + minor,
-                     std::greater_equal<int>());
+    return cudaArch.hasEqualOrGreaterBin(major, minor);
 #else
     (void)major;
     (void)minor;
@@ -170,6 +222,31 @@ bool cv::gpu::TargetArchs::hasEqualOrGreaterBin(int major, int minor)
 #endif
 }
 
+bool cv::gpu::deviceSupports(FeatureSet feature_set)
+{
+    static int versions[] =
+    {
+        -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1
+    };
+    static const int cache_size = static_cast<int>(sizeof(versions) / sizeof(versions[0]));
+
+    const int devId = getDevice();
+
+    int version;
+
+    if (devId < cache_size && versions[devId] >= 0)
+        version = versions[devId];
+    else
+    {
+        DeviceInfo dev(devId);
+        version = dev.majorVersion() * 10 + dev.minorVersion();
+        if (devId < cache_size)
+            versions[devId] = version;
+    }
+
+    return TargetArchs::builtWith(feature_set) && (version >= feature_set);
+}
+
 #if !defined (HAVE_CUDA)
 
 #define throw_nogpu CV_Error(CV_GpuNotSupported, "The library is compiled without CUDA support")

modules/gpu/include/opencv2/gpu/device/saturate_cast.hpp

@@ -58,35 +58,47 @@ namespace cv { namespace gpu { namespace device
 
     template<> __device__ __forceinline__ uchar saturate_cast<uchar>(schar v)
     {
-        return (uchar) ::max((int)v, 0);
+        uint res = 0;
+        int vi = v;
+        asm("cvt.sat.u8.s8 %0, %1;" : "=r"(res) : "r"(vi));
+        return res;
+    }
+    template<> __device__ __forceinline__ uchar saturate_cast<uchar>(short v)
+    {
+        uint res = 0;
+        asm("cvt.sat.u8.s16 %0, %1;" : "=r"(res) : "h"(v));
+        return res;
     }
     template<> __device__ __forceinline__ uchar saturate_cast<uchar>(ushort v)
     {
-        return (uchar) ::min((uint)v, (uint)UCHAR_MAX);
+        uint res = 0;
+        asm("cvt.sat.u8.u16 %0, %1;" : "=r"(res) : "h"(v));
+        return res;
     }
     template<> __device__ __forceinline__ uchar saturate_cast<uchar>(int v)
     {
-        return (uchar)((uint)v <= UCHAR_MAX ? v : v > 0 ? UCHAR_MAX : 0);
+        uint res = 0;
+        asm("cvt.sat.u8.s32 %0, %1;" : "=r"(res) : "r"(v));
+        return res;
     }
     template<> __device__ __forceinline__ uchar saturate_cast<uchar>(uint v)
     {
-        return (uchar) ::min(v, (uint)UCHAR_MAX);
+        uint res = 0;
+        asm("cvt.sat.u8.u32 %0, %1;" : "=r"(res) : "r"(v));
+        return res;
     }
-    template<> __device__ __forceinline__ uchar saturate_cast<uchar>(short v)
-    {
-        return saturate_cast<uchar>((uint)v);
-    }
-
     template<> __device__ __forceinline__ uchar saturate_cast<uchar>(float v)
     {
-        int iv = __float2int_rn(v);
-        return saturate_cast<uchar>(iv);
+        uint res = 0;
+        asm("cvt.rni.sat.u8.f32 %0, %1;" : "=r"(res) : "f"(v));
+        return res;
     }
     template<> __device__ __forceinline__ uchar saturate_cast<uchar>(double v)
     {
-    #if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 130
-        int iv = __double2int_rn(v);
-        return saturate_cast<uchar>(iv);
+    #if __CUDA_ARCH__ >= 130
+        uint res = 0;
+        asm("cvt.rni.sat.u8.f64 %0, %1;" : "=r"(res) : "d"(v));
+        return res;
     #else
         return saturate_cast<uchar>((float)v);
     #endif
@@ -94,35 +106,47 @@ namespace cv { namespace gpu { namespace device
 
     template<> __device__ __forceinline__ schar saturate_cast<schar>(uchar v)
     {
-        return (schar) ::min((int)v, SCHAR_MAX);
+        uint res = 0;
+        uint vi = v;
+        asm("cvt.sat.s8.u8 %0, %1;" : "=r"(res) : "r"(vi));
+        return res;
     }
-    template<> __device__ __forceinline__ schar saturate_cast<schar>(ushort v)
+    template<> __device__ __forceinline__ schar saturate_cast<schar>(short v)
     {
-        return (schar) ::min((uint)v, (uint)SCHAR_MAX);
+        uint res = 0;
+        asm("cvt.sat.s8.s16 %0, %1;" : "=r"(res) : "h"(v));
+        return res;
     }
-    template<> __device__ __forceinline__ schar saturate_cast<schar>(int v)
+    template<> __device__ __forceinline__ schar saturate_cast<schar>(ushort v)
     {
-        return (schar)((uint)(v-SCHAR_MIN) <= (uint)UCHAR_MAX ? v : v > 0 ? SCHAR_MAX : SCHAR_MIN);
+        uint res = 0;
+        asm("cvt.sat.s8.u16 %0, %1;" : "=r"(res) : "h"(v));
+        return res;
     }
-    template<> __device__ __forceinline__ schar saturate_cast<schar>(short v)
+    template<> __device__ __forceinline__ schar saturate_cast<schar>(int v)
     {
-        return saturate_cast<schar>((int)v);
+        uint res = 0;
+        asm("cvt.sat.s8.s32 %0, %1;" : "=r"(res) : "r"(v));
+        return res;
     }
     template<> __device__ __forceinline__ schar saturate_cast<schar>(uint v)
     {
-        return (schar) ::min(v, (uint)SCHAR_MAX);
+        uint res = 0;
+        asm("cvt.sat.s8.u32 %0, %1;" : "=r"(res) : "r"(v));
+        return res;
     }
-
     template<> __device__ __forceinline__ schar saturate_cast<schar>(float v)
     {
-        int iv = __float2int_rn(v);
-        return saturate_cast<schar>(iv);
+        uint res = 0;
+        asm("cvt.rni.sat.s8.f32 %0, %1;" : "=r"(res) : "f"(v));
+        return res;
     }
     template<> __device__ __forceinline__ schar saturate_cast<schar>(double v)
     {
-    #if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 130
-        int iv = __double2int_rn(v);
-        return saturate_cast<schar>(iv);
+    #if __CUDA_ARCH__ >= 130
+        uint res = 0;
+        asm("cvt.rni.sat.s8.f64 %0, %1;" : "=r"(res) : "d"(v));
+        return res;
     #else
         return saturate_cast<schar>((float)v);
     #endif
@@ -130,30 +154,41 @@ namespace cv { namespace gpu { namespace device
 
     template<> __device__ __forceinline__ ushort saturate_cast<ushort>(schar v)
     {
-        return (ushort) ::max((int)v, 0);
+        ushort res = 0;
+        int vi = v;
+        asm("cvt.sat.u16.s8 %0, %1;" : "=h"(res) : "r"(vi));
+        return res;
     }
     template<> __device__ __forceinline__ ushort saturate_cast<ushort>(short v)
     {
-        return (ushort) ::max((int)v, 0);
+        ushort res = 0;
+        asm("cvt.sat.u16.s16 %0, %1;" : "=h"(res) : "h"(v));
+        return res;
     }
     template<> __device__ __forceinline__ ushort saturate_cast<ushort>(int v)
     {
-        return (ushort)((uint)v <= (uint)USHRT_MAX ? v : v > 0 ? USHRT_MAX : 0);
+        ushort res = 0;
+        asm("cvt.sat.u16.s32 %0, %1;" : "=h"(res) : "r"(v));
+        return res;
     }
     template<> __device__ __forceinline__ ushort saturate_cast<ushort>(uint v)
     {
-        return (ushort) ::min(v, (uint)USHRT_MAX);
+        ushort res = 0;
+        asm("cvt.sat.u16.u32 %0, %1;" : "=h"(res) : "r"(v));
+        return res;
     }
     template<> __device__ __forceinline__ ushort saturate_cast<ushort>(float v)
     {
-        int iv = __float2int_rn(v);
-        return saturate_cast<ushort>(iv);
+        ushort res = 0;
+        asm("cvt.rni.sat.u16.f32 %0, %1;" : "=h"(res) : "f"(v));
+        return res;
     }
     template<> __device__ __forceinline__ ushort saturate_cast<ushort>(double v)
     {
-    #if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 130
-        int iv = __double2int_rn(v);
-        return saturate_cast<ushort>(iv);
+    #if __CUDA_ARCH__ >= 130
+        ushort res = 0;
+        asm("cvt.rni.sat.u16.f64 %0, %1;" : "=h"(res) : "d"(v));
+        return res;
     #else
         return saturate_cast<ushort>((float)v);
     #endif
@@ -161,31 +196,45 @@ namespace cv { namespace gpu { namespace device
 
     template<> __device__ __forceinline__ short saturate_cast<short>(ushort v)
     {
-        return (short) ::min((int)v, SHRT_MAX);
+        short res = 0;
+        asm("cvt.sat.s16.u16 %0, %1;" : "=h"(res) : "h"(v));
+        return res;
     }
     template<> __device__ __forceinline__ short saturate_cast<short>(int v)
     {
-        return (short)((uint)(v - SHRT_MIN) <= (uint)USHRT_MAX ? v : v > 0 ? SHRT_MAX : SHRT_MIN);
+        short res = 0;
+        asm("cvt.sat.s16.s32 %0, %1;" : "=h"(res) : "r"(v));
+        return res;
     }
     template<> __device__ __forceinline__ short saturate_cast<short>(uint v)
     {
-        return (short) ::min(v, (uint)SHRT_MAX);
+        short res = 0;
+        asm("cvt.sat.s16.u32 %0, %1;" : "=h"(res) : "r"(v));
+        return res;
     }
     template<> __device__ __forceinline__ short saturate_cast<short>(float v)
     {
-        int iv = __float2int_rn(v);
-        return saturate_cast<short>(iv);
+        short res = 0;
+        asm("cvt.rni.sat.s16.f32 %0, %1;" : "=h"(res) : "f"(v));
+        return res;
     }
     template<> __device__ __forceinline__ short saturate_cast<short>(double v)
     {
-    #if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 130
-        int iv = __double2int_rn(v);
-        return saturate_cast<short>(iv);
+    #if __CUDA_ARCH__ >= 130
+        short res = 0;
+        asm("cvt.rni.sat.s16.f64 %0, %1;" : "=h"(res) : "d"(v));
+        return res;
     #else
         return saturate_cast<short>((float)v);
     #endif
     }
 
+    template<> __device__ __forceinline__ int saturate_cast<int>(uint v)
+    {
+        int res = 0;
+        asm("cvt.sat.s32.u32 %0, %1;" : "=r"(res) : "r"(v));
+        return res;
+    }
     template<> __device__ __forceinline__ int saturate_cast<int>(float v)
     {
         return __float2int_rn(v);
@@ -199,6 +248,25 @@ namespace cv { namespace gpu { namespace device
     #endif
     }
 
+    template<> __device__ __forceinline__ uint saturate_cast<uint>(schar v)
+    {
+        uint res = 0;
+        int vi = v;
+        asm("cvt.sat.u32.s8 %0, %1;" : "=r"(res) : "r"(vi));
+        return res;
+    }
+    template<> __device__ __forceinline__ uint saturate_cast<uint>(short v)
+    {
+        uint res = 0;
+        asm("cvt.sat.u32.s16 %0, %1;" : "=r"(res) : "h"(v));
+        return res;
+    }
+    template<> __device__ __forceinline__ uint saturate_cast<uint>(int v)
+    {
+        uint res = 0;
+        asm("cvt.sat.u32.s32 %0, %1;" : "=r"(res) : "r"(v));
+        return res;
+    }
     template<> __device__ __forceinline__ uint saturate_cast<uint>(float v)
     {
         return __float2uint_rn(v);

modules/gpu/include/opencv2/gpu/device/utility.hpp

@@ -45,7 +45,6 @@
 
 #include "saturate_cast.hpp"
 #include "datamov_utils.hpp"
-#include "detail/reduction_detail.hpp"
 
 namespace cv { namespace gpu { namespace device
 {
@@ -156,29 +155,6 @@ namespace cv { namespace gpu { namespace device
         }
     };
 
-    ///////////////////////////////////////////////////////////////////////////////
-    // Reduction
-
-    template <int n, typename T, typename Op> __device__ __forceinline__ void reduce(volatile T* data, T& partial_reduction, int tid, const Op& op)
-    {
-        StaticAssert<n >= 8 && n <= 512>::check();
-        utility_detail::ReductionDispatcher<n <= 64>::reduce<n>(data, partial_reduction, tid, op);
-    }
-
-    template <int n, typename T, typename V, typename Pred>
-    __device__ __forceinline__ void reducePredVal(volatile T* sdata, T& myData, V* sval, V& myVal, int tid, const Pred& pred)
-    {
-        StaticAssert<n >= 8 && n <= 512>::check();
-        utility_detail::PredValReductionDispatcher<n <= 64>::reduce<n>(myData, myVal, sdata, sval, tid, pred);
-    }
-
-    template <int n, typename T, typename V1, typename V2, typename Pred>
-    __device__ __forceinline__ void reducePredVal2(volatile T* sdata, T& myData, V1* sval1, V1& myVal1, V2* sval2, V2& myVal2, int tid, const Pred& pred)
-    {
-        StaticAssert<n >= 8 && n <= 512>::check();
-        utility_detail::PredVal2ReductionDispatcher<n <= 64>::reduce<n>(myData, myVal1, myVal2, sdata, sval1, sval2, tid, pred);
-    }
-
     ///////////////////////////////////////////////////////////////////////////////
     // Solve linear system
 

modules/gpu/include/opencv2/gpu/device/vec_distance.hpp

@@ -43,7 +43,7 @@
 #ifndef __OPENCV_GPU_VEC_DISTANCE_HPP__
 #define __OPENCV_GPU_VEC_DISTANCE_HPP__
 
-#include "utility.hpp"
+#include "reduce.hpp"
 #include "functional.hpp"
 #include "detail/vec_distance_detail.hpp"
 
@@ -63,7 +63,7 @@ namespace cv { namespace gpu { namespace device
 
         template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(int* smem, int tid)
         {
-            reduce<THREAD_DIM>(smem, mySum, tid, plus<volatile int>());
+            reduce<THREAD_DIM>(smem, mySum, tid, plus<int>());
         }
 
         __device__ __forceinline__ operator int() const
@@ -87,7 +87,7 @@ namespace cv { namespace gpu { namespace device
 
         template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(float* smem, int tid)
         {
-            reduce<THREAD_DIM>(smem, mySum, tid, plus<volatile float>());
+            reduce<THREAD_DIM>(smem, mySum, tid, plus<float>());
         }
 
         __device__ __forceinline__ operator float() const
@@ -113,7 +113,7 @@ namespace cv { namespace gpu { namespace device
 
         template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(float* smem, int tid)
         {
-            reduce<THREAD_DIM>(smem, mySum, tid, plus<volatile float>());
+            reduce<THREAD_DIM>(smem, mySum, tid, plus<float>());
         }
 
         __device__ __forceinline__ operator float() const
@@ -138,7 +138,7 @@ namespace cv { namespace gpu { namespace device
 
         template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(int* smem, int tid)
         {
-            reduce<THREAD_DIM>(smem, mySum, tid, plus<volatile int>());
+            reduce<THREAD_DIM>(smem, mySum, tid, plus<int>());
         }
 
         __device__ __forceinline__ operator int() const

modules/gpu/include/opencv2/gpu/device/warp_shuffle.hpp

+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifndef __OPENCV_GPU_WARP_SHUFFLE_HPP__
+#define __OPENCV_GPU_WARP_SHUFFLE_HPP__
+
+namespace cv { namespace gpu { namespace device
+{
+    template <typename T>
+    __device__ __forceinline__ T shfl(T val, int srcLane, int width = warpSize)
+    {
+    #if __CUDA_ARCH__ >= 300
+        return __shfl(val, srcLane, width);
+    #else
+        return T();
+    #endif
+    }
+    __device__ __forceinline__ unsigned int shfl(unsigned int val, int srcLane, int width = warpSize)
+    {
+    #if __CUDA_ARCH__ >= 300
+        return (unsigned int) __shfl((int) val, srcLane, width);
+    #else
+        return 0;
+    #endif
+    }
+    __device__ __forceinline__ double shfl(double val, int srcLane, int width = warpSize)
+    {
+    #if __CUDA_ARCH__ >= 300
+        int lo = __double2loint(val);
+        int hi = __double2hiint(val);
+
+        lo = __shfl(lo, srcLane, width);
+        hi = __shfl(hi, srcLane, width);
+
+        return __hiloint2double(hi, lo);
+    #else
+        return 0.0;
+    #endif
+    }
+
+    template <typename T>
+    __device__ __forceinline__ T shfl_down(T val, unsigned int delta, int width = warpSize)
+    {
+    #if __CUDA_ARCH__ >= 300
+        return __shfl_down(val, delta, width);
+    #else
+        return T();
+    #endif
+    }
+    __device__ __forceinline__ unsigned int shfl_down(unsigned int val, unsigned int delta, int width = warpSize)
+    {
+    #if __CUDA_ARCH__ >= 300
+        return (unsigned int) __shfl_down((int) val, delta, width);
+    #else
+        return 0;
+    #endif
+    }
+    __device__ __forceinline__ double shfl_down(double val, unsigned int delta, int width = warpSize)
+    {
+    #if __CUDA_ARCH__ >= 300
+        int lo = __double2loint(val);
+        int hi = __double2hiint(val);
+
+        lo = __shfl_down(lo, delta, width);
+        hi = __shfl_down(hi, delta, width);
+
+        return __hiloint2double(hi, lo);
+    #else
+        return 0.0;
+    #endif
+    }
+
+    template <typename T>
+    __device__ __forceinline__ T shfl_up(T val, unsigned int delta, int width = warpSize)
+    {
+    #if __CUDA_ARCH__ >= 300
+        return __shfl_up(val, delta, width);
+    #else
+        return T();
+    #endif
+    }
+    __device__ __forceinline__ unsigned int shfl_up(unsigned int val, unsigned int delta, int width = warpSize)
+    {
+    #if __CUDA_ARCH__ >= 300
+        return (unsigned int) __shfl_up((int) val, delta, width);
+    #else
+        return 0;
+    #endif
+    }
+    __device__ __forceinline__ double shfl_up(double val, unsigned int delta, int width = warpSize)
+    {
+    #if __CUDA_ARCH__ >= 300
+        int lo = __double2loint(val);
+        int hi = __double2hiint(val);
+
+        lo = __shfl_up(lo, delta, width);
+        hi = __shfl_up(hi, delta, width);
+
+        return __hiloint2double(hi, lo);
+    #else
+        return 0.0;
+    #endif
+    }
+}}}
+
+#endif // __OPENCV_GPU_WARP_SHUFFLE_HPP__

modules/gpu/src/arithm.cpp

@@ -68,11 +68,16 @@ void cv::gpu::polarToCart(const GpuMat&, const GpuMat&, GpuMat&, GpuMat&, bool,
 void cv::gpu::gemm(const GpuMat& src1, const GpuMat& src2, double alpha, const GpuMat& src3, double beta, GpuMat& dst, int flags, Stream& stream)
 {
 #ifndef HAVE_CUBLAS
-    (void)src1; (void)src2; (void)alpha; (void)src3; (void)beta; (void)dst; (void)flags; (void)stream;
+    (void)src1;
+    (void)src2;
+    (void)alpha;
+    (void)src3;
+    (void)beta;
+    (void)dst;
+    (void)flags;
+    (void)stream;
     CV_Error(CV_StsNotImplemented, "The library was build without CUBLAS");
-
 #else
-
     // CUBLAS works with column-major matrices
 
     CV_Assert(src1.type() == CV_32FC1 || src1.type() == CV_32FC2 || src1.type() == CV_64FC1 || src1.type() == CV_64FC2);
@@ -80,7 +85,7 @@ void cv::gpu::gemm(const GpuMat& src1, const GpuMat& src2, double alpha, const G
 
     if (src1.depth() == CV_64F)
     {
-        if (!TargetArchs::builtWith(NATIVE_DOUBLE) || !DeviceInfo().supports(NATIVE_DOUBLE))
+        if (!deviceSupports(NATIVE_DOUBLE))
             CV_Error(CV_StsUnsupportedFormat, "The device doesn't support double");
     }
 
@@ -188,7 +193,6 @@ void cv::gpu::gemm(const GpuMat& src1, const GpuMat& src2, double alpha, const G
     }
 
     cublasSafeCall( cublasDestroy_v2(handle) );
-
 #endif
 }
 
@@ -227,7 +231,7 @@ void cv::gpu::transpose(const GpuMat& src, GpuMat& dst, Stream& s)
     }
     else // if (src.elemSize() == 8)
     {
-        if (!TargetArchs::builtWith(NATIVE_DOUBLE) || !DeviceInfo().supports(NATIVE_DOUBLE))
+        if (!deviceSupports(NATIVE_DOUBLE))
             CV_Error(CV_StsUnsupportedFormat, "The device doesn't support double");
 
         NppStStreamHandler h(stream);

modules/gpu/src/cuda/calib3d.cu

@@ -42,9 +42,10 @@
 
 #if !defined CUDA_DISABLER
 
-#include "internal_shared.hpp"
+#include "opencv2/gpu/device/common.hpp"
 #include "opencv2/gpu/device/transform.hpp"
 #include "opencv2/gpu/device/functional.hpp"
+#include "opencv2/gpu/device/reduce.hpp"
 
 namespace cv { namespace gpu { namespace device
 {
@@ -66,6 +67,8 @@ namespace cv { namespace gpu { namespace device
                         crot1.x * p.x + crot1.y * p.y + crot1.z * p.z + ctransl.y,
                         crot2.x * p.x + crot2.y * p.y + crot2.z * p.z + ctransl.z);
             }
+            __device__ __forceinline__ TransformOp() {}
+            __device__ __forceinline__ TransformOp(const TransformOp&) {}
         };
 
         void call(const PtrStepSz<float3> src, const float* rot,
@@ -103,6 +106,8 @@ namespace cv { namespace gpu { namespace device
                         (cproj0.x * t.x + cproj0.y * t.y) / t.z + cproj0.z,
                         (cproj1.x * t.x + cproj1.y * t.y) / t.z + cproj1.z);
             }
+            __device__ __forceinline__ ProjectOp() {}
+            __device__ __forceinline__ ProjectOp(const ProjectOp&) {}
         };
 
         void call(const PtrStepSz<float3> src, const float* rot,
@@ -134,6 +139,7 @@ namespace cv { namespace gpu { namespace device
             return x * x;
         }
 
+        template <int BLOCK_SIZE>
         __global__ void computeHypothesisScoresKernel(
                 const int num_points, const float3* object, const float2* image,
                 const float dist_threshold, int* g_num_inliers)
@@ -156,19 +162,11 @@ namespace cv { namespace gpu { namespace device
                     ++num_inliers;
             }
 
-            extern __shared__ float s_num_inliers[];
-            s_num_inliers[threadIdx.x] = num_inliers;
-            __syncthreads();
-
-            for (int step = blockDim.x / 2; step > 0; step >>= 1)
-            {
-                if (threadIdx.x < step)
-                    s_num_inliers[threadIdx.x] += s_num_inliers[threadIdx.x + step];
-                __syncthreads();
-            }
+            __shared__ int s_num_inliers[BLOCK_SIZE];
+            reduce<BLOCK_SIZE>(s_num_inliers, num_inliers, threadIdx.x, plus<int>());
 
             if (threadIdx.x == 0)
-                g_num_inliers[blockIdx.x] = s_num_inliers[0];
+                g_num_inliers[blockIdx.x] = num_inliers;
         }
 
         void computeHypothesisScores(
@@ -181,9 +179,8 @@ namespace cv { namespace gpu { namespace device
 
             dim3 threads(256);
             dim3 grid(num_hypotheses);
-            int smem_size = threads.x * sizeof(float);
 
-            computeHypothesisScoresKernel<<<grid, threads, smem_size>>>(
+            computeHypothesisScoresKernel<256><<<grid, threads>>>(
                     num_points, object, image, dist_threshold, hypothesis_scores);
             cudaSafeCall( cudaGetLastError() );
 
@@ -193,4 +190,4 @@ namespace cv { namespace gpu { namespace device
 }}} // namespace cv { namespace gpu { namespace device
 
 
-#endif /* CUDA_DISABLER */
\ No newline at end of file
+#endif /* CUDA_DISABLER */

modules/gpu/src/cuda/fgd_bgfg.cu

@@ -46,6 +46,8 @@
 #include "opencv2/gpu/device/vec_math.hpp"
 #include "opencv2/gpu/device/limits.hpp"
 #include "opencv2/gpu/device/utility.hpp"
+#include "opencv2/gpu/device/reduce.hpp"
+#include "opencv2/gpu/device/functional.hpp"
 #include "fgd_bgfg_common.hpp"
 
 using namespace cv::gpu;
@@ -181,57 +183,8 @@ namespace bgfg
         __shared__ unsigned int data1[MERGE_THREADBLOCK_SIZE];
         __shared__ unsigned int data2[MERGE_THREADBLOCK_SIZE];
 
-        data0[threadIdx.x] = sum0;
-        data1[threadIdx.x] = sum1;
-        data2[threadIdx.x] = sum2;
-        __syncthreads();
-
-        if (threadIdx.x < 128)
-        {
-            data0[threadIdx.x] = sum0 += data0[threadIdx.x + 128];
-            data1[threadIdx.x] = sum1 += data1[threadIdx.x + 128];
-            data2[threadIdx.x] = sum2 += data2[threadIdx.x + 128];
-        }
-        __syncthreads();
-
-        if (threadIdx.x < 64)
-        {
-            data0[threadIdx.x] = sum0 += data0[threadIdx.x + 64];
-            data1[threadIdx.x] = sum1 += data1[threadIdx.x + 64];
-            data2[threadIdx.x] = sum2 += data2[threadIdx.x + 64];
-        }
-        __syncthreads();
-
-        if (threadIdx.x < 32)
-        {
-            volatile unsigned int* vdata0 = data0;
-            volatile unsigned int* vdata1 = data1;
-            volatile unsigned int* vdata2 = data2;
-
-            vdata0[threadIdx.x] = sum0 += vdata0[threadIdx.x + 32];
-            vdata1[threadIdx.x] = sum1 += vdata1[threadIdx.x + 32];
-            vdata2[threadIdx.x] = sum2 += vdata2[threadIdx.x + 32];
-
-            vdata0[threadIdx.x] = sum0 += vdata0[threadIdx.x + 16];
-            vdata1[threadIdx.x] = sum1 += vdata1[threadIdx.x + 16];
-            vdata2[threadIdx.x] = sum2 += vdata2[threadIdx.x + 16];
-
-            vdata0[threadIdx.x] = sum0 += vdata0[threadIdx.x + 8];
-            vdata1[threadIdx.x] = sum1 += vdata1[threadIdx.x + 8];
-            vdata2[threadIdx.x] = sum2 += vdata2[threadIdx.x + 8];
-
-            vdata0[threadIdx.x] = sum0 += vdata0[threadIdx.x + 4];
-            vdata1[threadIdx.x] = sum1 += vdata1[threadIdx.x + 4];
-            vdata2[threadIdx.x] = sum2 += vdata2[threadIdx.x + 4];
-
-            vdata0[threadIdx.x] = sum0 += vdata0[threadIdx.x + 2];
-            vdata1[threadIdx.x] = sum1 += vdata1[threadIdx.x + 2];
-            vdata2[threadIdx.x] = sum2 += vdata2[threadIdx.x + 2];
-
-            vdata0[threadIdx.x] = sum0 += vdata0[threadIdx.x + 1];
-            vdata1[threadIdx.x] = sum1 += vdata1[threadIdx.x + 1];
-            vdata2[threadIdx.x] = sum2 += vdata2[threadIdx.x + 1];
-        }
+        plus<unsigned int> op;
+        reduce<MERGE_THREADBLOCK_SIZE>(smem_tuple(data0, data1, data2), thrust::tie(sum0, sum1, sum2), threadIdx.x, thrust::make_tuple(op, op, op));
 
         if(threadIdx.x == 0)
         {
@@ -245,9 +198,9 @@ namespace bgfg
     void calcDiffHistogram_gpu(PtrStepSzb prevFrame, PtrStepSzb curFrame,
                                unsigned int* hist0, unsigned int* hist1, unsigned int* hist2,
                                unsigned int* partialBuf0, unsigned int* partialBuf1, unsigned int* partialBuf2,
-                               int cc, cudaStream_t stream)
+                               bool cc20, cudaStream_t stream)
     {
-        const int HISTOGRAM_WARP_COUNT = cc < 20 ? 4 : 6;
+        const int HISTOGRAM_WARP_COUNT = cc20 ? 6 : 4;
         const int HISTOGRAM_THREADBLOCK_SIZE = HISTOGRAM_WARP_COUNT * WARP_SIZE;
 
         calcPartialHistogram<PT, CT><<<PARTIAL_HISTOGRAM_COUNT, HISTOGRAM_THREADBLOCK_SIZE, 0, stream>>>(
@@ -261,10 +214,10 @@ namespace bgfg
             cudaSafeCall( cudaDeviceSynchronize() );
     }
 
-    template void calcDiffHistogram_gpu<uchar3, uchar3>(PtrStepSzb prevFrame, PtrStepSzb curFrame, unsigned int* hist0, unsigned int* hist1, unsigned int* hist2, unsigned int* partialBuf0, unsigned int* partialBuf1, unsigned int* partialBuf2, int cc, cudaStream_t stream);
-    template void calcDiffHistogram_gpu<uchar3, uchar4>(PtrStepSzb prevFrame, PtrStepSzb curFrame, unsigned int* hist0, unsigned int* hist1, unsigned int* hist2, unsigned int* partialBuf0, unsigned int* partialBuf1, unsigned int* partialBuf2, int cc, cudaStream_t stream);
-    template void calcDiffHistogram_gpu<uchar4, uchar3>(PtrStepSzb prevFrame, PtrStepSzb curFrame, unsigned int* hist0, unsigned int* hist1, unsigned int* hist2, unsigned int* partialBuf0, unsigned int* partialBuf1, unsigned int* partialBuf2, int cc, cudaStream_t stream);
-    template void calcDiffHistogram_gpu<uchar4, uchar4>(PtrStepSzb prevFrame, PtrStepSzb curFrame, unsigned int* hist0, unsigned int* hist1, unsigned int* hist2, unsigned int* partialBuf0, unsigned int* partialBuf1, unsigned int* partialBuf2, int cc, cudaStream_t stream);
+    template void calcDiffHistogram_gpu<uchar3, uchar3>(PtrStepSzb prevFrame, PtrStepSzb curFrame, unsigned int* hist0, unsigned int* hist1, unsigned int* hist2, unsigned int* partialBuf0, unsigned int* partialBuf1, unsigned int* partialBuf2, bool cc20, cudaStream_t stream);
+    template void calcDiffHistogram_gpu<uchar3, uchar4>(PtrStepSzb prevFrame, PtrStepSzb curFrame, unsigned int* hist0, unsigned int* hist1, unsigned int* hist2, unsigned int* partialBuf0, unsigned int* partialBuf1, unsigned int* partialBuf2, bool cc20, cudaStream_t stream);
+    template void calcDiffHistogram_gpu<uchar4, uchar3>(PtrStepSzb prevFrame, PtrStepSzb curFrame, unsigned int* hist0, unsigned int* hist1, unsigned int* hist2, unsigned int* partialBuf0, unsigned int* partialBuf1, unsigned int* partialBuf2, bool cc20, cudaStream_t stream);
+    template void calcDiffHistogram_gpu<uchar4, uchar4>(PtrStepSzb prevFrame, PtrStepSzb curFrame, unsigned int* hist0, unsigned int* hist1, unsigned int* hist2, unsigned int* partialBuf0, unsigned int* partialBuf1, unsigned int* partialBuf2, bool cc20, cudaStream_t stream);
 
     /////////////////////////////////////////////////////////////////////////
     // calcDiffThreshMask
@@ -845,4 +798,4 @@ namespace bgfg
     template void updateBackgroundModel_gpu<uchar4, uchar4, uchar4>(PtrStepSzb prevFrame, PtrStepSzb curFrame, PtrStepSzb Ftd, PtrStepSzb Fbd, PtrStepSzb foreground, PtrStepSzb background, int deltaC, int deltaCC, float alpha1, float alpha2, float alpha3, int N1c, int N1cc, int N2c, int N2cc, float T, cudaStream_t stream);
 }
 
-#endif /* CUDA_DISABLER */
\ No newline at end of file
+#endif /* CUDA_DISABLER */

modules/gpu/src/cuda/fgd_bgfg_common.hpp

@@ -125,7 +125,7 @@ namespace bgfg
     void calcDiffHistogram_gpu(cv::gpu::PtrStepSzb prevFrame, cv::gpu::PtrStepSzb curFrame,
                                unsigned int* hist0, unsigned int* hist1, unsigned int* hist2,
                                unsigned int* partialBuf0, unsigned int* partialBuf1, unsigned int* partialBuf2,
-                               int cc, cudaStream_t stream);
+                               bool cc20, cudaStream_t stream);
 
     template <typename PT, typename CT>
     void calcDiffThreshMask_gpu(cv::gpu::PtrStepSzb prevFrame, cv::gpu::PtrStepSzb curFrame, uchar3 bestThres, cv::gpu::PtrStepSzb changeMask, cudaStream_t stream);

modules/gpu/src/cuda/global_motion.cu

@@ -43,12 +43,10 @@
 
 #if !defined CUDA_DISABLER
 
-#include "thrust/device_ptr.h"
-#include "thrust/remove.h"
-#include "thrust/functional.h"
-#include "internal_shared.hpp"
-
-using namespace thrust;
+#include <thrust/device_ptr.h>
+#include <thrust/remove.h>
+#include <thrust/functional.h>
+#include "opencv2/gpu/device/common.hpp"
 
 namespace cv { namespace gpu { namespace device { namespace globmotion {
 
@@ -61,10 +59,10 @@ int compactPoints(int N, float *points0, float *points1, const uchar *mask)
     thrust::device_ptr<float2> dpoints1((float2*)points1);
     thrust::device_ptr<const uchar> dmask(mask);
 
-    return thrust::remove_if(thrust::make_zip_iterator(thrust::make_tuple(dpoints0, dpoints1)),
+    return (int)(thrust::remove_if(thrust::make_zip_iterator(thrust::make_tuple(dpoints0, dpoints1)),
                              thrust::make_zip_iterator(thrust::make_tuple(dpoints0 + N, dpoints1 + N)),
                              dmask, thrust::not1(thrust::identity<uchar>()))
-           - make_zip_iterator(make_tuple(dpoints0, dpoints1));
+           - thrust::make_zip_iterator(make_tuple(dpoints0, dpoints1)));
 }
 
 
@@ -117,4 +115,4 @@ void calcWobbleSuppressionMaps(
 }}}}
 
 
-#endif /* CUDA_DISABLER */
\ No newline at end of file
+#endif /* CUDA_DISABLER */

modules/gpu/src/cuda/orb.cu

@@ -47,10 +47,11 @@
 
 #if !defined CUDA_DISABLER
 
+#include <thrust/device_ptr.h>
 #include <thrust/sort.h>
 
 #include "opencv2/gpu/device/common.hpp"
-#include "opencv2/gpu/device/utility.hpp"
+#include "opencv2/gpu/device/reduce.hpp"
 #include "opencv2/gpu/device/functional.hpp"
 
 namespace cv { namespace gpu { namespace device
@@ -75,9 +76,9 @@ namespace cv { namespace gpu { namespace device
 
         __global__ void HarrisResponses(const PtrStepb img, const short2* loc_, float* response, const int npoints, const int blockSize, const float harris_k)
         {
-            __shared__ int smem[8 * 32];
-
-            volatile int* srow = smem + threadIdx.y * blockDim.x;
+            __shared__ int smem0[8 * 32];
+            __shared__ int smem1[8 * 32];
+            __shared__ int smem2[8 * 32];
 
             const int ptidx = blockIdx.x * blockDim.y + threadIdx.y;
 
@@ -109,9 +110,12 @@ namespace cv { namespace gpu { namespace device
                     c += Ix * Iy;
                 }
 
-                reduce<32>(srow, a, threadIdx.x, plus<volatile int>());
-                reduce<32>(srow, b, threadIdx.x, plus<volatile int>());
-                reduce<32>(srow, c, threadIdx.x, plus<volatile int>());
+                int* srow0 = smem0 + threadIdx.y * blockDim.x;
+                int* srow1 = smem1 + threadIdx.y * blockDim.x;
+                int* srow2 = smem2 + threadIdx.y * blockDim.x;
+
+                plus<int> op;
+                reduce<32>(smem_tuple(srow0, srow1, srow2), thrust::tie(a, b, c), threadIdx.x, thrust::make_tuple(op, op, op));
 
                 if (threadIdx.x == 0)
                 {
@@ -151,9 +155,13 @@ namespace cv { namespace gpu { namespace device
 
         __global__ void IC_Angle(const PtrStepb image, const short2* loc_, float* angle, const int npoints, const int half_k)
         {
-            __shared__ int smem[8 * 32];
+            __shared__ int smem0[8 * 32];
+            __shared__ int smem1[8 * 32];
+
+            int* srow0 = smem0 + threadIdx.y * blockDim.x;
+            int* srow1 = smem1 + threadIdx.y * blockDim.x;
 
-            volatile int* srow = smem + threadIdx.y * blockDim.x;
+            plus<int> op;
 
             const int ptidx = blockIdx.x * blockDim.y + threadIdx.y;
 
@@ -167,7 +175,7 @@ namespace cv { namespace gpu { namespace device
                 for (int u = threadIdx.x - half_k; u <= half_k; u += blockDim.x)
                     m_10 += u * image(loc.y, loc.x + u);
 
-                reduce<32>(srow, m_10, threadIdx.x, plus<volatile int>());
+                reduce<32>(srow0, m_10, threadIdx.x, op);
 
                 for (int v = 1; v <= half_k; ++v)
                 {
@@ -185,8 +193,7 @@ namespace cv { namespace gpu { namespace device
                         m_sum += u * (val_plus + val_minus);
                     }
 
-                    reduce<32>(srow, v_sum, threadIdx.x, plus<volatile int>());
-                    reduce<32>(srow, m_sum, threadIdx.x, plus<volatile int>());
+                    reduce<32>(smem_tuple(srow0, srow1), thrust::tie(v_sum, m_sum), threadIdx.x, thrust::make_tuple(op, op));
 
                     m_10 += m_sum;
                     m_01 += v * v_sum;
@@ -419,4 +426,4 @@ namespace cv { namespace gpu { namespace device
     }
 }}}
 
-#endif /* CUDA_DISABLER */
\ No newline at end of file
+#endif /* CUDA_DISABLER */

modules/gpu/src/fgd_bgfg.cpp

@@ -336,7 +336,7 @@ namespace
 {
     void calcDiffHistogram(const cv::gpu::GpuMat& prevFrame, const cv::gpu::GpuMat& curFrame, cv::gpu::GpuMat& hist, cv::gpu::GpuMat& histBuf)
     {
-        typedef void (*func_t)(cv::gpu::PtrStepSzb prevFrame, cv::gpu::PtrStepSzb curFrame, unsigned int* hist0, unsigned int* hist1, unsigned int* hist2, unsigned int* partialBuf0, unsigned int* partialBuf1, unsigned int* partialBuf2, int cc, cudaStream_t stream);
+        typedef void (*func_t)(cv::gpu::PtrStepSzb prevFrame, cv::gpu::PtrStepSzb curFrame, unsigned int* hist0, unsigned int* hist1, unsigned int* hist2, unsigned int* partialBuf0, unsigned int* partialBuf1, unsigned int* partialBuf2, bool cc20, cudaStream_t stream);
         static const func_t funcs[4][4] =
         {
             {0,0,0,0},
@@ -348,14 +348,11 @@ namespace
         hist.create(3, 256, CV_32SC1);
         histBuf.create(3, bgfg::PARTIAL_HISTOGRAM_COUNT * bgfg::HISTOGRAM_BIN_COUNT, CV_32SC1);
 
-        cv::gpu::DeviceInfo devInfo;
-        int cc = devInfo.majorVersion() * 10 + devInfo.minorVersion();
-
         funcs[prevFrame.channels() - 1][curFrame.channels() - 1](
                     prevFrame, curFrame,
                     hist.ptr<unsigned int>(0), hist.ptr<unsigned int>(1), hist.ptr<unsigned int>(2),
                     histBuf.ptr<unsigned int>(0), histBuf.ptr<unsigned int>(1), histBuf.ptr<unsigned int>(2),
-                    cc, 0);
+                    cv::gpu::deviceSupports(cv::gpu::FEATURE_SET_COMPUTE_20), 0);
     }
 
     void calcRelativeVariance(unsigned int hist[3 * 256], double relativeVariance[3][bgfg::HISTOGRAM_BIN_COUNT])
@@ -526,15 +523,15 @@ namespace
 
         size_t total = all_contours.size();
 
-        _contours.create(total, 1, 0, -1, true);
+        _contours.create((int) total, 1, 0, -1, true);
 
         cv::SeqIterator<CvSeq*> it = all_contours.begin();
         for (size_t i = 0; i < total; ++i, ++it)
         {
             CvSeq* c = *it;
             ((CvContour*)c)->color = (int)i;
-            _contours.create((int)c->total, 1, CV_32SC2, i, true);
-            cv::Mat ci = _contours.getMat(i);
+            _contours.create((int)c->total, 1, CV_32SC2, (int)i, true);
+            cv::Mat ci = _contours.getMat((int)i);
             CV_Assert( ci.isContinuous() );
             cvCvtSeqToArray(c, ci.data);
         }

modules/gpu/src/hough.cpp

@@ -294,9 +294,8 @@ void cv::gpu::HoughCircles(const GpuMat& src, GpuMat& circles, HoughCirclesBuf&
 
     ensureSizeIsEnough(1, maxCircles, CV_32FC3, circles);
 
-    DeviceInfo devInfo;
     const int circlesCount = circlesAccumRadius_gpu(centers, centersCount, srcPoints, pointsCount, circles.ptr<float3>(), maxCircles,
-                                                    dp, minRadius, maxRadius, votesThreshold, devInfo.supports(FEATURE_SET_COMPUTE_20));
+                                                    dp, minRadius, maxRadius, votesThreshold, deviceSupports(FEATURE_SET_COMPUTE_20));
 
     if (circlesCount > 0)
         circles.cols = circlesCount;
@@ -531,7 +530,7 @@ namespace
         const func_t func = funcs[dx.depth()];
         CV_Assert(func != 0);
 
-        edgePointList.cols = edgePointList.step / sizeof(int);
+        edgePointList.cols = (int) (edgePointList.step / sizeof(int));
         ensureSizeIsEnough(2, edges.size().area(), CV_32SC1, edgePointList);
 
         edgePointList.cols = func(edges, dx, dy, edgePointList.ptr<unsigned int>(0), edgePointList.ptr<float>(1));