added mask support into gpu::minMax

diff --git a/modules/gpu/include/opencv2/gpu/gpu.hpp b/modules/gpu/include/opencv2/gpu/gpu.hpp
index dadad00..13f237b 100644 (file)
--- a/modules/gpu/include/opencv2/gpu/gpu.hpp
+++ b/modules/gpu/include/opencv2/gpu/gpu.hpp
@@ -425,10 +425,10 @@ namespace cv
         CV_EXPORTS Scalar sum(const GpuMat& m);\r
 \r
         //! finds global minimum and maximum array elements and returns their values\r
-        CV_EXPORTS void minMax(const GpuMat& src, double* minVal, double* maxVal=0);\r
+        CV_EXPORTS void minMax(const GpuMat& src, double* minVal, double* maxVal=0, const GpuMat& mask=GpuMat());\r
 \r
         //! finds global minimum and maximum array elements and returns their values\r
-        CV_EXPORTS void minMax(const GpuMat& src, double* minVal, double* maxVal, GpuMat& buf);\r
+        CV_EXPORTS void minMax(const GpuMat& src, double* minVal, double* maxVal, const GpuMat& mask, GpuMat& buf);\r
 \r
         //! finds global minimum and maximum array elements and returns their values with locations\r
         CV_EXPORTS void minMaxLoc(const GpuMat& src, double* minVal, double* maxVal=0, Point* minLoc=0, Point* maxLoc=0);\r

diff --git a/modules/gpu/src/arithm.cpp b/modules/gpu/src/arithm.cpp
index df7c550..c7ca547 100644 (file)
--- a/modules/gpu/src/arithm.cpp
+++ b/modules/gpu/src/arithm.cpp
@@ -65,8 +65,8 @@ double cv::gpu::norm(const GpuMat&, int) { throw_nogpu(); return 0.0; }
 double cv::gpu::norm(const GpuMat&, const GpuMat&, int) { throw_nogpu(); return 0.0; }\r
 void cv::gpu::flip(const GpuMat&, GpuMat&, int) { throw_nogpu(); }\r
 Scalar cv::gpu::sum(const GpuMat&) { throw_nogpu(); return Scalar(); }\r
-void cv::gpu::minMax(const GpuMat&, double*, double*) { throw_nogpu(); }\r
-void cv::gpu::minMax(const GpuMat&, double*, double*, GpuMat&) { throw_nogpu(); }\r
+void cv::gpu::minMax(const GpuMat&, double*, double*, const GpuMat&) { throw_nogpu(); }\r
+void cv::gpu::minMax(const GpuMat&, double*, double*, const GpuMat&, GpuMat&) { throw_nogpu(); }\r
 void cv::gpu::minMaxLoc(const GpuMat&, double*, double*, Point*, Point*) { throw_nogpu(); }\r
 void cv::gpu::minMaxLoc(const GpuMat&, double*, double*, Point*, Point*, GpuMat&, GpuMat&) { throw_nogpu(); }\r
 int cv::gpu::countNonZero(const GpuMat&) { throw_nogpu(); return 0; }\r
@@ -502,62 +502,68 @@ namespace cv { namespace gpu { namespace mathfunc { namespace minmax {
     void min_max_caller(const DevMem2D src, double* minval, double* maxval, PtrStep buf);\r
 \r
     template <typename T> \r
-    void min_max_caller_2steps(const DevMem2D src, double* minval, double* maxval, PtrStep buf);\r
+    void min_max_mask_caller(const DevMem2D src, const PtrStep mask, double* minval, double* maxval, PtrStep buf);\r
+\r
+    template <typename T> \r
+    void min_max_multipass_caller(const DevMem2D src, double* minval, double* maxval, PtrStep buf);\r
+\r
+    template <typename T> \r
+    void min_max_mask_multipass_caller(const DevMem2D src, const PtrStep mask, double* minval, double* maxval, PtrStep buf);\r
 \r
 }}}}\r
 \r
-void cv::gpu::minMax(const GpuMat& src, double* minVal, double* maxVal)\r
+void cv::gpu::minMax(const GpuMat& src, double* minVal, double* maxVal, const GpuMat& mask)\r
 {\r
     GpuMat buf;\r
-    minMax(src, minVal, maxVal, buf);\r
+    minMax(src, minVal, maxVal, mask, buf);\r
 }\r
 \r
-\r
-void cv::gpu::minMax(const GpuMat& src, double* minVal, double* maxVal, GpuMat& buf)\r
+void cv::gpu::minMax(const GpuMat& src, double* minVal, double* maxVal, const GpuMat& mask, GpuMat& buf)\r
 {\r
     using namespace mathfunc::minmax;\r
 \r
+    typedef void (*Caller)(const DevMem2D, double*, double*, PtrStep);\r
+    static const Caller callers[2][7] = \r
+    { { min_max_multipass_caller<unsigned char>, min_max_multipass_caller<signed char>, \r
+        min_max_multipass_caller<unsigned short>, min_max_multipass_caller<signed short>, \r
+        min_max_multipass_caller<int>, min_max_multipass_caller<float>, 0 },\r
+      { min_max_caller<unsigned char>, min_max_caller<signed char>, \r
+        min_max_caller<unsigned short>, min_max_caller<signed short>, \r
+        min_max_caller<int>, min_max_caller<float>, min_max_caller<double> } };\r
+\r
+    typedef void (*MaskedCaller)(const DevMem2D, const PtrStep, double*, double*, PtrStep);\r
+    static const MaskedCaller masked_callers[2][7] = \r
+    { { min_max_mask_multipass_caller<unsigned char>, min_max_mask_multipass_caller<signed char>, \r
+        min_max_mask_multipass_caller<unsigned short>, min_max_mask_multipass_caller<signed short>, \r
+        min_max_mask_multipass_caller<int>, min_max_mask_multipass_caller<float>, 0 },\r
+      { min_max_mask_caller<unsigned char>, min_max_mask_caller<signed char>, \r
+        min_max_mask_caller<unsigned short>, min_max_mask_caller<signed short>, \r
+        min_max_mask_caller<int>, min_max_mask_caller<float>, \r
+        min_max_mask_caller<double> } };\r
+\r
+\r
+    CV_Assert(src.channels() == 1);\r
+    CV_Assert(mask.empty() || (mask.type() == CV_8U && src.size() == mask.size()));\r
+    CV_Assert(src.type() != CV_64F || hasNativeDoubleSupport(getDevice()));\r
+\r
     double minVal_; if (!minVal) minVal = &minVal_;\r
     double maxVal_; if (!maxVal) maxVal = &maxVal_;\r
-\r
-    GpuMat src_ = src.reshape(1);\r
     \r
     Size bufSize;\r
     get_buf_size_required(src.elemSize(), bufSize.width, bufSize.height);\r
     buf.create(bufSize, CV_8U);\r
 \r
-    int device = getDevice();\r
-    if (hasAtomicsSupport(device))\r
+    if (mask.empty())\r
     {\r
-        switch (src_.type())\r
-        {\r
-        case CV_8U: min_max_caller<unsigned char>(src_, minVal, maxVal, buf); break;\r
-        case CV_8S: min_max_caller<signed char>(src_, minVal, maxVal, buf); break;\r
-        case CV_16U: min_max_caller<unsigned short>(src_, minVal, maxVal, buf); break;\r
-        case CV_16S: min_max_caller<signed short>(src_, minVal, maxVal, buf); break;\r
-        case CV_32S: min_max_caller<int>(src_, minVal, maxVal, buf); break;\r
-        case CV_32F: min_max_caller<float>(src_, minVal, maxVal, buf); break;\r
-        case CV_64F: \r
-            if (hasNativeDoubleSupport(device))\r
-            {\r
-                min_max_caller<double>(src_, minVal, maxVal, buf); \r
-                break;\r
-            }\r
-        default: CV_Error(CV_StsBadArg, "minMax: unsupported type");\r
-        }\r
+        Caller caller = callers[hasAtomicsSupport(getDevice())][src.type()];\r
+        if (!caller) CV_Error(CV_StsBadArg, "minMax: unsupported type");\r
+        caller(src, minVal, maxVal, buf);\r
     }\r
     else\r
     {\r
-        switch (src_.type())\r
-        {\r
-        case CV_8U: min_max_caller_2steps<unsigned char>(src_, minVal, maxVal, buf); break;\r
-        case CV_8S: min_max_caller_2steps<signed char>(src_, minVal, maxVal, buf); break;\r
-        case CV_16U: min_max_caller_2steps<unsigned short>(src_, minVal, maxVal, buf); break;\r
-        case CV_16S: min_max_caller_2steps<signed short>(src_, minVal, maxVal, buf); break;\r
-        case CV_32S: min_max_caller_2steps<int>(src_, minVal, maxVal, buf); break;\r
-        case CV_32F: min_max_caller_2steps<float>(src_, minVal, maxVal, buf); break;\r
-        default: CV_Error(CV_StsBadArg, "minMax: unsupported type");\r
-        }\r
+        MaskedCaller caller = masked_callers[hasAtomicsSupport(getDevice())][src.type()];\r
+        if (!caller) CV_Error(CV_StsBadArg, "minMax: unsupported type");\r
+        caller(src, mask, minVal, maxVal, buf);\r
     }\r
 }\r
 \r
@@ -575,7 +581,7 @@ namespace cv { namespace gpu { namespace mathfunc { namespace minmaxloc {
                             int minloc[2], int maxloc[2], PtrStep valbuf, PtrStep locbuf);\r
 \r
     template <typename T> \r
-    void min_max_loc_caller_2steps(const DevMem2D src, double* minval, double* maxval, \r
+    void min_max_loc_multipass_caller(const DevMem2D src, double* minval, double* maxval, \r
                                    int minloc[2], int maxloc[2], PtrStep valbuf, PtrStep locbuf);\r
 \r
 }}}}\r
@@ -627,12 +633,12 @@ void cv::gpu::minMaxLoc(const GpuMat& src, double* minVal, double* maxVal, Point
     {\r
         switch (src.type())\r
         {\r
-        case CV_8U: min_max_loc_caller_2steps<unsigned char>(src, minVal, maxVal, minLoc_, maxLoc_, valbuf, locbuf); break;\r
-        case CV_8S: min_max_loc_caller_2steps<signed char>(src, minVal, maxVal, minLoc_, maxLoc_, valbuf, locbuf); break;\r
-        case CV_16U: min_max_loc_caller_2steps<unsigned short>(src, minVal, maxVal, minLoc_, maxLoc_, valbuf, locbuf); break;\r
-        case CV_16S: min_max_loc_caller_2steps<signed short>(src, minVal, maxVal, minLoc_, maxLoc_, valbuf, locbuf); break;\r
-        case CV_32S: min_max_loc_caller_2steps<int>(src, minVal, maxVal, minLoc_, maxLoc_, valbuf, locbuf); break;\r
-        case CV_32F: min_max_loc_caller_2steps<float>(src, minVal, maxVal, minLoc_, maxLoc_, valbuf, locbuf); break;\r
+        case CV_8U: min_max_loc_multipass_caller<unsigned char>(src, minVal, maxVal, minLoc_, maxLoc_, valbuf, locbuf); break;\r
+        case CV_8S: min_max_loc_multipass_caller<signed char>(src, minVal, maxVal, minLoc_, maxLoc_, valbuf, locbuf); break;\r
+        case CV_16U: min_max_loc_multipass_caller<unsigned short>(src, minVal, maxVal, minLoc_, maxLoc_, valbuf, locbuf); break;\r
+        case CV_16S: min_max_loc_multipass_caller<signed short>(src, minVal, maxVal, minLoc_, maxLoc_, valbuf, locbuf); break;\r
+        case CV_32S: min_max_loc_multipass_caller<int>(src, minVal, maxVal, minLoc_, maxLoc_, valbuf, locbuf); break;\r
+        case CV_32F: min_max_loc_multipass_caller<float>(src, minVal, maxVal, minLoc_, maxLoc_, valbuf, locbuf); break;\r
         default: CV_Error(CV_StsBadArg, "minMaxLoc: unsupported type");\r
         }\r
     }\r
@@ -652,7 +658,7 @@ namespace cv { namespace gpu { namespace mathfunc { namespace countnonzero {
     int count_non_zero_caller(const DevMem2D src, PtrStep buf);\r
 \r
     template <typename T> \r
-    int count_non_zero_caller_2steps(const DevMem2D src, PtrStep buf);\r
+    int count_non_zero_multipass_caller(const DevMem2D src, PtrStep buf);\r
 \r
 }}}}\r
 \r
@@ -691,12 +697,12 @@ int cv::gpu::countNonZero(const GpuMat& src, GpuMat& buf)
     {\r
         switch (src.type())\r
         {\r
-        case CV_8U: return count_non_zero_caller_2steps<unsigned char>(src, buf);\r
-        case CV_8S: return count_non_zero_caller_2steps<signed char>(src, buf);\r
-        case CV_16U: return count_non_zero_caller_2steps<unsigned short>(src, buf);\r
-        case CV_16S: return count_non_zero_caller_2steps<signed short>(src, buf);\r
-        case CV_32S: return count_non_zero_caller_2steps<int>(src, buf);\r
-        case CV_32F: return count_non_zero_caller_2steps<float>(src, buf);\r
+        case CV_8U: return count_non_zero_multipass_caller<unsigned char>(src, buf);\r
+        case CV_8S: return count_non_zero_multipass_caller<signed char>(src, buf);\r
+        case CV_16U: return count_non_zero_multipass_caller<unsigned short>(src, buf);\r
+        case CV_16S: return count_non_zero_multipass_caller<signed short>(src, buf);\r
+        case CV_32S: return count_non_zero_multipass_caller<int>(src, buf);\r
+        case CV_32F: return count_non_zero_multipass_caller<float>(src, buf);\r
         }\r
     }\r
 \r

diff --git a/modules/gpu/src/cuda/mathfunc.cu b/modules/gpu/src/cuda/mathfunc.cu
index ce927d3..3f515e6 100644 (file)
--- a/modules/gpu/src/cuda/mathfunc.cu
+++ b/modules/gpu/src/cuda/mathfunc.cu
@@ -480,8 +480,8 @@ namespace cv { namespace gpu { namespace mathfunc
     }\r
 \r
 \r
-    template <int nthreads, typename T>\r
-    __global__ void min_max_kernel(const DevMem2D src, T* minval, T* maxval)\r
+    template <int nthreads, typename T, typename Mask>\r
+    __global__ void min_max_kernel(const DevMem2D src, Mask mask, T* minval, T* maxval)\r
     {\r
         typedef typename MinMaxTypeTraits<T>::best_type best_type;\r
         __shared__ best_type sminval[nthreads];\r
@@ -491,17 +491,21 @@ namespace cv { namespace gpu { namespace mathfunc
         unsigned int y0 = blockIdx.y * blockDim.y * ctheight + threadIdx.y;\r
         unsigned int tid = threadIdx.y * blockDim.x + threadIdx.x;\r
 \r
-        T val;\r
         T mymin = numeric_limits_gpu<T>::max();\r
         T mymax = numeric_limits_gpu<T>::min();\r
-        for (unsigned int y = 0; y < ctheight && y0 + y * blockDim.y < src.rows; ++y)\r
+        unsigned int y_end = min(y0 + (ctheight - 1) * blockDim.y + 1, src.rows);\r
+        unsigned int x_end = min(x0 + (ctwidth - 1) * blockDim.x + 1, src.cols);\r
+        for (unsigned int y = y0; y < y_end; y += blockDim.y)\r
         {\r
-            const T* ptr = (const T*)src.ptr(y0 + y * blockDim.y);\r
-            for (unsigned int x = 0; x < ctwidth && x0 + x * blockDim.x < src.cols; ++x)\r
+            const T* src_row = (const T*)src.ptr(y);\r
+            for (unsigned int x = x0; x < x_end; x += blockDim.x)\r
             {\r
-                val = ptr[x0 + x * blockDim.x];\r
-                mymin = min(mymin, val);\r
-                mymax = max(mymax, val);\r
+                T val = src_row[x];\r
+                if (mask(y, x)) \r
+                { \r
+                    mymin = min(mymin, val); \r
+                    mymax = max(mymax, val); \r
+                }\r
             }\r
         }\r
 \r
@@ -560,6 +564,35 @@ namespace cv { namespace gpu { namespace mathfunc
 \r
    \r
     template <typename T>\r
+    void min_max_mask_caller(const DevMem2D src, const PtrStep mask, double* minval, double* maxval, PtrStep buf)\r
+    {\r
+        dim3 threads, grid;\r
+        estimate_thread_cfg(threads, grid);\r
+        estimate_kernel_consts(src.cols, src.rows, threads, grid);\r
+\r
+        T* minval_buf = (T*)buf.ptr(0);\r
+        T* maxval_buf = (T*)buf.ptr(1);\r
+\r
+        min_max_kernel<256, T, Mask8U><<<grid, threads>>>(src, Mask8U(mask), minval_buf, maxval_buf);\r
+        cudaSafeCall(cudaThreadSynchronize());\r
+\r
+        T minval_, maxval_;\r
+        cudaSafeCall(cudaMemcpy(&minval_, minval_buf, sizeof(T), cudaMemcpyDeviceToHost));\r
+        cudaSafeCall(cudaMemcpy(&maxval_, maxval_buf, sizeof(T), cudaMemcpyDeviceToHost));\r
+        *minval = minval_;\r
+        *maxval = maxval_;\r
+    }  \r
+\r
+    template void min_max_mask_caller<unsigned char>(const DevMem2D, const PtrStep, double*, double*, PtrStep);\r
+    template void min_max_mask_caller<signed char>(const DevMem2D, const PtrStep, double*, double*, PtrStep);\r
+    template void min_max_mask_caller<unsigned short>(const DevMem2D, const PtrStep, double*, double*, PtrStep);\r
+    template void min_max_mask_caller<signed short>(const DevMem2D, const PtrStep, double*, double*, PtrStep);\r
+    template void min_max_mask_caller<int>(const DevMem2D, const PtrStep, double*, double*, PtrStep);\r
+    template void min_max_mask_caller<float>(const DevMem2D, const PtrStep, double*, double*, PtrStep);\r
+    template void min_max_mask_caller<double>(const DevMem2D, const PtrStep, double*, double*, PtrStep);\r
+\r
+\r
+    template <typename T>\r
     void min_max_caller(const DevMem2D src, double* minval, double* maxval, PtrStep buf)\r
     {\r
         dim3 threads, grid;\r
@@ -569,7 +602,7 @@ namespace cv { namespace gpu { namespace mathfunc
         T* minval_buf = (T*)buf.ptr(0);\r
         T* maxval_buf = (T*)buf.ptr(1);\r
 \r
-        min_max_kernel<256, T><<<grid, threads>>>(src, minval_buf, maxval_buf);\r
+        min_max_kernel<256, T, MaskTrue><<<grid, threads>>>(src, MaskTrue(), minval_buf, maxval_buf);\r
         cudaSafeCall(cudaThreadSynchronize());\r
 \r
         T minval_, maxval_;\r
@@ -584,13 +617,12 @@ namespace cv { namespace gpu { namespace mathfunc
     template void min_max_caller<unsigned short>(const DevMem2D, double*, double*, PtrStep);\r
     template void min_max_caller<signed short>(const DevMem2D, double*, double*, PtrStep);\r
     template void min_max_caller<int>(const DevMem2D, double*, double*, PtrStep);\r
-    template void min_max_caller<float>(const DevMem2D, double*, double*, PtrStep);\r
+    template void min_max_caller<float>(const DevMem2D, double*,double*, PtrStep);\r
     template void min_max_caller<double>(const DevMem2D, double*, double*, PtrStep);\r
 \r
 \r
-    // This kernel will be used only when compute capability is 1.0\r
     template <int nthreads, typename T>\r
-    __global__ void min_max_kernel_2ndstep(T* minval, T* maxval, int size)\r
+    __global__ void min_max_pass2_kernel(T* minval, T* maxval, int size)\r
     {\r
         typedef typename MinMaxTypeTraits<T>::best_type best_type;\r
         __shared__ best_type sminval[nthreads];\r
@@ -615,7 +647,36 @@ namespace cv { namespace gpu { namespace mathfunc
 \r
 \r
     template <typename T>\r
-    void min_max_caller_2steps(const DevMem2D src, double* minval, double* maxval, PtrStep buf)\r
+    void min_max_mask_multipass_caller(const DevMem2D src, const PtrStep mask, double* minval, double* maxval, PtrStep buf)\r
+    {\r
+        dim3 threads, grid;\r
+        estimate_thread_cfg(threads, grid);\r
+        estimate_kernel_consts(src.cols, src.rows, threads, grid);\r
+\r
+        T* minval_buf = (T*)buf.ptr(0);\r
+        T* maxval_buf = (T*)buf.ptr(1);\r
+\r
+        min_max_kernel<256, T, Mask8U><<<grid, threads>>>(src, Mask8U(mask), minval_buf, maxval_buf);\r
+        min_max_pass2_kernel<256, T><<<1, 256>>>(minval_buf, maxval_buf, grid.x * grid.y);\r
+        cudaSafeCall(cudaThreadSynchronize());\r
+\r
+        T minval_, maxval_;\r
+        cudaSafeCall(cudaMemcpy(&minval_, minval_buf, sizeof(T), cudaMemcpyDeviceToHost));\r
+        cudaSafeCall(cudaMemcpy(&maxval_, maxval_buf, sizeof(T), cudaMemcpyDeviceToHost));\r
+        *minval = minval_;\r
+        *maxval = maxval_;\r
+    }\r
+\r
+    template void min_max_mask_multipass_caller<unsigned char>(const DevMem2D, const PtrStep, double*, double*, PtrStep);\r
+    template void min_max_mask_multipass_caller<signed char>(const DevMem2D, const PtrStep, double*, double*, PtrStep);\r
+    template void min_max_mask_multipass_caller<unsigned short>(const DevMem2D, const PtrStep, double*, double*, PtrStep);\r
+    template void min_max_mask_multipass_caller<signed short>(const DevMem2D, const PtrStep, double*, double*, PtrStep);\r
+    template void min_max_mask_multipass_caller<int>(const DevMem2D, const PtrStep, double*, double*, PtrStep);\r
+    template void min_max_mask_multipass_caller<float>(const DevMem2D, const PtrStep, double*, double*, PtrStep);\r
+\r
+\r
+    template <typename T>\r
+    void min_max_multipass_caller(const DevMem2D src, double* minval, double* maxval, PtrStep buf)\r
     {\r
         dim3 threads, grid;\r
         estimate_thread_cfg(threads, grid);\r
@@ -624,8 +685,8 @@ namespace cv { namespace gpu { namespace mathfunc
         T* minval_buf = (T*)buf.ptr(0);\r
         T* maxval_buf = (T*)buf.ptr(1);\r
 \r
-        min_max_kernel<256, T><<<grid, threads>>>(src, minval_buf, maxval_buf);\r
-        min_max_kernel_2ndstep<256, T><<<1, 256>>>(minval_buf, maxval_buf, grid.x * grid.y);\r
+        min_max_kernel<256, T, MaskTrue><<<grid, threads>>>(src, MaskTrue(), minval_buf, maxval_buf);\r
+        min_max_pass2_kernel<256, T><<<1, 256>>>(minval_buf, maxval_buf, grid.x * grid.y);\r
         cudaSafeCall(cudaThreadSynchronize());\r
 \r
         T minval_, maxval_;\r
@@ -635,12 +696,12 @@ namespace cv { namespace gpu { namespace mathfunc
         *maxval = maxval_;\r
     }\r
 \r
-    template void min_max_caller_2steps<unsigned char>(const DevMem2D, double*, double*, PtrStep);\r
-    template void min_max_caller_2steps<signed char>(const DevMem2D, double*, double*, PtrStep);\r
-    template void min_max_caller_2steps<unsigned short>(const DevMem2D, double*, double*, PtrStep);\r
-    template void min_max_caller_2steps<signed short>(const DevMem2D, double*, double*, PtrStep);\r
-    template void min_max_caller_2steps<int>(const DevMem2D, double*, double*, PtrStep);\r
-    template void min_max_caller_2steps<float>(const DevMem2D, double*, double*, PtrStep);\r
+    template void min_max_multipass_caller<unsigned char>(const DevMem2D, double*, double*, PtrStep);\r
+    template void min_max_multipass_caller<signed char>(const DevMem2D, double*, double*, PtrStep);\r
+    template void min_max_multipass_caller<unsigned short>(const DevMem2D, double*, double*, PtrStep);\r
+    template void min_max_multipass_caller<signed short>(const DevMem2D, double*, double*, PtrStep);\r
+    template void min_max_multipass_caller<int>(const DevMem2D, double*, double*, PtrStep);\r
+    template void min_max_multipass_caller<float>(const DevMem2D, double*, double*, PtrStep);\r
 \r
     } // namespace minmax\r
 \r
@@ -861,7 +922,7 @@ namespace cv { namespace gpu { namespace mathfunc
 \r
     // This kernel will be used only when compute capability is 1.0\r
     template <int nthreads, typename T>\r
-    __global__ void min_max_loc_kernel_2ndstep(T* minval, T* maxval, unsigned int* minloc, unsigned int* maxloc, int size)\r
+    __global__ void min_max_loc_pass2_kernel(T* minval, T* maxval, unsigned int* minloc, unsigned int* maxloc, int size)\r
     {\r
         typedef typename MinMaxTypeTraits<T>::best_type best_type;\r
         __shared__ best_type sminval[nthreads];\r
@@ -892,7 +953,7 @@ namespace cv { namespace gpu { namespace mathfunc
 \r
 \r
     template <typename T>\r
-    void min_max_loc_caller_2steps(const DevMem2D src, double* minval, double* maxval, \r
+    void min_max_loc_multipass_caller(const DevMem2D src, double* minval, double* maxval, \r
                                    int minloc[2], int maxloc[2], PtrStep valbuf, PtrStep locbuf)\r
     {\r
         dim3 threads, grid;\r
@@ -905,7 +966,7 @@ namespace cv { namespace gpu { namespace mathfunc
         unsigned int* maxloc_buf = (unsigned int*)locbuf.ptr(1);\r
 \r
         min_max_loc_kernel<256, T><<<grid, threads>>>(src, minval_buf, maxval_buf, minloc_buf, maxloc_buf);\r
-        min_max_loc_kernel_2ndstep<256, T><<<1, 256>>>(minval_buf, maxval_buf, minloc_buf, maxloc_buf, grid.x * grid.y);\r
+        min_max_loc_pass2_kernel<256, T><<<1, 256>>>(minval_buf, maxval_buf, minloc_buf, maxloc_buf, grid.x * grid.y);\r
         cudaSafeCall(cudaThreadSynchronize());\r
 \r
         T minval_, maxval_;\r
@@ -921,12 +982,12 @@ namespace cv { namespace gpu { namespace mathfunc
         maxloc[1] = maxloc_ / src.cols; maxloc[0] = maxloc_ - maxloc[1] * src.cols;\r
     }\r
 \r
-    template void min_max_loc_caller_2steps<unsigned char>(const DevMem2D, double*, double*, int[2], int[2], PtrStep, PtrStep);\r
-    template void min_max_loc_caller_2steps<signed char>(const DevMem2D, double*, double*, int[2], int[2], PtrStep, PtrStep);\r
-    template void min_max_loc_caller_2steps<unsigned short>(const DevMem2D, double*, double*, int[2], int[2], PtrStep, PtrStep);\r
-    template void min_max_loc_caller_2steps<signed short>(const DevMem2D, double*, double*, int[2], int[2], PtrStep, PtrStep);\r
-    template void min_max_loc_caller_2steps<int>(const DevMem2D, double*, double*, int[2], int[2], PtrStep, PtrStep);\r
-    template void min_max_loc_caller_2steps<float>(const DevMem2D, double*, double*, int[2], int[2], PtrStep, PtrStep);\r
+    template void min_max_loc_multipass_caller<unsigned char>(const DevMem2D, double*, double*, int[2], int[2], PtrStep, PtrStep);\r
+    template void min_max_loc_multipass_caller<signed char>(const DevMem2D, double*, double*, int[2], int[2], PtrStep, PtrStep);\r
+    template void min_max_loc_multipass_caller<unsigned short>(const DevMem2D, double*, double*, int[2], int[2], PtrStep, PtrStep);\r
+    template void min_max_loc_multipass_caller<signed short>(const DevMem2D, double*, double*, int[2], int[2], PtrStep, PtrStep);\r
+    template void min_max_loc_multipass_caller<int>(const DevMem2D, double*, double*, int[2], int[2], PtrStep, PtrStep);\r
+    template void min_max_loc_multipass_caller<float>(const DevMem2D, double*, double*, int[2], int[2], PtrStep, PtrStep);\r
 \r
     } // namespace minmaxloc\r
 \r
@@ -1070,7 +1131,7 @@ namespace cv { namespace gpu { namespace mathfunc
 \r
 \r
     template <int nthreads, typename T>\r
-    __global__ void count_non_zero_kernel_2ndstep(unsigned int* count, int size)\r
+    __global__ void count_non_zero_pass2_kernel(unsigned int* count, int size)\r
     {\r
         __shared__ unsigned int scount[nthreads];\r
         unsigned int tid = threadIdx.y * blockDim.x + threadIdx.x;\r
@@ -1087,7 +1148,7 @@ namespace cv { namespace gpu { namespace mathfunc
 \r
 \r
     template <typename T>\r
-    int count_non_zero_caller_2steps(const DevMem2D src, PtrStep buf)\r
+    int count_non_zero_multipass_caller(const DevMem2D src, PtrStep buf)\r
     {\r
         dim3 threads, grid;\r
         estimate_thread_cfg(threads, grid);\r
@@ -1096,7 +1157,7 @@ namespace cv { namespace gpu { namespace mathfunc
         unsigned int* count_buf = (unsigned int*)buf.ptr(0);\r
 \r
         count_non_zero_kernel<256, T><<<grid, threads>>>(src, count_buf);\r
-        count_non_zero_kernel_2ndstep<256, T><<<1, 256>>>(count_buf, grid.x * grid.y);\r
+        count_non_zero_pass2_kernel<256, T><<<1, 256>>>(count_buf, grid.x * grid.y);\r
         cudaSafeCall(cudaThreadSynchronize());\r
 \r
         unsigned int count;\r
@@ -1105,12 +1166,12 @@ namespace cv { namespace gpu { namespace mathfunc
         return count;\r
     }  \r
 \r
-    template int count_non_zero_caller_2steps<unsigned char>(const DevMem2D, PtrStep);\r
-    template int count_non_zero_caller_2steps<signed char>(const DevMem2D, PtrStep);\r
-    template int count_non_zero_caller_2steps<unsigned short>(const DevMem2D, PtrStep);\r
-    template int count_non_zero_caller_2steps<signed short>(const DevMem2D, PtrStep);\r
-    template int count_non_zero_caller_2steps<int>(const DevMem2D, PtrStep);\r
-    template int count_non_zero_caller_2steps<float>(const DevMem2D, PtrStep);\r
+    template int count_non_zero_multipass_caller<unsigned char>(const DevMem2D, PtrStep);\r
+    template int count_non_zero_multipass_caller<signed char>(const DevMem2D, PtrStep);\r
+    template int count_non_zero_multipass_caller<unsigned short>(const DevMem2D, PtrStep);\r
+    template int count_non_zero_multipass_caller<signed short>(const DevMem2D, PtrStep);\r
+    template int count_non_zero_multipass_caller<int>(const DevMem2D, PtrStep);\r
+    template int count_non_zero_multipass_caller<float>(const DevMem2D, PtrStep);\r
 \r
     } // namespace countnonzero\r
 \r

diff --git a/tests/gpu/src/arithm.cpp b/tests/gpu/src/arithm.cpp
index b9f0b16..bf212f6 100644 (file)
--- a/tests/gpu/src/arithm.cpp
+++ b/tests/gpu/src/arithm.cpp
@@ -682,16 +682,16 @@ struct CV_GpuMinMaxTest: public CvTest
     {\r
         int depth_end;\r
         if (cv::gpu::hasNativeDoubleSupport(cv::gpu::getDevice())) depth_end = CV_64F; else depth_end = CV_32F;\r
-        for (int cn = 1; cn <= 4; ++cn)\r
-            for (int depth = CV_8U; depth <= depth_end; ++depth)\r
+        for (int depth = CV_8U; depth <= depth_end; ++depth)\r
+        {\r
+            for (int i = 0; i < 1; ++i)\r
             {\r
-                for (int i = 0; i < 1; ++i)\r
-                {\r
-                    int rows = 1 + rand() % 1000;\r
-                    int cols = 1 + rand() % 1000;\r
-                    test(rows, cols, cn, depth);\r
-                }\r
+                int rows = 1 + rand() % 1000;\r
+                int cols = 1 + rand() % 1000;\r
+                test(rows, cols, 1, depth);\r
+                test_masked(rows, cols, 1, depth);\r
             }\r
+        }\r
     }\r
 \r
     void test(int rows, int cols, int cn, int depth)\r
@@ -707,10 +707,59 @@ struct CV_GpuMinMaxTest: public CvTest
         double minVal, maxVal;\r
         cv::Point minLoc, maxLoc;\r
 \r
+        if (depth != CV_8S)\r
+        {\r
+            cv::minMaxLoc(src, &minVal, &maxVal, &minLoc, &maxLoc);\r
+        }\r
+        else \r
+        {\r
+            minVal = std::numeric_limits<double>::max();\r
+            maxVal = std::numeric_limits<double>::min();\r
+            for (int i = 0; i < src.rows; ++i)\r
+                for (int j = 0; j < src.cols; ++j)\r
+                {\r
+                    signed char val = src.at<signed char>(i, j);\r
+                    if (val < minVal) minVal = val;\r
+                    if (val > maxVal) maxVal = val;\r
+                }\r
+        }\r
+\r
+        double minVal_, maxVal_;\r
+        cv::Point minLoc_, maxLoc_;        \r
+        cv::gpu::minMax(cv::gpu::GpuMat(src), &minVal_, &maxVal_, cv::gpu::GpuMat(), buf);\r
+       \r
+        if (abs(minVal - minVal_) > 1e-3f)\r
+        {\r
+            ts->printf(CvTS::CONSOLE, "\nfail: minVal=%f minVal_=%f rows=%d cols=%d depth=%d cn=%d\n", minVal, minVal_, rows, cols, depth, cn);\r
+            ts->set_failed_test_info(CvTS::FAIL_INVALID_OUTPUT);\r
+        }\r
+        if (abs(maxVal - maxVal_) > 1e-3f)\r
+        {\r
+            ts->printf(CvTS::CONSOLE, "\nfail: maxVal=%f maxVal_=%f rows=%d cols=%d depth=%d cn=%d\n", maxVal, maxVal_, rows, cols, depth, cn);\r
+            ts->set_failed_test_info(CvTS::FAIL_INVALID_OUTPUT);\r
+        }\r
+    }  \r
+\r
+    void test_masked(int rows, int cols, int cn, int depth)\r
+    {\r
+        cv::Mat src(rows, cols, CV_MAKE_TYPE(depth, cn));\r
+        cv::RNG rng;\r
+        for (int i = 0; i < src.rows; ++i)\r
+        { \r
+            Mat row(1, src.cols * src.elemSize(), CV_8U, src.ptr(i));\r
+            rng.fill(row, RNG::UNIFORM, Scalar(0), Scalar(255));\r
+        }\r
+\r
+        cv::Mat mask(src.size(), CV_8U);\r
+        rng.fill(mask, RNG::UNIFORM, Scalar(0), Scalar(2));\r
+\r
+        double minVal, maxVal;\r
+        cv::Point minLoc, maxLoc;\r
+\r
         Mat src_ = src.reshape(1);\r
         if (depth != CV_8S)\r
         {\r
-            cv::minMaxLoc(src_, &minVal, &maxVal, &minLoc, &maxLoc);\r
+            cv::minMaxLoc(src_, &minVal, &maxVal, &minLoc, &maxLoc, mask);\r
         }\r
         else \r
         {\r
@@ -721,14 +770,14 @@ struct CV_GpuMinMaxTest: public CvTest
                 for (int j = 0; j < src_.cols; ++j)\r
                 {\r
                     char val = src_.at<char>(i, j);\r
-                    if (val < minVal) minVal = val;\r
-                    if (val > maxVal) maxVal = val;\r
+                    if (mask.at<unsigned char>(i, j)) { if (val < minVal) minVal = val; }\r
+                    if (mask.at<unsigned char>(i, j)) { if (val > maxVal) maxVal = val; }\r
                 }\r
         }\r
 \r
         double minVal_, maxVal_;\r
         cv::Point minLoc_, maxLoc_;        \r
-        cv::gpu::minMax(cv::gpu::GpuMat(src), &minVal_, &maxVal_, buf);\r
+        cv::gpu::minMax(cv::gpu::GpuMat(src), &minVal_, &maxVal_, cv::gpu::GpuMat(mask), buf);\r
        \r
         if (abs(minVal - minVal_) > 1e-3f)\r
         {\r