Fixed cv::cuda::reduce bug.

modules/cudaarithm/perf/perf_reductions.cpp

@@ -368,6 +368,8 @@ PERF_TEST_P(Sz_Depth_Cn_Code_Dim, Reduce,
 
         TEST_CYCLE() cv::cuda::reduce(d_src, dst, dim, reduceOp, CV_32F);
 
+        dst = dst.reshape(dst.channels(), 1);
+
         CUDA_SANITY_CHECK(dst);
     }
     else

modules/cudaarithm/src/cuda/reduce.cu

@@ -137,7 +137,7 @@ void cv::cuda::reduce(InputArray _src, OutputArray _dst, int dim, int reduceOp,
     if (dtype < 0)
         dtype = src.depth();
 
-    GpuMat dst = getOutputMat(_dst, 1, dim == 0 ? src.cols : src.rows, CV_MAKE_TYPE(CV_MAT_DEPTH(dtype), src.channels()), stream);
+    GpuMat dst = getOutputMat(_dst, dim == 0 ? 1 : src.rows, dim == 0 ? src.cols : 1, CV_MAKE_TYPE(CV_MAT_DEPTH(dtype), src.channels()), stream);
 
     if (dim == 0)
     {

modules/cudaarithm/test/test_reductions.cpp

@@ -877,14 +877,11 @@ CUDA_TEST_P(Reduce, Cols)
 {
     cv::Mat src = randomMat(size, type);
 
-    cv::cuda::GpuMat dst = createMat(cv::Size(src.rows, 1), dst_type, useRoi);
+    cv::cuda::GpuMat dst;
     cv::cuda::reduce(loadMat(src, useRoi), dst, 1, reduceOp, dst_depth);
 
     cv::Mat dst_gold;
     cv::reduce(src, dst_gold, 1, reduceOp, dst_depth);
-    dst_gold.cols = dst_gold.rows;
-    dst_gold.rows = 1;
-    dst_gold.step = dst_gold.cols * dst_gold.elemSize();
 
     EXPECT_MAT_NEAR(dst_gold, dst, dst_depth < CV_32F ? 0.0 : 0.02);
 }

modules/cudev/include/opencv2/cudev/grid/reduce_to_vec.hpp

@@ -182,7 +182,7 @@ __host__ void gridReduceToColumn_(const SrcPtr& src, GpuMat_<ResType>& dst, cons
 
     CV_Assert( getRows(mask) == rows && getCols(mask) == cols );
 
-    dst.create(1, rows);
+    cuda::createContinuous(rows, 1, dst.type(), dst);
 
     grid_reduce_to_vec_detail::reduceToColumn<Reductor, Policy>(shrinkPtr(src),
                                                                 dst[0],
@@ -197,7 +197,7 @@ __host__ void gridReduceToColumn_(const SrcPtr& src, GpuMat_<ResType>& dst, Stre
     const int rows = getRows(src);
     const int cols = getCols(src);
 
-    dst.create(1, rows);
+    cuda::createContinuous(rows, 1, dst.type(), dst);
 
     grid_reduce_to_vec_detail::reduceToColumn<Reductor, Policy>(shrinkPtr(src),
                                                                 dst[0],

modules/cudev/test/test_reduction.cu

@@ -228,9 +228,6 @@ TEST(ReduceToColumn, Sum)
 
     Mat dst_gold;
     cv::reduce(src, dst_gold, 1, REDUCE_SUM, CV_32S);
-    dst_gold.cols = dst_gold.rows;
-    dst_gold.rows = 1;
-    dst_gold.step = dst_gold.cols * dst_gold.elemSize();
 
     EXPECT_MAT_NEAR(dst_gold, dst, 0.0);
 }
@@ -247,9 +244,6 @@ TEST(ReduceToColumn, Avg)
 
     Mat dst_gold;
     cv::reduce(src, dst_gold, 1, REDUCE_AVG, CV_32F);
-    dst_gold.cols = dst_gold.rows;
-    dst_gold.rows = 1;
-    dst_gold.step = dst_gold.cols * dst_gold.elemSize();
 
     EXPECT_MAT_NEAR(dst_gold, dst, 1e-4);
 }
@@ -266,9 +260,6 @@ TEST(ReduceToColumn, Min)
 
     Mat dst_gold;
     cv::reduce(src, dst_gold, 1, REDUCE_MIN);
-    dst_gold.cols = dst_gold.rows;
-    dst_gold.rows = 1;
-    dst_gold.step = dst_gold.cols * dst_gold.elemSize();
 
     EXPECT_MAT_NEAR(dst_gold, dst, 0.0);
 }
@@ -285,9 +276,6 @@ TEST(ReduceToColumn, Max)
 
     Mat dst_gold;
     cv::reduce(src, dst_gold, 1, REDUCE_MAX);
-    dst_gold.cols = dst_gold.rows;
-    dst_gold.rows = 1;
-    dst_gold.step = dst_gold.cols * dst_gold.elemSize();
 
     EXPECT_MAT_NEAR(dst_gold, dst, 0.0);
 }