simplified gpu::columnSum test, it doesn't fail on Quadro anymore (when seed is 000001af5a11badd) after BFM test, but something definitely wrong with NPP_Staging's transpose

modules/gpu/src/imgproc_gpu.cpp

@@ -1156,25 +1156,25 @@ void cv::gpu::dft(const GpuMat& src, GpuMat& dst, Size dft_size, int flags)
     if (src_data.data != src.data)
         src.copyTo(src_data);
 
-    Size dft_size_ = dft_size;
+    Size dft_size_opt = dft_size;
     if (is_1d_input && !is_row_dft)
     {
         // If the source matrix is single column handle it as single row
-        dft_size_.width = std::max(dft_size.width, dft_size.height);
+        dft_size_opt.width = std::max(dft_size.width, dft_size.height);
-        dft_size_.height = std::min(dft_size.width, dft_size.height);
+        dft_size_opt.height = std::min(dft_size.width, dft_size.height);
     }
 
     cufftType dft_type = CUFFT_R2C;
     if (is_complex_input) 
         dft_type = is_complex_output ? CUFFT_C2C : CUFFT_C2R;
 
-    CV_Assert(dft_size_.width > 1);
+    CV_Assert(dft_size_opt.width > 1);
 
     cufftHandle plan;
     if (is_1d_input || is_row_dft)
-        cufftPlan1d(&plan, dft_size_.width, dft_type, dft_size_.height);
+        cufftPlan1d(&plan, dft_size_opt.width, dft_type, dft_size_opt.height);
     else
-        cufftPlan2d(&plan, dft_size_.height, dft_size_.width, dft_type);
+        cufftPlan2d(&plan, dft_size_opt.height, dft_size_opt.width, dft_type);
 
     if (is_complex_input)
     {
@@ -1194,7 +1194,8 @@ void cv::gpu::dft(const GpuMat& src, GpuMat& dst, Size dft_size, int flags)
     }
     else
     {
-        if (dft_size == dft_size_)
+        // We could swap dft_size for efficiency. Here we must reflect it
+        if (dft_size == dft_size_opt)
             createContinuous(Size(dft_size.width / 2 + 1, dft_size.height), CV_32FC2, dst);
         else
             createContinuous(Size(dft_size.width, dft_size.height / 2 + 1), CV_32FC2, dst);
@@ -1206,7 +1207,7 @@ void cv::gpu::dft(const GpuMat& src, GpuMat& dst, Size dft_size, int flags)
     cufftSafeCall(cufftDestroy(plan));
 
     if (is_scaled_dft)
-        multiply(dst, Scalar::all(1. / (dft_size.area())), dst);
+        multiply(dst, Scalar::all(1. / dft_size.area()), dst);
 }
 
 //////////////////////////////////////////////////////////////////////////////
@@ -1260,18 +1261,19 @@ void cv::gpu::convolve(const GpuMat& image, const GpuMat& templ, GpuMat& result,
     block_size.width = std::min(dft_size.width - templ.cols + 1, result.cols);
     block_size.height = std::min(dft_size.height - templ.rows + 1, result.rows);
 
-    GpuMat result_data = createContinuous(dft_size, CV_32F);
-
     int spect_len = dft_size.height * (dft_size.width / 2 + 1);
-    GpuMat image_spect(1, spect_len, CV_32FC2);
+    GpuMat image_spect = createContinuous(1, spect_len, CV_32FC2);
-    GpuMat templ_spect(1, spect_len, CV_32FC2);
+    GpuMat templ_spect = createContinuous(1, spect_len, CV_32FC2);
-    GpuMat result_spect(1, spect_len, CV_32FC2);
+    GpuMat result_spect = createContinuous(1, spect_len, CV_32FC2);
 
     cufftHandle planR2C, planC2R;
     cufftSafeCall(cufftPlan2d(&planC2R, dft_size.height, dft_size.width, CUFFT_C2R));
     cufftSafeCall(cufftPlan2d(&planR2C, dft_size.height, dft_size.width, CUFFT_R2C));
 
+    GpuMat image_block = createContinuous(dft_size, CV_32F);
     GpuMat templ_block = createContinuous(dft_size, CV_32F);
+    GpuMat result_data = createContinuous(dft_size, CV_32F);
+
     GpuMat templ_roi(templ.size(), CV_32F, templ.data, templ.step);
     copyMakeBorder(templ_roi, templ_block, 0, templ_block.rows - templ_roi.rows, 0, 
                    templ_block.cols - templ_roi.cols, 0);
@@ -1279,8 +1281,6 @@ void cv::gpu::convolve(const GpuMat& image, const GpuMat& templ, GpuMat& result,
     cufftSafeCall(cufftExecR2C(planR2C, templ_block.ptr<cufftReal>(), 
                                templ_spect.ptr<cufftComplex>()));
 
-    GpuMat image_block = createContinuous(dft_size, CV_32F);
-
     // Process all blocks of the result matrix
     for (int y = 0; y < result.rows; y += block_size.height)
     {

tests/gpu/src/gputest_main.cpp

@@ -47,7 +47,6 @@ const char* blacklist[] =
 {
     "GPU-AsyncGpuMatOperator",     // crash
     "GPU-NppImageCanny",            // NPP_TEXTURE_BIND_ERROR
-    "GPU-BruteForceMatcher", // often crashes when seed=000001af5a11badd
     0
 };
 

tests/gpu/src/imgproc_gpu.cpp

@@ -768,34 +768,37 @@ struct CV_GpuColumnSumTest: CvTest
     {
         try
         {
-            int n = 375;
+            int cols = 375;
-            int m = 1072;
+            int rows = 1072;
-            Mat src(n, m, CV_32F);
+
+            Mat src(rows, cols, CV_32F);
             RNG rng(1);
             rng.fill(src, RNG::UNIFORM, Scalar(0), Scalar(1));
-            Mat dst_gold, dst2_gold;
 
-            integral(src, dst_gold, dst2_gold);
+            GpuMat d_dst;
+            columnSum(GpuMat(src), d_dst);
             
-            GpuMat dsrc(src);
+            Mat dst = d_dst;
-            GpuMat buf;
+            for (int j = 0; j < src.cols; ++j)
-            GpuMat dst;
-            columnSum(dsrc, buf);
-            transpose(buf, dst);
-            columnSum(dst, buf);
-            transpose(buf, dst);
-            
-            Mat dst_ = dst;
-            for (int i = 0; i < dst_.rows; ++i)
             {
-                const double* dst_gold_data = (const double*)dst_gold.ptr(i + 1);
+                float a = src.at<float>(0, j);
-                for (int j = 0; j < dst_.cols; ++j)
+                float b = dst.at<float>(0, j);
-                {
-                    float a = (float)dst_gold_data[j + 1];
-                    float b = dst_.at<float>(i, j);
                 if (fabs(a - b) > 0.5f)
                 {
-                        ts->printf(CvTS::CONSOLE, "%d %d %f %f\n", i, j, a, b);
+                    ts->printf(CvTS::CONSOLE, "big diff at %d %d: %f %f\n", 0, j, a, b);
+                    ts->set_failed_test_info(CvTS::FAIL_INVALID_OUTPUT);
+                    return;
+                }
+            }
+            for (int i = 1; i < src.rows; ++i)
+            {
+                for (int j = 0; j < src.cols; ++j)
+                {
+                    float a = src.at<float>(i, j) += src.at<float>(i - 1, j);
+                    float b = dst.at<float>(i, j);
+                    if (fabs(a - b) > 0.5f)
+                    {
+                        ts->printf(CvTS::CONSOLE, "big diff at %d %d: %f %f\n", i, j, a, b);
                         ts->set_failed_test_info(CvTS::FAIL_INVALID_OUTPUT);
                         return;
                     }
@@ -804,6 +807,7 @@ struct CV_GpuColumnSumTest: CvTest
         }
         catch (const Exception& e)
         {
+            ts->printf(CvTS::CONSOLE, e.what());
             if (!check_and_treat_gpu_exception(e, ts)) throw;
             return;
         }