fix device pointer alignmen in sft-based integral integration

diff --git a/modules/gpu/src/imgproc.cpp b/modules/gpu/src/imgproc.cpp
index a4d1798..6dd0c71 100644 (file)
--- a/modules/gpu/src/imgproc.cpp
+++ b/modules/gpu/src/imgproc.cpp
@@ -548,12 +548,16 @@ void cv::gpu::integralBuffered(const GpuMat& src, GpuMat& sum, GpuMat& buffer, S
     cudaStream_t stream = StreamAccessor::getStream(s);\r
 \r
     DeviceInfo info;\r
+    cv::Size whole;\r
+    cv::Point offset;\r
 \r
-    if (info.supports(WARP_SHUFFLE_FUNCTIONS))\r
+    src.locateROI(whole, offset);\r
+\r
+    if (info.supports(WARP_SHUFFLE_FUNCTIONS) )\r
     {\r
         GpuMat srcAlligned;\r
 \r
-        if (src.cols % 16 == 0 && src.rows % 8 == 0)\r
+        if (src.cols % 16 == 0 && src.rows % 8 == 0 && offset.x % 16 == 0 && offset.y % 8 == 0)\r
             srcAlligned = src;\r
         else\r
         {\r
@@ -575,19 +579,18 @@ void cv::gpu::integralBuffered(const GpuMat& src, GpuMat& sum, GpuMat& buffer, S
             srcAlligned = buffer;\r
         }\r
 \r
-        sum.create(srcAlligned.rows + 1, srcAlligned.cols + 1, CV_32SC1);\r
+        sum.create(srcAlligned.rows + 1, srcAlligned.cols + 4, CV_32SC1);\r
 \r
         if (s)\r
             s.enqueueMemSet(sum, Scalar::all(0));\r
         else\r
             sum.setTo(Scalar::all(0));\r
 \r
-        GpuMat inner = sum(Rect(1, 1, srcAlligned.cols, srcAlligned.rows));\r
+        GpuMat inner = sum(Rect(4, 1, srcAlligned.cols, srcAlligned.rows));\r
 \r
         cv::gpu::device::imgproc::shfl_integral_gpu(srcAlligned, inner, stream);\r
 \r
-        if (srcAlligned.data != src.data)\r
-            sum = sum(Rect(0, 0, src.cols + 1, src.rows + 1));\r
+            sum = sum(Rect(3, 0, src.cols + 1, src.rows + 1));\r
     }\r
     else\r
     {\r