fixed gpu::device::transform with non-aligned sources (Ticket #1567)

modules/gpu/src/opencv2/gpu/device/common.hpp

@@ -72,6 +72,11 @@ namespace cv { namespace gpu
     {
         return reinterpret_cast<size_t>(ptr) % size == 0;
     }
+
+    static inline bool isAligned(size_t step, size_t size)
+    {
+        return step % size == 0;
+    }
 }}
 
 static inline void ___cudaSafeCall(cudaError_t err, const char *file, const int line, const char *func = "")

modules/gpu/src/opencv2/gpu/device/detail/transform_detail.hpp

@@ -347,7 +347,8 @@ namespace cv { namespace gpu { namespace device
 
                 StaticAssert<ft::smart_shift != 1>::check();
 
-                if (!isAligned(src.data, ft::smart_shift * sizeof(T)) || !isAligned(dst.data, ft::smart_shift * sizeof(D)))
+                if (!isAligned(src.data, ft::smart_shift * sizeof(T)) || !isAligned(src.step, ft::smart_shift * sizeof(T)) || 
+                    !isAligned(dst.data, ft::smart_shift * sizeof(D)) || !isAligned(dst.step, ft::smart_shift * sizeof(D)))
                 {
                     TransformDispatcher<false>::call(src, dst, op, mask, stream);
                     return;
@@ -370,7 +371,9 @@ namespace cv { namespace gpu { namespace device
 
                 StaticAssert<ft::smart_shift != 1>::check();
 
-                if (!isAligned(src1.data, ft::smart_shift * sizeof(T1)) || !isAligned(src2.data, ft::smart_shift * sizeof(T2)) || !isAligned(dst.data, ft::smart_shift * sizeof(D)))
+                if (!isAligned(src1.data, ft::smart_shift * sizeof(T1)) || !isAligned(src1.step, ft::smart_shift * sizeof(T1)) ||
+                    !isAligned(src2.data, ft::smart_shift * sizeof(T2)) || !isAligned(src2.step, ft::smart_shift * sizeof(T2)) ||
+                    !isAligned(dst.data, ft::smart_shift * sizeof(D)) || !isAligned(dst.step, ft::smart_shift * sizeof(D)))
                 {
                     TransformDispatcher<false>::call(src1, src2, dst, op, mask, stream);
                     return;