Merge pull request #16463 from alalek:core_strong_ptr_alignment

modules/core/include/opencv2/core/cvdef.h

@@ -340,6 +340,13 @@ enum CpuFeatures {
 
 #include "cv_cpu_dispatch.h"
 
+#if !defined(CV_STRONG_ALIGNMENT) && defined(__arm__) && !(defined(__aarch64__) || defined(_M_ARM64))
+// int*, int64* should be propertly aligned pointers on ARMv7
+#define CV_STRONG_ALIGNMENT 1
+#endif
+#if !defined(CV_STRONG_ALIGNMENT)
+#define CV_STRONG_ALIGNMENT 0
+#endif
 
 /* fundamental constants */
 #define CV_PI   3.1415926535897932384626433832795

modules/core/include/opencv2/core/hal/intrin_cpp.hpp

@@ -1458,16 +1458,23 @@ template<typename _Tp, int n> inline void v_zip( const v_reg<_Tp, n>& a0, const
 @return register object
 
 @note Returned type will be detected from passed pointer type, for example uchar ==> cv::v_uint8x16, int ==> cv::v_int32x4, etc.
+
+@note Alignment requirement:
+if CV_STRONG_ALIGNMENT=1 then passed pointer must be aligned (`sizeof(lane type)` should be enough).
+Do not cast pointer types without runtime check for pointer alignment (like `uchar*` => `int*`).
  */
 template<typename _Tp>
 inline v_reg<_Tp, V_TypeTraits<_Tp>::nlanes128> v_load(const _Tp* ptr)
 {
+#if CV_STRONG_ALIGNMENT
+    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
+#endif
     return v_reg<_Tp, V_TypeTraits<_Tp>::nlanes128>(ptr);
 }
 
 /** @brief Load register contents from memory (aligned)
 
-similar to cv::v_load, but source memory block should be aligned (to 16-byte boundary)
+similar to cv::v_load, but source memory block should be aligned (to 16-byte boundary in case of SIMD128, 32-byte - SIMD256, etc)
  */
 template<typename _Tp>
 inline v_reg<_Tp, V_TypeTraits<_Tp>::nlanes128> v_load_aligned(const _Tp* ptr)
@@ -1488,6 +1495,9 @@ v_int32x4 r = v_load_low(lo);
 template<typename _Tp>
 inline v_reg<_Tp, V_TypeTraits<_Tp>::nlanes128> v_load_low(const _Tp* ptr)
 {
+#if CV_STRONG_ALIGNMENT
+    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
+#endif
     v_reg<_Tp, V_TypeTraits<_Tp>::nlanes128> c;
     for( int i = 0; i < c.nlanes/2; i++ )
     {
@@ -1509,6 +1519,10 @@ v_int32x4 r = v_load_halves(lo, hi);
 template<typename _Tp>
 inline v_reg<_Tp, V_TypeTraits<_Tp>::nlanes128> v_load_halves(const _Tp* loptr, const _Tp* hiptr)
 {
+#if CV_STRONG_ALIGNMENT
+    CV_Assert(isAligned<sizeof(_Tp)>(loptr));
+    CV_Assert(isAligned<sizeof(_Tp)>(hiptr));
+#endif
     v_reg<_Tp, V_TypeTraits<_Tp>::nlanes128> c;
     for( int i = 0; i < c.nlanes/2; i++ )
     {
@@ -1531,6 +1545,9 @@ template<typename _Tp>
 inline v_reg<typename V_TypeTraits<_Tp>::w_type, V_TypeTraits<_Tp>::nlanes128 / 2>
 v_load_expand(const _Tp* ptr)
 {
+#if CV_STRONG_ALIGNMENT
+    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
+#endif
     typedef typename V_TypeTraits<_Tp>::w_type w_type;
     v_reg<w_type, V_TypeTraits<w_type>::nlanes128> c;
     for( int i = 0; i < c.nlanes; i++ )
@@ -1552,6 +1569,9 @@ template<typename _Tp>
 inline v_reg<typename V_TypeTraits<_Tp>::q_type, V_TypeTraits<_Tp>::nlanes128 / 4>
 v_load_expand_q(const _Tp* ptr)
 {
+#if CV_STRONG_ALIGNMENT
+    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
+#endif
     typedef typename V_TypeTraits<_Tp>::q_type q_type;
     v_reg<q_type, V_TypeTraits<q_type>::nlanes128> c;
     for( int i = 0; i < c.nlanes; i++ )
@@ -1572,6 +1592,9 @@ For all types except 64-bit. */
 template<typename _Tp, int n> inline void v_load_deinterleave(const _Tp* ptr, v_reg<_Tp, n>& a,
                                                             v_reg<_Tp, n>& b)
 {
+#if CV_STRONG_ALIGNMENT
+    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
+#endif
     int i, i2;
     for( i = i2 = 0; i < n; i++, i2 += 2 )
     {
@@ -1591,6 +1614,9 @@ For all types except 64-bit. */
 template<typename _Tp, int n> inline void v_load_deinterleave(const _Tp* ptr, v_reg<_Tp, n>& a,
                                                             v_reg<_Tp, n>& b, v_reg<_Tp, n>& c)
 {
+#if CV_STRONG_ALIGNMENT
+    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
+#endif
     int i, i3;
     for( i = i3 = 0; i < n; i++, i3 += 3 )
     {
@@ -1613,6 +1639,9 @@ inline void v_load_deinterleave(const _Tp* ptr, v_reg<_Tp, n>& a,
                                 v_reg<_Tp, n>& b, v_reg<_Tp, n>& c,
                                 v_reg<_Tp, n>& d)
 {
+#if CV_STRONG_ALIGNMENT
+    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
+#endif
     int i, i4;
     for( i = i4 = 0; i < n; i++, i4 += 4 )
     {
@@ -1636,6 +1665,9 @@ inline void v_store_interleave( _Tp* ptr, const v_reg<_Tp, n>& a,
                                const v_reg<_Tp, n>& b,
                                hal::StoreMode /*mode*/=hal::STORE_UNALIGNED)
 {
+#if CV_STRONG_ALIGNMENT
+    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
+#endif
     int i, i2;
     for( i = i2 = 0; i < n; i++, i2 += 2 )
     {
@@ -1657,6 +1689,9 @@ inline void v_store_interleave( _Tp* ptr, const v_reg<_Tp, n>& a,
                                 const v_reg<_Tp, n>& b, const v_reg<_Tp, n>& c,
                                 hal::StoreMode /*mode*/=hal::STORE_UNALIGNED)
 {
+#if CV_STRONG_ALIGNMENT
+    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
+#endif
     int i, i3;
     for( i = i3 = 0; i < n; i++, i3 += 3 )
     {
@@ -1679,6 +1714,9 @@ template<typename _Tp, int n> inline void v_store_interleave( _Tp* ptr, const v_
                                                             const v_reg<_Tp, n>& d,
                                                             hal::StoreMode /*mode*/=hal::STORE_UNALIGNED)
 {
+#if CV_STRONG_ALIGNMENT
+    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
+#endif
     int i, i4;
     for( i = i4 = 0; i < n; i++, i4 += 4 )
     {
@@ -1700,6 +1738,9 @@ Pointer can be unaligned. */
 template<typename _Tp, int n>
 inline void v_store(_Tp* ptr, const v_reg<_Tp, n>& a)
 {
+#if CV_STRONG_ALIGNMENT
+    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
+#endif
     for( int i = 0; i < n; i++ )
         ptr[i] = a.s[i];
 }
@@ -1707,6 +1748,9 @@ inline void v_store(_Tp* ptr, const v_reg<_Tp, n>& a)
 template<typename _Tp, int n>
 inline void v_store(_Tp* ptr, const v_reg<_Tp, n>& a, hal::StoreMode /*mode*/)
 {
+#if CV_STRONG_ALIGNMENT
+    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
+#endif
     v_store(ptr, a);
 }
 
@@ -1720,6 +1764,9 @@ Scheme:
 template<typename _Tp, int n>
 inline void v_store_low(_Tp* ptr, const v_reg<_Tp, n>& a)
 {
+#if CV_STRONG_ALIGNMENT
+    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
+#endif
     for( int i = 0; i < (n/2); i++ )
         ptr[i] = a.s[i];
 }
@@ -1734,6 +1781,9 @@ Scheme:
 template<typename _Tp, int n>
 inline void v_store_high(_Tp* ptr, const v_reg<_Tp, n>& a)
 {
+#if CV_STRONG_ALIGNMENT
+    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
+#endif
     for( int i = 0; i < (n/2); i++ )
         ptr[i] = a.s[i+(n/2)];
 }

modules/core/src/copy.cpp

@@ -563,12 +563,6 @@ Mat& Mat::setTo(InputArray _value, InputArray _mask)
     return *this;
 }
 
-#if CV_NEON && !defined(__aarch64__)
-#define CV_CHECK_ALIGNMENT 1
-#else
-#define CV_CHECK_ALIGNMENT 0
-#endif
-
 #if CV_SIMD128
 template<typename V> CV_ALWAYS_INLINE void flipHoriz_single( const uchar* src, size_t sstep, uchar* dst, size_t dstep, Size size, size_t esz )
 {
@@ -578,7 +572,7 @@ template<typename V> CV_ALWAYS_INLINE void flipHoriz_single( const uchar* src, s
     int width_1 = width & -v_uint8x16::nlanes;
     int i, j;
 
-#if CV_CHECK_ALIGNMENT
+#if CV_STRONG_ALIGNMENT
     CV_Assert(isAligned<sizeof(T)>(src, dst));
 #endif
 
@@ -630,7 +624,7 @@ template<typename T1, typename T2> CV_ALWAYS_INLINE void flipHoriz_double( const
     int end = (int)(size.width*esz);
     int width = (end + 1)/2;
 
-#if CV_CHECK_ALIGNMENT
+#if CV_STRONG_ALIGNMENT
     CV_Assert(isAligned<sizeof(T1)>(src, dst));
     CV_Assert(isAligned<sizeof(T2)>(src, dst));
 #endif
@@ -659,7 +653,7 @@ static void
 flipHoriz( const uchar* src, size_t sstep, uchar* dst, size_t dstep, Size size, size_t esz )
 {
 #if CV_SIMD
-#if CV_CHECK_ALIGNMENT
+#if CV_STRONG_ALIGNMENT
     size_t alignmentMark = ((size_t)src)|((size_t)dst)|sstep|dstep;
 #endif
     if (esz == 2 * v_uint8x16::nlanes)
@@ -712,7 +706,7 @@ flipHoriz( const uchar* src, size_t sstep, uchar* dst, size_t dstep, Size size,
         }
     }
     else if (esz == 8
-#if CV_CHECK_ALIGNMENT
+#if CV_STRONG_ALIGNMENT
             && isAligned<sizeof(uint64)>(alignmentMark)
 #endif
     )
@@ -720,7 +714,7 @@ flipHoriz( const uchar* src, size_t sstep, uchar* dst, size_t dstep, Size size,
         flipHoriz_single<v_uint64x2>(src, sstep, dst, dstep, size, esz);
     }
     else if (esz == 4
-#if CV_CHECK_ALIGNMENT
+#if CV_STRONG_ALIGNMENT
             && isAligned<sizeof(unsigned)>(alignmentMark)
 #endif
     )
@@ -728,7 +722,7 @@ flipHoriz( const uchar* src, size_t sstep, uchar* dst, size_t dstep, Size size,
         flipHoriz_single<v_uint32x4>(src, sstep, dst, dstep, size, esz);
     }
     else if (esz == 2
-#if CV_CHECK_ALIGNMENT
+#if CV_STRONG_ALIGNMENT
             && isAligned<sizeof(ushort)>(alignmentMark)
 #endif
     )
@@ -740,7 +734,7 @@ flipHoriz( const uchar* src, size_t sstep, uchar* dst, size_t dstep, Size size,
         flipHoriz_single<v_uint8x16>(src, sstep, dst, dstep, size, esz);
     }
     else if (esz == 24
-#if CV_CHECK_ALIGNMENT
+#if CV_STRONG_ALIGNMENT
             && isAligned<sizeof(uint64_t)>(alignmentMark)
 #endif
     )
@@ -766,7 +760,7 @@ flipHoriz( const uchar* src, size_t sstep, uchar* dst, size_t dstep, Size size,
             }
         }
     }
-#if !CV_CHECK_ALIGNMENT
+#if !CV_STRONG_ALIGNMENT
     else if (esz == 12)
     {
         flipHoriz_double<uint64_t,uint>(src, sstep, dst, dstep, size, esz);
@@ -815,7 +809,7 @@ flipVert( const uchar* src0, size_t sstep, uchar* dst0, size_t dstep, Size size,
     {
         int i = 0;
 #if CV_SIMD
-#if CV_CHECK_ALIGNMENT
+#if CV_STRONG_ALIGNMENT
         if (isAligned<sizeof(int)>(src0, src1, dst0, dst1))
 #endif
         {
@@ -827,7 +821,7 @@ flipVert( const uchar* src0, size_t sstep, uchar* dst0, size_t dstep, Size size,
                 vx_store((int*)(dst1 + i), t0);
             }
         }
-#if CV_CHECK_ALIGNMENT
+#if CV_STRONG_ALIGNMENT
         else
         {
             for (; i <= size.width - CV_SIMD_WIDTH; i += CV_SIMD_WIDTH)

modules/dnn/src/onnx/onnx_importer.cpp

@@ -148,6 +148,7 @@ Mat getMatFromTensor(opencv_onnx::TensorProto& tensor_proto)
         else
         {
             const char* val = tensor_proto.raw_data().c_str();
+#if CV_STRONG_ALIGNMENT
             // Aligned pointer is required: https://github.com/opencv/opencv/issues/16373
             // this doesn't work: typedef int64_t CV_DECL_ALIGNED(1) unaligned_int64_t;
             AutoBuffer<int64_t, 16> aligned_val;
@@ -158,6 +159,7 @@ Mat getMatFromTensor(opencv_onnx::TensorProto& tensor_proto)
                 memcpy(aligned_val.data(), val, sz);
                 val = (const char*)aligned_val.data();
             }
+#endif
             const int64_t* src = reinterpret_cast<const int64_t*>(val);
             convertInt64ToInt32(src, dst, blob.total());
         }