Merge pull request opencv#25364 from mshabunin:fix-unaligned-filter

imgproc: fix unaligned memory access in filters and Gaussian blur opencv#25364

* filter/SIMD: removed parts which casted 8u pointers to int causing unaligned memory access on RISC-V platform.
* GaussianBlur/fixed_point: replaced casts from s16 to u32 with union operations

Performance comparison:
- [x] check performance on x86_64 - (4 threads, `-DCPU_BASELINE=AVX2`, GCC 11.4, Ubuntu 22) - [report_imgproc_x86_64.ods](https://github.com/opencv/opencv/files/14904702/report_x86_64.ods)
- [x] check performance on AArch64 - (4 cores of RK3588, GCC 11.4 aarch64, Raspbian) - [report_imgproc_aarch64.ods](https://github.com/opencv/opencv/files/14908437/report_aarch64.ods)

Note: for some reason my performance results are quite unstable, unaffected functions show speedups and slowdowns in many cases. Filter2D and GaussianBlur seem to be OK.

Slightly related PR: opencv/ci-gha-workflow#165

Author: mshabunin

modules/imgproc/src/filter.simd.hpp

@@ -86,7 +86,6 @@ Ptr<BaseFilter> getLinearFilter(
 
 #ifndef CV_CPU_OPTIMIZATION_DECLARATIONS_ONLY
 
-typedef int CV_DECL_ALIGNED(1) unaligned_int;
 #define VEC_ALIGN CV_MALLOC_ALIGN
 
 int FilterEngine__start(FilterEngine& this_, const Size &_wholeSize, const Size &sz, const Point &ofs)
@@ -1083,21 +1082,6 @@ struct SymmColumnVec_32s8u
                 v_pack_u_store(dst + i, v_pack(v_round(s0), v_round(s1)));
                 i += VTraits<v_uint16>::vlanes();
             }
-#if CV_SIMD_WIDTH > 16
-            while( i <= width - 4 /*VTraits<v_int32x4>::vlanes()*/ )
-#else
-            if( i <= width - VTraits<v_int32>::vlanes() )
-#endif
-            {
-                v_float32 s0 = v_muladd(v_cvt_f32(vx_load(src[0] + i)), vx_setall_f32(ky[0]), vx_setall_f32(delta));
-                s0 = v_muladd(v_cvt_f32(v_add(vx_load(src[1] + i), vx_load(src[-1] + i))), vx_setall_f32(ky[1]), s0);
-                for( k = 2; k <= ksize2; k++ )
-                    s0 = v_muladd(v_cvt_f32(v_add(vx_load(src[k] + i), vx_load(src[-k] + i))), vx_setall_f32(ky[k]), s0);
-                v_int32 s32 = v_round(s0);
-                v_int16 s16 = v_pack(s32, s32);
-                *(unaligned_int*)(dst + i) = v_get0(v_reinterpret_as_s32(v_pack_u(s16, s16)));
-                i += 4 /*v_int32x4::nlanes*/ ;
-            }
         }
         else
         {
@@ -1139,20 +1123,6 @@ struct SymmColumnVec_32s8u
                 v_pack_u_store(dst + i, v_pack(v_round(s0), v_round(s1)));
                 i += VTraits<v_uint16>::vlanes();
             }
-#if CV_SIMD_WIDTH > 16
-            while( i <= width - 4 /*VTraits<v_int32x4>::vlanes()*/ )
-#else
-            if( i <= width - VTraits<v_int32>::vlanes() )
-#endif
-            {
-                v_float32 s0 = v_muladd(v_cvt_f32(v_sub(vx_load(src[1] + i), vx_load(src[-1] + i))), vx_setall_f32(ky[1]), vx_setall_f32(delta));
-                for (k = 2; k <= ksize2; k++)
-                    s0 = v_muladd(v_cvt_f32(v_sub(vx_load(src[k] + i), vx_load(src[-k] + i))), vx_setall_f32(ky[k]), s0);
-                v_int32 s32 = v_round(s0);
-                v_int16 s16 = v_pack(s32, s32);
-                *(unaligned_int*)(dst + i) = v_get0(v_reinterpret_as_s32(v_pack_u(s16, s16)));
-                i += 4 /*v_int32x4::nlanes*/ ;
-            }
         }
         return i;
     }
@@ -2236,20 +2206,6 @@ struct FilterVec_8u
             v_pack_u_store(dst + i, v_pack(v_round(s0), v_round(s1)));
             i += VTraits<v_uint16>::vlanes();
         }
-#if CV_SIMD_WIDTH > 16
-        while( i <= width - 4 /*VTraits<v_int32x4>::vlanes()*/ )
-#else
-        if( i <= width - VTraits<v_int32>::vlanes() )
-#endif
-        {
-            v_float32 s0 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(vx_load_expand_q(src[0] + i))), vx_setall_f32(kf[0]), vx_setall_f32(delta));
-            for( k = 1; k < nz; k++ )
-                s0 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(vx_load_expand_q(src[k] + i))), vx_setall_f32(kf[k]), s0);
-            v_int32 s32 = v_round(s0);
-            v_int16 s16 = v_pack(s32, s32);
-            *(unaligned_int*)(dst + i) = v_get0(v_reinterpret_as_s32(v_pack_u(s16, s16)));
-            i += 4 /*VTraits<v_int32x4>::vlanes()*/ ;
-        }
         return i;
     }
 

modules/imgproc/src/fixedpoint.inl.hpp

@@ -370,7 +370,7 @@ class ufixedpoint16
     static CV_ALWAYS_INLINE ufixedpoint16 one() { return ufixedpoint16((uint16_t)(1 << fixedShift)); }
 
     static CV_ALWAYS_INLINE ufixedpoint16 fromRaw(uint16_t v) { return ufixedpoint16(v); }
-    CV_ALWAYS_INLINE uint16_t raw() { return val; }
+    CV_ALWAYS_INLINE uint16_t raw() const { return val; }
 };
 
 }

modules/imgproc/src/smooth.simd.hpp

@@ -1634,6 +1634,15 @@ void vlineSmooth(const FT* const * src, const FT* m, int n, ET* dst, int len)
         dst[i] = val;
     }
 }
+
+inline uint32_t read_pair_as_u32(const ufixedpoint16 * mem)
+{
+    union Cv32sufX2 { uint32_t v32; int16_t v16[2]; } res;
+    res.v16[0] = mem->raw();
+    res.v16[1] = (mem + 1)->raw();
+    return res.v32;
+}
+
 template <>
 void vlineSmooth<uint8_t, ufixedpoint16>(const ufixedpoint16* const * src, const ufixedpoint16* m, int n, uint8_t* dst, int len)
 {
@@ -1655,7 +1664,7 @@ void vlineSmooth<uint8_t, ufixedpoint16>(const ufixedpoint16* const * src, const
         v_int16 v_src00, v_src10, v_src01, v_src11, v_src02, v_src12, v_src03, v_src13;
         v_int16 v_tmp0, v_tmp1;
 
-        v_int16 v_mul = v_reinterpret_as_s16(vx_setall_u32(*((uint32_t*)m)));
+        v_int16 v_mul = v_reinterpret_as_s16(vx_setall_u32(read_pair_as_u32(m)));
 
         const int16_t* src0 = (const int16_t*)src[0] + i;
         const int16_t* src1 = (const int16_t*)src[1] + i;
@@ -1683,7 +1692,7 @@ void vlineSmooth<uint8_t, ufixedpoint16>(const ufixedpoint16* const * src, const
         int j = 2;
         for (; j < n - 1; j+=2)
         {
-            v_mul = v_reinterpret_as_s16(vx_setall_u32(*((uint32_t*)(m+j))));
+            v_mul = v_reinterpret_as_s16(vx_setall_u32(read_pair_as_u32(m + j)));
 
             const int16_t* srcj0 = (const int16_t*)src[j] + i;
             const int16_t* srcj1 = (const int16_t*)src[j + 1] + i;