some improvements of existing sse3 optimization of bilateral filter in case of…

some improvements of existing sse3 optimization of bilateral filter in case of 8uc3. Now perf tests take 6120ms instead of previous 7250ms (1.18x speed-up)

modules/imgproc/src/smooth.cpp

@@ -1787,6 +1787,7 @@ public:
                     #if CV_SSE3
                     if( haveSSE3 )
                     {
+                        const __m128i izero = _mm_setzero_si128();
                         const __m128 _b0 = _mm_set1_ps(static_cast<float>(b0));
                         const __m128 _g0 = _mm_set1_ps(static_cast<float>(g0));
                         const __m128 _r0 = _mm_set1_ps(static_cast<float>(r0));
@@ -1794,14 +1795,17 @@ public:
 
                         for( ; k <= maxk - 4; k += 4 )
                         {
-                            const uchar* sptr_k  = sptr + j + space_ofs[k];
-                            const uchar* sptr_k1 = sptr + j + space_ofs[k+1];
-                            const uchar* sptr_k2 = sptr + j + space_ofs[k+2];
-                            const uchar* sptr_k3 = sptr + j + space_ofs[k+3];
-
-                            __m128 _b = _mm_set_ps(sptr_k3[0],sptr_k2[0],sptr_k1[0],sptr_k[0]);
-                            __m128 _g = _mm_set_ps(sptr_k3[1],sptr_k2[1],sptr_k1[1],sptr_k[1]);
-                            __m128 _r = _mm_set_ps(sptr_k3[2],sptr_k2[2],sptr_k1[2],sptr_k[2]);
+                            const int* const sptr_k0  = reinterpret_cast<const int*>(sptr + j + space_ofs[k]);
+                            const int* const sptr_k1  = reinterpret_cast<const int*>(sptr + j + space_ofs[k+1]);
+                            const int* const sptr_k2  = reinterpret_cast<const int*>(sptr + j + space_ofs[k+2]);
+                            const int* const sptr_k3  = reinterpret_cast<const int*>(sptr + j + space_ofs[k+3]);
+
+                            __m128 _b = _mm_cvtepi32_ps(_mm_unpacklo_epi16(_mm_unpacklo_epi8(_mm_cvtsi32_si128(sptr_k0[0]), izero), izero));
+                            __m128 _g = _mm_cvtepi32_ps(_mm_unpacklo_epi16(_mm_unpacklo_epi8(_mm_cvtsi32_si128(sptr_k1[0]), izero), izero));
+                            __m128 _r = _mm_cvtepi32_ps(_mm_unpacklo_epi16(_mm_unpacklo_epi8(_mm_cvtsi32_si128(sptr_k2[0]), izero), izero));
+                            __m128 _z = _mm_cvtepi32_ps(_mm_unpacklo_epi16(_mm_unpacklo_epi8(_mm_cvtsi32_si128(sptr_k3[0]), izero), izero));
+
+                            _MM_TRANSPOSE4_PS(_b, _g, _r, _z);
 
                             __m128 bt = _mm_andnot_ps(_signMask, _mm_sub_ps(_b,_b0));
                             __m128 gt = _mm_andnot_ps(_signMask, _mm_sub_ps(_g,_g0));