improve MomentsInTile_SIMD<uchar, int, int>()

modules/imgproc/src/moments.cpp

@@ -227,16 +227,15 @@ struct MomentsInTile_SIMD<uchar, int, int>
 
         if( useSIMD )
         {
-            __m128i qx_init = _mm_setr_epi16(0, 1, 2, 3, 4, 5, 6, 7);
             __m128i dx = _mm_set1_epi16(8);
-            __m128i z = _mm_setzero_si128(), qx0 = z, qx1 = z, qx2 = z, qx3 = z, qx = qx_init;
+            __m128i z = _mm_setzero_si128(), qx0 = z, qx1 = z, qx2 = z, qx3 = z, qx = _mm_setr_epi16(0, 1, 2, 3, 4, 5, 6, 7);
 
             for( ; x <= len - 8; x += 8 )
             {
                 __m128i p = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i*)(ptr + x)), z);
                 __m128i sx = _mm_mullo_epi16(qx, qx);
 
-                qx0 = _mm_add_epi32(qx0, _mm_sad_epu8(p, z));
+                qx0 = _mm_add_epi16(qx0, p);
                 qx1 = _mm_add_epi32(qx1, _mm_madd_epi16(p, qx));
                 qx2 = _mm_add_epi32(qx2, _mm_madd_epi16(p, sx));
                 qx3 = _mm_add_epi32(qx3, _mm_madd_epi16( _mm_mullo_epi16(p, qx), sx));
@@ -244,14 +243,21 @@ struct MomentsInTile_SIMD<uchar, int, int>
                 qx = _mm_add_epi16(qx, dx);
             }
 
-            _mm_store_si128((__m128i*)buf, qx0);
-            x0 = buf[0] + buf[1] + buf[2] + buf[3];
-            _mm_store_si128((__m128i*)buf, qx1);
-            x1 = buf[0] + buf[1] + buf[2] + buf[3];
-            _mm_store_si128((__m128i*)buf, qx2);
-            x2 = buf[0] + buf[1] + buf[2] + buf[3];
-            _mm_store_si128((__m128i*)buf, qx3);
-            x3 = buf[0] + buf[1] + buf[2] + buf[3];
+            __m128i qx01_lo = _mm_unpacklo_epi32(qx0, qx1);
+            __m128i qx23_lo = _mm_unpacklo_epi32(qx2, qx3);
+            __m128i qx01_hi = _mm_unpackhi_epi32(qx0, qx1);
+            __m128i qx23_hi = _mm_unpackhi_epi32(qx2, qx3);
+            qx01_lo = _mm_add_epi32(qx01_lo, qx01_hi);
+            qx23_lo = _mm_add_epi32(qx23_lo, qx23_hi);
+            __m128i qx0123_lo = _mm_unpacklo_epi64(qx01_lo, qx23_lo);
+            __m128i qx0123_hi = _mm_unpackhi_epi64(qx01_lo, qx23_lo);
+            qx0123_lo = _mm_add_epi32(qx0123_lo, qx0123_hi);
+            _mm_store_si128((__m128i*)buf, qx0123_lo);
+
+            x0 = (buf[0] & 0xffff) + (buf[0] >> 16);
+            x1 = buf[1];
+            x2 = buf[2];
+            x3 = buf[3];
         }
 
         return x;