213               SetrM128i(LoadUnaligned16(&src[x]), LoadUnaligned16(&src[x + 8]));  in FilterHorizontal()
216           const __m256i src_long2 = SetrM128i(LoadUnaligned16(&src[x + 16]),  in FilterHorizontal()
217                                               LoadUnaligned16(&src[x + 24]));  in FilterHorizontal()
252             SetrM128i(LoadUnaligned16(&src[0]), LoadUnaligned16(&src[8]));  in FilterHorizontal()
256             SetrM128i(LoadUnaligned16(&src[src_stride]),  in FilterHorizontal()
257                       LoadUnaligned16(&src[8 + src_stride]));  in FilterHorizontal()
269         const __m256i src_long = SetrM128i(LoadUnaligned16(&src[0]),  in FilterHorizontal()
270                                            LoadUnaligned16(&src[src_stride]));  in FilterHorizontal()
274             LoadUnaligned16(&src[8]), LoadUnaligned16(&src[8 + src_stride]));  in FilterHorizontal()
292           SetrM128i(LoadUnaligned16(&src[0]), LoadUnaligned16(&src[8]));  in FilterHorizontal()
303       const __m128i this_row = LoadUnaligned16(&src[0]);  in FilterHorizontal()
304       const __m128i next_row = LoadUnaligned16(&src[src_stride]);  in FilterHorizontal()
319         const __m128i this_row = LoadUnaligned16(&src[0]);  in FilterHorizontal()
320         const __m128i next_row = LoadUnaligned16(&src[src_stride]);  in FilterHorizontal()
337       const __m256i src_long = _mm256_castsi128_si256(LoadUnaligned16(&src[0]));  in FilterHorizontal()
348       const __m128i this_row = LoadUnaligned16(&src[0]);  in FilterHorizontal()
349       const __m128i next_row = LoadUnaligned16(&src[src_stride]);  in FilterHorizontal()
357         const __m128i this_row = LoadUnaligned16(&src[0]);  in FilterHorizontal()
358         const __m128i next_row = LoadUnaligned16(&src[src_stride]);  in FilterHorizontal()
375       const __m256i src_long = _mm256_castsi128_si256(LoadUnaligned16(&src[0]));  in FilterHorizontal()
871   srcs[0] = _mm256_castsi128_si256(LoadUnaligned16(src_x));  in FilterVertical16xH()
874     srcs[1] = _mm256_castsi128_si256(LoadUnaligned16(src_x));  in FilterVertical16xH()
878     srcs[2] = _mm256_castsi128_si256(LoadUnaligned16(src_x));  in FilterVertical16xH()
883       srcs[3] = _mm256_castsi128_si256(LoadUnaligned16(src_x));  in FilterVertical16xH()
887       srcs[4] = _mm256_castsi128_si256(LoadUnaligned16(src_x));  in FilterVertical16xH()
892         srcs[5] = _mm256_castsi128_si256(LoadUnaligned16(src_x));  in FilterVertical16xH()
896         srcs[6] = _mm256_castsi128_si256(LoadUnaligned16(src_x));  in FilterVertical16xH()
906     srcs[next_row - 1] = _mm256_castsi128_si256(LoadUnaligned16(src_x));  in FilterVertical16xH()
912     srcs[next_row] = _mm256_castsi128_si256(LoadUnaligned16(src_x));  in FilterVertical16xH()