204       dst0 = _mm_loadl_epi64((__m128i*)(dst + 0 * BPS));  in Transform()
205       dst1 = _mm_loadl_epi64((__m128i*)(dst + 1 * BPS));  in Transform()
206       dst2 = _mm_loadl_epi64((__m128i*)(dst + 2 * BPS));  in Transform()
207       dst3 = _mm_loadl_epi64((__m128i*)(dst + 3 * BPS));  in Transform()
210       dst0 = _mm_cvtsi32_si128(*(int*)(dst + 0 * BPS));  in Transform()
211       dst1 = _mm_cvtsi32_si128(*(int*)(dst + 1 * BPS));  in Transform()
212       dst2 = _mm_cvtsi32_si128(*(int*)(dst + 2 * BPS));  in Transform()
213       dst3 = _mm_cvtsi32_si128(*(int*)(dst + 3 * BPS));  in Transform()
233       _mm_storel_epi64((__m128i*)(dst + 0 * BPS), dst0);  in Transform()
234       _mm_storel_epi64((__m128i*)(dst + 1 * BPS), dst1);  in Transform()
235       _mm_storel_epi64((__m128i*)(dst + 2 * BPS), dst2);  in Transform()
236       _mm_storel_epi64((__m128i*)(dst + 3 * BPS), dst3);  in Transform()
239       *(int*)(dst + 0 * BPS) = _mm_cvtsi128_si32(dst0);  in Transform()
240       *(int*)(dst + 1 * BPS) = _mm_cvtsi128_si32(dst1);  in Transform()
241       *(int*)(dst + 2 * BPS) = _mm_cvtsi128_si32(dst2);  in Transform()
242       *(int*)(dst + 3 * BPS) = _mm_cvtsi128_si32(dst3);  in Transform()
265   __m128i dst0 = _mm_cvtsi32_si128(*(int*)(dst + 0 * BPS));  in TransformAC3()
266   __m128i dst1 = _mm_cvtsi32_si128(*(int*)(dst + 1 * BPS));  in TransformAC3()
267   __m128i dst2 = _mm_cvtsi32_si128(*(int*)(dst + 2 * BPS));  in TransformAC3()
268   __m128i dst3 = _mm_cvtsi32_si128(*(int*)(dst + 3 * BPS));  in TransformAC3()
285   *(int*)(dst + 0 * BPS) = _mm_cvtsi128_si32(dst0);  in TransformAC3()
286   *(int*)(dst + 1 * BPS) = _mm_cvtsi128_si32(dst1);  in TransformAC3()
287   *(int*)(dst + 2 * BPS) = _mm_cvtsi128_si32(dst2);  in TransformAC3()
288   *(int*)(dst + 3 * BPS) = _mm_cvtsi128_si32(dst3);  in TransformAC3()