268       WebPUint32ToMem(&dst[0 * BPS], _mm_cvtsi128_si32(ref0));  in ITransform()
269       WebPUint32ToMem(&dst[1 * BPS], _mm_cvtsi128_si32(ref1));  in ITransform()
270       WebPUint32ToMem(&dst[2 * BPS], _mm_cvtsi128_si32(ref2));  in ITransform()
271       WebPUint32ToMem(&dst[3 * BPS], _mm_cvtsi128_si32(ref3));  in ITransform()
813     WebPUint32ToMem(dst + i * BPS, vals);  in VE4()
823   WebPUint32ToMem(dst + 0 * BPS, 0x01010101U * AVG3(X, I, J));  in HE4()
824   WebPUint32ToMem(dst + 1 * BPS, 0x01010101U * AVG3(I, J, K));  in HE4()
825   WebPUint32ToMem(dst + 2 * BPS, 0x01010101U * AVG3(J, K, L));  in HE4()
826   WebPUint32ToMem(dst + 3 * BPS, 0x01010101U * AVG3(K, L, L));  in HE4()
846   WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(               abcdefg    ));  in LD4()
847   WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1)));  in LD4()
848   WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2)));  in LD4()
849   WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3)));  in LD4()
868   WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(               abcd    ));  in VR4()
869   WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(               efgh    ));  in VR4()
870   WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_slli_si128(abcd, 1)));  in VR4()
871   WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_slli_si128(efgh, 1)));  in VR4()
894   WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(               avg1    ));  in VL4()
895   WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(               avg4    ));  in VL4()
896   WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(avg1, 1)));  in VL4()
897   WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(avg4, 1)));  in VL4()
914   WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(               abcdefg    ));  in RD4()
915   WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1)));  in RD4()
916   WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2)));  in RD4()
917   WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3)));  in RD4()
967     WebPUint32ToMem(dst, _mm_cvtsi128_si32(out));  in TM4()