188       WebPUint32ToMem(&dst[0 * BPS], _mm_cvtsi128_si32(ref0));  in ITransform_SSE2()
189       WebPUint32ToMem(&dst[1 * BPS], _mm_cvtsi128_si32(ref1));  in ITransform_SSE2()
190       WebPUint32ToMem(&dst[2 * BPS], _mm_cvtsi128_si32(ref2));  in ITransform_SSE2()
191       WebPUint32ToMem(&dst[3 * BPS], _mm_cvtsi128_si32(ref3));  in ITransform_SSE2()
728     WebPUint32ToMem(dst + i * BPS, vals);  in VE4_SSE2()
739   WebPUint32ToMem(dst + 0 * BPS, 0x01010101U * AVG3(X, I, J));  in HE4_SSE2()
740   WebPUint32ToMem(dst + 1 * BPS, 0x01010101U * AVG3(I, J, K));  in HE4_SSE2()
741   WebPUint32ToMem(dst + 2 * BPS, 0x01010101U * AVG3(J, K, L));  in HE4_SSE2()
742   WebPUint32ToMem(dst + 3 * BPS, 0x01010101U * AVG3(K, L, L));  in HE4_SSE2()
763   WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(               abcdefg    ));  in LD4_SSE2()
764   WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1)));  in LD4_SSE2()
765   WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2)));  in LD4_SSE2()
766   WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3)));  in LD4_SSE2()
785   WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(               abcd    ));  in VR4_SSE2()
786   WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(               efgh    ));  in VR4_SSE2()
787   WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_slli_si128(abcd, 1)));  in VR4_SSE2()
788   WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_slli_si128(efgh, 1)));  in VR4_SSE2()
811   WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(               avg1    ));  in VL4_SSE2()
812   WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(               avg4    ));  in VL4_SSE2()
813   WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(avg1, 1)));  in VL4_SSE2()
814   WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(avg4, 1)));  in VL4_SSE2()
832   WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(               abcdefg    ));  in RD4_SSE2()
833   WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1)));  in RD4_SSE2()
834   WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2)));  in RD4_SSE2()
835   WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3)));  in RD4_SSE2()
885     WebPUint32ToMem(dst, _mm_cvtsi128_si32(out));  in TM4_SSE2()