455         __m128i y1, y2, y3, y4;  in rsdIntrinsicYuv_K()  local
464         y4 = _mm_shuffle_epi8(y3, T4x4);  in rsdIntrinsicYuv_K()
465         _mm_storeu_si128((__m128i *)dst, y4);  in rsdIntrinsicYuv_K()
517         __m128i y1, y2, y3, y4;  in rsdIntrinsicYuvR_K()  local
526         y4 = _mm_shuffle_epi8(y3, T4x4);  in rsdIntrinsicYuvR_K()
527         _mm_storeu_si128((__m128i *)dst, y4);  in rsdIntrinsicYuvR_K()
578         __m128i y1, y2, y3, y4;  in rsdIntrinsicYuv2_K()  local
587         y4 = _mm_shuffle_epi8(y3, T4x4);  in rsdIntrinsicYuv2_K()
588         _mm_storeu_si128((__m128i *)dst, y4);  in rsdIntrinsicYuv2_K()
598                                           const void *y3, const void *y4,  in rsdIntrinsicConvolve5x5_K()  argument
677         p32 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(*((int32_t *)y4)), _mm_setzero_si128());  in rsdIntrinsicConvolve5x5_K()
678         p33 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(*((int32_t *)y4+1)), _mm_setzero_si128());  in rsdIntrinsicConvolve5x5_K()
679         p34 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(*((int32_t *)y4+2)), _mm_setzero_si128());  in rsdIntrinsicConvolve5x5_K()
680         p35 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(*((int32_t *)y4+3)), _mm_setzero_si128());  in rsdIntrinsicConvolve5x5_K()
681         p36 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(*((int32_t *)y4+4)), _mm_setzero_si128());  in rsdIntrinsicConvolve5x5_K()
682         p37 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(*((int32_t *)y4+5)), _mm_setzero_si128());  in rsdIntrinsicConvolve5x5_K()
683         p38 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(*((int32_t *)y4+6)), _mm_setzero_si128());  in rsdIntrinsicConvolve5x5_K()
684         p39 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(*((int32_t *)y4+7)), _mm_setzero_si128());  in rsdIntrinsicConvolve5x5_K()
755         y4 = (const char *)y4 + 16;  in rsdIntrinsicConvolve5x5_K()