80                                           const void *y1, const void *y2,  in rsdIntrinsicConvolve3x3_K()  argument
108         p8 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(*((int32_t *)y2)), _mm_setzero_si128());  in rsdIntrinsicConvolve3x3_K()
109         p9 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(*((int32_t *)y2+1)), _mm_setzero_si128());  in rsdIntrinsicConvolve3x3_K()
110         p10 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(*((int32_t *)y2+2)), _mm_setzero_si128());  in rsdIntrinsicConvolve3x3_K()
111         p11 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(*((int32_t *)y2+3)), _mm_setzero_si128());  in rsdIntrinsicConvolve3x3_K()
137         y2 = (const char *)y2 + 8;  in rsdIntrinsicConvolve3x3_K()
154     __m128i x2, y2, z2, w2;  in rsdIntrinsicColorMatrix4x4_K()  local
171         y2 = _mm_madd_epi16(xy, _mm_shuffle_epi32(c0, 0x55));  in rsdIntrinsicColorMatrix4x4_K()
176         y2 = _mm_add_epi32(y2, _mm_madd_epi16(zw, _mm_shuffle_epi32(c2, 0x55)));  in rsdIntrinsicColorMatrix4x4_K()
181         y2 = _mm_srai_epi32(y2, 8);  in rsdIntrinsicColorMatrix4x4_K()
185         x2 = packus_epi32(x2, y2);  in rsdIntrinsicColorMatrix4x4_K()
210     __m128i x2, y2, z2, w2;  in rsdIntrinsicColorMatrix3x3_K()  local
227         y2 =  _mm_madd_epi16(xy, _mm_shuffle_epi32(c0, 0x55));  in rsdIntrinsicColorMatrix3x3_K()
231         y2 = _mm_add_epi32(y2, _mm_madd_epi16(zw, _mm_shuffle_epi32(c2, 0x55)));  in rsdIntrinsicColorMatrix3x3_K()
235         y2 = _mm_srai_epi32(y2, 8);  in rsdIntrinsicColorMatrix3x3_K()
239         x2 = packus_epi32(x2, y2);  in rsdIntrinsicColorMatrix3x3_K()
262     __m128i x2, y2, z2, w2;  in rsdIntrinsicColorMatrixDot_K()  local
287         y2 = x2;  in rsdIntrinsicColorMatrixDot_K()
291         x2 = packus_epi32(x2, y2);  in rsdIntrinsicColorMatrixDot_K()
455         __m128i y1, y2, y3, y4;  in rsdIntrinsicYuv_K()  local
458         y2 = packus_epi32(B, A);  in rsdIntrinsicYuv_K()
459         y3 = _mm_packus_epi16(y1, y2);  in rsdIntrinsicYuv_K()
517         __m128i y1, y2, y3, y4;  in rsdIntrinsicYuvR_K()  local
520         y2 = packus_epi32(B, A);  in rsdIntrinsicYuvR_K()
521         y3 = _mm_packus_epi16(y1, y2);  in rsdIntrinsicYuvR_K()
578         __m128i y1, y2, y3, y4;  in rsdIntrinsicYuv2_K()  local
581         y2 = packus_epi32(B, A);  in rsdIntrinsicYuv2_K()
582         y3 = _mm_packus_epi16(y1, y2);  in rsdIntrinsicYuv2_K()
597                                           const void *y1, const void *y2,  in rsdIntrinsicConvolve5x5_K()  argument
659         p16 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(*((int32_t *)y2)), _mm_setzero_si128());  in rsdIntrinsicConvolve5x5_K()
660         p17 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(*((int32_t *)y2+1)), _mm_setzero_si128());  in rsdIntrinsicConvolve5x5_K()
661         p18 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(*((int32_t *)y2+2)), _mm_setzero_si128());  in rsdIntrinsicConvolve5x5_K()
662         p19 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(*((int32_t *)y2+3)), _mm_setzero_si128());  in rsdIntrinsicConvolve5x5_K()
663         p20 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(*((int32_t *)y2+4)), _mm_setzero_si128());  in rsdIntrinsicConvolve5x5_K()
664         p21 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(*((int32_t *)y2+5)), _mm_setzero_si128());  in rsdIntrinsicConvolve5x5_K()
665         p22 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(*((int32_t *)y2+6)), _mm_setzero_si128());  in rsdIntrinsicConvolve5x5_K()
666         p23 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(*((int32_t *)y2+7)), _mm_setzero_si128());  in rsdIntrinsicConvolve5x5_K()
753         y2 = (const char *)y2 + 16;  in rsdIntrinsicConvolve5x5_K()