86     __m128i o0, o1;  in rsdIntrinsicConvolve3x3_K()  local
114         o1 = _mm_madd_epi16(_mm_unpacklo_epi16(p1, p2), c0);  in rsdIntrinsicConvolve3x3_K()
117         o1 = _mm_add_epi32(o1, _mm_madd_epi16(_mm_unpacklo_epi16(p3, p5), c2));  in rsdIntrinsicConvolve3x3_K()
120         o1 = _mm_add_epi32(o1, _mm_madd_epi16(_mm_unpacklo_epi16(p6, p7), c4));  in rsdIntrinsicConvolve3x3_K()
123         o1 = _mm_add_epi32(o1, _mm_madd_epi16(_mm_unpacklo_epi16(p9, p10), c6));  in rsdIntrinsicConvolve3x3_K()
126         o1 = _mm_add_epi32(o1, _mm_madd_epi16(_mm_unpacklo_epi16(p11, _mm_setzero_si128()), c8));  in rsdIntrinsicConvolve3x3_K()
129         o1 = _mm_srai_epi32(o1, 8);  in rsdIntrinsicConvolve3x3_K()
131         o0 = packus_epi32(o0, o1);  in rsdIntrinsicConvolve3x3_K()
609     __m128i o0, o1, o2, o3;  in rsdIntrinsicConvolve5x5_K()  local
701         o1 =                   _mm_madd_epi16( _mm_unpacklo_epi16(p1, p2),  c0);  in rsdIntrinsicConvolve5x5_K()
702         o1 = _mm_add_epi32(o1, _mm_madd_epi16( _mm_unpacklo_epi16(p3,p4),  c2));  in rsdIntrinsicConvolve5x5_K()
703         o1 = _mm_add_epi32(o1, _mm_madd_epi16( _mm_unpacklo_epi16(p5, p9),  c4));  in rsdIntrinsicConvolve5x5_K()
704         o1 = _mm_add_epi32(o1, _mm_madd_epi16( _mm_unpacklo_epi16(p10,p11),  c6));  in rsdIntrinsicConvolve5x5_K()
705         o1 = _mm_add_epi32(o1, _mm_madd_epi16( _mm_unpacklo_epi16(p12,p13),  c8));  in rsdIntrinsicConvolve5x5_K()
706         o1 = _mm_add_epi32(o1, _mm_madd_epi16( _mm_unpacklo_epi16(p17,p18), c10));  in rsdIntrinsicConvolve5x5_K()
707         o1 = _mm_add_epi32(o1, _mm_madd_epi16( _mm_unpacklo_epi16(p19,p20), c12));  in rsdIntrinsicConvolve5x5_K()
708         o1 = _mm_add_epi32(o1, _mm_madd_epi16( _mm_unpacklo_epi16(p21,p25), c14));  in rsdIntrinsicConvolve5x5_K()
709         o1 = _mm_add_epi32(o1, _mm_madd_epi16( _mm_unpacklo_epi16(p26, p27), c16));  in rsdIntrinsicConvolve5x5_K()
710         o1 = _mm_add_epi32(o1, _mm_madd_epi16( _mm_unpacklo_epi16(p28, p29), c18));  in rsdIntrinsicConvolve5x5_K()
711         o1 = _mm_add_epi32(o1, _mm_madd_epi16( _mm_unpacklo_epi16(p33, p34), c20));  in rsdIntrinsicConvolve5x5_K()
712         o1 = _mm_add_epi32(o1, _mm_madd_epi16( _mm_unpacklo_epi16(p35, p36), c22));  in rsdIntrinsicConvolve5x5_K()
713         o1 = _mm_add_epi32(o1, _mm_madd_epi16( _mm_unpacklo_epi16(p37, _mm_setzero_si128()), c24));  in rsdIntrinsicConvolve5x5_K()
714         o1 = _mm_srai_epi32(o1, 8);  in rsdIntrinsicConvolve5x5_K()
746         o0 = packus_epi32(o0, o1);  in rsdIntrinsicConvolve5x5_K()