86     __m128i o0, o1;  in rsdIntrinsicConvolve3x3_K()  local
113         o0 = _mm_madd_epi16(_mm_unpacklo_epi16(p0, p1), c0);  in rsdIntrinsicConvolve3x3_K()
116         o0 = _mm_add_epi32(o0, _mm_madd_epi16(_mm_unpacklo_epi16(p2, p4), c2));  in rsdIntrinsicConvolve3x3_K()
119         o0 = _mm_add_epi32(o0, _mm_madd_epi16(_mm_unpacklo_epi16(p5, p6), c4));  in rsdIntrinsicConvolve3x3_K()
122         o0 = _mm_add_epi32(o0, _mm_madd_epi16(_mm_unpacklo_epi16(p8, p9), c6));  in rsdIntrinsicConvolve3x3_K()
125         o0 = _mm_add_epi32(o0, _mm_madd_epi16(_mm_unpacklo_epi16(p10, _mm_setzero_si128()), c8));  in rsdIntrinsicConvolve3x3_K()
128         o0 = _mm_srai_epi32(o0, 8);  in rsdIntrinsicConvolve3x3_K()
131         o0 = packus_epi32(o0, o1);  in rsdIntrinsicConvolve3x3_K()
132         o0 = _mm_packus_epi16(o0, o0);  in rsdIntrinsicConvolve3x3_K()
133         _mm_storel_epi64((__m128i *)dst, o0);  in rsdIntrinsicConvolve3x3_K()
609     __m128i o0, o1, o2, o3;  in rsdIntrinsicConvolve5x5_K()  local
686         o0 =                   _mm_madd_epi16( _mm_unpacklo_epi16(p0, p1),  c0);  in rsdIntrinsicConvolve5x5_K()
687         o0 = _mm_add_epi32(o0, _mm_madd_epi16( _mm_unpacklo_epi16(p2, p3),  c2));  in rsdIntrinsicConvolve5x5_K()
688         o0 = _mm_add_epi32(o0, _mm_madd_epi16( _mm_unpacklo_epi16(p4, p8),  c4));  in rsdIntrinsicConvolve5x5_K()
689         o0 = _mm_add_epi32(o0, _mm_madd_epi16( _mm_unpacklo_epi16(p9,p10),  c6));  in rsdIntrinsicConvolve5x5_K()
690         o0 = _mm_add_epi32(o0, _mm_madd_epi16( _mm_unpacklo_epi16(p11, p12),  c8));  in rsdIntrinsicConvolve5x5_K()
691         o0 = _mm_add_epi32(o0, _mm_madd_epi16( _mm_unpacklo_epi16(p16, p17), c10));  in rsdIntrinsicConvolve5x5_K()
692         o0 = _mm_add_epi32(o0, _mm_madd_epi16( _mm_unpacklo_epi16(p18, p19), c12));  in rsdIntrinsicConvolve5x5_K()
693         o0 = _mm_add_epi32(o0, _mm_madd_epi16( _mm_unpacklo_epi16(p20, p24), c14));  in rsdIntrinsicConvolve5x5_K()
694         o0 = _mm_add_epi32(o0, _mm_madd_epi16( _mm_unpacklo_epi16(p25,p26), c16));  in rsdIntrinsicConvolve5x5_K()
695         o0 = _mm_add_epi32(o0, _mm_madd_epi16( _mm_unpacklo_epi16(p27, p28), c18));  in rsdIntrinsicConvolve5x5_K()
696         o0 = _mm_add_epi32(o0, _mm_madd_epi16( _mm_unpacklo_epi16(p32, p33), c20));  in rsdIntrinsicConvolve5x5_K()
697         o0 = _mm_add_epi32(o0, _mm_madd_epi16( _mm_unpacklo_epi16(p34, p35), c22));  in rsdIntrinsicConvolve5x5_K()
698         o0 = _mm_add_epi32(o0, _mm_madd_epi16( _mm_unpacklo_epi16(p36, _mm_setzero_si128()), c24));  in rsdIntrinsicConvolve5x5_K()
699         o0 = _mm_srai_epi32(o0, 8);  in rsdIntrinsicConvolve5x5_K()
746         o0 = packus_epi32(o0, o1);  in rsdIntrinsicConvolve5x5_K()
748         o0 = _mm_packus_epi16(o0, o2);  in rsdIntrinsicConvolve5x5_K()
749         _mm_storeu_si128((__m128i *)dst, o0);  in rsdIntrinsicConvolve5x5_K()