83     __m128i c0, c2, c4, c6, c8;  in rsdIntrinsicConvolve3x3_K()  local
91     c2 = _mm_shuffle_epi32(x, 0x55);  in rsdIntrinsicConvolve3x3_K()
116         o0 = _mm_add_epi32(o0, _mm_madd_epi16(_mm_unpacklo_epi16(p2, p4), c2));  in rsdIntrinsicConvolve3x3_K()
117         o1 = _mm_add_epi32(o1, _mm_madd_epi16(_mm_unpacklo_epi16(p3, p5), c2));  in rsdIntrinsicConvolve3x3_K()
151     __m128i c0, c1, c2, c3;  in rsdIntrinsicColorMatrix4x4_K()  local
161     c2 = _mm_loadl_epi64((const __m128i *)(coef+8));  in rsdIntrinsicColorMatrix4x4_K()
163     c2 = _mm_unpacklo_epi16(c2, c3);  in rsdIntrinsicColorMatrix4x4_K()
175         x2 = _mm_add_epi32(x2, _mm_madd_epi16(zw, _mm_shuffle_epi32(c2, 0x00)));  in rsdIntrinsicColorMatrix4x4_K()
176         y2 = _mm_add_epi32(y2, _mm_madd_epi16(zw, _mm_shuffle_epi32(c2, 0x55)));  in rsdIntrinsicColorMatrix4x4_K()
177         z2 = _mm_add_epi32(z2, _mm_madd_epi16(zw, _mm_shuffle_epi32(c2, 0xaa)));  in rsdIntrinsicColorMatrix4x4_K()
178         w2 = _mm_add_epi32(w2, _mm_madd_epi16(zw, _mm_shuffle_epi32(c2, 0xff)));  in rsdIntrinsicColorMatrix4x4_K()
207     __m128i c0, c1, c2, c3;  in rsdIntrinsicColorMatrix3x3_K()  local
217     c2 = _mm_loadl_epi64((const __m128i *)(coef+8));  in rsdIntrinsicColorMatrix3x3_K()
219     c2 = _mm_unpacklo_epi16(c2, c3);  in rsdIntrinsicColorMatrix3x3_K()
230         x2 = _mm_add_epi32(x2, _mm_madd_epi16(zw, _mm_shuffle_epi32(c2, 0x00)));  in rsdIntrinsicColorMatrix3x3_K()
231         y2 = _mm_add_epi32(y2, _mm_madd_epi16(zw, _mm_shuffle_epi32(c2, 0x55)));  in rsdIntrinsicColorMatrix3x3_K()
232         z2 = _mm_add_epi32(z2, _mm_madd_epi16(zw, _mm_shuffle_epi32(c2, 0xaa)));  in rsdIntrinsicColorMatrix3x3_K()
259     __m128i c0, c1, c2, c3;  in rsdIntrinsicColorMatrixDot_K()  local
271     c2 = _mm_loadl_epi64((const __m128i *)(coef+8));  in rsdIntrinsicColorMatrixDot_K()
272     c2 = _mm_shufflelo_epi16(c2, 0);  in rsdIntrinsicColorMatrixDot_K()
275     c2 = _mm_unpacklo_epi16(c2, c3);  in rsdIntrinsicColorMatrixDot_K()
284         x2 = _mm_add_epi32(x2, _mm_madd_epi16(zw, c2));  in rsdIntrinsicColorMatrixDot_K()
414     __m128i c0, c1, c2, c3, c4;  in rsdIntrinsicYuv_K()  local
421     c2 = _mm_set1_epi32(param[2]);  /* -100 */  in rsdIntrinsicYuv_K()
446         G = _mm_add_epi32(Y, mullo_epi32(U, c2));  in rsdIntrinsicYuv_K()
476     __m128i c0, c1, c2, c3, c4;  in rsdIntrinsicYuvR_K()  local
483     c2 = _mm_set1_epi32(param[2]);  /* -100 */  in rsdIntrinsicYuvR_K()
508         G = _mm_add_epi32(Y, mullo_epi32(U, c2));  in rsdIntrinsicYuvR_K()
538     __m128i c0, c1, c2, c3, c4;  in rsdIntrinsicYuv2_K()  local
545     c2 = _mm_set1_epi32(param[2]);  /* -100 */  in rsdIntrinsicYuv2_K()
569         G = _mm_add_epi32(Y, mullo_epi32(U, c2));  in rsdIntrinsicYuv2_K()
600     __m128i c0, c2, c4, c6, c8, c10, c12;  in rsdIntrinsicConvolve5x5_K()  local
613     c2  = _mm_shuffle_epi32(x, 0x55);  in rsdIntrinsicConvolve5x5_K()
686         o0 = _mm_add_epi32(o0, _mm_madd_epi16( _mm_unpacklo_epi16(p2, p3),  c2));  in rsdIntrinsicConvolve5x5_K()
701         o1 = _mm_add_epi32(o1, _mm_madd_epi16( _mm_unpacklo_epi16(p3,p4),  c2));  in rsdIntrinsicConvolve5x5_K()
716         o2 = _mm_add_epi32(o2, _mm_madd_epi16( _mm_unpacklo_epi16(p4, p5),  c2));  in rsdIntrinsicConvolve5x5_K()
731         o3 = _mm_add_epi32(o3, _mm_madd_epi16( _mm_unpacklo_epi16(p5, p6),  c2));  in rsdIntrinsicConvolve5x5_K()