352             auto _23 =                _mm_loadu_si128((const __m128i*)(ptr +  6))    ;  in load3()  local
356             _2 = _23; _3 = _mm_srli_si128(_23, 6),  in load3()
376         __m128i _01, _23, _45, _67;  in load4()  local
379             _01 = _23 = _45 = _67 = _mm_setzero_si128();  in load4()
382             if (tail > 2) { _23 = _mm_loadl_pd(_23, src+2); }  in load4()
383             if (tail > 3) { _23 = _mm_loadh_pd(_23, src+3); }  in load4()
389             _23 = _mm_loadu_si128(((__m128i*)ptr) + 1);  in load4()
394         auto _02 = _mm_unpacklo_epi16(_01, _23),  // r0 r2 g0 g2 b0 b2 a0 a2  in load4()
395              _13 = _mm_unpackhi_epi16(_01, _23),  // r1 r3 g1 g3 b1 b3 a1 a3  in load4()
416              _23 = _mm_unpackhi_epi32(rg0123, ba0123),  in store4()  local
424             if (tail > 2) { _mm_storel_pd(dst+2, _23); }  in store4()
425             if (tail > 3) { _mm_storeh_pd(dst+3, _23); }  in store4()
431             _mm_storeu_si128((__m128i*)ptr + 1, _23);  in store4()
482               _23 = _mm256_permute2f128_ps(_26, _37, 32),  in store4()  local
486             _mm256_storeu_ps(ptr+ 8, _23);  in store4()
560                  _23 = _mm_srli_si128(_mm_loadu_si128((const __m128i*)(ptr + 4)), 4);  in load3()  local
565             _2 = _23;  in load3()
566             _3 = _mm_srli_si128(_23, 6);  in load3()
583         __m128i _01, _23;  in load4()  local
585             _01 = _23 = _mm_setzero_si128();  in load4()
589             if (tail > 2) { _23 = _mm_loadl_pd(_23, src + 2); } // r2 g2 b2 a2 00 00 00 00  in load4()
592             _23 = _mm_loadu_si128(((__m128i*)ptr) + 1); // r2 g2 b2 a2 r3 g3 b3 a3  in load4()
595         auto _02 = _mm_unpacklo_epi16(_01, _23),  // r0 r2 g0 g2 b0 b2 a0 a2  in load4()
596              _13 = _mm_unpackhi_epi16(_01, _23);  // r1 r3 g1 g3 b1 b3 a1 a3  in load4()