310             __m128i v_r0 = _mm_packs_epi32(_mm_loadu_si128((__m128i const *)(row0 + x)),  in operator ()()  local
340             __m128i v_r0 = _mm_packs_epi32(_mm_loadu_si128((__m128i const *)(row0 + x)),  in operator ()()  local
374             __m128i v_r0 = _mm_loadu_si128((__m128i const *)(row0 + x)),  in operator ()()  local
399             __m128i v_r0 = _mm_loadu_si128((__m128i const *)(row0 + x)),  in operator ()()  local
434             __m128i v_r0 = _mm_loadu_si128((__m128i const *)(row0 + x)),  in operator ()()  local
459             __m128i v_r0 = _mm_loadu_si128((__m128i const *)(row0 + x)),  in operator ()()  local
499             __m128 v_r0 = _mm_loadu_ps(row0 + x);  in operator ()()  local
532 …uint16x8_t v_r0 = vcombine_u16(vqmovn_u32(vld1q_u32(row0 + x)), vqmovn_u32(vld1q_u32(row0 + x + 4)…  in operator ()()  local
637             float32x4_t v_r0 = vld1q_f32(row0 + x);  in operator ()()  local
673 …uint16x8_t v_r0 = vcombine_u16(vqmovn_u32(vld1q_u32(row0 + x)), vqmovn_u32(vld1q_u32(row0 + x + 4)…  in operator ()()  local
697 …uint16x8_t v_r0 = vcombine_u16(vqmovn_u32(vld1q_u32(row0 + x)), vqmovn_u32(vld1q_u32(row0 + x + 4)…  in operator ()()  local
724 …     uint32x4_t v_r0 = vld1q_u32(row0 + x), v_r1 = vld1q_u32(row1 + x), v_r2 = vld1q_u32(row2 + x);  in operator ()()  local
745 …     uint32x4_t v_r0 = vld1q_u32(row0 + x), v_r1 = vld1q_u32(row1 + x), v_r2 = vld1q_u32(row2 + x);  in operator ()()  local
770 …      int32x4_t v_r0 = vld1q_s32(row0 + x), v_r1 = vld1q_s32(row1 + x), v_r2 = vld1q_s32(row2 + x);  in operator ()()  local
791 …      int32x4_t v_r0 = vld1q_s32(row0 + x), v_r1 = vld1q_s32(row1 + x), v_r2 = vld1q_s32(row2 + x);  in operator ()()  local
816             float32x4_t v_r0 = vld1q_f32(row0 + x);  in operator ()()  local