canny.cpp | 363 __m128i v_dx = _mm_loadu_si128((const __m128i *)(_dx + j)); local 365 v_dx = _mm_max_epi16(v_dx, _mm_sub_epi16(v_zero, v_dx)); 368 __m128i v_norm = _mm_add_epi32(_mm_unpacklo_epi16(v_dx, v_zero), _mm_unpacklo_epi16(v_dy, v_zero)); 371 v_norm = _mm_add_epi32(_mm_unpackhi_epi16(v_dx, v_zero), _mm_unpackhi_epi16(v_dy, v_zero)); 378 int16x8_t v_dx = vld1q_s16(_dx + j), v_dy = vld1q_s16(_dy + j); local 379 vst1q_s32(_norm + j, vaddq_s32(vabsq_s32(vmovl_s16(vget_low_s16(v_dx))), 381 vst1q_s32(_norm + j + 4, vaddq_s32(vabsq_s32(vmovl_s16(vget_high_s16(v_dx))), 396 __m128i v_dx = _mm_loadu_si128((const __m128i *)(_dx + j)) local 412 int16x8_t v_dx = vld1q_s16(_dx + j), v_dy = vld1q_s16(_dy + j); local 767 __m128i v_dx = _mm_loadu_si128((const __m128i *)(_dx + j)); local 782 int16x8_t v_dx = vld1q_s16(_dx + j), v_dy = vld1q_s16(_dy + j); local 800 __m128i v_dx = _mm_loadu_si128((const __m128i *)(_dx + j)); local 816 int16x8_t v_dx = vld1q_s16(_dx + j), v_dy = vld1q_s16(_dy + j); local [all...] |