pyramids.cpp | 318 __m128i v_dst00 = _mm_adds_epu16(_mm_adds_epu16(v_r0, v_r2), _mm_adds_epu16(v_2r1, v_4r1)); local 332 _mm_storeu_si128((__m128i *)(dst0 + x), _mm_packus_epi16(_mm_srli_epi16(_mm_adds_epu16(v_dst00, v_delta), 6), 378 __m128i v_dst00 = _mm_add_epi32(_mm_add_epi32(v_r0, v_r2), _mm_add_epi32(v_2r1, v_4r1)); local 390 _mm_packs_epi32(_mm_srai_epi32(_mm_add_epi32(v_dst00, v_delta), 6), 438 __m128i v_dst00 = _mm_add_epi32(_mm_add_epi32(v_r0, v_r2), _mm_add_epi32(v_2r1, v_4r1)); local 450 _mm_packus_epi32(_mm_srli_epi32(_mm_add_epi32(v_dst00, v_delta), 6), 678 uint16x8_t v_dst00 = vaddq_u16(vaddq_u16(v_r0, v_r2), vaddq_u16(v_2r1, v_4r1)); local 689 vst1q_u8(dst0 + x, vcombine_u8(vqmovn_u16(vshrq_n_u16(vaddq_u16(v_dst00, v_delta), 6)), 726 uint32x4_t v_dst00 = vaddq_u32(vaddq_u32(v_r0, v_r2), vaddq_u32(v_2r1, v_4r1)); local 772 int32x4_t v_dst00 = vaddq_s32(vaddq_s32(v_r0, v_r2), vaddq_s32(v_2r1, v_4r1)); local [all...] |