Lines Matching refs:v_X1
5287 __m128i v_x1 = _mm_packs_epi32(_mm_srai_epi32(_mm_add_epi32(v_X0, _mm_loadu_si128((__m128i const *)(adelta + x + x1 + 8))), AB_BITS),
5295 _mm_interleave_epi16(v_x0, v_x1, v_y0, v_y1);
5298 _mm_storeu_si128((__m128i *)(xy + x1 * 2 + 8), v_x1);
5355 int32x4_t v_X1 = vshrq_n_s32(vaddq_s32(v__X0, vld1q_s32(adelta + x + x1 + 4)), AB_BITS - INTER_BITS);
5359 v_xy.val[0] = vcombine_s16(vqmovn_s32(vshrq_n_s32(v_X0, INTER_BITS)), vqmovn_s32(vshrq_n_s32(v_X1, INTER_BITS)));
5367 vandq_s32(v_X1, v_mask)));
5751 __m128d v_x1 = _mm_set_pd(1, 0);
5758 __m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
5760 __m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
5761 __m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
5762 v_x1 = _mm_add_pd(v_x1, v_2);
5764 v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
5766 __m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
5767 __m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
5768 v_x1 = _mm_add_pd(v_x1, v_2);
5777 __m128i v_X1, v_Y1;
5779 __m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
5781 __m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
5782 __m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
5783 v_x1 = _mm_add_pd(v_x1, v_2);
5785 v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
5787 __m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
5788 __m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
5789 v_x1 = _mm_add_pd(v_x1, v_2);
5791 v_X1 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
5800 __m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
5802 __m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
5803 __m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
5804 v_x1 = _mm_add_pd(v_x1, v_2);
5806 v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
5808 __m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
5809 __m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
5810 v_x1 = _mm_add_pd(v_x1, v_2);
5821 __m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
5823 __m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
5824 __m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
5825 v_x1 = _mm_add_pd(v_x1, v_2);
5827 v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
5829 __m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
5830 __m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
5831 v_x1 = _mm_add_pd(v_x1, v_2);
5840 v_X0 = _mm_packs_epi32(v_X0, v_X1);
5841 v_X1 = _mm_packs_epi32(v_X2, v_X3);
5845 _mm_interleave_epi16(v_X0, v_X1, v_Y0, v_Y1);
5848 _mm_storeu_si128((__m128i *)(xy + x1 * 2 + 8), v_X1);
5879 __m128d v_x1 = _mm_set_pd(1, 0);
5886 v_x1), v_W0);
5888 __m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
5889 __m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
5890 v_x1 = _mm_add_pd(v_x1, v_2);
5892 v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
5894 __m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
5895 __m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
5896 v_x1 = _mm_add_pd(v_x1, v_2);
5905 __m128i v_X1, v_Y1;
5907 __m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
5909 __m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
5910 __m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
5911 v_x1 = _mm_add_pd(v_x1, v_2);
5913 v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
5915 __m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
5916 __m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
5917 v_x1 = _mm_add_pd(v_x1, v_2);
5919 v_X1 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
5928 __m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
5930 __m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
5931 __m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
5932 v_x1 = _mm_add_pd(v_x1, v_2);
5934 v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
5936 __m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
5937 __m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
5938 v_x1 = _mm_add_pd(v_x1, v_2);
5949 __m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
5951 __m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
5952 __m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
5953 v_x1 = _mm_add_pd(v_x1, v_2);
5955 v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
5957 __m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
5958 __m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
5959 v_x1 = _mm_add_pd(v_x1, v_2);
5971 _mm_and_si128(v_X1, v_itsi1));
5981 v_X0 = _mm_packs_epi32(_mm_srai_epi32(v_X0, INTER_BITS), _mm_srai_epi32(v_X1, INTER_BITS));
5982 v_X1 = _mm_packs_epi32(_mm_srai_epi32(v_X2, INTER_BITS), _mm_srai_epi32(v_X3, INTER_BITS));
5986 _mm_interleave_epi16(v_X0, v_X1, v_Y0, v_Y1);
5989 _mm_storeu_si128((__m128i *)(xy + x1 * 2 + 8), v_X1);
