| /external/libhevc/common/x86/ |
| D | ihevc_itrans_recon_32x32_ssse3_intr.c | 304 m_temp_reg_14 = _mm_madd_epi16(m_temp_reg_1, m_coeff3); in ihevc_itrans_recon_32x32_ssse3()
327 m_temp_reg_14 = _mm_madd_epi16(m_temp_reg_1, m_coeff3); in ihevc_itrans_recon_32x32_ssse3()
368 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff1); in ihevc_itrans_recon_32x32_ssse3()
385 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_14, m_coeff1); in ihevc_itrans_recon_32x32_ssse3()
398 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff2); in ihevc_itrans_recon_32x32_ssse3()
412 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_14, m_coeff2); in ihevc_itrans_recon_32x32_ssse3()
426 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff3); in ihevc_itrans_recon_32x32_ssse3()
440 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_14, m_coeff3); in ihevc_itrans_recon_32x32_ssse3()
457 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff4); in ihevc_itrans_recon_32x32_ssse3()
471 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_14, m_coeff4); in ihevc_itrans_recon_32x32_ssse3()
[all …]
|
| D | ihevc_32x32_itrans_recon_sse42_intr.c | 295 m_temp_reg_14 = _mm_madd_epi16(m_temp_reg_1, m_coeff3); in ihevc_itrans_recon_32x32_sse42()
318 m_temp_reg_14 = _mm_madd_epi16(m_temp_reg_1, m_coeff3); in ihevc_itrans_recon_32x32_sse42()
358 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff1); in ihevc_itrans_recon_32x32_sse42()
374 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_14, m_coeff1); in ihevc_itrans_recon_32x32_sse42()
387 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff2); in ihevc_itrans_recon_32x32_sse42()
401 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_14, m_coeff2); in ihevc_itrans_recon_32x32_sse42()
415 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff3); in ihevc_itrans_recon_32x32_sse42()
429 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_14, m_coeff3); in ihevc_itrans_recon_32x32_sse42()
446 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff4); in ihevc_itrans_recon_32x32_sse42()
460 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_14, m_coeff4); in ihevc_itrans_recon_32x32_sse42()
[all …]
|
| D | ihevc_16x16_itrans_recon_sse42_intr.c | 256 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff1); in ihevc_itrans_recon_16x16_sse42()
275 m_temp_reg_31 = _mm_madd_epi16(m_temp_reg_11, m_coeff1); in ihevc_itrans_recon_16x16_sse42()
294 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff3); in ihevc_itrans_recon_16x16_sse42()
311 m_temp_reg_31 = _mm_madd_epi16(m_temp_reg_11, m_coeff3); in ihevc_itrans_recon_16x16_sse42()
331 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff1); in ihevc_itrans_recon_16x16_sse42()
347 m_temp_reg_31 = _mm_madd_epi16(m_temp_reg_11, m_coeff1); in ihevc_itrans_recon_16x16_sse42()
365 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff3); in ihevc_itrans_recon_16x16_sse42()
380 m_temp_reg_31 = _mm_madd_epi16(m_temp_reg_11, m_coeff3); in ihevc_itrans_recon_16x16_sse42()
406 m_temp_reg_20 = _mm_madd_epi16(m_temp_reg_0, m_coeff1); in ihevc_itrans_recon_16x16_sse42()
407 m_temp_reg_22 = _mm_madd_epi16(m_temp_reg_0, m_coeff2); in ihevc_itrans_recon_16x16_sse42()
[all …]
|
| D | ihevc_itrans_recon_16x16_ssse3_intr.c | 262 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff1); in ihevc_itrans_recon_16x16_ssse3()
281 m_temp_reg_31 = _mm_madd_epi16(m_temp_reg_11, m_coeff1); in ihevc_itrans_recon_16x16_ssse3()
300 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff3); in ihevc_itrans_recon_16x16_ssse3()
317 m_temp_reg_31 = _mm_madd_epi16(m_temp_reg_11, m_coeff3); in ihevc_itrans_recon_16x16_ssse3()
337 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff1); in ihevc_itrans_recon_16x16_ssse3()
353 m_temp_reg_31 = _mm_madd_epi16(m_temp_reg_11, m_coeff1); in ihevc_itrans_recon_16x16_ssse3()
371 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_10, m_coeff3); in ihevc_itrans_recon_16x16_ssse3()
386 m_temp_reg_31 = _mm_madd_epi16(m_temp_reg_11, m_coeff3); in ihevc_itrans_recon_16x16_ssse3()
412 m_temp_reg_20 = _mm_madd_epi16(m_temp_reg_0, m_coeff1); in ihevc_itrans_recon_16x16_ssse3()
413 m_temp_reg_22 = _mm_madd_epi16(m_temp_reg_0, m_coeff2); in ihevc_itrans_recon_16x16_ssse3()
[all …]
|
| D | ihevc_itrans_recon_sse42_intr.c | 887 m_temp_reg_10 = _mm_madd_epi16(m_temp_reg_0, m_coeff1); in ihevc_itrans_recon_8x8_sse42()
888 m_temp_reg_12 = _mm_madd_epi16(m_temp_reg_0, m_coeff2); in ihevc_itrans_recon_8x8_sse42()
905 m_temp_reg_16 = _mm_madd_epi16(m_temp_reg_4, m_coeff1); in ihevc_itrans_recon_8x8_sse42()
906 m_temp_reg_14 = _mm_madd_epi16(m_temp_reg_4, m_coeff2); in ihevc_itrans_recon_8x8_sse42()
941 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_60, m_coeff1); in ihevc_itrans_recon_8x8_sse42()
967 m_temp_reg_32 = _mm_madd_epi16(m_temp_reg_60, m_coeff3); in ihevc_itrans_recon_8x8_sse42()
996 m_temp_reg_34 = _mm_madd_epi16(m_temp_reg_60, m_coeff1); in ihevc_itrans_recon_8x8_sse42()
1026 m_temp_reg_36 = _mm_madd_epi16(m_temp_reg_60, m_coeff3); in ihevc_itrans_recon_8x8_sse42()
1095 m_temp_reg_10 = _mm_madd_epi16(m_temp_reg_0, m_coeff1); in ihevc_itrans_recon_8x8_sse42()
1096 m_temp_reg_12 = _mm_madd_epi16(m_temp_reg_0, m_coeff2); in ihevc_itrans_recon_8x8_sse42()
[all …]
|
| D | ihevc_itrans_recon_ssse3_intr.c | 1066 m_temp_reg_10 = _mm_madd_epi16(m_temp_reg_0, m_coeff1); in ihevc_itrans_recon_8x8_ssse3()
1067 m_temp_reg_12 = _mm_madd_epi16(m_temp_reg_0, m_coeff2); in ihevc_itrans_recon_8x8_ssse3()
1084 m_temp_reg_16 = _mm_madd_epi16(m_temp_reg_4, m_coeff1); in ihevc_itrans_recon_8x8_ssse3()
1085 m_temp_reg_14 = _mm_madd_epi16(m_temp_reg_4, m_coeff2); in ihevc_itrans_recon_8x8_ssse3()
1120 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_60, m_coeff1); in ihevc_itrans_recon_8x8_ssse3()
1146 m_temp_reg_32 = _mm_madd_epi16(m_temp_reg_60, m_coeff3); in ihevc_itrans_recon_8x8_ssse3()
1175 m_temp_reg_34 = _mm_madd_epi16(m_temp_reg_60, m_coeff1); in ihevc_itrans_recon_8x8_ssse3()
1205 m_temp_reg_36 = _mm_madd_epi16(m_temp_reg_60, m_coeff3); in ihevc_itrans_recon_8x8_ssse3()
1286 m_temp_reg_10 = _mm_madd_epi16(m_temp_reg_0, m_coeff1); in ihevc_itrans_recon_8x8_ssse3()
1287 m_temp_reg_12 = _mm_madd_epi16(m_temp_reg_0, m_coeff2); in ihevc_itrans_recon_8x8_ssse3()
[all …]
|
| /external/libvpx/libvpx/vp9/encoder/x86/ |
| D | vp9_dct_intrin_sse2.c | 89 u[0] = _mm_madd_epi16(v[0], k__cospi_p16_p16); // 0 in fdct4_sse2()
90 u[1] = _mm_madd_epi16(v[0], k__cospi_p16_m16); // 2 in fdct4_sse2()
91 u[2] = _mm_madd_epi16(v[1], k__cospi_p08_p24); // 1 in fdct4_sse2()
92 u[3] = _mm_madd_epi16(v[1], k__cospi_p24_m08); // 3 in fdct4_sse2()
125 v[0] = _mm_madd_epi16(u[0], k__sinpi_p01_p02); // s0 + s2 in fadst4_sse2()
126 v[1] = _mm_madd_epi16(u[1], k__sinpi_p03_p04); // s4 + s5 in fadst4_sse2()
127 v[2] = _mm_madd_epi16(u[2], k__sinpi_p03_p03); // x1 in fadst4_sse2()
128 v[3] = _mm_madd_epi16(u[0], k__sinpi_p04_m01); // s1 - s3 in fadst4_sse2()
129 v[4] = _mm_madd_epi16(u[1], k__sinpi_m03_p02); // -s4 + s6 in fadst4_sse2()
130 v[5] = _mm_madd_epi16(u[3], k__sinpi_p03_p03); // s4 in fadst4_sse2()
[all …]
|
| /external/libaom/libaom/aom_dsp/x86/ |
| D | aom_subpixel_8t_intrin_sse2.c | 50 __m128i d1 = _mm_madd_epi16(ss_1_1, secondFilters); in aom_filter_block1d16_h4_sse2()
51 __m128i d2 = _mm_madd_epi16(ss_2_1, thirdFilters); in aom_filter_block1d16_h4_sse2()
58 d1 = _mm_madd_epi16(ss_1_2, secondFilters); in aom_filter_block1d16_h4_sse2()
59 d2 = _mm_madd_epi16(ss_2_2, thirdFilters); in aom_filter_block1d16_h4_sse2()
74 d1 = _mm_madd_epi16(ss_1_1, secondFilters); in aom_filter_block1d16_h4_sse2()
75 d2 = _mm_madd_epi16(ss_2_1, thirdFilters); in aom_filter_block1d16_h4_sse2()
82 d1 = _mm_madd_epi16(ss_1_2, secondFilters); in aom_filter_block1d16_h4_sse2()
83 d2 = _mm_madd_epi16(ss_2_2, thirdFilters); in aom_filter_block1d16_h4_sse2()
171 tmp_0 = _mm_madd_epi16(resReg23_lo_1, secondFilters); in aom_filter_block1d16_v4_sse2()
172 tmp_1 = _mm_madd_epi16(resReg23_lo_2, secondFilters); in aom_filter_block1d16_v4_sse2()
[all …]
|
| D | sum_squares_sse2.c | 42 const __m128i v_sq_01_d = _mm_madd_epi16(v_val_01_w, v_val_01_w); in sum_squares_i16_4x4_sse2()
43 const __m128i v_sq_23_d = _mm_madd_epi16(v_val_23_w, v_val_23_w); in sum_squares_i16_4x4_sse2()
98 const __m128i v_sq_0_d = _mm_madd_epi16(v_val_0_w, v_val_0_w); in aom_sum_squares_2d_i16_nxn_sse2()
99 const __m128i v_sq_1_d = _mm_madd_epi16(v_val_1_w, v_val_1_w); in aom_sum_squares_2d_i16_nxn_sse2()
100 const __m128i v_sq_2_d = _mm_madd_epi16(v_val_2_w, v_val_2_w); in aom_sum_squares_2d_i16_nxn_sse2()
101 const __m128i v_sq_3_d = _mm_madd_epi16(v_val_3_w, v_val_3_w); in aom_sum_squares_2d_i16_nxn_sse2()
163 const __m128i v_sq_0_d = _mm_madd_epi16(v_val_0_w, v_val_0_w); in aom_sum_squares_i16_64n_sse2()
164 const __m128i v_sq_1_d = _mm_madd_epi16(v_val_1_w, v_val_1_w); in aom_sum_squares_i16_64n_sse2()
165 const __m128i v_sq_2_d = _mm_madd_epi16(v_val_2_w, v_val_2_w); in aom_sum_squares_i16_64n_sse2()
166 const __m128i v_sq_3_d = _mm_madd_epi16(v_val_3_w, v_val_3_w); in aom_sum_squares_i16_64n_sse2()
[all …]
|
| D | highbd_convolve_sse2.c | 65 resReg23_lo = _mm_madd_epi16(srcReg23_lo, secondFilters); in aom_highbd_filter_block1d4_v4_sse2()
66 resReg34_lo = _mm_madd_epi16(srcReg34_lo, secondFilters); in aom_highbd_filter_block1d4_v4_sse2()
67 resReg45_lo = _mm_madd_epi16(srcReg45_lo, thirdFilters); in aom_highbd_filter_block1d4_v4_sse2()
68 resReg56_lo = _mm_madd_epi16(srcReg56_lo, thirdFilters); in aom_highbd_filter_block1d4_v4_sse2()
136 ss_23 = _mm_madd_epi16(ss_23, secondFilters); in aom_highbd_filter_block1d4_h4_sse2()
137 ss_45 = _mm_madd_epi16(ss_45, thirdFilters); in aom_highbd_filter_block1d4_h4_sse2()
210 resReg23_lo = _mm_madd_epi16(srcReg23_lo, secondFilters); in aom_highbd_filter_block1d8_v4_sse2()
211 resReg34_lo = _mm_madd_epi16(srcReg34_lo, secondFilters); in aom_highbd_filter_block1d8_v4_sse2()
212 resReg45_lo = _mm_madd_epi16(srcReg45_lo, thirdFilters); in aom_highbd_filter_block1d8_v4_sse2()
213 resReg56_lo = _mm_madd_epi16(srcReg56_lo, thirdFilters); in aom_highbd_filter_block1d8_v4_sse2()
[all …]
|
| D | fwd_txfm_impl_sse2.h | 116 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16); in FDCT8x8_2D()
117 const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16); in FDCT8x8_2D()
118 const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16); in FDCT8x8_2D()
119 const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16); in FDCT8x8_2D()
120 const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08); in FDCT8x8_2D()
121 const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08); in FDCT8x8_2D()
122 const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24); in FDCT8x8_2D()
123 const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24); in FDCT8x8_2D()
160 const __m128i e0 = _mm_madd_epi16(d0, k__cospi_p16_m16); in FDCT8x8_2D()
161 const __m128i e1 = _mm_madd_epi16(d1, k__cospi_p16_m16); in FDCT8x8_2D()
[all …]
|
| /external/libvpx/libvpx/vpx_dsp/x86/ |
| D | inv_txfm_sse2.c | 121 v[0] = _mm_madd_epi16(in[0], k__sinpi_1_3); // s_1 * x0 + s_3 * x1 in iadst4_sse2()
122 v[1] = _mm_madd_epi16(in[1], k__sinpi_4_2); // s_4 * x2 + s_2 * x3 in iadst4_sse2()
123 v[2] = _mm_madd_epi16(in[0], k__sinpi_2_3); // s_2 * x0 + s_3 * x1 in iadst4_sse2()
124 v[3] = _mm_madd_epi16(in[1], k__sinpi_1_4); // s_1 * x2 + s_4 * x3 in iadst4_sse2()
125 v[4] = _mm_madd_epi16(in[0], k__sinpi_12_n3); // (s_1 + s_2) * x0 - s_3 * x1 in iadst4_sse2()
133 u[2] = _mm_madd_epi16(in[0], k__sinpi_1_3); in iadst4_sse2()
264 u[0] = _mm_madd_epi16(s[0], k__cospi_p02_p30); in iadst8_sse2()
265 u[1] = _mm_madd_epi16(s[1], k__cospi_p02_p30); in iadst8_sse2()
266 u[2] = _mm_madd_epi16(s[0], k__cospi_p30_m02); in iadst8_sse2()
267 u[3] = _mm_madd_epi16(s[1], k__cospi_p30_m02); in iadst8_sse2()
[all …]
|
| D | sum_squares_sse2.c | 26 sq[0] = _mm_madd_epi16(s[0], s[0]); in vpx_sum_squares_2d_i16_sse2()
27 sq[1] = _mm_madd_epi16(s[1], s[1]); in vpx_sum_squares_2d_i16_sse2()
64 const __m128i v_sq_0_d = _mm_madd_epi16(v_val_0_w, v_val_0_w); in vpx_sum_squares_2d_i16_sse2()
65 const __m128i v_sq_1_d = _mm_madd_epi16(v_val_1_w, v_val_1_w); in vpx_sum_squares_2d_i16_sse2()
66 const __m128i v_sq_2_d = _mm_madd_epi16(v_val_2_w, v_val_2_w); in vpx_sum_squares_2d_i16_sse2()
67 const __m128i v_sq_3_d = _mm_madd_epi16(v_val_3_w, v_val_3_w); in vpx_sum_squares_2d_i16_sse2()
68 const __m128i v_sq_4_d = _mm_madd_epi16(v_val_4_w, v_val_4_w); in vpx_sum_squares_2d_i16_sse2()
69 const __m128i v_sq_5_d = _mm_madd_epi16(v_val_5_w, v_val_5_w); in vpx_sum_squares_2d_i16_sse2()
70 const __m128i v_sq_6_d = _mm_madd_epi16(v_val_6_w, v_val_6_w); in vpx_sum_squares_2d_i16_sse2()
71 const __m128i v_sq_7_d = _mm_madd_epi16(v_val_7_w, v_val_7_w); in vpx_sum_squares_2d_i16_sse2()
|
| D | fwd_dct32x32_impl_sse2.h | 405 const __m128i s2_20_2 = _mm_madd_epi16(s2_20_0, k__cospi_p16_m16); in FDCT32x32_2D()
406 const __m128i s2_20_3 = _mm_madd_epi16(s2_20_1, k__cospi_p16_m16); in FDCT32x32_2D()
407 const __m128i s2_21_2 = _mm_madd_epi16(s2_21_0, k__cospi_p16_m16); in FDCT32x32_2D()
408 const __m128i s2_21_3 = _mm_madd_epi16(s2_21_1, k__cospi_p16_m16); in FDCT32x32_2D()
409 const __m128i s2_22_2 = _mm_madd_epi16(s2_22_0, k__cospi_p16_m16); in FDCT32x32_2D()
410 const __m128i s2_22_3 = _mm_madd_epi16(s2_22_1, k__cospi_p16_m16); in FDCT32x32_2D()
411 const __m128i s2_23_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_m16); in FDCT32x32_2D()
412 const __m128i s2_23_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_m16); in FDCT32x32_2D()
413 const __m128i s2_24_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_p16); in FDCT32x32_2D()
414 const __m128i s2_24_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_p16); in FDCT32x32_2D()
[all …]
|
| D | convolve_sse2.h | 41 const __m128i madd_1 = _mm_madd_epi16(src_1_half, *ker_1); in mm_madd_add_epi8_sse2()
42 const __m128i madd_2 = _mm_madd_epi16(src_2_half, *ker_2); in mm_madd_add_epi8_sse2()
53 const __m128i madd_1 = _mm_madd_epi16(*src_1, *ker_1); in mm_madd_add_epi16_sse2()
54 const __m128i madd_2 = _mm_madd_epi16(*src_2, *ker_2); in mm_madd_add_epi16_sse2()
61 const __m128i madd_1 = _mm_madd_epi16(*src_0, *ker); in mm_madd_packs_epi16_sse2()
62 const __m128i madd_2 = _mm_madd_epi16(*src_1, *ker); in mm_madd_packs_epi16_sse2()
|
| /external/libaom/libaom/av1/encoder/x86/ |
| D | wedge_utils_sse2.c | 63 const __m128i v_t0l_d = _mm_madd_epi16(v_rd0l_w, v_m0l_w); in av1_wedge_sse_from_residuals_sse2()
64 const __m128i v_t0h_d = _mm_madd_epi16(v_rd0h_w, v_m0h_w); in av1_wedge_sse_from_residuals_sse2()
65 const __m128i v_t1l_d = _mm_madd_epi16(v_rd1l_w, v_m1l_w); in av1_wedge_sse_from_residuals_sse2()
66 const __m128i v_t1h_d = _mm_madd_epi16(v_rd1h_w, v_m1h_w); in av1_wedge_sse_from_residuals_sse2()
71 const __m128i v_sq0_d = _mm_madd_epi16(v_t0_w, v_t0_w); in av1_wedge_sse_from_residuals_sse2()
72 const __m128i v_sq1_d = _mm_madd_epi16(v_t1_w, v_t1_w); in av1_wedge_sse_from_residuals_sse2()
139 const __m128i v_p0_d = _mm_madd_epi16(v_d0_w, v_m0_w); in av1_wedge_sign_from_residuals_sse2()
140 const __m128i v_p1_d = _mm_madd_epi16(v_d1_w, v_m1_w); in av1_wedge_sign_from_residuals_sse2()
141 const __m128i v_p2_d = _mm_madd_epi16(v_d2_w, v_m2_w); in av1_wedge_sign_from_residuals_sse2()
142 const __m128i v_p3_d = _mm_madd_epi16(v_d3_w, v_m3_w); in av1_wedge_sign_from_residuals_sse2()
[all …]
|
| /external/libmpeg2/common/x86/ |
| D | impeg2_idct_recon_sse42_intr.c | 225 m_temp_reg_10 = _mm_madd_epi16(m_temp_reg_0, m_coeff1); in impeg2_idct_recon_sse42()
226 m_temp_reg_12 = _mm_madd_epi16(m_temp_reg_0, m_coeff2); in impeg2_idct_recon_sse42()
243 m_temp_reg_16 = _mm_madd_epi16(m_temp_reg_4, m_coeff1); in impeg2_idct_recon_sse42()
244 m_temp_reg_14 = _mm_madd_epi16(m_temp_reg_4, m_coeff2); in impeg2_idct_recon_sse42()
279 m_temp_reg_30 = _mm_madd_epi16(m_temp_reg_60, m_coeff1); in impeg2_idct_recon_sse42()
305 m_temp_reg_32 = _mm_madd_epi16(m_temp_reg_60, m_coeff3); in impeg2_idct_recon_sse42()
334 m_temp_reg_34 = _mm_madd_epi16(m_temp_reg_60, m_coeff1); in impeg2_idct_recon_sse42()
364 m_temp_reg_36 = _mm_madd_epi16(m_temp_reg_60, m_coeff3); in impeg2_idct_recon_sse42()
433 m_temp_reg_10 = _mm_madd_epi16(m_temp_reg_0, m_coeff1); in impeg2_idct_recon_sse42()
434 m_temp_reg_12 = _mm_madd_epi16(m_temp_reg_0, m_coeff2); in impeg2_idct_recon_sse42()
[all …]
|
| /external/flac/libFLAC/ |
| D | lpc_intrin_sse2.c | 86 summ = _mm_madd_epi16(q11, _mm_loadu_si128((const __m128i*)(data+i-12))); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
87 …mull = _mm_madd_epi16(q10, _mm_loadu_si128((const __m128i*)(data+i-11))); summ = _mm_add_epi32(sum… in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
88 …mull = _mm_madd_epi16(q9, _mm_loadu_si128((const __m128i*)(data+i-10))); summ = _mm_add_epi32(summ… in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
89 …mull = _mm_madd_epi16(q8, _mm_loadu_si128((const __m128i*)(data+i-9))); summ = _mm_add_epi32(summ,… in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
90 …mull = _mm_madd_epi16(q7, _mm_loadu_si128((const __m128i*)(data+i-8))); summ = _mm_add_epi32(summ,… in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
91 …mull = _mm_madd_epi16(q6, _mm_loadu_si128((const __m128i*)(data+i-7))); summ = _mm_add_epi32(summ,… in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
92 …mull = _mm_madd_epi16(q5, _mm_loadu_si128((const __m128i*)(data+i-6))); summ = _mm_add_epi32(summ,… in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
93 …mull = _mm_madd_epi16(q4, _mm_loadu_si128((const __m128i*)(data+i-5))); summ = _mm_add_epi32(summ,… in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
94 …mull = _mm_madd_epi16(q3, _mm_loadu_si128((const __m128i*)(data+i-4))); summ = _mm_add_epi32(summ,… in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
95 …mull = _mm_madd_epi16(q2, _mm_loadu_si128((const __m128i*)(data+i-3))); summ = _mm_add_epi32(summ,… in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
[all …]
|
| /external/libaom/libaom/av1/common/x86/ |
| D | highbd_wiener_convolve_ssse3.c | 76 const __m128i res_0 = _mm_madd_epi16(data, coeff_01); in av1_highbd_wiener_convolve_add_src_ssse3()
78 _mm_madd_epi16(_mm_alignr_epi8(data2, data, 4), coeff_23); in av1_highbd_wiener_convolve_add_src_ssse3()
80 _mm_madd_epi16(_mm_alignr_epi8(data2, data, 8), coeff_45); in av1_highbd_wiener_convolve_add_src_ssse3()
82 _mm_madd_epi16(_mm_alignr_epi8(data2, data, 12), coeff_67); in av1_highbd_wiener_convolve_add_src_ssse3()
91 _mm_madd_epi16(_mm_alignr_epi8(data2, data, 2), coeff_01); in av1_highbd_wiener_convolve_add_src_ssse3()
93 _mm_madd_epi16(_mm_alignr_epi8(data2, data, 6), coeff_23); in av1_highbd_wiener_convolve_add_src_ssse3()
95 _mm_madd_epi16(_mm_alignr_epi8(data2, data, 10), coeff_45); in av1_highbd_wiener_convolve_add_src_ssse3()
97 _mm_madd_epi16(_mm_alignr_epi8(data2, data, 14), coeff_67); in av1_highbd_wiener_convolve_add_src_ssse3()
154 const __m128i res_0 = _mm_madd_epi16(src_0, coeff_01); in av1_highbd_wiener_convolve_add_src_ssse3()
155 const __m128i res_2 = _mm_madd_epi16(src_2, coeff_23); in av1_highbd_wiener_convolve_add_src_ssse3()
[all …]
|
| D | wiener_convolve_sse2.c | 74 const __m128i res_0 = _mm_madd_epi16(src_0, coeff_01); in av1_wiener_convolve_add_src_sse2()
76 const __m128i res_2 = _mm_madd_epi16(src_2, coeff_23); in av1_wiener_convolve_add_src_sse2()
78 const __m128i res_4 = _mm_madd_epi16(src_4, coeff_45); in av1_wiener_convolve_add_src_sse2()
80 const __m128i res_6 = _mm_madd_epi16(src_6, coeff_67); in av1_wiener_convolve_add_src_sse2()
89 const __m128i res_1 = _mm_madd_epi16(src_1, coeff_01); in av1_wiener_convolve_add_src_sse2()
91 const __m128i res_3 = _mm_madd_epi16(src_3, coeff_23); in av1_wiener_convolve_add_src_sse2()
93 const __m128i res_5 = _mm_madd_epi16(src_5, coeff_45); in av1_wiener_convolve_add_src_sse2()
95 const __m128i res_7 = _mm_madd_epi16(src_7, coeff_67); in av1_wiener_convolve_add_src_sse2()
152 const __m128i res_0 = _mm_madd_epi16(src_0, coeff_01); in av1_wiener_convolve_add_src_sse2()
153 const __m128i res_2 = _mm_madd_epi16(src_2, coeff_23); in av1_wiener_convolve_add_src_sse2()
[all …]
|
| D | jnt_convolve_ssse3.c | 88 const __m128i res_0 = _mm_madd_epi16(src_lo, coeff_01); in av1_dist_wtd_convolve_2d_ssse3()
90 const __m128i res_2 = _mm_madd_epi16(src_2, coeff_23); in av1_dist_wtd_convolve_2d_ssse3()
92 const __m128i res_4 = _mm_madd_epi16(src_4, coeff_45); in av1_dist_wtd_convolve_2d_ssse3()
94 const __m128i res_6 = _mm_madd_epi16(src_6, coeff_67); in av1_dist_wtd_convolve_2d_ssse3()
103 const __m128i res_1 = _mm_madd_epi16(src_1, coeff_01); in av1_dist_wtd_convolve_2d_ssse3()
105 const __m128i res_3 = _mm_madd_epi16(src_3, coeff_23); in av1_dist_wtd_convolve_2d_ssse3()
107 const __m128i res_5 = _mm_madd_epi16(src_5, coeff_45); in av1_dist_wtd_convolve_2d_ssse3()
109 const __m128i res_7 = _mm_madd_epi16(src_7, coeff_67); in av1_dist_wtd_convolve_2d_ssse3()
165 const __m128i res_0 = _mm_madd_epi16(src_0, coeff_01); in av1_dist_wtd_convolve_2d_ssse3()
166 const __m128i res_2 = _mm_madd_epi16(src_2, coeff_23); in av1_dist_wtd_convolve_2d_ssse3()
[all …]
|
| D | highbd_convolve_2d_sse4.c | 240 const __m128i res_0 = _mm_madd_epi16(data, coeff_01); in av1_highbd_dist_wtd_convolve_2d_sse4_1()
242 _mm_madd_epi16(_mm_alignr_epi8(data2, data, 4), coeff_23); in av1_highbd_dist_wtd_convolve_2d_sse4_1()
244 _mm_madd_epi16(_mm_alignr_epi8(data2, data, 8), coeff_45); in av1_highbd_dist_wtd_convolve_2d_sse4_1()
246 _mm_madd_epi16(_mm_alignr_epi8(data2, data, 12), coeff_67); in av1_highbd_dist_wtd_convolve_2d_sse4_1()
255 _mm_madd_epi16(_mm_alignr_epi8(data2, data, 2), coeff_01); in av1_highbd_dist_wtd_convolve_2d_sse4_1()
257 _mm_madd_epi16(_mm_alignr_epi8(data2, data, 6), coeff_23); in av1_highbd_dist_wtd_convolve_2d_sse4_1()
259 _mm_madd_epi16(_mm_alignr_epi8(data2, data, 10), coeff_45); in av1_highbd_dist_wtd_convolve_2d_sse4_1()
261 _mm_madd_epi16(_mm_alignr_epi8(data2, data, 14), coeff_67); in av1_highbd_dist_wtd_convolve_2d_sse4_1()
317 const __m128i res_0 = _mm_madd_epi16(src_0, coeff_01); in av1_highbd_dist_wtd_convolve_2d_sse4_1()
318 const __m128i res_2 = _mm_madd_epi16(src_2, coeff_23); in av1_highbd_dist_wtd_convolve_2d_sse4_1()
[all …]
|
| D | convolve_2d_sse2.c | 76 const __m128i res_0 = _mm_madd_epi16(src_0, coeff_01); in av1_convolve_2d_sr_sse2()
78 const __m128i res_2 = _mm_madd_epi16(src_2, coeff_23); in av1_convolve_2d_sr_sse2()
80 const __m128i res_4 = _mm_madd_epi16(src_4, coeff_45); in av1_convolve_2d_sr_sse2()
82 const __m128i res_6 = _mm_madd_epi16(src_6, coeff_67); in av1_convolve_2d_sr_sse2()
91 const __m128i res_1 = _mm_madd_epi16(src_1, coeff_01); in av1_convolve_2d_sr_sse2()
93 const __m128i res_3 = _mm_madd_epi16(src_3, coeff_23); in av1_convolve_2d_sr_sse2()
95 const __m128i res_5 = _mm_madd_epi16(src_5, coeff_45); in av1_convolve_2d_sr_sse2()
97 const __m128i res_7 = _mm_madd_epi16(src_7, coeff_67); in av1_convolve_2d_sr_sse2()
157 const __m128i res_0 = _mm_madd_epi16(src_0, coeff_01); in av1_convolve_2d_sr_sse2()
158 const __m128i res_2 = _mm_madd_epi16(src_2, coeff_23); in av1_convolve_2d_sr_sse2()
[all …]
|
| /external/libavc/common/x86/ |
| D | ih264_inter_pred_filters_ssse3.c | 898 res_t1_4x32b = _mm_madd_epi16(src_r0_8x16b, coeff0_1_8x16b); in ih264_inter_pred_luma_horz_hpel_vert_hpel_ssse3()
899 res_t2_4x32b = _mm_madd_epi16(src_t1_8x16b, coeff2_3_8x16b); in ih264_inter_pred_luma_horz_hpel_vert_hpel_ssse3()
900 res_t3_4x32b = _mm_madd_epi16(src_t2_8x16b, coeff4_5_8x16b); in ih264_inter_pred_luma_horz_hpel_vert_hpel_ssse3()
912 res_t1_4x32b = _mm_madd_epi16(src_r1_8x16b, coeff0_1_8x16b); in ih264_inter_pred_luma_horz_hpel_vert_hpel_ssse3()
913 res_t2_4x32b = _mm_madd_epi16(src_t1_8x16b, coeff2_3_8x16b); in ih264_inter_pred_luma_horz_hpel_vert_hpel_ssse3()
914 res_t3_4x32b = _mm_madd_epi16(src_t2_8x16b, coeff4_5_8x16b); in ih264_inter_pred_luma_horz_hpel_vert_hpel_ssse3()
1091 res_t1_4x32b = _mm_madd_epi16(src_r0r1_8x16b, coeff0_1_8x16b); in ih264_inter_pred_luma_horz_hpel_vert_hpel_ssse3()
1092 res_t2_4x32b = _mm_madd_epi16(src_r2r3_8x16b, coeff2_3_8x16b); in ih264_inter_pred_luma_horz_hpel_vert_hpel_ssse3()
1093 res_t3_4x32b = _mm_madd_epi16(src_r4r5_8x16b, coeff4_5_8x16b); in ih264_inter_pred_luma_horz_hpel_vert_hpel_ssse3()
1104 res_t1_4x32b = _mm_madd_epi16(src_r0r1_8x16b, coeff0_1_8x16b); in ih264_inter_pred_luma_horz_hpel_vert_hpel_ssse3()
[all …]
|
| /external/libavc/encoder/x86/ |
| D | ih264e_half_pel_ssse3.c | 407 res_t1_4x32b = _mm_madd_epi16(src_r0r1_8x16b, coeff0_1_8x16b); in ih264e_sixtap_filter_2dvh_vert_ssse3()
408 res_t2_4x32b = _mm_madd_epi16(src_r2r3_8x16b, coeff2_3_8x16b); in ih264e_sixtap_filter_2dvh_vert_ssse3()
409 res_t3_4x32b = _mm_madd_epi16(src_r4r5_8x16b, coeff4_5_8x16b); in ih264e_sixtap_filter_2dvh_vert_ssse3()
420 res_t1_4x32b = _mm_madd_epi16(src_r0r1_8x16b, coeff0_1_8x16b); in ih264e_sixtap_filter_2dvh_vert_ssse3()
421 res_t2_4x32b = _mm_madd_epi16(src_r2r3_8x16b, coeff2_3_8x16b); in ih264e_sixtap_filter_2dvh_vert_ssse3()
422 res_t3_4x32b = _mm_madd_epi16(src_r4r5_8x16b, coeff4_5_8x16b); in ih264e_sixtap_filter_2dvh_vert_ssse3()
445 res_t1_4x32b = _mm_madd_epi16(src_r0r1_8x16b, coeff0_1_8x16b); in ih264e_sixtap_filter_2dvh_vert_ssse3()
446 res_t2_4x32b = _mm_madd_epi16(src_r2r3_8x16b, coeff2_3_8x16b); in ih264e_sixtap_filter_2dvh_vert_ssse3()
447 res_t3_4x32b = _mm_madd_epi16(src_r4r5_8x16b, coeff4_5_8x16b); in ih264e_sixtap_filter_2dvh_vert_ssse3()
458 res_t1_4x32b = _mm_madd_epi16(src_r0r1_8x16b, coeff0_1_8x16b); in ih264e_sixtap_filter_2dvh_vert_ssse3()
[all …]
|