Home
last modified time | relevance | path

Searched refs:_mm_madd_epi16 (Results 1 – 25 of 52) sorted by relevance

123

/external/libhevc/common/x86/
Dihevc_itrans_recon_32x32_ssse3_intr.c304 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 …]
Dihevc_32x32_itrans_recon_sse42_intr.c295 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 …]
Dihevc_itrans_recon_16x16_ssse3_intr.c262 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 …]
Dihevc_16x16_itrans_recon_sse42_intr.c256 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 …]
Dihevc_itrans_recon_sse42_intr.c887 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 …]
Dihevc_itrans_recon_ssse3_intr.c1066 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 …]
Dihevc_inter_pred_filters_ssse3_intr.c2686 s0_8x16b = _mm_madd_epi16(s3_0_16x8b, coeff0_1_8x16b); in ihevc_inter_pred_luma_vert_w16inp_ssse3()
2696 s1_8x16b = _mm_madd_epi16(s3_1_16x8b, coeff2_3_8x16b); in ihevc_inter_pred_luma_vert_w16inp_ssse3()
2706 s2_8x16b = _mm_madd_epi16(s3_2_16x8b, coeff4_5_8x16b); in ihevc_inter_pred_luma_vert_w16inp_ssse3()
2716 s3_8x16b = _mm_madd_epi16(s3_3_16x8b, coeff6_7_8x16b); in ihevc_inter_pred_luma_vert_w16inp_ssse3()
2747 s20_8x16b = _mm_madd_epi16(s3_1_16x8b, coeff0_1_8x16b); in ihevc_inter_pred_luma_vert_w16inp_ssse3()
2748 s21_8x16b = _mm_madd_epi16(s3_2_16x8b, coeff2_3_8x16b); in ihevc_inter_pred_luma_vert_w16inp_ssse3()
2749 s22_8x16b = _mm_madd_epi16(s3_3_16x8b, coeff4_5_8x16b); in ihevc_inter_pred_luma_vert_w16inp_ssse3()
2759 s23_8x16b = _mm_madd_epi16(s3_4_16x8b, coeff6_7_8x16b); in ihevc_inter_pred_luma_vert_w16inp_ssse3()
2793 s10_8x16b = _mm_madd_epi16(s4_0_16x8b, coeff0_1_8x16b); in ihevc_inter_pred_luma_vert_w16inp_ssse3()
2797 s11_8x16b = _mm_madd_epi16(s4_1_16x8b, coeff2_3_8x16b); in ihevc_inter_pred_luma_vert_w16inp_ssse3()
[all …]
/external/libvpx/libvpx/vp9/encoder/x86/
Dvp9_dct_intrin_sse2.c88 u[0] = _mm_madd_epi16(v[0], k__cospi_p16_p16); // 0 in fdct4_sse2()
89 u[1] = _mm_madd_epi16(v[0], k__cospi_p16_m16); // 2 in fdct4_sse2()
90 u[2] = _mm_madd_epi16(v[1], k__cospi_p08_p24); // 1 in fdct4_sse2()
91 u[3] = _mm_madd_epi16(v[1], k__cospi_p24_m08); // 3 in fdct4_sse2()
124 v[0] = _mm_madd_epi16(u[0], k__sinpi_p01_p02); // s0 + s2 in fadst4_sse2()
125 v[1] = _mm_madd_epi16(u[1], k__sinpi_p03_p04); // s4 + s5 in fadst4_sse2()
126 v[2] = _mm_madd_epi16(u[2], k__sinpi_p03_p03); // x1 in fadst4_sse2()
127 v[3] = _mm_madd_epi16(u[0], k__sinpi_p04_m01); // s1 - s3 in fadst4_sse2()
128 v[4] = _mm_madd_epi16(u[1], k__sinpi_m03_p02); // -s4 + s6 in fadst4_sse2()
129 v[5] = _mm_madd_epi16(u[3], k__sinpi_p03_p03); // s4 in fadst4_sse2()
[all …]
Dvp9_dct_ssse3.c106 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16); in vp9_fdct8x8_quant_ssse3()
107 const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16); in vp9_fdct8x8_quant_ssse3()
108 const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16); in vp9_fdct8x8_quant_ssse3()
109 const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16); in vp9_fdct8x8_quant_ssse3()
111 const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08); in vp9_fdct8x8_quant_ssse3()
112 const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08); in vp9_fdct8x8_quant_ssse3()
113 const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24); in vp9_fdct8x8_quant_ssse3()
114 const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24); in vp9_fdct8x8_quant_ssse3()
161 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04); in vp9_fdct8x8_quant_ssse3()
162 const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04); in vp9_fdct8x8_quant_ssse3()
[all …]
/external/libmpeg2/common/x86/
Dimpeg2_idct_recon_sse42_intr.c225 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 …]
Dicv_variance_ssse3.c140 sqr_r0 = _mm_madd_epi16(ssrc_r0, ssrc_r0); in icv_variance_8x4_ssse3()
141 sqr_r1 = _mm_madd_epi16(ssrc_r1, ssrc_r1); in icv_variance_8x4_ssse3()
142 sqr_r2 = _mm_madd_epi16(ssrc_r2, ssrc_r2); in icv_variance_8x4_ssse3()
143 sqr_r3 = _mm_madd_epi16(ssrc_r3, ssrc_r3); in icv_variance_8x4_ssse3()
/external/libvpx/libvpx/vpx_dsp/x86/
Dinv_txfm_sse2.c122 v[0] = _mm_madd_epi16(u[0], k__sinpi_p01_p04); // s0 + s3 in iadst4_sse2()
123 v[1] = _mm_madd_epi16(u[1], k__sinpi_p03_p02); // s2 + s5 in iadst4_sse2()
124 v[2] = _mm_madd_epi16(u[2], k__sinpi_p03_p03); // x2 in iadst4_sse2()
125 v[3] = _mm_madd_epi16(u[0], k__sinpi_p02_m01); // s1 - s4 in iadst4_sse2()
126 v[4] = _mm_madd_epi16(u[1], k__sinpi_p03_m04); // s2 - s6 in iadst4_sse2()
127 v[5] = _mm_madd_epi16(u[3], k__sinpi_p03_p03); // s2 in iadst4_sse2()
285 u0 = _mm_madd_epi16(s0, k__cospi_p02_p30); in iadst8_sse2()
286 u1 = _mm_madd_epi16(s1, k__cospi_p02_p30); in iadst8_sse2()
287 u2 = _mm_madd_epi16(s0, k__cospi_p30_m02); in iadst8_sse2()
288 u3 = _mm_madd_epi16(s1, k__cospi_p30_m02); in iadst8_sse2()
[all …]
Dsum_squares_sse2.c28 const __m128i v_sq_0_d = _mm_madd_epi16(v_val_0_w, v_val_0_w); in vpx_sum_squares_2d_i16_4x4_sse2()
29 const __m128i v_sq_1_d = _mm_madd_epi16(v_val_1_w, v_val_1_w); in vpx_sum_squares_2d_i16_4x4_sse2()
30 const __m128i v_sq_2_d = _mm_madd_epi16(v_val_2_w, v_val_2_w); in vpx_sum_squares_2d_i16_4x4_sse2()
31 const __m128i v_sq_3_d = _mm_madd_epi16(v_val_3_w, v_val_3_w); in vpx_sum_squares_2d_i16_4x4_sse2()
78 const __m128i v_sq_0_d = _mm_madd_epi16(v_val_0_w, v_val_0_w); in vpx_sum_squares_2d_i16_nxn_sse2()
79 const __m128i v_sq_1_d = _mm_madd_epi16(v_val_1_w, v_val_1_w); in vpx_sum_squares_2d_i16_nxn_sse2()
80 const __m128i v_sq_2_d = _mm_madd_epi16(v_val_2_w, v_val_2_w); in vpx_sum_squares_2d_i16_nxn_sse2()
81 const __m128i v_sq_3_d = _mm_madd_epi16(v_val_3_w, v_val_3_w); in vpx_sum_squares_2d_i16_nxn_sse2()
82 const __m128i v_sq_4_d = _mm_madd_epi16(v_val_4_w, v_val_4_w); in vpx_sum_squares_2d_i16_nxn_sse2()
83 const __m128i v_sq_5_d = _mm_madd_epi16(v_val_5_w, v_val_5_w); in vpx_sum_squares_2d_i16_nxn_sse2()
[all …]
Dfwd_dct32x32_impl_sse2.h403 const __m128i s2_20_2 = _mm_madd_epi16(s2_20_0, k__cospi_p16_m16); in FDCT32x32_2D()
404 const __m128i s2_20_3 = _mm_madd_epi16(s2_20_1, k__cospi_p16_m16); in FDCT32x32_2D()
405 const __m128i s2_21_2 = _mm_madd_epi16(s2_21_0, k__cospi_p16_m16); in FDCT32x32_2D()
406 const __m128i s2_21_3 = _mm_madd_epi16(s2_21_1, k__cospi_p16_m16); in FDCT32x32_2D()
407 const __m128i s2_22_2 = _mm_madd_epi16(s2_22_0, k__cospi_p16_m16); in FDCT32x32_2D()
408 const __m128i s2_22_3 = _mm_madd_epi16(s2_22_1, k__cospi_p16_m16); in FDCT32x32_2D()
409 const __m128i s2_23_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_m16); in FDCT32x32_2D()
410 const __m128i s2_23_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_m16); in FDCT32x32_2D()
411 const __m128i s2_24_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_p16); in FDCT32x32_2D()
412 const __m128i s2_24_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_p16); in FDCT32x32_2D()
[all …]
Dfwd_txfm_impl_sse2.h150 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_A); in FDCT4x4_2D()
151 const __m128i u2 = _mm_madd_epi16(t0, k__cospi_B); in FDCT4x4_2D()
152 const __m128i u1 = _mm_madd_epi16(t1, k__cospi_C); in FDCT4x4_2D()
153 const __m128i u3 = _mm_madd_epi16(t1, k__cospi_D); in FDCT4x4_2D()
206 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_E); in FDCT4x4_2D()
207 const __m128i u1 = _mm_madd_epi16(t0, k__cospi_F); in FDCT4x4_2D()
208 const __m128i u2 = _mm_madd_epi16(t1, k__cospi_G); in FDCT4x4_2D()
209 const __m128i u3 = _mm_madd_epi16(t1, k__cospi_H); in FDCT4x4_2D()
341 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16); in FDCT8x8_2D()
342 const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16); in FDCT8x8_2D()
[all …]
/external/flac/libFLAC/
Dlpc_intrin_sse2.c86 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/libavc/common/x86/
Dih264_inter_pred_filters_ssse3.c898 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 …]
Dih264_iquant_itrans_recon_ssse3.c150 …temp4 = _mm_madd_epi16(src_r0, temp4); //a00*b00*q00 a10*b10*q10 a20*b20*q20 a30*b30 q30 -- 32 bit… in ih264_iquant_itrans_recon_4x4_ssse3()
151 temp5 = _mm_madd_epi16(src_r1, temp5); in ih264_iquant_itrans_recon_4x4_ssse3()
152 temp6 = _mm_madd_epi16(src_r2, temp6); in ih264_iquant_itrans_recon_4x4_ssse3()
153 temp7 = _mm_madd_epi16(src_r3, temp7); in ih264_iquant_itrans_recon_4x4_ssse3()
421 …temp5 = _mm_madd_epi16(src_r0_1, scalemat_r0_1); // a00*b00*q0 a01*b01*q1 a02*b02*q2 a03*b03*q3 --… in ih264_iquant_itrans_recon_8x8_ssse3()
422 …temp7 = _mm_madd_epi16(src_r0_2, scalemat_r0_2); // a04*b04*q4 a05*b05*q5 a06*b06*q6 a07*b07*q7 --… in ih264_iquant_itrans_recon_8x8_ssse3()
443 …temp5 = _mm_madd_epi16(src_r0_1, scalemat_r0_1); // a00*b00*q0 a01*b01*q1 a02*b02*q2 a03*b03*q3 --… in ih264_iquant_itrans_recon_8x8_ssse3()
444 …temp7 = _mm_madd_epi16(src_r0_2, scalemat_r0_2); // a04*b04*q4 a05*b05*q5 a06*b06*q6 a07*b07*q7 --… in ih264_iquant_itrans_recon_8x8_ssse3()
464 …temp5 = _mm_madd_epi16(src_r0_1, scalemat_r0_1); // a00*b00*q0 a01*b01*q1 a02*b02*q2 a03*b03*q3 --… in ih264_iquant_itrans_recon_8x8_ssse3()
465 …temp7 = _mm_madd_epi16(src_r0_2, scalemat_r0_2); // a04*b04*q4 a05*b05*q5 a06*b06*q6 a07*b07*q7 --… in ih264_iquant_itrans_recon_8x8_ssse3()
[all …]
Dih264_ihadamard_scaling_ssse3.c180 src_r0 = _mm_madd_epi16(src_r0, mult_val); in ih264_ihadamard_scaling_4x4_ssse3()
181 src_r1 = _mm_madd_epi16(src_r1, mult_val); in ih264_ihadamard_scaling_4x4_ssse3()
182 src_r2 = _mm_madd_epi16(src_r2, mult_val); in ih264_ihadamard_scaling_4x4_ssse3()
183 src_r3 = _mm_madd_epi16(src_r3, mult_val); in ih264_ihadamard_scaling_4x4_ssse3()
/external/libavc/encoder/x86/
Dih264e_half_pel_ssse3.c407 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 …]
/external/libopus/celt/x86/
Dpitch_sse4_1.c67 inVec1_76543210 = _mm_madd_epi16(inVec1_76543210, inVec2_76543210); in celt_inner_prod_sse4_1()
68 inVec1_FEDCBA98 = _mm_madd_epi16(inVec1_FEDCBA98, inVec2_FEDCBA98); in celt_inner_prod_sse4_1()
81 inVec1_76543210 = _mm_madd_epi16(inVec1_76543210, inVec2_76543210); in celt_inner_prod_sse4_1()
135 sum0 = _mm_add_epi32(sum0, _mm_madd_epi16(vecX, vecY0)); in xcorr_kernel_sse4_1()
136 sum1 = _mm_add_epi32(sum1, _mm_madd_epi16(vecX, vecY1)); in xcorr_kernel_sse4_1()
137 sum2 = _mm_add_epi32(sum2, _mm_madd_epi16(vecX, vecY2)); in xcorr_kernel_sse4_1()
138 sum3 = _mm_add_epi32(sum3, _mm_madd_epi16(vecX, vecY3)); in xcorr_kernel_sse4_1()
Dpitch_sse2.c65 inVec1_76543210 = _mm_madd_epi16(inVec1_76543210, inVec2_76543210); in celt_inner_prod_sse2()
66 inVec1_FEDCBA98 = _mm_madd_epi16(inVec1_FEDCBA98, inVec2_FEDCBA98); in celt_inner_prod_sse2()
79 inVec1_76543210 = _mm_madd_epi16(inVec1_76543210, inVec2_76543210); in celt_inner_prod_sse2()
/external/neven/Embedded/common/src/b_BasicEm/
DMathSSE2.c63 m_XMM0 = _mm_madd_epi16( m_XMM0, m_XMM1 ); in bbs_dotProduct_64SSE2()
68 m_XMM2 = _mm_madd_epi16( m_XMM2, m_XMM3 ); in bbs_dotProduct_64SSE2()
76 m_XMM4 = _mm_madd_epi16( m_XMM4, m_XMM5 ); in bbs_dotProduct_64SSE2()
79 m_XMM6 = _mm_madd_epi16( m_XMM6, m_XMM8 ); in bbs_dotProduct_64SSE2()
172 m_XMM0 = _mm_madd_epi16( m_XMM0, m_XMM2 ); in bbs_dotProduct_128SSE2()
178 m_XMM6 = _mm_madd_epi16( m_XMM6, m_XMM3 ); in bbs_dotProduct_128SSE2()
276 m_XMM0 = _mm_madd_epi16( m_XMM0, m_XMM2 ); in bbs_dotProduct_u128SSE2()
282 m_XMM6 = _mm_madd_epi16( m_XMM6, m_XMM3 ); in bbs_dotProduct_u128SSE2()
/external/webp/src/dsp/
Dssim_sse2.c37 const __m128i sum1 = _mm_madd_epi16(C0, C0); in SubtractAndSquare_SSE2()
38 const __m128i sum2 = _mm_madd_epi16(C1, C1); in SubtractAndSquare_SSE2()
118 xxm = _mm_add_epi32(xxm, _mm_madd_epi16(a1, wa1)); \
119 xym = _mm_add_epi32(xym, _mm_madd_epi16(a1, wb1)); \
120 yym = _mm_add_epi32(yym, _mm_madd_epi16(b1, wb1)); \
/external/gemmlowp/internal/
Dpack_sse.h98 xmm2 = _mm_madd_epi16(xmm1, one); in Pack()
104 xmm2 = _mm_madd_epi16(xmm1, one); in Pack()
108 xmm2 = _mm_madd_epi16(xmm1, one); in Pack()
112 xmm2 = _mm_madd_epi16(xmm1, one); in Pack()

123