| /external/libaom/libaom/av1/encoder/x86/ |
| D | av1_fwd_txfm1d_sse4.c | 25 buf1[31] = _mm_sub_epi32(input[startidx], input[endidx]); in av1_fdct32_sse4_1()
29 buf1[30] = _mm_sub_epi32(input[startidx], input[endidx]); in av1_fdct32_sse4_1()
33 buf1[29] = _mm_sub_epi32(input[startidx], input[endidx]); in av1_fdct32_sse4_1()
37 buf1[28] = _mm_sub_epi32(input[startidx], input[endidx]); in av1_fdct32_sse4_1()
41 buf1[27] = _mm_sub_epi32(input[startidx], input[endidx]); in av1_fdct32_sse4_1()
45 buf1[26] = _mm_sub_epi32(input[startidx], input[endidx]); in av1_fdct32_sse4_1()
49 buf1[25] = _mm_sub_epi32(input[startidx], input[endidx]); in av1_fdct32_sse4_1()
53 buf1[24] = _mm_sub_epi32(input[startidx], input[endidx]); in av1_fdct32_sse4_1()
57 buf1[23] = _mm_sub_epi32(input[startidx], input[endidx]); in av1_fdct32_sse4_1()
61 buf1[22] = _mm_sub_epi32(input[startidx], input[endidx]); in av1_fdct32_sse4_1()
[all …]
|
| D | highbd_fwd_txfm_sse4.c | 75 s3 = _mm_sub_epi32(in[0], in[endidx]); in fdct4x4_sse4_1()
78 s2 = _mm_sub_epi32(in[num_col], in[endidx]); in fdct4x4_sse4_1()
84 v0 = _mm_sub_epi32(u0, u1); in fdct4x4_sse4_1()
102 v2 = _mm_sub_epi32(v1, v0); in fdct4x4_sse4_1()
154 s7 = _mm_sub_epi32(t, in[idx]); in fadst4x4_sse4_1()
159 t = _mm_sub_epi32(s1, s3); in fadst4x4_sse4_1()
165 s2 = _mm_sub_epi32(x2, x3); in fadst4x4_sse4_1()
166 t = _mm_sub_epi32(x2, x0); in fadst4x4_sse4_1()
535 v[7] = _mm_sub_epi32(in[startidx], in[endidx]); // v[7] in fdct4x8_sse4_1()
539 u[6] = _mm_sub_epi32(in[startidx], in[endidx]); in fdct4x8_sse4_1()
[all …]
|
| /external/libvpx/libvpx/vpx_dsp/x86/ |
| D | highbd_idct32x32_add_sse2.c | 34 step1[10] = _mm_sub_epi32(step2[9], step2[10]); in highbd_idct32_4x32_quarter_2_stage_4_to_6()
35 step1[11] = _mm_sub_epi32(step2[8], step2[11]); in highbd_idct32_4x32_quarter_2_stage_4_to_6()
36 step1[12] = _mm_sub_epi32(step2[15], step2[12]); in highbd_idct32_4x32_quarter_2_stage_4_to_6()
37 step1[13] = _mm_sub_epi32(step2[14], step2[13]); in highbd_idct32_4x32_quarter_2_stage_4_to_6()
59 step2[18] = _mm_sub_epi32(step1[17], step1[18]); in highbd_idct32_4x32_quarter_3_4_stage_4_to_7()
60 step2[19] = _mm_sub_epi32(step1[16], step1[19]); in highbd_idct32_4x32_quarter_3_4_stage_4_to_7()
61 step2[20] = _mm_sub_epi32(step1[20], step1[23]); // step2[20] = -step2[20] in highbd_idct32_4x32_quarter_3_4_stage_4_to_7()
62 step2[21] = _mm_sub_epi32(step1[21], step1[22]); // step2[21] = -step2[21] in highbd_idct32_4x32_quarter_3_4_stage_4_to_7()
68 step2[26] = _mm_sub_epi32(step1[26], step1[25]); // step2[26] = -step2[26] in highbd_idct32_4x32_quarter_3_4_stage_4_to_7()
69 step2[27] = _mm_sub_epi32(step1[27], step1[24]); // step2[27] = -step2[27] in highbd_idct32_4x32_quarter_3_4_stage_4_to_7()
[all …]
|
| D | highbd_idct32x32_add_sse4.c | 38 step1[10] = _mm_sub_epi32(step2[9], step2[10]); in highbd_idct32_4x32_quarter_2_stage_4_to_6()
39 step1[11] = _mm_sub_epi32(step2[8], step2[11]); in highbd_idct32_4x32_quarter_2_stage_4_to_6()
40 step1[12] = _mm_sub_epi32(step2[15], step2[12]); in highbd_idct32_4x32_quarter_2_stage_4_to_6()
41 step1[13] = _mm_sub_epi32(step2[14], step2[13]); in highbd_idct32_4x32_quarter_2_stage_4_to_6()
63 step2[18] = _mm_sub_epi32(step1[17], step1[18]); in highbd_idct32_4x32_quarter_3_4_stage_4_to_7()
64 step2[19] = _mm_sub_epi32(step1[16], step1[19]); in highbd_idct32_4x32_quarter_3_4_stage_4_to_7()
65 step2[20] = _mm_sub_epi32(step1[23], step1[20]); in highbd_idct32_4x32_quarter_3_4_stage_4_to_7()
66 step2[21] = _mm_sub_epi32(step1[22], step1[21]); in highbd_idct32_4x32_quarter_3_4_stage_4_to_7()
72 step2[26] = _mm_sub_epi32(step1[25], step1[26]); in highbd_idct32_4x32_quarter_3_4_stage_4_to_7()
73 step2[27] = _mm_sub_epi32(step1[24], step1[27]); in highbd_idct32_4x32_quarter_3_4_stage_4_to_7()
[all …]
|
| D | highbd_idct16x16_add_sse2.c | 24 out[2] = _mm_sub_epi32(in[1], in[2]); in highbd_idct16_4col_stage5()
25 out[3] = _mm_sub_epi32(in[0], in[3]); in highbd_idct16_4col_stage5()
29 out[10] = _mm_sub_epi32(in[9], in[10]); in highbd_idct16_4col_stage5()
30 out[11] = _mm_sub_epi32(in[8], in[11]); in highbd_idct16_4col_stage5()
31 out[12] = _mm_sub_epi32(in[15], in[12]); in highbd_idct16_4col_stage5()
32 out[13] = _mm_sub_epi32(in[14], in[13]); in highbd_idct16_4col_stage5()
43 out[4] = _mm_sub_epi32(in[3], in[4]); in highbd_idct16_4col_stage6()
44 out[5] = _mm_sub_epi32(in[2], in[5]); in highbd_idct16_4col_stage6()
45 out[6] = _mm_sub_epi32(in[1], in[6]); in highbd_idct16_4col_stage6()
46 out[7] = _mm_sub_epi32(in[0], in[7]); in highbd_idct16_4col_stage6()
[all …]
|
| D | highbd_idct16x16_add_sse4.c | 25 out[2] = _mm_sub_epi32(in[1], in[2]); in highbd_idct16_4col_stage5()
26 out[3] = _mm_sub_epi32(in[0], in[3]); in highbd_idct16_4col_stage5()
30 out[10] = _mm_sub_epi32(in[9], in[10]); in highbd_idct16_4col_stage5()
31 out[11] = _mm_sub_epi32(in[8], in[11]); in highbd_idct16_4col_stage5()
32 out[12] = _mm_sub_epi32(in[15], in[12]); in highbd_idct16_4col_stage5()
33 out[13] = _mm_sub_epi32(in[14], in[13]); in highbd_idct16_4col_stage5()
44 out[4] = _mm_sub_epi32(in[3], in[4]); in highbd_idct16_4col_stage6()
45 out[5] = _mm_sub_epi32(in[2], in[5]); in highbd_idct16_4col_stage6()
46 out[6] = _mm_sub_epi32(in[1], in[6]); in highbd_idct16_4col_stage6()
47 out[7] = _mm_sub_epi32(in[0], in[7]); in highbd_idct16_4col_stage6()
[all …]
|
| D | post_proc_sse2.c | 94 sumsq_0 = _mm_sub_epi32(sumsq_0, _mm_unpacklo_epi16(above_sq, zero)); in vpx_mbpost_proc_down_sse2()
95 sumsq_1 = _mm_sub_epi32(sumsq_1, _mm_unpackhi_epi16(above_sq, zero)); in vpx_mbpost_proc_down_sse2()
103 mask_0 = _mm_sub_epi32(mask_0, sumsq_0); in vpx_mbpost_proc_down_sse2()
105 mask_1 = _mm_sub_epi32(mask_1, sumsq_1); in vpx_mbpost_proc_down_sse2()
110 mask_0 = _mm_sub_epi32(mask_0, _mm_unpacklo_epi16(multmp_0, multmp_1)); in vpx_mbpost_proc_down_sse2()
111 mask_1 = _mm_sub_epi32(mask_1, _mm_unpackhi_epi16(multmp_0, multmp_1)); in vpx_mbpost_proc_down_sse2()
114 mask_0 = _mm_sub_epi32(mask_0, f); in vpx_mbpost_proc_down_sse2()
115 mask_1 = _mm_sub_epi32(mask_1, f); in vpx_mbpost_proc_down_sse2()
|
| D | inv_txfm_sse2.c | 132 u[1] = _mm_sub_epi32(v[2], v[3]); in iadst4_sse2()
134 u[3] = _mm_sub_epi32(v[1], v[3]); in iadst4_sse2()
290 w[8] = _mm_sub_epi32(u[0], u[8]); in iadst8_sse2()
291 w[9] = _mm_sub_epi32(u[1], u[9]); in iadst8_sse2()
292 w[10] = _mm_sub_epi32(u[2], u[10]); in iadst8_sse2()
293 w[11] = _mm_sub_epi32(u[3], u[11]); in iadst8_sse2()
294 w[12] = _mm_sub_epi32(u[4], u[12]); in iadst8_sse2()
295 w[13] = _mm_sub_epi32(u[5], u[13]); in iadst8_sse2()
296 w[14] = _mm_sub_epi32(u[6], u[14]); in iadst8_sse2()
297 w[15] = _mm_sub_epi32(u[7], u[15]); in iadst8_sse2()
[all …]
|
| /external/flac/src/libFLAC/ |
| D | fixed_intrin_sse2.c | 72 last_error = _mm_sub_epi32(last_error, _mm_cvtsi32_si128(itmp)); // 0 0 le0 le1 in FLAC__fixed_compute_best_predictor_intrin_sse2()
75 last_error = _mm_sub_epi32(last_error, _mm_cvtsi32_si128(itmp)); // 0 le0 le1 le2 in FLAC__fixed_compute_best_predictor_intrin_sse2()
78 last_error = _mm_sub_epi32(last_error, _mm_cvtsi32_si128(itmp)); // le0 le1 le2 le3 in FLAC__fixed_compute_best_predictor_intrin_sse2()
86 err1 = _mm_sub_epi32(err1, last_error); in FLAC__fixed_compute_best_predictor_intrin_sse2()
88 err1 = _mm_sub_epi32(err1, last_error); in FLAC__fixed_compute_best_predictor_intrin_sse2()
90 err1 = _mm_sub_epi32(err1, last_error); in FLAC__fixed_compute_best_predictor_intrin_sse2()
92 err1 = _mm_sub_epi32(err1, last_error); // e1 e2 e3 e4 in FLAC__fixed_compute_best_predictor_intrin_sse2()
96 err1 = _mm_sub_epi32(err1, last_error); // e1 e2 e3 e4 in FLAC__fixed_compute_best_predictor_intrin_sse2()
104 err0 = _mm_sub_epi32(err0, tmp); in FLAC__fixed_compute_best_predictor_intrin_sse2()
107 err1 = _mm_sub_epi32(err1, tmp); in FLAC__fixed_compute_best_predictor_intrin_sse2()
[all …]
|
| D | fixed_intrin_ssse3.c | 72 last_error = _mm_sub_epi32(last_error, _mm_cvtsi32_si128(itmp)); // 0 0 le0 le1 in FLAC__fixed_compute_best_predictor_intrin_ssse3()
75 last_error = _mm_sub_epi32(last_error, _mm_cvtsi32_si128(itmp)); // 0 le0 le1 le2 in FLAC__fixed_compute_best_predictor_intrin_ssse3()
78 last_error = _mm_sub_epi32(last_error, _mm_cvtsi32_si128(itmp)); // le0 le1 le2 le3 in FLAC__fixed_compute_best_predictor_intrin_ssse3()
86 err1 = _mm_sub_epi32(err1, last_error); in FLAC__fixed_compute_best_predictor_intrin_ssse3()
88 err1 = _mm_sub_epi32(err1, last_error); in FLAC__fixed_compute_best_predictor_intrin_ssse3()
90 err1 = _mm_sub_epi32(err1, last_error); in FLAC__fixed_compute_best_predictor_intrin_ssse3()
92 err1 = _mm_sub_epi32(err1, last_error); // e1 e2 e3 e4 in FLAC__fixed_compute_best_predictor_intrin_ssse3()
96 err1 = _mm_sub_epi32(err1, last_error); // e1 e2 e3 e4 in FLAC__fixed_compute_best_predictor_intrin_ssse3()
163 last_error = _mm_sub_epi32(last_error, _mm_cvtsi32_si128(itmp)); // 0 0 le0 le1 in FLAC__fixed_compute_best_predictor_wide_intrin_ssse3()
166 last_error = _mm_sub_epi32(last_error, _mm_cvtsi32_si128(itmp)); // 0 le0 le1 le2 in FLAC__fixed_compute_best_predictor_wide_intrin_ssse3()
[all …]
|
| /external/libhevc/common/x86/ |
| D | ihevc_16x16_itrans_recon_sse42_intr.c | 263 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_24, m_temp_reg_30); in ihevc_itrans_recon_16x16_sse42()
282 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_25, m_temp_reg_31); in ihevc_itrans_recon_16x16_sse42()
301 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_26, m_temp_reg_30); in ihevc_itrans_recon_16x16_sse42()
318 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_27, m_temp_reg_31); in ihevc_itrans_recon_16x16_sse42()
336 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_26, m_temp_reg_30); in ihevc_itrans_recon_16x16_sse42()
352 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_27, m_temp_reg_31); in ihevc_itrans_recon_16x16_sse42()
370 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_24, m_temp_reg_30); in ihevc_itrans_recon_16x16_sse42()
385 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_25, m_temp_reg_31); in ihevc_itrans_recon_16x16_sse42()
446 m_temp_reg_46 = _mm_sub_epi32(m_temp_reg_24, m_temp_reg_20); in ihevc_itrans_recon_16x16_sse42()
451 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_40, m_temp_reg_30); in ihevc_itrans_recon_16x16_sse42()
[all …]
|
| D | ihevc_itrans_recon_16x16_ssse3_intr.c | 269 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_24, m_temp_reg_30); in ihevc_itrans_recon_16x16_ssse3()
288 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_25, m_temp_reg_31); in ihevc_itrans_recon_16x16_ssse3()
307 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_26, m_temp_reg_30); in ihevc_itrans_recon_16x16_ssse3()
324 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_27, m_temp_reg_31); in ihevc_itrans_recon_16x16_ssse3()
342 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_26, m_temp_reg_30); in ihevc_itrans_recon_16x16_ssse3()
358 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_27, m_temp_reg_31); in ihevc_itrans_recon_16x16_ssse3()
376 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_24, m_temp_reg_30); in ihevc_itrans_recon_16x16_ssse3()
391 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_25, m_temp_reg_31); in ihevc_itrans_recon_16x16_ssse3()
450 m_temp_reg_46 = _mm_sub_epi32(m_temp_reg_24, m_temp_reg_20); in ihevc_itrans_recon_16x16_ssse3()
455 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_40, m_temp_reg_30); in ihevc_itrans_recon_16x16_ssse3()
[all …]
|
| D | ihevc_itrans_recon_ssse3_intr.c | 186 m_temp_reg_14 = _mm_sub_epi32(m_temp_reg_0, m_temp_reg_2); in ihevc_itrans_recon_4x4_ttype1_ssse3()
211 m_temp_reg_12 = _mm_sub_epi32(m_temp_reg_0, m_temp_reg_3); in ihevc_itrans_recon_4x4_ttype1_ssse3()
224 m_temp_reg_23 = _mm_sub_epi32(m_temp_reg_20, m_temp_reg_10); in ihevc_itrans_recon_4x4_ttype1_ssse3()
225 m_temp_reg_30 = _mm_sub_epi32(m_temp_reg_23, m_temp_reg_21); in ihevc_itrans_recon_4x4_ttype1_ssse3()
230 m_temp_reg_23 = _mm_sub_epi32(m_temp_reg_20, m_temp_reg_11); in ihevc_itrans_recon_4x4_ttype1_ssse3()
231 m_temp_reg_31 = _mm_sub_epi32(m_temp_reg_23, m_temp_reg_21); in ihevc_itrans_recon_4x4_ttype1_ssse3()
238 m_temp_reg_23 = _mm_sub_epi32(m_temp_reg_20, m_temp_reg_11); in ihevc_itrans_recon_4x4_ttype1_ssse3()
239 m_temp_reg_32 = _mm_sub_epi32(m_temp_reg_23, m_temp_reg_21); in ihevc_itrans_recon_4x4_ttype1_ssse3()
244 m_temp_reg_23 = _mm_sub_epi32(m_temp_reg_20, m_temp_reg_12); in ihevc_itrans_recon_4x4_ttype1_ssse3()
245 m_temp_reg_33 = _mm_sub_epi32(m_temp_reg_23, m_temp_reg_21); in ihevc_itrans_recon_4x4_ttype1_ssse3()
[all …]
|
| D | ihevc_itrans_recon_sse42_intr.c | 172 m_temp_reg_14 = _mm_sub_epi32(m_temp_reg_0, m_temp_reg_2); in ihevc_itrans_recon_4x4_ttype1_sse42()
192 m_temp_reg_12 = _mm_sub_epi32(m_temp_reg_0, m_temp_reg_3); in ihevc_itrans_recon_4x4_ttype1_sse42()
227 m_temp_reg_21 = _mm_sub_epi32(m_temp_reg_33, m_temp_reg_32); in ihevc_itrans_recon_4x4_ttype1_sse42()
231 m_temp_reg_4 = _mm_sub_epi32(m_rdng_factor, m_temp_reg_13); in ihevc_itrans_recon_4x4_ttype1_sse42()
267 m_temp_reg_4 = _mm_sub_epi32(m_temp_reg_20, m_temp_reg_21); in ihevc_itrans_recon_4x4_ttype1_sse42()
287 m_temp_reg_2 = _mm_sub_epi32(m_temp_reg_20, m_temp_reg_23); in ihevc_itrans_recon_4x4_ttype1_sse42()
318 m_temp_reg_21 = _mm_sub_epi32(m_temp_reg_33, m_temp_reg_32); in ihevc_itrans_recon_4x4_ttype1_sse42()
321 m_temp_reg_4 = _mm_sub_epi32(m_rdng_factor, m_temp_reg_3); in ihevc_itrans_recon_4x4_ttype1_sse42()
520 m_temp_reg_31 = _mm_sub_epi32(m_temp_reg_10, m_temp_reg_11); in ihevc_itrans_recon_4x4_sse42()
535 m_temp_reg_33 = _mm_sub_epi32(m_temp_reg_12, m_temp_reg_13); in ihevc_itrans_recon_4x4_sse42()
[all …]
|
| D | ihevc_32x32_itrans_recon_sse42_intr.c | 363 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_40, m_temp_reg_30); in ihevc_itrans_recon_32x32_sse42()
377 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_44, m_temp_reg_30); in ihevc_itrans_recon_32x32_sse42()
390 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_41, m_temp_reg_30); in ihevc_itrans_recon_32x32_sse42()
404 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_45, m_temp_reg_30); in ihevc_itrans_recon_32x32_sse42()
418 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_42, m_temp_reg_30); in ihevc_itrans_recon_32x32_sse42()
432 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_46, m_temp_reg_30); in ihevc_itrans_recon_32x32_sse42()
449 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_43, m_temp_reg_30); in ihevc_itrans_recon_32x32_sse42()
463 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_47, m_temp_reg_30); in ihevc_itrans_recon_32x32_sse42()
478 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_43, m_temp_reg_30); in ihevc_itrans_recon_32x32_sse42()
491 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_47, m_temp_reg_30); in ihevc_itrans_recon_32x32_sse42()
[all …]
|
| D | ihevc_itrans_recon_32x32_ssse3_intr.c | 374 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_40, m_temp_reg_30); in ihevc_itrans_recon_32x32_ssse3()
388 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_44, m_temp_reg_30); in ihevc_itrans_recon_32x32_ssse3()
401 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_41, m_temp_reg_30); in ihevc_itrans_recon_32x32_ssse3()
415 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_45, m_temp_reg_30); in ihevc_itrans_recon_32x32_ssse3()
429 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_42, m_temp_reg_30); in ihevc_itrans_recon_32x32_ssse3()
443 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_46, m_temp_reg_30); in ihevc_itrans_recon_32x32_ssse3()
460 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_43, m_temp_reg_30); in ihevc_itrans_recon_32x32_ssse3()
474 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_47, m_temp_reg_30); in ihevc_itrans_recon_32x32_ssse3()
489 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_43, m_temp_reg_30); in ihevc_itrans_recon_32x32_ssse3()
502 m_temp_reg_35 = _mm_sub_epi32(m_temp_reg_47, m_temp_reg_30); in ihevc_itrans_recon_32x32_ssse3()
[all …]
|
| /external/libavc/common/x86/ |
| D | ih264_ihadamard_scaling_ssse3.c | 138 temp2 = _mm_sub_epi32(src_r1, src_r2); in ih264_ihadamard_scaling_4x4_ssse3()
139 temp3 = _mm_sub_epi32(src_r0, src_r3); in ih264_ihadamard_scaling_4x4_ssse3()
143 src_r2 = _mm_sub_epi32(temp0, temp1); in ih264_ihadamard_scaling_4x4_ssse3()
144 src_r3 = _mm_sub_epi32(temp3, temp2); in ih264_ihadamard_scaling_4x4_ssse3()
167 temp2 = _mm_sub_epi32(src_r1, src_r2); in ih264_ihadamard_scaling_4x4_ssse3()
168 temp3 = _mm_sub_epi32(src_r0, src_r3); in ih264_ihadamard_scaling_4x4_ssse3()
172 src_r2 = _mm_sub_epi32(temp0, temp1); in ih264_ihadamard_scaling_4x4_ssse3()
173 src_r3 = _mm_sub_epi32(temp3, temp2); in ih264_ihadamard_scaling_4x4_ssse3()
|
| D | ih264_ihadamard_scaling_sse42.c | 136 temp2 = _mm_sub_epi32(src_r1, src_r2); in ih264_ihadamard_scaling_4x4_sse42()
137 temp3 = _mm_sub_epi32(src_r0, src_r3); in ih264_ihadamard_scaling_4x4_sse42()
141 src_r2 = _mm_sub_epi32(temp0, temp1); in ih264_ihadamard_scaling_4x4_sse42()
142 src_r3 = _mm_sub_epi32(temp3, temp2); in ih264_ihadamard_scaling_4x4_sse42()
165 temp2 = _mm_sub_epi32(src_r1, src_r2); in ih264_ihadamard_scaling_4x4_sse42()
166 temp3 = _mm_sub_epi32(src_r0, src_r3); in ih264_ihadamard_scaling_4x4_sse42()
170 src_r2 = _mm_sub_epi32(temp0, temp1); in ih264_ihadamard_scaling_4x4_sse42()
171 src_r3 = _mm_sub_epi32(temp3, temp2); in ih264_ihadamard_scaling_4x4_sse42()
|
| D | ih264_iquant_itrans_recon_sse42.c | 196 temp1 = _mm_sub_epi32(resq_r0, resq_r2); in ih264_iquant_itrans_recon_4x4_sse42()
199 temp2 = _mm_sub_epi32(temp2, resq_r3); //(w1>>1) - w3 in ih264_iquant_itrans_recon_4x4_sse42()
209 resq_r2 = _mm_sub_epi32(temp1, temp2); in ih264_iquant_itrans_recon_4x4_sse42()
211 resq_r3 = _mm_sub_epi32(temp0, temp3); in ih264_iquant_itrans_recon_4x4_sse42()
248 temp1 = _mm_sub_epi32(resq_r0, resq_r2); in ih264_iquant_itrans_recon_4x4_sse42()
251 temp2 = _mm_sub_epi32(temp2, resq_r3); in ih264_iquant_itrans_recon_4x4_sse42()
267 temp6 = _mm_sub_epi32(temp1, temp2); in ih264_iquant_itrans_recon_4x4_sse42()
272 temp7 = _mm_sub_epi32(temp0, temp3); in ih264_iquant_itrans_recon_4x4_sse42()
446 temp1 = _mm_sub_epi32(resq_r0, resq_r2); in ih264_iquant_itrans_recon_chroma_4x4_sse42()
449 temp2 = _mm_sub_epi32(temp2, resq_r3); //(w1>>1) - w3 in ih264_iquant_itrans_recon_chroma_4x4_sse42()
[all …]
|
| /external/libmpeg2/common/x86/ |
| D | impeg2_idct_recon_sse42_intr.c | 264 m_temp_reg_44 = _mm_sub_epi32(m_temp_reg_12, m_temp_reg_16); in impeg2_idct_recon_sse42()
267 m_temp_reg_46 = _mm_sub_epi32(m_temp_reg_10, m_temp_reg_14); in impeg2_idct_recon_sse42()
295 m_temp_reg_66 = _mm_sub_epi32(m_temp_reg_40, m_temp_reg_30); in impeg2_idct_recon_sse42()
325 m_temp_reg_66 = _mm_sub_epi32(m_temp_reg_42, m_temp_reg_32); in impeg2_idct_recon_sse42()
355 m_temp_reg_66 = _mm_sub_epi32(m_temp_reg_44, m_temp_reg_34); in impeg2_idct_recon_sse42()
381 m_temp_reg_66 = _mm_sub_epi32(m_temp_reg_46, m_temp_reg_36); in impeg2_idct_recon_sse42()
472 m_temp_reg_44 = _mm_sub_epi32(m_temp_reg_12, m_temp_reg_16); in impeg2_idct_recon_sse42()
475 m_temp_reg_46 = _mm_sub_epi32(m_temp_reg_10, m_temp_reg_14); in impeg2_idct_recon_sse42()
507 m_temp_reg_66 = _mm_sub_epi32(m_temp_reg_40, m_temp_reg_30); in impeg2_idct_recon_sse42()
529 m_temp_reg_32 = _mm_sub_epi32(m_temp_reg_22, m_temp_reg_26); in impeg2_idct_recon_sse42()
[all …]
|
| /external/XNNPACK/src/qs8-requantization/ |
| D | precise-sse2.c | 58 const __m128i x_abs0123 = _mm_sub_epi32(_mm_xor_si128(x, x_neg_mask), x_neg_mask); in xnn_qs8_requantize_precise__sse2()
59 const __m128i y_abs0123 = _mm_sub_epi32(_mm_xor_si128(y, y_neg_mask), y_neg_mask); in xnn_qs8_requantize_precise__sse2()
60 const __m128i z_abs0123 = _mm_sub_epi32(_mm_xor_si128(z, z_neg_mask), z_neg_mask); in xnn_qs8_requantize_precise__sse2()
61 const __m128i w_abs0123 = _mm_sub_epi32(_mm_xor_si128(w, w_neg_mask), w_neg_mask); in xnn_qs8_requantize_precise__sse2()
101 const __m128i x_scaled = _mm_sub_epi32(_mm_xor_si128(x_abs_scaled, x_neg_mask), x_neg_mask); in xnn_qs8_requantize_precise__sse2()
102 const __m128i y_scaled = _mm_sub_epi32(_mm_xor_si128(y_abs_scaled, y_neg_mask), y_neg_mask); in xnn_qs8_requantize_precise__sse2()
103 const __m128i z_scaled = _mm_sub_epi32(_mm_xor_si128(z_abs_scaled, z_neg_mask), z_neg_mask); in xnn_qs8_requantize_precise__sse2()
104 const __m128i w_scaled = _mm_sub_epi32(_mm_xor_si128(w_abs_scaled, w_neg_mask), w_neg_mask); in xnn_qs8_requantize_precise__sse2()
|
| /external/XNNPACK/src/qu8-requantization/ |
| D | precise-sse2.c | 58 const __m128i x_abs0123 = _mm_sub_epi32(_mm_xor_si128(x, x_neg_mask), x_neg_mask); in xnn_qu8_requantize_precise__sse2()
59 const __m128i y_abs0123 = _mm_sub_epi32(_mm_xor_si128(y, y_neg_mask), y_neg_mask); in xnn_qu8_requantize_precise__sse2()
60 const __m128i z_abs0123 = _mm_sub_epi32(_mm_xor_si128(z, z_neg_mask), z_neg_mask); in xnn_qu8_requantize_precise__sse2()
61 const __m128i w_abs0123 = _mm_sub_epi32(_mm_xor_si128(w, w_neg_mask), w_neg_mask); in xnn_qu8_requantize_precise__sse2()
101 const __m128i x_scaled = _mm_sub_epi32(_mm_xor_si128(x_abs_scaled, x_neg_mask), x_neg_mask); in xnn_qu8_requantize_precise__sse2()
102 const __m128i y_scaled = _mm_sub_epi32(_mm_xor_si128(y_abs_scaled, y_neg_mask), y_neg_mask); in xnn_qu8_requantize_precise__sse2()
103 const __m128i z_scaled = _mm_sub_epi32(_mm_xor_si128(z_abs_scaled, z_neg_mask), z_neg_mask); in xnn_qu8_requantize_precise__sse2()
104 const __m128i w_scaled = _mm_sub_epi32(_mm_xor_si128(w_abs_scaled, w_neg_mask), w_neg_mask); in xnn_qu8_requantize_precise__sse2()
|
| /external/XNNPACK/src/qs8-vaddc/gen/ |
| D | minmax-sse41-mul32-ld32-x32.c | 66 …vacc0123 = _mm_sub_epi32(_mm_sra_epi32(vacc0123, vshift), _mm_cmpgt_epi32(vrem0123, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x32()
67 …vacc4567 = _mm_sub_epi32(_mm_sra_epi32(vacc4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x32()
68 …vacc89AB = _mm_sub_epi32(_mm_sra_epi32(vacc89AB, vshift), _mm_cmpgt_epi32(vrem89AB, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x32()
69 …vaccCDEF = _mm_sub_epi32(_mm_sra_epi32(vaccCDEF, vshift), _mm_cmpgt_epi32(vremCDEF, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x32()
70 …vaccGHIJ = _mm_sub_epi32(_mm_sra_epi32(vaccGHIJ, vshift), _mm_cmpgt_epi32(vremGHIJ, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x32()
71 …vaccKLMN = _mm_sub_epi32(_mm_sra_epi32(vaccKLMN, vshift), _mm_cmpgt_epi32(vremKLMN, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x32()
72 …vaccOPQR = _mm_sub_epi32(_mm_sra_epi32(vaccOPQR, vshift), _mm_cmpgt_epi32(vremOPQR, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x32()
73 …vaccSTUV = _mm_sub_epi32(_mm_sra_epi32(vaccSTUV, vshift), _mm_cmpgt_epi32(vremSTUV, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x32()
109 …vacc0123 = _mm_sub_epi32(_mm_sra_epi32(vacc0123, vshift), _mm_cmpgt_epi32(vrem0123, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x32()
110 …vacc4567 = _mm_sub_epi32(_mm_sra_epi32(vacc4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x32()
|
| D | minmax-xop-mul32-ld32-x32.c | 71 …vacc0123 = _mm_sub_epi32(_mm_sra_epi32(vacc0123, vshift), _mm_cmpgt_epi32(vrem0123, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x32()
72 …vacc4567 = _mm_sub_epi32(_mm_sra_epi32(vacc4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x32()
73 …vacc89AB = _mm_sub_epi32(_mm_sra_epi32(vacc89AB, vshift), _mm_cmpgt_epi32(vrem89AB, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x32()
74 …vaccCDEF = _mm_sub_epi32(_mm_sra_epi32(vaccCDEF, vshift), _mm_cmpgt_epi32(vremCDEF, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x32()
75 …vaccGHIJ = _mm_sub_epi32(_mm_sra_epi32(vaccGHIJ, vshift), _mm_cmpgt_epi32(vremGHIJ, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x32()
76 …vaccKLMN = _mm_sub_epi32(_mm_sra_epi32(vaccKLMN, vshift), _mm_cmpgt_epi32(vremKLMN, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x32()
77 …vaccOPQR = _mm_sub_epi32(_mm_sra_epi32(vaccOPQR, vshift), _mm_cmpgt_epi32(vremOPQR, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x32()
78 …vaccSTUV = _mm_sub_epi32(_mm_sra_epi32(vaccSTUV, vshift), _mm_cmpgt_epi32(vremSTUV, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x32()
114 …vacc0123 = _mm_sub_epi32(_mm_sra_epi32(vacc0123, vshift), _mm_cmpgt_epi32(vrem0123, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x32()
115 …vacc4567 = _mm_sub_epi32(_mm_sra_epi32(vacc4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x32()
|
| D | minmax-xop-mul32-ld32-x24.c | 65 …vacc0123 = _mm_sub_epi32(_mm_sra_epi32(vacc0123, vshift), _mm_cmpgt_epi32(vrem0123, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x24()
66 …vacc4567 = _mm_sub_epi32(_mm_sra_epi32(vacc4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x24()
67 …vacc89AB = _mm_sub_epi32(_mm_sra_epi32(vacc89AB, vshift), _mm_cmpgt_epi32(vrem89AB, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x24()
68 …vaccCDEF = _mm_sub_epi32(_mm_sra_epi32(vaccCDEF, vshift), _mm_cmpgt_epi32(vremCDEF, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x24()
69 …vaccGHIJ = _mm_sub_epi32(_mm_sra_epi32(vaccGHIJ, vshift), _mm_cmpgt_epi32(vremGHIJ, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x24()
70 …vaccKLMN = _mm_sub_epi32(_mm_sra_epi32(vaccKLMN, vshift), _mm_cmpgt_epi32(vremKLMN, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x24()
103 …vacc0123 = _mm_sub_epi32(_mm_sra_epi32(vacc0123, vshift), _mm_cmpgt_epi32(vrem0123, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x24()
104 …vacc4567 = _mm_sub_epi32(_mm_sra_epi32(vacc4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_thr… in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x24()
|