Home
last modified time | relevance | path

Searched refs:_mm256_cvtepi16_epi32 (Results 1 – 25 of 60) sorted by relevance

123

/aosp_15_r20/external/XNNPACK/src/qs8-dwconv/gen/
H A Dup32x25-minmax-fp32-avx2-mul16-vpmovsx.c178 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x… in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
179 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod0x89ABCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
182 …vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x… in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
183 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod0xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
193 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x… in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
194 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod1x89ABCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
197 …vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x… in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
198 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod1xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
208 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2x… in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
209 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod2x89ABCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
[all …]
H A Dup16x25-minmax-fp32-avx2-mul16-vpmovsx.c174 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x… in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx()
175 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod0x89ABCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx()
183 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x… in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx()
184 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod1x89ABCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx()
192 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2x… in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx()
193 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod2x89ABCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx()
201 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod3x… in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx()
202 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod3x89ABCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx()
210 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod4x… in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx()
211 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod4x89ABCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx()
[all …]
H A Dup32x9-minmax-fp32-avx2-mul16-vpmovsx.c98 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x… in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpmovsx()
99 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod0x89ABCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpmovsx()
102 …vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x… in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpmovsx()
103 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod0xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpmovsx()
113 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x… in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpmovsx()
114 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod1x89ABCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpmovsx()
117 …vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x… in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpmovsx()
118 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod1xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpmovsx()
128 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2x… in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpmovsx()
129 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod2x89ABCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpmovsx()
[all …]
H A Dup16x9-minmax-fp32-avx2-mul16-vpmovsx.c94 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x… in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpmovsx()
95 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod0x89ABCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpmovsx()
103 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x… in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpmovsx()
104 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod1x89ABCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpmovsx()
112 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2x… in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpmovsx()
113 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod2x89ABCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpmovsx()
121 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod3x… in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpmovsx()
122 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod3x89ABCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpmovsx()
130 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod4x… in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpmovsx()
131 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod4x89ABCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpmovsx()
[all …]
/aosp_15_r20/external/XNNPACK/src/qc8-dwconv/gen/
H A Dup32x25-minmax-fp32-avx2-mul16-vpmovsx.c178 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x… in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
179 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod0x89ABCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
182 …vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x… in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
183 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod0xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
193 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x… in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
194 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod1x89ABCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
197 …vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x… in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
198 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod1xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
208 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2x… in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
209 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod2x89ABCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx()
[all …]
H A Dup16x25-minmax-fp32-avx2-mul16-vpmovsx.c174 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x… in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx()
175 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod0x89ABCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx()
183 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x… in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx()
184 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod1x89ABCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx()
192 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2x… in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx()
193 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod2x89ABCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx()
201 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod3x… in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx()
202 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod3x89ABCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx()
210 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod4x… in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx()
211 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod4x89ABCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx()
[all …]
H A Dup32x9-minmax-fp32-avx2-mul16-vpmovsx.c98 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x… in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpmovsx()
99 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod0x89ABCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpmovsx()
102 …vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x… in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpmovsx()
103 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod0xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpmovsx()
113 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x… in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpmovsx()
114 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod1x89ABCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpmovsx()
117 …vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x… in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpmovsx()
118 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod1xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpmovsx()
128 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2x… in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpmovsx()
129 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod2x89ABCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpmovsx()
[all …]
H A Dup16x9-minmax-fp32-avx2-mul16-vpmovsx.c94 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x… in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpmovsx()
95 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod0x89ABCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpmovsx()
103 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x… in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpmovsx()
104 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod1x89ABCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpmovsx()
112 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2x… in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpmovsx()
113 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod2x89ABCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpmovsx()
121 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod3x… in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpmovsx()
122 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod3x89ABCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpmovsx()
130 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod4x… in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpmovsx()
131 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod4x89ABCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpmovsx()
[all …]
/aosp_15_r20/external/libaom/aom_dsp/x86/
H A Dhighbd_adaptive_quantize_avx2.c25 *zbin = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)zbin_ptr)); in highbd_load_b_values_avx2()
27 *round = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)round_ptr)); in highbd_load_b_values_avx2()
28 *quant = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)quant_ptr)); in highbd_load_b_values_avx2()
30 _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)dequant_ptr)); in highbd_load_b_values_avx2()
31 *shift = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)shift_ptr)); in highbd_load_b_values_avx2()
311 zbin = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)zbin_ptr)); in aom_highbd_quantize_b_32x32_adaptive_avx2()
312 round = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)round_ptr)); in aom_highbd_quantize_b_32x32_adaptive_avx2()
313 quant = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)quant_ptr)); in aom_highbd_quantize_b_32x32_adaptive_avx2()
314 dequant = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)dequant_ptr)); in aom_highbd_quantize_b_32x32_adaptive_avx2()
316 _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)quant_shift_ptr)); in aom_highbd_quantize_b_32x32_adaptive_avx2()
H A Davg_intrin_avx2.c390 src32[0] = _mm256_cvtepi16_epi32(src16[0]); in aom_highbd_hadamard_8x8_avx2()
391 src32[1] = _mm256_cvtepi16_epi32(src16[1]); in aom_highbd_hadamard_8x8_avx2()
392 src32[2] = _mm256_cvtepi16_epi32(src16[2]); in aom_highbd_hadamard_8x8_avx2()
393 src32[3] = _mm256_cvtepi16_epi32(src16[3]); in aom_highbd_hadamard_8x8_avx2()
394 src32[4] = _mm256_cvtepi16_epi32(src16[4]); in aom_highbd_hadamard_8x8_avx2()
395 src32[5] = _mm256_cvtepi16_epi32(src16[5]); in aom_highbd_hadamard_8x8_avx2()
396 src32[6] = _mm256_cvtepi16_epi32(src16[6]); in aom_highbd_hadamard_8x8_avx2()
397 src32[7] = _mm256_cvtepi16_epi32(src16[7]); in aom_highbd_hadamard_8x8_avx2()
/aosp_15_r20/external/libvpx/vpx_dsp/x86/
H A Davg_intrin_avx2.c109 src32[0] = _mm256_cvtepi16_epi32(src16[0]); in vpx_highbd_hadamard_8x8_avx2()
110 src32[1] = _mm256_cvtepi16_epi32(src16[1]); in vpx_highbd_hadamard_8x8_avx2()
111 src32[2] = _mm256_cvtepi16_epi32(src16[2]); in vpx_highbd_hadamard_8x8_avx2()
112 src32[3] = _mm256_cvtepi16_epi32(src16[3]); in vpx_highbd_hadamard_8x8_avx2()
113 src32[4] = _mm256_cvtepi16_epi32(src16[4]); in vpx_highbd_hadamard_8x8_avx2()
114 src32[5] = _mm256_cvtepi16_epi32(src16[5]); in vpx_highbd_hadamard_8x8_avx2()
115 src32[6] = _mm256_cvtepi16_epi32(src16[6]); in vpx_highbd_hadamard_8x8_avx2()
116 src32[7] = _mm256_cvtepi16_epi32(src16[7]); in vpx_highbd_hadamard_8x8_avx2()
/aosp_15_r20/external/tensorflow/tensorflow/lite/kernels/internal/optimized/integer_ops/
H A Dsub.h58 _mm256_cvtepi16_epi32(_mm256_castsi256_si128(input1_val_original)); in SubElementwiseInt16()
60 _mm256_cvtepi16_epi32(_mm256_extracti128_si256(input1_val_original, 1)); in SubElementwiseInt16()
62 _mm256_cvtepi16_epi32(_mm256_castsi256_si128(input2_val_original)); in SubElementwiseInt16()
64 _mm256_cvtepi16_epi32(_mm256_extracti128_si256(input2_val_original, 1)); in SubElementwiseInt16()
H A Dleaky_relu.h54 __m256i input_low = _mm256_cvtepi16_epi32(_mm256_castsi256_si128(input)); in QuantizeLeakyRelu()
56 _mm256_cvtepi16_epi32(_mm256_extracti128_si256(input, 1)); in QuantizeLeakyRelu()
H A Dadd.h196 _mm256_cvtepi16_epi32(_mm256_castsi256_si128(input1_val_original)); in AddElementwiseInt16()
198 _mm256_cvtepi16_epi32(_mm256_extracti128_si256(input1_val_original, 1)); in AddElementwiseInt16()
200 _mm256_cvtepi16_epi32(_mm256_castsi256_si128(input2_val_original)); in AddElementwiseInt16()
202 _mm256_cvtepi16_epi32(_mm256_extracti128_si256(input2_val_original, 1)); in AddElementwiseInt16()
/aosp_15_r20/external/libopenapv/src/avx/
H A Doapv_tq_avx.c351 __m256i coef_row_ext = _mm256_cvtepi16_epi32(coef_row); in oapv_quant_avx()
408 __m256i cur_q_matrix = _mm256_cvtepi16_epi32(_mm_loadu_si128((__m128i*)(q_matrix + i))); in oapv_dquant_avx()
409 __m256i coef_8_val_act = _mm256_cvtepi16_epi32(_mm_loadu_si128((__m128i*)(coef + i))); in oapv_dquant_avx()
431 __m256i cur_q_matrix = _mm256_cvtepi16_epi32(_mm_loadu_si128((__m128i*)(q_matrix + i))); in oapv_dquant_avx()
432 __m256i coef_8_val_act = _mm256_cvtepi16_epi32(_mm_loadu_si128((__m128i*)(coef + i))); in oapv_dquant_avx()
/aosp_15_r20/external/XNNPACK/src/qs8-dwconv/
H A Dunipass-avx2-mul16-vpmovsx.c.in66 …vacc${ABC[C:C+8]} = _mm256_add_epi32(vacc${ABC[C:C+8]}, _mm256_cvtepi16_epi32(_mm256_castsi256_si1…
67 …vacc${ABC[C+8:C+16]} = _mm256_add_epi32(vacc${ABC[C+8:C+16]}, _mm256_cvtepi16_epi32(vprod${K}x${AB…
131 …vacc${ABC[0:8]} = _mm256_add_epi32(vacc${ABC[0:8]}, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(v…
132 …vacc${ABC[8:16]} = _mm256_add_epi32(vacc${ABC[8:16]}, _mm256_cvtepi16_epi32(vprod${K}x${ABC[8:16]}…
/aosp_15_r20/external/libopus/silk/x86/
H A DNSQ_del_dec_avx2.c1044 __m256i in_v = _mm256_cvtepi16_epi32(_mm_loadu_si128((__m128i_u*)&in_ptr[-8])); in silk_LPC_analysis_filter_avx2()
1045 __m256i B_v = _mm256_cvtepi16_epi32(_mm_loadu_si128((__m128i_u*)& B[0])); in silk_LPC_analysis_filter_avx2()
1049 in_v = _mm256_cvtepi16_epi32(_mm_loadu_si128((__m128i_u*)&in_ptr[-16])); in silk_LPC_analysis_filter_avx2()
1050 B_v = _mm256_cvtepi16_epi32(_mm_loadu_si128((__m128i_u*)&B [8])); in silk_LPC_analysis_filter_avx2()
1055 in_v = _mm256_cvtepi16_epi32(_mm_loadu_si32(&in_ptr[-10])); in silk_LPC_analysis_filter_avx2()
1056 B_v = _mm256_cvtepi16_epi32(_mm_loadu_si32(&B [8])); in silk_LPC_analysis_filter_avx2()
/aosp_15_r20/external/libaom/av1/common/x86/
H A Dresize_avx2.c226 const __m256i dst_00 = _mm256_cvtepi16_epi32(CAST_LOW(dst_0)); in resize_convolve()
228 _mm256_cvtepi16_epi32(_mm256_extracti128_si256(dst_0, 1)); in resize_convolve()
229 const __m256i dst_10 = _mm256_cvtepi16_epi32(CAST_LOW(dst_1)); in resize_convolve()
231 _mm256_cvtepi16_epi32(_mm256_extracti128_si256(dst_1, 1)); in resize_convolve()
H A Dhighbd_warp_affine_avx2.c102 tmp[k + 7] = _mm256_cvtepi16_epi32(_mm_set1_epi16( in av1_highbd_warp_affine_avx2()
113 tmp[k + 7] = _mm256_cvtepi16_epi32( in av1_highbd_warp_affine_avx2()
/aosp_15_r20/external/libaom/av1/encoder/x86/
H A Dhighbd_fwd_txfm_avx2.c48 out[0] = _mm256_cvtepi16_epi32(out1[0]); in load_buffer_8x8_avx2()
49 out[1] = _mm256_cvtepi16_epi32(out1[1]); in load_buffer_8x8_avx2()
50 out[2] = _mm256_cvtepi16_epi32(out1[2]); in load_buffer_8x8_avx2()
51 out[3] = _mm256_cvtepi16_epi32(out1[3]); in load_buffer_8x8_avx2()
52 out[4] = _mm256_cvtepi16_epi32(out1[4]); in load_buffer_8x8_avx2()
53 out[5] = _mm256_cvtepi16_epi32(out1[5]); in load_buffer_8x8_avx2()
54 out[6] = _mm256_cvtepi16_epi32(out1[6]); in load_buffer_8x8_avx2()
55 out[7] = _mm256_cvtepi16_epi32(out1[7]); in load_buffer_8x8_avx2()
58 out[0] = _mm256_cvtepi16_epi32(mm_reverse_epi16(out1[0])); in load_buffer_8x8_avx2()
59 out[1] = _mm256_cvtepi16_epi32(mm_reverse_epi16(out1[1])); in load_buffer_8x8_avx2()
[all …]
/aosp_15_r20/prebuilts/clang/host/linux-x86/clang-r522817/lib/clang/18/include/
Davx512vlbf16intrin.h438 (__m256i)_mm256_cvtepi16_epi32((__m128i)__A), 16)); in _mm256_cvtpbh_ps()
509 (__m256i)__S, (__mmask8)__U, (__m256i)_mm256_cvtepi16_epi32((__m128i)__A), in _mm256_mask_cvtpbh_ps()
/aosp_15_r20/prebuilts/clang/host/linux-x86/clang-r536225/lib/clang/19/include/
Davx512vlbf16intrin.h438 (__m256i)_mm256_cvtepi16_epi32((__m128i)__A), 16)); in _mm256_cvtpbh_ps()
509 (__m256i)__S, (__mmask8)__U, (__m256i)_mm256_cvtepi16_epi32((__m128i)__A), in _mm256_mask_cvtpbh_ps()
/aosp_15_r20/prebuilts/clang-tools/linux-x86/lib64/clang/19/include/
H A Davx512vlbf16intrin.h438 (__m256i)_mm256_cvtepi16_epi32((__m128i)__A), 16)); in _mm256_cvtpbh_ps()
509 (__m256i)__S, (__mmask8)__U, (__m256i)_mm256_cvtepi16_epi32((__m128i)__A), in _mm256_mask_cvtpbh_ps()
/aosp_15_r20/prebuilts/clang/host/linux-x86/clang-r530567b/lib/clang/19/include/
Davx512vlbf16intrin.h438 (__m256i)_mm256_cvtepi16_epi32((__m128i)__A), 16)); in _mm256_cvtpbh_ps()
509 (__m256i)__S, (__mmask8)__U, (__m256i)_mm256_cvtepi16_epi32((__m128i)__A), in _mm256_mask_cvtpbh_ps()
/aosp_15_r20/prebuilts/clang/host/linux-x86/clang-r530567/lib/clang/19/include/
Davx512vlbf16intrin.h438 (__m256i)_mm256_cvtepi16_epi32((__m128i)__A), 16)); in _mm256_cvtpbh_ps()
509 (__m256i)__S, (__mmask8)__U, (__m256i)_mm256_cvtepi16_epi32((__m128i)__A), in _mm256_mask_cvtpbh_ps()

123