Home
last modified time | relevance | path

Searched refs:vz6 (Results 1 – 25 of 52) sorted by relevance

123

/external/XNNPACK/src/f32-sigmoid/gen/
Davx2-rr1-p5-div-x56.c56 const __m256 vz6 = _mm256_or_ps(vx6, vsign_mask); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x56() local
64 __m256 vn6 = _mm256_fmadd_ps(vz6, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x56()
88 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2, vz6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x56()
160 vf6 = _mm256_andnot_ps(_mm256_cmp_ps(vz6, vdenorm_cutoff, _CMP_LT_OS), vf6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x56()
Davx2-rr1-p5-div-x64.c57 const __m256 vz6 = _mm256_or_ps(vx6, vsign_mask); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x64() local
66 __m256 vn6 = _mm256_fmadd_ps(vz6, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x64()
93 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2, vz6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x64()
174 vf6 = _mm256_andnot_ps(_mm256_cmp_ps(vz6, vdenorm_cutoff, _CMP_LT_OS), vf6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x64()
Davx2-rr1-p5-nr1fma-x56.c56 const __m256 vz6 = _mm256_or_ps(vx6, vsign_mask); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x56() local
64 __m256 vn6 = _mm256_fmadd_ps(vz6, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x56()
88 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2, vz6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x56()
177 vf6 = _mm256_andnot_ps(_mm256_cmp_ps(vz6, vdenorm_cutoff, _CMP_LT_OS), vf6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x56()
Davx512f-rr1-lut16-p3-perm-scalef-div-x112.c56 const __m512 vz6 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx6), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112() local
64 __m512 vn6 = _mm512_fmadd_ps(vz6, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
88 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2, vz6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
Davx512f-rr1-p5-scalef-div-x112.c53 const __m512 vz6 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx6), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x112() local
61 __m512 vn6 = _mm512_mul_ps(vz6, vlog2e); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x112()
77 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2, vz6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x112()
Davx-rr2-p5-div-x56.c57 const __m256 vz6 = _mm256_or_ps(vx6, vsign_mask); in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x56() local
65 __m256 vn6 = _mm256_add_ps(_mm256_mul_ps(vz6, vlog2e), vmagic_bias); in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x56()
103 __m256 vt6 = _mm256_add_ps(_mm256_mul_ps(vn6, vminus_ln2_hi), vz6); in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x56()
183 vf6 = _mm256_andnot_ps(_mm256_cmp_ps(vz6, vdenorm_cutoff, _CMP_LT_OS), vf6); in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x56()
Davx2-rr1-p5-nr2fma-x56.c56 const __m256 vz6 = _mm256_or_ps(vx6, vsign_mask); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x56() local
64 __m256 vn6 = _mm256_fmadd_ps(vz6, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x56()
88 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2, vz6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x56()
184 vf6 = _mm256_andnot_ps(_mm256_cmp_ps(vz6, vdenorm_cutoff, _CMP_LT_OS), vf6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x56()
Davx2-rr1-p5-div-x72.c58 const __m256 vz6 = _mm256_or_ps(vx6, vsign_mask); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x72() local
68 __m256 vn6 = _mm256_fmadd_ps(vz6, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x72()
98 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2, vz6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x72()
188 vf6 = _mm256_andnot_ps(_mm256_cmp_ps(vz6, vdenorm_cutoff, _CMP_LT_OS), vf6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x72()
Davx2-rr1-p5-nr1fma-x64.c57 const __m256 vz6 = _mm256_or_ps(vx6, vsign_mask); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x64() local
66 __m256 vn6 = _mm256_fmadd_ps(vz6, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x64()
93 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2, vz6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x64()
193 vf6 = _mm256_andnot_ps(_mm256_cmp_ps(vz6, vdenorm_cutoff, _CMP_LT_OS), vf6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x64()
Davx2-rr1-p5-div-x80.c59 const __m256 vz6 = _mm256_or_ps(vx6, vsign_mask); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x80() local
70 __m256 vn6 = _mm256_fmadd_ps(vz6, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x80()
103 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2, vz6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x80()
202 vf6 = _mm256_andnot_ps(_mm256_cmp_ps(vz6, vdenorm_cutoff, _CMP_LT_OS), vf6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x80()
Davx512f-rr2-lut32-p2-perm2-scalef-div-x112.c62 const __m512 vz6 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx6), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112() local
70 __m512 vn6 = _mm512_fmadd_ps(vz6, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
94 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vz6); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
Davx512f-rr1-p5-scalef-div-x128.c54 const __m512 vz6 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx6), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128() local
63 __m512 vn6 = _mm512_mul_ps(vz6, vlog2e); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
81 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2, vz6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
Davx-rr2-p5-div-x64.c58 const __m256 vz6 = _mm256_or_ps(vx6, vsign_mask); in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x64() local
67 __m256 vn6 = _mm256_add_ps(_mm256_mul_ps(vz6, vlog2e), vmagic_bias); in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x64()
110 __m256 vt6 = _mm256_add_ps(_mm256_mul_ps(vn6, vminus_ln2_hi), vz6); in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x64()
200 vf6 = _mm256_andnot_ps(_mm256_cmp_ps(vz6, vdenorm_cutoff, _CMP_LT_OS), vf6); in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x64()
Davx2-rr1-p5-nr2fma-x64.c57 const __m256 vz6 = _mm256_or_ps(vx6, vsign_mask); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x64() local
66 __m256 vn6 = _mm256_fmadd_ps(vz6, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x64()
93 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2, vz6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x64()
201 vf6 = _mm256_andnot_ps(_mm256_cmp_ps(vz6, vdenorm_cutoff, _CMP_LT_OS), vf6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x64()
Davx2-rr1-p5-nr2fma-x72.c58 const __m256 vz6 = _mm256_or_ps(vx6, vsign_mask); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x72() local
68 __m256 vn6 = _mm256_fmadd_ps(vz6, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x72()
98 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2, vz6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x72()
218 vf6 = _mm256_andnot_ps(_mm256_cmp_ps(vz6, vdenorm_cutoff, _CMP_LT_OS), vf6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x72()
Davx2-rr1-p5-nr1fma-x72.c58 const __m256 vz6 = _mm256_or_ps(vx6, vsign_mask); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x72() local
68 __m256 vn6 = _mm256_fmadd_ps(vz6, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x72()
98 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2, vz6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x72()
209 vf6 = _mm256_andnot_ps(_mm256_cmp_ps(vz6, vdenorm_cutoff, _CMP_LT_OS), vf6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x72()
Davx2-rr1-p5-nr1fma-x80.c59 const __m256 vz6 = _mm256_or_ps(vx6, vsign_mask); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x80() local
70 __m256 vn6 = _mm256_fmadd_ps(vz6, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x80()
103 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2, vz6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x80()
225 vf6 = _mm256_andnot_ps(_mm256_cmp_ps(vz6, vdenorm_cutoff, _CMP_LT_OS), vf6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x80()
Davx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c62 const __m512 vz6 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx6), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112() local
70 __m512 vn6 = _mm512_fmadd_ps(vz6, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
94 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vz6); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
Davx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c56 const __m512 vz6 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx6), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x112() local
64 __m512 vn6 = _mm512_fmadd_ps(vz6, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x112()
88 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2, vz6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x112()
Davx512f-rr1-p5-scalef-nr1fma-x112.c53 const __m512 vz6 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx6), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x112() local
61 __m512 vn6 = _mm512_mul_ps(vz6, vlog2e); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x112()
77 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2, vz6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x112()
Davx512f-rr2-lut32-p2-perm2-scalef-div-x128.c63 const __m512 vz6 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx6), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128() local
72 __m512 vn6 = _mm512_fmadd_ps(vz6, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
99 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vz6); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
Davx512f-rr1-lut16-p3-perm-scalef-div-x128.c57 const __m512 vz6 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx6), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x128() local
66 __m512 vn6 = _mm512_fmadd_ps(vz6, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x128()
93 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2, vz6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x128()
Davx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c57 const __m512 vz6 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx6), vsign_mask)); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128() local
66 __m512 vn6 = _mm512_fmadd_ps(vz6, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128()
93 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2, vz6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128()
/external/XNNPACK/src/f32-velu/gen/
Dvelu-avx512f-rr1-lut16-p3-perm-x112.c58 const __m512 vz6 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx6, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112() local
66 __m512 vn6 = _mm512_fmadd_ps(vz6, vlog2e, vmagic_bias); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
104 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2, vz6); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
Dvelu-avx512f-rr1-p6-x112.c58 const __m512 vz6 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx6, vprescale)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112() local
66 __m512 vn6 = _mm512_fmadd_ps(vz6, vlog2e, vmagic_bias); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
89 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2, vz6); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()

123