| /external/XNNPACK/src/f32-sigmoid/gen/ |
| D | avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c | 35 const __m512 vtable_lo = _mm512_set_ps( in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128() local
75 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
76 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
77 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
78 const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
79 const __m512 vl4 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn4), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
80 const __m512 vl5 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn5), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
81 const __m512 vl6 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn6), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
82 const __m512 vl7 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn7), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
191 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
[all …]
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c | 35 const __m512 vtable_lo = _mm512_set_ps( in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96() local
69 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96()
70 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96()
71 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96()
72 const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96()
73 const __m512 vl4 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn4), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96()
74 const __m512 vl5 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn5), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96()
161 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96()
193 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96()
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c | 35 const __m512 vtable_lo = _mm512_set_ps( in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112() local
72 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
73 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
74 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
75 const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
76 const __m512 vl4 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn4), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
77 const __m512 vl5 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn5), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
78 const __m512 vl6 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn6), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
176 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
208 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c | 35 const __m512 vtable_lo = _mm512_set_ps( in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x80() local
66 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x80()
67 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x80()
68 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x80()
69 const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x80()
70 const __m512 vl4 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn4), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x80()
146 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x80()
178 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x80()
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c | 35 const __m512 vtable_lo = _mm512_set_ps( in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112() local
72 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
73 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
74 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
75 const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
76 const __m512 vl4 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn4), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
77 const __m512 vl5 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn5), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
78 const __m512 vl6 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn6), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
193 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
228 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c | 35 const __m512 vtable_lo = _mm512_set_ps( in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x64() local
63 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x64()
64 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x64()
65 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x64()
66 const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x64()
131 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x64()
163 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x64()
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c | 35 const __m512 vtable_lo = _mm512_set_ps( in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128() local
75 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
76 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
77 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
78 const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
79 const __m512 vl4 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn4), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
80 const __m512 vl5 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn5), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
81 const __m512 vl6 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn6), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
82 const __m512 vl7 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn7), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
210 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
[all …]
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c | 35 const __m512 vtable_lo = _mm512_set_ps( in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x48() local
60 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x48()
61 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x48()
62 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x48()
116 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x48()
148 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x48()
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c | 35 const __m512 vtable_lo = _mm512_set_ps( in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80() local
66 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
67 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
68 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
69 const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
70 const __m512 vl4 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn4), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
159 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
194 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c | 35 const __m512 vtable_lo = _mm512_set_ps( in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96() local
69 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
70 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
71 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
72 const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
73 const __m512 vl4 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn4), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
74 const __m512 vl5 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn5), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
176 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
211 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c | 35 const __m512 vtable_lo = _mm512_set_ps( in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64() local
63 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64()
64 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64()
65 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64()
66 const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64()
142 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64()
177 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64()
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c | 35 const __m512 vtable_lo = _mm512_set_ps( in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x32() local
57 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x32()
58 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x32()
101 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x32()
133 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x32()
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c | 35 const __m512 vtable_lo = _mm512_set_ps( in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x48() local
60 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x48()
61 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x48()
62 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x48()
125 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x48()
160 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x48()
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c | 35 const __m512 vtable_lo = _mm512_set_ps( in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x32() local
57 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x32()
58 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x32()
108 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x32()
143 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x32()
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c | 35 const __m512 vtable_lo = _mm512_set_ps( in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x16() local
53 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x16()
85 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x16()
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c | 35 const __m512 vtable_lo = _mm512_set_ps( in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x16() local
53 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x16()
88 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x16()
|
| /external/XNNPACK/src/math/ |
| D | exp-avx512f-rr2-lut32-p2-perm2-scalef.c | 33 const __m512 vtable_lo = _mm512_set_ps( in xnn_math_f32_exp__avx512f_rr2_lut32_p2_perm2_scalef() local
51 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_math_f32_exp__avx512f_rr2_lut32_p2_perm2_scalef()
|
| D | sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c | 32 const __m512 vtable_lo = _mm512_set_ps( in xnn_math_f32_sigmoid__avx512f_rr1_lut32_p2_perm2_scalef_div() local
66 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_math_f32_sigmoid__avx512f_rr1_lut32_p2_perm2_scalef_div()
|
| D | sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c | 32 const __m512 vtable_lo = _mm512_set_ps( in xnn_math_f32_sigmoid__avx512f_rr1_lut32_p2_perm2_scalef_nr1fma() local
66 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_math_f32_sigmoid__avx512f_rr1_lut32_p2_perm2_scalef_nr1fma()
|
| D | sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c | 32 const __m512 vtable_lo = _mm512_set_ps( in xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_div() local
67 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_div()
|
| D | sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c | 32 const __m512 vtable_lo = _mm512_set_ps( in xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma() local
67 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma()
|
| D | sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c | 32 const __m512 vtable_lo = _mm512_set_ps( in xnn_math_f32_sigmoid__avx512f_rr1_lut32_p2_perm2_scalef_nr1fma1adj() local
66 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_math_f32_sigmoid__avx512f_rr1_lut32_p2_perm2_scalef_nr1fma1adj()
|
| D | sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c | 32 const __m512 vtable_lo = _mm512_set_ps( in xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma1adj() local
67 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma1adj()
|
| D | exp-avx512f-rr2-lut32-p2-perm2.c | 38 const __m512 vtable_lo = _mm512_set_ps( in xnn_math_f32_exp__avx512f_rr2_lut32_p2_perm2() local
79 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_math_f32_exp__avx512f_rr2_lut32_p2_perm2()
|
| /external/XNNPACK/src/f32-sigmoid/ |
| D | avx512f-lut32-p2-perm2-scalef.c.in | 37 const __m512 vtable_lo = _mm512_set_ps( variable
65 …const __m512 vl${ABC[N]} = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn${ABC[N]}), vta…
127 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi);
170 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi);
|