| /external/XNNPACK/src/f32-vsigmoid/gen/ |
| D | vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c | 30 const __m512 vtable_lo = _mm512_load_ps(params->avx512_rr2_lut32_p2.table_lo); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128() local
67 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
68 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
69 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
70 const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
71 const __m512 vl4 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn4), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
72 const __m512 vl5 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn5), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
73 const __m512 vl6 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn6), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
74 const __m512 vl7 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn7), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
183 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
[all …]
|
| D | vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c | 30 const __m512 vtable_lo = _mm512_load_ps(params->avx512_rr2_lut32_p2.table_lo); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112() local
64 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
65 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
66 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
67 const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
68 const __m512 vl4 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn4), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
69 const __m512 vl5 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn5), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
70 const __m512 vl6 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn6), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
168 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
200 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
|
| D | vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c | 30 const __m512 vtable_lo = _mm512_load_ps(params->avx512_rr2_lut32_p2.table_lo); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x80() local
58 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x80()
59 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x80()
60 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x80()
61 const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x80()
62 const __m512 vl4 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn4), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x80()
138 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x80()
170 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x80()
|
| D | vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c | 30 const __m512 vtable_lo = _mm512_load_ps(params->avx512_rr2_lut32_p2.table_lo); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96() local
61 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96()
62 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96()
63 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96()
64 const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96()
65 const __m512 vl4 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn4), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96()
66 const __m512 vl5 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn5), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96()
153 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96()
185 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96()
|
| D | vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c | 30 const __m512 vtable_lo = _mm512_load_ps(params->avx512_rr2_lut32_p2.table_lo); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128() local
67 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
68 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
69 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
70 const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
71 const __m512 vl4 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn4), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
72 const __m512 vl5 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn5), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
73 const __m512 vl6 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn6), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
74 const __m512 vl7 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn7), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
202 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
[all …]
|
| D | vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c | 30 const __m512 vtable_lo = _mm512_load_ps(params->avx512_rr2_lut32_p2.table_lo); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x64() local
55 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x64()
56 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x64()
57 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x64()
58 const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x64()
123 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x64()
155 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x64()
|
| D | vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c | 30 const __m512 vtable_lo = _mm512_load_ps(params->avx512_rr2_lut32_p2.table_lo); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96() local
61 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
62 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
63 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
64 const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
65 const __m512 vl4 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn4), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
66 const __m512 vl5 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn5), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
168 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
203 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96()
|
| D | vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c | 30 const __m512 vtable_lo = _mm512_load_ps(params->avx512_rr2_lut32_p2.table_lo); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112() local
64 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
65 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
66 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
67 const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
68 const __m512 vl4 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn4), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
69 const __m512 vl5 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn5), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
70 const __m512 vl6 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn6), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
185 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
220 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
|
| D | vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c | 30 const __m512 vtable_lo = _mm512_load_ps(params->avx512_rr2_lut32_p2.table_lo); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x48() local
52 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x48()
53 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x48()
54 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x48()
108 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x48()
140 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x48()
|
| D | vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c | 30 const __m512 vtable_lo = _mm512_load_ps(params->avx512_rr2_lut32_p2.table_lo); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64() local
55 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64()
56 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64()
57 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64()
58 const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64()
134 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64()
169 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64()
|
| D | vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c | 30 const __m512 vtable_lo = _mm512_load_ps(params->avx512_rr2_lut32_p2.table_lo); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80() local
58 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
59 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
60 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
61 const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
62 const __m512 vl4 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn4), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
151 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
186 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80()
|
| D | vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c | 30 const __m512 vtable_lo = _mm512_load_ps(params->avx512_rr2_lut32_p2.table_lo); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x48() local
52 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x48()
53 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x48()
54 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x48()
117 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x48()
152 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x48()
|
| D | vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c | 30 const __m512 vtable_lo = _mm512_load_ps(params->avx512_rr2_lut32_p2.table_lo); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x32() local
49 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x32()
50 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x32()
93 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x32()
125 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x32()
|
| D | vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c | 30 const __m512 vtable_lo = _mm512_load_ps(params->avx512_rr2_lut32_p2.table_lo); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x32() local
49 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x32()
50 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x32()
100 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x32()
135 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x32()
|
| D | vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c | 30 const __m512 vtable_lo = _mm512_load_ps(params->avx512_rr2_lut32_p2.table_lo); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x16() local
45 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x16()
77 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x16()
|
| D | vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c | 30 const __m512 vtable_lo = _mm512_load_ps(params->avx512_rr2_lut32_p2.table_lo); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x16() local
45 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x16()
80 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x16()
|
| /external/XNNPACK/src/math/ |
| D | exp-f32-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-f32-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-f32-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-f32-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-f32-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-f32-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-f32-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-f32-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-vsigmoid/ |
| D | avx512f-rr2-lut32-p2-perm2-scalef.c.in | 31 const __m512 vtable_lo = _mm512_load_ps(params->avx512_rr2_lut32_p2.table_lo); variable
53 …const __m512 vl${ABC[N]} = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn${ABC[N]}), vta…
111 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi);
151 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi);
|