Home
last modified time | relevance | path

Searched refs:_mm512_sub_ps (Results 1 – 25 of 145) sorted by relevance

123456

/external/XNNPACK/src/f32-raddextexp/gen/
Davx512f-p5-scalef-x192-acc6.c199 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
200 const __m512 vdelta_acce1 = _mm512_sub_ps(vacce1, vmax_e1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
201 const __m512 vdelta_acce2 = _mm512_sub_ps(vacce2, vmax_e2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
202 const __m512 vdelta_acce3 = _mm512_sub_ps(vacce3, vmax_e3); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
203 const __m512 vdelta_acce4 = _mm512_sub_ps(vacce4, vmax_e4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
204 const __m512 vdelta_acce5 = _mm512_sub_ps(vacce5, vmax_e5); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
205 const __m512 vdelta_e0 = _mm512_sub_ps(vn0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
206 const __m512 vdelta_e1 = _mm512_sub_ps(vn1, vmax_e1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
207 const __m512 vdelta_e2 = _mm512_sub_ps(vn2, vmax_e2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
208 const __m512 vdelta_e3 = _mm512_sub_ps(vn3, vmax_e3); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
[all …]
Davx512f-p5-scalef-x160-acc5.c177 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5()
178 const __m512 vdelta_acce1 = _mm512_sub_ps(vacce1, vmax_e1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5()
179 const __m512 vdelta_acce2 = _mm512_sub_ps(vacce2, vmax_e2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5()
180 const __m512 vdelta_acce3 = _mm512_sub_ps(vacce3, vmax_e3); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5()
181 const __m512 vdelta_acce4 = _mm512_sub_ps(vacce4, vmax_e4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5()
182 const __m512 vdelta_e0 = _mm512_sub_ps(vn0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5()
183 const __m512 vdelta_e1 = _mm512_sub_ps(vn1, vmax_e1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5()
184 const __m512 vdelta_e2 = _mm512_sub_ps(vn2, vmax_e2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5()
185 const __m512 vdelta_e3 = _mm512_sub_ps(vn3, vmax_e3); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5()
186 const __m512 vdelta_e4 = _mm512_sub_ps(vn4, vmax_e4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5()
[all …]
Davx512f-p5-scalef-x128-acc4.c155 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4()
156 const __m512 vdelta_acce1 = _mm512_sub_ps(vacce1, vmax_e1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4()
157 const __m512 vdelta_acce2 = _mm512_sub_ps(vacce2, vmax_e2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4()
158 const __m512 vdelta_acce3 = _mm512_sub_ps(vacce3, vmax_e3); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4()
159 const __m512 vdelta_e0 = _mm512_sub_ps(vn0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4()
160 const __m512 vdelta_e1 = _mm512_sub_ps(vn1, vmax_e1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4()
161 const __m512 vdelta_e2 = _mm512_sub_ps(vn2, vmax_e2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4()
162 const __m512 vdelta_e3 = _mm512_sub_ps(vn3, vmax_e3); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4()
163 const __m512 vdelta_e4 = _mm512_sub_ps(vn4, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4()
164 const __m512 vdelta_e5 = _mm512_sub_ps(vn5, vmax_e1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4()
[all …]
Davx512f-p5-scalef-x144-acc3.c163 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3()
164 const __m512 vdelta_acce1 = _mm512_sub_ps(vacce1, vmax_e1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3()
165 const __m512 vdelta_acce2 = _mm512_sub_ps(vacce2, vmax_e2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3()
166 const __m512 vdelta_e0 = _mm512_sub_ps(vn0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3()
167 const __m512 vdelta_e1 = _mm512_sub_ps(vn1, vmax_e1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3()
168 const __m512 vdelta_e2 = _mm512_sub_ps(vn2, vmax_e2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3()
169 const __m512 vdelta_e3 = _mm512_sub_ps(vn3, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3()
170 const __m512 vdelta_e4 = _mm512_sub_ps(vn4, vmax_e1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3()
171 const __m512 vdelta_e5 = _mm512_sub_ps(vn5, vmax_e2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3()
172 const __m512 vdelta_e6 = _mm512_sub_ps(vn6, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3()
[all …]
Davx512f-p5-scalef-x192-acc3.c193 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3()
194 const __m512 vdelta_acce1 = _mm512_sub_ps(vacce1, vmax_e1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3()
195 const __m512 vdelta_acce2 = _mm512_sub_ps(vacce2, vmax_e2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3()
196 const __m512 vdelta_e0 = _mm512_sub_ps(vn0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3()
197 const __m512 vdelta_e1 = _mm512_sub_ps(vn1, vmax_e1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3()
198 const __m512 vdelta_e2 = _mm512_sub_ps(vn2, vmax_e2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3()
199 const __m512 vdelta_e3 = _mm512_sub_ps(vn3, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3()
200 const __m512 vdelta_e4 = _mm512_sub_ps(vn4, vmax_e1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3()
201 const __m512 vdelta_e5 = _mm512_sub_ps(vn5, vmax_e2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3()
202 const __m512 vdelta_e6 = _mm512_sub_ps(vn6, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3()
[all …]
Davx512f-p5-scalef-x128-acc2.c151 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2()
152 const __m512 vdelta_acce1 = _mm512_sub_ps(vacce1, vmax_e1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2()
153 const __m512 vdelta_e0 = _mm512_sub_ps(vn0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2()
154 const __m512 vdelta_e1 = _mm512_sub_ps(vn1, vmax_e1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2()
155 const __m512 vdelta_e2 = _mm512_sub_ps(vn2, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2()
156 const __m512 vdelta_e3 = _mm512_sub_ps(vn3, vmax_e1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2()
157 const __m512 vdelta_e4 = _mm512_sub_ps(vn4, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2()
158 const __m512 vdelta_e5 = _mm512_sub_ps(vn5, vmax_e1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2()
159 const __m512 vdelta_e6 = _mm512_sub_ps(vn6, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2()
160 const __m512 vdelta_e7 = _mm512_sub_ps(vn7, vmax_e1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2()
[all …]
Davx512f-p5-scalef-x160-acc2.c171 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2()
172 const __m512 vdelta_acce1 = _mm512_sub_ps(vacce1, vmax_e1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2()
173 const __m512 vdelta_e0 = _mm512_sub_ps(vn0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2()
174 const __m512 vdelta_e1 = _mm512_sub_ps(vn1, vmax_e1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2()
175 const __m512 vdelta_e2 = _mm512_sub_ps(vn2, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2()
176 const __m512 vdelta_e3 = _mm512_sub_ps(vn3, vmax_e1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2()
177 const __m512 vdelta_e4 = _mm512_sub_ps(vn4, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2()
178 const __m512 vdelta_e5 = _mm512_sub_ps(vn5, vmax_e1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2()
179 const __m512 vdelta_e6 = _mm512_sub_ps(vn6, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2()
180 const __m512 vdelta_e7 = _mm512_sub_ps(vn7, vmax_e1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2()
[all …]
Davx512f-p5-scalef-x192-acc2.c191 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
192 const __m512 vdelta_acce1 = _mm512_sub_ps(vacce1, vmax_e1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
193 const __m512 vdelta_e0 = _mm512_sub_ps(vn0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
194 const __m512 vdelta_e1 = _mm512_sub_ps(vn1, vmax_e1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
195 const __m512 vdelta_e2 = _mm512_sub_ps(vn2, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
196 const __m512 vdelta_e3 = _mm512_sub_ps(vn3, vmax_e1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
197 const __m512 vdelta_e4 = _mm512_sub_ps(vn4, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
198 const __m512 vdelta_e5 = _mm512_sub_ps(vn5, vmax_e1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
199 const __m512 vdelta_e6 = _mm512_sub_ps(vn6, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
200 const __m512 vdelta_e7 = _mm512_sub_ps(vn7, vmax_e1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
[all …]
Davx512f-p5-scalef-x192.c189 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
190 const __m512 vdelta_e0 = _mm512_sub_ps(vn0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
191 const __m512 vdelta_e1 = _mm512_sub_ps(vn1, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
192 const __m512 vdelta_e2 = _mm512_sub_ps(vn2, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
193 const __m512 vdelta_e3 = _mm512_sub_ps(vn3, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
194 const __m512 vdelta_e4 = _mm512_sub_ps(vn4, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
195 const __m512 vdelta_e5 = _mm512_sub_ps(vn5, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
196 const __m512 vdelta_e6 = _mm512_sub_ps(vn6, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
197 const __m512 vdelta_e7 = _mm512_sub_ps(vn7, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
198 const __m512 vdelta_e8 = _mm512_sub_ps(vn8, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
[all …]
Davx512f-p5-scalef-x160.c169 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160()
170 const __m512 vdelta_e0 = _mm512_sub_ps(vn0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160()
171 const __m512 vdelta_e1 = _mm512_sub_ps(vn1, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160()
172 const __m512 vdelta_e2 = _mm512_sub_ps(vn2, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160()
173 const __m512 vdelta_e3 = _mm512_sub_ps(vn3, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160()
174 const __m512 vdelta_e4 = _mm512_sub_ps(vn4, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160()
175 const __m512 vdelta_e5 = _mm512_sub_ps(vn5, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160()
176 const __m512 vdelta_e6 = _mm512_sub_ps(vn6, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160()
177 const __m512 vdelta_e7 = _mm512_sub_ps(vn7, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160()
178 const __m512 vdelta_e8 = _mm512_sub_ps(vn8, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160()
[all …]
Davx512f-p5-scalef-x144.c159 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144()
160 const __m512 vdelta_e0 = _mm512_sub_ps(vn0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144()
161 const __m512 vdelta_e1 = _mm512_sub_ps(vn1, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144()
162 const __m512 vdelta_e2 = _mm512_sub_ps(vn2, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144()
163 const __m512 vdelta_e3 = _mm512_sub_ps(vn3, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144()
164 const __m512 vdelta_e4 = _mm512_sub_ps(vn4, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144()
165 const __m512 vdelta_e5 = _mm512_sub_ps(vn5, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144()
166 const __m512 vdelta_e6 = _mm512_sub_ps(vn6, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144()
167 const __m512 vdelta_e7 = _mm512_sub_ps(vn7, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144()
168 const __m512 vdelta_e8 = _mm512_sub_ps(vn8, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144()
[all …]
Davx512f-p5-scalef-x128.c149 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128()
150 const __m512 vdelta_e0 = _mm512_sub_ps(vn0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128()
151 const __m512 vdelta_e1 = _mm512_sub_ps(vn1, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128()
152 const __m512 vdelta_e2 = _mm512_sub_ps(vn2, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128()
153 const __m512 vdelta_e3 = _mm512_sub_ps(vn3, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128()
154 const __m512 vdelta_e4 = _mm512_sub_ps(vn4, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128()
155 const __m512 vdelta_e5 = _mm512_sub_ps(vn5, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128()
156 const __m512 vdelta_e6 = _mm512_sub_ps(vn6, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128()
157 const __m512 vdelta_e7 = _mm512_sub_ps(vn7, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128()
199 const __m512 vdelta_acce = _mm512_sub_ps(vacce, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128()
[all …]
/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/
Davx512f-p5-scalef-x192.c58 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
59 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
60 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
61 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
62 const __m512 vx4 = _mm512_sub_ps(vi4, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
63 const __m512 vx5 = _mm512_sub_ps(vi5, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
64 const __m512 vx6 = _mm512_sub_ps(vi6, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
65 const __m512 vx7 = _mm512_sub_ps(vi7, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
66 const __m512 vx8 = _mm512_sub_ps(vi8, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
67 const __m512 vx9 = _mm512_sub_ps(vi9, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
[all …]
Davx512f-p5-scalef-x192-acc2.c59 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
60 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
61 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
62 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
63 const __m512 vx4 = _mm512_sub_ps(vi4, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
64 const __m512 vx5 = _mm512_sub_ps(vi5, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
65 const __m512 vx6 = _mm512_sub_ps(vi6, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
66 const __m512 vx7 = _mm512_sub_ps(vi7, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
67 const __m512 vx8 = _mm512_sub_ps(vi8, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
68 const __m512 vx9 = _mm512_sub_ps(vi9, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
[all …]
Davx512f-p5-scalef-x192-acc3.c60 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
61 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
62 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
63 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
64 const __m512 vx4 = _mm512_sub_ps(vi4, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
65 const __m512 vx5 = _mm512_sub_ps(vi5, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
66 const __m512 vx6 = _mm512_sub_ps(vi6, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
67 const __m512 vx7 = _mm512_sub_ps(vi7, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
68 const __m512 vx8 = _mm512_sub_ps(vi8, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
69 const __m512 vx9 = _mm512_sub_ps(vi9, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
[all …]
Davx512f-p5-scalef-x160.c56 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
57 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
58 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
59 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
60 const __m512 vx4 = _mm512_sub_ps(vi4, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
61 const __m512 vx5 = _mm512_sub_ps(vi5, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
62 const __m512 vx6 = _mm512_sub_ps(vi6, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
63 const __m512 vx7 = _mm512_sub_ps(vi7, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
64 const __m512 vx8 = _mm512_sub_ps(vi8, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
65 const __m512 vx9 = _mm512_sub_ps(vi9, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
[all …]
/external/XNNPACK/src/f32-raddexpminusmax/gen/
Davx512f-p5-scalef-x192.c57 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
58 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
59 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
60 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
61 const __m512 vx4 = _mm512_sub_ps(vi4, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
62 const __m512 vx5 = _mm512_sub_ps(vi5, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
63 const __m512 vx6 = _mm512_sub_ps(vi6, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
64 const __m512 vx7 = _mm512_sub_ps(vi7, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
65 const __m512 vx8 = _mm512_sub_ps(vi8, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
66 const __m512 vx9 = _mm512_sub_ps(vi9, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
[all …]
Davx512f-p5-scalef-x192-acc3.c59 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
60 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
61 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
62 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
63 const __m512 vx4 = _mm512_sub_ps(vi4, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
64 const __m512 vx5 = _mm512_sub_ps(vi5, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
65 const __m512 vx6 = _mm512_sub_ps(vi6, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
66 const __m512 vx7 = _mm512_sub_ps(vi7, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
67 const __m512 vx8 = _mm512_sub_ps(vi8, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
68 const __m512 vx9 = _mm512_sub_ps(vi9, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
[all …]
Davx512f-p5-scalef-x192-acc6.c62 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
63 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
64 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
65 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
66 const __m512 vx4 = _mm512_sub_ps(vi4, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
67 const __m512 vx5 = _mm512_sub_ps(vi5, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
68 const __m512 vx6 = _mm512_sub_ps(vi6, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
69 const __m512 vx7 = _mm512_sub_ps(vi7, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
70 const __m512 vx8 = _mm512_sub_ps(vi8, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
71 const __m512 vx9 = _mm512_sub_ps(vi9, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
[all …]
Davx512f-p5-scalef-x192-acc2.c58 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
59 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
60 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
61 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
62 const __m512 vx4 = _mm512_sub_ps(vi4, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
63 const __m512 vx5 = _mm512_sub_ps(vi5, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
64 const __m512 vx6 = _mm512_sub_ps(vi6, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
65 const __m512 vx7 = _mm512_sub_ps(vi7, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
66 const __m512 vx8 = _mm512_sub_ps(vi8, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
67 const __m512 vx9 = _mm512_sub_ps(vi9, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
[all …]
Davx512f-p5-scalef-x160.c55 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
56 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
57 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
58 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
59 const __m512 vx4 = _mm512_sub_ps(vi4, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
60 const __m512 vx5 = _mm512_sub_ps(vi5, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
61 const __m512 vx6 = _mm512_sub_ps(vi6, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
62 const __m512 vx7 = _mm512_sub_ps(vi7, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
63 const __m512 vx8 = _mm512_sub_ps(vi8, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
64 const __m512 vx9 = _mm512_sub_ps(vi9, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
[all …]
Davx512f-p5-scalef-x160-acc2.c56 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
57 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
58 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
59 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
60 const __m512 vx4 = _mm512_sub_ps(vi4, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
61 const __m512 vx5 = _mm512_sub_ps(vi5, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
62 const __m512 vx6 = _mm512_sub_ps(vi6, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
63 const __m512 vx7 = _mm512_sub_ps(vi7, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
64 const __m512 vx8 = _mm512_sub_ps(vi8, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
65 const __m512 vx9 = _mm512_sub_ps(vi9, vi_max); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
[all …]
/external/XNNPACK/src/f32-vscaleexpminusmax/gen/
Davx512f-p5-scalef-x192.c58 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
59 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
60 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
61 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
62 const __m512 vx4 = _mm512_sub_ps(vi4, vi_max); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
63 const __m512 vx5 = _mm512_sub_ps(vi5, vi_max); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
64 const __m512 vx6 = _mm512_sub_ps(vi6, vi_max); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
65 const __m512 vx7 = _mm512_sub_ps(vi7, vi_max); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
66 const __m512 vx8 = _mm512_sub_ps(vi8, vi_max); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
67 const __m512 vx9 = _mm512_sub_ps(vi9, vi_max); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
[all …]
Davx512f-p5-scalef-x176.c57 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
58 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
59 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
60 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
61 const __m512 vx4 = _mm512_sub_ps(vi4, vi_max); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
62 const __m512 vx5 = _mm512_sub_ps(vi5, vi_max); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
63 const __m512 vx6 = _mm512_sub_ps(vi6, vi_max); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
64 const __m512 vx7 = _mm512_sub_ps(vi7, vi_max); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
65 const __m512 vx8 = _mm512_sub_ps(vi8, vi_max); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
66 const __m512 vx9 = _mm512_sub_ps(vi9, vi_max); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
[all …]
/external/XNNPACK/src/f32-vbinary/gen/
Dvsqrdiffc-avx512f-x32.c39 __m512 vy0123456789ABCDEF = _mm512_sub_ps(va0123456789ABCDEF, vb); in xnn_f32_vsqrdiffc_ukernel__avx512f_x32()
40 __m512 vyGHIJKLMNOPQRSTUV = _mm512_sub_ps(vaGHIJKLMNOPQRSTUV, vb); in xnn_f32_vsqrdiffc_ukernel__avx512f_x32()
54 __m512 vy = _mm512_sub_ps(va, vb); in xnn_f32_vsqrdiffc_ukernel__avx512f_x32()
68 __m512 vy = _mm512_sub_ps(va, vb); in xnn_f32_vsqrdiffc_ukernel__avx512f_x32()

123456