Home
last modified time | relevance | path

Searched refs:vc0 (Results 1 – 25 of 98) sorted by relevance

1234

/external/XNNPACK/src/f32-raddexpminusmax/gen/
Davx512f-p5-scalef-x192-acc2.c30 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2() local
166 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
167 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
168 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
169 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
170 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
171 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
172 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
173 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
174 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
[all …]
Davx512f-p5-scalef-x192.c30 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192() local
165 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
166 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
167 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
168 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
169 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
170 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
171 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
172 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
173 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
[all …]
Davx512f-p5-scalef-x192-acc3.c30 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3() local
167 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
168 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
169 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
170 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
171 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
172 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
173 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
174 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
175 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
[all …]
Davx512f-p5-scalef-x192-acc6.c30 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6() local
170 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
171 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
172 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
173 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
174 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
175 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
176 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
177 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
178 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
[all …]
Davx512f-p5-scalef-x160-acc5.c30 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5() local
151 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
152 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
153 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
154 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
155 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
156 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
157 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
158 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
159 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
[all …]
Davx512f-p5-scalef-x160.c30 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160() local
147 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
148 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
149 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
150 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
151 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
152 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
153 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
154 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
155 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
[all …]
Davx512f-p5-scalef-x160-acc2.c30 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2() local
148 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
149 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
150 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
151 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
152 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
153 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
154 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
155 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
156 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
[all …]
Davx512f-p5-scalef-x144.c30 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144() local
138 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
139 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
140 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
141 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
142 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
143 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
144 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
145 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
146 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
[all …]
Davx512f-p5-scalef-x144-acc3.c30 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3() local
140 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
141 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
142 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
143 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
144 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
145 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
146 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
147 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
148 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
[all …]
/external/XNNPACK/src/f32-vscaleexpminusmax/gen/
Davx512f-p5-scalef-x192.c31 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192() local
166 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
167 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
168 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
169 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
170 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
171 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
172 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
173 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
174 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
[all …]
Davx512f-p5-scalef-x176.c31 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176() local
157 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
158 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
159 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
160 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
161 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
162 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
163 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
164 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
165 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
[all …]
Davx512f-p5-scalef-x160.c31 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160() local
148 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
149 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
150 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
151 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
152 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
153 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
154 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
155 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
156 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
[all …]
Davx512f-p5-scalef-x144.c31 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144() local
139 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
140 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
141 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
142 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
143 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
144 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
145 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
146 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
147 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
[all …]
/external/XNNPACK/src/f32-vscaleextexp/gen/
Davx512f-p5-scalef-x192.c32 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192() local
153 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
154 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
155 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
156 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
157 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
158 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
159 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
160 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
161 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
[all …]
Davx512f-p5-scalef-x176.c32 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176() local
145 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
146 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
147 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
148 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
149 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
150 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
151 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
152 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
153 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
[all …]
Davx512f-p5-scalef-x160.c32 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160() local
137 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
138 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
139 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
140 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
141 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
142 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
143 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
144 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
145 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
[all …]
Davx512f-p5-scalef-x144.c32 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144() local
129 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
130 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
131 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
132 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
133 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
134 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
135 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
136 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
137 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
[all …]
/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/
Davx512f-p5-scalef-x192-acc2.c31 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2() local
167 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
168 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
169 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
170 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
171 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
172 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
173 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
174 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
175 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
[all …]
Davx512f-p5-scalef-x192-acc3.c31 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3() local
168 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
169 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
170 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
171 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
172 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
173 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
174 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
175 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
176 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
[all …]
Davx512f-p5-scalef-x192.c31 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192() local
166 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
167 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
168 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
169 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
170 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
171 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
172 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
173 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
174 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
[all …]
Davx512f-p5-scalef-x160-acc2.c31 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2() local
149 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
150 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
151 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
152 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
153 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
154 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
155 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
156 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
157 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
[all …]
Davx512f-p5-scalef-x160.c31 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160() local
148 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
149 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
150 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
151 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
152 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
153 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
154 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
155 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
156 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
[all …]
Davx512f-p5-scalef-x144.c31 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144() local
139 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
140 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
141 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
142 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
143 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
144 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
145 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
146 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
147 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
[all …]
Davx512f-p5-scalef-x144-acc3.c31 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3() local
141 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
142 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
143 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
144 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
145 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
146 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
147 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
148 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
149 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
[all …]
/external/XNNPACK/src/f32-raddextexp/gen/
Davx512f-p5-scalef-x192.c31 const __m512 vc0 = _mm512_set1_ps(1.0f); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192() local
153 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
154 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
155 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
156 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
157 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
158 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
159 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
160 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
161 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
[all …]

1234