Home
last modified time | relevance | path

Searched refs:vt11 (Results 1 – 25 of 28) sorted by relevance

12

/external/XNNPACK/src/f32-raddexpminusmax/gen/
Davx2-p5-x96.c129 __m256 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_hi, vx11); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96() local
142 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_lo, vt11); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96()
156 __m256 vp11 = _mm256_fmadd_ps(vc5, vt11, vc4); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96()
169 vp11 = _mm256_fmadd_ps(vp11, vt11, vc3); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96()
182 vp11 = _mm256_fmadd_ps(vp11, vt11, vc2); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96()
195 vp11 = _mm256_fmadd_ps(vp11, vt11, vc1); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96()
212 vt11 = _mm256_mul_ps(vt11, vs11); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96()
225 __m256 vf11 = _mm256_fmadd_ps(vt11, vp11, vs11); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96()
Davx2-p5-x96-acc2.c130 __m256 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_hi, vx11); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc2() local
143 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_lo, vt11); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc2()
157 __m256 vp11 = _mm256_fmadd_ps(vc5, vt11, vc4); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc2()
170 vp11 = _mm256_fmadd_ps(vp11, vt11, vc3); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc2()
183 vp11 = _mm256_fmadd_ps(vp11, vt11, vc2); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc2()
196 vp11 = _mm256_fmadd_ps(vp11, vt11, vc1); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc2()
213 vt11 = _mm256_mul_ps(vt11, vs11); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc2()
226 __m256 vf11 = _mm256_fmadd_ps(vt11, vp11, vs11); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc2()
Davx2-p5-x96-acc3.c131 __m256 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_hi, vx11); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc3() local
144 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_lo, vt11); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc3()
158 __m256 vp11 = _mm256_fmadd_ps(vc5, vt11, vc4); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc3()
171 vp11 = _mm256_fmadd_ps(vp11, vt11, vc3); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc3()
184 vp11 = _mm256_fmadd_ps(vp11, vt11, vc2); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc3()
197 vp11 = _mm256_fmadd_ps(vp11, vt11, vc1); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc3()
214 vt11 = _mm256_mul_ps(vt11, vs11); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc3()
227 __m256 vf11 = _mm256_fmadd_ps(vt11, vp11, vs11); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc3()
Davx2-p5-x96-acc6.c134 __m256 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_hi, vx11); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc6() local
147 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_lo, vt11); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc6()
161 __m256 vp11 = _mm256_fmadd_ps(vc5, vt11, vc4); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc6()
174 vp11 = _mm256_fmadd_ps(vp11, vt11, vc3); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc6()
187 vp11 = _mm256_fmadd_ps(vp11, vt11, vc2); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc6()
200 vp11 = _mm256_fmadd_ps(vp11, vt11, vc1); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc6()
217 vt11 = _mm256_mul_ps(vt11, vs11); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc6()
230 __m256 vf11 = _mm256_fmadd_ps(vt11, vp11, vs11); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc6()
Davx512f-p5-scalef-x192-acc2.c98 __m512 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_hi, vx11); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2() local
111 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_lo, vt11); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
125 __m512 vp11 = _mm512_fmadd_ps(vc5, vt11, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
138 vp11 = _mm512_fmadd_ps(vp11, vt11, vc3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
151 vp11 = _mm512_fmadd_ps(vp11, vt11, vc2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
164 vp11 = _mm512_fmadd_ps(vp11, vt11, vc1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
177 vp11 = _mm512_fmadd_ps(vp11, vt11, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
Davx512f-p5-scalef-x192.c97 __m512 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_hi, vx11); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192() local
110 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_lo, vt11); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
124 __m512 vp11 = _mm512_fmadd_ps(vc5, vt11, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
137 vp11 = _mm512_fmadd_ps(vp11, vt11, vc3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
150 vp11 = _mm512_fmadd_ps(vp11, vt11, vc2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
163 vp11 = _mm512_fmadd_ps(vp11, vt11, vc1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
176 vp11 = _mm512_fmadd_ps(vp11, vt11, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
Davx512f-p5-scalef-x192-acc3.c99 __m512 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_hi, vx11); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3() local
112 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_lo, vt11); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
126 __m512 vp11 = _mm512_fmadd_ps(vc5, vt11, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
139 vp11 = _mm512_fmadd_ps(vp11, vt11, vc3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
152 vp11 = _mm512_fmadd_ps(vp11, vt11, vc2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
165 vp11 = _mm512_fmadd_ps(vp11, vt11, vc1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
178 vp11 = _mm512_fmadd_ps(vp11, vt11, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
Davx512f-p5-scalef-x192-acc6.c102 __m512 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_hi, vx11); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6() local
115 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_lo, vt11); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
129 __m512 vp11 = _mm512_fmadd_ps(vc5, vt11, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
142 vp11 = _mm512_fmadd_ps(vp11, vt11, vc3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
155 vp11 = _mm512_fmadd_ps(vp11, vt11, vc2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
168 vp11 = _mm512_fmadd_ps(vp11, vt11, vc1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
181 vp11 = _mm512_fmadd_ps(vp11, vt11, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/
Davx2-p5-x96-acc6.c135 __m256 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_hi, vx11); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc6() local
148 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_lo, vt11); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc6()
162 __m256 vp11 = _mm256_fmadd_ps(vc5, vt11, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc6()
175 vp11 = _mm256_fmadd_ps(vp11, vt11, vc3); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc6()
188 vp11 = _mm256_fmadd_ps(vp11, vt11, vc2); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc6()
201 vp11 = _mm256_fmadd_ps(vp11, vt11, vc1); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc6()
218 vt11 = _mm256_mul_ps(vt11, vs11); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc6()
231 __m256 vf11 = _mm256_fmadd_ps(vt11, vp11, vs11); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc6()
Davx2-p5-x96-acc3.c132 __m256 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_hi, vx11); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc3() local
145 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_lo, vt11); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc3()
159 __m256 vp11 = _mm256_fmadd_ps(vc5, vt11, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc3()
172 vp11 = _mm256_fmadd_ps(vp11, vt11, vc3); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc3()
185 vp11 = _mm256_fmadd_ps(vp11, vt11, vc2); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc3()
198 vp11 = _mm256_fmadd_ps(vp11, vt11, vc1); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc3()
215 vt11 = _mm256_mul_ps(vt11, vs11); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc3()
228 __m256 vf11 = _mm256_fmadd_ps(vt11, vp11, vs11); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc3()
Davx2-p5-x96-acc2.c131 __m256 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_hi, vx11); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc2() local
144 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_lo, vt11); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc2()
158 __m256 vp11 = _mm256_fmadd_ps(vc5, vt11, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc2()
171 vp11 = _mm256_fmadd_ps(vp11, vt11, vc3); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc2()
184 vp11 = _mm256_fmadd_ps(vp11, vt11, vc2); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc2()
197 vp11 = _mm256_fmadd_ps(vp11, vt11, vc1); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc2()
214 vt11 = _mm256_mul_ps(vt11, vs11); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc2()
227 __m256 vf11 = _mm256_fmadd_ps(vt11, vp11, vs11); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc2()
Davx2-p5-x96.c130 __m256 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_hi, vx11); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96() local
143 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_lo, vt11); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96()
157 __m256 vp11 = _mm256_fmadd_ps(vc5, vt11, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96()
170 vp11 = _mm256_fmadd_ps(vp11, vt11, vc3); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96()
183 vp11 = _mm256_fmadd_ps(vp11, vt11, vc2); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96()
196 vp11 = _mm256_fmadd_ps(vp11, vt11, vc1); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96()
213 vt11 = _mm256_mul_ps(vt11, vs11); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96()
226 __m256 vf11 = _mm256_fmadd_ps(vt11, vp11, vs11); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96()
Davx512f-p5-scalef-x192-acc2.c99 __m512 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_hi, vx11); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2() local
112 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_lo, vt11); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
126 __m512 vp11 = _mm512_fmadd_ps(vc5, vt11, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
139 vp11 = _mm512_fmadd_ps(vp11, vt11, vc3); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
152 vp11 = _mm512_fmadd_ps(vp11, vt11, vc2); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
165 vp11 = _mm512_fmadd_ps(vp11, vt11, vc1); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
178 vp11 = _mm512_fmadd_ps(vp11, vt11, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
Davx512f-p5-scalef-x192-acc3.c100 __m512 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_hi, vx11); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3() local
113 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_lo, vt11); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
127 __m512 vp11 = _mm512_fmadd_ps(vc5, vt11, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
140 vp11 = _mm512_fmadd_ps(vp11, vt11, vc3); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
153 vp11 = _mm512_fmadd_ps(vp11, vt11, vc2); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
166 vp11 = _mm512_fmadd_ps(vp11, vt11, vc1); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
179 vp11 = _mm512_fmadd_ps(vp11, vt11, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
Davx512f-p5-scalef-x192.c98 __m512 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_hi, vx11); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192() local
111 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_lo, vt11); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
125 __m512 vp11 = _mm512_fmadd_ps(vc5, vt11, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
138 vp11 = _mm512_fmadd_ps(vp11, vt11, vc3); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
151 vp11 = _mm512_fmadd_ps(vp11, vt11, vc2); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
164 vp11 = _mm512_fmadd_ps(vp11, vt11, vc1); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
177 vp11 = _mm512_fmadd_ps(vp11, vt11, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
Davx512f-p5-scalef-x192-acc6.c103 __m512 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_hi, vx11); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6() local
116 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_lo, vt11); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
130 __m512 vp11 = _mm512_fmadd_ps(vc5, vt11, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
143 vp11 = _mm512_fmadd_ps(vp11, vt11, vc3); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
156 vp11 = _mm512_fmadd_ps(vp11, vt11, vc2); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
169 vp11 = _mm512_fmadd_ps(vp11, vt11, vc1); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
182 vp11 = _mm512_fmadd_ps(vp11, vt11, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
/external/XNNPACK/src/f32-vscaleexpminusmax/gen/
Davx2-p5-x96.c131 __m256 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_hi, vx11); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96() local
144 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_lo, vt11); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96()
158 __m256 vp11 = _mm256_fmadd_ps(vc5, vt11, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96()
171 vp11 = _mm256_fmadd_ps(vp11, vt11, vc3); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96()
184 vp11 = _mm256_fmadd_ps(vp11, vt11, vc2); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96()
197 vp11 = _mm256_fmadd_ps(vp11, vt11, vc1); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96()
214 vt11 = _mm256_mul_ps(vt11, vs11); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96()
227 __m256 vf11 = _mm256_fmadd_ps(vt11, vp11, vs11); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96()
Davx512f-p5-scalef-x192.c98 __m512 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_hi, vx11); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192() local
111 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_lo, vt11); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
125 __m512 vp11 = _mm512_fmadd_ps(vc5, vt11, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
138 vp11 = _mm512_fmadd_ps(vp11, vt11, vc3); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
151 vp11 = _mm512_fmadd_ps(vp11, vt11, vc2); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
164 vp11 = _mm512_fmadd_ps(vp11, vt11, vc1); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
177 vp11 = _mm512_fmadd_ps(vp11, vt11, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
/external/XNNPACK/src/f32-vscaleextexp/gen/
Davx512f-p5-scalef-x192.c85 __m512 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_hi, vx11); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192() local
98 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_lo, vt11); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
112 __m512 vp11 = _mm512_fmadd_ps(vc5, vt11, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
125 vp11 = _mm512_fmadd_ps(vp11, vt11, vc3); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
138 vp11 = _mm512_fmadd_ps(vp11, vt11, vc2); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
151 vp11 = _mm512_fmadd_ps(vp11, vt11, vc1); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
164 vp11 = _mm512_fmadd_ps(vp11, vt11, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
Davx2-p5-x96.c91 __m256 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_hi, vx11); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x96() local
104 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_lo, vt11); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x96()
118 __m256 vp11 = _mm256_fmadd_ps(vc5, vt11, vc4); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x96()
131 vp11 = _mm256_fmadd_ps(vp11, vt11, vc3); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x96()
144 vp11 = _mm256_fmadd_ps(vp11, vt11, vc2); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x96()
157 vp11 = _mm256_fmadd_ps(vp11, vt11, vc1); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x96()
170 vp11 = _mm256_fmadd_ps(vp11, vt11, vc0); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x96()
/external/XNNPACK/src/f32-raddextexp/gen/
Davx512f-p5-scalef-x192.c85 __m512 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_hi, vx11); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192() local
98 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_lo, vt11); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
112 __m512 vp11 = _mm512_fmadd_ps(vc5, vt11, vc4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
125 vp11 = _mm512_fmadd_ps(vp11, vt11, vc3); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
138 vp11 = _mm512_fmadd_ps(vp11, vt11, vc2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
151 vp11 = _mm512_fmadd_ps(vp11, vt11, vc1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
164 vp11 = _mm512_fmadd_ps(vp11, vt11, vc0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
Davx512f-p5-scalef-x192-acc2.c87 __m512 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_hi, vx11); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2() local
100 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_lo, vt11); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
114 __m512 vp11 = _mm512_fmadd_ps(vc5, vt11, vc4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
127 vp11 = _mm512_fmadd_ps(vp11, vt11, vc3); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
140 vp11 = _mm512_fmadd_ps(vp11, vt11, vc2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
153 vp11 = _mm512_fmadd_ps(vp11, vt11, vc1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
166 vp11 = _mm512_fmadd_ps(vp11, vt11, vc0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
Davx512f-p5-scalef-x192-acc3.c89 __m512 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_hi, vx11); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3() local
102 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_lo, vt11); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3()
116 __m512 vp11 = _mm512_fmadd_ps(vc5, vt11, vc4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3()
129 vp11 = _mm512_fmadd_ps(vp11, vt11, vc3); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3()
142 vp11 = _mm512_fmadd_ps(vp11, vt11, vc2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3()
155 vp11 = _mm512_fmadd_ps(vp11, vt11, vc1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3()
168 vp11 = _mm512_fmadd_ps(vp11, vt11, vc0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3()
Davx512f-p5-scalef-x192-acc6.c95 __m512 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_hi, vx11); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6() local
108 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_lo, vt11); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
122 __m512 vp11 = _mm512_fmadd_ps(vc5, vt11, vc4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
135 vp11 = _mm512_fmadd_ps(vp11, vt11, vc3); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
148 vp11 = _mm512_fmadd_ps(vp11, vt11, vc2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
161 vp11 = _mm512_fmadd_ps(vp11, vt11, vc1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
174 vp11 = _mm512_fmadd_ps(vp11, vt11, vc0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
Davx2-p5-x96.c89 __m256 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_hi, vx11); in xnn_f32_raddextexp_ukernel__avx2_p5_x96() local
102 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_lo, vt11); in xnn_f32_raddextexp_ukernel__avx2_p5_x96()
116 __m256 vp11 = _mm256_fmadd_ps(vc5, vt11, vc4); in xnn_f32_raddextexp_ukernel__avx2_p5_x96()
129 vp11 = _mm256_fmadd_ps(vp11, vt11, vc3); in xnn_f32_raddextexp_ukernel__avx2_p5_x96()
142 vp11 = _mm256_fmadd_ps(vp11, vt11, vc2); in xnn_f32_raddextexp_ukernel__avx2_p5_x96()
155 vp11 = _mm256_fmadd_ps(vp11, vt11, vc1); in xnn_f32_raddextexp_ukernel__avx2_p5_x96()
168 vp11 = _mm256_fmadd_ps(vp11, vt11, vc0); in xnn_f32_raddextexp_ukernel__avx2_p5_x96()

12