Home
last modified time | relevance | path

Searched refs:vt10 (Results 1 – 25 of 32) sorted by relevance

12

/external/XNNPACK/src/f32-vscaleexpminusmax/gen/
Davx2-p5-x88.c125 __m256 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_hi, vx10); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x88() local
137 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_lo, vt10); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x88()
150 __m256 vp10 = _mm256_fmadd_ps(vc5, vt10, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x88()
162 vp10 = _mm256_fmadd_ps(vp10, vt10, vc3); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x88()
174 vp10 = _mm256_fmadd_ps(vp10, vt10, vc2); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x88()
186 vp10 = _mm256_fmadd_ps(vp10, vt10, vc1); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x88()
202 vt10 = _mm256_mul_ps(vt10, vs10); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x88()
214 __m256 vf10 = _mm256_fmadd_ps(vt10, vp10, vs10); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x88()
Davx512f-p5-scalef-x176.c94 __m512 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_hi, vx10); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176() local
106 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_lo, vt10); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
119 __m512 vp10 = _mm512_fmadd_ps(vc5, vt10, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
131 vp10 = _mm512_fmadd_ps(vp10, vt10, vc3); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
143 vp10 = _mm512_fmadd_ps(vp10, vt10, vc2); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
155 vp10 = _mm512_fmadd_ps(vp10, vt10, vc1); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
167 vp10 = _mm512_fmadd_ps(vp10, vt10, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
Davx2-p5-x96.c130 __m256 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_hi, vx10); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96() local
143 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_lo, vt10); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96()
157 __m256 vp10 = _mm256_fmadd_ps(vc5, vt10, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96()
170 vp10 = _mm256_fmadd_ps(vp10, vt10, vc3); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96()
183 vp10 = _mm256_fmadd_ps(vp10, vt10, vc2); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96()
196 vp10 = _mm256_fmadd_ps(vp10, vt10, vc1); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96()
213 vt10 = _mm256_mul_ps(vt10, vs10); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96()
226 __m256 vf10 = _mm256_fmadd_ps(vt10, vp10, vs10); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96()
Davx512f-p5-scalef-x192.c97 __m512 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_hi, vx10); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192() local
110 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_lo, vt10); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
124 __m512 vp10 = _mm512_fmadd_ps(vc5, vt10, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
137 vp10 = _mm512_fmadd_ps(vp10, vt10, vc3); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
150 vp10 = _mm512_fmadd_ps(vp10, vt10, vc2); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
163 vp10 = _mm512_fmadd_ps(vp10, vt10, vc1); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
176 vp10 = _mm512_fmadd_ps(vp10, vt10, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
/external/XNNPACK/src/f32-vscaleextexp/gen/
Davx512f-p5-scalef-x176.c82 __m512 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_hi, vx10); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176() local
94 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_lo, vt10); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
107 __m512 vp10 = _mm512_fmadd_ps(vc5, vt10, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
119 vp10 = _mm512_fmadd_ps(vp10, vt10, vc3); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
131 vp10 = _mm512_fmadd_ps(vp10, vt10, vc2); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
143 vp10 = _mm512_fmadd_ps(vp10, vt10, vc1); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
155 vp10 = _mm512_fmadd_ps(vp10, vt10, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
Davx512f-p5-scalef-x192.c84 __m512 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_hi, vx10); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192() local
97 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_lo, vt10); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
111 __m512 vp10 = _mm512_fmadd_ps(vc5, vt10, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
124 vp10 = _mm512_fmadd_ps(vp10, vt10, vc3); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
137 vp10 = _mm512_fmadd_ps(vp10, vt10, vc2); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
150 vp10 = _mm512_fmadd_ps(vp10, vt10, vc1); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
163 vp10 = _mm512_fmadd_ps(vp10, vt10, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
/external/XNNPACK/src/f32-raddexpminusmax/gen/
Davx2-p5-x96-acc3.c130 __m256 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_hi, vx10); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc3() local
143 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_lo, vt10); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc3()
157 __m256 vp10 = _mm256_fmadd_ps(vc5, vt10, vc4); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc3()
170 vp10 = _mm256_fmadd_ps(vp10, vt10, vc3); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc3()
183 vp10 = _mm256_fmadd_ps(vp10, vt10, vc2); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc3()
196 vp10 = _mm256_fmadd_ps(vp10, vt10, vc1); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc3()
213 vt10 = _mm256_mul_ps(vt10, vs10); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc3()
226 __m256 vf10 = _mm256_fmadd_ps(vt10, vp10, vs10); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc3()
Davx2-p5-x96-acc6.c133 __m256 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_hi, vx10); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc6() local
146 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_lo, vt10); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc6()
160 __m256 vp10 = _mm256_fmadd_ps(vc5, vt10, vc4); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc6()
173 vp10 = _mm256_fmadd_ps(vp10, vt10, vc3); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc6()
186 vp10 = _mm256_fmadd_ps(vp10, vt10, vc2); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc6()
199 vp10 = _mm256_fmadd_ps(vp10, vt10, vc1); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc6()
216 vt10 = _mm256_mul_ps(vt10, vs10); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc6()
229 __m256 vf10 = _mm256_fmadd_ps(vt10, vp10, vs10); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc6()
Davx2-p5-x96.c128 __m256 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_hi, vx10); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96() local
141 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_lo, vt10); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96()
155 __m256 vp10 = _mm256_fmadd_ps(vc5, vt10, vc4); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96()
168 vp10 = _mm256_fmadd_ps(vp10, vt10, vc3); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96()
181 vp10 = _mm256_fmadd_ps(vp10, vt10, vc2); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96()
194 vp10 = _mm256_fmadd_ps(vp10, vt10, vc1); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96()
211 vt10 = _mm256_mul_ps(vt10, vs10); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96()
224 __m256 vf10 = _mm256_fmadd_ps(vt10, vp10, vs10); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96()
Davx2-p5-x96-acc2.c129 __m256 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_hi, vx10); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc2() local
142 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_lo, vt10); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc2()
156 __m256 vp10 = _mm256_fmadd_ps(vc5, vt10, vc4); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc2()
169 vp10 = _mm256_fmadd_ps(vp10, vt10, vc3); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc2()
182 vp10 = _mm256_fmadd_ps(vp10, vt10, vc2); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc2()
195 vp10 = _mm256_fmadd_ps(vp10, vt10, vc1); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc2()
212 vt10 = _mm256_mul_ps(vt10, vs10); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc2()
225 __m256 vf10 = _mm256_fmadd_ps(vt10, vp10, vs10); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc2()
Davx512f-p5-scalef-x192-acc6.c101 __m512 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_hi, vx10); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6() local
114 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_lo, vt10); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
128 __m512 vp10 = _mm512_fmadd_ps(vc5, vt10, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
141 vp10 = _mm512_fmadd_ps(vp10, vt10, vc3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
154 vp10 = _mm512_fmadd_ps(vp10, vt10, vc2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
167 vp10 = _mm512_fmadd_ps(vp10, vt10, vc1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
180 vp10 = _mm512_fmadd_ps(vp10, vt10, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
Davx512f-p5-scalef-x192-acc3.c98 __m512 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_hi, vx10); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3() local
111 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_lo, vt10); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
125 __m512 vp10 = _mm512_fmadd_ps(vc5, vt10, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
138 vp10 = _mm512_fmadd_ps(vp10, vt10, vc3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
151 vp10 = _mm512_fmadd_ps(vp10, vt10, vc2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
164 vp10 = _mm512_fmadd_ps(vp10, vt10, vc1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
177 vp10 = _mm512_fmadd_ps(vp10, vt10, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
Davx512f-p5-scalef-x192.c96 __m512 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_hi, vx10); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192() local
109 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_lo, vt10); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
123 __m512 vp10 = _mm512_fmadd_ps(vc5, vt10, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
136 vp10 = _mm512_fmadd_ps(vp10, vt10, vc3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
149 vp10 = _mm512_fmadd_ps(vp10, vt10, vc2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
162 vp10 = _mm512_fmadd_ps(vp10, vt10, vc1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
175 vp10 = _mm512_fmadd_ps(vp10, vt10, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
Davx512f-p5-scalef-x192-acc2.c97 __m512 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_hi, vx10); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2() local
110 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_lo, vt10); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
124 __m512 vp10 = _mm512_fmadd_ps(vc5, vt10, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
137 vp10 = _mm512_fmadd_ps(vp10, vt10, vc3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
150 vp10 = _mm512_fmadd_ps(vp10, vt10, vc2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
163 vp10 = _mm512_fmadd_ps(vp10, vt10, vc1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
176 vp10 = _mm512_fmadd_ps(vp10, vt10, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/
Davx2-p5-x96-acc2.c130 __m256 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_hi, vx10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc2() local
143 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_lo, vt10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc2()
157 __m256 vp10 = _mm256_fmadd_ps(vc5, vt10, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc2()
170 vp10 = _mm256_fmadd_ps(vp10, vt10, vc3); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc2()
183 vp10 = _mm256_fmadd_ps(vp10, vt10, vc2); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc2()
196 vp10 = _mm256_fmadd_ps(vp10, vt10, vc1); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc2()
213 vt10 = _mm256_mul_ps(vt10, vs10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc2()
226 __m256 vf10 = _mm256_fmadd_ps(vt10, vp10, vs10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc2()
Davx2-p5-x96-acc3.c131 __m256 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_hi, vx10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc3() local
144 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_lo, vt10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc3()
158 __m256 vp10 = _mm256_fmadd_ps(vc5, vt10, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc3()
171 vp10 = _mm256_fmadd_ps(vp10, vt10, vc3); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc3()
184 vp10 = _mm256_fmadd_ps(vp10, vt10, vc2); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc3()
197 vp10 = _mm256_fmadd_ps(vp10, vt10, vc1); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc3()
214 vt10 = _mm256_mul_ps(vt10, vs10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc3()
227 __m256 vf10 = _mm256_fmadd_ps(vt10, vp10, vs10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc3()
Davx2-p5-x96-acc6.c134 __m256 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_hi, vx10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc6() local
147 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_lo, vt10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc6()
161 __m256 vp10 = _mm256_fmadd_ps(vc5, vt10, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc6()
174 vp10 = _mm256_fmadd_ps(vp10, vt10, vc3); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc6()
187 vp10 = _mm256_fmadd_ps(vp10, vt10, vc2); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc6()
200 vp10 = _mm256_fmadd_ps(vp10, vt10, vc1); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc6()
217 vt10 = _mm256_mul_ps(vt10, vs10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc6()
230 __m256 vf10 = _mm256_fmadd_ps(vt10, vp10, vs10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc6()
Davx2-p5-x96.c129 __m256 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_hi, vx10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96() local
142 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_lo, vt10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96()
156 __m256 vp10 = _mm256_fmadd_ps(vc5, vt10, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96()
169 vp10 = _mm256_fmadd_ps(vp10, vt10, vc3); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96()
182 vp10 = _mm256_fmadd_ps(vp10, vt10, vc2); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96()
195 vp10 = _mm256_fmadd_ps(vp10, vt10, vc1); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96()
212 vt10 = _mm256_mul_ps(vt10, vs10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96()
225 __m256 vf10 = _mm256_fmadd_ps(vt10, vp10, vs10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96()
Davx512f-p5-scalef-x192.c97 __m512 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_hi, vx10); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192() local
110 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_lo, vt10); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
124 __m512 vp10 = _mm512_fmadd_ps(vc5, vt10, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
137 vp10 = _mm512_fmadd_ps(vp10, vt10, vc3); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
150 vp10 = _mm512_fmadd_ps(vp10, vt10, vc2); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
163 vp10 = _mm512_fmadd_ps(vp10, vt10, vc1); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
176 vp10 = _mm512_fmadd_ps(vp10, vt10, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
Davx512f-p5-scalef-x192-acc3.c99 __m512 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_hi, vx10); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3() local
112 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_lo, vt10); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
126 __m512 vp10 = _mm512_fmadd_ps(vc5, vt10, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
139 vp10 = _mm512_fmadd_ps(vp10, vt10, vc3); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
152 vp10 = _mm512_fmadd_ps(vp10, vt10, vc2); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
165 vp10 = _mm512_fmadd_ps(vp10, vt10, vc1); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
178 vp10 = _mm512_fmadd_ps(vp10, vt10, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
Davx512f-p5-scalef-x192-acc2.c98 __m512 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_hi, vx10); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2() local
111 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_lo, vt10); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
125 __m512 vp10 = _mm512_fmadd_ps(vc5, vt10, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
138 vp10 = _mm512_fmadd_ps(vp10, vt10, vc3); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
151 vp10 = _mm512_fmadd_ps(vp10, vt10, vc2); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
164 vp10 = _mm512_fmadd_ps(vp10, vt10, vc1); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
177 vp10 = _mm512_fmadd_ps(vp10, vt10, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
Davx512f-p5-scalef-x192-acc6.c102 __m512 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_hi, vx10); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6() local
115 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_lo, vt10); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
129 __m512 vp10 = _mm512_fmadd_ps(vc5, vt10, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
142 vp10 = _mm512_fmadd_ps(vp10, vt10, vc3); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
155 vp10 = _mm512_fmadd_ps(vp10, vt10, vc2); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
168 vp10 = _mm512_fmadd_ps(vp10, vt10, vc1); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
181 vp10 = _mm512_fmadd_ps(vp10, vt10, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
/external/XNNPACK/src/f32-raddextexp/gen/
Davx512f-p5-scalef-x192.c84 __m512 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_hi, vx10); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192() local
97 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_lo, vt10); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
111 __m512 vp10 = _mm512_fmadd_ps(vc5, vt10, vc4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
124 vp10 = _mm512_fmadd_ps(vp10, vt10, vc3); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
137 vp10 = _mm512_fmadd_ps(vp10, vt10, vc2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
150 vp10 = _mm512_fmadd_ps(vp10, vt10, vc1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
163 vp10 = _mm512_fmadd_ps(vp10, vt10, vc0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
Davx512f-p5-scalef-x192-acc2.c86 __m512 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_hi, vx10); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2() local
99 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_lo, vt10); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
113 __m512 vp10 = _mm512_fmadd_ps(vc5, vt10, vc4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
126 vp10 = _mm512_fmadd_ps(vp10, vt10, vc3); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
139 vp10 = _mm512_fmadd_ps(vp10, vt10, vc2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
152 vp10 = _mm512_fmadd_ps(vp10, vt10, vc1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
165 vp10 = _mm512_fmadd_ps(vp10, vt10, vc0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
Davx512f-p5-scalef-x192-acc3.c88 __m512 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_hi, vx10); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3() local
101 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_lo, vt10); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3()
115 __m512 vp10 = _mm512_fmadd_ps(vc5, vt10, vc4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3()
128 vp10 = _mm512_fmadd_ps(vp10, vt10, vc3); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3()
141 vp10 = _mm512_fmadd_ps(vp10, vt10, vc2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3()
154 vp10 = _mm512_fmadd_ps(vp10, vt10, vc1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3()
167 vp10 = _mm512_fmadd_ps(vp10, vt10, vc0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3()

12