Home
last modified time | relevance | path

Searched refs:vacce1 (Results 1 – 16 of 16) sorted by relevance

/external/XNNPACK/src/f32-raddextexp/gen/
Davx512f-p5-scalef-x128-acc2.c43 __m512 vacce1 = vminus_inf; in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2() local
143 __m512 vmax_e1 = _mm512_max_ps(vacce1, vn1); 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()
175 vacce1 = vmax_e1; in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2()
179 const __m512 vmax_acce01 = _mm512_max_ps(vacce0, vacce1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2()
182 const __m512 vdelta_acce1 = _mm512_sub_ps(vacce1, vmax_acce01); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2()
Davx512f-p5-scalef-x144-acc3.c44 __m512 vacce1 = vminus_inf; in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3() local
154 __m512 vmax_e1 = _mm512_max_ps(vacce1, vn1); 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()
191 vacce1 = vmax_e1; in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3()
196 const __m512 vmax_acce01 = _mm512_max_ps(vacce0, vacce1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3()
201 const __m512 vdelta_acce1 = _mm512_sub_ps(vacce1, vmax_acce012); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3()
Davx512f-p5-scalef-x160-acc2.c43 __m512 vacce1 = vminus_inf; in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2() local
161 __m512 vmax_e1 = _mm512_max_ps(vacce1, vn1); 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()
199 vacce1 = vmax_e1; in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2()
203 const __m512 vmax_acce01 = _mm512_max_ps(vacce0, vacce1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2()
206 const __m512 vdelta_acce1 = _mm512_sub_ps(vacce1, vmax_acce01); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2()
Davx512f-p5-scalef-x128-acc4.c45 __m512 vacce1 = vminus_inf; in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4() local
147 __m512 vmax_e1 = _mm512_max_ps(vacce1, vn1); 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()
183 vacce1 = vmax_e1; in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4()
189 const __m512 vmax_acce01 = _mm512_max_ps(vacce0, vacce1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4()
194 const __m512 vdelta_acce1 = _mm512_sub_ps(vacce1, vmax_acce0123); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4()
Davx512f-p5-scalef-x192-acc2.c43 __m512 vacce1 = vminus_inf; in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2() local
179 __m512 vmax_e1 = _mm512_max_ps(vacce1, vn1); 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()
223 vacce1 = vmax_e1; in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
227 const __m512 vmax_acce01 = _mm512_max_ps(vacce0, vacce1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
230 const __m512 vdelta_acce1 = _mm512_sub_ps(vacce1, vmax_acce01); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
Davx512f-p5-scalef-x160-acc5.c46 __m512 vacce1 = vminus_inf; in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5() local
167 __m512 vmax_e1 = _mm512_max_ps(vacce1, vn1); 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()
211 vacce1 = vmax_e1; in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5()
218 const __m512 vmax_acce01 = _mm512_max_ps(vacce0, vacce1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5()
225 const __m512 vdelta_acce1 = _mm512_sub_ps(vacce1, vmax_acce01234); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5()
Davx2-p5-x64-acc2.c47 __m256 vacce1 = vminus_inf; in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc2() local
147 __m256 vmax_e1 = _mm256_max_ps(vacce1, vn1); in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc2()
160 const __m256 vdelta_acce1 = _mm256_max_ps(_mm256_sub_ps(vacce1, vmax_e1), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc2()
199 vacce1 = vmax_e1; in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc2()
203 const __m256 vmax_acce01 = _mm256_max_ps(vacce0, vacce1); in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc2()
206 const __m256 vdelta_acce1 = _mm256_max_ps(_mm256_sub_ps(vacce1, vmax_acce01), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc2()
Davx2-p5-x64-acc4.c49 __m256 vacce1 = vminus_inf; in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc4() local
151 __m256 vmax_e1 = _mm256_max_ps(vacce1, vn1); in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc4()
164 const __m256 vdelta_acce1 = _mm256_max_ps(_mm256_sub_ps(vacce1, vmax_e1), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc4()
209 vacce1 = vmax_e1; in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc4()
215 const __m256 vmax_acce01 = _mm256_max_ps(vacce0, vacce1); in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc4()
220 const __m256 vdelta_acce1 = _mm256_max_ps(_mm256_sub_ps(vacce1, vmax_acce0123), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc4()
Davx2-p5-x80-acc2.c47 __m256 vacce1 = vminus_inf; in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2() local
165 __m256 vmax_e1 = _mm256_max_ps(vacce1, vn1); in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2()
180 const __m256 vdelta_acce1 = _mm256_max_ps(_mm256_sub_ps(vacce1, vmax_e1), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2()
225 vacce1 = vmax_e1; in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2()
229 const __m256 vmax_acce01 = _mm256_max_ps(vacce0, vacce1); in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2()
232 const __m256 vdelta_acce1 = _mm256_max_ps(_mm256_sub_ps(vacce1, vmax_acce01), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2()
Davx512f-p5-scalef-x192-acc3.c44 __m512 vacce1 = vminus_inf; in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3() local
181 __m512 vmax_e1 = _mm512_max_ps(vacce1, vn1); 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()
227 vacce1 = vmax_e1; in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3()
232 const __m512 vmax_acce01 = _mm512_max_ps(vacce0, vacce1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3()
237 const __m512 vdelta_acce1 = _mm512_sub_ps(vacce1, vmax_acce012); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3()
Davx2-p5-x72-acc3.c48 __m256 vacce1 = vminus_inf; in xnn_f32_raddextexp_ukernel__avx2_p5_x72_acc3() local
158 __m256 vmax_e1 = _mm256_max_ps(vacce1, vn1); in xnn_f32_raddextexp_ukernel__avx2_p5_x72_acc3()
172 const __m256 vdelta_acce1 = _mm256_max_ps(_mm256_sub_ps(vacce1, vmax_e1), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x72_acc3()
217 vacce1 = vmax_e1; in xnn_f32_raddextexp_ukernel__avx2_p5_x72_acc3()
222 const __m256 vmax_acce01 = _mm256_max_ps(vacce0, vacce1); in xnn_f32_raddextexp_ukernel__avx2_p5_x72_acc3()
227 const __m256 vdelta_acce1 = _mm256_max_ps(_mm256_sub_ps(vacce1, vmax_acce012), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x72_acc3()
Davx512f-p5-scalef-x192-acc6.c47 __m512 vacce1 = vminus_inf; in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6() local
187 __m512 vmax_e1 = _mm512_max_ps(vacce1, vn1); 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()
239 vacce1 = vmax_e1; in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
247 const __m512 vmax_acce01 = _mm512_max_ps(vacce0, vacce1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
254 const __m512 vdelta_acce1 = _mm512_sub_ps(vacce1, vmax_acce012345); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
Davx2-p5-x80-acc5.c50 __m256 vacce1 = vminus_inf; in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc5() local
171 __m256 vmax_e1 = _mm256_max_ps(vacce1, vn1); in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc5()
186 const __m256 vdelta_acce1 = _mm256_max_ps(_mm256_sub_ps(vacce1, vmax_e1), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc5()
240 vacce1 = vmax_e1; in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc5()
247 const __m256 vmax_acce01 = _mm256_max_ps(vacce0, vacce1); in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc5()
254 const __m256 vdelta_acce1 = _mm256_max_ps(_mm256_sub_ps(vacce1, vmax_acce01234), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc5()
Davx2-p5-x96-acc3.c48 __m256 vacce1 = vminus_inf; in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc3() local
185 __m256 vmax_e1 = _mm256_max_ps(vacce1, vn1); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc3()
202 const __m256 vdelta_acce1 = _mm256_max_ps(_mm256_sub_ps(vacce1, vmax_e1), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc3()
256 vacce1 = vmax_e1; in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc3()
261 const __m256 vmax_acce01 = _mm256_max_ps(vacce0, vacce1); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc3()
266 const __m256 vdelta_acce1 = _mm256_max_ps(_mm256_sub_ps(vacce1, vmax_acce012), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc3()
Davx2-p5-x96-acc2.c47 __m256 vacce1 = vminus_inf; in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc2() local
183 __m256 vmax_e1 = _mm256_max_ps(vacce1, vn1); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc2()
200 const __m256 vdelta_acce1 = _mm256_max_ps(_mm256_sub_ps(vacce1, vmax_e1), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc2()
251 vacce1 = vmax_e1; in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc2()
255 const __m256 vmax_acce01 = _mm256_max_ps(vacce0, vacce1); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc2()
258 const __m256 vdelta_acce1 = _mm256_max_ps(_mm256_sub_ps(vacce1, vmax_acce01), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc2()
Davx2-p5-x96-acc6.c51 __m256 vacce1 = vminus_inf; in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc6() local
191 __m256 vmax_e1 = _mm256_max_ps(vacce1, vn1); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc6()
208 const __m256 vdelta_acce1 = _mm256_max_ps(_mm256_sub_ps(vacce1, vmax_e1), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc6()
271 vacce1 = vmax_e1; in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc6()
279 const __m256 vmax_acce01 = _mm256_max_ps(vacce0, vacce1); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc6()
286 const __m256 vdelta_acce1 = _mm256_max_ps(_mm256_sub_ps(vacce1, vmax_acce012345), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc6()