Home
last modified time | relevance | path

Searched refs:vt9 (Results 1 – 25 of 63) sorted by relevance

123

/external/XNNPACK/src/f32-vscaleexpminusmax/gen/
Davx2-p5-x80.c119 __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x80() local
130 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_lo, vt9); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x80()
142 __m256 vp9 = _mm256_fmadd_ps(vc5, vt9, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x80()
153 vp9 = _mm256_fmadd_ps(vp9, vt9, vc3); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x80()
164 vp9 = _mm256_fmadd_ps(vp9, vt9, vc2); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x80()
175 vp9 = _mm256_fmadd_ps(vp9, vt9, vc1); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x80()
190 vt9 = _mm256_mul_ps(vt9, vs9); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x80()
201 __m256 vf9 = _mm256_fmadd_ps(vt9, vp9, vs9); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x80()
Davx2-p5-x88.c124 __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x88() local
136 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_lo, vt9); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x88()
149 __m256 vp9 = _mm256_fmadd_ps(vc5, vt9, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x88()
161 vp9 = _mm256_fmadd_ps(vp9, vt9, vc3); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x88()
173 vp9 = _mm256_fmadd_ps(vp9, vt9, vc2); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x88()
185 vp9 = _mm256_fmadd_ps(vp9, vt9, vc1); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x88()
201 vt9 = _mm256_mul_ps(vt9, vs9); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x88()
213 __m256 vf9 = _mm256_fmadd_ps(vt9, vp9, vs9); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x88()
Davx512f-p5-scalef-x160.c90 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160() local
101 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_lo, vt9); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
113 __m512 vp9 = _mm512_fmadd_ps(vc5, vt9, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
124 vp9 = _mm512_fmadd_ps(vp9, vt9, vc3); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
135 vp9 = _mm512_fmadd_ps(vp9, vt9, vc2); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
146 vp9 = _mm512_fmadd_ps(vp9, vt9, vc1); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
157 vp9 = _mm512_fmadd_ps(vp9, vt9, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
Davx512f-p5-scalef-x176.c93 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176() local
105 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_lo, vt9); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
118 __m512 vp9 = _mm512_fmadd_ps(vc5, vt9, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
130 vp9 = _mm512_fmadd_ps(vp9, vt9, vc3); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
142 vp9 = _mm512_fmadd_ps(vp9, vt9, vc2); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
154 vp9 = _mm512_fmadd_ps(vp9, vt9, vc1); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
166 vp9 = _mm512_fmadd_ps(vp9, vt9, vc0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
Davx2-p5-x96.c129 __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96() local
142 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_lo, vt9); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96()
156 __m256 vp9 = _mm256_fmadd_ps(vc5, vt9, vc4); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96()
169 vp9 = _mm256_fmadd_ps(vp9, vt9, vc3); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96()
182 vp9 = _mm256_fmadd_ps(vp9, vt9, vc2); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96()
195 vp9 = _mm256_fmadd_ps(vp9, vt9, vc1); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96()
212 vt9 = _mm256_mul_ps(vt9, vs9); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96()
225 __m256 vf9 = _mm256_fmadd_ps(vt9, vp9, vs9); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96()
/external/XNNPACK/src/f32-raddexpminusmax/gen/
Davx2-p5-x80-acc2.c118 __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2() local
129 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_lo, vt9); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2()
141 __m256 vp9 = _mm256_fmadd_ps(vc5, vt9, vc4); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2()
152 vp9 = _mm256_fmadd_ps(vp9, vt9, vc3); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2()
163 vp9 = _mm256_fmadd_ps(vp9, vt9, vc2); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2()
174 vp9 = _mm256_fmadd_ps(vp9, vt9, vc1); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2()
189 vt9 = _mm256_mul_ps(vt9, vs9); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2()
200 __m256 vf9 = _mm256_fmadd_ps(vt9, vp9, vs9); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2()
Davx2-p5-x80.c117 __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80() local
128 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_lo, vt9); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80()
140 __m256 vp9 = _mm256_fmadd_ps(vc5, vt9, vc4); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80()
151 vp9 = _mm256_fmadd_ps(vp9, vt9, vc3); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80()
162 vp9 = _mm256_fmadd_ps(vp9, vt9, vc2); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80()
173 vp9 = _mm256_fmadd_ps(vp9, vt9, vc1); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80()
188 vt9 = _mm256_mul_ps(vt9, vs9); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80()
199 __m256 vf9 = _mm256_fmadd_ps(vt9, vp9, vs9); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80()
Davx2-p5-x80-acc5.c121 __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc5() local
132 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_lo, vt9); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc5()
144 __m256 vp9 = _mm256_fmadd_ps(vc5, vt9, vc4); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc5()
155 vp9 = _mm256_fmadd_ps(vp9, vt9, vc3); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc5()
166 vp9 = _mm256_fmadd_ps(vp9, vt9, vc2); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc5()
177 vp9 = _mm256_fmadd_ps(vp9, vt9, vc1); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc5()
192 vt9 = _mm256_mul_ps(vt9, vs9); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc5()
203 __m256 vf9 = _mm256_fmadd_ps(vt9, vp9, vs9); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc5()
Davx512f-p5-scalef-x160-acc2.c90 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2() local
101 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_lo, vt9); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
113 __m512 vp9 = _mm512_fmadd_ps(vc5, vt9, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
124 vp9 = _mm512_fmadd_ps(vp9, vt9, vc3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
135 vp9 = _mm512_fmadd_ps(vp9, vt9, vc2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
146 vp9 = _mm512_fmadd_ps(vp9, vt9, vc1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
157 vp9 = _mm512_fmadd_ps(vp9, vt9, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
Davx512f-p5-scalef-x160-acc5.c93 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5() local
104 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_lo, vt9); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
116 __m512 vp9 = _mm512_fmadd_ps(vc5, vt9, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
127 vp9 = _mm512_fmadd_ps(vp9, vt9, vc3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
138 vp9 = _mm512_fmadd_ps(vp9, vt9, vc2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
149 vp9 = _mm512_fmadd_ps(vp9, vt9, vc1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
160 vp9 = _mm512_fmadd_ps(vp9, vt9, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
Davx512f-p5-scalef-x160.c89 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160() local
100 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_lo, vt9); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
112 __m512 vp9 = _mm512_fmadd_ps(vc5, vt9, vc4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
123 vp9 = _mm512_fmadd_ps(vp9, vt9, vc3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
134 vp9 = _mm512_fmadd_ps(vp9, vt9, vc2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
145 vp9 = _mm512_fmadd_ps(vp9, vt9, vc1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
156 vp9 = _mm512_fmadd_ps(vp9, vt9, vc0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
Davx2-p5-x96-acc2.c128 __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc2() local
141 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_lo, vt9); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc2()
155 __m256 vp9 = _mm256_fmadd_ps(vc5, vt9, vc4); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc2()
168 vp9 = _mm256_fmadd_ps(vp9, vt9, vc3); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc2()
181 vp9 = _mm256_fmadd_ps(vp9, vt9, vc2); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc2()
194 vp9 = _mm256_fmadd_ps(vp9, vt9, vc1); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc2()
211 vt9 = _mm256_mul_ps(vt9, vs9); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc2()
224 __m256 vf9 = _mm256_fmadd_ps(vt9, vp9, vs9); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc2()
/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/
Davx2-p5-x80-acc5.c122 __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80_acc5() local
133 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_lo, vt9); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80_acc5()
145 __m256 vp9 = _mm256_fmadd_ps(vc5, vt9, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80_acc5()
156 vp9 = _mm256_fmadd_ps(vp9, vt9, vc3); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80_acc5()
167 vp9 = _mm256_fmadd_ps(vp9, vt9, vc2); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80_acc5()
178 vp9 = _mm256_fmadd_ps(vp9, vt9, vc1); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80_acc5()
193 vt9 = _mm256_mul_ps(vt9, vs9); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80_acc5()
204 __m256 vf9 = _mm256_fmadd_ps(vt9, vp9, vs9); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80_acc5()
Davx2-p5-x80.c118 __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80() local
129 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_lo, vt9); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80()
141 __m256 vp9 = _mm256_fmadd_ps(vc5, vt9, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80()
152 vp9 = _mm256_fmadd_ps(vp9, vt9, vc3); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80()
163 vp9 = _mm256_fmadd_ps(vp9, vt9, vc2); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80()
174 vp9 = _mm256_fmadd_ps(vp9, vt9, vc1); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80()
189 vt9 = _mm256_mul_ps(vt9, vs9); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80()
200 __m256 vf9 = _mm256_fmadd_ps(vt9, vp9, vs9); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80()
Davx2-p5-x80-acc2.c119 __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80_acc2() local
130 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_lo, vt9); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80_acc2()
142 __m256 vp9 = _mm256_fmadd_ps(vc5, vt9, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80_acc2()
153 vp9 = _mm256_fmadd_ps(vp9, vt9, vc3); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80_acc2()
164 vp9 = _mm256_fmadd_ps(vp9, vt9, vc2); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80_acc2()
175 vp9 = _mm256_fmadd_ps(vp9, vt9, vc1); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80_acc2()
190 vt9 = _mm256_mul_ps(vt9, vs9); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80_acc2()
201 __m256 vf9 = _mm256_fmadd_ps(vt9, vp9, vs9); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80_acc2()
Davx512f-p5-scalef-x160-acc2.c91 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2() local
102 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_lo, vt9); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
114 __m512 vp9 = _mm512_fmadd_ps(vc5, vt9, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
125 vp9 = _mm512_fmadd_ps(vp9, vt9, vc3); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
136 vp9 = _mm512_fmadd_ps(vp9, vt9, vc2); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
147 vp9 = _mm512_fmadd_ps(vp9, vt9, vc1); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
158 vp9 = _mm512_fmadd_ps(vp9, vt9, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
Davx512f-p5-scalef-x160.c90 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160() local
101 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_lo, vt9); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
113 __m512 vp9 = _mm512_fmadd_ps(vc5, vt9, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
124 vp9 = _mm512_fmadd_ps(vp9, vt9, vc3); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
135 vp9 = _mm512_fmadd_ps(vp9, vt9, vc2); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
146 vp9 = _mm512_fmadd_ps(vp9, vt9, vc1); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
157 vp9 = _mm512_fmadd_ps(vp9, vt9, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
Davx512f-p5-scalef-x160-acc5.c94 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5() local
105 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_lo, vt9); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
117 __m512 vp9 = _mm512_fmadd_ps(vc5, vt9, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
128 vp9 = _mm512_fmadd_ps(vp9, vt9, vc3); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
139 vp9 = _mm512_fmadd_ps(vp9, vt9, vc2); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
150 vp9 = _mm512_fmadd_ps(vp9, vt9, vc1); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
161 vp9 = _mm512_fmadd_ps(vp9, vt9, vc0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
Davx2-p5-x96-acc2.c129 __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc2() local
142 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_lo, vt9); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc2()
156 __m256 vp9 = _mm256_fmadd_ps(vc5, vt9, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc2()
169 vp9 = _mm256_fmadd_ps(vp9, vt9, vc3); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc2()
182 vp9 = _mm256_fmadd_ps(vp9, vt9, vc2); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc2()
195 vp9 = _mm256_fmadd_ps(vp9, vt9, vc1); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc2()
212 vt9 = _mm256_mul_ps(vt9, vs9); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc2()
225 __m256 vf9 = _mm256_fmadd_ps(vt9, vp9, vs9); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc2()
Davx2-p5-x96.c128 __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96() local
141 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_lo, vt9); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96()
155 __m256 vp9 = _mm256_fmadd_ps(vc5, vt9, vc4); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96()
168 vp9 = _mm256_fmadd_ps(vp9, vt9, vc3); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96()
181 vp9 = _mm256_fmadd_ps(vp9, vt9, vc2); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96()
194 vp9 = _mm256_fmadd_ps(vp9, vt9, vc1); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96()
211 vt9 = _mm256_mul_ps(vt9, vs9); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96()
224 __m256 vf9 = _mm256_fmadd_ps(vt9, vp9, vs9); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96()
/external/XNNPACK/src/f32-vscaleextexp/gen/
Davx512f-p5-scalef-x160.c79 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160() local
90 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_lo, vt9); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
102 __m512 vp9 = _mm512_fmadd_ps(vc5, vt9, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
113 vp9 = _mm512_fmadd_ps(vp9, vt9, vc3); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
124 vp9 = _mm512_fmadd_ps(vp9, vt9, vc2); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
135 vp9 = _mm512_fmadd_ps(vp9, vt9, vc1); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
146 vp9 = _mm512_fmadd_ps(vp9, vt9, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
Davx512f-p5-scalef-x176.c81 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176() local
93 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_lo, vt9); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
106 __m512 vp9 = _mm512_fmadd_ps(vc5, vt9, vc4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
118 vp9 = _mm512_fmadd_ps(vp9, vt9, vc3); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
130 vp9 = _mm512_fmadd_ps(vp9, vt9, vc2); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
142 vp9 = _mm512_fmadd_ps(vp9, vt9, vc1); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
154 vp9 = _mm512_fmadd_ps(vp9, vt9, vc0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
/external/XNNPACK/src/f32-raddextexp/gen/
Davx512f-p5-scalef-x160-acc2.c81 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2() local
92 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_lo, vt9); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2()
104 __m512 vp9 = _mm512_fmadd_ps(vc5, vt9, vc4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2()
115 vp9 = _mm512_fmadd_ps(vp9, vt9, vc3); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2()
126 vp9 = _mm512_fmadd_ps(vp9, vt9, vc2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2()
137 vp9 = _mm512_fmadd_ps(vp9, vt9, vc1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2()
148 vp9 = _mm512_fmadd_ps(vp9, vt9, vc0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2()
Davx512f-p5-scalef-x160.c79 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160() local
90 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_lo, vt9); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160()
102 __m512 vp9 = _mm512_fmadd_ps(vc5, vt9, vc4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160()
113 vp9 = _mm512_fmadd_ps(vp9, vt9, vc3); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160()
124 vp9 = _mm512_fmadd_ps(vp9, vt9, vc2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160()
135 vp9 = _mm512_fmadd_ps(vp9, vt9, vc1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160()
146 vp9 = _mm512_fmadd_ps(vp9, vt9, vc0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160()
/external/XNNPACK/src/f32-velu/gen/
Dvelu-avx2-rr1-p6-x80.c107 __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2, vz9); in xnn_f32_velu_ukernel__avx2_rr1_p6_x80() local
118 __m256 vp9 = _mm256_fmadd_ps(vc6, vt9, vc5); in xnn_f32_velu_ukernel__avx2_rr1_p6_x80()
129 vp9 = _mm256_fmadd_ps(vp9, vt9, vc4); in xnn_f32_velu_ukernel__avx2_rr1_p6_x80()
140 vp9 = _mm256_fmadd_ps(vp9, vt9, vc3); in xnn_f32_velu_ukernel__avx2_rr1_p6_x80()
151 vp9 = _mm256_fmadd_ps(vp9, vt9, vc2); in xnn_f32_velu_ukernel__avx2_rr1_p6_x80()
171 vp9 = _mm256_mul_ps(vp9, vt9); in xnn_f32_velu_ukernel__avx2_rr1_p6_x80()
172 vt9 = _mm256_mul_ps(vt9, vs9); in xnn_f32_velu_ukernel__avx2_rr1_p6_x80()
193 vp9 = _mm256_fmadd_ps(vp9, vt9, vt9); in xnn_f32_velu_ukernel__avx2_rr1_p6_x80()

123