| /external/XNNPACK/src/f32-vscaleexpminusmax/gen/ |
| D | avx2-p5-x88.c | 214 __m256 vf10 = _mm256_fmadd_ps(vt10, vp10, vs10); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x88() local
228 vf10 = _mm256_andnot_ps(_mm256_cmp_ps(vx10, vdenorm_cutoff, _CMP_LT_OS), vf10); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x88()
241 vf10 = _mm256_mul_ps(vf10, vscale); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x88()
254 _mm256_storeu_ps(output + 80, vf10); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x88()
|
| D | avx2-p5-x96.c | 226 __m256 vf10 = _mm256_fmadd_ps(vt10, vp10, vs10); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96() local
241 vf10 = _mm256_andnot_ps(_mm256_cmp_ps(vx10, vdenorm_cutoff, _CMP_LT_OS), vf10); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96()
255 vf10 = _mm256_mul_ps(vf10, vscale); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96()
269 _mm256_storeu_ps(output + 80, vf10); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96()
|
| D | avx512f-p5-scalef-x176.c | 182 __m512 vf10 = _mm512_scalef_ps(vp10, vn10); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176() local
195 vf10 = _mm512_mul_ps(vf10, vscale); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
209 _mm512_storeu_ps(output + 160, vf10); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
|
| D | avx512f-p5-scalef-x192.c | 192 __m512 vf10 = _mm512_scalef_ps(vp10, vn10); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192() local
206 vf10 = _mm512_mul_ps(vf10, vscale); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
221 _mm512_storeu_ps(output + 160, vf10); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
|
| /external/XNNPACK/src/f32-vscaleextexp/gen/ |
| D | avx512f-p5-scalef-x176.c | 173 __m512 vf10 = _mm512_mul_ps(vp10, vscalev); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176() local
198 vf10 = _mm512_scalef_ps(vf10, ve10); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
212 _mm512_storeu_ps(y + 160, vf10); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
|
| D | avx512f-p5-scalef-x192.c | 182 __m512 vf10 = _mm512_mul_ps(vp10, vscalev); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192() local
209 vf10 = _mm512_scalef_ps(vf10, ve10); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
224 _mm512_storeu_ps(y + 160, vf10); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
|
| D | avx2-p5-x88.c | 179 __m256 vf10 = _mm256_mul_ps(vp10, vscalev); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x88() local
235 vf10 = _mm256_mul_ps(vf10, vs10); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x88()
248 _mm256_storeu_ps(y + 80, vf10); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x88()
|
| D | avx2-p5-x96.c | 188 __m256 vf10 = _mm256_mul_ps(vp10, vscalev); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x96() local
248 vf10 = _mm256_mul_ps(vf10, vs10); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x96()
262 _mm256_storeu_ps(y + 80, vf10); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x96()
|
| /external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
| D | avx2-rr1-p5-x96-acc2.c | 195 __m256 vf10 = _mm256_fmadd_ps(vt10, vp10, vs10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_rr1_p5_x96_acc2() local
208 vf10 = _mm256_andnot_ps(_mm256_cmp_ps(vx10, vdenorm_cutoff, _CMP_LT_OS), vf10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_rr1_p5_x96_acc2()
221 _mm256_storeu_ps(output + 80, vf10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_rr1_p5_x96_acc2()
235 vacc0 = _mm256_add_ps(vacc0, vf10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_rr1_p5_x96_acc2()
|
| D | avx2-rr1-p5-x96.c | 194 __m256 vf10 = _mm256_fmadd_ps(vt10, vp10, vs10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_rr1_p5_x96() local
207 vf10 = _mm256_andnot_ps(_mm256_cmp_ps(vx10, vdenorm_cutoff, _CMP_LT_OS), vf10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_rr1_p5_x96()
220 _mm256_storeu_ps(output + 80, vf10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_rr1_p5_x96()
234 vacc0 = _mm256_add_ps(vacc0, vf10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_rr1_p5_x96()
|
| D | avx2-rr1-p5-x96-acc3.c | 196 __m256 vf10 = _mm256_fmadd_ps(vt10, vp10, vs10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_rr1_p5_x96_acc3() local
209 vf10 = _mm256_andnot_ps(_mm256_cmp_ps(vx10, vdenorm_cutoff, _CMP_LT_OS), vf10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_rr1_p5_x96_acc3()
222 _mm256_storeu_ps(output + 80, vf10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_rr1_p5_x96_acc3()
236 vacc1 = _mm256_add_ps(vacc1, vf10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_rr1_p5_x96_acc3()
|
| D | avx2-rr1-p5-x96-acc6.c | 199 __m256 vf10 = _mm256_fmadd_ps(vt10, vp10, vs10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_rr1_p5_x96_acc6() local
212 vf10 = _mm256_andnot_ps(_mm256_cmp_ps(vx10, vdenorm_cutoff, _CMP_LT_OS), vf10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_rr1_p5_x96_acc6()
225 _mm256_storeu_ps(output + 80, vf10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_rr1_p5_x96_acc6()
239 vacc4 = _mm256_add_ps(vacc4, vf10); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_rr1_p5_x96_acc6()
|
| D | avx512f-rr1-p5-scalef-x192-acc2.c | 169 const __m512 vf10 = _mm512_scalef_ps(vp10, vn10); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_rr1_p5_scalef_x192_acc2() local
182 _mm512_storeu_ps(output + 160, vf10); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_rr1_p5_scalef_x192_acc2()
196 vacc0 = _mm512_add_ps(vacc0, vf10); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_rr1_p5_scalef_x192_acc2()
|
| D | avx512f-rr1-p5-scalef-x192.c | 168 const __m512 vf10 = _mm512_scalef_ps(vp10, vn10); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_rr1_p5_scalef_x192() local
181 _mm512_storeu_ps(output + 160, vf10); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_rr1_p5_scalef_x192()
195 vacc0 = _mm512_add_ps(vacc0, vf10); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_rr1_p5_scalef_x192()
|
| D | avx512f-rr1-p5-scalef-x192-acc3.c | 170 const __m512 vf10 = _mm512_scalef_ps(vp10, vn10); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_rr1_p5_scalef_x192_acc3() local
183 _mm512_storeu_ps(output + 160, vf10); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_rr1_p5_scalef_x192_acc3()
197 vacc1 = _mm512_add_ps(vacc1, vf10); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_rr1_p5_scalef_x192_acc3()
|
| D | avx512f-rr1-p5-scalef-x192-acc6.c | 173 const __m512 vf10 = _mm512_scalef_ps(vp10, vn10); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_rr1_p5_scalef_x192_acc6() local
186 _mm512_storeu_ps(output + 160, vf10); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_rr1_p5_scalef_x192_acc6()
200 vacc4 = _mm512_add_ps(vacc4, vf10); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_rr1_p5_scalef_x192_acc6()
|
| /external/XNNPACK/src/f32-raddexpminusmax/gen/ |
| D | avx2-p5-x96.c | 224 __m256 vf10 = _mm256_fmadd_ps(vt10, vp10, vs10); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96() local
239 vf10 = _mm256_andnot_ps(_mm256_cmp_ps(vx10, vdenorm_cutoff, _CMP_LT_OS), vf10); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96()
253 vacc0 = _mm256_add_ps(vacc0, vf10); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96()
|
| D | avx2-p5-x96-acc6.c | 229 __m256 vf10 = _mm256_fmadd_ps(vt10, vp10, vs10); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc6() local
244 vf10 = _mm256_andnot_ps(_mm256_cmp_ps(vx10, vdenorm_cutoff, _CMP_LT_OS), vf10); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc6()
258 vacc4 = _mm256_add_ps(vacc4, vf10); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc6()
|
| D | avx2-p5-x96-acc2.c | 225 __m256 vf10 = _mm256_fmadd_ps(vt10, vp10, vs10); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc2() local
240 vf10 = _mm256_andnot_ps(_mm256_cmp_ps(vx10, vdenorm_cutoff, _CMP_LT_OS), vf10); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc2()
254 vacc0 = _mm256_add_ps(vacc0, vf10); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc2()
|
| D | avx2-p5-x96-acc3.c | 226 __m256 vf10 = _mm256_fmadd_ps(vt10, vp10, vs10); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc3() local
241 vf10 = _mm256_andnot_ps(_mm256_cmp_ps(vx10, vdenorm_cutoff, _CMP_LT_OS), vf10); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc3()
255 vacc1 = _mm256_add_ps(vacc1, vf10); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc3()
|
| D | avx512f-p5-scalef-x192-acc2.c | 192 const __m512 vf10 = _mm512_scalef_ps(vp10, vn10); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2() local
206 vacc0 = _mm512_add_ps(vacc0, vf10); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
|
| D | avx512f-p5-scalef-x192-acc3.c | 193 const __m512 vf10 = _mm512_scalef_ps(vp10, vn10); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3() local
207 vacc1 = _mm512_add_ps(vacc1, vf10); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
|
| D | avx512f-p5-scalef-x192.c | 191 const __m512 vf10 = _mm512_scalef_ps(vp10, vn10); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192() local
205 vacc0 = _mm512_add_ps(vacc0, vf10); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
|
| D | avx512f-p5-scalef-x192-acc6.c | 196 const __m512 vf10 = _mm512_scalef_ps(vp10, vn10); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6() local
210 vacc4 = _mm512_add_ps(vacc4, vf10); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
|