/external/XNNPACK/src/f32-raddextexp/gen/ |
D | avx512f-p5-scalef-x128.c | 199 const __m512 vdelta_acce = _mm512_sub_ps(vacce, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128() local 231 const __m512 vdelta_acce = _mm512_sub_ps(vacce, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128() local 240 const __m512 vdelta_acce = _mm512_sub_ps(vacce, _mm512_set1_ps(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128() local
|
D | avx512f-p5-scalef-x144.c | 211 const __m512 vdelta_acce = _mm512_sub_ps(vacce, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144() local 243 const __m512 vdelta_acce = _mm512_sub_ps(vacce, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144() local 252 const __m512 vdelta_acce = _mm512_sub_ps(vacce, _mm512_set1_ps(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144() local
|
D | avx512f-p5-scalef-x128-acc2.c | 210 const __m512 vdelta_acce = _mm512_sub_ps(vacce, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2() local 242 const __m512 vdelta_acce = _mm512_sub_ps(vacce, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2() local 251 const __m512 vdelta_acce = _mm512_sub_ps(vacce, _mm512_set1_ps(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2() local
|
D | avx512f-p5-scalef-x160.c | 223 const __m512 vdelta_acce = _mm512_sub_ps(vacce, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160() local 255 const __m512 vdelta_acce = _mm512_sub_ps(vacce, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160() local 264 const __m512 vdelta_acce = _mm512_sub_ps(vacce, _mm512_set1_ps(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160() local
|
D | avx2-p5-x64.c | 224 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x64() local 269 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x64() local 286 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x64() local
|
D | avx512f-p5-scalef-x160-acc2.c | 234 const __m512 vdelta_acce = _mm512_sub_ps(vacce, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2() local 266 const __m512 vdelta_acce = _mm512_sub_ps(vacce, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2() local 275 const __m512 vdelta_acce = _mm512_sub_ps(vacce, _mm512_set1_ps(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2() local
|
D | avx512f-p5-scalef-x144-acc3.c | 231 const __m512 vdelta_acce = _mm512_sub_ps(vacce, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3() local 263 const __m512 vdelta_acce = _mm512_sub_ps(vacce, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3() local 272 const __m512 vdelta_acce = _mm512_sub_ps(vacce, _mm512_set1_ps(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3() local
|
D | avx2-p5-x72.c | 237 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x72() local 282 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x72() local 299 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x72() local
|
D | avx512f-p5-scalef-x128-acc4.c | 226 const __m512 vdelta_acce = _mm512_sub_ps(vacce, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4() local 258 const __m512 vdelta_acce = _mm512_sub_ps(vacce, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4() local 267 const __m512 vdelta_acce = _mm512_sub_ps(vacce, _mm512_set1_ps(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4() local
|
D | avx512f-p5-scalef-x192.c | 247 const __m512 vdelta_acce = _mm512_sub_ps(vacce, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192() local 279 const __m512 vdelta_acce = _mm512_sub_ps(vacce, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192() local 288 const __m512 vdelta_acce = _mm512_sub_ps(vacce, _mm512_set1_ps(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192() local
|
D | avx512f-p5-scalef-x192-acc2.c | 258 const __m512 vdelta_acce = _mm512_sub_ps(vacce, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2() local 290 const __m512 vdelta_acce = _mm512_sub_ps(vacce, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2() local 299 const __m512 vdelta_acce = _mm512_sub_ps(vacce, _mm512_set1_ps(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2() local
|
D | avx2-p5-x80.c | 250 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x80() local 295 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x80() local 312 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x80() local
|
D | avx2-p5-x64-acc2.c | 239 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc2() local 284 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc2() local 301 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc2() local
|
D | avx512f-p5-scalef-x192-acc3.c | 267 const __m512 vdelta_acce = _mm512_sub_ps(vacce, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3() local 299 const __m512 vdelta_acce = _mm512_sub_ps(vacce, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3() local 308 const __m512 vdelta_acce = _mm512_sub_ps(vacce, _mm512_set1_ps(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3() local
|
D | avx2-p5-x96.c | 276 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96() local 321 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96() local 338 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96() local
|
D | avx2-p5-x80-acc2.c | 265 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2() local 310 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2() local 327 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2() local
|
D | avx512f-p5-scalef-x160-acc5.c | 259 const __m512 vdelta_acce = _mm512_sub_ps(vacce, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5() local 291 const __m512 vdelta_acce = _mm512_sub_ps(vacce, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5() local 300 const __m512 vdelta_acce = _mm512_sub_ps(vacce, _mm512_set1_ps(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5() local
|
D | avx2-p5-x64-acc4.c | 259 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc4() local 304 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc4() local 321 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc4() local
|
D | avx2-p5-x72-acc3.c | 263 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x72_acc3() local 308 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x72_acc3() local 325 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x72_acc3() local
|
D | avx512f-p5-scalef-x192-acc6.c | 290 const __m512 vdelta_acce = _mm512_sub_ps(vacce, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6() local 322 const __m512 vdelta_acce = _mm512_sub_ps(vacce, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6() local 331 const __m512 vdelta_acce = _mm512_sub_ps(vacce, _mm512_set1_ps(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6() local
|
D | avx2-p5-x96-acc2.c | 291 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc2() local 336 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc2() local 353 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc2() local
|
D | avx2-p5-x80-acc5.c | 296 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc5() local 341 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc5() local 358 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc5() local
|
D | avx2-p5-x96-acc3.c | 302 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc3() local 347 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc3() local 364 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc3() local
|
D | avx2-p5-x96-acc6.c | 331 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc6() local 376 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc6() local 393 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc6() local
|