/external/XNNPACK/src/f32-raddextexp/gen/ |
D | avx512f-p5-scalef-x128.c | 200 const __m512 vdelta_e = _mm512_sub_ps(vn, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128() local 232 const __m512 vdelta_e = _mm512_sub_ps(vn, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128() local
|
D | avx512f-p5-scalef-x144.c | 212 const __m512 vdelta_e = _mm512_sub_ps(vn, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144() local 244 const __m512 vdelta_e = _mm512_sub_ps(vn, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144() local
|
D | avx512f-p5-scalef-x128-acc2.c | 211 const __m512 vdelta_e = _mm512_sub_ps(vn, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2() local 243 const __m512 vdelta_e = _mm512_sub_ps(vn, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2() local
|
D | avx512f-p5-scalef-x160.c | 224 const __m512 vdelta_e = _mm512_sub_ps(vn, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160() local 256 const __m512 vdelta_e = _mm512_sub_ps(vn, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160() local
|
D | avx2-p5-x64.c | 225 const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x64() local 268 const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x64() local
|
D | avx512f-p5-scalef-x160-acc2.c | 235 const __m512 vdelta_e = _mm512_sub_ps(vn, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2() local 267 const __m512 vdelta_e = _mm512_sub_ps(vn, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2() local
|
D | avx512f-p5-scalef-x144-acc3.c | 232 const __m512 vdelta_e = _mm512_sub_ps(vn, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3() local 264 const __m512 vdelta_e = _mm512_sub_ps(vn, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3() local
|
D | avx2-p5-x72.c | 238 const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x72() local 281 const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x72() local
|
D | avx512f-p5-scalef-x128-acc4.c | 227 const __m512 vdelta_e = _mm512_sub_ps(vn, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4() local 259 const __m512 vdelta_e = _mm512_sub_ps(vn, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4() local
|
D | avx512f-p5-scalef-x192.c | 248 const __m512 vdelta_e = _mm512_sub_ps(vn, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192() local 280 const __m512 vdelta_e = _mm512_sub_ps(vn, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192() local
|
D | avx512f-p5-scalef-x192-acc2.c | 259 const __m512 vdelta_e = _mm512_sub_ps(vn, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2() local 291 const __m512 vdelta_e = _mm512_sub_ps(vn, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2() local
|
D | avx2-p5-x80.c | 251 const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x80() local 294 const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x80() local
|
D | avx2-p5-x64-acc2.c | 240 const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc2() local 283 const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc2() local
|
D | avx512f-p5-scalef-x192-acc3.c | 268 const __m512 vdelta_e = _mm512_sub_ps(vn, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3() local 300 const __m512 vdelta_e = _mm512_sub_ps(vn, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3() local
|
D | avx2-p5-x96.c | 277 const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96() local 320 const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96() local
|
D | avx2-p5-x80-acc2.c | 266 const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2() local 309 const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2() local
|
D | avx512f-p5-scalef-x160-acc5.c | 260 const __m512 vdelta_e = _mm512_sub_ps(vn, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5() local 292 const __m512 vdelta_e = _mm512_sub_ps(vn, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5() local
|
D | avx2-p5-x64-acc4.c | 260 const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc4() local 303 const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc4() local
|
D | avx2-p5-x72-acc3.c | 264 const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x72_acc3() local 307 const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x72_acc3() local
|
D | avx512f-p5-scalef-x192-acc6.c | 291 const __m512 vdelta_e = _mm512_sub_ps(vn, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6() local 323 const __m512 vdelta_e = _mm512_sub_ps(vn, vmax_e); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6() local
|
D | avx2-p5-x96-acc2.c | 292 const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc2() local 335 const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc2() local
|
D | avx2-p5-x80-acc5.c | 297 const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc5() local 340 const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc5() local
|
D | avx2-p5-x96-acc3.c | 303 const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc3() local 346 const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc3() local
|
D | avx2-p5-x96-acc6.c | 332 const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc6() local 375 const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc6() local
|