/external/XNNPACK/src/f32-raddextexp/gen/ |
D | avx512f-p5-scalef-x128-acc2.c | 151 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2() local 181 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_acce01); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2() local
|
D | avx512f-p5-scalef-x160-acc2.c | 171 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2() local 205 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_acce01); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2() local
|
D | avx512f-p5-scalef-x144-acc3.c | 163 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3() local 200 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_acce012); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3() local
|
D | avx512f-p5-scalef-x128-acc4.c | 155 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4() local 193 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_acce0123); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4() local
|
D | avx512f-p5-scalef-x192-acc2.c | 191 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2() local 229 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_acce01); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2() local
|
D | avx2-p5-x64-acc2.c | 159 const __m256 vdelta_acce0 = _mm256_max_ps(_mm256_sub_ps(vacce0, vmax_e0), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc2() local 205 const __m256 vdelta_acce0 = _mm256_max_ps(_mm256_sub_ps(vacce0, vmax_acce01), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc2() local
|
D | avx512f-p5-scalef-x192-acc3.c | 193 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3() local 236 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_acce012); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3() local
|
D | avx2-p5-x80-acc2.c | 179 const __m256 vdelta_acce0 = _mm256_max_ps(_mm256_sub_ps(vacce0, vmax_e0), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2() local 231 const __m256 vdelta_acce0 = _mm256_max_ps(_mm256_sub_ps(vacce0, vmax_acce01), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2() local
|
D | avx512f-p5-scalef-x160-acc5.c | 177 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5() local 224 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_acce01234); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5() local
|
D | avx2-p5-x64-acc4.c | 163 const __m256 vdelta_acce0 = _mm256_max_ps(_mm256_sub_ps(vacce0, vmax_e0), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc4() local 219 const __m256 vdelta_acce0 = _mm256_max_ps(_mm256_sub_ps(vacce0, vmax_acce0123), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc4() local
|
D | avx2-p5-x72-acc3.c | 171 const __m256 vdelta_acce0 = _mm256_max_ps(_mm256_sub_ps(vacce0, vmax_e0), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x72_acc3() local 226 const __m256 vdelta_acce0 = _mm256_max_ps(_mm256_sub_ps(vacce0, vmax_acce012), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x72_acc3() local
|
D | avx512f-p5-scalef-x192-acc6.c | 199 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6() local 253 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_acce012345); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6() local
|
D | avx2-p5-x96-acc2.c | 199 const __m256 vdelta_acce0 = _mm256_max_ps(_mm256_sub_ps(vacce0, vmax_e0), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc2() local 257 const __m256 vdelta_acce0 = _mm256_max_ps(_mm256_sub_ps(vacce0, vmax_acce01), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc2() local
|
D | avx2-p5-x80-acc5.c | 185 const __m256 vdelta_acce0 = _mm256_max_ps(_mm256_sub_ps(vacce0, vmax_e0), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc5() local 253 const __m256 vdelta_acce0 = _mm256_max_ps(_mm256_sub_ps(vacce0, vmax_acce01234), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc5() local
|
D | avx2-p5-x96-acc3.c | 201 const __m256 vdelta_acce0 = _mm256_max_ps(_mm256_sub_ps(vacce0, vmax_e0), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc3() local 265 const __m256 vdelta_acce0 = _mm256_max_ps(_mm256_sub_ps(vacce0, vmax_acce012), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc3() local
|
D | avx2-p5-x96-acc6.c | 207 const __m256 vdelta_acce0 = _mm256_max_ps(_mm256_sub_ps(vacce0, vmax_e0), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc6() local 285 const __m256 vdelta_acce0 = _mm256_max_ps(_mm256_sub_ps(vacce0, vmax_acce012345), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc6() local
|
D | avx512f-p5-scalef-x128.c | 149 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128() local
|
D | avx512f-p5-scalef-x144.c | 159 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144() local
|
D | avx512f-p5-scalef-x160.c | 169 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160() local
|
D | avx2-p5-x64.c | 157 const __m256 vdelta_acce0 = _mm256_max_ps(_mm256_sub_ps(vacce0, vmax_e0), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x64() local
|
D | avx2-p5-x72.c | 167 const __m256 vdelta_acce0 = _mm256_max_ps(_mm256_sub_ps(vacce0, vmax_e0), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x72() local
|
D | avx512f-p5-scalef-x192.c | 189 const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_e0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192() local
|
D | avx2-p5-x80.c | 177 const __m256 vdelta_acce0 = _mm256_max_ps(_mm256_sub_ps(vacce0, vmax_e0), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x80() local
|
D | avx2-p5-x96.c | 197 const __m256 vdelta_acce0 = _mm256_max_ps(_mm256_sub_ps(vacce0, vmax_e0), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96() local
|