| /external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
| D | avx512f-p5-scalef-x192.c | 87 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
88 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
89 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
90 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
91 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
92 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
93 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
94 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
95 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
96 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
[all …]
|
| D | avx512f-p5-scalef-x192-acc2.c | 88 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
89 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
90 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
91 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
92 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
93 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
94 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
95 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
96 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
97 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
[all …]
|
| D | avx512f-p5-scalef-x192-acc3.c | 89 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
90 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
91 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
92 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
93 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
94 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
95 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
96 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
97 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
98 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
[all …]
|
| D | avx512f-p5-scalef-x160.c | 81 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
82 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
83 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
84 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
85 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
86 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
87 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
88 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
89 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
90 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
[all …]
|
| D | avx512f-p5-scalef-x160-acc2.c | 82 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
83 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
84 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
85 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
86 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
87 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
88 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
89 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
90 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
91 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
[all …]
|
| D | avx512f-p5-scalef-x192-acc6.c | 92 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
93 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
94 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
95 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
96 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
97 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
98 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
99 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
100 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
101 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
[all …]
|
| D | avx512f-p5-scalef-x144-acc3.c | 80 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
81 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
82 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
83 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
84 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
85 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
86 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
87 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
88 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
90 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
[all …]
|
| D | avx512f-p5-scalef-x144.c | 78 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
79 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
80 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
81 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
82 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
83 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
84 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
85 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
86 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
88 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
[all …]
|
| /external/XNNPACK/src/f32-raddexpminusmax/gen/ |
| D | avx512f-p5-scalef-x192.c | 86 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
87 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
88 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
89 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
90 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
91 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
92 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
93 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
94 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
95 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
[all …]
|
| D | avx512f-p5-scalef-x192-acc3.c | 88 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
89 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
90 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
91 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
92 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
93 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
94 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
95 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
96 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
97 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
[all …]
|
| D | avx512f-p5-scalef-x192-acc6.c | 91 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
92 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
93 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
94 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
95 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
96 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
97 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
98 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
99 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
100 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
[all …]
|
| D | avx512f-p5-scalef-x192-acc2.c | 87 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
88 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
89 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
90 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
91 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
92 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
93 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
94 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
95 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
96 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
[all …]
|
| D | avx512f-p5-scalef-x160.c | 80 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
81 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
82 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
83 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
84 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
85 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
86 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
87 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
88 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
89 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
[all …]
|
| D | avx512f-p5-scalef-x160-acc2.c | 81 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
82 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
83 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
84 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
85 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
86 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
87 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
88 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
89 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
90 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
[all …]
|
| D | avx512f-p5-scalef-x160-acc5.c | 84 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
85 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
86 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
87 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
88 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
89 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
90 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
91 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
92 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
93 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
[all …]
|
| D | avx512f-p5-scalef-x144.c | 77 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
78 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
79 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
80 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
81 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
82 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
83 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
84 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
85 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
87 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
[all …]
|
| /external/XNNPACK/src/f32-vscaleexpminusmax/gen/ |
| D | avx512f-p5-scalef-x192.c | 87 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
88 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
89 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
90 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
91 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
92 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
93 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
94 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
95 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
96 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
[all …]
|
| D | avx512f-p5-scalef-x176.c | 84 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
85 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
86 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
87 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
88 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
89 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
90 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
91 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
92 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
93 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
[all …]
|
| D | avx512f-p5-scalef-x160.c | 81 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
82 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
83 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
84 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
85 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
86 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
87 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
88 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
89 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
90 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
[all …]
|
| /external/XNNPACK/src/f32-vscaleextexp/gen/ |
| D | avx512f-p5-scalef-x192.c | 74 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
75 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
76 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
77 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
78 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
79 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
80 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
81 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
82 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
83 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
[all …]
|
| D | avx512f-p5-scalef-x176.c | 72 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
73 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
74 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
75 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
76 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
77 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
78 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
79 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
80 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
81 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
[all …]
|
| D | avx512f-p5-scalef-x160.c | 70 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
71 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
72 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
73 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
74 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
75 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
76 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
77 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
78 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
79 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
[all …]
|
| D | avx512f-p5-scalef-x144.c | 68 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
69 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
70 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
71 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
72 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
73 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
74 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
75 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
76 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
78 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
[all …]
|
| /external/XNNPACK/src/f32-raddextexp/gen/ |
| D | avx512f-p5-scalef-x192.c | 74 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
75 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
76 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
77 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
78 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
79 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
80 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
81 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
82 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
83 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
[all …]
|
| D | avx512f-p5-scalef-x192-acc2.c | 76 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
77 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
78 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
79 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
80 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
81 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
82 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
83 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
84 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
85 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
[all …]
|