| /external/XNNPACK/src/f32-raddexpminusmax/gen/ |
| D | avx512f-p5-scalef-x192-acc2.c | 28 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2() local
100 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
101 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
102 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
103 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
104 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
105 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
106 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
107 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
108 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
[all …]
|
| D | avx512f-p5-scalef-x192.c | 28 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192() local
99 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
100 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
101 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
102 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
103 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
104 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
105 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
106 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
107 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
[all …]
|
| D | avx512f-p5-scalef-x192-acc3.c | 28 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3() local
101 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
102 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
103 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
104 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
105 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
106 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
107 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
108 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
109 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
[all …]
|
| D | avx512f-p5-scalef-x192-acc6.c | 28 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6() local
104 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
105 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
106 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
107 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
108 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
109 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
110 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
111 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
112 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6()
[all …]
|
| D | avx512f-p5-scalef-x160-acc5.c | 28 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5() local
95 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
96 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
97 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
98 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
99 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
100 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
101 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
102 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
103 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5()
[all …]
|
| D | avx512f-p5-scalef-x160.c | 28 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160() local
91 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
92 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
93 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
94 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
95 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
96 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
97 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
98 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
99 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160()
[all …]
|
| D | avx512f-p5-scalef-x160-acc2.c | 28 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2() local
92 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
93 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
94 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
95 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
96 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
97 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
98 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
99 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
100 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
[all …]
|
| D | avx512f-p5-scalef-x144.c | 28 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144() local
87 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
88 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
89 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
90 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
91 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
92 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
93 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
94 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
95 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144()
[all …]
|
| D | avx512f-p5-scalef-x144-acc3.c | 28 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3() local
89 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
90 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
91 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
92 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
93 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
94 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
95 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
96 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
97 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
[all …]
|
| /external/XNNPACK/src/f32-vscaleexpminusmax/gen/ |
| D | avx512f-p5-scalef-x192.c | 29 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192() local
100 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
101 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
102 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
103 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
104 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
105 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
106 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
107 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
108 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192()
[all …]
|
| D | avx512f-p5-scalef-x176.c | 29 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176() local
96 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
97 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
98 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
99 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
100 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
101 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
102 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
103 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
104 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176()
[all …]
|
| D | avx512f-p5-scalef-x160.c | 29 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160() local
92 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
93 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
94 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
95 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
96 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
97 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
98 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
99 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
100 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160()
[all …]
|
| D | avx512f-p5-scalef-x144.c | 29 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144() local
88 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
89 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
90 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
91 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
92 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
93 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
94 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
95 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
96 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144()
[all …]
|
| /external/XNNPACK/src/f32-vscaleextexp/gen/ |
| D | avx512f-p5-scalef-x192.c | 30 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192() local
87 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
88 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
89 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
90 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
91 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
92 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
93 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
94 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
95 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192()
[all …]
|
| D | avx512f-p5-scalef-x176.c | 30 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176() local
84 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
85 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
86 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
87 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
88 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
89 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
90 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
91 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
92 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176()
[all …]
|
| D | avx512f-p5-scalef-x160.c | 30 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160() local
81 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
82 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
83 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
84 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
85 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
86 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
87 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
88 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
89 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160()
[all …]
|
| D | avx512f-p5-scalef-x144.c | 30 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144() local
78 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
79 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
80 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
81 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
82 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
83 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
84 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
85 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
86 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144()
[all …]
|
| /external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
| D | avx512f-p5-scalef-x192-acc2.c | 29 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2() local
101 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
102 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
103 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
104 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
105 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
106 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
107 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
108 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
109 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
[all …]
|
| D | avx512f-p5-scalef-x192-acc3.c | 29 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3() local
102 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
103 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
104 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
105 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
106 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
107 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
108 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
109 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
110 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3()
[all …]
|
| D | avx512f-p5-scalef-x192.c | 29 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192() local
100 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
101 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
102 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
103 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
104 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
105 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
106 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
107 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
108 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
[all …]
|
| D | avx512f-p5-scalef-x160-acc2.c | 29 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2() local
93 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
94 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
95 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
96 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
97 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
98 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
99 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
100 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
101 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2()
[all …]
|
| D | avx512f-p5-scalef-x160.c | 29 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160() local
92 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
93 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
94 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
95 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
96 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
97 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
98 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
99 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
100 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160()
[all …]
|
| D | avx512f-p5-scalef-x144.c | 29 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144() local
88 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
89 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
90 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
91 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
92 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
93 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
94 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
95 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
96 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144()
[all …]
|
| D | avx512f-p5-scalef-x144-acc3.c | 29 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3() local
90 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
91 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
92 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
93 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
94 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
95 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
96 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
97 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
98 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3()
[all …]
|
| /external/XNNPACK/src/f32-raddextexp/gen/ |
| D | avx512f-p5-scalef-x192.c | 29 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192() local
87 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
88 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
89 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
90 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
91 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
92 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
93 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
94 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
95 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
[all …]
|