27   const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
28   const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
29   const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
31   const __m512 vc0 = _mm512_set1_ps(1.0f);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
32   const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
33   const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
34   const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
35   const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
36   const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
38   const __m512 vminus_inf = _mm512_set1_ps(-INFINITY);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
40   __m512 vaccv0 = _mm512_setzero_ps();  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
41   __m512 vaccv1 = _mm512_setzero_ps();  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
42   __m512 vaccv2 = _mm512_setzero_ps();  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
43   __m512 vaccv3 = _mm512_setzero_ps();  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
44   __m512 vaccv4 = _mm512_setzero_ps();  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
45   __m512 vaccv5 = _mm512_setzero_ps();  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
46   __m512 vacce0 = vminus_inf;  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
47   __m512 vacce1 = vminus_inf;  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
48   __m512 vacce2 = vminus_inf;  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
49   __m512 vacce3 = vminus_inf;  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
50   __m512 vacce4 = vminus_inf;  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
51   __m512 vacce5 = vminus_inf;  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
54     const __m512 vx0 = _mm512_loadu_ps(x);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
55     const __m512 vx1 = _mm512_loadu_ps(x + 16);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
56     const __m512 vx2 = _mm512_loadu_ps(x + 32);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
57     const __m512 vx3 = _mm512_loadu_ps(x + 48);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
58     const __m512 vx4 = _mm512_loadu_ps(x + 64);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
59     const __m512 vx5 = _mm512_loadu_ps(x + 80);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
60     const __m512 vx6 = _mm512_loadu_ps(x + 96);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
61     const __m512 vx7 = _mm512_loadu_ps(x + 112);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
62     const __m512 vx8 = _mm512_loadu_ps(x + 128);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
63     const __m512 vx9 = _mm512_loadu_ps(x + 144);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
64     const __m512 vx10 = _mm512_loadu_ps(x + 160);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
65     const __m512 vx11 = _mm512_loadu_ps(x + 176);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
69     const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
70     const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
71     const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
72     const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
73     const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
74     const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
75     const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
76     const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
77     const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
78     const __m512 vn9 = _mm512_roundscale_ps(_mm512_mul_ps(vx9, vlog2e), 0);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
79     const __m512 vn10 = _mm512_roundscale_ps(_mm512_mul_ps(vx10, vlog2e), 0);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
80     const __m512 vn11 = _mm512_roundscale_ps(_mm512_mul_ps(vx11, vlog2e), 0);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
84     __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
85     __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
86     __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
87     __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
88     __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
89     __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
90     __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
91     __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
92     __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
93     __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
94     __m512 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_hi, vx10);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
95     __m512 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_hi, vx11);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
111     __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
112     __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
113     __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
114     __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
115     __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
116     __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
117     __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
118     __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
119     __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
120     __m512 vp9 = _mm512_fmadd_ps(vc5, vt9, vc4);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
121     __m512 vp10 = _mm512_fmadd_ps(vc5, vt10, vc4);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
122     __m512 vp11 = _mm512_fmadd_ps(vc5, vt11, vc4);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
186     __m512 vmax_e0 = _mm512_max_ps(vacce0, vn0);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
187     __m512 vmax_e1 = _mm512_max_ps(vacce1, vn1);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
188     __m512 vmax_e2 = _mm512_max_ps(vacce2, vn2);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
189     __m512 vmax_e3 = _mm512_max_ps(vacce3, vn3);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
190     __m512 vmax_e4 = _mm512_max_ps(vacce4, vn4);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
191     __m512 vmax_e5 = _mm512_max_ps(vacce5, vn5);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
199     const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_e0);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
200     const __m512 vdelta_acce1 = _mm512_sub_ps(vacce1, vmax_e1);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
201     const __m512 vdelta_acce2 = _mm512_sub_ps(vacce2, vmax_e2);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
202     const __m512 vdelta_acce3 = _mm512_sub_ps(vacce3, vmax_e3);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
203     const __m512 vdelta_acce4 = _mm512_sub_ps(vacce4, vmax_e4);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
204     const __m512 vdelta_acce5 = _mm512_sub_ps(vacce5, vmax_e5);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
205     const __m512 vdelta_e0 = _mm512_sub_ps(vn0, vmax_e0);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
206     const __m512 vdelta_e1 = _mm512_sub_ps(vn1, vmax_e1);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
207     const __m512 vdelta_e2 = _mm512_sub_ps(vn2, vmax_e2);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
208     const __m512 vdelta_e3 = _mm512_sub_ps(vn3, vmax_e3);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
209     const __m512 vdelta_e4 = _mm512_sub_ps(vn4, vmax_e4);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
210     const __m512 vdelta_e5 = _mm512_sub_ps(vn5, vmax_e5);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
211     const __m512 vdelta_e6 = _mm512_sub_ps(vn6, vmax_e0);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
212     const __m512 vdelta_e7 = _mm512_sub_ps(vn7, vmax_e1);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
213     const __m512 vdelta_e8 = _mm512_sub_ps(vn8, vmax_e2);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
214     const __m512 vdelta_e9 = _mm512_sub_ps(vn9, vmax_e3);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
215     const __m512 vdelta_e10 = _mm512_sub_ps(vn10, vmax_e4);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
216     const __m512 vdelta_e11 = _mm512_sub_ps(vn11, vmax_e5);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
247   const __m512 vmax_acce01 = _mm512_max_ps(vacce0, vacce1);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
248   const __m512 vmax_acce23 = _mm512_max_ps(vacce2, vacce3);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
249   const __m512 vmax_acce45 = _mm512_max_ps(vacce4, vacce5);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
250   const __m512 vmax_acce0123 = _mm512_max_ps(vmax_acce01, vmax_acce23);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
251   const __m512 vmax_acce012345 = _mm512_max_ps(vmax_acce0123, vmax_acce45);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
253   const __m512 vdelta_acce0 = _mm512_sub_ps(vacce0, vmax_acce012345);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
254   const __m512 vdelta_acce1 = _mm512_sub_ps(vacce1, vmax_acce012345);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
255   const __m512 vdelta_acce2 = _mm512_sub_ps(vacce2, vmax_acce012345);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
256   const __m512 vdelta_acce3 = _mm512_sub_ps(vacce3, vmax_acce012345);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
257   const __m512 vdelta_acce4 = _mm512_sub_ps(vacce4, vmax_acce012345);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
258   const __m512 vdelta_acce5 = _mm512_sub_ps(vacce5, vmax_acce012345);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
260   __m512 vaccv = _mm512_scalef_ps(vaccv0, vdelta_acce0);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
266   __m512 vacce = vmax_acce012345;  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
270     const __m512 vx = _mm512_loadu_ps(x);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
274     const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
278     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
282     __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
289     const __m512 vmax_e = _mm512_max_ps(vacce, vn);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
290     const __m512 vdelta_acce = _mm512_sub_ps(vacce, vmax_e);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
291     const __m512 vdelta_e = _mm512_sub_ps(vn, vmax_e);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
303     const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
306     const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
310     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
314     __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
321     const __m512 vmax_e = _mm512_mask_max_ps(vacce, vmask, vacce, vn);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
322     const __m512 vdelta_acce = _mm512_sub_ps(vacce, vmax_e);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
323     const __m512 vdelta_e = _mm512_sub_ps(vn, vmax_e);  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()
331   const __m512 vdelta_acce = _mm512_sub_ps(vacce, _mm512_set1_ps(vmax_acce));  in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc6()