28   const __m512 vhalf = _mm512_set1_ps(params->fma.half);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
30     const __m512 vx0 = _mm512_loadu_ps(x);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
31     const __m512 vx1 = _mm512_loadu_ps(x + 16);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
32     const __m512 vx2 = _mm512_loadu_ps(x + 32);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
33     const __m512 vx3 = _mm512_loadu_ps(x + 48);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
34     const __m512 vx4 = _mm512_loadu_ps(x + 64);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
35     const __m512 vx5 = _mm512_loadu_ps(x + 80);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
36     const __m512 vx6 = _mm512_loadu_ps(x + 96);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
37     const __m512 vx7 = _mm512_loadu_ps(x + 112);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
40     const __m512 vrsqrtx0 = _mm512_rsqrt14_ps(vx0);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
41     const __m512 vrsqrtx1 = _mm512_rsqrt14_ps(vx1);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
42     const __m512 vrsqrtx2 = _mm512_rsqrt14_ps(vx2);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
43     const __m512 vrsqrtx3 = _mm512_rsqrt14_ps(vx3);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
44     const __m512 vrsqrtx4 = _mm512_rsqrt14_ps(vx4);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
45     const __m512 vrsqrtx5 = _mm512_rsqrt14_ps(vx5);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
46     const __m512 vrsqrtx6 = _mm512_rsqrt14_ps(vx6);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
47     const __m512 vrsqrtx7 = _mm512_rsqrt14_ps(vx7);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
49     __m512 vsqrtx0 = _mm512_mul_ps(vrsqrtx0, vx0);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
50     __m512 vhalfrsqrtx0 = _mm512_mul_ps(vrsqrtx0, vhalf);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
51     __m512 vsqrtx1 = _mm512_mul_ps(vrsqrtx1, vx1);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
52     __m512 vhalfrsqrtx1 = _mm512_mul_ps(vrsqrtx1, vhalf);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
53     __m512 vsqrtx2 = _mm512_mul_ps(vrsqrtx2, vx2);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
54     __m512 vhalfrsqrtx2 = _mm512_mul_ps(vrsqrtx2, vhalf);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
55     __m512 vsqrtx3 = _mm512_mul_ps(vrsqrtx3, vx3);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
56     __m512 vhalfrsqrtx3 = _mm512_mul_ps(vrsqrtx3, vhalf);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
57     __m512 vsqrtx4 = _mm512_mul_ps(vrsqrtx4, vx4);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
58     __m512 vhalfrsqrtx4 = _mm512_mul_ps(vrsqrtx4, vhalf);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
59     __m512 vsqrtx5 = _mm512_mul_ps(vrsqrtx5, vx5);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
60     __m512 vhalfrsqrtx5 = _mm512_mul_ps(vrsqrtx5, vhalf);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
61     __m512 vsqrtx6 = _mm512_mul_ps(vrsqrtx6, vx6);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
62     __m512 vhalfrsqrtx6 = _mm512_mul_ps(vrsqrtx6, vhalf);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
63     __m512 vsqrtx7 = _mm512_mul_ps(vrsqrtx7, vx7);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
64     __m512 vhalfrsqrtx7 = _mm512_mul_ps(vrsqrtx7, vhalf);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
66     const __m512 vresidual0 = _mm512_fnmadd_ps(vsqrtx0, vhalfrsqrtx0, vhalf);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
67     const __m512 vresidual1 = _mm512_fnmadd_ps(vsqrtx1, vhalfrsqrtx1, vhalf);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
68     const __m512 vresidual2 = _mm512_fnmadd_ps(vsqrtx2, vhalfrsqrtx2, vhalf);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
69     const __m512 vresidual3 = _mm512_fnmadd_ps(vsqrtx3, vhalfrsqrtx3, vhalf);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
70     const __m512 vresidual4 = _mm512_fnmadd_ps(vsqrtx4, vhalfrsqrtx4, vhalf);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
71     const __m512 vresidual5 = _mm512_fnmadd_ps(vsqrtx5, vhalfrsqrtx5, vhalf);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
72     const __m512 vresidual6 = _mm512_fnmadd_ps(vsqrtx6, vhalfrsqrtx6, vhalf);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
73     const __m512 vresidual7 = _mm512_fnmadd_ps(vsqrtx7, vhalfrsqrtx7, vhalf);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
92     const __m512 vadjustment0 = _mm512_fnmadd_ps(vsqrtx0, vsqrtx0, vx0);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
93     const __m512 vadjustment1 = _mm512_fnmadd_ps(vsqrtx1, vsqrtx1, vx1);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
94     const __m512 vadjustment2 = _mm512_fnmadd_ps(vsqrtx2, vsqrtx2, vx2);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
95     const __m512 vadjustment3 = _mm512_fnmadd_ps(vsqrtx3, vsqrtx3, vx3);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
96     const __m512 vadjustment4 = _mm512_fnmadd_ps(vsqrtx4, vsqrtx4, vx4);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
97     const __m512 vadjustment5 = _mm512_fnmadd_ps(vsqrtx5, vsqrtx5, vx5);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
98     const __m512 vadjustment6 = _mm512_fnmadd_ps(vsqrtx6, vsqrtx6, vx6);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
99     const __m512 vadjustment7 = _mm512_fnmadd_ps(vsqrtx7, vsqrtx7, vx7);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
101     const __m512 vy0 = _mm512_fmadd_ps(vhalfrsqrtx0, vadjustment0, vsqrtx0);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
102     const __m512 vy1 = _mm512_fmadd_ps(vhalfrsqrtx1, vadjustment1, vsqrtx1);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
103     const __m512 vy2 = _mm512_fmadd_ps(vhalfrsqrtx2, vadjustment2, vsqrtx2);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
104     const __m512 vy3 = _mm512_fmadd_ps(vhalfrsqrtx3, vadjustment3, vsqrtx3);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
105     const __m512 vy4 = _mm512_fmadd_ps(vhalfrsqrtx4, vadjustment4, vsqrtx4);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
106     const __m512 vy5 = _mm512_fmadd_ps(vhalfrsqrtx5, vadjustment5, vsqrtx5);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
107     const __m512 vy6 = _mm512_fmadd_ps(vhalfrsqrtx6, vadjustment6, vsqrtx6);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
108     const __m512 vy7 = _mm512_fmadd_ps(vhalfrsqrtx7, vadjustment7, vsqrtx7);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
121     const __m512 vx = _mm512_loadu_ps(x);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
124     const __m512 vrsqrtx = _mm512_rsqrt14_ps(vx);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
125     __m512 vsqrtx = _mm512_mul_ps(vrsqrtx, vx);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
126     __m512 vhalfrsqrtx = _mm512_mul_ps(vrsqrtx, vhalf);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
127     const __m512 vresidual = _mm512_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
130     const __m512 vadjustment = _mm512_fnmadd_ps(vsqrtx, vsqrtx, vx);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
131     const __m512 vy = _mm512_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
143     const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
144     const __m512 vrsqrtx = _mm512_rsqrt14_ps(vx);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
145     __m512 vsqrtx = _mm512_mul_ps(vrsqrtx, vx);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
146     __m512 vhalfrsqrtx = _mm512_mul_ps(vrsqrtx, vhalf);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
147     const __m512 vresidual = _mm512_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
150     const __m512 vadjustment = _mm512_fnmadd_ps(vsqrtx, vsqrtx, vx);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
151     const __m512 vy = _mm512_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);  in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()