111     vn0123 = wasm_f32x4_sub(vn0123, vmagic_bias);  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
112     vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias);  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
113     vn89AB = wasm_f32x4_sub(vn89AB, vmagic_bias);  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
114     vnCDEF = wasm_f32x4_sub(vnCDEF, vmagic_bias);  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
115     vnGHIJ = wasm_f32x4_sub(vnGHIJ, vmagic_bias);  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
135     vp0123 = wasm_f32x4_sub(vt0123, wasm_f32x4_mul(vp0123, vt0123));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
136     vp4567 = wasm_f32x4_sub(vt4567, wasm_f32x4_mul(vp4567, vt4567));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
137     vp89AB = wasm_f32x4_sub(vt89AB, wasm_f32x4_mul(vp89AB, vt89AB));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
138     vpCDEF = wasm_f32x4_sub(vtCDEF, wasm_f32x4_mul(vpCDEF, vtCDEF));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
139     vpGHIJ = wasm_f32x4_sub(vtGHIJ, wasm_f32x4_mul(vpGHIJ, vtGHIJ));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
141     const v128_t vy0123 = wasm_f32x4_sub(vs0123, wasm_f32x4_mul(vs0123, vp0123));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
142     const v128_t vy4567 = wasm_f32x4_sub(vs4567, wasm_f32x4_mul(vs4567, vp4567));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
143     const v128_t vy89AB = wasm_f32x4_sub(vs89AB, wasm_f32x4_mul(vs89AB, vp89AB));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
144     const v128_t vyCDEF = wasm_f32x4_sub(vsCDEF, wasm_f32x4_mul(vsCDEF, vpCDEF));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
145     const v128_t vyGHIJ = wasm_f32x4_sub(vsGHIJ, wasm_f32x4_mul(vsGHIJ, vpGHIJ));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
165     vf0123 = wasm_v128_bitselect(vf0123, wasm_f32x4_sub(vone, vf0123), wasm_i32x4_shr(vx0123, 31));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
166     vf4567 = wasm_v128_bitselect(vf4567, wasm_f32x4_sub(vone, vf4567), wasm_i32x4_shr(vx4567, 31));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
167     vf89AB = wasm_v128_bitselect(vf89AB, wasm_f32x4_sub(vone, vf89AB), wasm_i32x4_shr(vx89AB, 31));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
168     vfCDEF = wasm_v128_bitselect(vfCDEF, wasm_f32x4_sub(vone, vfCDEF), wasm_i32x4_shr(vxCDEF, 31));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
169     vfGHIJ = wasm_v128_bitselect(vfGHIJ, wasm_f32x4_sub(vone, vfGHIJ), wasm_i32x4_shr(vxGHIJ, 31));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
197     vn = wasm_f32x4_sub(vn, vmagic_bias);  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
203     vp = wasm_f32x4_sub(vt, wasm_f32x4_mul(vp, vt));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
205     const v128_t vy = wasm_f32x4_sub(vs, wasm_f32x4_mul(vs, vp));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
210     vf = wasm_v128_bitselect(vf, wasm_f32x4_sub(vone, vf), wasm_i32x4_shr(vx, 31));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
233     vn = wasm_f32x4_sub(vn, vmagic_bias);  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
239     vp = wasm_f32x4_sub(vt, wasm_f32x4_mul(vp, vt));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
241     const v128_t vy = wasm_f32x4_sub(vs, wasm_f32x4_mul(vs, vp));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()
246     vf = wasm_v128_bitselect(vf, wasm_f32x4_sub(vone, vf), wasm_i32x4_shr(vx, 31));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20()