124     vn0123 = wasm_f32x4_sub(vn0123, vmagic_bias);  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
125     vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias);  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
126     vn89AB = wasm_f32x4_sub(vn89AB, vmagic_bias);  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
127     vnCDEF = wasm_f32x4_sub(vnCDEF, vmagic_bias);  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
128     vnGHIJ = wasm_f32x4_sub(vnGHIJ, vmagic_bias);  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
129     vnKLMN = wasm_f32x4_sub(vnKLMN, vmagic_bias);  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
152     vp0123 = wasm_f32x4_sub(vt0123, wasm_f32x4_mul(vp0123, vt0123));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
153     vp4567 = wasm_f32x4_sub(vt4567, wasm_f32x4_mul(vp4567, vt4567));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
154     vp89AB = wasm_f32x4_sub(vt89AB, wasm_f32x4_mul(vp89AB, vt89AB));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
155     vpCDEF = wasm_f32x4_sub(vtCDEF, wasm_f32x4_mul(vpCDEF, vtCDEF));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
156     vpGHIJ = wasm_f32x4_sub(vtGHIJ, wasm_f32x4_mul(vpGHIJ, vtGHIJ));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
157     vpKLMN = wasm_f32x4_sub(vtKLMN, wasm_f32x4_mul(vpKLMN, vtKLMN));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
159     const v128_t vy0123 = wasm_f32x4_sub(vs0123, wasm_f32x4_mul(vs0123, vp0123));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
160     const v128_t vy4567 = wasm_f32x4_sub(vs4567, wasm_f32x4_mul(vs4567, vp4567));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
161     const v128_t vy89AB = wasm_f32x4_sub(vs89AB, wasm_f32x4_mul(vs89AB, vp89AB));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
162     const v128_t vyCDEF = wasm_f32x4_sub(vsCDEF, wasm_f32x4_mul(vsCDEF, vpCDEF));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
163     const v128_t vyGHIJ = wasm_f32x4_sub(vsGHIJ, wasm_f32x4_mul(vsGHIJ, vpGHIJ));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
164     const v128_t vyKLMN = wasm_f32x4_sub(vsKLMN, wasm_f32x4_mul(vsKLMN, vpKLMN));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
187     vf0123 = wasm_v128_bitselect(vf0123, wasm_f32x4_sub(vone, vf0123), wasm_i32x4_shr(vx0123, 31));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
188     vf4567 = wasm_v128_bitselect(vf4567, wasm_f32x4_sub(vone, vf4567), wasm_i32x4_shr(vx4567, 31));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
189     vf89AB = wasm_v128_bitselect(vf89AB, wasm_f32x4_sub(vone, vf89AB), wasm_i32x4_shr(vx89AB, 31));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
190     vfCDEF = wasm_v128_bitselect(vfCDEF, wasm_f32x4_sub(vone, vfCDEF), wasm_i32x4_shr(vxCDEF, 31));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
191     vfGHIJ = wasm_v128_bitselect(vfGHIJ, wasm_f32x4_sub(vone, vfGHIJ), wasm_i32x4_shr(vxGHIJ, 31));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
192     vfKLMN = wasm_v128_bitselect(vfKLMN, wasm_f32x4_sub(vone, vfKLMN), wasm_i32x4_shr(vxKLMN, 31));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
221     vn = wasm_f32x4_sub(vn, vmagic_bias);  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
227     vp = wasm_f32x4_sub(vt, wasm_f32x4_mul(vp, vt));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
229     const v128_t vy = wasm_f32x4_sub(vs, wasm_f32x4_mul(vs, vp));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
234     vf = wasm_v128_bitselect(vf, wasm_f32x4_sub(vone, vf), wasm_i32x4_shr(vx, 31));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
257     vn = wasm_f32x4_sub(vn, vmagic_bias);  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
263     vp = wasm_f32x4_sub(vt, wasm_f32x4_mul(vp, vt));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
265     const v128_t vy = wasm_f32x4_sub(vs, wasm_f32x4_mul(vs, vp));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()
270     vf = wasm_v128_bitselect(vf, wasm_f32x4_sub(vone, vf), wasm_i32x4_shr(vx, 31));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24()