98     vn0123 = wasm_f32x4_sub(vn0123, vmagic_bias);  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
99     vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias);  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
100     vn89AB = wasm_f32x4_sub(vn89AB, vmagic_bias);  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
101     vnCDEF = wasm_f32x4_sub(vnCDEF, vmagic_bias);  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
118     vp0123 = wasm_f32x4_sub(vt0123, wasm_f32x4_mul(vp0123, vt0123));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
119     vp4567 = wasm_f32x4_sub(vt4567, wasm_f32x4_mul(vp4567, vt4567));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
120     vp89AB = wasm_f32x4_sub(vt89AB, wasm_f32x4_mul(vp89AB, vt89AB));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
121     vpCDEF = wasm_f32x4_sub(vtCDEF, wasm_f32x4_mul(vpCDEF, vtCDEF));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
123     const v128_t vy0123 = wasm_f32x4_sub(vs0123, wasm_f32x4_mul(vs0123, vp0123));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
124     const v128_t vy4567 = wasm_f32x4_sub(vs4567, wasm_f32x4_mul(vs4567, vp4567));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
125     const v128_t vy89AB = wasm_f32x4_sub(vs89AB, wasm_f32x4_mul(vs89AB, vp89AB));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
126     const v128_t vyCDEF = wasm_f32x4_sub(vsCDEF, wasm_f32x4_mul(vsCDEF, vpCDEF));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
143     vf0123 = wasm_v128_bitselect(vf0123, wasm_f32x4_sub(vone, vf0123), wasm_i32x4_shr(vx0123, 31));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
144     vf4567 = wasm_v128_bitselect(vf4567, wasm_f32x4_sub(vone, vf4567), wasm_i32x4_shr(vx4567, 31));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
145     vf89AB = wasm_v128_bitselect(vf89AB, wasm_f32x4_sub(vone, vf89AB), wasm_i32x4_shr(vx89AB, 31));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
146     vfCDEF = wasm_v128_bitselect(vfCDEF, wasm_f32x4_sub(vone, vfCDEF), wasm_i32x4_shr(vxCDEF, 31));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
173     vn = wasm_f32x4_sub(vn, vmagic_bias);  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
179     vp = wasm_f32x4_sub(vt, wasm_f32x4_mul(vp, vt));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
181     const v128_t vy = wasm_f32x4_sub(vs, wasm_f32x4_mul(vs, vp));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
186     vf = wasm_v128_bitselect(vf, wasm_f32x4_sub(vone, vf), wasm_i32x4_shr(vx, 31));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
209     vn = wasm_f32x4_sub(vn, vmagic_bias);  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
215     vp = wasm_f32x4_sub(vt, wasm_f32x4_mul(vp, vt));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
217     const v128_t vy = wasm_f32x4_sub(vs, wasm_f32x4_mul(vs, vp));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()
222     vf = wasm_v128_bitselect(vf, wasm_f32x4_sub(vone, vf), wasm_i32x4_shr(vx, 31));  in xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16()