56       const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4;  in xnn_f32_prelu_ukernel__neon_2x16()  local
69       float32x4_t vacc1x0123 = vmulq_f32(vi1x0123, vw0123);  in xnn_f32_prelu_ukernel__neon_2x16()
70       const uint32x4_t vm1x0123 = vcltq_s32(vreinterpretq_s32_f32(vi1x0123), vmovq_n_s32(0));  in xnn_f32_prelu_ukernel__neon_2x16()
82       vacc1x0123 = vbslq_f32(vm1x0123, vacc1x0123, vi1x0123);  in xnn_f32_prelu_ukernel__neon_2x16()
101       const float32x4_t vi1x0123 = vld1q_f32(i1);  in xnn_f32_prelu_ukernel__neon_2x16()  local
106       float32x4_t vacc1x0123 = vmulq_f32(vi1x0123, vw0123);  in xnn_f32_prelu_ukernel__neon_2x16()
107       const uint32x4_t vm1x0123 = vcltq_s32(vreinterpretq_s32_f32(vi1x0123), vmovq_n_s32(0));  in xnn_f32_prelu_ukernel__neon_2x16()
110       vacc1x0123 = vbslq_f32(vm1x0123, vacc1x0123, vi1x0123);  in xnn_f32_prelu_ukernel__neon_2x16()
120       const float32x4_t vi1x0123 = vld1q_f32(i1);  in xnn_f32_prelu_ukernel__neon_2x16()  local
125       float32x4_t vacc1x0123 = vmulq_f32(vi1x0123, vw0123);  in xnn_f32_prelu_ukernel__neon_2x16()
126       const uint32x4_t vm1x0123 = vcltq_s32(vreinterpretq_s32_f32(vi1x0123), vmovq_n_s32(0));  in xnn_f32_prelu_ukernel__neon_2x16()
129       vacc1x0123 = vbslq_f32(vm1x0123, vacc1x0123, vi1x0123);  in xnn_f32_prelu_ukernel__neon_2x16()