184         const float32x4_t vk01c0x0123 = vld1q_f32(w + 40);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()  local
186         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk01c0x0123, vget_high_f32(vi0x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
187         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c0x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
189         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk01c0x0123, vget_low_f32(vi0x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
190         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk01c0x0123, vget_low_f32(vi2x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
444         const float32x4_t vk01c0x0123 = vld1q_f32(w + 40);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()  local
446         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c0x0123, vget_high_f32(vi0x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
447         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
635         const float32x4_t vk01c0x0123 = vld1q_f32(w + 40);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()  local
637         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c0x0123, vget_high_f32(vi0x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
638         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()