248         const float32x4_t vk21c2x0123 = vld1q_f32(w + 72);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()  local
250         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk21c2x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
251         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk21c2x0123, vget_low_f32(vi4x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
253         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk21c2x0123, vget_high_f32(vi2x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
254         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk21c2x0123, vget_high_f32(vi4x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
491         const float32x4_t vk21c2x0123 = vld1q_f32(w + 72);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()  local
493         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c2x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
494         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c2x0123, vget_low_f32(vi4x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
682         const float32x4_t vk21c2x0123 = vld1q_f32(w + 72);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()  local
684         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c2x0123, vi2x1, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
685         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c2x0123, vi4x1, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()